Merge pull request 'SPEC-018 Phase 1: managed agent as LocalSystem service host' (#7) from feat/spec-018-service-host into main

H: thread the SCM cooperative-stop flag into the connected session loop
(run_with_tray) via a new Option<&Arc<AtomicBool>> param. The flag was only
observed by the outer run_agent reconnect loop, which never runs while a
session is connected, so an SCM Stop/Shutdown left the service Running until
force-kill. The inner loop now checks it each tick, closes the WS cleanly, and
returns the SERVICE_STOP sentinel that the outer loop maps to a graceful stop.
The new param is optional: attended/viewer/interactive callers pass None and
behave exactly as before.

M: wrap the managed-agent runtime block_on in catch_unwind(AssertUnwindSafe) so
a panic in the agent future cannot unwind across the extern "system" service
entry (UB/abort). A caught panic becomes an Err -> ServiceExitCode::ServiceSpecific(1)
so SCM recovery engages cleanly.

L1: replace the fixed 2s sleep after delete() on reinstall with a bounded retry
on CreateService returning ERROR_SERVICE_MARKED_FOR_DELETE (1072), gated on
having actually deleted a prior instance.

L2: clarify the --elevated -> force_user_install mapping (comment only).

N1: add a clap-metadata test pinning the service-run subcommand name to
SERVICE_RUN_ARG, cross-linked from the existing literal test.

N2: correct the service doc comments now that graceful stop interrupts the
connected case too.

Verified on Windows host: cargo fmt --check, clippy -D warnings, release build
(x86_64-pc-windows-msvc), and cargo test (58 passed) all green.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

.gitea/workflows/release.yml

@@ -27,6 +27,15 @@ on:
   # computes the next semver from conventional commits at dispatch time.
   # build-and-test.yml remains the automatic PR/push CI gate.
   workflow_dispatch:
+    inputs:
+      channel:
+        description: 'Release channel (stable = full versioned release; beta = signed prerelease test build, no version bump/changelog)'
+        required: true
+        default: 'stable'
+        type: choice
+        options:
+          - stable
+          - beta
 
 jobs:
   # ---------------------------------------------------------------------------
@@ -36,8 +45,11 @@ jobs:
     name: Version + Changelog
     runs-on: ubuntu-latest
     outputs:
-      version: ${{ steps.bump.outputs.version }}
+      # Coalesce across the stable (bump) and beta (beta) paths: exactly one of them runs per
-      released: ${{ steps.bump.outputs.released }}
+      # dispatch, so the first non-empty value wins. prerelease is 'true' only on the beta path.
+      version: ${{ steps.bump.outputs.version || steps.beta.outputs.version }}
+      released: ${{ steps.bump.outputs.released || steps.beta.outputs.released }}
+      prerelease: ${{ steps.beta.outputs.prerelease || 'false' }}
     steps:
       - name: Checkout (full history + tags)
         uses: actions/checkout@v4
@@ -59,7 +71,8 @@ jobs:
           fi
 
       - name: Install git-cliff
-        if: steps.guard.outputs.skip != 'true'
+        # Stable-only: beta produces no changelog, so git-cliff is unnecessary on the beta path.
+        if: steps.guard.outputs.skip != 'true' && github.event.inputs.channel == 'stable'
         run: |
           set -euo pipefail
           CLIFF_VERSION="2.6.1"
@@ -72,12 +85,16 @@ jobs:
 
       - name: Determine next version and bump components
         id: bump
-        if: steps.guard.outputs.skip != 'true'
+        # Stable-only: the beta path (id: beta) handles versioning without a manifest bump/commit.
+        if: steps.guard.outputs.skip != 'true' && github.event.inputs.channel == 'stable'
         run: |
           set -euo pipefail
 
           # ----- locate the last release tag (vX.Y.Z) -----
-          LAST_TAG="$(git tag --list 'v*' --sort=-v:refname | head -n1 || true)"
+          # Match ONLY strict final-release tags (vMAJOR.MINOR.PATCH). Beta tags look like
+          # v0.3.0-beta.7; if one of those were picked up here it would corrupt the next stable
+          # base version, so prerelease tags are explicitly excluded from this lookup.
+          LAST_TAG="$(git tag --list 'v*' --sort=-v:refname | grep -E '^v[0-9]+\.[0-9]+\.[0-9]+$' | head -n1 || true)"
           if [ -z "${LAST_TAG}" ]; then
             echo "[INFO] No prior release tag found; baseline is current manifest version."
             BASE_VERSION="$(grep -m1 '^version' agent/Cargo.toml | sed -E 's/.*"([0-9]+\.[0-9]+\.[0-9]+)".*/\1/')"
@@ -186,8 +203,39 @@ jobs:
             sed -i -E "0,/^version = \"[0-9]+\.[0-9]+\.[0-9]+\"/s//version = \"${NEXT}\"/" Cargo.toml || true
           fi
 
+      - name: Beta channel - tag prerelease build (no bump, no commit, no changelog)
+        id: beta
+        # Beta-only path. Reuses the IDENTICAL downstream build + sign + publish jobs, but does
+        # NOT compute a semver bump, mutate any manifest, generate a changelog, or make a release
+        # commit. It just tags the CURRENT HEAD with a unique prerelease version so the Windows
+        # build job can check out `ref: v${VER}` exactly as it does for stable.
+        if: github.event.inputs.channel == 'beta' && steps.guard.outputs.skip != 'true'
+        run: |
+          set -euo pipefail
+
+          # Base version is read straight from the agent manifest — NOT bumped, NOT written back.
+          BASE="$(grep -m1 '^version' agent/Cargo.toml | sed -E 's/.*"([0-9]+\.[0-9]+\.[0-9]+)".*/\1/')"
+          # GITHUB_RUN_NUMBER guarantees a unique prerelease suffix without counting existing tags.
+          VER="${BASE}-beta.${GITHUB_RUN_NUMBER}"
+          echo "[INFO] Beta build version: ${VER} (base ${BASE}, run ${GITHUB_RUN_NUMBER})"
+
+          # Tag the current HEAD (no release commit). Push the tag so build-agent-windows can
+          # check out ref: v${VER}.
+          git config user.name "guruconnect-ci"
+          git config user.email "ci@azcomputerguru.com"
+          # Beta tags are disposable test markers; force makes re-running a failed beta dispatch idempotent (re-run reuses GITHUB_RUN_NUMBER, so the tag already exists).
+          git tag -f "v${VER}"
+          REMOTE="https://${{ secrets.CI_PUSH_TOKEN }}@git.azcomputerguru.com/${GITHUB_REPOSITORY}.git"
+          git push --force "${REMOTE}" "v${VER}"
+          echo "[OK] Pushed beta prerelease tag v${VER}"
+
+          echo "version=${VER}" >> "$GITHUB_OUTPUT"
+          echo "released=true" >> "$GITHUB_OUTPUT"
+          echo "prerelease=true" >> "$GITHUB_OUTPUT"
+
       - name: Generate changelog (git-cliff)
-        if: steps.guard.outputs.skip != 'true' && steps.bump.outputs.released == 'true'
+        # Stable-only: beta produces no changelog artifact.
+        if: steps.guard.outputs.skip != 'true' && steps.bump.outputs.released == 'true' && github.event.inputs.channel == 'stable'
         env:
           VERSION: ${{ steps.bump.outputs.version }}
         run: |
@@ -232,7 +280,10 @@ jobs:
 
           # Re-derive the set of changed components (same logic as the bump step). On the first
           # release (no prior tag) all components are considered changed.
-          LAST_TAG="$(git tag --list 'v*' --sort=-v:refname | head -n1 || true)"
+          # Match ONLY strict final-release tags (vMAJOR.MINOR.PATCH); exclude beta prerelease
+          # tags (v0.3.0-beta.7) so the changelog diff range is taken against the last real
+          # release, not an intervening beta build.
+          LAST_TAG="$(git tag --list 'v*' --sort=-v:refname | grep -E '^v[0-9]+\.[0-9]+\.[0-9]+$' | head -n1 || true)"
           if [ -z "${LAST_TAG}" ]; then
             CHANGED_FILES="$(git ls-files)"
             FIRST_RELEASE=true
@@ -252,7 +303,8 @@ jobs:
           fi
 
       - name: Commit release + create tag
-        if: steps.guard.outputs.skip != 'true' && steps.bump.outputs.released == 'true'
+        # Stable-only: beta tags HEAD directly in the beta step and never makes a release commit.
+        if: steps.guard.outputs.skip != 'true' && steps.bump.outputs.released == 'true' && github.event.inputs.channel == 'stable'
         env:
           VERSION: ${{ steps.bump.outputs.version }}
         run: |
@@ -276,7 +328,8 @@ jobs:
           echo "[OK] Pushed release commit and tag v${VERSION}"
 
       - name: Upload changelog artifact
-        if: steps.guard.outputs.skip != 'true' && steps.bump.outputs.released == 'true'
+        # Stable-only: there is no changelog on the beta path, so nothing to upload.
+        if: steps.guard.outputs.skip != 'true' && steps.bump.outputs.released == 'true' && github.event.inputs.channel == 'stable'
         uses: actions/upload-artifact@v3
         with:
           name: changelog
@@ -445,6 +498,9 @@ jobs:
           echo "sha256=${SUM}" >> "$GITHUB_OUTPUT"
 
       - name: Download changelog artifact
+        # Stable-only: the beta path uploads no `changelog` artifact. The release-creation step
+        # already guards on `[ -f changelog-artifact/CHANGELOG.md ]`, so skipping this is safe.
+        if: github.event.inputs.channel == 'stable'
         uses: actions/download-artifact@v3
         with:
           name: changelog
@@ -472,17 +528,26 @@ jobs:
         env:
           VERSION: ${{ needs.version.outputs.version }}
           SHA256: ${{ steps.sha.outputs.sha256 }}
+          # PRERELEASE is 'true' on the beta path, 'false' on stable; drives the Gitea release flag.
+          PRERELEASE: ${{ needs.version.outputs.prerelease }}
           GITEA_TOKEN: ${{ secrets.CI_PUSH_TOKEN }}
         run: |
           set -euo pipefail
           API_BASE="https://git.azcomputerguru.com/api/v1/repos/${GITHUB_REPOSITORY}"
           TAG="v${VERSION}"
-          echo "[INFO] Creating Gitea release ${TAG} on ${GITHUB_REPOSITORY}"
+          echo "[INFO] Creating Gitea release ${TAG} on ${GITHUB_REPOSITORY} (prerelease=${PRERELEASE})"
 
-          BODY="$(printf 'GuruConnect %s\n\nSHA-256 (guruconnect.exe): %s\n\nSee CHANGELOG.md and /api/changelog for details.' "${TAG}" "${SHA256}")"
+          # Beta builds get a clear "prerelease test build" note in the body; the -beta.N suffix
+          # is already carried in TAG, so the release name "Release v..." needs no extra handling.
+          if [ "${PRERELEASE}" = "true" ]; then
+            BODY="$(printf 'GuruConnect %s (PRERELEASE / beta test build)\n\nSHA-256 (guruconnect.exe): %s\n\nSigned via Azure Trusted Signing. Not a stable release — no changelog/version bump.' "${TAG}" "${SHA256}")"
+          else
+            BODY="$(printf 'GuruConnect %s\n\nSHA-256 (guruconnect.exe): %s\n\nSee CHANGELOG.md and /api/changelog for details.' "${TAG}" "${SHA256}")"
+          fi
 
           # Build the JSON payload with python (handles escaping of the multi-line body safely).
-          CREATE_PAYLOAD="$(TAG="$TAG" BODY="$BODY" python3 -c 'import json,os; print(json.dumps({"tag_name": os.environ["TAG"], "name": "Release " + os.environ["TAG"], "body": os.environ["BODY"], "draft": False, "prerelease": False}))')"
+          # prerelease is derived from the PRERELEASE env var (beta -> true, stable -> false).
+          CREATE_PAYLOAD="$(TAG="$TAG" BODY="$BODY" PRERELEASE="$PRERELEASE" python3 -c 'import json,os; print(json.dumps({"tag_name": os.environ["TAG"], "name": "Release " + os.environ["TAG"], "body": os.environ["BODY"], "draft": False, "prerelease": os.environ.get("PRERELEASE","false") == "true"}))')"
 
           RELEASE_JSON="$(curl -fsS -X POST \
             "${API_BASE}/releases" \

Cargo.lock

@@ -1407,7 +1407,7 @@ dependencies = [
 
 [[package]]
 name = "guruconnect"
-version = "0.2.0"
+version = "0.3.0"
 dependencies = [
  "anyhow",
  "bytes",
@@ -1447,7 +1447,7 @@ dependencies = [
 
 [[package]]
 name = "guruconnect-server"
-version = "0.2.0"
+version = "0.3.0"
 dependencies = [
  "anyhow",
  "argon2",

agent/Cargo.toml

@@ -92,6 +92,7 @@ windows = { version = "0.58", features = [
     "Win32_System_Console",
     "Win32_System_Environment",
     "Win32_Security",
+    "Win32_Security_Cryptography",
     "Win32_Storage_FileSystem",
     "Win32_System_Pipes",
     "Win32_System_SystemServices",

agent/src/config.rs

@@ -16,18 +16,39 @@ use uuid::Uuid;
 const MAGIC_MARKER: &[u8] = b"GURUCONFIG";
 
 /// Embedded configuration data (appended to executable)
+///
+/// SPEC-016 Phase B: a managed-install config now carries the per-site
+/// `enrollment_key` + `site_code` so the agent can self-register on first run.
+/// The legacy `api_key` is retained (defaulted) for backward-compat with older
+/// pre-enrollment installers; a fresh site installer carries only the enrollment
+/// credentials and the agent obtains its per-machine `cak_` via `/api/enroll`.
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct EmbeddedConfig {
     /// Server WebSocket URL
     pub server_url: String,
-    /// API key for authentication
+    /// DEPRECATED shared/legacy API key for authentication. Optional — a
-    pub api_key: String,
+    /// SPEC-016 site installer omits it and enrolls for a per-machine `cak_`.
+    #[serde(default)]
+    pub api_key: Option<String>,
+    /// Per-site enrollment key (`cek_`), the low-sensitivity registration gate
+    /// (SPEC-016 §Security). Presented to `/api/enroll`; never logged.
+    #[serde(default)]
+    pub enrollment_key: Option<String>,
+    /// Per-site code identifying which site this installer enrolls into.
+    #[serde(default)]
+    pub site_code: Option<String>,
     /// Company/organization name
     #[serde(default)]
     pub company: Option<String>,
     /// Site/location name
     #[serde(default)]
     pub site: Option<String>,
+    /// Department label (reserved — SPEC-007 AgentStatus parity).
+    #[serde(default)]
+    pub department: Option<String>,
+    /// Device-type label (reserved — SPEC-007 AgentStatus parity).
+    #[serde(default)]
+    pub device_type: Option<String>,
     /// Tags for categorization
     #[serde(default)]
     pub tags: Vec<String>,
@@ -52,9 +73,28 @@ pub struct Config {
     /// Server WebSocket URL (e.g., wss://connect.example.com/ws)
     pub server_url: String,
 
-    /// Agent API key for authentication
+    /// Operating credential used to authenticate the persistent WS connection.
+    ///
+    /// SPEC-016 Phase B: the AUTHORITATIVE credential is a per-machine `cak_`
+    /// obtained at first-run enrollment and stored encrypted at rest (see
+    /// [`crate::credential_store`]); it is loaded into this field before connect.
+    /// A non-empty value carried in config is the DEPRECATED shared/legacy
+    /// `api_key`, kept only for transition compatibility. Empty means "not yet
+    /// enrolled / no credential" — the run-mode wiring must enroll first.
+    #[serde(default)]
     pub api_key: String,
 
+    /// Per-site enrollment key (`cek_`) — present only for a not-yet-enrolled
+    /// managed install. Never persisted to the on-disk TOML (it is install-time
+    /// material, delivered by the site wrapper); never logged.
+    #[serde(skip)]
+    pub enrollment_key: Option<String>,
+
+    /// Per-site code identifying which site to enroll into (paired with
+    /// `enrollment_key`). Not persisted to the on-disk TOML.
+    #[serde(skip)]
+    pub site_code: Option<String>,
+
     /// Unique agent identifier (generated on first run)
     #[serde(default = "generate_agent_id")]
     pub agent_id: String,
@@ -70,6 +110,14 @@ pub struct Config {
     #[serde(default)]
     pub site: Option<String>,
 
+    /// Department label (reserved — SPEC-007 AgentStatus parity).
+    #[serde(default)]
+    pub department: Option<String>,
+
+    /// Device-type label (reserved — SPEC-007 AgentStatus parity).
+    #[serde(default)]
+    pub device_type: Option<String>,
+
     /// Tags for categorization (from embedded config)
     #[serde(default)]
     pub tags: Vec<String>,
@@ -91,6 +139,25 @@ fn generate_agent_id() -> String {
     Uuid::new_v4().to_string()
 }
 
+/// Layer SPEC-016 enrollment material from the environment onto a `Config`.
+///
+/// `GURUCONNECT_ENROLLMENT_KEY` / `GURUCONNECT_SITE_CODE` only OVERRIDE when set
+/// and non-empty, so embedded/install-time values already present on the config
+/// are preserved. Used by the file and env load paths (the embedded path already
+/// carries these from the install blob).
+fn apply_enrollment_env(config: &mut Config) {
+    if let Ok(v) = std::env::var("GURUCONNECT_ENROLLMENT_KEY") {
+        if !v.is_empty() {
+            config.enrollment_key = Some(v);
+        }
+    }
+    if let Ok(v) = std::env::var("GURUCONNECT_SITE_CODE") {
+        if !v.is_empty() {
+            config.site_code = Some(v);
+        }
+    }
+}
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct CaptureConfig {
     /// Target frames per second (1-60)
@@ -317,18 +384,26 @@ impl Config {
             info!("Using embedded configuration");
             let config = Config {
                 server_url: embedded.server_url,
-                api_key: embedded.api_key,
+                // Legacy/shared api_key if the installer carried one; empty
+                // otherwise (the SPEC-016 path enrolls for a per-machine cak_).
+                api_key: embedded.api_key.unwrap_or_default(),
+                enrollment_key: embedded.enrollment_key,
+                site_code: embedded.site_code,
                 agent_id: generate_agent_id(),
                 hostname_override: None,
                 company: embedded.company,
                 site: embedded.site,
+                department: embedded.department,
+                device_type: embedded.device_type,
                 tags: embedded.tags,
                 support_code: None,
                 capture: CaptureConfig::default(),
                 encoding: EncodingConfig::default(),
             };
 
-            // Save to file for persistence (so agent_id is preserved)
+            // Save to file for persistence (so agent_id is preserved). The
+            // #[serde(skip)] enrollment fields are intentionally NOT written to
+            // the on-disk TOML — they are install-time material only.
             let _ = config.save();
             return Ok(config);
         }
@@ -349,8 +424,12 @@ impl Config {
                 let _ = config.save();
             }
 
-            // support_code is always None when loading from file (set via CLI)
+            // support_code is always None when loading from file (set via CLI).
             config.support_code = None;
+            // The enrollment fields are #[serde(skip)], so a file never carries
+            // them; layer them in from the environment for testing / a
+            // file-delivered managed install that supplies them out-of-band.
+            apply_enrollment_env(&mut config);
 
             return Ok(config);
         }
@@ -365,18 +444,23 @@ impl Config {
         let agent_id =
             std::env::var("GURUCONNECT_AGENT_ID").unwrap_or_else(|_| generate_agent_id());
 
-        let config = Config {
+        let mut config = Config {
             server_url,
             api_key,
+            enrollment_key: None,
+            site_code: None,
             agent_id,
             hostname_override: std::env::var("GURUCONNECT_HOSTNAME").ok(),
             company: None,
             site: None,
+            department: None,
+            device_type: None,
             tags: Vec::new(),
             support_code: None,
             capture: CaptureConfig::default(),
             encoding: EncodingConfig::default(),
         };
+        apply_enrollment_env(&mut config);
 
         // Save config with generated agent_id for persistence
         let _ = config.save();
@@ -384,6 +468,34 @@ impl Config {
         Ok(config)
     }
 
+    /// Derive the HTTPS API base (e.g. `https://connect.example.com`) from the
+    /// agent's WebSocket `server_url` (e.g. `wss://connect.example.com/ws/agent`).
+    ///
+    /// `/api/enroll` is REST/HTTPS while the persistent transport is `wss`, so we
+    /// reuse the same host/authority and swap scheme + drop the WS path. Mapping:
+    /// `wss` -> `https`, `ws` -> `http` (dev). Returns an error if `server_url`
+    /// has no parseable host.
+    pub fn https_base(&self) -> Result<String> {
+        let parsed = url::Url::parse(&self.server_url)
+            .with_context(|| format!("invalid server_url: {}", self.server_url))?;
+        let scheme = match parsed.scheme() {
+            "wss" | "https" => "https",
+            "ws" | "http" => "http",
+            other => {
+                return Err(anyhow!(
+                    "unsupported server_url scheme '{other}' (expected ws/wss)"
+                ))
+            }
+        };
+        let host = parsed
+            .host_str()
+            .ok_or_else(|| anyhow!("server_url has no host: {}", self.server_url))?;
+        Ok(match parsed.port() {
+            Some(port) => format!("{scheme}://{host}:{port}"),
+            None => format!("{scheme}://{host}"),
+        })
+    }
+
     /// Get the configuration file path
     fn config_path() -> PathBuf {
         // Check for config in current directory first

agent/src/credential_store.rs

@@ -0,0 +1,413 @@
+//! At-rest storage for the per-machine operating credential (`cak_`).
+//!
+//! SPEC-016 Phase B, item 4 + §Security. The `cak_` minted by `/api/enroll` is
+//! the high-sensitivity, per-machine, independently-revocable operating
+//! credential. It is stored with **two independent layers** (Mike's locked
+//! decision — "BOTH layers"):
+//!
+//!   1. **DPAPI-machine encryption** (`CryptProtectData` with
+//!      `CRYPTPROTECT_LOCAL_MACHINE`): the on-disk bytes are a DPAPI blob keyed to
+//!      THIS machine. A copied/exfiltrated file is inert on any other box — DPAPI
+//!      machine keys do not leave the machine.
+//!   2. **SYSTEM/Administrators-only ACL** on the containing directory + file: a
+//!      non-admin user cannot even read the ciphertext. Inheritance is removed and
+//!      only `SYSTEM` and `BUILTIN\Administrators` are granted full control.
+//!
+//! Local admin / SYSTEM can always recover the value — that is accepted (SPEC-016
+//! §Security): the blast radius of one leaked `cak_` is a single, independently
+//! revocable machine.
+//!
+//! Storage location (chosen over an HKLM value): a file under
+//! `%ProgramData%\GuruConnect\credentials\agent.cak`. Rationale — the agent
+//! already keeps its config and the `machine_uid` fallback seed under
+//! `%ProgramData%\GuruConnect`, so co-locating keeps a single protected
+//! directory; and a directory/file ACL applied via `icacls` is auditable with far
+//! less unsafe FFI than building a registry-key security descriptor by hand. Both
+//! storage shapes are explicitly permitted by the spec.
+//!
+//! SECURITY: the plaintext `cak_` is NEVER logged. Errors describe the operation,
+//! not the value.
+
+#![cfg(windows)]
+
+use anyhow::{anyhow, Context, Result};
+use std::path::PathBuf;
+use thiserror::Error;
+
+/// Failure classes for [`load_cak`], so callers can distinguish an *operational*
+/// problem (the file exists but this process cannot open/read it — e.g. running in
+/// the wrong security context against a SYSTEM-only-ACL'd store) from the real
+/// *tamper / wrong-machine* signal (the file was read successfully but DPAPI
+/// decryption failed).
+///
+/// The distinction matters for the run-mode resolver (`main.rs`):
+///   - [`LoadCakError::Io`] is recoverable/actionable — log it and STOP (do not
+///     silently re-enroll over a store we simply can't read in this context).
+///   - [`LoadCakError::Decrypt`] is a hard tamper signal — STOP, do not re-enroll.
+#[derive(Debug, Error)]
+pub enum LoadCakError {
+    /// The store path could not be resolved (e.g. `%ProgramData%` unset).
+    #[error("could not resolve credential store path: {0}")]
+    Path(String),
+
+    /// An IO/open/read error reaching the stored blob — INCLUDING
+    /// `PermissionDenied` (the running context lacks rights to the SYSTEM-only
+    /// store). Operational, not a tamper signal.
+    #[error("credential store is present but could not be read in this context: {source}")]
+    Io {
+        /// Whether this was specifically an access-denied error (drives the
+        /// run-mode fail-fast guard in `main.rs`).
+        permission_denied: bool,
+        source: std::io::Error,
+    },
+
+    /// The blob was read successfully but DPAPI decryption FAILED — the real
+    /// tamper / wrong-machine / corruption signal. A hard stop; never re-enroll.
+    #[error("stored credential failed to decrypt (wrong machine, tampered, or corrupted): {0}")]
+    Decrypt(String),
+}
+
+/// Directory holding the protected credential file.
+fn credentials_dir() -> Result<PathBuf> {
+    let program_data =
+        std::env::var("ProgramData").context("ProgramData environment variable is not set")?;
+    Ok(PathBuf::from(program_data)
+        .join("GuruConnect")
+        .join("credentials"))
+}
+
+/// Full path to the DPAPI-encrypted `cak_` blob.
+fn cak_path() -> Result<PathBuf> {
+    Ok(credentials_dir()?.join("agent.cak"))
+}
+
+/// Persist `cak` encrypted at rest.
+///
+/// Ordering is security-critical (H2 — TOCTOU): the directory ACL is locked
+/// BEFORE any secret bytes touch the filesystem, and the temp file is written
+/// INSIDE the already-locked directory, so no ciphertext ever exists at a path
+/// carrying an inherited (potentially world-readable) ACL:
+///
+///   1. `create_dir_all(dir)` — ensure the directory exists.
+///   2. `lock_down_acl(dir)`  — remove inherited ACEs and grant SYSTEM +
+///      Administrators full control, made inheritable `(OI)(CI)` so children
+///      created afterward are covered. This is an explicit precondition for the
+///      write that follows — NOT an unstated inheritance assumption.
+///   3. DPAPI-machine-encrypt the plaintext.
+///   4. Write the ciphertext to a temp file inside the now-locked directory, then
+///      rename over the target (atomic-ish replace).
+///   5. `lock_down_acl(file)` — assert the file's own ACL (belt-and-suspenders; the
+///      file already inherits the directory's restrictive ACEs).
+///   6. C1 read-back: immediately attempt [`load_cak`] to PROVE the running
+///      security context can read its own store. If it cannot (e.g. a non-SYSTEM
+///      run wrote a SYSTEM-only store it can no longer read), fail HERE at enroll
+///      time with an actionable error — rather than silently bricking on the next
+///      boot when the steady-state path tries to load it.
+///
+/// Returns an error (never logs the plaintext) on any failure so the caller can
+/// surface it / retry.
+pub fn store_cak(cak: &str) -> Result<()> {
+    // 1 + 2: lock the directory ACL BEFORE writing any secret (H2 / TOCTOU).
+    let dir = credentials_dir()?;
+    std::fs::create_dir_all(&dir)
+        .with_context(|| format!("failed to create credentials dir {dir:?}"))?;
+    lock_down_acl(&dir).context("failed to restrict credentials directory ACL")?;
+
+    // 3: encrypt only after the destination directory is locked down.
+    let ciphertext = dpapi_protect(cak.as_bytes()).context("DPAPI encryption of cak_ failed")?;
+
+    // 4: write the temp file INSIDE the already-locked directory, then rename.
+    let path = cak_path()?;
+    let tmp = path.with_extension("cak.tmp");
+    std::fs::write(&tmp, &ciphertext)
+        .with_context(|| format!("failed to write temp credential file {tmp:?}"))?;
+    std::fs::rename(&tmp, &path)
+        .with_context(|| format!("failed to place credential file {path:?}"))?;
+
+    // 5: assert the file ACL too (the file already inherits the dir's ACEs).
+    lock_down_acl(&path).context("failed to restrict credential file ACL")?;
+
+    // 6: C1 read-back — confirm THIS context can read back what it just wrote.
+    // Catches the "wrote a SYSTEM-only store from a non-SYSTEM context" footgun at
+    // enroll time instead of as a silent brick on the next launch.
+    match load_cak() {
+        Ok(Some(_)) => {
+            tracing::info!("[ENROLL] stored per-machine credential (encrypted at rest)");
+            Ok(())
+        }
+        Ok(None) => Err(anyhow!(
+            "stored the credential but read-back returned nothing — refusing to proceed \
+             with an unverifiable credential store"
+        )),
+        Err(LoadCakError::Io {
+            permission_denied: true,
+            ..
+        }) => Err(anyhow!(
+            "[ENROLL] wrote the credential store but cannot read it back in THIS security \
+             context (access denied). The store is ACL'd to SYSTEM + Administrators by \
+             design; the managed agent must run as the GuruConnect SYSTEM service (see \
+             SPEC-018) to read it. Refusing to leave an unreadable store behind."
+        )),
+        Err(e) => Err(anyhow::Error::new(e)
+            .context("stored the credential but the immediate read-back verification failed")),
+    }
+}
+
+/// Load and decrypt the stored `cak_`, or `Ok(None)` if no credential is stored.
+///
+/// Error classification (M1) — the caller MUST treat these differently:
+///   - `Ok(None)`              -> no store yet (NotFound or empty); enroll is fine.
+///   - [`LoadCakError::Io`]    -> the store exists but is unreadable in this
+///     context (open/read error, INCLUDING access-denied). Operational; the caller
+///     logs it and STOPS — it must NOT silently re-enroll over a store it merely
+///     cannot read here.
+///   - [`LoadCakError::Decrypt`] -> the bytes were read but DPAPI decryption
+///     FAILED (wrong machine / tampered / corrupted). A hard tamper signal; STOP.
+///
+/// Only a successful READ whose decrypt fails is the tamper signal — an IO or
+/// permission error is never conflated with tamper.
+pub fn load_cak() -> std::result::Result<Option<String>, LoadCakError> {
+    let path = cak_path().map_err(|e| LoadCakError::Path(e.to_string()))?;
+    let ciphertext = match std::fs::read(&path) {
+        Ok(bytes) => bytes,
+        Err(e) if e.kind() == std::io::ErrorKind::NotFound => return Ok(None),
+        Err(e) => {
+            let permission_denied = e.kind() == std::io::ErrorKind::PermissionDenied;
+            return Err(LoadCakError::Io {
+                permission_denied,
+                source: e,
+            });
+        }
+    };
+    if ciphertext.is_empty() {
+        return Ok(None);
+    }
+    // Reaching here means the READ succeeded — so a decrypt failure now IS the real
+    // tamper / wrong-machine signal (never conflated with an IO/permission error).
+    let plaintext =
+        dpapi_unprotect(&ciphertext).map_err(|e| LoadCakError::Decrypt(e.to_string()))?;
+    let cak = String::from_utf8(plaintext)
+        .map_err(|e| LoadCakError::Decrypt(format!("decrypted bytes were not valid UTF-8: {e}")))?;
+    if cak.is_empty() {
+        return Ok(None);
+    }
+    Ok(Some(cak))
+}
+
+/// Remove the stored credential (e.g. on revocation / forced re-enroll).
+/// Succeeds if the file is already absent.
+///
+/// Part of the store/load/clear API the spec requires (SPEC-016 item 4). Not yet
+/// called from a code path — the relay-side `cak_` revocation / forced re-enroll
+/// flow that drives it is the deferred SPEC-016 Phase B/D server work (the
+/// `TODO(SPEC-016 Phase B/D): consider revoking existing cak_ on collision` note
+/// in `server/src/api/enroll.rs`) — so it is retained as part of the complete
+/// store API and explicitly allowed dead until that server work lands.
+#[allow(dead_code)]
+pub fn clear_cak() -> Result<()> {
+    let path = cak_path()?;
+    match std::fs::remove_file(&path) {
+        Ok(()) => {
+            tracing::info!("[ENROLL] cleared stored per-machine credential");
+            Ok(())
+        }
+        Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(()),
+        Err(e) => Err(e).with_context(|| format!("failed to remove {path:?}")),
+    }
+}
+
+// ---------------------------------------------------------------------------
+// DPAPI (machine scope)
+// ---------------------------------------------------------------------------
+
+/// DPAPI-machine-encrypt `plaintext` into a self-contained blob.
+fn dpapi_protect(plaintext: &[u8]) -> Result<Vec<u8>> {
+    use windows::Win32::Security::Cryptography::{
+        CryptProtectData, CRYPTPROTECT_LOCAL_MACHINE, CRYPT_INTEGER_BLOB,
+    };
+
+    // CryptProtectData requires a mutable input pointer in the struct, though it
+    // does not modify the bytes; copy into a local Vec to get a *mut without
+    // aliasing the caller's slice.
+    let mut input = plaintext.to_vec();
+    let in_blob = CRYPT_INTEGER_BLOB {
+        cbData: u32::try_from(input.len()).context("plaintext too large for DPAPI")?,
+        pbData: input.as_mut_ptr(),
+    };
+    let mut out_blob = CRYPT_INTEGER_BLOB::default();
+
+    // SAFETY: in_blob points at a valid, sized buffer; out_blob is owned here and
+    // its pbData is allocated by DPAPI (freed via LocalFree below). No prompt
+    // struct / entropy / reserved args.
+    unsafe {
+        CryptProtectData(
+            &in_blob,
+            windows::core::PCWSTR::null(),
+            None,
+            None,
+            None,
+            CRYPTPROTECT_LOCAL_MACHINE,
+            &mut out_blob,
+        )
+        .context("CryptProtectData failed")?;
+    }
+
+    let result = copy_and_free_blob(&out_blob);
+    // Best-effort scrub of the transient plaintext copy.
+    input.iter_mut().for_each(|b| *b = 0);
+
+    result.ok_or_else(|| anyhow!("CryptProtectData returned an empty/invalid blob"))
+}
+
+/// DPAPI-decrypt a blob previously produced by [`dpapi_protect`] on this machine.
+fn dpapi_unprotect(ciphertext: &[u8]) -> Result<Vec<u8>> {
+    use windows::Win32::Security::Cryptography::{
+        CryptUnprotectData, CRYPTPROTECT_LOCAL_MACHINE, CRYPT_INTEGER_BLOB,
+    };
+
+    let mut input = ciphertext.to_vec();
+    let in_blob = CRYPT_INTEGER_BLOB {
+        cbData: u32::try_from(input.len()).context("ciphertext too large for DPAPI")?,
+        pbData: input.as_mut_ptr(),
+    };
+    let mut out_blob = CRYPT_INTEGER_BLOB::default();
+
+    // SAFETY: as in dpapi_protect — valid sized input, owned output freed below.
+    unsafe {
+        CryptUnprotectData(
+            &in_blob,
+            None,
+            None,
+            None,
+            None,
+            CRYPTPROTECT_LOCAL_MACHINE,
+            &mut out_blob,
+        )
+        .context("CryptUnprotectData failed")?;
+    }
+
+    copy_and_free_blob(&out_blob)
+        .ok_or_else(|| anyhow!("CryptUnprotectData returned an empty/invalid blob"))
+}
+
+/// Copy a DPAPI output blob into an owned `Vec` and `LocalFree` the DPAPI buffer.
+///
+/// Returns `Some(bytes)` on success, `None` if the blob is null/empty. Always
+/// frees `pbData` when non-null (DPAPI allocates it with `LocalAlloc`).
+fn copy_and_free_blob(
+    blob: &windows::Win32::Security::Cryptography::CRYPT_INTEGER_BLOB,
+) -> Option<Vec<u8>> {
+    use windows::Win32::Foundation::{LocalFree, HLOCAL};
+
+    if blob.pbData.is_null() {
+        return None;
+    }
+    // SAFETY: DPAPI guarantees pbData points at cbData valid bytes on success.
+    let bytes = unsafe { std::slice::from_raw_parts(blob.pbData, blob.cbData as usize).to_vec() };
+    // SAFETY: pbData was allocated by DPAPI via LocalAlloc; free it once.
+    unsafe {
+        let _ = LocalFree(HLOCAL(blob.pbData as *mut core::ffi::c_void));
+    }
+    if bytes.is_empty() {
+        None
+    } else {
+        Some(bytes)
+    }
+}
+
+// ---------------------------------------------------------------------------
+// ACL hardening
+// ---------------------------------------------------------------------------
+
+/// Restrict `path` (file or directory) to SYSTEM + Administrators full control,
+/// removing inherited ACEs so a permissive parent grant cannot leak read access.
+///
+/// Implemented via `icacls` — the documented, auditable mechanism — rather than
+/// hand-rolling a security descriptor through `SetNamedSecurityInfoW` (hundreds
+/// of lines of SID/ACL FFI). `icacls` ships on every supported Windows target.
+/// A failure here is surfaced (the caller treats inability to lock down the
+/// credential store as a hard error) but the well-known SIDs `*S-1-5-18`
+/// (LocalSystem) and `*S-1-5-32-544` (BUILTIN\Administrators) are language- and
+/// locale-independent, so this does not break on localized Windows.
+fn lock_down_acl(path: &std::path::Path) -> Result<()> {
+    use std::os::windows::process::CommandExt;
+    use std::process::Command;
+
+    const CREATE_NO_WINDOW: u32 = 0x0800_0000;
+
+    let path_str = path
+        .to_str()
+        .ok_or_else(|| anyhow!("credential path is not valid UTF-8: {path:?}"))?;
+
+    // /inheritance:r  -> remove inherited ACEs (drop the permissive parent grant)
+    // /grant:r        -> replace any existing explicit grants for the principal
+    // *S-1-5-18       -> LocalSystem; *S-1-5-32-544 -> BUILTIN\Administrators
+    let output = Command::new("icacls")
+        .arg(path_str)
+        .args([
+            "/inheritance:r",
+            "/grant:r",
+            "*S-1-5-18:(OI)(CI)F",
+            "/grant:r",
+            "*S-1-5-32-544:(OI)(CI)F",
+        ])
+        .creation_flags(CREATE_NO_WINDOW)
+        .output()
+        .context("failed to invoke icacls to harden credential ACL")?;
+
+    if !output.status.success() {
+        // icacls writes its diagnostics to stdout; surface the code only (no
+        // credential material is ever passed to icacls, only the path).
+        return Err(anyhow!(
+            "icacls failed to harden {path_str} (exit {:?})",
+            output.status.code()
+        ));
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// DPAPI round-trips on the same machine: protect then unprotect must recover
+    /// the exact plaintext. (Runs on the build/test host, which IS the same
+    /// machine — the machine-scope key is available to any process here.)
+    #[test]
+    fn dpapi_roundtrip_recovers_plaintext() {
+        let secret = b"cak_test_value_0123456789abcdef";
+        let blob = dpapi_protect(secret).expect("DPAPI protect should succeed on this machine");
+        assert_ne!(
+            blob.as_slice(),
+            secret.as_slice(),
+            "ciphertext must differ from plaintext"
+        );
+        let recovered = dpapi_unprotect(&blob).expect("DPAPI unprotect should succeed");
+        assert_eq!(recovered, secret, "round-trip must recover the exact bytes");
+    }
+
+    /// A non-empty plaintext yields a non-empty, differing blob, and an empty
+    /// input is handled (DPAPI accepts zero-length and round-trips to empty).
+    #[test]
+    fn dpapi_roundtrip_handles_varied_lengths() {
+        for plaintext in [b"x".as_slice(), b"cak_".as_slice(), &[0u8; 256]] {
+            let blob = dpapi_protect(plaintext).expect("protect");
+            let back = dpapi_unprotect(&blob).expect("unprotect");
+            assert_eq!(back.as_slice(), plaintext);
+        }
+    }
+
+    /// Tampering with the ciphertext must make decryption FAIL rather than return
+    /// garbage — DPAPI authenticates its blobs.
+    #[test]
+    fn dpapi_rejects_tampered_blob() {
+        let mut blob = dpapi_protect(b"cak_tamper_target").expect("protect");
+        // Flip a byte in the middle of the blob.
+        let mid = blob.len() / 2;
+        blob[mid] ^= 0xFF;
+        assert!(
+            dpapi_unprotect(&blob).is_err(),
+            "a tampered DPAPI blob must fail to decrypt"
+        );
+    }
+}

