claudetools/session-logs/2026-05-29-session.md

Session Log — 2026-05-29

User

• User: Mike Swanson (mike)
• Machine: GURU-5070
• Role: admin

Session Summary

Shaped a pre-implementation spec for native integrated remote control in the GuruRMM ecosystem, then restructured how the guru-connect product is tracked in the monorepo. The session began as a "fix lint errors" request that was redirected into a GuruRMM feature request for the guru-connect (GC) project: native, integrated remote control comparable to ScreenConnect/Splashtop, built entirely on our own Rust stack to avoid third-party agents and supply-chain exposure.

Research established that GC already implements the full remote-control engine (DXGI/GDI capture, input injection, viewer, guruconnect:// protocol handler, persistent/unattended + support-code/attended modes, protobuf over WSS) and that GuruRMM already has the orchestration rails (per-agent command dispatch, stable device_id identity, the AgentDetail action-button pattern, and a half-built generic tunnel scaffold). Two parallel Explore agents mapped the exact integration surfaces with file:line references. The feature is therefore ~80% wiring against existing capability, not greenfield. Architecture decisions were captured via the user: broker model (RMM orchestrates the separate GC agent), both unattended and attended access, multi-monitor in scope, file transfer / session recording / non-Windows agents out of scope, priority P2.

The /shape-spec skill produced four files in projects/msp-tools/guru-connect/specs/native-remote-control/ (shape, plan, references, standards). The user then clarified that GC is a standalone product with its own pipeline/cadence, and the real intent is a durable, versioned integration contract so the two products stay integration-compatible without coupling. The spec was rewritten around a GC-owned, semver'd integration contract (/api/integration/v1/, capability discovery, embedded session viewer). A concrete blocker was identified: GC's security_headers.rs sets frame-ancestors 'none', which must be relaxed to a scoped RMM-origin allowlist for the embedded viewer. RMM-side hints were added (ADR-008 + docs/GURU_CONNECT_INTEGRATION.md) recording that RMM consumes GC via the contract and does no active dev on GC. The spec and hints were committed across both repos (commit-only, no push).

The user then asked to wire GC as a submodule like guru-rmm. Investigation revealed the remote azcomputerguru/guru-connect repo was ~4 months stale (frozen 2026-01-18) while the local monorepo copy was far ahead (entire middleware/metrics/utils modules, token blacklist, Phase-1 security/deploy work, the new spec). Per the user's decisions (publish local to the existing repo as a snapshot commit; preserve history), the Gitea Agent published the local working state to GC main (fast-forward 5b7cf5f..e3e95f8, history preserved, KEEP paths .gitignore/.cargo/server/static/downloads retained), then converted the vendored directory into a submodule pinned at e3e95f8. Confirmed that GC deploy.yml triggers only on v*.*.* tags / manual dispatch, so the push ran CI build/test but did not deploy to production.

Finally, the user confirmed RMM and GC are the only versionable products; everything else stays in the monorepo. This policy was recorded to memory (project_versionable_products.md).

Key Decisions

• Broker architecture: RMM orchestrates the separate GC agent (two agents coexist) rather than merging GC into the RMM agent — reuses GC's existing engine, ships sooner, keeps GC standalone.
• The deliverable is a GC-owned, semver'd integration contract + capability discovery, not one-off broker wiring — so the two products stay in-sync via the contract without sharing pipelines or releasing in lockstep.
• Stable cross-product identity = RMM device_id passed as the GC agent_id, so brokered sessions deterministically match the endpoint.
• Supply-chain guard made concrete: the RMM agent downloads the GC binary only from GC's release channel and verifies SHA-256 before launch (reusing GC's releases.checksum_sha256).
• Embedded viewer over native-only: relax frame-ancestors/X-Frame-Options on the viewer route to a scoped RMM-origin allowlist; keep 'none' everywhere else.
• Spec lives in the GC repo (GC owns the contract); RMM gets ADR-008 + a pointer doc reminding it not to perform active dev on GC.
• Submodule reconciliation: publish the local (authoritative) state up to the stale GC repo as a snapshot commit on top of existing main (preserve history), then submodule-add — nothing lost.
• Only GuruRMM and GuruConnect are versionable products (own repos/submodules); all other projects stay in the claudetools monorepo. Split only for an independent pipeline OR a versioned external consumer.
• All git operations committed but NOT pushed (claudetools), per the established pattern of leaving the push to the user; the GC repo push was mandatory for the submodule to resolve.

Problems Encountered

• Initial "fix lint errors" request was ambiguous (clean tree, multiple lintable projects). Asked which project; user redirected to the GC feature request instead.
• CLAUDE.md warns that a Gitea repo named guru-connect is an "abandoned duplicate." Verified by inspecting the remote repo's contents (proto/guruconnect.proto, agent/, server/, dashboard/) that azcomputerguru/guru-connect is the real GC product, not the abandoned RMM duplicate the warning refers to.
• The remote GC repo was 4 months stale and the local monorepo copy had diverged substantially (whole modules + Phase-1 work never pushed). A naive submodule add would have reverted that work. Resolved by diffing local vs remote, surfacing the divergence, and publishing local→remote before converting.
• Production-deploy risk on push: checked GC's .gitea/workflows; confirmed deploy.yml triggers only on v*.*.* tags / workflow_dispatch, so pushing to main runs CI but does not deploy.

Configuration Changes

Created (committed afbe5a8, then moved into the GC repo via the submodule conversion):

• projects/msp-tools/guru-connect/specs/native-remote-control/shape.md
• projects/msp-tools/guru-connect/specs/native-remote-control/plan.md
• projects/msp-tools/guru-connect/specs/native-remote-control/references.md
• projects/msp-tools/guru-connect/specs/native-remote-control/standards.md

guru-rmm submodule (committed 7701d26 in the submodule):

• Modified docs/ARCHITECTURE_DECISIONS.md — added ADR-008 (GC is a separate product consumed via versioned contract)
• Created docs/GURU_CONNECT_INTEGRATION.md — RMM-side boundary/pointer doc

Repo structure:

• .gitmodules — added projects/msp-tools/guru-connect submodule entry (branch main)
• projects/msp-tools/guru-connect — converted from vendored directory to submodule (gitlink mode 160000 at e3e95f8)

Memory:

• Created .claude/memory/project_versionable_products.md
• Updated .claude/memory/MEMORY.md index (Project section)

Credentials & Secrets

None discovered or created this session. The spec references secrets to be sourced from env/SOPS at implementation time (CONNECT_INTEGRATION_KEY, CONNECT_SERVER_URL, per-machine GC agent keys, CONNECT_EMBED_ALLOWED_ORIGINS) — none provisioned yet.

Infrastructure & Servers

• Gitea (internal): http://172