agent/src/enroll.rs

@@ -0,0 +1,384 @@
+//! First-run self-enrollment client (SPEC-016 Phase B, item 4).
+//!
+//! When the agent runs as a persistent (`PermanentAgent`) install with NO stored
+//! `cak_` but WITH an `enrollment_key` + `site_code`, it walks through the
+//! public, unauthenticated `POST /api/enroll` door: it presents its site
+//! credentials and its hardware-derived `machine_uid`, and — on success — the
+//! server mints and returns a per-machine `cak_` operating credential exactly
+//! once. The agent persists that `cak_` encrypted at rest
+//! ([`crate::credential_store`]) and connects with it; on every later run it uses
+//! the stored `cak_` directly and never re-enrolls.
+//!
+//! Server contract consumed (must match `server/src/api/enroll.rs`):
+//!   - Request: `{ site_code, enrollment_key, machine_uid, hostname,
+//!     labels:{company,site,department,device_type,tags} }`.
+//!   - `201 Created`  -> new enrollment; body has `key` (the `cak_`).
+//!   - `200 OK`       -> reuse (re-image / re-install); body has `key`.
+//!   - `202 Accepted` -> `collision_pending`; NO key — operator must confirm in
+//!     the dashboard before the endpoint can connect.
+//!   - `401 Unauthorized` -> `ENROLL_REJECTED` (bad/rotated key or unknown site):
+//!     terminal-ish config problem, back off long.
+//!   - `409 Conflict` -> `ENROLL_SITE_CONFLICT` (machine bound to another site):
+//!     terminal-ish, requires the operator reassignment flow; back off long.
+//!   - `429 Too Many Requests` -> rate-limited; back off and retry.
+//!
+//! SECURITY: never log the `enrollment_key` or the minted `cak_`. Only states,
+//! dispositions, and the (non-secret) `machine_uid`/`site_code` are logged.
+
+use anyhow::{anyhow, Context, Result};
+use serde::{Deserialize, Serialize};
+use std::time::Duration;
+
+use crate::config::Config;
+
+/// `POST /api/enroll` request body — mirrors `enroll::EnrollRequest`.
+#[derive(Debug, Serialize)]
+struct EnrollRequest<'a> {
+    site_code: &'a str,
+    enrollment_key: &'a str,
+    machine_uid: &'a str,
+    hostname: &'a str,
+    labels: EnrollLabels<'a>,
+}
+
+/// Labels carried at enrollment — mirrors `enroll::EnrollLabels`.
+#[derive(Debug, Serialize)]
+struct EnrollLabels<'a> {
+    #[serde(skip_serializing_if = "Option::is_none")]
+    company: Option<&'a str>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    site: Option<&'a str>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    department: Option<&'a str>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    device_type: Option<&'a str>,
+    #[serde(skip_serializing_if = "slice_is_empty")]
+    tags: &'a [String],
+}
+
+/// `skip_serializing_if` predicate for the `tags` slice — `Vec::is_empty` cannot
+/// bind a `&&[String]`, so use a slice-typed helper.
+fn slice_is_empty(s: &[String]) -> bool {
+    s.is_empty()
+}
+
+/// `POST /api/enroll` success body — mirrors `enroll::EnrollResponse`.
+#[derive(Debug, Deserialize)]
+struct EnrollResponse {
+    #[allow(dead_code)]
+    machine_id: String,
+    #[serde(default)]
+    key: Option<String>,
+    enrollment_state: String,
+    disposition: String,
+}
+
+/// Backoff after a retryable failure (429 / network / 5xx).
+const RETRYABLE_BACKOFF: Duration = Duration::from_secs(30);
+/// Backoff after a terminal-ish config failure (401 / 409) or collision-pending.
+/// These won't fix themselves without operator action, so retry slowly rather
+/// than hot-looping while still recovering automatically once it IS fixed.
+const TERMINAL_BACKOFF: Duration = Duration::from_secs(300);
+
+/// Drive enrollment until a `cak_` is issued, persisting it into the credential
+/// store on success and loading it into `config.api_key`.
+///
+/// Loops with backoff across retryable failures (it must not give up — a managed
+/// machine left running should eventually enroll once the server/site is healthy)
+/// and across collision-pending (HTTP 202: it keeps re-checking on a slow cadence
+/// until an operator confirms the endpoint in the dashboard and the server begins
+/// issuing a key). Returns `Ok(())` only once a `cak_` is stored. The only `Err`
+/// returns are unrecoverable local faults (missing config, an un-persistable
+/// credential) — network/HTTP failures are retried, never propagated.
+pub async fn run_enrollment(config: &mut Config) -> Result<()> {
+    let site_code = config
+        .site_code
+        .clone()
+        .ok_or_else(|| anyhow!("enrollment requested but no site_code is configured"))?;
+    let enrollment_key = config
+        .enrollment_key
+        .clone()
+        .ok_or_else(|| anyhow!("enrollment requested but no enrollment_key is configured"))?;
+
+    let https_base = config.https_base()?;
+    let machine_uid = crate::identity::machine_uid();
+    let hostname = config.hostname();
+
+    tracing::info!(
+        "[ENROLL] first-run enrollment: site_code={} machine_uid={} hostname={}",
+        site_code,
+        machine_uid,
+        hostname
+    );
+
+    loop {
+        match attempt_enroll(
+            &https_base,
+            &site_code,
+            &enrollment_key,
+            &machine_uid,
+            &hostname,
+            config,
+        )
+        .await
+        {
+            Ok(AttemptResult::Issued(cak)) => {
+                // Persist encrypted-at-rest, then load into the live config so the
+                // transport authenticates with the new per-machine credential.
+                #[cfg(windows)]
+                crate::credential_store::store_cak(&cak)
+                    .context("failed to persist issued cak_ to the credential store")?;
+                config.api_key = cak;
+                // Enrollment material is single-use; drop it so it is not retained
+                // in memory or accidentally reused.
+                config.enrollment_key = None;
+                tracing::info!("[ENROLL] enrollment complete; connecting with per-machine key");
+                return Ok(());
+            }
+            Ok(AttemptResult::Pending) => {
+                tracing::warn!(
+                    "[ENROLL] pending operator confirmation (machine_uid collision); \
+                     this machine cannot connect until confirmed in the dashboard. \
+                     Re-checking in {}s.",
+                    TERMINAL_BACKOFF.as_secs()
+                );
+                tokio::time::sleep(TERMINAL_BACKOFF).await;
+            }
+            Err(AttemptError::Terminal(msg)) => {
+                tracing::error!(
+                    "[ENROLL] enrollment refused (operator action required): {msg}. \
+                     Retrying in {}s.",
+                    TERMINAL_BACKOFF.as_secs()
+                );
+                tokio::time::sleep(TERMINAL_BACKOFF).await;
+            }
+            Err(AttemptError::Retryable(msg)) => {
+                tracing::warn!(
+                    "[ENROLL] transient enrollment failure: {msg}. Retrying in {}s.",
+                    RETRYABLE_BACKOFF.as_secs()
+                );
+                tokio::time::sleep(RETRYABLE_BACKOFF).await;
+            }
+        }
+    }
+}
+
+/// Result of one HTTP enrollment attempt.
+enum AttemptResult {
+    /// A `cak_` was issued (201/200). Carries the plaintext (never logged).
+    Issued(String),
+    /// Collision-gated (202): no key issued.
+    Pending,
+}
+
+/// Failure classes that drive the backoff policy.
+enum AttemptError {
+    /// 401/409 — won't fix without operator action; back off long but keep trying.
+    Terminal(String),
+    /// 429 / network / 5xx / decode — transient; short backoff.
+    Retryable(String),
+}
+
+/// Make one `POST /api/enroll` call and classify the response per the contract.
+async fn attempt_enroll(
+    https_base: &str,
+    site_code: &str,
+    enrollment_key: &str,
+    machine_uid: &str,
+    hostname: &str,
+    config: &Config,
+) -> std::result::Result<AttemptResult, AttemptError> {
+    let url = format!("{}/api/enroll", https_base.trim_end_matches('/'));
+
+    let body = EnrollRequest {
+        site_code,
+        enrollment_key,
+        machine_uid,
+        hostname,
+        labels: EnrollLabels {
+            company: config.company.as_deref().filter(|s| !s.is_empty()),
+            site: config.site.as_deref().filter(|s| !s.is_empty()),
+            department: config.department.as_deref().filter(|s| !s.is_empty()),
+            device_type: config.device_type.as_deref().filter(|s| !s.is_empty()),
+            tags: &config.tags,
+        },
+    };
+
+    let client = build_client().map_err(|e| AttemptError::Retryable(e.to_string()))?;
+
+    let response = client
+        .post(&url)
+        .json(&body)
+        .timeout(Duration::from_secs(30))
+        .send()
+        .await
+        .map_err(|e| AttemptError::Retryable(format!("request to {url} failed: {e}")))?;
+
+    let status = response.status();
+    match status.as_u16() {
+        // New (201) or reuse (200): body carries the cak_.
+        200 | 201 => {
+            let parsed: EnrollResponse = response
+                .json()
+                .await
+                .map_err(|e| AttemptError::Retryable(format!("malformed success body: {e}")))?;
+            match parsed.key {
+                Some(cak) if !cak.is_empty() => {
+                    tracing::info!(
+                        "[ENROLL] server accepted enrollment: state={} disposition={}",
+                        parsed.enrollment_state,
+                        parsed.disposition
+                    );
+                    Ok(AttemptResult::Issued(cak))
+                }
+                // 2xx with no key is contract-violating for the active path; treat
+                // as retryable so we don't silently spin or crash.
+                _ => Err(AttemptError::Retryable(format!(
+                    "server returned {} with no key (state={}, disposition={})",
+                    status, parsed.enrollment_state, parsed.disposition
+                ))),
+            }
+        }
+
+        // Collision-gated: pending operator confirmation, no key.
+        202 => {
+            // Body decode is best-effort here; the status alone is authoritative.
+            Ok(AttemptResult::Pending)
+        }
+
+        // Bad/rotated enrollment key or unknown site code.
+        401 => Err(AttemptError::Terminal(
+            "ENROLL_REJECTED — the site code or enrollment key is invalid or rotated; \
+             this installer needs a current per-site key"
+                .to_string(),
+        )),
+
+        // Machine already enrolled at a different site.
+        409 => Err(AttemptError::Terminal(
+            "ENROLL_SITE_CONFLICT — this machine is already enrolled at another site; \
+             a deliberate move requires the operator-initiated reassignment flow"
+                .to_string(),
+        )),
+
+        // Rate-limited / locked out — honor Retry-After if present, else default.
+        429 => {
+            let retry_after = response
+                .headers()
+                .get(reqwest::header::RETRY_AFTER)
+                .and_then(|v| v.to_str().ok())
+                .and_then(|s| s.parse::<u64>().ok());
+            Err(AttemptError::Retryable(match retry_after {
+                Some(secs) => format!("RATE_LIMITED (retry-after {secs}s)"),
+                None => "RATE_LIMITED".to_string(),
+            }))
+        }
+
+        // 5xx or anything else — transient from the agent's perspective.
+        _ => Err(AttemptError::Retryable(format!(
+            "unexpected enrollment response: HTTP {status}"
+        ))),
+    }
+}
+
+/// Build the HTTP client for enrollment, matching the update path's TLS posture
+/// (`rustls`, with an opt-in dev-insecure escape hatch in debug builds only).
+fn build_client() -> Result<reqwest::Client> {
+    reqwest::Client::builder()
+        .danger_accept_invalid_certs(dev_insecure_tls())
+        .build()
+        .context("failed to build enrollment HTTP client")
+}
+
+/// Dev-only TLS bypass — identical policy to `update::dev_insecure_tls`: only in
+/// debug builds AND only when `GURUCONNECT_DEV_INSECURE_TLS` is set. NEVER active
+/// in a release build.
+fn dev_insecure_tls() -> bool {
+    if cfg!(debug_assertions) && std::env::var("GURUCONNECT_DEV_INSECURE_TLS").is_ok() {
+        tracing::warn!(
+            "[ENROLL] TLS verification DISABLED (dev-insecure mode) — DO NOT use in production"
+        );
+        true
+    } else {
+        false
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The request body must serialize to exactly the field names the Phase A
+    /// server deserializes (`enroll::EnrollRequest` / `EnrollLabels`). A drift here
+    /// is a silent enrollment failure, so pin the wire shape.
+    #[test]
+    fn request_serializes_to_the_server_contract() {
+        let tags = vec!["prod".to_string()];
+        let req = EnrollRequest {
+            site_code: "ACME-HQ",
+            enrollment_key: "cek_secret",
+            machine_uid: "muid_abc",
+            hostname: "WS-01",
+            labels: EnrollLabels {
+                company: Some("Acme"),
+                site: Some("HQ"),
+                department: Some("IT"),
+                device_type: Some("workstation"),
+                tags: &tags,
+            },
+        };
+        let v: serde_json::Value = serde_json::to_value(&req).unwrap();
+        assert_eq!(v["site_code"], "ACME-HQ");
+        assert_eq!(v["enrollment_key"], "cek_secret");
+        assert_eq!(v["machine_uid"], "muid_abc");
+        assert_eq!(v["hostname"], "WS-01");
+        assert_eq!(v["labels"]["company"], "Acme");
+        assert_eq!(v["labels"]["site"], "HQ");
+        assert_eq!(v["labels"]["department"], "IT");
+        assert_eq!(v["labels"]["device_type"], "workstation");
+        assert_eq!(v["labels"]["tags"][0], "prod");
+    }
+
+    /// Empty optional labels are omitted (the server defaults them), and an empty
+    /// tag list is not serialized — keeping the body minimal for a thin installer.
+    #[test]
+    fn request_omits_empty_optional_labels() {
+        let tags: Vec<String> = Vec::new();
+        let req = EnrollRequest {
+            site_code: "S",
+            enrollment_key: "cek_x",
+            machine_uid: "muid_x",
+            hostname: "H",
+            labels: EnrollLabels {
+                company: None,
+                site: None,
+                department: None,
+                device_type: None,
+                tags: &tags,
+            },
+        };
+        let v: serde_json::Value = serde_json::to_value(&req).unwrap();
+        let labels = v["labels"].as_object().unwrap();
+        assert!(!labels.contains_key("company"));
+        assert!(!labels.contains_key("department"));
+        assert!(!labels.contains_key("tags"));
+    }
+
+    /// The success response decoder must accept both a key-bearing active body and
+    /// a keyless pending body (mirrors `EnrollResponse` with `skip_serializing_if`).
+    #[test]
+    fn response_decodes_active_and_pending_shapes() {
+        let active: EnrollResponse = serde_json::from_str(
+            r#"{"machine_id":"m1","key":"cak_live","enrollment_state":"active","disposition":"new"}"#,
+        )
+        .unwrap();
+        assert_eq!(active.key.as_deref(), Some("cak_live"));
+        assert_eq!(active.enrollment_state, "active");
+
+        let pending: EnrollResponse = serde_json::from_str(
+            r#"{"machine_id":"m2","enrollment_state":"pending","disposition":"collision_pending"}"#,
+        )
+        .unwrap();
+        assert!(pending.key.is_none());
+        assert_eq!(pending.disposition, "collision_pending");
+    }
+}

agent/src/identity.rs

@@ -9,22 +9,48 @@
 //! **recomputable**: the same machine yields the same id on every call with no
 //! persistence required.
 //!
-//! - **Windows:** SHA-256 hash of the OS machine GUID read from
+//! - **Windows:** SHA-256 of a hardware identity string. The id is derived from
-//!   `HKLM\SOFTWARE\Microsoft\Cryptography\MachineGuid` (a `REG_SZ`). The raw
+//!   the **hardware salt ONLY** whenever any durable hardware signal is readable:
-//!   GUID is never returned — only the opaque `muid_<hex>` derived from it.
+//!   the **SMBIOS system UUID** (`Win32_ComputerSystemProduct.UUID`), or — when
-//! - **Non-Windows (and Windows registry failure):** a random UUID persisted in
+//!   that is absent / all-zeros / all-FFs (some OEMs/hypervisors) — the
-//!   the agent's data directory, read back on subsequent runs so it is stable
+//!   **motherboard serial** (`Win32_BaseBoard.SerialNumber`) plus the **primary
-//!   across calls and process restarts.
+//!   disk serial**. A fixed namespace string is mixed in for domain separation.
+//!   The OS machine GUID
+//!   (`HKLM\SOFTWARE\Microsoft\Cryptography\MachineGuid`, a `REG_SZ`) is used
+//!   ONLY as a last-resort signal when NO hardware salt is readable. The raw
+//!   signals are never returned — only the opaque `muid_<hex>` derived from them.
+//! - **Non-Windows (and Windows with no readable signal at all):** a random UUID
+//!   persisted in the agent's data directory, read back on subsequent runs so it
+//!   is stable across calls and process restarts.
+//!
+//! **Stability contract (SPEC-016 item 1):**
+//! - **Salted path (hardware signal present) is re-image-stable:** the digest
+//!   mixes only durable hardware signals (SMBIOS UUID, or board + disk serial) and
+//!   a fixed namespace — NOT the `MachineGuid`, which Windows regenerates on every
+//!   OS install/re-image. So the `machine_uid` survives both a reboot AND an OS
+//!   re-image on the SAME hardware (the re-image dedup goal), while distinct
+//!   physical boxes stay distinct.
+//! - **MachineGuid-only path is the volatile floor:** when no hardware salt is
+//!   readable, the id anchors on the `MachineGuid` alone. This is stable across
+//!   reboots but NOT across a re-image (the GUID is regenerated). This degraded
+//!   path is logged at WARN so the server-side collision gate operator has a clue.
 //!
 //! This module deliberately does NOT change `agent_id`/`generate_agent_id`.
 //! `machine_uid` is reported *alongside* `agent_id`; the server-side dedup that
-//! consumes it is a separate task.
+//! consumes it lives in `POST /api/enroll` (SPEC-016 Phase A) and the relay
+//! connect path.
 
 use std::sync::OnceLock;
 
 /// Prefix marking the value as an opaque machine-uid (vs. a raw GUID/UUID).
 const MUID_PREFIX: &str = "muid_";
 
+/// Fixed namespace mixed into the hardware-salted derivation for domain
+/// separation: it ties the digest to *this* identity scheme so the same raw
+/// hardware serial can never collide with an unrelated digest, and it documents
+/// the derivation version. It is NOT a secret — it is a constant.
+const MUID_NAMESPACE: &str = "guruconnect:machine_uid:v1";
+
 /// Cached value — `machine_uid()` reads the registry / a file, so compute once
 /// and reuse for the lifetime of the process.
 static MACHINE_UID: OnceLock<String> = OnceLock::new();
@@ -32,10 +58,11 @@ static MACHINE_UID: OnceLock<String> = OnceLock::new();
 /// Return a deterministic, recomputable opaque machine identifier.
 ///
 /// The result is non-empty and prefixed with [`MUID_PREFIX`]. It is cached after
-/// the first call. On Windows it is derived purely from the OS machine GUID (no
+/// the first call. On Windows it is derived from a durable hardware salt when one
-/// persistence). If the Windows registry read fails — or on any non-Windows
+/// is readable (re-image-stable; see the module docs), falling back to the OS
-/// platform — it degrades to a persisted random UUID (today's-behavior-equivalent
+/// machine GUID alone (reboot-stable floor) and finally — when no signal at all is
-/// stability) rather than panicking.
+/// readable, or on any non-Windows platform — a persisted random UUID, rather than
+/// panicking.
 pub fn machine_uid() -> String {
     MACHINE_UID.get_or_init(compute_machine_uid).clone()
 }
@@ -56,23 +83,265 @@ fn derive_uid(raw: &str) -> String {
 
 #[cfg(windows)]
 fn compute_machine_uid() -> String {
+    // PRIMARY signal (SPEC-016 item 1): a durable hardware salt — SMBIOS system
+    // UUID if usable, else motherboard + disk serial. When ANY hardware salt is
+    // readable we derive the uid from the salt ALONE (plus a fixed namespace),
+    // deliberately EXCLUDING the MachineGuid: Windows regenerates the MachineGuid
+    // on every OS install/re-image, so mixing it in would break re-image dedup.
+    // The salted digest survives both reboot AND re-image on the same hardware.
+    if let Some(salt) = hardware_salt() {
+        tracing::info!("machine_uid derived from durable hardware salt (re-image-stable)");
+        return derive_uid(&format!("{MUID_NAMESPACE}|{salt}"));
+    }
+
+    // LAST-RESORT signal: no hardware salt is readable, so anchor on the OS
+    // MachineGuid alone. This is the volatile FLOOR — stable across reboots but
+    // NOT across an OS re-image (the GUID is regenerated). We WARN so the
+    // server-side collision-gate operator knows this endpoint's uid is not
+    // re-image-stable. The MachineGuid itself is never logged.
     match read_machine_guid() {
-        Ok(guid) if !guid.trim().is_empty() => derive_uid(guid.trim()),
+        Ok(guid) if !guid.trim().is_empty() => {
+            tracing::warn!(
+                "machine_uid: no durable hardware salt readable; anchoring on MachineGuid \
+                 ONLY — this id is reboot-stable but NOT re-image-stable"
+            );
+            derive_uid(&format!("{MUID_NAMESPACE}|machineguid:{}", guid.trim()))
+        }
         Ok(_) => {
             tracing::warn!(
-                "MachineGuid registry value was empty; falling back to persisted machine_uid"
+                "machine_uid: no hardware salt and MachineGuid registry value was empty; \
+                 falling back to persisted machine_uid"
             );
             persisted_uid()
         }
         Err(e) => {
             tracing::warn!(
-                "Failed to read MachineGuid from registry ({e}); falling back to persisted machine_uid"
+                "machine_uid: no hardware salt and failed to read MachineGuid ({e}); \
+                 falling back to persisted machine_uid"
             );
             persisted_uid()
         }
     }
 }
 
+/// Collect the durable hardware salt for the `machine_uid` (Windows only).
+///
+/// This is the PRIMARY identity signal: when it returns `Some(salt)`, the caller
+/// derives the uid from the salt ALONE (re-image-stable). Returns `Some(salt)`
+/// where `salt` is a deterministic, normalized concatenation of usable hardware
+/// signals, or `None` when nothing durable is readable (in which case the caller
+/// degrades to anchoring on the MachineGuid alone — the volatile floor).
+///
+/// Order of preference, per SPEC-016 item 1:
+///   1. SMBIOS system UUID (`Win32_ComputerSystemProduct.UUID`) — when present and
+///      not a degenerate placeholder (all-zeros / all-FFs, which some OEMs and
+///      hypervisor templates emit).
+///   2. Fallback: motherboard serial (`Win32_BaseBoard.SerialNumber`) + primary
+///      disk serial — combined so a single weak signal does not stand alone.
+///
+/// Each component is read via a narrow PowerShell CIM query (see
+/// [`query_cim_property`]); the values are normalized (trimmed, upper-cased) so
+/// trivial formatting drift never changes the digest.
+#[cfg(windows)]
+fn hardware_salt() -> Option<String> {
+    if let Some(uuid) = smbios_uuid() {
+        return Some(format!("smbios:{uuid}"));
+    }
+
+    // SMBIOS UUID unusable — fall back to board + disk serial. Use whichever of
+    // the two are readable; require at least one to be present, otherwise there
+    // is no durable salt and we return None.
+    let board = normalize_signal(query_cim_property("Win32_BaseBoard", "SerialNumber").as_deref());
+    let disk = primary_disk_serial();
+
+    match (board, disk) {
+        (Some(b), Some(d)) => Some(format!("board:{b}|disk:{d}")),
+        (Some(b), None) => Some(format!("board:{b}")),
+        (None, Some(d)) => Some(format!("disk:{d}")),
+        (None, None) => None,
+    }
+}
+
+/// The SMBIOS system UUID, or `None` if absent or a degenerate placeholder.
+///
+/// Some OEMs ship an all-zeros UUID and some hypervisor templates clone an
+/// all-FFs (or all-zeros) UUID; either is worthless as a distinguishing signal,
+/// so we reject both and let the caller fall back to board/disk serial.
+#[cfg(windows)]
+fn smbios_uuid() -> Option<String> {
+    let raw =
+        normalize_signal(query_cim_property("Win32_ComputerSystemProduct", "UUID").as_deref())?;
+
+    // Reject degenerate placeholders (ignoring dashes): all-zeros or all-FFs.
+    let hex: String = raw.chars().filter(|c| *c != '-').collect();
+    let all_zero = !hex.is_empty() && hex.chars().all(|c| c == '0');
+    let all_ff = !hex.is_empty() && hex.chars().all(|c| c == 'F');
+    if hex.is_empty() || all_zero || all_ff {
+        tracing::debug!("SMBIOS UUID is absent or a degenerate placeholder; using fallback salt");
+        return None;
+    }
+    Some(raw)
+}
+
+/// The serial number of the primary (boot/index-0) physical disk, normalized.
+///
+/// Prefers the disk whose `Index == 0` (the conventional boot disk); falls back
+/// to the first disk that reports any serial. Returns `None` if no disk reports a
+/// usable serial.
+#[cfg(windows)]
+fn primary_disk_serial() -> Option<String> {
+    // One narrow query: index + serial for all physical disks, sorted by index,
+    // emitted as `index<TAB>serial` lines. Parse the lowest-index non-empty serial.
+    let script = "Get-CimInstance -ClassName Win32_DiskDrive | \
+                  Sort-Object Index | \
+                  ForEach-Object { \"$($_.Index)`t$($_.SerialNumber)\" }";
+    let out = run_powershell(script)?;
+    for line in out.lines() {
+        let mut parts = line.splitn(2, '\t');
+        let _index = parts.next();
+        if let Some(serial) = parts.next() {
+            if let Some(n) = normalize_signal(Some(serial)) {
+                return Some(n);
+            }
+        }
+    }
+    None
+}
+
+/// Read a single property of a single-instance CIM class via PowerShell.
+///
+/// Returns the raw (untrimmed) first non-empty line of output, or `None`. This is
+/// a deliberately narrow shell-out rather than a full WMI/COM binding: the agent
+/// already has no WMI crate, and a COM `IWbemServices` binding for two scalar
+/// reads would be far more code and unsafe surface for no benefit. PowerShell's
+/// CIM cmdlets are present on every supported Windows target (7 SP1+/2008 R2+
+/// ship WMI; CIM cmdlets ship from PowerShell 3.0 / WMF 3.0, universally present
+/// on currently-supported builds).
+#[cfg(windows)]
+fn query_cim_property(class: &str, property: &str) -> Option<String> {
+    // `(Get-CimInstance -ClassName X).Property` — single scalar, no formatting.
+    let script = format!("(Get-CimInstance -ClassName {class}).{property}");
+    let out = run_powershell(&script)?;
+    out.lines()
+        .map(str::trim)
+        .find(|l| !l.is_empty())
+        .map(str::to_string)
+}
+
+/// Wall-clock bound on a single PowerShell hardware-signal query.
+///
+/// A wedged WMI/CIM provider can hang indefinitely; without a bound that would
+/// hang agent startup forever. On timeout we kill the child and treat the signal
+/// as missing (fall back through the chain) — never panic.
+#[cfg(windows)]
+const POWERSHELL_QUERY_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(10);
+
+/// Run a short PowerShell snippet and capture stdout, or `None` on any failure
+/// (including a wall-clock timeout).
+///
+/// Hidden window (`CREATE_NO_WINDOW`) so an interactive desktop never flashes a
+/// console; `-NonInteractive -NoProfile` for determinism and speed. The call is
+/// spawned and waited on with a [`POWERSHELL_QUERY_TIMEOUT`] bound so a stuck WMI
+/// provider cannot wedge startup; on timeout the child is killed and the signal is
+/// treated as missing. Never logs the captured output (it carries hardware
+/// identifiers).
+#[cfg(windows)]
+fn run_powershell(script: &str) -> Option<String> {
+    use std::io::Read;
+    use std::os::windows::process::CommandExt;
+    use std::process::{Command, Stdio};
+    use std::time::Instant;
+
+    // CREATE_NO_WINDOW — avoid a console flash on the interactive desktop.
+    const CREATE_NO_WINDOW: u32 = 0x0800_0000;
+
+    let mut child = match Command::new("powershell.exe")
+        .args([
+            "-NonInteractive",
+            "-NoProfile",
+            "-ExecutionPolicy",
+            "Bypass",
+            "-Command",
+            script,
+        ])
+        .stdin(Stdio::null())
+        .stdout(Stdio::piped())
+        .stderr(Stdio::null())
+        .creation_flags(CREATE_NO_WINDOW)
+        .spawn()
+    {
+        Ok(c) => c,
+        Err(e) => {
+            tracing::debug!("could not run hardware-signal query ({e}); ignoring this signal");
+            return None;
+        }
+    };
+
+    // Poll for exit with a wall-clock bound. We spin with a short sleep rather than
+    // a reader thread: the queries are infrequent (startup only) and the loop keeps
+    // the timeout logic simple and panic-free.
+    let deadline = Instant::now() + POWERSHELL_QUERY_TIMEOUT;
+    let status = loop {
+        match child.try_wait() {
+            Ok(Some(status)) => break status,
+            Ok(None) => {
+                if Instant::now() >= deadline {
+                    // Wedged provider: kill and treat as a missing signal.
+                    let _ = child.kill();
+                    let _ = child.wait();
+                    tracing::debug!(
+                        "hardware-signal query exceeded {}s timeout; killed and ignoring this signal",
+                        POWERSHELL_QUERY_TIMEOUT.as_secs()
+                    );
+                    return None;
+                }
+                std::thread::sleep(std::time::Duration::from_millis(50));
+            }
+            Err(e) => {
+                tracing::debug!("error waiting on hardware-signal query ({e}); ignoring");
+                let _ = child.kill();
+                let _ = child.wait();
+                return None;
+            }
+        }
+    };
+
+    if !status.success() {
+        tracing::debug!(
+            "hardware-signal query exited with status {:?}; ignoring this signal",
+            status.code()
+        );
+        return None;
+    }
+
+    // The process exited; drain its captured stdout.
+    let mut buf = Vec::new();
+    if let Some(mut out) = child.stdout.take() {
+        if let Err(e) = out.read_to_end(&mut buf) {
+            tracing::debug!("error reading hardware-signal query output ({e}); ignoring");
+            return None;
+        }
+    }
+    let s = String::from_utf8_lossy(&buf).trim().to_string();
+    if s.is_empty() {
+        None
+    } else {
+        Some(s)
+    }
+}
+
+/// Normalize a raw hardware signal: trim, upper-case, drop if empty. Upper-casing
+/// makes the digest stable against vendor case drift; trimming removes stray
+/// whitespace WMI sometimes pads serials with.
+#[cfg(windows)]
+fn normalize_signal(raw: Option<&str>) -> Option<String> {
+    let v = raw?.trim();
+    if v.is_empty() {
+        return None;
+    }
+    Some(v.to_uppercase())
+}
+
 #[cfg(not(windows))]
 fn compute_machine_uid() -> String {
     // No OS machine GUID available — use the persisted random UUID, hashed for a
@@ -297,4 +566,108 @@ mod tests {
         assert_eq!(a, b, "compute_machine_uid must be deterministic on Windows");
         assert!(a.starts_with(MUID_PREFIX));
     }
+
+    /// Pin the EXACT derivation strings that `compute_machine_uid` builds, so these
+    /// pure-function tests track the production logic. Keep in lock-step with
+    /// `compute_machine_uid`.
+    #[cfg(windows)]
+    fn salted_uid(salt: &str) -> String {
+        derive_uid(&format!("{MUID_NAMESPACE}|{salt}"))
+    }
+    #[cfg(windows)]
+    fn machineguid_only_uid(guid: &str) -> String {
+        derive_uid(&format!("{MUID_NAMESPACE}|machineguid:{guid}"))
+    }
+
+    /// H1 RE-IMAGE STABILITY: when a hardware salt is present, the uid is derived
+    /// from the salt ALONE — the MachineGuid is NOT part of the input. So holding
+    /// the hardware signals fixed while varying the MachineGuid MUST yield the SAME
+    /// uid. This is exactly the re-image case: an OS re-image regenerates the
+    /// MachineGuid but leaves SMBIOS UUID / board+disk serial unchanged, and the
+    /// machine_uid must not move (otherwise dedup breaks). We prove it by showing
+    /// the salted derivation has no MachineGuid term to vary.
+    #[cfg(windows)]
+    #[test]
+    fn salted_uid_is_reimage_stable_independent_of_machine_guid() {
+        let salt = "smbios:4C4C4544-0043-3010-8052-B4C04F564231";
+        // "Before re-image" and "after re-image": MachineGuid differs, but the
+        // salt-derived uid takes no MachineGuid input, so both are identical.
+        let before = salted_uid(salt);
+        let after = salted_uid(salt);
+        assert_eq!(
+            before, after,
+            "salted uid must be stable across a re-image (no MachineGuid term)"
+        );
+
+        // Contrast: the MachineGuid-only floor DOES move when the GUID changes —
+        // demonstrating WHY the salted path must exclude it for re-image stability.
+        let guid_a = machineguid_only_uid("11111111-2222-3333-4444-555555555555");
+        let guid_b = machineguid_only_uid("99999999-8888-7777-6666-555555555555");
+        assert_ne!(
+            guid_a, guid_b,
+            "MachineGuid-only floor is volatile across re-image (expected)"
+        );
+
+        // And the salted uid must differ from the MachineGuid-only floor for the
+        // same box: the two derivation paths are domain-separated.
+        assert_ne!(before, guid_a);
+    }
+
+    /// The hardware-salted derivation is `derive_uid` over a deterministic,
+    /// namespaced concatenation: identical signals MUST yield an identical uid and
+    /// any changed signal MUST change it. Pins the SPEC-016 determinism contract
+    /// independent of the (machine-specific) live hardware reads.
+    #[cfg(windows)]
+    #[test]
+    fn salted_derivation_is_deterministic_and_signal_sensitive() {
+        let with_smbios = salted_uid("smbios:AAAA-BBBB");
+        let with_smbios_again = salted_uid("smbios:AAAA-BBBB");
+        let with_board = salted_uid("board:SN123|disk:DSK9");
+
+        // Same inputs -> same uid.
+        assert_eq!(with_smbios, with_smbios_again);
+        // Different salt composition -> different uid (distinct boxes stay distinct).
+        assert_ne!(with_smbios, with_board);
+    }
+
+    /// All-zero and all-FF SMBIOS UUIDs are degenerate placeholders that some OEMs
+    /// and hypervisor templates emit; the normalizer + placeholder check must
+    /// reject them so the derivation falls through to board/disk serial. We
+    /// exercise the rejection predicate directly (it is pure) rather than the
+    /// live WMI read.
+    #[cfg(windows)]
+    #[test]
+    fn degenerate_smbios_uuids_are_rejected() {
+        // Replicate the predicate `smbios_uuid` applies after normalization.
+        fn is_degenerate(raw: &str) -> bool {
+            let Some(norm) = normalize_signal(Some(raw)) else {
+                return true;
+            };
+            let hex: String = norm.chars().filter(|c| *c != '-').collect();
+            hex.is_empty()
+                || (!hex.is_empty() && hex.chars().all(|c| c == '0'))
+                || (!hex.is_empty() && hex.chars().all(|c| c == 'F'))
+        }
+
+        assert!(is_degenerate("00000000-0000-0000-0000-000000000000"));
+        assert!(is_degenerate("FFFFFFFF-FFFF-FFFF-FFFF-FFFFFFFFFFFF"));
+        assert!(is_degenerate("ffffffff-ffff-ffff-ffff-ffffffffffff")); // case-insensitive via normalize
+        assert!(is_degenerate("   "));
+        // A real, mixed UUID is NOT degenerate.
+        assert!(!is_degenerate("4C4C4544-0043-3010-8052-B4C04F564231"));
+    }
+
+    /// `normalize_signal` trims, upper-cases, and drops empties — so case/space
+    /// drift in a vendor serial never perturbs the digest.
+    #[cfg(windows)]
+    #[test]
+    fn normalize_signal_is_stable_against_drift() {
+        assert_eq!(
+            normalize_signal(Some("  abc123 ")),
+            Some("ABC123".to_string())
+        );
+        assert_eq!(normalize_signal(Some("ABC123")), Some("ABC123".to_string()));
+        assert_eq!(normalize_signal(Some("   ")), None);
+        assert_eq!(normalize_signal(None), None);
+    }
 }

agent/src/install.rs

@@ -290,6 +290,18 @@ pub fn install(force_user_install: bool) -> Result<()> {
     // Register protocol handler
     register_protocol_handler(elevated)?;
 
+    // SPEC-018: a MANAGED install (embedded config => persistent agent) installs
+    // the LocalSystem service as its single autostart and removes the per-user
+    // HKCU\…\Run entry. Attended (support-code) and viewer installs are untouched:
+    // they have no embedded config and continue to use the HKCU Run / protocol
+    // handler paths exactly as before.
+    #[cfg(windows)]
+    {
+        if crate::config::Config::has_embedded_config() {
+            install_managed_service(&exe_path)?;
+        }
+    }
+
     info!("Installation complete!");
     if elevated {
         info!("Installed system-wide to: {}", install_path.display());
@@ -300,6 +312,64 @@ pub fn install(force_user_install: bool) -> Result<()> {
     Ok(())
 }
 
+/// SPEC-018: install the managed agent as a LocalSystem service and swap out the
+/// legacy per-user `HKCU\…\Run` autostart so the service is the single managed
+/// autostart (no double-run).
+///
+/// Installing a LocalSystem service requires Administrator. If the SCM rejects the
+/// create (not elevated), we surface the error rather than silently leaving the
+/// machine with no managed autostart — a managed deployment is expected to run the
+/// install elevated. The HKCU Run entry is removed best-effort regardless.
+#[cfg(windows)]
+pub fn install_managed_service(exe_path: &std::path::Path) -> Result<()> {
+    info!("Managed install: registering LocalSystem service (SPEC-018)");
+
+    crate::service::install_service(exe_path)
+        .map_err(|e| anyhow!("failed to install the managed agent service: {e:#}"))?;
+
+    // Start the service now so the agent comes up immediately on first install
+    // rather than only on the next boot. Best-effort: the service is auto-start, so
+    // a transient start failure still self-heals on reboot.
+    if let Err(e) = crate::service::start_service() {
+        warn!(
+            "managed service installed but did not start now ({e:#}); \
+             it is auto-start and will run on next boot"
+        );
+    }
+
+    // Remove the legacy per-user autostart so the agent does not also launch in the
+    // user's session (which would double-run alongside the service).
+    if let Err(e) = crate::startup::remove_from_startup() {
+        warn!(
+            "managed service installed, but failed to remove the legacy HKCU Run \
+             autostart (harmless if it was never present): {}",
+            e
+        );
+    } else {
+        info!("removed legacy HKCU Run autostart (service is now the managed autostart)");
+    }
+
+    Ok(())
+}
+
+/// SPEC-018: remove the managed agent service and any legacy HKCU Run autostart.
+/// Idempotent — succeeds if neither is present.
+#[cfg(windows)]
+pub fn uninstall_managed_service() -> Result<()> {
+    info!("Managed uninstall: removing LocalSystem service (SPEC-018)");
+
+    // Best-effort removal of the legacy autostart first (cheap, no SCM).
+    if let Err(e) = crate::startup::remove_from_startup() {
+        warn!(
+            "failed to remove legacy HKCU Run autostart during uninstall: {}",
+            e
+        );
+    }
+
+    crate::service::uninstall_service()
+        .map_err(|e| anyhow!("failed to uninstall the managed agent service: {e:#}"))
+}
+
 /// Check if the guruconnect:// protocol handler is registered
 #[cfg(windows)]
 pub fn is_protocol_handler_registered() -> bool {

agent/src/main.rs

@@ -16,11 +16,16 @@ mod capture;
 mod chat;
 mod config;
 mod consent;
+#[cfg(windows)]
+mod credential_store;
 mod encoder;
+mod enroll;
 mod identity;
 mod input;
 mod install;
 mod sas_client;
+#[cfg(windows)]
+mod service;
 mod session;
 mod startup;
 mod transport;
@@ -179,6 +184,12 @@ enum Commands {
     /// Show detailed version and build information
     #[command(name = "version-info")]
     VersionInfo,
+
+    /// Internal: entry point invoked by the Windows Service Control Manager to run
+    /// the managed agent as a LocalSystem service (SPEC-018). Not for interactive
+    /// use — running it by hand fails because there is no controlling SCM.
+    #[command(name = "service-run", hide = true)]
+    ServiceRun,
 }
 
 fn main() -> Result<()> {
@@ -223,7 +234,24 @@ fn main() -> Result<()> {
         Some(Commands::Install {
             user_only,
             elevated,
-        }) => run_install(user_only || elevated),
+        }) => {
+            // `run_install`'s parameter is `force_user_install` — when true it
+            // skips the UAC re-elevation attempt and installs in-place with
+            // whatever rights this process already has.
+            //
+            // - `user_only`: the user explicitly asked for a per-user install;
+            //   honour it directly.
+            // - `elevated`: this is the internal, already-elevated re-exec spawned
+            //   by `try_elevate_and_install` ("install --elevated"). It must NOT
+            //   attempt to elevate AGAIN (that would loop / re-prompt), so we pass
+            //   force=true here too. This is correct even though it routes through
+            //   the "user install" parameter, because the re-exec genuinely runs
+            //   elevated: `is_elevated()` returns true inside `install()`, so the
+            //   path resolves to Program Files and the LocalSystem service installs
+            //   normally. The flag only suppresses re-elevation; it does not force a
+            //   per-user (non-elevated) install when we are already elevated.
+            run_install(user_only || elevated)
+        }
         Some(Commands::Uninstall) => run_uninstall(),
         Some(Commands::Launch { url }) => run_launch(&url),
         Some(Commands::VersionInfo) => {
@@ -233,6 +261,21 @@ fn main() -> Result<()> {
             println!("{}", build_info::full_version());
             Ok(())
         }
+        Some(Commands::ServiceRun) => {
+            // SPEC-018 Phase 1: SCM-invoked entry. Hand off to the service
+            // dispatcher, which calls back into the control loop and runs the
+            // managed-agent logic as SYSTEM. Blocks until the service stops.
+            #[cfg(windows)]
+            {
+                service::run_dispatcher()
+            }
+            #[cfg(not(windows))]
+            {
+                Err(anyhow::anyhow!(
+                    "service-run is a Windows-only entry point (SPEC-018)"
+                ))
+            }
+        }
         None => {
             // No subcommand - detect mode from filename or embedded config
             // Legacy: if support_code arg provided, use that
@@ -261,16 +304,31 @@ fn main() -> Result<()> {
                     run_agent_mode(Some(code))
                 }
                 RunMode::PermanentAgent => {
-                    // Embedded config found - run as permanent agent
+                    // Embedded config found - managed/persistent agent.
                     info!("Permanent agent mode detected (embedded config)");
-                    if !install::is_protocol_handler_registered() {
+
-                        // First run - install then run as agent
+                    // SPEC-018: managed mode runs as the LocalSystem service, not as
-                        info!("First run - installing agent");
+                    // an interactive process. The service is the single autostart.
-                        if let Err(e) = install::install(false) {
+                    //   - If the service is already installed, the service is (or
-                            warn!("Installation failed: {}", e);
+                    //     will be) running the agent — this interactive invocation
-                        }
+                    //     must NOT spawn a second agent. Exit quietly.
+                    //   - On first run, install (which installs + starts the service
+                    //     and removes the legacy HKCU Run entry), then exit and let
+                    //     the service carry the agent as SYSTEM.
+                    #[cfg(windows)]
+                    {
+                        run_permanent_agent_managed()
+                    }
+                    #[cfg(not(windows))]
+                    {
+                        if !install::is_protocol_handler_registered() {
+                            info!("First run - installing agent");
+                            if let Err(e) = install::install(false) {
+                                warn!("Installation failed: {}", e);
+                            }
+                        }
+                        run_agent_mode(None)
                     }
-                    run_agent_mode(None)
                 }
                 RunMode::Default => {
                     // No special mode detected - use legacy logic
@@ -323,7 +381,224 @@ fn run_agent_mode(support_code: Option<String>) -> Result<()> {
 
     // Run the agent
     let rt = tokio::runtime::Runtime::new()?;
-    rt.block_on(run_agent(config))
+    rt.block_on(async move {
+        // SPEC-016 Phase B: resolve the operating credential before connecting.
+        // Support sessions are unaffected — they authenticate by support code, not
+        // by a per-machine cak_, so we only resolve enrollment for a managed agent.
+        if config.support_code.is_none() {
+            resolve_agent_credential(&mut config).await?;
+        }
+        run_agent(config, None).await
+    })
+}
+
+/// SPEC-018 Phase 1: run the managed/persistent agent as the LocalSystem service.
+///
+/// Invoked from the service control loop ([`service::run_service`]) once the
+/// service has reported `Running`. This is the same persistent-agent logic as
+/// [`run_agent_mode`] (load config, resolve/enroll the per-machine `cak_` per
+/// SPEC-016, hold the relay connection) — but it runs **as SYSTEM**, so the
+/// SYSTEM-ACL'd `cak_` store is finally readable in-context, and it observes the
+/// SCM `shutdown` flag for a graceful stop.
+///
+/// Returns `Ok(())` when the agent loop exits because a stop was requested, and
+/// `Err` only on an unrecoverable *local* fault (e.g. no usable credential and no
+/// enrollment material) — network errors are retried inside the loop and never
+/// surface here.
+///
+/// Phase 2 seam: this is where the session broker is wired in — the runtime
+/// started here will own the broker that spawns the per-session capture/input
+/// worker (`CreateProcessAsUserW`) and the IPC server. Phase 1 connects/enrolls
+/// only; it does not capture a desktop (a Session-0 SYSTEM process cannot).
+#[cfg(windows)]
+pub fn run_managed_agent_service(
+    shutdown: std::sync::Arc<std::sync::atomic::AtomicBool>,
+) -> Result<()> {
+    info!("Loading managed-agent configuration (running as SYSTEM)");
+
+    let mut config = config::Config::load()?;
+    // The service ONLY ever runs the managed/persistent path. A support session is
+    // an interactive, user-launched flow and must never be carried by the service.
+    config.support_code = None;
+
+    info!("Server: {}", config.server_url);
+    if let Some(ref company) = config.company {
+        info!("Company: {}", company);
+    }
+    if let Some(ref site) = config.site {
+        info!("Site: {}", site);
+    }
+
+    let rt = tokio::runtime::Runtime::new()?;
+
+    // SPEC-018 (finding M): this future runs across the `extern "system"` service
+    // entry point (ffi_service_main -> service_main -> run_service -> here). A
+    // panic that unwound across that FFI boundary is undefined behaviour (the C
+    // ABI cannot carry a Rust unwind) and would abort the process instead of
+    // taking the intended ServiceSpecific(1) fault path. Catch it here and convert
+    // it into an `Err`, which `run_service` maps to ServiceExitCode::ServiceSpecific(1)
+    // so the SCM applies its configured recovery (restart) cleanly. `Running` is
+    // already reported before we get here, so a fault does not strand StartPending.
+    let outcome = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
+        rt.block_on(async move {
+            // SPEC-016 Phase B: resolve the operating credential before connecting.
+            // Running as SYSTEM, the SYSTEM+Administrators-ACL'd cak_ store is now
+            // readable in-context, so the Phase B fail-fast guard is not hit on this
+            // path (it remains as a safety net for any non-SYSTEM invocation).
+            resolve_agent_credential(&mut config).await?;
+            run_agent(config, Some(shutdown)).await
+        })
+    }));
+
+    match outcome {
+        Ok(result) => result,
+        Err(panic) => {
+            // Recover a human-readable message from the panic payload for the log;
+            // do not re-panic (that would unwind across the FFI boundary again).
+            let detail = panic
+                .downcast_ref::<&str>()
+                .map(|s| s.to_string())
+                .or_else(|| panic.downcast_ref::<String>().cloned())
+                .unwrap_or_else(|| "non-string panic payload".to_string());
+            error!("managed-agent runtime panicked: {detail}");
+            Err(anyhow::anyhow!("managed-agent runtime panicked: {detail}"))
+        }
+    }
+}
+
+/// SPEC-018 Phase 1: handle an interactive launch of a MANAGED agent binary (one
+/// carrying embedded config, detected as [`config::RunMode::PermanentAgent`]).
+///
+/// Managed mode runs as the LocalSystem service, never as an interactive process:
+///   - If the service is already installed, the service is (or will be) running
+///     the agent as SYSTEM, so this interactive invocation must NOT spawn a second
+///     agent — it exits quietly.
+///   - On first run, install (which installs + starts the service and removes the
+///     legacy `HKCU\…\Run` autostart), then exit and let the service carry the
+///     agent. If the service install fails (e.g. not elevated), fall back to
+///     running the agent in-process for this run so the machine is not left with no
+///     agent at all.
+#[cfg(windows)]
+fn run_permanent_agent_managed() -> Result<()> {
+    if service::is_service_installed() {
+        info!(
+            "Managed service already installed; the service runs the agent as SYSTEM — \
+             this interactive instance has nothing to do"
+        );
+        return Ok(());
+    }
+
+    info!("First run - installing managed agent service");
+    if let Err(e) = install::install(false) {
+        warn!(
+            "Managed service install failed ({e:#}); falling back to in-process agent for this run"
+        );
+        return run_agent_mode(None);
+    }
+
+    info!("Managed agent service installed; handing off to the service");
+    Ok(())
+}
+
+/// Resolve the per-machine operating credential for a managed agent (SPEC-016
+/// Phase B, run-mode wiring).
+///
+/// Precedence:
+///   1. A `cak_` already stored encrypted at rest -> load it and connect with it
+///      (the steady-state path; no network call, no re-enroll).
+///   2. No stored `cak_` but an `enrollment_key` + `site_code` are present ->
+///      run first-run enrollment to obtain + persist a `cak_`, then connect.
+///   3. Neither a stored `cak_` nor enrollment material, but a non-empty
+///      `api_key` is configured -> use it as the DEPRECATED shared/legacy key
+///      (transition compatibility only; logged at WARNING).
+///   4. Nothing usable -> error; a managed agent cannot authenticate.
+async fn resolve_agent_credential(config: &mut config::Config) -> Result<()> {
+    // 1. Stored per-machine cak_ (steady state).
+    #[cfg(windows)]
+    {
+        use credential_store::LoadCakError;
+        match credential_store::load_cak() {
+            Ok(Some(cak)) => {
+                info!("Using stored per-machine credential (cak_)");
+                config.api_key = cak;
+                // Any leftover enrollment material is now moot.
+                config.enrollment_key = None;
+                return Ok(());
+            }
+            Ok(None) => {
+                info!("No stored per-machine credential; will enroll if configured");
+            }
+            // C1 / M1 — the store exists but THIS security context cannot read it
+            // (access-denied against the SYSTEM-only ACL). This is the brick the
+            // C1 guard prevents: a non-SYSTEM run could write the store but never
+            // read it back. Fail fast with an actionable message; do NOT loop and
+            // do NOT silently re-enroll. The SYSTEM+Administrators ACL is correct
+            // for the target (Option A) and is deliberately kept.
+            //
+            // SPEC-018 (this spec): the managed agent now runs as the GuruConnect
+            // SYSTEM service ([`run_managed_agent_service`]), so on the production
+            // managed path the store IS readable in-context and this branch is NOT
+            // hit. The guard is intentionally retained as a harmless safety net for
+            // any non-SYSTEM invocation (e.g. someone running the managed binary
+            // interactively): it still fails fast with an actionable message rather
+            // than bricking. Do NOT remove it in Phase 1.
+            Err(LoadCakError::Io {
+                permission_denied: true,
+                source,
+            }) => {
+                return Err(anyhow::anyhow!(
+                    "[ENROLL] credential store is not accessible in this context \
+                     ({source}) — the managed agent must run as the GuruConnect SYSTEM \
+                     service (see SPEC-018). Refusing to re-enroll."
+                ));
+            }
+            // M1 — other IO error reaching the store (not access-denied): also
+            // operational, not a tamper signal. Surface it; do not re-enroll over a
+            // store we simply could not read.
+            Err(e @ LoadCakError::Io { .. }) => {
+                return Err(anyhow::Error::new(e).context(
+                    "[ENROLL] credential store present but unreadable (IO error); \
+                     refusing to re-enroll over it",
+                ));
+            }
+            Err(e @ LoadCakError::Path(_)) => {
+                return Err(anyhow::Error::new(e)
+                    .context("[ENROLL] could not resolve the credential store path"));
+            }
+            // M1 — the bytes were read but failed to DECRYPT: the real tamper /
+            // wrong-machine signal. Hard stop; never silently re-enroll over it.
+            Err(e @ LoadCakError::Decrypt(_)) => {
+                return Err(anyhow::Error::new(e).context(
+                    "[ENROLL] stored credential failed to decrypt — possible tamper or \
+                     copy from another machine; refusing to silently re-enroll",
+                ));
+            }
+        }
+    }
+
+    // 2. First-run enrollment (the SPEC-016 zero-touch path). run_enrollment only
+    // returns once a cak_ is stored (it retries network/429/collision-pending
+    // internally); a returned Err is an unrecoverable local fault.
+    if config.enrollment_key.is_some() && config.site_code.is_some() {
+        info!("Enrollment material present; running first-run enrollment");
+        enroll::run_enrollment(config).await?;
+        return Ok(());
+    }
+
+    // 3. DEPRECATED shared/legacy api_key fallback (transition only).
+    if !config.api_key.is_empty() {
+        warn!(
+            "Connecting with a DEPRECATED shared/legacy api_key. Migrate this agent \
+             to a per-site enrollment (SPEC-016); the shared key path will be removed."
+        );
+        return Ok(());
+    }
+
+    // 4. Nothing usable.
+    Err(anyhow::anyhow!(
+        "no operating credential available: no stored cak_, no enrollment_key/site_code, \
+         and no legacy api_key — this managed agent cannot authenticate"
+    ))
 }
 
 /// Run in viewer mode (connect to remote session)
@@ -376,7 +651,22 @@ fn run_install(force_user_install: bool) -> Result<()> {
 fn run_uninstall() -> Result<()> {
     info!("Uninstalling GuruConnect...");
 
-    // Remove from startup
+    // SPEC-018: remove the managed LocalSystem service and the legacy HKCU Run
+    // autostart. Idempotent — no error if the service was never installed (an
+    // attended/viewer install has no service), so this is safe for every install
+    // shape. Requires Administrator to delete the service; a non-elevated uninstall
+    // still clears the per-user autostart below.
+    #[cfg(windows)]
+    {
+        if let Err(e) = install::uninstall_managed_service() {
+            warn!(
+                "Failed to remove managed service (may require Administrator): {}",
+                e
+            );
+        }
+    }
+
+    // Remove from startup (covers non-elevated / attended / viewer installs).
     if let Err(e) = startup::remove_from_startup() {
         warn!("Failed to remove from startup: {}", e);
     }
@@ -474,31 +764,62 @@ fn cleanup_on_exit() {
     }
 }
 
-/// Run the agent main loop
+/// Run the agent main loop.
-async fn run_agent(config: config::Config) -> Result<()> {
+///
+/// `service_shutdown`, when present, is the SCM cooperative-stop flag (SPEC-018):
+/// the managed-agent service passes it so the loop exits promptly on
+/// `Stop`/`Shutdown`. It is `None` for the interactive/user-launched paths, which
+/// stop via the tray exit / server control messages instead.
+async fn run_agent(
+    config: config::Config,
+    service_shutdown: Option<std::sync::Arc<std::sync::atomic::AtomicBool>>,
+) -> Result<()> {
+    use std::sync::atomic::Ordering;
+
     let elevated = install::is_elevated();
+    let running_as_service = service_shutdown.is_some();
     let mut session = session::SessionManager::new(config.clone(), elevated);
     let is_support_session = config.support_code.is_some();
     let hostname = config.hostname();
 
-    // Add to startup
+    // Helper: has the SCM asked us to stop?
-    if let Err(e) = startup::add_to_startup() {
+    let stop_requested = |flag: &Option<std::sync::Arc<std::sync::atomic::AtomicBool>>| -> bool {
+        flag.as_ref()
+            .map(|f| f.load(Ordering::SeqCst))
+            .unwrap_or(false)
+    };
+
+    // Autostart persistence:
+    //   - As the SYSTEM service (SPEC-018), the SERVICE itself is the managed
+    //     autostart — do NOT write the per-user HKCU\…\Run entry (that would be a
+    //     second, redundant autostart, and writing it from SYSTEM lands in the
+    //     wrong hive). The service install/uninstall owns lifecycle.
+    //   - Interactive/user-launched runs keep the existing HKCU Run behavior.
+    if running_as_service {
+        info!("Running as the GuruConnect SYSTEM service; service is the autostart (skipping HKCU Run)");
+    } else if let Err(e) = startup::add_to_startup() {
         warn!("Failed to add to startup: {}", e);
     }
 
-    // Create tray icon
+    // A Session-0 SYSTEM service has no interactive desktop, so a tray icon is
-    let tray = match tray::TrayController::new(
+    // both impossible and meaningless there (SPEC-018 Phase 2 moves the user-facing
-        &hostname,
+    // surface into the per-session worker). Only create the tray off the service.
-        config.support_code.as_deref(),
+    let tray = if running_as_service {
-        is_support_session,
+        None
-    ) {
+    } else {
-        Ok(t) => {
+        match tray::TrayController::new(
-            info!("Tray icon created");
+            &hostname,
-            Some(t)
+            config.support_code.as_deref(),
-        }
+            is_support_session,
-        Err(e) => {
+        ) {
-            warn!("Failed to create tray icon: {}", e);
+            Ok(t) => {
-            None
+                info!("Tray icon created");
+                Some(t)
+            }
+            Err(e) => {
+                warn!("Failed to create tray icon: {}", e);
+                None
+            }
         }
     };
 
@@ -507,6 +828,12 @@ async fn run_agent(config: config::Config) -> Result<()> {
 
     // Connect to server and run main loop
     loop {
+        // SPEC-018: honour an SCM stop request before (re)connecting.
+        if stop_requested(&service_shutdown) {
+            info!("Service stop requested; exiting agent loop");
+            return Ok(());
+        }
+
         info!("Connecting to server...");
 
         if is_support_session {
@@ -528,11 +855,22 @@ async fn run_agent(config: config::Config) -> Result<()> {
                 }
 
                 if let Err(e) = session
-                    .run_with_tray(tray.as_ref(), chat_ctrl.as_ref())
+                    .run_with_tray(tray.as_ref(), chat_ctrl.as_ref(), service_shutdown.as_ref())
                     .await
                 {
                     let error_msg = e.to_string();
 
+                    // SPEC-018 (finding H): the connected session loop broke
+                    // because the SCM asked the service to stop. The loop already
+                    // closed the WebSocket cleanly; treat this as a graceful stop
+                    // (no reconnect) so the service transitions StopPending ->
+                    // Stopped. Only the service path can produce this (it is the
+                    // only caller that passes a shutdown flag).
+                    if error_msg.contains(session::SERVICE_STOP_SENTINEL) {
+                        info!("Service stop requested during session; exiting agent loop");
+                        return Ok(());
+                    }
+
                     if error_msg.contains("USER_EXIT") {
                         info!("Session ended by user");
                         cleanup_on_exit();
@@ -605,6 +943,47 @@ async fn run_agent(config: config::Config) -> Result<()> {
         }
 
         info!("Reconnecting in 5 seconds...");
-        tokio::time::sleep(tokio::time::Duration::from_secs(5)).await;
+        // SPEC-018: poll the SCM stop flag during the backoff so a service stop is
+        // honoured within ~250ms instead of waiting the full reconnect delay.
+        if service_shutdown.is_some() {
+            for _ in 0..20 {
+                if stop_requested(&service_shutdown) {
+                    info!("Service stop requested during reconnect backoff; exiting agent loop");
+                    return Ok(());
+                }
+                tokio::time::sleep(tokio::time::Duration::from_millis(250)).await;
+            }
+        } else {
+            tokio::time::sleep(tokio::time::Duration::from_secs(5)).await;
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use clap::CommandFactory;
+
+    /// SPEC-018 finding N1: pin the clap subcommand name to the constant the SCM
+    /// is registered with. The service is installed with `SERVICE_RUN_ARG` as its
+    /// launch argument; when the SCM starts it, clap must route that exact token
+    /// into [`Commands::ServiceRun`]. If the `#[command(name = "service-run")]`
+    /// attribute and the constant ever drift apart, the SCM would start the binary
+    /// but clap would fail to match the subcommand and the process would fall
+    /// through to default (non-service) mode and exit. Asserting against the live
+    /// clap metadata (not a second string literal) makes that drift impossible.
+    #[test]
+    #[cfg(windows)]
+    fn service_run_subcommand_matches_scm_launch_arg() {
+        let cmd = Cli::command();
+        let has_matching_subcommand = cmd
+            .get_subcommands()
+            .any(|sc| sc.get_name() == service::SERVICE_RUN_ARG);
+        assert!(
+            has_matching_subcommand,
+            "no clap subcommand named '{}' (the SCM launch arg); the ServiceRun \
+             #[command(name = ...)] attribute drifted from service::SERVICE_RUN_ARG",
+            service::SERVICE_RUN_ARG
+        );
     }
 }

agent/src/service/mod.rs

@@ -0,0 +1,520 @@
+//! Windows SYSTEM service host for the managed GuruConnect agent (SPEC-018).
+//!
+//! # Phase 1 scope (this module)
+//!
+//! Phase 1 proves the *managed/persistent* agent can run as **LocalSystem** in
+//! the isolated Session 0 across reboots and at the login screen:
+//!
+//!   1. Register the agent with the Service Control Manager (SCM) and run, when
+//!      started, the **existing persistent-agent logic** (`RunMode::PermanentAgent`
+//!      path) *as SYSTEM* — i.e. resolve/enroll the per-machine `cak_` (SPEC-016,
+//!      now readable because the SYSTEM-ACL'd store is in-context) and hold the
+//!      relay WSS connection.
+//!   2. Report a correct service lifecycle to the SCM (`StartPending` ->
+//!      `Running` -> `StopPending` -> `Stopped`) and handle `Stop`/`Shutdown`
+//!      gracefully. The control handler sets a shared shutdown flag; the agent
+//!      runtime observes it both between reconnect attempts AND inside the
+//!      connected session loop (SPEC-018 finding H), so a stop received while a
+//!      session is live breaks out promptly, closes the WS connection cleanly,
+//!      and exits — rather than waiting for the SCM to force-kill.
+//!   3. Provide install/uninstall of the service (LocalSystem, auto-start, crash
+//!      recovery) so managed mode uses the service as its single autostart
+//!      instead of the per-user `HKCU\…\Run` entry.
+//!
+//! # Phase 2 (deliberately NOT built here — see SPEC-018 §Scope)
+//!
+//! A SYSTEM service lives in Session 0 and **cannot** capture or inject the
+//! interactive desktop directly. Phase 1 therefore enrolls and connects but does
+//! **NOT** capture a desktop yet. The following are Phase 2 and are intentionally
+//! absent; the seams where they attach are called out inline below:
+//!
+//!   - the **session broker** (`WTSEnumerateSessionsW` /
+//!     `WTSGetActiveConsoleSessionId` / `WTSQueryUserToken`),
+//!   - the **per-session capture/input worker** spawned via `CreateProcessAsUserW`
+//!     into `winsta0\default`,
+//!   - **service <-> worker IPC** (the per-session ACL'd named pipe), and
+//!   - **`SERVICE_CONTROL_SESSIONCHANGE`** reaction (logon/logoff/console-connect
+//!     retarget).
+//!
+//! Phase 1 registers the control handler for `Stop`/`Shutdown`/`Interrogate`
+//! only. When Phase 2 lands, the broker hangs off the same control handler
+//! (adding `SESSIONCHANGE`) and off the same agent runtime started here.
+
+#![cfg(windows)]
+
+use std::ffi::OsString;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::Arc;
+use std::time::Duration;
+
+use anyhow::{Context, Result};
+use tracing::{error, info, warn};
+
+use windows_service::{
+    define_windows_service,
+    service::{
+        ServiceAccess, ServiceControl, ServiceControlAccept, ServiceErrorControl, ServiceExitCode,
+        ServiceInfo, ServiceStartType, ServiceState, ServiceStatus, ServiceType,
+    },
+    service_control_handler::{self, ServiceControlHandlerResult},
+    service_dispatcher,
+    service_manager::{ServiceManager, ServiceManagerAccess},
+};
+
+/// Internal service name registered with the SCM (no spaces; used by `sc`,
+/// `ServiceManager`, and the control handler).
+pub const SERVICE_NAME: &str = "GuruConnectAgent";
+
+/// Human-facing display name shown in `services.msc`.
+pub const SERVICE_DISPLAY_NAME: &str = "GuruConnect Managed Agent";
+
+/// Service description shown in `services.msc`.
+pub const SERVICE_DESCRIPTION: &str =
+    "Runs the managed GuruConnect remote-support agent as LocalSystem so it is \
+     reachable at the login screen and across reboots (SPEC-018).";
+
+/// Hidden subcommand the SCM invokes to enter the service control loop. The
+/// service is registered with this as its launch argument (see [`install_service`]),
+/// and `main.rs` routes it into [`run_dispatcher`].
+pub const SERVICE_RUN_ARG: &str = "service-run";
+
+/// Hint we give the SCM for how long start/stop transitions may take before it
+/// should consider the service hung.
+const TRANSITION_WAIT: Duration = Duration::from_secs(10);
+
+// The `windows-service` dispatcher requires a `extern "system"` entry point with
+// a fixed ABI; this macro generates `ffi_service_main`, which trampolines into
+// our safe `service_main`.
+define_windows_service!(ffi_service_main, service_main);
+
+/// Enter the SCM dispatcher (called from `main.rs` for the `service-run`
+/// subcommand). Blocks until the service stops. This must be invoked by the SCM,
+/// not interactively — `service_dispatcher::start` fails with
+/// `ERROR_FAILED_SERVICE_CONTROLLER_CONNECT` (1063) if there is no controlling
+/// SCM, which is the expected outcome of running `guruconnect service-run` by hand.
+pub fn run_dispatcher() -> Result<()> {
+    service_dispatcher::start(SERVICE_NAME, ffi_service_main)
+        .context("failed to connect to the service control dispatcher (must be started by the SCM)")
+}
+
+/// SCM-invoked service body. Any error is logged; the function cannot return an
+/// error to the SCM directly, so [`run_service`] reports a failed exit code on the
+/// status handle before returning.
+fn service_main(_arguments: Vec<OsString>) {
+    if let Err(e) = run_service() {
+        error!("service exited with error: {e:#}");
+    }
+}
+
+/// Drive the full service lifecycle: register the control handler, report
+/// `Running`, run the persistent agent until a stop is requested, then report
+/// `Stopped`.
+fn run_service() -> Result<()> {
+    info!("GuruConnect managed agent service starting (running as SYSTEM in session 0)");
+
+    // Cooperative shutdown flag flipped by the SCM control handler and observed by
+    // the agent runtime. `AtomicBool` keeps the handler closure trivially `Send`
+    // and avoids holding a lock inside an SCM callback.
+    let shutdown = Arc::new(AtomicBool::new(false));
+    let shutdown_for_handler = shutdown.clone();
+
+    let event_handler = move |control_event| -> ServiceControlHandlerResult {
+        match control_event {
+            // SPEC-018 Phase 1: graceful stop. Phase 2 adds
+            // `ServiceControl::SessionChange(_)` here to drive the session broker
+            // (retarget the capture/input worker on logon/logoff/console-connect);
+            // we intentionally do not accept SESSIONCHANGE yet.
+            ServiceControl::Stop | ServiceControl::Shutdown => {
+                info!("received {control_event:?}; signalling agent to shut down");
+                // Set the cooperative-stop flag. The agent runtime observes it on
+                // every idle tick of the connected session loop and between
+                // reconnect attempts (SPEC-018 finding H), so it breaks out and
+                // closes the WebSocket cleanly within ~100ms even if a session is
+                // currently connected.
+                shutdown_for_handler.store(true, Ordering::SeqCst);
+                ServiceControlHandlerResult::NoError
+            }
+            ServiceControl::Interrogate => ServiceControlHandlerResult::NoError,
+            _ => ServiceControlHandlerResult::NotImplemented,
+        }
+    };
+
+    let status_handle = service_control_handler::register(SERVICE_NAME, event_handler)
+        .context("failed to register the service control handler")?;
+
+    // Report StartPending while we spin up the runtime and connect.
+    set_status(
+        &status_handle,
+        ServiceState::StartPending,
+        ServiceControlAccept::empty(),
+        TRANSITION_WAIT,
+    );
+
+    // Report Running and accept Stop + Shutdown. We report Running before the
+    // first connect attempt completes because the agent loop reconnects forever;
+    // "the service is up and trying" is the correct steady state, and blocking the
+    // SCM on the first relay handshake would risk a start timeout on a slow boot.
+    set_status(
+        &status_handle,
+        ServiceState::Running,
+        ServiceControlAccept::STOP | ServiceControlAccept::SHUTDOWN,
+        Duration::default(),
+    );
+    info!("service reported Running; entering managed-agent control loop");
+
+    // Run the existing persistent-agent logic as SYSTEM. This is the Phase 1
+    // payload: resolve/enroll the cak_ (SPEC-016) and hold the relay connection.
+    let run_result = crate::run_managed_agent_service(shutdown.clone());
+
+    if let Err(e) = &run_result {
+        // The agent loop only returns Err on an unrecoverable LOCAL fault (e.g. no
+        // usable credential and nothing to enroll with). Network errors are
+        // retried inside the loop and never surface here. Report the failure to
+        // the SCM so recovery actions (restart) engage.
+        error!("managed-agent control loop terminated with error: {e:#}");
+    } else {
+        info!("managed-agent control loop exited cleanly on stop request");
+    }
+
+    // Transition StopPending -> Stopped.
+    set_status(
+        &status_handle,
+        ServiceState::StopPending,
+        ServiceControlAccept::empty(),
+        TRANSITION_WAIT,
+    );
+
+    let exit_code = match run_result {
+        Ok(()) => ServiceExitCode::Win32(0),
+        // ERROR_SERVICE_SPECIFIC_ERROR-style: surface a non-zero service-specific
+        // code so the SCM treats the exit as a failure and applies recovery.
+        Err(_) => ServiceExitCode::ServiceSpecific(1),
+    };
+
+    set_status_with_exit(
+        &status_handle,
+        ServiceState::Stopped,
+        ServiceControlAccept::empty(),
+        Duration::default(),
+        exit_code,
+    );
+    info!("service reported Stopped");
+
+    Ok(())
+}
+
+/// Report a status with a zero (success) exit code.
+fn set_status(
+    handle: &service_control_handler::ServiceStatusHandle,
+    state: ServiceState,
+    accepted: ServiceControlAccept,
+    wait_hint: Duration,
+) {
+    set_status_with_exit(
+        handle,
+        state,
+        accepted,
+        wait_hint,
+        ServiceExitCode::Win32(0),
+    );
+}
+
+/// Report a status to the SCM. A failure to report is logged (best-effort) — we
+/// cannot do anything actionable about it and must not panic inside the service.
+fn set_status_with_exit(
+    handle: &service_control_handler::ServiceStatusHandle,
+    state: ServiceState,
+    accepted: ServiceControlAccept,
+    wait_hint: Duration,
+    exit_code: ServiceExitCode,
+) {
+    let status = ServiceStatus {
+        service_type: ServiceType::OWN_PROCESS,
+        current_state: state,
+        controls_accepted: accepted,
+        exit_code,
+        checkpoint: 0,
+        wait_hint,
+        process_id: None,
+    };
+    if let Err(e) = handle.set_service_status(status) {
+        warn!("failed to report service status {state:?} to the SCM: {e}");
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Install / uninstall (used by install.rs for managed mode)
+// ---------------------------------------------------------------------------
+
+/// Install (or reinstall) the managed agent as a LocalSystem auto-start service
+/// pointing at `exe_path` with the [`SERVICE_RUN_ARG`] launch argument.
+///
+/// Idempotent: if the service already exists it is stopped and deleted first,
+/// then recreated, so an upgrade picks up a new binary path / config. Configures
+/// crash recovery (restart on failure) via `sc failure`.
+///
+/// Requires Administrator (SCM `CREATE_SERVICE`). Returns an error otherwise.
+pub fn install_service(exe_path: &std::path::Path) -> Result<()> {
+    let manager = ServiceManager::local_computer(
+        None::<&str>,
+        ServiceManagerAccess::CONNECT | ServiceManagerAccess::CREATE_SERVICE,
+    )
+    .context("failed to connect to the Service Control Manager (run as Administrator)")?;
+
+    // Remove any prior installation so the binary path / args are refreshed.
+    let mut deleted_existing = false;
+    if let Ok(existing) = manager.open_service(
+        SERVICE_NAME,
+        ServiceAccess::QUERY_STATUS | ServiceAccess::STOP | ServiceAccess::DELETE,
+    ) {
+        info!("existing {SERVICE_NAME} service found; removing before reinstall");
+        stop_if_running(&existing);
+        existing
+            .delete()
+            .context("failed to delete the existing service before reinstall")?;
+        drop(existing);
+        deleted_existing = true;
+    }
+
+    let service_info = ServiceInfo {
+        name: OsString::from(SERVICE_NAME),
+        display_name: OsString::from(SERVICE_DISPLAY_NAME),
+        service_type: ServiceType::OWN_PROCESS,
+        start_type: ServiceStartType::AutoStart,
+        error_control: ServiceErrorControl::Normal,
+        executable_path: exe_path.to_path_buf(),
+        launch_arguments: vec![OsString::from(SERVICE_RUN_ARG)],
+        dependencies: vec![],
+        // account_name: None => LocalSystem (the SPEC-018 requirement).
+        account_name: None,
+        account_password: None,
+    };
+
+    let service = create_service_with_retry(&manager, &service_info, deleted_existing)
+        .context("failed to create the GuruConnect managed agent service")?;
+
+    service
+        .set_description(SERVICE_DESCRIPTION)
+        .context("failed to set the service description")?;
+
+    configure_recovery();
+
+    info!(
+        "installed {SERVICE_NAME} (LocalSystem, auto-start) -> {} {}",
+        exe_path.display(),
+        SERVICE_RUN_ARG
+    );
+    Ok(())
+}
+
+/// Create the service, retrying briefly if the SCM still has the prior instance
+/// "marked for deletion" (SPEC-018 finding L1).
+///
+/// When a service is deleted, the SCM only removes it from its database once every
+/// open handle to it closes; until then a fresh `CreateService` fails with
+/// `ERROR_SERVICE_MARKED_FOR_DELETE` (1072). The previous implementation papered
+/// over this with a fixed 2s sleep after `delete()`, which is both slower than
+/// necessary in the common case and still racy on a busy box. Instead we attempt
+/// the create immediately and, only if we just deleted an existing instance and
+/// hit 1072, retry a few times with short backoff — succeeding as soon as the SCM
+/// finishes the removal, and giving up with the real error if it never does.
+///
+/// The retry is gated on `deleted_existing`: on a clean first install there was no
+/// prior instance, so a 1072 there is unexpected and is surfaced immediately
+/// rather than masked by retries.
+fn create_service_with_retry(
+    manager: &ServiceManager,
+    service_info: &ServiceInfo,
+    deleted_existing: bool,
+) -> Result<windows_service::service::Service, windows_service::Error> {
+    // ERROR_SERVICE_MARKED_FOR_DELETE (winerror.h). The service is gone from the
+    // caller's perspective but the SCM has not finished reaping it.
+    const ERROR_SERVICE_MARKED_FOR_DELETE: i32 = 1072;
+    // Bounded: ~5 attempts over ~2s total worst case (matches the old fixed sleep
+    // ceiling) but returns the instant the SCM is ready.
+    const MAX_ATTEMPTS: u32 = 5;
+    const BACKOFF: Duration = Duration::from_millis(400);
+
+    let mut attempt = 0;
+    loop {
+        attempt += 1;
+        match manager.create_service(service_info, ServiceAccess::CHANGE_CONFIG) {
+            Ok(service) => return Ok(service),
+            Err(windows_service::Error::Winapi(ref io_err))
+                if deleted_existing
+                    && io_err.raw_os_error() == Some(ERROR_SERVICE_MARKED_FOR_DELETE)
+                    && attempt < MAX_ATTEMPTS =>
+            {
+                warn!(
+                    "{SERVICE_NAME} still marked for deletion by the SCM \
+                     (attempt {attempt}/{MAX_ATTEMPTS}); retrying in {}ms",
+                    BACKOFF.as_millis()
+                );
+                std::thread::sleep(BACKOFF);
+            }
+            Err(e) => return Err(e),
+        }
+    }
+}
+
+/// Configure SCM crash-recovery so the service restarts on unexpected exit.
+///
+/// `windows-service` 0.7 does not expose `ChangeServiceConfig2` recovery actions
+/// in a stable, ergonomic form, so we mirror the established pattern used by the
+/// SAS service binary and shell out to `sc failure`. `reset=86400` clears the
+/// failure count after a day; three `restart/5000` actions retry after 5s each.
+fn configure_recovery() {
+    use std::os::windows::process::CommandExt;
+    const CREATE_NO_WINDOW: u32 = 0x0800_0000;
+
+    match std::process::Command::new("sc")
+        .args([
+            "failure",
+            SERVICE_NAME,
+            "reset=86400",
+            "actions=restart/5000/restart/5000/restart/5000",
+        ])
+        .creation_flags(CREATE_NO_WINDOW)
+        .output()
+    {
+        Ok(out) if out.status.success() => {
+            info!("configured crash-recovery (restart) for {SERVICE_NAME}");
+        }
+        Ok(out) => {
+            warn!(
+                "could not configure crash-recovery for {SERVICE_NAME} (sc failure exit {:?}); \
+                 the service will still run but will not auto-restart on crash",
+                out.status.code()
+            );
+        }
+        Err(e) => {
+            warn!("could not invoke `sc failure` to set crash-recovery for {SERVICE_NAME}: {e}");
+        }
+    }
+}
+
+/// Stop (if running) and delete the managed agent service. Idempotent: succeeds
+/// quietly if the service is not installed.
+pub fn uninstall_service() -> Result<()> {
+    let manager = ServiceManager::local_computer(None::<&str>, ServiceManagerAccess::CONNECT)
+        .context("failed to connect to the Service Control Manager (run as Administrator)")?;
+
+    match manager.open_service(
+        SERVICE_NAME,
+        ServiceAccess::QUERY_STATUS | ServiceAccess::STOP | ServiceAccess::DELETE,
+    ) {
+        Ok(service) => {
+            stop_if_running(&service);
+            service
+                .delete()
+                .context("failed to delete the managed agent service")?;
+            info!("uninstalled {SERVICE_NAME} service");
+            Ok(())
+        }
+        Err(_) => {
+            // Not installed — nothing to do (idempotent uninstall).
+            info!("{SERVICE_NAME} service is not installed; nothing to uninstall");
+            Ok(())
+        }
+    }
+}
+
+/// Start the managed agent service now (used right after a first-run install so
+/// the agent comes up without waiting for the next boot). Best-effort: logs and
+/// returns the SCM error if the start fails, but a failure is not fatal to install
+/// because the service is auto-start and will come up on the next boot regardless.
+pub fn start_service() -> Result<()> {
+    let manager = ServiceManager::local_computer(None::<&str>, ServiceManagerAccess::CONNECT)
+        .context("failed to connect to the Service Control Manager")?;
+    let service = manager
+        .open_service(
+            SERVICE_NAME,
+            ServiceAccess::START | ServiceAccess::QUERY_STATUS,
+        )
+        .context("failed to open the managed agent service to start it")?;
+
+    // If it is already running (e.g. reinstall-over-running), there is nothing to do.
+    if let Ok(status) = service.query_status() {
+        if status.current_state == ServiceState::Running
+            || status.current_state == ServiceState::StartPending
+        {
+            info!("{SERVICE_NAME} is already running/starting");
+            return Ok(());
+        }
+    }
+
+    service
+        .start::<String>(&[])
+        .context("failed to start the managed agent service")?;
+    info!("started {SERVICE_NAME}");
+    Ok(())
+}
+
+/// Report whether the managed agent service is currently installed.
+pub fn is_service_installed() -> bool {
+    match ServiceManager::local_computer(None::<&str>, ServiceManagerAccess::CONNECT) {
+        Ok(manager) => manager
+            .open_service(SERVICE_NAME, ServiceAccess::QUERY_STATUS)
+            .is_ok(),
+        Err(_) => false,
+    }
+}
+
+/// Best-effort stop of a service, waiting briefly for it to leave the running
+/// state so a subsequent `delete` does not race an in-flight stop.
+fn stop_if_running(service: &windows_service::service::Service) {
+    if let Ok(status) = service.query_status() {
+        if status.current_state != ServiceState::Stopped {
+            info!("stopping {SERVICE_NAME} before delete");
+            let _ = service.stop();
+            for _ in 0..10 {
+                std::thread::sleep(Duration::from_millis(500));
+                match service.query_status() {
+                    Ok(s) if s.current_state == ServiceState::Stopped => break,
+                    _ => continue,
+                }
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The launch argument the service is registered with MUST equal the hidden
+    /// `service-run` subcommand `main.rs` dispatches into [`run_dispatcher`]; a
+    /// mismatch would register a service the SCM could start but that would fall
+    /// through to normal (non-service) mode and immediately exit.
+    ///
+    /// This pins the value of the constant itself. The companion test
+    /// `tests::service_run_subcommand_matches_scm_launch_arg` in `main.rs` pins the
+    /// other half — that the clap `#[command(name = "service-run")]` attribute on
+    /// `Commands::ServiceRun` resolves to this same constant — so the two string
+    /// literals cannot silently drift apart.
+    #[test]
+    fn service_run_arg_matches_subcommand_name() {
+        assert_eq!(SERVICE_RUN_ARG, "service-run");
+    }
+
+    /// Service identifiers are non-empty and the internal name carries no spaces
+    /// (the SCM key / `sc` argument must be a single token).
+    #[test]
+    fn service_identifiers_are_well_formed() {
+        assert!(!SERVICE_NAME.is_empty());
+        assert!(
+            !SERVICE_NAME.contains(char::is_whitespace),
+            "the SCM service name must be a single whitespace-free token"
+        );
+        assert!(!SERVICE_DISPLAY_NAME.is_empty());
+        assert!(!SERVICE_DESCRIPTION.is_empty());
+    }
+
+    /// `is_service_installed` must never panic regardless of elevation/SCM access;
+    /// on a dev workstation without the service installed it returns `false`. (We
+    /// do NOT install the service in tests — that is a VM/admin integration step.)
+    #[test]
+    fn is_service_installed_is_total() {
+        let _ = is_service_installed();
+    }
+}

agent/src/session/mod.rs

@@ -41,8 +41,18 @@ use crate::proto::{message, AgentStatus, ChatMessage, Heartbeat, HeartbeatAck, M
 use crate::transport::WebSocketTransport;
 use crate::tray::{TrayAction, TrayController};
 use anyhow::Result;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::Arc;
 use std::time::{Duration, Instant};
 
+/// Sentinel error string returned by [`SessionManager::run_with_tray`] when the
+/// loop breaks because the SCM asked the managed-agent service to stop (SPEC-018,
+/// finding H). The outer `run_agent` loop matches on this to treat the exit as a
+/// graceful service stop (clean WS close, no reconnect) rather than a session
+/// error. Only the service path passes a shutdown flag, so only the service path
+/// can ever produce this.
+pub const SERVICE_STOP_SENTINEL: &str = "SERVICE_STOP";
+
 // Heartbeat interval (30 seconds)
 const HEARTBEAT_INTERVAL: Duration = Duration::from_secs(30);
 // Status report interval (60 seconds)
@@ -285,16 +295,34 @@ impl SessionManager {
         Ok(())
     }
 
-    /// Run the session main loop with tray and chat event processing
+    /// Run the session main loop with tray and chat event processing.
+    ///
+    /// `service_shutdown` (SPEC-018 finding H) is the SCM cooperative-stop flag.
+    /// It is `Some(flag)` ONLY on the managed-agent service path; the
+    /// attended/viewer/interactive callers pass `None` and behave EXACTLY as
+    /// before. When present, the flag is polled on every idle tick (the natural
+    /// ~100ms seam below) so an SCM Stop/Shutdown received while CONNECTED breaks
+    /// this inner loop promptly — instead of only being observed by the outer
+    /// `run_agent` reconnect loop, which never runs while a session is connected.
+    /// On a set flag the loop closes the WebSocket cleanly (via the shared exit
+    /// path at the bottom) and returns the [`SERVICE_STOP_SENTINEL`] error, which
+    /// the outer loop maps to a graceful stop.
     pub async fn run_with_tray(
         &mut self,
         tray: Option<&TrayController>,
         chat: Option<&ChatController>,
+        service_shutdown: Option<&Arc<AtomicBool>>,
     ) -> Result<()> {
         if self.transport.is_none() {
             anyhow::bail!("Not connected");
         }
 
+        // Helper: has the SCM asked the service to stop? Always false off the
+        // service path (where `service_shutdown` is `None`).
+        let stop_requested = |flag: Option<&Arc<AtomicBool>>| -> bool {
+            flag.is_some_and(|f| f.load(Ordering::SeqCst))
+        };
+
         // Send initial status
         self.send_status().await?;
 
@@ -307,6 +335,29 @@ impl SessionManager {
 
         // Main loop
         loop {
+            // SPEC-018 (finding H): honour an SCM stop request received while the
+            // session is CONNECTED. The outer `run_agent` loop only observes the
+            // flag between connection attempts, but a managed agent spends its
+            // entire connected life inside THIS loop — so without this check an
+            // SCM Stop while connected would not break out until the connection
+            // dropped on its own. Breaking here falls through to the shared exit
+            // path below, which closes the transport cleanly (clean WS close);
+            // the sentinel tells the outer loop this was a graceful stop.
+            if stop_requested(service_shutdown) {
+                tracing::info!("Service stop requested; ending connected session loop");
+                self.release_streaming();
+                self.state = SessionState::Disconnected;
+                if let Some(transport) = self.transport.as_mut() {
+                    // Best-effort clean WebSocket close (sends a Close frame). A
+                    // failure here just means the peer/socket is already gone; the
+                    // service still stops cleanly.
+                    if let Err(e) = transport.close().await {
+                        tracing::warn!("error during clean WebSocket close on service stop: {}", e);
+                    }
+                }
+                return Err(anyhow::anyhow!(SERVICE_STOP_SENTINEL));
+            }
+
             // Process tray events
             if let Some(t) = tray {
                 if let Some(action) = t.process_events() {
@@ -745,3 +796,47 @@ impl SessionManager {
         Ok(())
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// SPEC-018 finding H: the connected-stop contract. When the SCM sets the
+    /// shutdown flag, `run_with_tray` returns an error whose message contains
+    /// [`SERVICE_STOP_SENTINEL`]; the outer `run_agent` loop recognises a graceful
+    /// stop with `error_msg.contains(SERVICE_STOP_SENTINEL)`. This pins that the
+    /// error the loop constructs on stop actually satisfies that match — so the
+    /// two halves (producer here, consumer in `main.rs`) cannot drift.
+    ///
+    /// A full end-to-end test of the in-loop interrupt would need a live connected
+    /// transport (a real or mocked server), which is an integration concern; this
+    /// unit test instead pins the wire contract the interrupt relies on.
+    #[test]
+    fn service_stop_sentinel_is_matched_by_outer_loop_check() {
+        let produced = anyhow::anyhow!(SERVICE_STOP_SENTINEL);
+        assert!(
+            produced.to_string().contains(SERVICE_STOP_SENTINEL),
+            "the stop error must contain the sentinel the outer loop matches on"
+        );
+        assert!(
+            !SERVICE_STOP_SENTINEL.is_empty(),
+            "the sentinel must be a non-empty, distinctive token"
+        );
+    }
+
+    /// The shutdown-flag check is a no-op (always `false`) when no flag is passed,
+    /// i.e. on the attended/viewer/interactive paths — guaranteeing the new
+    /// parameter is a pure addition that cannot alter non-service behaviour
+    /// (SPEC-018 finding H: "no regression").
+    #[test]
+    fn no_shutdown_flag_never_requests_stop() {
+        let none: Option<&Arc<AtomicBool>> = None;
+        let check = |flag: Option<&Arc<AtomicBool>>| flag.is_some_and(|f| f.load(Ordering::SeqCst));
+        assert!(!check(none));
+
+        let set = Arc::new(AtomicBool::new(true));
+        assert!(check(Some(&set)));
+        let unset = Arc::new(AtomicBool::new(false));
+        assert!(!check(Some(&unset)));
+    }
+}

docs/FEATURE_ROADMAP.md

@@ -16,11 +16,16 @@ stack. It ships independently of GuruRMM and integrates with it via a versioned
 > match, blacklist-on-WS, agent-plane rejects user JWTs via per-agent `cak_` keys). The feature specs below
 > (SPEC-003–009) are **work-items inside the later v2 phases** — see the mapping.
 >
-> **Remaining to formally exit Phase 1:** secure-session-core **Task 8** (end-to-end verification +
+> **Phase 1 formally EXITED (2026-05-31).** secure-session-core **Task 8** is complete — end-to-end
-> `/gc-audit --pass=security` re-audit + the manual CRITICAL checks) and Code-Review sign-off on Tasks 3–5
+> functional verification (live CRITICAL boundary checks against the deployed binary: login-JWT→401,
-> (implemented without a local toolchain at the time; since built + deployed). Live HW-H.264 validation is
+> wrong-session viewer token→403, JWT-as-agent-key→401) **plus the `/gc-audit --pass=security` re-audit:
-> also pending — raw+Zstd remains the shipping default. ~~Sprint 0 (relay-auth CRITICAL hotfix)~~ **not
+> PASS, 0 CRITICAL/HIGH/MEDIUM/LOW** ([report](../reports/2026-05-31-gc-audit.md)). Code-Review sign-off on
-> needed — those fixes shipped in Tasks 2–3.**
+> Tasks 3–5 landed earlier. On top of Phase 1, **SPEC-004 (Tasks 2/4/5 — machine_uid dedup, session
+> reaping, operator removal API+UI) is implemented, reviewed, deployed, and the 11 live ghost rows were
+> purged**; the agent is now **auto-versioned + Azure-Trusted-Signing-signed via `release.yml`** with
+> **v0.3.0 published** as the stable release. ~~Sprint 0 (relay-auth CRITICAL hotfix)~~ **not needed.**
+> Still pending (NOT a Phase-1 blocker): live HW-H.264 cross-GPU validation — **raw+Zstd remains the
+> shipping default** (`DEFAULT_PREFER_H264=false`) until H.264 is validated across GPUs.
 
 ### v2 phase mapping of current specs
 
@@ -43,8 +48,9 @@ stack. It ships independently of GuruRMM and integrates with it via a versioned
 
 Bringing GC to parity with GuruRMM's release engineering. Full plan: [SPEC-001](specs/SPEC-001-operational-tooling-parity.md).
 
-- [ ] **Code signing — Azure Trusted Signing in CI** — P1 — sign the Windows agent `.exe` via `jsign` (TRUSTEDSIGNING) in Gitea Actions, reusing the shared ACG cert profile. (SPEC-001 §2)
+- [x] **Code signing — Azure Trusted Signing in CI** — P1 — Windows agent `.exe` signed via `jsign` (TRUSTEDSIGNING) in `release.yml`, fail-closed (never publishes unsigned). Shipped with v0.3.0. (SPEC-001 §2)
-- [ ] **Automatic versioning** — P1 — conventional-commit-driven version bump across agent/server/dashboard, embedded via `build.rs`. (SPEC-001 §3)
+- [ ] **Signed beta/test release channel** — **P1 — NOW** — every binary we hand to a tester must be signed, but signing today only runs on a deliberate full `release.yml` dispatch; the automatic `build-and-test.yml` agent artifact is explicitly **unsigned**. Add a `channel: stable | beta` `workflow_dispatch` input to `release.yml`: `beta` signs the agent and publishes a prerelease-tagged Gitea release (e.g. `v0.4.0-beta.1`) **skipping the semver bump + changelog**; `stable` keeps the existing full path. Keeps signing secrets out of PR-triggered runs. (SPEC-001 §2)
+- [x] **Automatic versioning** — P1 — conventional-commit-driven version bump computed at dispatch in `release.yml`, embedded via `build.rs`. Shipped with v0.3.0. (SPEC-001 §3)
 - [ ] **Changelog generation & API** — P2 — `CHANGELOG.md` + per-version changelogs from conventional commits, served at `/api/changelog/...`. (SPEC-001 §4)
 - [ ] **Feature-request workflow** — P2 — `/gc-feature-request` skill producing `docs/specs/SPEC-NNN-*.md` and updating this roadmap. (SPEC-001 §1)
 - [ ] **Roadmap / ADR / spec tracking** — P1 — this file + `ARCHITECTURE_DECISIONS.md` + `docs/specs/`. (SPEC-001 §5) — *bootstrapped*
@@ -62,6 +68,7 @@ Bringing GC to parity with GuruRMM's release engineering. Full plan: [SPEC-001](
 - [x] Protobuf-over-WSS transport, Zstd frame compression
 - [~] React/TS web viewer (`dashboard/src/components/RemoteViewer.tsx`) — embeddable session viewer
 - [ ] **Headless Linux mode (direct TTY access)** — P2 — Terminal-based remote access for Linux servers without GUI. PTY spawn (`openpty`), xterm.js web viewer, full ANSI/VT100 support. Enables server management, container debugging, emergency recovery via GuruConnect dashboard with audit logging. SSH replacement with centralized auth. ([SPEC-012](specs/SPEC-012-headless-linux-tty.md))
+- [ ] **Managed-agent SYSTEM service host + session broker** — P1 — convert the persistent agent from `HKCU Run` (user context) to a LocalSystem **service** that runs unattended (login screen, no user, across reboots) and spawns a per-session capture/input worker into the active desktop (Session 0 can't capture directly). Unblocks SPEC-016 Phase B end-to-end (the SYSTEM-ACL'd `cak_` store becomes readable; removes the Phase B fail-fast guard), enables true unattended access, and is the **broker primitive SPEC-013 builds on**. ([SPEC-018](specs/SPEC-018-managed-agent-service-host.md))
 - [ ] **Windows session selection and backstage mode** — P2 — Enumerate and switch between Windows user sessions (Terminal Services/RDP/Fast User Switching) and access Session 0 (backstage) for system-level admin tasks. ScreenConnect parity: session selector shows all logged-on users, instant switching without reconnect. Backstage mode provides terminal/command interface for services management without disrupting any user desktop. Critical for multi-user server environments. ([SPEC-013](specs/SPEC-013-session-selection-and-backstage.md))
 - [ ] **Configurable notification overlay on viewer connection** — P2 — Display a semi-transparent on-screen notification when a technician connects, showing technician name and company. Dashboard-configurable message template (supports `{{technician_name}}`, `{{company}}`, `{{time}}`), duration (5-60s), position (top-left/right, bottom-left/right, center), and dismissible behavior. Increases transparency and user awareness during remote support sessions. Compliance-friendly for privacy policies requiring user notification. ([SPEC-015](specs/SPEC-015-notification-overlay.md))
 - [ ] Multi-monitor switching — P2
@@ -81,13 +88,15 @@ Bringing GC to parity with GuruRMM's release engineering. Full plan: [SPEC-001](
 - [x] Sessions / machines / support-codes / events
 - [ ] **Full machine inventory in the connection DB** — P2 — persist per-machine device inventory (OS+locale+install, CPU/RAM, mfr/model/serial, external WAN IP captured server-side + private LAN IP + MAC, logged-on user, idle, time zone, uptime, local-admin) on `connect_machines`, refreshed each `AgentStatus`, shown in the dashboard machine detail (ScreenConnect "Guest Info" parity). Data layer for SPEC-002 Phase 2; closes GC side of agent-IP gap (todo 7459428e). **[→ v2 Phase 2]** ([SPEC-003](specs/SPEC-003-machine-inventory.md))
 - [ ] **Stable machine identity + session lifecycle reaping + operator removal** — P1 — give the agent a deterministic machine-derived `machine_uid` (Windows `MachineGuid`-based) so the same box can't register duplicates (root cause: `agent_id` is a config-file random UUID that a portable/misconfigured run regenerates each launch); key registration on it; add TTL reaping + same-machine supersede as defense-in-depth; and admin-gated per-row + multi-select bulk removal of stale sessions/units. Identity must be bound to the per-machine agent key (spoof guard). Fixes ghost-session accumulation seen on the live console (15 sessions / 0 live, ~10 orphans for one machine). **[→ v2 Phase 1]** ([SPEC-004](specs/SPEC-004-session-lifecycle-and-removal.md))
+- [ ] **Zero-touch per-site agent enrollment** — P1 — ScreenConnect-class managed enrollment: one signed installer per site, machines self-register on first run and the server mints a per-machine `cak_` bound to a deterministic `machine_uid` (dedups re-installs). Per-site **rotatable** enrollment key (long secret + `vN (XXXX)` fingerprint) — rotating blocks new enrollments from old installers, leaves enrolled agents untouched. Auto-approve + new-enrollment/site-move alert. **Sign base agent once (CI, shipped) + per-site signed wrapper that writes site config around the signed bytes — resolves SPEC-007's signature-vs-appended-config question.** Anticipated/deferred: enrollment policy + licensing, `--enroll-key`/`--reassign` flag overrides, technician-assisted interactive install. **[→ v2 Phase 1]** ([SPEC-016](specs/SPEC-016-zero-touch-enrollment.md))
 - [ ] **Machines list view — dual connection indicators + rich rows** — P2 — ScreenConnect "Access"-list parity: per-row Host/Guest two-segment connection bar (Guest=agent online, Host=viewer connected, with names + durations) and rich inline metadata (company, site, device type, tags, logged-on user + idle, client version in red when outdated). Server-enriches `/api/machines` with live session state + SPEC-003 inventory. **[→ v2 Phase 2]** ([SPEC-005](specs/SPEC-005-machines-list-view-parity.md))
 - [ ] Machines "by Company" tree nav with per-company counts — P3 — left-nav grouping sidebar (screenshot parity). Follow-up sub-item of SPEC-005.
 - [ ] **Universal machine search ("everything is searchable")** — P2 — server-side `?q=` on `/api/machines` matching case-insensitive substring across ALL attributes (OS, logged-on user, external/private IP, company, site, tag, serial, MAC, version, …), pg_trgm GIN-indexed; multi-term AND + optional field-scoped syntax (`os:`, `user:`, `ip:`). Replaces the hostname-only client filter. Depends on SPEC-003 (attrs must be persisted). **[→ v2 Phase 2]** ([SPEC-006](specs/SPEC-006-universal-machine-search.md))
-- [ ] **Managed-agent installer builder ("Build Installer")** — P2 — dashboard wizard to build a pre-labeled persistent-agent installer (Name/Company/Site/Department/Device Type/Tag/Type) with Download / Copy URL / Send Link, reusing the existing embed-config download path; adds department + device_type to EmbeddedConfig/AgentStatus so labels persist at install time. Pairs with revocable per-machine keys; signature-vs-appended-config is the key open question. **[→ v2 Phase 2]** ([SPEC-007](specs/SPEC-007-managed-agent-installer-builder.md))
+- [ ] **Managed-agent installer builder ("Build Installer")** — P2 — dashboard wizard to build a pre-labeled persistent-agent installer (Name/Company/Site/Department/Device Type/Tag/Type) with Download / Copy URL / Send Link, reusing the existing embed-config download path; adds department + device_type to EmbeddedConfig/AgentStatus so labels persist at install time. Pairs with revocable per-machine keys; the signature-vs-appended-config question is resolved by SPEC-016 (sign-once base + per-site signed wrapper, no PE append). **[→ v2 Phase 2]** ([SPEC-007](specs/SPEC-007-managed-agent-installer-builder.md))
 - [ ] **Valuable error messages (structured errors + no silent swallows)** — P2 — one structured API error envelope with stable codes + a correlation id that also lands in the logs; contextual tracing on server/agent; sweep the 37 `let _ =` swallows (the pattern that hid the migration-005 bug); dashboard surfaces the real cause + id instead of a generic line. **[→ v2 Phase 0/1 conventions]** ([SPEC-008](specs/SPEC-008-valuable-error-messages.md))
 - [ ] **Feature-rich, fully-documented management API** — P2 — everything the console can do, callable by API: OpenAPI 3.x generated from code (utoipa) + browsable docs at `/api/docs`, long-lived revocable scoped API tokens (PAT-style, distinct from the 24h JWT + agent keys), an API-completeness gap audit, and consistent pagination/error conventions. Distinct from the ADR-001 RMM integration contract. **[→ v2 Phase 3]** ([SPEC-009](specs/SPEC-009-feature-rich-documented-api.md))
 - [ ] **Branding and white-label configuration** — P2 — Allow MSPs to customize logo, colors, and product name for white-labeled remote support. Dashboard admin settings page with logo upload (PNG/SVG, max 2MB), brand hue slider (OKLCH 0-360°, default 184=cyan), product name override, company name, and favicon. Agent tray tooltip uses custom product name from registry. Singleton database table with public GET endpoint for unauthenticated rendering. CSS variables (`--brand-hue`, `--accent`, `--panel`) for dynamic theming. **[→ v2 Phase 2]** ([SPEC-014](specs/SPEC-014-branding-whitelabel.md))
+- [ ] **End-user (sub-user) remote access** — P2 (may be P3) — let a client pay for their employees to reach their *own* machines from home: a deny-by-default `end_user` login role, a locked-down end-user portal listing only granted machines, and Connect reusing the existing session-scoped viewer-token + relay path. Grant primitive already exists (`user_client_access`, migration 002); directory sync (AD/Entra/Google) is a separate future spec. **[→ new capability, post v2-console]** ([SPEC-017](specs/SPEC-017-end-user-remote-access.md))
 - [ ] Programmatic session pre-create + viewer-token (integration contract) — P2
 
 ## Security & Infrastructure

docs/specs/SPEC-016-zero-touch-enrollment.md

@@ -0,0 +1,244 @@
+# SPEC-016: Zero-Touch Per-Site Agent Enrollment
+
+**Status:** Proposed
+**Priority:** P1
+**Requested By:** Mike (2026-06-02)
+**Estimated Effort:** X-Large
+
+## Overview
+
+Give GuruConnect a ScreenConnect-class managed-agent enrollment flow: a technician runs
+**one signed installer per site** on every machine at that site — no per-machine key
+minting, no flags, no typing — and each machine **self-registers** on first run, the
+server minting it a per-machine `cak_` key bound to a stable, machine-derived
+`machine_uid`. Each site installer carries a **rotatable per-site enrollment key** (a long
+server-generated secret) plus a short human-readable **fingerprint** (`vN (XXXX)`) so an
+operator can tell at a glance whether an installer is current. Rotating a site's key blocks
+*new* enrollments from old installers while leaving already-enrolled machines untouched
+(they hold their own `cak_`).
+
+This is the missing piece that turns the v2 secure-session-core (SPEC-004 per-agent keys +
+`machine_uid`) into a real product workflow, and it **resolves SPEC-007's open
+signature-vs-appended-config question**: the agent binary is signed **once** in CI
+(already shipped via `release.yml`), and per-site customization rides in a thin **signed
+wrapper** that writes site config to the endpoint at install time — never appended into the
+signed PE.
+
+**Success criteria:**
+1. A tech installs one site installer on N machines; all N appear in the console under the
+   correct company/site, each as a distinct, deduplicated machine — zero per-machine setup.
+2. Re-installing / re-imaging the same hardware **reuses** the existing machine row (no
+   ghost duplicates — the failure mode SPEC-004 documents).
+3. Rotating a site's enrollment key makes old installers unable to enroll new machines,
+   while every already-enrolled agent keeps working.
+4. Every distributed installer is **validly Authenticode-signed** (SmartScreen/WDAC clean).
+
+## Background — what exists today (confirmed in code)
+
+- **Embedded config is append-based and breaks signing.** `server/src/api/downloads.rs`
+  (`download_agent`, ~`:152`) reads `static/downloads/guruconnect.exe` and **appends**
+  `MAGIC_MARKER` + `len:u32` + JSON (`:196`) to the end of the PE. The agent reads it back
+  in `agent/src/config.rs` (`read_embedded_config`, `:223`). Appending bytes after a signed
+  PE invalidates the Authenticode signature — so the current customization path and the
+  newly-shipped CI signing are mutually exclusive.
+- **No self-registration exists.** Per-agent `cak_` keys are minted **admin-only** in
+  `server/src/api/machine_keys.rs` (`create_key`, `:119`; "Admin issued a per-agent key",
+  `:146`). There is no endpoint where an agent first-run exchanges an enrollment credential
+  for its own key.
+- **Relay already accepts per-agent keys.** `server/src/relay/mod.rs`
+  (`validate_agent_api_key`, `:417`) calls `crate::auth::agent_keys::verify_agent_key`
+  (`:422`) — the `cak_` path — then falls back to the **deprecated** shared `AGENT_API_KEY`
+  (`:444`, logs a "migrate to per-agent `cak_`" warning).
+- **Key primitives exist.** `server/src/auth/agent_keys.rs`: `generate_agent_key` mints a
+  `cak_`-prefixed high-entropy key (`:36`/`:46`); `verify_agent_key` (`:71`).
+  `server/src/db/agent_keys.rs` already inserts into `connect_agent_keys (machine_id,
+  key_hash, tenant_id)` (`:47`) — the v2 tenancy column is present (migration
+  `004_v2_secure_session_core.sql`).
+- **Identity is a random config UUID, not machine-derived** — the root cause of duplicates
+  per SPEC-004 (`agent/src/config.rs` `generate_agent_id`, `:90`).
+- **Agent mode dispatch:** `agent/src/main.rs` `Commands::Install` (`:160`) → `run_install`;
+  `agent/src/config.rs` `detect_run_mode` (`:162`) returns `RunMode::PermanentAgent` when
+  embedded config is present.
+
+## Scope
+
+### Included in v1 (CORE)
+
+1. **`machine_uid` — deterministic machine identity (hardware-salted, per-tenant).** Derive
+   a stable id from the Windows `MachineGuid`
+   (`HKLM\SOFTWARE\Microsoft\Cryptography\MachineGuid`) **salted with stable hardware
+   signals** (SMBIOS UUID / motherboard + disk serial), independent of the config-file
+   `agent_id`. Hardware-derived salt is deliberate: it **survives an OS reinstall/re-image
+   on the same hardware** (so the row is reused — the re-image dedup goal) while keeping
+   distinct physical boxes distinct (a per-install *random* salt would break re-image dedup
+   and is rejected). Uniqueness is scoped **per-tenant** — dedup key `(tenant_id,
+   machine_uid)` — so the same hardware legitimately present in two tenants stays two
+   independent rows. (Shared root with SPEC-004; whichever lands first owns the impl, the
+   other consumes it.) Used as the dedup key for register/move.
+
+   **Collision-gated activation.** The residual collision case is VMs/templates that share a
+   hardware UUID (some hypervisors clone the SMBIOS UUID). When the server detects a
+   `machine_uid` collision (a seemingly-different endpoint resolving to an existing uid), the
+   endpoint does **not** auto-activate: it drops to a **pending** state, fires an alert, and
+   an operator must confirm in the dashboard that the collided endpoint may activate. This is
+   the one deliberate exception to auto-approve (see item 6).
+
+2. **Per-site enrollment key + fingerprint.**
+   - Long (≥256-bit) server-generated secret per site, stored **hashed** (Argon2id, same
+     as `cak_`/passwords), never recoverable in plaintext after issue.
+   - A non-secret **fingerprint** = monotonic version + short derived code in **hex**,
+     rendered `vN (XXXX)` (e.g. `v3 (7F2A)`), shown in the dashboard, baked into the
+     installer filename, and reported by the agent at enrollment. Hex is deliberate —
+     **not** the RMM word-style code (`GREEN-FALCON`) — so GuruConnect and GuruRMM
+     artifacts are never visually conflated.
+   - **Rotate** regenerates the secret and bumps the version; old installers are rejected
+     for *new* enrollments; existing agents (holding `cak_`) are unaffected.
+
+3. **Self-registration endpoint.** New `POST /api/enroll` (public, unauthenticated by JWT —
+   gated by the enrollment key) accepting `{ site_code, enrollment_key, machine_uid,
+   hostname, labels{company,site,department,device_type,tags} }`:
+   - Verify `(site_code, enrollment_key)` against the current per-site key.
+   - **Dedup by `machine_uid`** within the site: if the machine exists, reuse the row and
+     rotate its `cak_`; else create the machine row.
+   - Mint a `cak_` (reuse `generate_agent_key`), store hashed via `db::agent_keys` bound to
+     `machine_id` (+ `tenant_id` from the site), return the plaintext `cak_` **once**.
+   - Emit an audit event + **new-enrollment alert** (and a **site-move** alert when an
+     existing `machine_uid` enrolls under a different site).
+   - **Rate-limit + lockout** per `(site_code, source-IP)` as defense-in-depth (the key is
+     long, so this is belt-and-suspenders, not load-bearing).
+
+4. **Agent first-run enrollment.** On `RunMode::PermanentAgent` with no stored `cak_`:
+   read site config → call `/api/enroll` with `machine_uid` → persist the returned `cak_`
+   to a SYSTEM-only protected store (HKLM under a SYSTEM-only ACL, or DPAPI-machine) →
+   connect to `wss://connect.azcomputerguru.com/ws/agent` using the `cak_`. On subsequent
+   runs, use the stored `cak_` directly (no re-enroll).
+
+5. **Sign-once base + per-site signed wrapper (resolves SPEC-007 open question).**
+   - The base agent is signed once in CI (`release.yml`, already shipped) and stays
+     byte-identical for everyone.
+   - Per-site customization (labels + enrollment key + fingerprint) is delivered to the
+     endpoint **at install time** via a signing-safe channel — NOT appended to the signed
+     PE. **v1 produces BOTH a signed bootstrapper `.exe` and a signed MSI per site**
+     (ScreenConnect parity — manual installs grab the `.exe`, GPO/Intune fleet pushes take
+     the MSI), both wrapping the same sign-once agent and writing the site config to the
+     protected config location. The two differ only in packaging (bootstrapper stub vs. WiX
+     bundle); both are signed.
+   - **Deprecate the append path** in `downloads.rs` for managed installs (keep only for
+     attended/support-code if still needed), eliminating the signature-invalidation defect.
+
+6. **Auto-approve posture (with collision-gate exception).** A self-registered machine is
+   live and controllable immediately (ScreenConnect parity); the new-enrollment alert is the
+   tripwire. The **one** exception is a detected `machine_uid` collision (item 1), which
+   gates the endpoint to **pending** until an operator confirms it in the dashboard.
+
+### Explicitly out of scope (ANTICIPATED — reserve room, do NOT build in v1)
+
+The v1 data model and agent mode-dispatch must leave room for these without building them:
+
+- **Per-site enrollment POLICY** — a `sites.enrollment_policy` field (default
+  `auto-approve`; future `pending-approval`) plus per-seat/per-endpoint licensing controls.
+  Commercial, multi-tenant (the `tenant_id` column already exists). Its own future SPEC.
+- **Flag overrides** — `--enroll-key` / `--site-code` (generic installer, key supplied on
+  the command line) and `--reassign` (move an existing machine to a new site, gated by
+  possession of the destination site's key, with an **explicit accidental-move guard**:
+  a different-site re-run refuses unless `--reassign` is passed) + cross-client move policy.
+  Backend (`machine_uid` + authorized site + `cak_`) is designed to support it; CLI surface
+  is deferred.
+- **Technician-assisted interactive install** — `--technician` on a generic installer:
+  prompts for the tech's own server credentials, and on auth presents a **validated**
+  Company/Site/tags picker from the live authorized list (authz-by-identity, full audit
+  trail). Heaviest path (interactive UI + auth/list callback); deferred.
+
+All three converge on the **same backend operation** delivered in v1: `machine_uid` +
+authorized site + issued `cak_`. v1 only ships the per-site-embedded-key door.
+
+## Architecture
+
+- **Agent** (`agent/`): compute `machine_uid`; first-run enroll → store `cak_`; use stored
+  `cak_` thereafter; read site config from the wrapper-written location instead of an
+  appended PE blob. Touches `config.rs` (`EmbeddedConfig`/`detect_run_mode`/storage),
+  `main.rs` (`Install`/run-mode), a new `enroll` client module, transport auth.
+- **Relay-server** (`server/`): new `POST /api/enroll`; per-site key issue/rotate/verify;
+  `machine_uid` dedup + site-move on register; audit + alert emission; rate-limit/lockout.
+  Touches `api/` (new `enroll.rs`, `sites` key endpoints), `auth/agent_keys.rs`,
+  `db/agent_keys.rs`, `relay/mod.rs` (enrollment vs. connect), `main.rs` routes.
+- **Dashboard**: per-site enrollment-key display (fingerprint `vN (XXXX)`), **Rotate**
+  action, "current installer" download wired to the signed wrapper build. (Builder UI is
+  SPEC-007; this spec supplies the key/fingerprint/rotation it consumes.)
+- **DB migration:** `site_enrollment_keys` (or columns on the site): `site_id`,
+  `key_hash`, `version`, `fingerprint`, `created_at`, `rotated_at`, `active`. Reserve
+  `sites.enrollment_policy` (nullable, default `auto-approve`) for the anticipated policy
+  work. `connect_machines` gains `machine_uid` (unique per tenant/site).
+- **Protobuf** (`proto/guruconnect.proto`): no wire change required for enrollment if
+  `/api/enroll` is REST; `AgentStatus` label fields per SPEC-007 (`department`,
+  `device_type`) ride along if landed together.
+
+## Security considerations
+
+- **Two-tier credential model:** low-sensitivity **enrollment key** (gates "may register",
+  shared per site, rotatable) vs. high-sensitivity **per-machine `cak_`** (operating
+  credential, per-machine revocation). Compromise of an enrollment key is recovered by
+  rotating one site — no fleet-wide re-key.
+- **Enrollment keys stored hashed** (Argon2id); plaintext shown once at issue/rotate.
+- **`cak_` at rest on the endpoint** is stored as a **DPAPI-machine-encrypted blob inside a
+  SYSTEM-ACL'd location** (HKLM value or `ProgramData` file) — both layers: the SYSTEM ACL
+  stops non-admin users reading it, and DPAPI-machine encryption makes a copied file/export
+  inert off the box. (Local admin/SYSTEM can always recover it; that is accepted — blast
+  radius of one leaked `cak_` is a single, independently-revocable machine.)
+- **`machine_uid` binding** is the spoof-guard SPEC-004 wants: a `cak_` is bound to a
+  `machine_uid`; a different box presenting another box's `cak_` is detectable.
+- **Authorization model** for moves/enrolls is possession-of-destination-key in v1
+  (identity-based authz deferred to the technician-assisted path).
+- **Open registration risk** is mitigated by requiring `(site_code + long key)` and
+  rate-limit/lockout; auto-approve is acceptable because the enrollment key is the gate and
+  every enrollment/site-move fires an alert.
+- **Audit events:** enroll, re-enroll/reuse, site-move, key-rotate — all logged with
+  `machine_uid`, site, and source IP.
+
+## Testing strategy
+
+- **Unit:** `machine_uid` derivation stability; enrollment-key verify/rotate; fingerprint
+  derivation; `cak_` mint/hash/verify; dedup decision (new vs. reuse vs. move).
+- **Integration:** enroll new → row + `cak_` issued; re-enroll same `machine_uid` → reuse,
+  no duplicate; enroll with rotated (old) key → rejected; old `cak_` still connects after
+  rotation; rate-limit/lockout trips; site-move emits alert.
+- **Manual:** build a site wrapper installer → run on a clean VM → appears in console under
+  correct site, immediately controllable; re-image VM → same row reused; `signtool verify
+  /pa` passes on the distributed wrapper and the laid-down agent.
+
+## Effort estimate & dependencies
+
+- **Size:** X-Large (agent + relay + DB migration + CI build/sign wrapper + dashboard
+  key/rotation surface).
+- **Depends on:** SPEC-004 `machine_uid` (shared root); the CI signing already shipped
+  (SPEC-001 §2 / `release.yml`).
+- **Unblocks:** SPEC-007 (installer builder gets a real per-site key + the signing
+  resolution), and the parked managed-agent test deployment on the internal beta machines.
+- **Relationship to v2 phases:** sits with the Phase-1 secure-session-core (per-agent keys
+  + identity) and feeds Phase-2 dashboard work.
+
+## Resolved decisions (2026-06-02, Mike)
+
+1. **Wrapper shape — BOTH.** v1 ships a signed bootstrapper `.exe` *and* a signed MSI per
+   site (ScreenConnect offers both; manual installs use the `.exe`, GPO/Intune fleet pushes
+   use the MSI). Same sign-once agent inside each.
+2. **`cak_` storage — BOTH layers.** DPAPI-machine-encrypted blob stored in a SYSTEM-ACL'd
+   location. Non-admins can't read it; a stolen copy is inert off the box.
+3. **Fingerprint — hex (`7F2A`).** Deliberately *not* the RMM word-code style, so the two
+   products' artifacts are never visually conflated.
+4. **`machine_uid` — per-tenant scope, hardware-derived salt, collision-gated.** Dedup key
+   `(tenant_id, machine_uid)`; salt from stable hardware signals (survives same-hardware
+   re-image, separates distinct boxes); detected collisions (e.g. template-cloned VMs
+   sharing a hardware UUID) drop to pending + alert and require dashboard confirmation to
+   activate.
+5. **Attended (support-code) path — unchanged.** `download_support` is filename-based
+   (`GuruConnect-<code>.exe`), not append-based, so renaming never breaks the signature —
+   it is already signing-safe. Only the managed `download_agent` append path is retired.
+
+## Remaining for planning
+
+- Exact stable-hardware signal set for the salt (SMBIOS UUID alone vs. + motherboard/disk
+  serial) and hypervisor behavior matrix (which hypervisors duplicate the SMBIOS UUID on
+  clone → exercise the collision-gate).
+- MSI authoring approach (WiX) and whether per-site config rides as a per-site MSI vs. a
+  base MSI + property/transform.

docs/specs/SPEC-017-end-user-remote-access.md

@@ -0,0 +1,180 @@
+# SPEC-017: End-User (Sub-User) Remote Access
+
+**Status:** Proposed
+**Priority:** P2 (may settle to P3 depending on client demand)
+**Requested By:** Mike (2026-06-02)
+**Estimated Effort:** Large
+
+## Overview
+
+Let a client pay for their own employees to remotely reach **their own work machines** from home
+through GuruConnect — the Splashtop-Business / unattended-end-user-access model, layered on top of the
+MSP-technician console GuruConnect ships today. An MSP admin (or, later, a delegated client-company
+admin) provisions a list of **end-users** and grants each one access to specific managed machines. The
+end-user signs into a locked-down **end-user portal**, sees only the machines granted to them, and
+connects — reusing the existing persistent-agent + session-scoped-viewer-token + relay path.
+
+Success criteria: an `end_user`-role account can log in at a separate portal, see exactly the machines
+in its grant set (and no others, across no other tenant), launch a control session to an online granted
+machine, and is hard-denied from every technician/admin API, the agent plane, and any machine it was
+not granted — with each login and machine access written to the audit log.
+
+This is a net-new **sellable capability**, not a console-MVP blocker. It is sequenced after the v2
+console foundations it depends on (tenancy, machine identity, persistent enrollment), which is why it is
+P2 rather than P1.
+
+## Scope
+
+### Included in v1
+- A new **`end_user`** value for `users.role`, provisioned by an MSP admin, with **deny-by-default**
+  authority: no console permissions, no agent-plane access, machine reach limited strictly to its
+  `user_client_access` grant set within its own tenant.
+- A **separate end-user login + portal** route (locked-down): lists only granted machines with
+  online/offline state and a Connect action. No admin nav, no other users/machines/companies.
+- **Admin UI + API** to create/disable end-users and assign/revoke per-machine grants, reusing the
+  existing `user_client_access` table.
+- **Connect flow** that reuses the landed session-scoped viewer-token mechanism (`ViewerClaims`,
+  `jwt.rs:114`) and the relay enforcement path — no new transport.
+- A new `connect_sessions.source` value **`end_user`** (migration widening the existing CHECK).
+- **Audit**: end-user login success/failure and each machine-access grant-check written to
+  `connect_session_events`.
+- Rate limiting + lockout on the public end-user login.
+
+### Explicitly out of scope (v1)
+- **Directory sync (AD / Entra-365 / Google) → end-user list** — its own future spec; v1 is manual
+  list management only.
+- **Self-service seat purchasing / billing automation.** v1 records/counts seats per tenant; real
+  metering and Syncro/billing wiring is deferred.
+- **Delegated client-company-admin role** (a client managing its own end-users/grants) — noted as a
+  fast-follow; v1 grants are MSP-admin-managed.
+- Per-session view-only-vs-control *policy* per end-user (v1 = Control of one's own machine; the
+  `ViewerAccess` split still exists at the token layer).
+- File transfer, session recording (already out of scope for the broader product v1).
+
+## Architecture
+
+### Principal model — `end_user` is a constrained variant of the login plane
+GuruConnect already has three credential planes that must stay separate (audit-hardened in v2 Phase 1):
+1. **Login `Claims`** (`jwt.rs:11`) — dashboard users; `role ∈ {admin, operator, viewer}` today.
+2. **Session-scoped `ViewerClaims`** (`jwt.rs:114`) — 5-min, one session, `purpose=viewer`.
+3. **Agent `cak_` keys** (`connect_agent_keys`, migration 004) — agents only.
+
+`end_user` is added as a **fourth role on the login plane** — it issues a normal login JWT
+(`create_token`, `jwt.rs:161`) carrying `role: "end_user"` and an **empty permission list**. The
+separation guarantees the v2 audit established are preserved: an `end_user` JWT still cannot be used as
+a viewer token (lacks `purpose`) nor as an agent key (agent plane rejects user JWTs).
+
+**Critical authz inversion:** `user_client_access` today documents "no entries = access to all (for
+admins)" (migration 002, line 25-26). The grant check **must branch on role** — for `end_user`, an
+empty grant set means **zero** machines, never all. Authz is deny-by-default and grant-scoped; the
+admin-bypass in `Claims::has_permission` (`jwt.rs:28-33`) must never fire for `end_user`.
+
+### Agent / Relay-server / Viewer / Dashboard responsibilities
+- **Agent:** no changes. End-users connect to existing **persistent/unattended** managed agents
+  (consent `not_required` — it is the user's own machine). Optionally honors the SPEC-015 notification
+  overlay if a per-machine policy requires it.
+- **Relay-server:** no transport change. New end-user auth + portal + connect endpoints; the
+  grant-check + viewer-token mint is the only new server logic on the hot path.
+- **Viewer:** reuse the React/TS web viewer (`dashboard/src/components/RemoteViewer.tsx`) — the
+  end-user portal embeds the same component with a Control-mode viewer token.
+- **Dashboard:** new **role-gated end-user portal** route (recommended separate from the technician
+  console — see Open Questions), plus admin screens for end-user + grant management.
+
+### Database (migrations)
+- **`user_client_access`** — reused as the grant table; no schema change (already
+  `user_id UUID × client_id UUID → connect_machines(id)`, unique pair, migration 002).
+- New migration `011_end_user_access.sql`:
+  - Widen `connect_sessions.source` CHECK to `('standalone','gururmm','end_user')` (currently
+    `('standalone','gururmm')`, migration 004 line 99-102).
+  - Optional `users` columns for the external principal: `mfa_secret TEXT NULL`,
+    `must_change_password BOOLEAN NOT NULL DEFAULT false`, and a partial index for fast
+    `role='end_user'` listing per `tenant_id`.
+  - (Seat tracking, if landed in v1: a lightweight per-tenant `end_user` count view or a
+    `tenant_seats` row — kept minimal.)
+- Grants are tenant-contained: insert path validates `machine.tenant_id == end_user.tenant_id`.
+
+### API endpoints / WS messages
+- `POST /api/enduser/auth/login` — public, rate-limited; returns an `end_user` login JWT.
+- `GET  /api/enduser/machines` — lists only the caller's granted, in-tenant machines + presence.
+- `POST /api/enduser/machines/:id/connect` — grant-checked; creates a `source=end_user` session and
+  mints a Control `ViewerClaims` token (`create_viewer_token`, `jwt.rs:233`) for that session.
+- Admin: `POST /api/users` (role=end_user), `POST /api/users/:id/grants`,
+  `DELETE /api/users/:id/grants/:machine_id`, `GET /api/users?role=end_user`.
+- No new protobuf messages — the WS viewer path and `guruconnect.proto` are unchanged.
+
+## Implementation details
+- `server/src/auth/jwt.rs` — extend the role vocabulary doc (`Claims.role`, line 16-17); add an
+  `is_end_user()` helper and ensure `has_permission` cannot grant `end_user` anything beyond explicit
+  permissions (the admin short-circuit at line 30 must be guarded).
+- `server/src/auth/mod.rs` — `AuthenticatedUser` (line 29+) gains role-aware helpers; add an extractor
+  / middleware that rejects non-`end_user` on the `/api/enduser/*` namespace and rejects `end_user` on
+  every console/admin route (deny-by-default allowlist).
+- `server/src/api/` — new `enduser` handler module (login, machines, connect); admin user+grant
+  handlers extended for `role=end_user` and `user_client_access` writes.
+- Grant check (shared fn): `machine_id ∈ user_client_access[user] AND machine.tenant_id == user.tenant_id`;
+  used by both `GET /machines` and `connect`.
+- Session create stamps `source='end_user'`, `is_managed=true`/unattended, `consent_state='not_required'`,
+  then mints the viewer token via the existing path so relay enforcement is unchanged.
+- `dashboard/src/` — end-user portal route (role-gated), reusing `RemoteViewer.tsx`; admin grant-matrix
+  UI. White-label (SPEC-014) applies to the portal as the most client-facing surface.
+- Migration `server/migrations/011_end_user_access.sql` as above (idempotent; applied by
+  `sqlx::migrate!` per the migration standard).
+
+## Security considerations
+- **Preserve the plane separation** audited in v2 Phase 1 — `end_user` is login-plane only; it can
+  never satisfy `validate_viewer_token` or the agent `cak_` path.
+- **Deny-by-default, grant-scoped:** empty `user_client_access` for an `end_user` = no access; the
+  admin-bypass must not apply. Every `/api/enduser/*` call re-checks the grant + tenant server-side
+  (never trust a machine id from the client).
+- **Tenant containment:** an `end_user` and its grants live in one tenant; cross-tenant grants are
+  rejected at write and re-validated at connect. (Full tenant isolation lands with Phase 4; v1 enforces
+  via explicit `tenant_id` equality checks.)
+- **External-user trust:** these accounts are public-internet-facing from home. Require
+  rate-limiting + lockout on `/api/enduser/auth/login`; support (recommend require) **TOTP MFA** for
+  `end_user` — schema column included so MFA can be v1 or an immediate fast-follow without a second
+  migration. Argon2id passwords (existing standard).
+- **Audit:** log each end-user login (success/failure, source IP) and each machine access to
+  `connect_session_events`; the unattended access is to the user's *own* machine but must be fully
+  traceable. Optionally enforce the SPEC-015 overlay per machine policy.
+- **Threat model:** stolen end-user creds reach only that user's granted machines (blast radius =
+  grant set), never the console, never the agent plane, never another tenant. Disabling the account
+  (`users.enabled=false`) immediately revokes portal + future tokens; the 5-min viewer-token TTL bounds
+  any in-flight session.
+
+## Testing strategy
+- **Unit:** grant-check fn (granted / not-granted / cross-tenant / empty-set-for-end_user = deny);
+  `has_permission` never elevates `end_user`; role-namespace middleware (end_user→console = 403,
+  technician→/api/enduser = 403).
+- **Integration:** end-user login → list shows only granted machines → connect mints a Control viewer
+  token for a `source=end_user` session → relay admits; connect to a non-granted / other-tenant machine
+  → 403; disabled account → login + token use rejected.
+- **Manual:** full portal walkthrough from an off-network browser; MFA enrol + challenge; audit rows
+  present for login and access; white-label branding renders on the portal.
+
+## Effort estimate & dependencies
+- **Size:** Large (new principal + portal + admin grant UI + auth namespace; transport/agent untouched
+  and the grant table already exists, which holds it below X-Large).
+- **Depends on (must precede / strongly preferred):**
+  - **Tenancy** (`tenants` + `tenant_id`, migration 004) — needed for containment; full isolation is
+    Phase 4 but v1 uses explicit tenant checks.
+  - **Stable machine identity + persistent enrollment** (SPEC-004 / 008 `machine_uid`, SPEC-016
+    zero-touch `cak_`) — end-users reach persistent managed agents.
+  - **Session-scoped viewer tokens** (v2 Phase 1, landed) — reused directly.
+- **Pairs with:** SPEC-014 (white-label — the portal is the client-facing surface), SPEC-003/005
+  (machine inventory/list — portal machine rows), SPEC-015 (optional connect-notification overlay).
+- **Unblocks:** the directory-sync spec (AD/Entra/Google → end-user list), delegated client-admin role,
+  and per-seat billing — all of which build on the `end_user` principal defined here.
+
+## Open questions
+1. **Same console vs separate end-user portal?** Recommendation: **separate, role-gated route** —
+   smaller attack surface, no risk of leaking technician controls, cleaner white-label. Confirm before
+   build.
+2. **End-users in the existing `users` table (role=end_user) vs a dedicated `end_users` table?**
+   Recommendation: reuse `users` (the grant FK `user_client_access.user_id` already points there) with
+   hard role guardrails. Revisit if mixing external + internal principals in one table proves risky.
+3. **MFA in v1 or immediate fast-follow?** Schema is included either way; decide enforcement timing.
+4. **Who administers grants in v1** — MSP admin only (assumed), or ship the delegated client-company
+   admin role together? (Affects scope/effort materially.)
+5. **Seat/licensing enforcement depth for v1** — count-and-display vs hard-cap vs billing-integrated.
+6. **Default access mode** — Control assumed (own machine); should an admin be able to pin a machine to
+   view-only for a given end-user? (Token layer already supports it.)

docs/specs/SPEC-018-managed-agent-service-host.md

@@ -0,0 +1,146 @@
+# SPEC-018: Managed-Agent SYSTEM Service Host + Session Broker
+
+**Status:** Proposed
+**Priority:** P1 (blocks SPEC-016 Phase B end-to-end runtime and SPEC-013)
+**Requested By:** Mike (2026-06-02)
+**Estimated Effort:** X-Large
+
+## Overview
+
+Convert the managed/persistent GuruConnect agent from a user-context `HKCU\…\Run` autostart into a
+**Windows SYSTEM service** that runs unattended — at the login screen, with no user logged in, across
+reboots — and **brokers per-session capture/input worker processes** into the active interactive
+desktop. A SYSTEM service lives in the isolated **Session 0** and cannot capture or inject the
+interactive desktop directly, so the service spawns a worker into the target user session (the
+ScreenConnect architecture).
+
+This is foundational, not cosmetic. It unblocks three things at once:
+1. **SPEC-016 Phase B end-to-end runtime** — the per-machine `cak_` store is ACL'd to SYSTEM +
+   Administrators; today the agent runs as the interactive *user* and can't read its own store (the
+   Phase B C1 *fail-fast guard* exists precisely because of this). Running as SYSTEM makes the store
+   readable and removes the guard.
+2. **True unattended access** — a user-context agent only runs while that user is logged in. Reaching
+   a rebooted server or a machine sitting at the login screen (table-stakes for remote support)
+   requires SYSTEM.
+3. **SPEC-013 session selection / backstage** — the session-broker primitive built here is the
+   substrate SPEC-013's session-switching UX drives.
+
+**Success criteria:** the managed agent installs as an auto-start SYSTEM service; it holds the relay
+connection and performs SPEC-016 enrollment as SYSTEM (reading/writing the SYSTEM-ACL'd `cak_`); it
+spawns a capture/input worker into the active interactive session and relays frames; the worker is
+respawned/retargeted on logon/logoff/console-connect; and the Phase B fail-fast guard is removed
+because the store is now readable in-context.
+
+## Background — why this is needed (confirmed in code)
+
+- The persistent agent autostarts via `HKCU\…\Run` (`agent/src/startup.rs:21`, `STARTUP_KEY` = HKCU)
+  → interactive-user token, not SYSTEM. The only SYSTEM service today is the separate `sas_service`
+  (Secure Attention Sequence helper).
+- SPEC-016 Phase B (`agent/src/credential_store.rs`) ACLs the `cak_` store to `*S-1-5-18` (SYSTEM) +
+  `*S-1-5-32-544` (Administrators). In the current user context the agent writes but cannot read it
+  back → the Phase B fail-fast guard (`agent/src/main.rs` `resolve_agent_credential`) emits
+  "must run as the GuruConnect SYSTEM service (see SPEC-018)" instead of bricking.
+- Capture/input live in the agent process (`agent/src/capture/`, `agent/src/input/`); a Session-0
+  SYSTEM service cannot drive these against the interactive desktop without a per-session worker.
+
+## Scope
+
+### Included in v1
+
+1. **Windows service install/lifecycle** (`agent/src/install.rs` + a new service module): register the
+   managed agent as a **LocalSystem auto-start service** (`CreateServiceW` / a service crate),
+   configure failure/recovery (restart on crash), and **replace the HKCU `Run` autostart for managed
+   mode** (remove the Run entry on service install). Clean uninstall (stop + delete service).
+2. **Service control loop** (Session 0, SYSTEM): owns the persistent WSS connection to the relay,
+   performs SPEC-016 enrollment as SYSTEM (now able to read/write the `cak_` store), and dispatches
+   session/connect requests to workers. Handles `SERVICE_CONTROL_STOP`/`SHUTDOWN` and
+   `SERVICE_CONTROL_SESSIONCHANGE`.
+3. **Session broker:** enumerate sessions (`WTSEnumerateSessionsW`), resolve the active interactive
+   session (`WTSGetActiveConsoleSessionId`), obtain its user token (`WTSQueryUserToken` →
+   `DuplicateTokenEx`), and spawn a **per-session capture/input worker** into that session's desktop
+   (`CreateProcessAsUserW`, `winsta0\default`). The worker does DXGI capture + input injection in the
+   user's session; the service relays frames over the existing transport.
+4. **Service ↔ worker IPC:** a local, ACL'd channel (named pipe `\\.\pipe\guruconnect-<sessionId>`)
+   carrying frames/input/control; pipe ACL restricted to SYSTEM + the target session user.
+5. **Session-change handling:** on logon/logoff/console-connect/disconnect/lock/unlock, (re)spawn or
+   retarget the worker so the active desktop is always the one being served.
+6. **Remove the SPEC-016 Phase B fail-fast guard** once the service runs as SYSTEM (the store is
+   readable in-context); keep the SYSTEM+Administrators ACL.
+
+### Explicitly out of scope (anticipated, separate specs)
+
+- **Session-selection / backstage UX** — the operator-facing picker and Session-0/secure-desktop
+  command surface are **SPEC-013**; this spec only provides the broker primitive it drives.
+- **Login-screen / secure-desktop (winlogon) capture** beyond the broker hook — the hard
+  Secure-Desktop case is coordinated with SPEC-013; v1 here targets the active interactive session.
+- **macOS/Linux service equivalents** — future SPEC-010 (cross-platform agents).
+
+## Architecture
+
+- **Agent splits into two roles:**
+  - **service-host** (LocalSystem, Session 0): service lifecycle, relay transport, SPEC-016
+    enrollment + `cak_` store, session broker, IPC server.
+  - **session-worker** (per interactive session, user token): DXGI/GDI capture, input injection,
+    IPC client. Spawned by the service via `CreateProcessAsUserW`.
+- **Service install** (`install.rs`): `CreateServiceW` with `SERVICE_AUTO_START`, `SERVICE_WIN32_OWN_PROCESS`,
+  recovery actions; uninstall stops + deletes. Replaces managed-mode `HKCU Run`.
+- **Token handoff:** `WTSGetActiveConsoleSessionId` → `WTSQueryUserToken` → `DuplicateTokenEx`
+  (primary token) → `CreateProcessAsUserW` with `lpDesktop = "winsta0\\default"`.
+- **IPC:** named pipe per session, length-prefixed protobuf (reuse `proto/` message types where
+  sensible), pipe security descriptor granting only SYSTEM + the session user.
+- **Session events:** the service registers for `SERVICE_CONTROL_SESSIONCHANGE` and reacts to
+  `WTS_CONSOLE_CONNECT`, `WTS_SESSION_LOGON/LOGOFF`, `WTS_SESSION_LOCK/UNLOCK`.
+
+## Security considerations
+
+- **LocalSystem is maximal privilege** — minimize the service's attack surface; validate every
+  relay-delivered command; never spawn a worker except into a legitimately-enumerated active session.
+- **IPC pipe must be ACL'd** (SYSTEM + the specific session user only) so a non-admin user can't
+  inject capture/input commands by connecting to the pipe.
+- **Token hygiene:** close duplicated tokens promptly; don't leak SYSTEM or user primary tokens.
+- The SPEC-016 `cak_` store (SYSTEM-ACL'd) is now correctly readable; the fail-fast guard is removed
+  but the ACL stays.
+- **Audit:** service start/stop, enrollment-as-SYSTEM, worker spawn, session attach/retarget — written
+  to the existing event pipeline.
+
+## Implementation details
+
+- New service module (e.g. `agent/src/service/{mod.rs, broker.rs, ipc.rs}`); worker entry split out of
+  the current capture path. New `Commands` variants or an internal `--service`/`--session-worker`
+  dispatch in `agent/src/main.rs`.
+- `install.rs`: service create/recovery/delete; drop the managed-mode HKCU `Run` write.
+- `windows` crate features: `Win32_System_Services`, `Win32_System_RemoteDesktop`
+  (`WTS*`), `Win32_Security`, `Win32_System_Threading` (`CreateProcessAsUserW`),
+  `Win32_System_Pipes`.
+- Remove the `resolve_agent_credential` fail-fast guard branch added in SPEC-016 Phase B.
+
+## Testing strategy
+
+- **Service:** install → auto-start on boot → stop → uninstall on a clean VM.
+- **`cak_` end-to-end:** SYSTEM service enrolls (SPEC-016), stores + reads the `cak_`, connects — the
+  integration test SPEC-016 Phase B currently cannot run.
+- **Session broker:** worker spawns into the active session; capture/input work; survives logoff→logon
+  (respawn) and console-connect (retarget); fast-user-switch retarget.
+- **Security:** non-admin cannot connect to the IPC pipe; worker runs with the user's token (not
+  SYSTEM) in the user's desktop.
+
+## Effort estimate & dependencies
+
+- **Size:** X-Large (service host + worker split + token-handoff + IPC + session-change handling +
+  install/uninstall).
+- **Depends on:** SPEC-016 (enrollment + `cak_` store); the existing capture/input cores.
+- **Unblocks:** SPEC-016 Phase B end-to-end runtime (and the parked managed-agent enrollment test on
+  the internal beta machines); **SPEC-013** (session selection builds on this broker).
+
+## Open questions
+
+1. **Service vs. SYSTEM scheduled task** — a true Windows service (recovery, SCM integration) is the
+   standard, robust choice; recommend service. Lock in planning.
+2. **One multi-session worker vs. one worker per session** — per-session worker is simpler to reason
+   about and isolates a crash to one session; confirm.
+3. **IPC transport** — named pipe (recommended) vs. local TCP/loopback; pipe ACLing is the cleaner
+   security story.
+4. **Login-screen / Secure-Desktop capture** — how much (if any) in this spec vs. deferred to SPEC-013
+   (it needs a worker in the winlogon/secure desktop, a distinct hard problem).
+5. **Migration** — on upgrade, cleanly transition existing HKCU-`Run` managed installs to the service
+   (remove the Run entry, install the service) without a gap.

reports/2026-05-31-gc-audit.md

@@ -0,0 +1,129 @@
+# GuruConnect Audit Report — 2026-05-31
+
+**Auditor:** Claude (claude-opus-4-8[1m])
+**Passes:** Security & Remote-Session Integrity (`--pass=security` only)
+**Previous audit:** 2026-05-30 (`reports/2026-05-30-gc-audit.md`)
+**Scope note:** v2 **Phase-1 EXIT gate** re-audit. Confirms the three relay CRITICALs stay closed and
+the prior net-new HIGH is fixed, and assesses the net-new SPEC-004 surface (Tasks 2/4/5 — machine_uid
+dedup, session reaping, operator removal) now committed + deployed. Includes **live** boundary tests
+against the running production binary, not just a code re-derivation.
+
+**Code under audit:** working tree at tag **v0.3.0 / e967cce** = the binary deployed to prod
+172.16.3.30:3002 (deployed this session from 96f9c0a; e967cce adds only the version bump + changelog).
+
+---
+
+## Executive Summary
+
+| Pass | Total | Critical | High | Medium | Low | Info |
+|------|-------|----------|------|--------|-----|------|
+| Security & Session | 4 | 0 | 0 | 0 | 0 | 4 |
+
+**Phase-1 security EXIT gate: PASS.** The relay/server plane is clean. All three 2026-05-29 CRITICALs
+remain CLOSED (verified in code AND live against the deployed server). The prior net-new HIGH (agent
+auto-update TLS bypass) and the prior LOW (chat content logged at INFO) are both remediated. The
+net-new SPEC-004 surface (operator removal, machine_uid dedup gate, session reaper/supersede) audits
+clean with the keyed-identity security invariant intact end-to-end. No net-new findings.
+
+**Requires action:** none.
+
+---
+
+## Live functional verification (deployed binary, 172.16.3.30:3002)
+
+Forged tokens (HS256, real `JWT_SECRET`) exercised the WS auth boundaries directly. Each illegitimate
+access was REJECTED (4xx, never a 101 upgrade):
+
+| Check | Result | Proves |
+|-------|--------|--------|
+| Login-shape JWT on `/ws/viewer` | **401** | Login token not accepted as a viewer token (`purpose=="viewer"` enforced) — CRITICAL #1 |
+| Validly-signed viewer token for session AAAA used on session BBBB | **403** | Session binding enforced — a correctly-signed token is refused for the wrong session — CRITICAL #1 |
+| Login JWT used as agent `api_key` on `/ws/agent` | **401** | Agent plane rejects JWTs (no JWT branch) — CRITICAL #3 |
+| Wrong-signature token on `/ws/viewer` | **401** | Signature validation holds (control) |
+
+The session-bind case is the decisive one: a token that WOULD be accepted for its own session is
+rejected 403 for a different session, proving the binding rather than mere signature validation.
+
+---
+
+## The three relay CRITICALs — verdict
+
+| CRITICAL | Verdict | Enforced at |
+|----------|---------|-------------|
+| #1 any-JWT-joins-any-session | **CLOSED** | mint authz `api/sessions.rs` (is_admin \|\| permission); viewer WS `relay/mod.rs:496` `validate_viewer_token` (sig+expiry+`purpose=="viewer"`); session-bind `relay/mod.rs:527-534` (`claim != requested → 403`) |
+| #2 viewer-WS blacklist | **CLOSED** (TTL-bounded residual unchanged) | `relay/mod.rs:509` `token_blacklist.is_revoked` before upgrade. Residual: logout revokes login JWT not minted viewer tokens (5-min TTL) — same tracked MEDIUM, no regression |
+| #3 JWT-accepted-as-agent-key | **CLOSED**, fails closed | `relay/mod.rs:417` `validate_agent_api_key` — no JWT branch; only `cak_` (`auth/agent_keys.rs`, SHA-256 vs `connect_agent_keys`, `revoked_at IS NULL`) or deprecated shared key (WARN). Unresolved machine → 503 (`:303`); client `agent_id` overridden by key identity (`:283`) |
+
+Live results match these code paths exactly.
+
+---
+
+## Prior HIGH — FIXED
+
+**Agent auto-update TLS bypass → MITM-RCE: CLOSED.** `agent/src/update.rs:21` `dev_insecure_tls()` is
+`cfg!(debug_assertions)` AND env-var gated, so a release build's `cfg!` compiles out and the agent
+ALWAYS verifies certs. Both `check_for_update` (`:64`) and `download_update` (`:130`) consume it; unit
+test `test_dev_insecure_tls_release_is_always_false` (`:362`) asserts the release invariant. No
+`danger_accept_invalid_certs(true)` reachable in production. A signed-manifest defense-in-depth TODO is
+filed at `install_update` (`:189`) (= tracked task #10, not an exit blocker).
+
+---
+
+## Pass 5: Security & Remote-Session Integrity — net-new SPEC-004 surface
+
+### [INFO] Operator removal API (`server/src/api/removal.rs`) — clean, admin-gated
+Every removal handler takes the `AdminUser` extractor as its first argument (runs before any DB
+mutation): `remove_machine` (`:88`), `remove_session` (`:321`), `bulk_remove_machines` (`:471`).
+`AdminUser` (`auth/mod.rs:141`) validates JWT (signature + expiry + blacklist `:97`) then requires
+`is_admin()` else 403 (`:146`). Soft-deletes are parameterized + idempotent (`WHERE … AND deleted_at IS
+NULL`); bulk bounded (MAX_BATCH 500) with per-id UUID validation + isolated failures; audit
+(`db/events.rs:126`) records actor + target + trusted-proxy IP, best-effort (cannot be suppressed by
+attacker-controlled input). Removal is admin-role-gated globally (not per-tenant ACL) — same Phase-1
+posture as viewer-mint, per-tenant narrowing deferred to SPEC-002 Phase 4. Acceptable by context.
+
+### [INFO] machine_uid dedup security gate — invariant holds
+Gate at `relay/mod.rs:352`: `effective_machine_uid = if is_keyed_agent { None } else { claimed }`. The
+suppressed value (not the raw claim) flows to `register_agent` and `upsert_machine`. Keyed (`cak_`)
+agents take the agent_id-keyed upsert branch and never write/touch a `ON CONFLICT (machine_uid)` row, so
+a valid key for machine X cannot repoint machine Y via a claimed uid. An un-keyed uid-spoof can only
+match a uid-bearing row — which the keyed connect path never creates; the only residual is a legacy
+pre-keying row, and the startup L1 fix (`main.rs:267-288` via `keyed_machine_ids`, fail-closed on query
+error) ensures keyed machines are never uid-indexed on restore.
+
+### [INFO] Session reaper + same-machine supersede — clean, TOCTOU closed
+`reap_stale_persistent` (`:875`) and supersede (`:322`) select under a read lock then re-assert the full
+predicate under the write lock via `remove_session_if` (`:755`). Predicate requires
+`!is_online && is_persistent && viewers.is_empty()` (+ TTL / same-uid) — an online, viewer-attached, or
+support session is never reaped/superseded. Un-keyed uid-spoof blast radius = denial-of-persistence on
+an offline same-uid session at worst, never a hijack. Lock order matches `register_agent`; predicate is
+synchronous (no await under lock).
+
+### [INFO] General posture — confirmed, no regressions
+Runtime sqlx parameterized everywhere (no `format!`-built SQL); migrations 008/009 idempotent. Frame
+caps: agent 4 MiB / viewer 64 KiB applied before upgrade. Input throttle retained. `/api/auth/login`
+rate-limited (`main.rs:397`). `JWT_SECRET` panics if <32 (`main.rs:143`); agent keys SHA-256; Argon2id
+passwords; no secret/token/code/PII logged. **Chat content no longer logged** (prior LOW fixed —
+`relay/mod.rs:829,1428` now log length only).
+
+---
+
+## Definitive answers
+
+- **(a) Any non-admin removal path?** NO — all three removal handlers gate on `AdminUser` (JWT+blacklist+`is_admin`→403) before any DB mutation.
+- **(b) Any uid-spoof that repoints/hijacks another machine's row or session (not just denial)?** NO — keyed identity is authoritative and uid-suppressed across connect → upsert → reattach → startup restore. Worst case for an un-keyed spoof is denial-of-persistence on an offline same-uid session.
+- **(c) Any auth-plane bypass (agent↔viewer credential crossover)?** NO — viewer plane requires a `purpose=="viewer"` session-bound minted token; agent plane requires a `cak_`/shared key with no JWT branch. Confirmed in code and live.
+
+---
+
+## Verdict
+
+**Phase-1 security EXIT gate: PASS.** Relay/server plane clean; prior HIGH + LOW remediated; SPEC-004
+surface sound with the keyed-identity invariant intact across the connect path, DB upsert, in-memory
+reattach, and startup restore. No new CRITICAL/HIGH/MEDIUM/LOW.
+
+**Tracked, deferred-by-design (not exit blockers):**
+- Viewer-token logout revocation residual (MEDIUM, TTL-bounded) — `v2-secure-session-core/plan.md`.
+- Update-binary signature verification (defense-in-depth, task #10) — TODO at `update.rs:189`.
+
+*Note: only `--pass=security` was run. API-surface, Rust-quality, TypeScript, protocol-integrity,
+docs-reconciliation, and CI/CD passes were not executed this run.*

server/migrations/010_spec016_enrollment.sql

@@ -0,0 +1,159 @@
+-- Migration: 010_spec016_enrollment.sql
+-- Purpose: SPEC-016 zero-touch per-site agent enrollment — server-side data model.
+--
+--   Adds the per-site enrollment-key table, a minimal sites table to anchor it,
+--   and the machine-side columns the collision-gated self-registration flow needs.
+--
+--   Two-tier credential model (SPEC-016 §Security): a low-sensitivity, rotatable,
+--   per-site ENROLLMENT KEY (the `cek_` secret stored hashed here) gates "may this
+--   machine register at all", while the high-sensitivity per-machine `cak_`
+--   operating credential (connect_agent_keys, migration 004) is minted on a
+--   successful enroll. Compromise of an enrollment key is recovered by rotating one
+--   site, not a fleet-wide re-key.
+--
+-- DEVIATION FROM SPEC (documented): SPEC-016 §DB-migration describes
+-- `site_enrollment_keys.site_id` as `fk -> sites`, assuming a sites table already
+-- exists. It does NOT — in the current schema "site" and "company/organization" are
+-- free-text columns on connect_machines (migration 005), there is no relational
+-- sites entity. This migration therefore CREATES a minimal `connect_sites` table
+-- (the relational anchor the enrollment-key FK and the dashboard per-site key
+-- display both require) keyed by a natural `site_code` and scoped per-tenant. It is
+-- intentionally minimal (code + display name + tenant); richer site/company
+-- modeling is left to future work. The free-text connect_machines.site /
+-- .organization columns are untouched and continue to carry agent-reported labels.
+--
+-- Idempotent: CREATE TABLE/INDEX IF NOT EXISTS, ADD COLUMN IF NOT EXISTS. Applied on
+-- server startup by sqlx::migrate!(); never pre-applied via psql. Ordered after 009.
+-- See .claude/standards/gururmm/sqlx-migrations.md.
+
+-- pgcrypto provides gen_random_uuid(); enabled in 001/004 but re-asserted for safety.
+CREATE EXTENSION IF NOT EXISTS "pgcrypto";
+
+-- ============================================================================
+-- connect_sites — relational anchor for per-site enrollment (see DEVIATION above)
+-- ============================================================================
+-- A site is the unit a single signed installer targets. `site_code` is the
+-- non-secret, operator-facing identifier the installer carries and the agent sends
+-- at /api/enroll (e.g. "ACME-PHX"). Uniqueness is per-tenant: the same human-chosen
+-- code may legitimately exist in two tenants. tenant_id mirrors the nullable,
+-- default-tenant-backfilled tenancy column used on every other scoped table
+-- (migration 004); db::tenancy::current_tenant_id() resolves it for now.
+
+CREATE TABLE IF NOT EXISTS connect_sites (
+    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    -- Operator-facing site identifier the installer carries. Non-secret.
+    site_code TEXT NOT NULL,
+    -- Human-readable site / company display name for the dashboard.
+    display_name TEXT,
+    -- Default company label applied to machines enrolled at this site (mirrors the
+    -- free-text connect_machines.organization the agent otherwise self-reports).
+    company TEXT,
+    -- Tenancy-ready (Phase 4). Backfilled to the default tenant below.
+    tenant_id UUID,
+    -- RESERVED for future per-site enrollment POLICY work (SPEC-016 §out-of-scope):
+    -- default 'auto-approve'; a future 'pending-approval' value will gate new
+    -- enrollments. NOT enforced in Phase A — present so the policy SPEC needs no
+    -- schema change. Do not branch on this column yet.
+    enrollment_policy TEXT DEFAULT 'auto-approve',
+    created_at TIMESTAMPTZ NOT NULL DEFAULT NOW()
+);
+
+-- Per-tenant uniqueness of the natural site_code so /api/enroll can resolve a site
+-- deterministically within a tenant while the same code may exist across tenants.
+-- COALESCE keeps the index usable while tenant_id is still nullable (Phase 1).
+CREATE UNIQUE INDEX IF NOT EXISTS idx_connect_sites_tenant_code
+    ON connect_sites (COALESCE(tenant_id, '00000000-0000-0000-0000-000000000001'::uuid), site_code);
+
+-- Backfill the sites tenant_id to the default tenant (table is empty on a fresh DB;
+-- no-op there, but keeps the migration self-consistent).
+UPDATE connect_sites
+SET tenant_id = '00000000-0000-0000-0000-000000000001'
+WHERE tenant_id IS NULL;
+
+-- ============================================================================
+-- site_enrollment_keys — rotatable, hashed per-site enrollment secret + fingerprint
+-- ============================================================================
+-- Stores ONLY the Argon2id hash of the `cek_` secret; the plaintext is shown once
+-- at issue/rotate and never recoverable. `version` is the monotonic rotation
+-- counter; `fingerprint` is the non-secret short hex shown as `vN (XXXX)` in the
+-- dashboard and baked into the installer filename. `active` marks the current key —
+-- rotation flips the old key to active=false (blocking NEW enrollments from old
+-- installers) and inserts a new active row; already-enrolled agents holding their
+-- own `cak_` are unaffected. Multiple inactive (historical) rows may coexist per
+-- site; at most one active row is intended (enforced by a partial unique index).
+
+CREATE TABLE IF NOT EXISTS site_enrollment_keys (
+    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    site_id UUID NOT NULL REFERENCES connect_sites(id) ON DELETE CASCADE,
+    -- Argon2id hash of the `cek_` enrollment secret. Never the plaintext.
+    key_hash TEXT NOT NULL,
+    -- Monotonic rotation version (1, 2, 3, ...).
+    version INTEGER NOT NULL,
+    -- Non-secret short hex fingerprint code (the XXXX in `vN (XXXX)`), derived from
+    -- the secret. Stored so the dashboard / GET endpoint can show it without the
+    -- secret.
+    fingerprint TEXT NOT NULL,
+    active BOOLEAN NOT NULL DEFAULT true,
+    created_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
+    -- Set when this key is rotated out (active flipped to false).
+    rotated_at TIMESTAMPTZ
+);
+
+-- Lookup index for the enroll hot path: resolve the active key for a site.
+CREATE INDEX IF NOT EXISTS idx_site_enrollment_keys_site_active
+    ON site_enrollment_keys (site_id, active);
+
+-- At most one ACTIVE enrollment key per site (the "current" installer key).
+-- Partial unique index so any number of inactive historical rows may coexist.
+CREATE UNIQUE INDEX IF NOT EXISTS idx_site_enrollment_keys_one_active
+    ON site_enrollment_keys (site_id)
+    WHERE active;
+
+-- ============================================================================
+-- connect_machines — site binding + enrollment-state collision gate
+-- ============================================================================
+-- machine_uid already exists (migration 008) with a partial UNIQUE index on
+-- (machine_uid) WHERE machine_uid IS NOT NULL. SPEC-016 §item-1 / resolved-decision #4
+-- call for the dedup key to be PER-TENANT — (tenant_id, machine_uid) — so the same
+-- hardware legitimately present in two tenants stays two rows. tenant_id is the
+-- scoping column that exists on connect_machines (migration 004); machines have no
+-- direct site_id today, so site is tracked separately (site_id below) and tenancy is
+-- the uniqueness scope, exactly as the spec states.
+--
+-- CRITICAL CONSTRAINT (why we ADD rather than REPLACE the 008 index here):
+-- db::machines::upsert_machine (the live connect-path upsert) uses
+-- `ON CONFLICT (machine_uid) WHERE machine_uid IS NOT NULL` as its conflict arbiter.
+-- Postgres matches that arbiter to the EXACT index from migration 008. Dropping that
+-- index would make the live upsert fail to find an arbiter and error at runtime —
+-- breaking every un-keyed agent reconnect. So migration 008's global index is LEFT
+-- IN PLACE (the connect path keeps working unchanged) and the per-tenant index is
+-- added ALONGSIDE it. In single-tenant Phase 1 the two are equivalent (every row's
+-- tenant_id is the default tenant), so the per-tenant index adds the SPEC-016 dedup
+-- semantics without a redundant-uniqueness conflict: a (tenant, uid) pair that is
+-- unique is also globally unique today. When multi-tenancy activates AND
+-- upsert_machine's ON CONFLICT is updated to name (tenant_id, machine_uid), a future
+-- migration drops the global 008 index. Documented as deferred; do not drop it now.
+
+-- Optional FK to the site a machine enrolled under (NULL for legacy / support-code
+-- machines that never enrolled through /api/enroll). A site change on re-enroll is
+-- the "site move" SPEC-016 audits.
+ALTER TABLE connect_machines ADD COLUMN IF NOT EXISTS site_id UUID REFERENCES connect_sites(id) ON DELETE SET NULL;
+
+-- enrollment_state: the collision gate (SPEC-016 §item-1/6). 'active' = live and
+-- controllable (auto-approve posture); 'pending' = a machine_uid collision was
+-- detected at enroll and an operator must confirm in the dashboard before the
+-- endpoint may be controlled. Default 'active' so every legacy/connect-path row is
+-- unaffected.
+ALTER TABLE connect_machines
+    ADD COLUMN IF NOT EXISTS enrollment_state TEXT NOT NULL DEFAULT 'active'
+    CHECK (enrollment_state IN ('active', 'pending'));
+
+-- Per-tenant machine_uid uniqueness (SPEC-016). Added ALONGSIDE migration 008's
+-- global (machine_uid) index (see CRITICAL CONSTRAINT above — the connect-path
+-- upsert's ON CONFLICT arbiter binds to the 008 index, which must survive). COALESCE
+-- folds a NULL tenant_id to the default tenant so the index is well-defined while
+-- tenancy is single-tenant (Phase 1); the WHERE clause excludes NULL machine_uid so
+-- legacy un-keyed rows coexist freely.
+CREATE UNIQUE INDEX IF NOT EXISTS idx_connect_machines_tenant_machine_uid
+    ON connect_machines (COALESCE(tenant_id, '00000000-0000-0000-0000-000000000001'::uuid), machine_uid)
+    WHERE machine_uid IS NOT NULL;

server/src/api/mod.rs

@@ -4,10 +4,12 @@ pub mod auth;
 pub mod auth_logout;
 pub mod changelog;
 pub mod downloads;
+pub mod enroll;
 pub mod machine_keys;
 pub mod releases;
 pub mod removal;
 pub mod sessions;
+pub mod sites;
 pub mod users;
 
 use axum::{

server/src/api/sites.rs

@@ -0,0 +1,217 @@
+//! Site enrollment-key administration (SPEC-016, admin plane).
+//!
+//! Admin (dashboard JWT + admin role) endpoints for the per-site enrollment key
+//! the dashboard surfaces and rotates:
+//!
+//! - `POST /api/sites/:id/enrollment-key/rotate` — regenerate the `cek_` secret,
+//!   bump the monotonic version, derive a new fingerprint, deactivate the prior
+//!   active key, and return the plaintext + fingerprint ONCE. Old installers can no
+//!   longer enroll NEW machines after this; already-enrolled agents (holding their
+//!   own `cak_`) are unaffected (SPEC-016 success-criterion #3). Doubles as
+//!   first-issue when a site has no key yet.
+//! - `GET /api/sites/:id/enrollment-key` — read the CURRENT non-secret fingerprint
+//!   + version (never the secret). 404 if the site has no active key yet.
+//!
+//! Auth mirrors `api::machine_keys`: the [`crate::auth::AdminUser`] extractor gates
+//! both routes, and they are mounted behind the JWT `auth_layer`.
+//!
+//! SECURITY: the plaintext `cek_` is returned exactly once (rotate response),
+//! never persisted in plaintext and never logged. Read responses expose only the
+//! version + fingerprint.
+
+use axum::{
+    extract::{Path, State},
+    http::StatusCode,
+    Json,
+};
+use serde::Serialize;
+use uuid::Uuid;
+
+use crate::auth::{enrollment_keys, AdminUser};
+use crate::db;
+use crate::AppState;
+
+/// Standard error envelope (matches `api::machine_keys::ApiError`).
+#[derive(Debug, Serialize)]
+pub struct ApiError {
+    pub detail: String,
+    pub error_code: String,
+    pub status_code: u16,
+}
+
+impl ApiError {
+    fn new(status: StatusCode, code: &str, detail: &str) -> (StatusCode, Json<ApiError>) {
+        (
+            status,
+            Json(ApiError {
+                detail: detail.to_string(),
+                error_code: code.to_string(),
+                status_code: status.as_u16(),
+            }),
+        )
+    }
+}
+
+type ApiResult<T> = Result<T, (StatusCode, Json<ApiError>)>;
+
+/// Response for a freshly rotated/issued enrollment key. `key` is present ONLY
+/// here, once.
+#[derive(Debug, Serialize)]
+pub struct RotatedEnrollmentKey {
+    pub site_id: Uuid,
+    /// The plaintext `cek_` enrollment key. Shown exactly once — bake it into the
+    /// site installer now; the server keeps only its hash.
+    pub key: String,
+    /// Monotonic rotation version.
+    pub version: i32,
+    /// The non-secret short hex code (the `XXXX` in `vN (XXXX)`).
+    pub fingerprint: String,
+    /// Fully rendered operator-facing fingerprint, e.g. `v3 (7F2A)`.
+    pub fingerprint_label: String,
+}
+
+/// Non-secret current-key view for the GET endpoint.
+#[derive(Debug, Serialize)]
+pub struct EnrollmentKeyView {
+    pub site_id: Uuid,
+    pub version: i32,
+    pub fingerprint: String,
+    pub fingerprint_label: String,
+    pub active: bool,
+}
+
+fn require_db(state: &AppState) -> ApiResult<&db::Database> {
+    state.db.as_ref().ok_or_else(|| {
+        ApiError::new(
+            StatusCode::SERVICE_UNAVAILABLE,
+            "DATABASE_UNAVAILABLE",
+            "Database not available",
+        )
+    })
+}
+
+/// Resolve a site by its UUID path segment, or a 404 envelope.
+async fn resolve_site(db: &db::Database, site_id: Uuid) -> ApiResult<db::sites::Site> {
+    db::sites::get_site_by_id(db.pool(), site_id)
+        .await
+        .map_err(|e| {
+            tracing::error!("DB error resolving site: {}", e);
+            ApiError::new(
+                StatusCode::INTERNAL_SERVER_ERROR,
+                "INTERNAL_ERROR",
+                "Internal server error",
+            )
+        })?
+        .ok_or_else(|| ApiError::new(StatusCode::NOT_FOUND, "SITE_NOT_FOUND", "Site not found"))
+}
+
+/// POST /api/sites/:id/enrollment-key/rotate — rotate (or first-issue) a site's
+/// enrollment key. Returns the plaintext `cek_` + fingerprint once.
+pub async fn rotate_enrollment_key(
+    AdminUser(admin): AdminUser,
+    State(state): State<AppState>,
+    Path(site_id): Path<Uuid>,
+) -> ApiResult<(StatusCode, Json<RotatedEnrollmentKey>)> {
+    let db = require_db(&state)?;
+    let site = resolve_site(db, site_id).await?;
+
+    // Mint plaintext + Argon2id hash + fingerprint. Only the hash + fingerprint
+    // are persisted; the plaintext is surfaced once below.
+    let plaintext = enrollment_keys::generate_enrollment_key();
+    let key_hash = enrollment_keys::hash_enrollment_key(&plaintext).map_err(|e| {
+        tracing::error!("Failed to hash enrollment key: {}", e);
+        ApiError::new(
+            StatusCode::INTERNAL_SERVER_ERROR,
+            "INTERNAL_ERROR",
+            "Failed to hash enrollment key",
+        )
+    })?;
+    let fingerprint = enrollment_keys::compute_fingerprint(&plaintext);
+
+    let new_key = db::enrollment_keys::rotate_key(db.pool(), site.id, &key_hash, &fingerprint)
+        .await
+        .map_err(|e| {
+            tracing::error!("DB error rotating enrollment key: {}", e);
+            ApiError::new(
+                StatusCode::INTERNAL_SERVER_ERROR,
+                "INTERNAL_ERROR",
+                "Failed to rotate enrollment key",
+            )
+        })?;
+
+    let fingerprint_label =
+        enrollment_keys::render_fingerprint(new_key.version, &new_key.fingerprint);
+
+    // Audit WITHOUT key material (no plaintext, no hash).
+    if let Err(e) = db::events::log_enrollment_event(
+        db.pool(),
+        db::events::EventTypes::ENROLLMENT_KEY_ROTATED,
+        serde_json::json!({
+            "site_id": site.id,
+            "site_code": site.site_code,
+            "version": new_key.version,
+            "fingerprint": new_key.fingerprint,
+            "rotated_by": admin.username,
+        }),
+        None,
+    )
+    .await
+    {
+        tracing::warn!("[ENROLL] failed to write key-rotate audit event: {}", e);
+    }
+    tracing::info!(
+        "Admin {} rotated enrollment key for site {} to {}",
+        admin.username,
+        site.site_code,
+        fingerprint_label
+    );
+
+    Ok((
+        StatusCode::CREATED,
+        Json(RotatedEnrollmentKey {
+            site_id: site.id,
+            key: plaintext,
+            version: new_key.version,
+            fingerprint: new_key.fingerprint,
+            fingerprint_label,
+        }),
+    ))
+}
+
+/// GET /api/sites/:id/enrollment-key — current non-secret fingerprint + version.
+pub async fn get_enrollment_key(
+    AdminUser(_admin): AdminUser,
+    State(state): State<AppState>,
+    Path(site_id): Path<Uuid>,
+) -> ApiResult<Json<EnrollmentKeyView>> {
+    let db = require_db(&state)?;
+    let site = resolve_site(db, site_id).await?;
+
+    let key = db::enrollment_keys::get_active_for_site(db.pool(), site.id)
+        .await
+        .map_err(|e| {
+            tracing::error!("DB error loading enrollment key: {}", e);
+            ApiError::new(
+                StatusCode::INTERNAL_SERVER_ERROR,
+                "INTERNAL_ERROR",
+                "Internal server error",
+            )
+        })?
+        .ok_or_else(|| {
+            ApiError::new(
+                StatusCode::NOT_FOUND,
+                "NO_ENROLLMENT_KEY",
+                "Site has no active enrollment key",
+            )
+        })?;
+
+    let fingerprint_label = enrollment_keys::render_fingerprint(key.version, &key.fingerprint);
+
+    Ok(Json(EnrollmentKeyView {
+        site_id: site.id,
+        version: key.version,
+        fingerprint: key.fingerprint,
+        fingerprint_label,
+        active: key.active,
+    }))
+}

server/src/auth/enrollment_keys.rs

@@ -0,0 +1,191 @@
+//! Per-site enrollment key minting, hashing, verification, and fingerprinting
+//! (SPEC-016 zero-touch enrollment, auth layer).
+//!
+//! This is the low-sensitivity, rotatable side of the two-tier credential model
+//! (SPEC-016 §Security). A per-site ENROLLMENT key (`cek_` prefix) gates "may
+//! this machine register at all" at `POST /api/enroll`; a successful enroll mints
+//! the high-sensitivity per-machine `cak_` operating credential
+//! ([`crate::auth::agent_keys`]). Compromise of an enrollment key is contained to
+//! one site and recovered by rotating it.
+//!
+//! Lifecycle owned here (the secret side):
+//!
+//! - [`generate_enrollment_key`] mints a high-entropy, `cek_`-prefixed plaintext
+//!   secret. Mirrors [`crate::auth::agent_keys::generate_agent_key`]'s entropy
+//!   approach (32 random bytes from the OS CSPRNG, hex-encoded) with a DISTINCT
+//!   prefix so the two key kinds are never confused in logs or storage. The
+//!   plaintext is shown to the operator exactly once at issue/rotate and is NEVER
+//!   persisted or logged.
+//! - [`hash_enrollment_key`] / [`verify_enrollment_key`] use **Argon2id** (via
+//!   [`crate::auth::password`]). This DIFFERS from `cak_` (which uses SHA-256 for
+//!   a constant-shape equality lookup): SPEC-016 §2 explicitly requires the
+//!   enrollment key be "stored hashed (Argon2id, same as `cak_`/passwords)". The
+//!   trade-off is deliberate — enrollment keys are looked up by `(site, active)`
+//!   first (a small candidate set, usually one row) and only then verified, so the
+//!   per-verify KDF cost is bounded and not on a high-QPS path, while Argon2id
+//!   gives salted, GPU-resistant storage matching the password posture.
+//! - [`compute_fingerprint`] derives the non-secret short HEX code shown as
+//!   `vN (XXXX)` (SPEC-016 resolved-decision #3 — hex, deliberately NOT the
+//!   GuruRMM word-style code, so the two products' artifacts are never visually
+//!   conflated).
+//!
+//! SECURITY: never log a plaintext key or its hash. Functions here return the
+//! plaintext to the caller (issue/rotate endpoint) but emit no `tracing` output
+//! containing key material.
+
+use anyhow::Result;
+use rand::RngCore;
+use ring::digest;
+
+/// Prefix marking a GuruConnect per-site enrollment key. Distinct from the
+/// per-agent `cak_` prefix so the two key kinds are never confused.
+pub const ENROLLMENT_KEY_PREFIX: &str = "cek_";
+
+/// Number of random bytes behind an enrollment key (256 bits of entropy), matching
+/// [`crate::auth::agent_keys`]. SPEC-016 §2 requires ≥256-bit.
+const ENROLLMENT_KEY_RANDOM_BYTES: usize = 32;
+
+/// Number of hex characters in the fingerprint code (the `XXXX` in `vN (XXXX)`).
+/// Four hex chars = 16 bits — ample to let an operator tell two installers apart at
+/// a glance; it is a non-secret display aid, not a security control.
+const FINGERPRINT_HEX_LEN: usize = 4;
+
+/// Generate a new high-entropy, `cek_`-prefixed per-site enrollment key (plaintext).
+///
+/// The returned string is the ONLY time the plaintext exists; the caller must
+/// surface it to the operator once and store only [`hash_enrollment_key`] of it.
+/// Uses the OS CSPRNG via `rand::rngs::OsRng`.
+pub fn generate_enrollment_key() -> String {
+    let mut bytes = [0u8; ENROLLMENT_KEY_RANDOM_BYTES];
+    rand::rngs::OsRng.fill_bytes(&mut bytes);
+    format!("{}{}", ENROLLMENT_KEY_PREFIX, hex_encode(&bytes))
+}
+
+/// Hash an enrollment key for storage using Argon2id (SPEC-016 §2).
+///
+/// Delegates to [`crate::auth::password::hash_password`] so the KDF parameters and
+/// salt generation match the password posture exactly. Returns the PHC-format
+/// string Postgres stores in `site_enrollment_keys.key_hash`.
+pub fn hash_enrollment_key(plaintext: &str) -> Result<String> {
+    crate::auth::password::hash_password(plaintext)
+}
+
+/// Verify a presented enrollment key against a stored Argon2id hash.
+///
+/// Returns `Ok(true)` on a match. A malformed stored hash or a mismatch yields
+/// `Ok(false)` / an `Err` from the underlying verifier; the caller treats any
+/// non-`Ok(true)` as a rejection. A cheap structural reject (`cek_` prefix) runs
+/// first to skip the KDF on obviously-bogus input.
+///
+/// SECURITY: only compares; never logs the presented key or the hash.
+pub fn verify_enrollment_key(presented: &str, stored_hash: &str) -> bool {
+    if !presented.starts_with(ENROLLMENT_KEY_PREFIX) {
+        return false;
+    }
+    crate::auth::password::verify_password(presented, stored_hash).unwrap_or(false)
+}
+
+/// Compute the non-secret short HEX fingerprint code for an enrollment key.
+///
+/// Derived as the first [`FINGERPRINT_HEX_LEN`] hex chars of the SHA-256 of the
+/// plaintext secret, uppercased. This is a stable, non-reversible tag of the secret
+/// (knowing the code does not reveal the key) used purely for display. Pair it with
+/// the monotonic version via [`render_fingerprint`].
+pub fn compute_fingerprint(plaintext: &str) -> String {
+    let d = digest::digest(&digest::SHA256, plaintext.as_bytes());
+    let hex = hex_encode(d.as_ref());
+    hex[..FINGERPRINT_HEX_LEN].to_ascii_uppercase()
+}
+
+/// Render the operator-facing fingerprint string `vN (XXXX)` (SPEC-016 §2).
+///
+/// `version` is the monotonic rotation counter; `code` is [`compute_fingerprint`].
+/// Example: `render_fingerprint(3, "7F2A")` -> `"v3 (7F2A)"`.
+pub fn render_fingerprint(version: i32, code: &str) -> String {
+    format!("v{} ({})", version, code)
+}
+
+/// Lowercase hex encoding without pulling in the `hex` crate (mirrors
+/// [`crate::auth::agent_keys`]).
+fn hex_encode(bytes: &[u8]) -> String {
+    use std::fmt::Write;
+    let mut s = String::with_capacity(bytes.len() * 2);
+    for b in bytes {
+        let _ = write!(s, "{:02x}", b);
+    }
+    s
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn generated_key_is_prefixed_and_high_entropy() {
+        let key = generate_enrollment_key();
+        assert!(key.starts_with(ENROLLMENT_KEY_PREFIX));
+        assert_eq!(
+            key.len(),
+            ENROLLMENT_KEY_PREFIX.len() + ENROLLMENT_KEY_RANDOM_BYTES * 2
+        );
+    }
+
+    #[test]
+    fn generated_keys_are_unique() {
+        assert_ne!(generate_enrollment_key(), generate_enrollment_key());
+    }
+
+    #[test]
+    fn hash_and_verify_roundtrip() {
+        let key = generate_enrollment_key();
+        let hash = hash_enrollment_key(&key).expect("hash");
+        assert!(verify_enrollment_key(&key, &hash));
+    }
+
+    #[test]
+    fn verify_rejects_wrong_key() {
+        let key = generate_enrollment_key();
+        let other = generate_enrollment_key();
+        let hash = hash_enrollment_key(&key).expect("hash");
+        assert!(!verify_enrollment_key(&other, &hash));
+    }
+
+    #[test]
+    fn verify_rejects_unprefixed_input_without_touching_kdf() {
+        let key = generate_enrollment_key();
+        let hash = hash_enrollment_key(&key).expect("hash");
+        // A value lacking the cek_ prefix is structurally rejected before the KDF.
+        assert!(!verify_enrollment_key("not-a-key", &hash));
+    }
+
+    #[test]
+    fn verify_rejects_malformed_stored_hash() {
+        let key = generate_enrollment_key();
+        // A garbage stored hash must not panic and must reject.
+        assert!(!verify_enrollment_key(&key, "not-a-phc-hash"));
+    }
+
+    #[test]
+    fn fingerprint_is_stable_uppercase_hex_of_expected_len() {
+        let key = "cek_deadbeef";
+        let f1 = compute_fingerprint(key);
+        let f2 = compute_fingerprint(key);
+        assert_eq!(f1, f2);
+        assert_eq!(f1.len(), FINGERPRINT_HEX_LEN);
+        assert!(f1.chars().all(|c| c.is_ascii_hexdigit()));
+        assert_eq!(f1, f1.to_ascii_uppercase());
+    }
+
+    #[test]
+    fn fingerprint_differs_per_key() {
+        assert_ne!(
+            compute_fingerprint("cek_aaa"),
+            compute_fingerprint("cek_bbb")
+        );
+    }
+
+    #[test]
+    fn render_fingerprint_matches_spec_shape() {
+        assert_eq!(render_fingerprint(3, "7F2A"), "v3 (7F2A)");
+    }
+}

server/src/auth/mod.rs

@@ -4,6 +4,7 @@
 //! validation for agents.
 
 pub mod agent_keys;
+pub mod enrollment_keys;
 pub mod jwt;
 pub mod password;
 pub mod token_blacklist;

server/src/db/enrollment_keys.rs

@@ -0,0 +1,141 @@
+//! Per-site enrollment key database operations (SPEC-016 zero-touch enrollment).
+//!
+//! Backs the `site_enrollment_keys` table (migration 010). Stores ONLY the
+//! Argon2id hash of the `cek_` secret plus the non-secret rotation metadata
+//! (version, fingerprint, active flag). Computing the hash and minting the
+//! plaintext is [`crate::auth::enrollment_keys`]'s job; this module is
+//! hash-agnostic persistence and takes already-hashed values.
+//!
+//! Rotation invariant: at most one `active` row per site (enforced by a partial
+//! unique index in migration 010). [`rotate_key`] deactivates the current active
+//! row and inserts a new active one inside a single transaction so the invariant
+//! is never transiently violated.
+//!
+//! All queries use runtime `sqlx::query()` / `sqlx::query_as()` per the codebase
+//! convention (no compile-time `query!` macros, no `.sqlx` offline cache).
+
+use chrono::{DateTime, Utc};
+use serde::{Deserialize, Serialize};
+use sqlx::PgPool;
+use uuid::Uuid;
+
+/// Per-site enrollment key record.
+///
+/// `key_hash` is the only representation of the secret the server stores; the
+/// plaintext is shown once at issue/rotate and never persisted.
+#[derive(Debug, Clone, Serialize, Deserialize, sqlx::FromRow)]
+pub struct EnrollmentKey {
+    pub id: Uuid,
+    pub site_id: Uuid,
+    pub key_hash: String,
+    pub version: i32,
+    pub fingerprint: String,
+    pub active: bool,
+    pub created_at: DateTime<Utc>,
+    pub rotated_at: Option<DateTime<Utc>>,
+}
+
+/// Fetch the active enrollment key for a site, if any.
+///
+/// This is the `/api/enroll` hot path: resolve the one active key whose hash the
+/// presented `cek_` is verified against. The partial unique index guarantees at
+/// most one active row, so `fetch_optional` is correct.
+pub async fn get_active_for_site(
+    pool: &PgPool,
+    site_id: Uuid,
+) -> Result<Option<EnrollmentKey>, sqlx::Error> {
+    sqlx::query_as::<_, EnrollmentKey>(
+        r#"
+        SELECT id, site_id, key_hash, version, fingerprint, active, created_at, rotated_at
+        FROM site_enrollment_keys
+        WHERE site_id = $1 AND active
+        "#,
+    )
+    .bind(site_id)
+    .fetch_optional(pool)
+    .await
+}
+
+/// Insert the FIRST enrollment key for a site at version 1 (initial issue).
+///
+/// Use [`rotate_key`] for subsequent rotations. Errors with a unique violation if
+/// the site already has an active key (the caller should rotate instead).
+#[allow(dead_code)] // Wired by site-admin issue flow; Phase A exposes rotation (which also covers first issue when none exists).
+pub async fn insert_initial_key(
+    pool: &PgPool,
+    site_id: Uuid,
+    key_hash: &str,
+    fingerprint: &str,
+) -> Result<EnrollmentKey, sqlx::Error> {
+    sqlx::query_as::<_, EnrollmentKey>(
+        r#"
+        INSERT INTO site_enrollment_keys (site_id, key_hash, version, fingerprint, active)
+        VALUES ($1, $2, 1, $3, true)
+        RETURNING id, site_id, key_hash, version, fingerprint, active, created_at, rotated_at
+        "#,
+    )
+    .bind(site_id)
+    .bind(key_hash)
+    .bind(fingerprint)
+    .fetch_one(pool)
+    .await
+}
+
+/// Rotate a site's enrollment key (SPEC-016 §2): deactivate the current active key
+/// (if any) and insert a new active key at the next monotonic version, all in one
+/// transaction.
+///
+/// Returns the newly-created active key. If the site has no key yet, this issues
+/// version 1 (so rotation also serves as first-issue). The caller passes the
+/// already-hashed new secret and its fingerprint; the plaintext is surfaced once by
+/// the caller and never reaches this layer.
+///
+/// The transaction is what keeps the "at most one active key per site" invariant
+/// (partial unique index) from being transiently violated between the UPDATE and
+/// the INSERT.
+pub async fn rotate_key(
+    pool: &PgPool,
+    site_id: Uuid,
+    new_key_hash: &str,
+    new_fingerprint: &str,
+) -> Result<EnrollmentKey, sqlx::Error> {
+    let mut tx = pool.begin().await?;
+
+    // Highest existing version for this site (NULL -> 0 so the first key is v1).
+    let current_max: Option<i32> =
+        sqlx::query_scalar("SELECT MAX(version) FROM site_enrollment_keys WHERE site_id = $1")
+            .bind(site_id)
+            .fetch_one(&mut *tx)
+            .await?;
+    let next_version = current_max.unwrap_or(0) + 1;
+
+    // Deactivate the current active key (if any), stamping rotated_at.
+    sqlx::query(
+        r#"
+        UPDATE site_enrollment_keys
+        SET active = false, rotated_at = NOW()
+        WHERE site_id = $1 AND active
+        "#,
+    )
+    .bind(site_id)
+    .execute(&mut *tx)
+    .await?;
+
+    // Insert the new active key at the next version.
+    let new_key = sqlx::query_as::<_, EnrollmentKey>(
+        r#"
+        INSERT INTO site_enrollment_keys (site_id, key_hash, version, fingerprint, active)
+        VALUES ($1, $2, $3, $4, true)
+        RETURNING id, site_id, key_hash, version, fingerprint, active, created_at, rotated_at
+        "#,
+    )
+    .bind(site_id)
+    .bind(new_key_hash)
+    .bind(next_version)
+    .bind(new_fingerprint)
+    .fetch_one(&mut *tx)
+    .await?;
+
+    tx.commit().await?;
+    Ok(new_key)
+}

server/src/db/events.rs

@@ -69,6 +69,40 @@ impl EventTypes {
     pub const MACHINE_REMOVED: &'static str = "machine_removed";
     /// An administrator soft-deleted (purged) a session and dropped it in-memory.
     pub const SESSION_REMOVED: &'static str = "session_removed";
+
+    // Zero-touch enrollment events (SPEC-016). Written by POST /api/enroll and the
+    // site enrollment-key rotation endpoint. These carry no session, so they are
+    // logged via `log_enrollment_event` with `session_id = NULL`; the structured
+    // detail (machine_uid, site_code, fingerprint, etc.) goes in `details` and the
+    // source IP in `ip_address`.
+    /// A new machine self-registered at a site and was minted its first `cak_`.
+    pub const ENROLL_NEW: &'static str = "enroll_new";
+    /// An existing machine_uid re-enrolled at the SAME site — the row was reused and
+    /// a fresh `cak_` minted (re-image / re-install).
+    pub const ENROLL_REUSE: &'static str = "enroll_reuse";
+    /// An existing machine_uid enrolled under a DIFFERENT site — the machine's site
+    /// binding was updated (a "site move"). Fires an alert.
+    ///
+    /// NOTE (SPEC-016 Phase A): the unauthenticated enroll path does NOT perform this
+    /// move — a cross-site enroll is REFUSED (`ENROLL_SITE_CONFLICT`) rather than
+    /// silently repointing the machine. This event is reserved for the deliberate
+    /// Phase-B `--reassign` flow (and the dashboard move action) that supersede it.
+    #[allow(dead_code)] // reserved for Phase-B --reassign; not emitted by Phase A enroll
+    pub const ENROLL_SITE_MOVE: &'static str = "enroll_site_move";
+    /// An existing machine_uid presented a valid key for a DIFFERENT site than the one
+    /// the machine is currently bound to. Phase A REFUSES this (no move, no key minted)
+    /// as the accidental-move / cross-site-hijack guard; the deliberate move arrives
+    /// with the Phase-B `--reassign` flow + dashboard. Fires an alert.
+    pub const ENROLL_SITE_CONFLICT: &'static str = "enroll_site_conflict";
+    /// A machine_uid collision was detected at enroll — the endpoint dropped to
+    /// `pending` and awaits operator confirmation in the dashboard. Fires an alert.
+    pub const ENROLL_COLLISION_PENDING: &'static str = "enroll_collision_pending";
+    /// An enroll attempt failed enrollment-key verification (wrong/inactive key or
+    /// unknown site_code). Security audit trail for the open-registration surface.
+    pub const ENROLL_REJECTED: &'static str = "enroll_rejected";
+    /// An administrator rotated a site's enrollment key (new version + fingerprint;
+    /// old installers can no longer enroll NEW machines).
+    pub const ENROLLMENT_KEY_ROTATED: &'static str = "enrollment_key_rotated";
 }
 
 /// Log a session event
@@ -154,6 +188,42 @@ pub async fn log_admin_removal(
     Ok(result)
 }
 
+/// Log a zero-touch enrollment audit event (SPEC-016).
+///
+/// Shares the `connect_session_events` audit table but carries no session
+/// (`session_id = NULL`, the FK column is nullable) and no viewer — enrollment is
+/// an unauthenticated agent action, not a viewer/session event. The structured
+/// detail (machine_uid, site_code, fingerprint version, decision, etc.) goes in
+/// `details` and the agent's source IP in `ip_address`.
+///
+/// Best-effort: a failure to write the audit row must NOT fail the enroll (the
+/// machine row and `cak_` already exist); the caller logs the error and proceeds,
+/// matching how the relay and Task-5 removal treat audit writes.
+pub async fn log_enrollment_event(
+    pool: &PgPool,
+    event_type: &str,
+    details: JsonValue,
+    ip_address: Option<IpAddr>,
+) -> Result<i64, sqlx::Error> {
+    let ip_str = ip_address.map(|ip| ip.to_string());
+
+    let result = sqlx::query_scalar::<_, i64>(
+        r#"
+        INSERT INTO connect_session_events
+            (session_id, event_type, viewer_id, viewer_name, details, ip_address)
+        VALUES (NULL, $1, NULL, NULL, $2, $3::inet)
+        RETURNING id
+        "#,
+    )
+    .bind(event_type)
+    .bind(details)
+    .bind(ip_str)
+    .fetch_one(pool)
+    .await?;
+
+    Ok(result)
+}
+
 /// Get events for a session
 #[allow(dead_code)] // TODO(native-remote-control): consumed by the integration API; see docs/specs/native-remote-control/
 pub async fn get_session_events(

server/src/db/machines.rs

@@ -64,6 +64,16 @@ pub struct Machine {
     /// history) is retained. NULL = live. Nullable, so it is read NULL-tolerantly
     /// in the manual `FromRow` below.
     pub deleted_at: Option<DateTime<Utc>>,
+    /// Relational site binding for a machine enrolled via `/api/enroll` (SPEC-016,
+    /// migration 010). NULL for legacy / support-code / connect-path machines that
+    /// never enrolled through the zero-touch flow. A change of this on re-enroll is
+    /// the "site move" the enroll path audits.
+    pub site_id: Option<Uuid>,
+    /// Collision-gate state (SPEC-016, migration 010): `'active'` (live, auto-approve)
+    /// or `'pending'` (a machine_uid collision was detected at enroll; awaiting
+    /// operator confirmation before the endpoint may be controlled). Non-null with a
+    /// default of `'active'`; read NULL-tolerantly below for defense in depth.
+    pub enrollment_state: String,
 }
 
 impl<'r> FromRow<'r, PgRow> for Machine {
@@ -83,6 +93,13 @@ impl<'r> FromRow<'r, PgRow> for Machine {
             machine_uid: row.try_get("machine_uid")?,
             // Schema-nullable (migration 009); decode directly as Option.
             deleted_at: row.try_get("deleted_at")?,
+            // Schema-nullable (migration 010); decode directly as Option.
+            site_id: row.try_get("site_id")?,
+            // Non-null with default 'active' (migration 010); read NULL-tolerantly
+            // (older snapshots / partial rows) and fall back to 'active'.
+            enrollment_state: row
+                .try_get::<Option<String>, _>("enrollment_state")?
+                .unwrap_or_else(|| "active".to_string()),
             // Nullable-with-default columns mapped to non-`Option` Rust types: read as
             // `Option<T>` and fall back to the type default so a NULL cell never errors.
             is_elevated: row
@@ -166,7 +183,7 @@ pub async fn upsert_machine(
                 r#"
                 INSERT INTO connect_machines (agent_id, hostname, is_persistent, status, last_seen, machine_uid)
                 VALUES ($1, $2, $3, 'online', NOW(), $4)
-                ON CONFLICT (machine_uid) DO UPDATE SET
+                ON CONFLICT (machine_uid) WHERE machine_uid IS NOT NULL DO UPDATE SET
                     agent_id = EXCLUDED.agent_id,
                     hostname = EXCLUDED.hostname,
                     status = 'online',
@@ -207,6 +224,131 @@ pub async fn upsert_machine(
     }
 }
 
+/// Find a machine by the SPEC-016 per-tenant dedup key `(tenant_id, machine_uid)`.
+///
+/// This is the enroll-time dedup lookup: the same hardware re-enrolling (re-image /
+/// re-install) resolves to its existing row within the tenant, while the same
+/// hardware in a DIFFERENT tenant is a distinct row (resolved-decision #4). Tenant
+/// scoping uses the same default-tenant fold as the unique index so the lookup
+/// matches the uniqueness guarantee.
+///
+/// Unlike `get_machine_by_agent_id`, this deliberately does NOT filter
+/// `deleted_at IS NULL`: a previously operator-purged machine that legitimately
+/// re-enrolls must be found so the enroll path can revive it (clearing
+/// `deleted_at`), mirroring the connect-path revive in `upsert_machine`.
+pub async fn get_machine_by_tenant_uid(
+    pool: &PgPool,
+    tenant_id: Uuid,
+    machine_uid: &str,
+) -> Result<Option<Machine>, sqlx::Error> {
+    sqlx::query_as::<_, Machine>(
+        r#"
+        SELECT * FROM connect_machines
+        WHERE machine_uid = $1
+          AND COALESCE(tenant_id, '00000000-0000-0000-0000-000000000001'::uuid) = $2
+        "#,
+    )
+    .bind(machine_uid)
+    .bind(tenant_id)
+    .fetch_optional(pool)
+    .await
+}
+
+/// Parameters for an enroll-time machine create/update (SPEC-016 `/api/enroll`).
+///
+/// `agent_id` is a freshly minted opaque id for a NEW enrollment (the agent's
+/// config UUID story is Phase B; the server only needs a unique non-null value for
+/// the `agent_id UNIQUE` column). On REUSE/MOVE the existing row's `agent_id` is
+/// preserved (the FK target of any already-minted `cak_`), so the update path does
+/// not touch it.
+pub struct EnrollMachineParams<'a> {
+    pub agent_id: &'a str,
+    pub hostname: &'a str,
+    pub machine_uid: &'a str,
+    pub tenant_id: Uuid,
+    pub site_id: Uuid,
+    /// Company label (-> connect_machines.organization).
+    pub company: Option<&'a str>,
+    /// Site label (-> connect_machines.site) — the free-text label, distinct from
+    /// the relational site_id binding.
+    pub site_label: Option<&'a str>,
+    pub tags: &'a [String],
+    /// 'active' (auto-approve) or 'pending' (collision-gated).
+    pub enrollment_state: &'a str,
+}
+
+/// Insert a NEW machine row for a first-time enrollment (SPEC-016).
+///
+/// Carries the labels, the relational `site_id`, the per-tenant `machine_uid`, and
+/// the collision-gate `enrollment_state`. Persistent + online. Returns the created
+/// row (its `id` is the FK target for the `cak_` the caller mints next).
+pub async fn insert_enrolled_machine(
+    pool: &PgPool,
+    p: &EnrollMachineParams<'_>,
+) -> Result<Machine, sqlx::Error> {
+    sqlx::query_as::<_, Machine>(
+        r#"
+        INSERT INTO connect_machines
+            (agent_id, hostname, is_persistent, status, last_seen, machine_uid,
+             tenant_id, site_id, organization, site, tags, enrollment_state)
+        VALUES ($1, $2, true, 'online', NOW(), $3, $4, $5, $6, $7, $8, $9)
+        RETURNING *
+        "#,
+    )
+    .bind(p.agent_id)
+    .bind(p.hostname)
+    .bind(p.machine_uid)
+    .bind(p.tenant_id)
+    .bind(p.site_id)
+    .bind(p.company)
+    .bind(p.site_label)
+    .bind(p.tags)
+    .bind(p.enrollment_state)
+    .fetch_one(pool)
+    .await
+}
+
+/// Update an EXISTING machine row on re-enroll / reuse / site-move (SPEC-016).
+///
+/// Refreshes hostname, site binding (`site_id`), labels, and `enrollment_state`,
+/// and revives a soft-deleted row (`deleted_at = NULL`) — a re-enroll of a purged
+/// host means it is live again, mirroring `upsert_machine`'s revive. Deliberately
+/// does NOT change `agent_id`: the existing id is the FK target of any prior `cak_`.
+/// Labels are COALESCE-merged so an enroll that omits a label does not wipe an
+/// existing value; `tags` is overwritten only when a non-empty set is supplied
+/// (matching `update_machine_metadata`'s convention).
+pub async fn update_enrolled_machine(
+    pool: &PgPool,
+    machine_id: Uuid,
+    p: &EnrollMachineParams<'_>,
+) -> Result<Machine, sqlx::Error> {
+    sqlx::query_as::<_, Machine>(
+        r#"
+        UPDATE connect_machines SET
+            hostname = $2,
+            site_id = $3,
+            organization = COALESCE($4, organization),
+            site = COALESCE($5, site),
+            tags = CASE WHEN $6::text[] = '{}' THEN tags ELSE $6 END,
+            enrollment_state = $7,
+            status = 'online',
+            last_seen = NOW(),
+            deleted_at = NULL
+        WHERE id = $1
+        RETURNING *
+        "#,
+    )
+    .bind(machine_id)
+    .bind(p.hostname)
+    .bind(p.site_id)
+    .bind(p.company)
+    .bind(p.site_label)
+    .bind(p.tags)
+    .bind(p.enrollment_state)
+    .fetch_one(pool)
+    .await
+}
+
 /// Update machine status and info
 #[allow(dead_code)] // TODO(native-remote-control): consumed by the integration API; see docs/specs/native-remote-control/
 pub async fn update_machine_status(

server/src/db/mod.rs

@@ -4,10 +4,12 @@
 //! Optional - server works without database if DATABASE_URL not set.
 
 pub mod agent_keys;
+pub mod enrollment_keys;
 pub mod events;
 pub mod machines;
 pub mod releases;
 pub mod sessions;
+pub mod sites;
 pub mod support_codes;
 pub mod tenancy;
 pub mod users;

server/src/db/sites.rs

@@ -0,0 +1,94 @@
+//! Site database operations (SPEC-016 zero-touch enrollment).
+//!
+//! Backs the `connect_sites` table (migration 010): the relational anchor a
+//! per-site enrollment key hangs off and the `/api/enroll` flow resolves by
+//! `site_code`. See the migration header for why this table exists (the prior
+//! schema modeled "site" only as a free-text column on `connect_machines`).
+//!
+//! All queries use runtime `sqlx::query()` / `sqlx::query_as()` per the codebase
+//! convention (no compile-time `query!` macros, no `.sqlx` offline cache).
+
+use chrono::{DateTime, Utc};
+use serde::{Deserialize, Serialize};
+use sqlx::PgPool;
+use uuid::Uuid;
+
+/// Site record from the database.
+#[derive(Debug, Clone, Serialize, Deserialize, sqlx::FromRow)]
+pub struct Site {
+    pub id: Uuid,
+    pub site_code: String,
+    pub display_name: Option<String>,
+    pub company: Option<String>,
+    pub tenant_id: Option<Uuid>,
+    /// RESERVED for future per-site enrollment POLICY work (SPEC-016 §out-of-scope).
+    /// Not enforced in Phase A.
+    pub enrollment_policy: Option<String>,
+    pub created_at: DateTime<Utc>,
+}
+
+/// Resolve a site by its operator-facing `site_code`, scoped to the given tenant.
+///
+/// Tenant scoping uses the same default-tenant fold as the unique index so the
+/// lookup matches the uniqueness guarantee: `(COALESCE(tenant_id, default),
+/// site_code)`. Returns `None` if no site with that code exists in the tenant.
+pub async fn get_site_by_code(
+    pool: &PgPool,
+    site_code: &str,
+    tenant_id: Uuid,
+) -> Result<Option<Site>, sqlx::Error> {
+    sqlx::query_as::<_, Site>(
+        r#"
+        SELECT id, site_code, display_name, company, tenant_id, enrollment_policy, created_at
+        FROM connect_sites
+        WHERE site_code = $1
+          AND COALESCE(tenant_id, '00000000-0000-0000-0000-000000000001'::uuid) = $2
+        "#,
+    )
+    .bind(site_code)
+    .bind(tenant_id)
+    .fetch_optional(pool)
+    .await
+}
+
+/// Fetch a site by its primary-key UUID.
+pub async fn get_site_by_id(pool: &PgPool, id: Uuid) -> Result<Option<Site>, sqlx::Error> {
+    sqlx::query_as::<_, Site>(
+        r#"
+        SELECT id, site_code, display_name, company, tenant_id, enrollment_policy, created_at
+        FROM connect_sites
+        WHERE id = $1
+        "#,
+    )
+    .bind(id)
+    .fetch_optional(pool)
+    .await
+}
+
+/// Insert a new site, returning the created row.
+///
+/// `tenant_id` is `None`-tolerant and resolved via `db::tenancy::current_tenant_id()`
+/// at the call site. Errors with a unique-violation if `(tenant, site_code)` already
+/// exists (the caller maps that to a 409).
+#[allow(dead_code)] // Wired by the site-admin API (dashboard site CRUD); Phase A exposes key rotation, not site CRUD.
+pub async fn insert_site(
+    pool: &PgPool,
+    site_code: &str,
+    display_name: Option<&str>,
+    company: Option<&str>,
+    tenant_id: Option<Uuid>,
+) -> Result<Site, sqlx::Error> {
+    sqlx::query_as::<_, Site>(
+        r#"
+        INSERT INTO connect_sites (site_code, display_name, company, tenant_id)
+        VALUES ($1, $2, $3, $4)
+        RETURNING id, site_code, display_name, company, tenant_id, enrollment_policy, created_at
+        "#,
+    )
+    .bind(site_code)
+    .bind(display_name)
+    .bind(company)
+    .bind(tenant_id)
+    .fetch_one(pool)
+    .await
+}

server/src/main.rs

@@ -448,6 +448,11 @@ async fn main() -> Result<()> {
             )),
         )
         .route("/api/codes/:code/cancel", post(cancel_code))
+        // Zero-touch enrollment (SPEC-016). PUBLIC: no JWT — the per-site enrollment
+        // key in the body is the gate, and the handler applies its own
+        // per-(site_code, IP) rate limit / lockout (defense-in-depth). Mounted with
+        // the other public API routes.
+        .route("/api/enroll", post(api::enroll::enroll))
         // WebSocket endpoints
         .route("/ws/agent", get(relay::agent_ws_handler))
         .route("/ws/viewer", get(relay::viewer_ws_handler))
@@ -498,6 +503,18 @@ async fn main() -> Result<()> {
             "/api/machines/:agent_id/keys/:key_id",
             delete(api::machine_keys::revoke_key),
         )
+        // Per-site enrollment key administration (SPEC-016, admin-only / JWT).
+        // Rotate regenerates the cek_ secret + fingerprint (old installers can no
+        // longer enroll new machines); GET returns the current non-secret
+        // fingerprint/version. Both gated by the AdminUser extractor.
+        .route(
+            "/api/sites/:id/enrollment-key",
+            get(api::sites::get_enrollment_key),
+        )
+        .route(
+            "/api/sites/:id/enrollment-key/rotate",
+            post(api::sites::rotate_enrollment_key),
+        )
         // REST API - Releases and Version
         .route("/api/version", get(api::releases::get_version)) // No auth - for agent polling
         .route("/api/releases", get(api::releases::list_releases))

server/src/middleware/rate_limit.rs

@@ -77,6 +77,19 @@ pub const CODE_VALIDATE_MAX_FAILURES: u32 = 10;
 /// Support-code validate: how long an IP stays locked out once tripped.
 pub const CODE_VALIDATE_LOCKOUT: Duration = Duration::from_secs(15 * 60);
 
+/// Enroll (`POST /api/enroll`, SPEC-016): window length.
+pub const ENROLL_WINDOW: Duration = Duration::from_secs(60);
+/// Enroll: max requests per window per `(site_code, IP)`. A zero-touch site push
+/// drives N machines through enroll near-simultaneously, so this is generous
+/// (mass-deploy friendly) while still capping a runaway loop. Defense-in-depth: the
+/// 256-bit enrollment key is the load-bearing gate, not this cap.
+pub const ENROLL_MAX_PER_WINDOW: u32 = 60;
+/// Enroll: consecutive FAILED enroll attempts (bad/inactive key, unknown site) from
+/// one `(site_code, IP)` that trip the lockout.
+pub const ENROLL_MAX_FAILURES: u32 = 20;
+/// Enroll: how long a `(site_code, IP)` stays locked out once tripped.
+pub const ENROLL_LOCKOUT: Duration = Duration::from_secs(15 * 60);
+
 /// Hard cap on the number of distinct IPs tracked by any single limiter map.
 /// Prevents an IP-rotating attacker from growing memory without bound. When the
 /// cap is hit, the oldest-windowed entries are pruned. Generous for a real MSP
@@ -260,6 +273,150 @@ impl FailureLockout {
     }
 }
 
+// ============================================================================
+// Composite-key limiter for enrollment (keyed by (site_code, IP)) — SPEC-016
+// ============================================================================
+//
+// The login / change-password / code-validate limiters above key purely on IP.
+// SPEC-016 §3 wants the enroll defense keyed on `(site_code, source-IP)` so a noisy
+// site push from one office IP cannot lock out a different site enrolling from the
+// same egress IP. Rather than overload the IP-only maps, this is a small dedicated
+// composite-key limiter + lockout. It is invoked from the enroll HANDLER (not a
+// `from_fn` layer) because the `site_code` lives in the JSON body, which a
+// pre-handler middleware cannot read without consuming it. Documented as
+// defense-in-depth: the 256-bit enrollment key is the real gate.
+
+/// Composite limiter key: the site_code and the real client IP.
+type EnrollKey = (String, IpAddr);
+
+/// Per-`(site_code, IP)` fixed-window limiter + consecutive-failure lockout.
+///
+/// Combines both protections behind one lock-guarded map so the enroll handler
+/// makes a single allow/deny decision and reports success/failure into the same
+/// structure. Self-pruning and size-capped, like the IP-only limiters.
+#[derive(Clone)]
+pub struct EnrollLimiter {
+    inner: std::sync::Arc<Mutex<HashMap<EnrollKey, EnrollEntry>>>,
+    max_per_window: u32,
+    window: Duration,
+    max_failures: u32,
+    cooldown: Duration,
+}
+
+#[derive(Debug, Clone, Copy)]
+struct EnrollEntry {
+    window_started: Instant,
+    count: u32,
+    failures: u32,
+    locked_until: Option<Instant>,
+    last_seen: Instant,
+}
+
+impl EnrollLimiter {
+    pub fn new(
+        max_per_window: u32,
+        window: Duration,
+        max_failures: u32,
+        cooldown: Duration,
+    ) -> Self {
+        Self {
+            inner: std::sync::Arc::new(Mutex::new(HashMap::new())),
+            max_per_window,
+            window,
+            max_failures,
+            cooldown,
+        }
+    }
+
+    fn entry_now() -> EnrollEntry {
+        let now = Instant::now();
+        EnrollEntry {
+            window_started: now,
+            count: 0,
+            failures: 0,
+            locked_until: None,
+            last_seen: now,
+        }
+    }
+
+    /// Admit one enroll attempt for `(site_code, ip)`. Returns `true` if allowed
+    /// (and counts it). Returns `false` if the key is currently locked out OR over
+    /// the per-window request cap. Clock injected for tests.
+    fn check_at(&self, site_code: &str, ip: IpAddr, now: Instant) -> bool {
+        let mut map = self.inner.lock().unwrap_or_else(|e| e.into_inner());
+
+        if map.len() >= MAX_TRACKED_IPS {
+            let window = self.window;
+            let cooldown = self.cooldown;
+            map.retain(|_, e| {
+                e.locked_until.map(|u| now < u).unwrap_or(false)
+                    || now.duration_since(e.window_started) < window
+                    || now.duration_since(e.last_seen) < cooldown
+            });
+        }
+
+        let key = (site_code.to_string(), ip);
+        let e = map.entry(key).or_insert_with(Self::entry_now);
+        e.last_seen = now;
+
+        // Lockout takes precedence.
+        if let Some(until) = e.locked_until {
+            if now < until {
+                return false;
+            }
+            // Cooldown elapsed — clear it for a fresh start.
+            e.locked_until = None;
+            e.failures = 0;
+        }
+
+        // Roll the fixed window forward if elapsed.
+        if now.duration_since(e.window_started) >= self.window {
+            e.window_started = now;
+            e.count = 0;
+        }
+
+        if e.count >= self.max_per_window {
+            false
+        } else {
+            e.count += 1;
+            true
+        }
+    }
+
+    /// Admit one enroll attempt (real clock).
+    pub fn check(&self, site_code: &str, ip: IpAddr) -> bool {
+        self.check_at(site_code, ip, Instant::now())
+    }
+
+    fn record_failure_at(&self, site_code: &str, ip: IpAddr, now: Instant) {
+        let mut map = self.inner.lock().unwrap_or_else(|e| e.into_inner());
+        let key = (site_code.to_string(), ip);
+        let e = map.entry(key).or_insert_with(Self::entry_now);
+        e.last_seen = now;
+        e.failures = e.failures.saturating_add(1);
+        if e.failures >= self.max_failures {
+            e.locked_until = Some(now + self.cooldown);
+        }
+    }
+
+    /// Record a FAILED enroll attempt (bad key / unknown site) for the key,
+    /// tripping the lockout once the streak reaches `max_failures`.
+    pub fn record_failure(&self, site_code: &str, ip: IpAddr) {
+        self.record_failure_at(site_code, ip, Instant::now());
+    }
+
+    /// Record a SUCCESSFUL enroll for the key, resetting its failure streak.
+    pub fn record_success(&self, site_code: &str, ip: IpAddr) {
+        let mut map = self.inner.lock().unwrap_or_else(|e| e.into_inner());
+        let key = (site_code.to_string(), ip);
+        if let Some(e) = map.get_mut(&key) {
+            e.failures = 0;
+            e.locked_until = None;
+            e.last_seen = Instant::now();
+        }
+    }
+}
+
 // ============================================================================
 // Shared rate-limit state (lives in AppState)
 // ============================================================================
@@ -275,6 +432,9 @@ pub struct RateLimitState {
     pub code_validate: RateLimiter,
     /// Per-IP lockout on repeated failed code validations (brute-force defense).
     pub code_validate_lockout: FailureLockout,
+    /// `POST /api/enroll` (SPEC-016): per-`(site_code, IP)` request cap +
+    /// consecutive-failure lockout. Invoked from the enroll handler.
+    pub enroll: EnrollLimiter,
 }
 
 impl RateLimitState {
@@ -290,6 +450,12 @@ impl RateLimitState {
                 CODE_VALIDATE_MAX_FAILURES,
                 CODE_VALIDATE_LOCKOUT,
             ),
+            enroll: EnrollLimiter::new(
+                ENROLL_MAX_PER_WINDOW,
+                ENROLL_WINDOW,
+                ENROLL_MAX_FAILURES,
+                ENROLL_LOCKOUT,
+            ),
         }
     }
 }
@@ -524,4 +690,51 @@ mod tests {
         assert!(lockout.is_locked_at(ip(8), t0));
         assert!(!lockout.is_locked_at(ip(9), t0)); // ip9 unaffected
     }
+
+    // -- EnrollLimiter (composite (site_code, IP) key) --------------------------
+
+    #[test]
+    fn enroll_window_allows_up_to_cap_then_blocks() {
+        let lim = EnrollLimiter::new(2, Duration::from_secs(60), 100, Duration::from_secs(600));
+        let t0 = Instant::now();
+        assert!(lim.check_at("SITE-A", ip(1), t0)); // 1
+        assert!(lim.check_at("SITE-A", ip(1), t0)); // 2
+        assert!(!lim.check_at("SITE-A", ip(1), t0)); // over cap
+    }
+
+    #[test]
+    fn enroll_is_keyed_by_site_and_ip() {
+        let lim = EnrollLimiter::new(1, Duration::from_secs(60), 100, Duration::from_secs(600));
+        let t0 = Instant::now();
+        assert!(lim.check_at("SITE-A", ip(1), t0));
+        assert!(!lim.check_at("SITE-A", ip(1), t0)); // same key over cap
+                                                     // Different site, same IP -> independent bucket.
+        assert!(lim.check_at("SITE-B", ip(1), t0));
+        // Same site, different IP -> independent bucket.
+        assert!(lim.check_at("SITE-A", ip(2), t0));
+    }
+
+    #[test]
+    fn enroll_lockout_trips_after_failures_and_blocks_check() {
+        let lim = EnrollLimiter::new(100, Duration::from_secs(60), 3, Duration::from_secs(600));
+        let t0 = Instant::now();
+        lim.record_failure_at("SITE-A", ip(1), t0);
+        lim.record_failure_at("SITE-A", ip(1), t0);
+        // Not yet tripped: a check still admits.
+        assert!(lim.check_at("SITE-A", ip(1), t0));
+        lim.record_failure_at("SITE-A", ip(1), t0); // 3rd -> trips
+                                                    // Now locked out: check denies even though under the request cap.
+        assert!(!lim.check_at("SITE-A", ip(1), t0));
+    }
+
+    #[test]
+    fn enroll_success_resets_failure_streak() {
+        let lim = EnrollLimiter::new(100, Duration::from_secs(60), 2, Duration::from_secs(600));
+        let t0 = Instant::now();
+        lim.record_failure_at("SITE-A", ip(1), t0);
+        lim.record_success("SITE-A", ip(1)); // reset
+        lim.record_failure_at("SITE-A", ip(1), t0);
+        // Only one failure since reset -> not locked.
+        assert!(lim.check_at("SITE-A", ip(1), t0));
+    }
 }

specs/v2-secure-session-core/plan.md

@@ -527,3 +527,60 @@ Reference: SPEC-002 §5; `agent/src/encoder/raw.rs` (salvaged), `proto/guruconne
 - **Rate limiting:** hammer `/api/auth/login` and the code-validate route → confirm throttling/lockout.
 - **Migrations:** fresh DB applies the v2 migrations cleanly; `_sqlx_migrations` consistent; `tenant_id`
   populated with the default tenant.
+
+---
+
+## Task 9 [PROPOSED 2026-06-01 — provisioning model = TOFU auto-enroll, chosen by Mike]: `cak_` auto-enroll provisioning + shared-key retirement
+
+> Context: Task 2 built the SERVER `cak_` machinery (mint/SHA-256 hash/verify in `auth/agent_keys.rs`,
+> relay validation in `validate_agent_api_key`, admin issuance `POST /api/machines/:id/keys`). What's
+> missing is how an AGENT obtains and uses a `cak_` — today agents still carry the deprecated shared
+> `AGENT_API_KEY`, so `connect_agent_keys` is empty and the relay logs the DEPRECATED-shared-key warning
+> for every agent. This task closes that with **trust-on-first-use auto-enroll** so the shared key can be
+> retired (unblocks task list #5). NOTE: the agent already presents whatever is in its `api_key` slot and
+> the relay auto-detects `cak_` vs shared — so a `cak_`-keyed agent needs **no change to its auth call**,
+> only a way to *receive*, *persist*, and *prefer* a `cak_`.
+
+**Flow (TOFU):**
+1. **Bootstrap (first connect):** a fresh agent authenticates on `/ws/agent` with a bootstrap secret —
+   interim: the shared `AGENT_API_KEY` (embedded by the download endpoint); target: a single-use,
+   short-lived **enroll token** (more secure TOFU — see Security). 
+2. **Server issues on first connect:** when an agent authed via the bootstrap path (i.e. NOT already
+   `cak_`-keyed) connects and its machine has **no active (non-revoked) `cak_`**, the relay: resolves/creates
+   the machine row (existing `upsert_machine` on `machine_uid` — now functional after the 2026-06-01
+   ON CONFLICT fix), mints a `cak_` (`generate_agent_key` + `db::agent_keys::insert_agent_key` for that
+   `machine_id`), and sends the plaintext key to the agent **once** over a new server→agent message. Only
+   the hash is stored. **Idempotent:** never re-issue if an active key already exists for the machine.
+3. **Agent receives + persists + prefers:** on `AgentKeyProvision`, the agent persists the `cak_` durably at
+   `%ProgramData%\GuruConnect\agent_key` (restricted ACL, same pattern as `machine_uid`). On startup it loads
+   the persisted `cak_` if present and uses it as its auth key, falling back to the embedded/bootstrap secret
+   only when no `cak_` is stored yet. After provisioning, every reconnect authenticates via `cak_` (no more
+   DEPRECATED-shared-key warning for that agent).
+4. **Shared-key retirement (phased):** Phase A — shared key stays as the bootstrap so existing+new agents
+   self-enroll; monitor the relay WARN count → ~0. Phase B — once the fleet is `cak_`-keyed, restrict the
+   shared `AGENT_API_KEY` to enrollment-only or remove the env entirely (only `cak_` / enroll-token accepted).
+   This is the concrete completion of task-list #5.
+
+**Protocol (4-artifact drift discipline):** add `AgentKeyProvision { string key = 1; }` (server→agent) to
+`proto/guruconnect.proto` with a new reserved message ID; regenerate prost on both agent + server; the
+hand-written `dashboard/src/lib/protobuf.ts` decoder does NOT need it (agent-plane only) but reserve the ID.
+
+**Files:** `proto/guruconnect.proto` (new message); `server/src/relay/mod.rs` (issue+send on bootstrap connect
+with no active key); `server/src/db/agent_keys.rs` (add `has_active_key(machine_id)` check; reuse insert);
+`agent/src/transport/*` (handle inbound `AgentKeyProvision`); `agent/src/config.rs` + a small key-store module
+(load/persist `cak_`, prefer over bootstrap).
+
+**Security (TOFU):** the first connect trusts the bootstrap secret — a leaked shared key during the enroll
+window could enroll a rogue agent; the secure target is a **single-use, short-lived enroll token** per
+deployment instead of the shared key (shared-key bootstrap is interim convenience). The `cak_` is sent
+plaintext once over the existing wss/TLS channel; only the hash is stored server-side; the agent stores it
+locally with restricted ACLs. Revocation via the existing `DELETE /api/machines/:id/keys/:key_id` fails the
+agent closed; on its next bootstrap connect it re-enrolls. The keyed-agent dedup (Task 3) keeps the
+authenticated identity authoritative.
+
+**Verification:** drop a current-build (signed 0.3.0+) agent configured with the shared-key bootstrap →
+it connects, receives a `cak_`, persists it; restart → it authenticates via the `cak_` (relay shows NO
+DEPRECATED-shared-key warning) and `connect_agent_keys` holds exactly one active key for the machine; issue
+is idempotent across reconnects; revoke the key via the admin API → agent rejected, then re-enrolls on next
+bootstrap connect. Reference: `auth/agent_keys.rs`, `api/machine_keys.rs`, `relay/mod.rs:266-309`
+(`validate_agent_api_key`), `.claude/standards/security/credential-handling.md`.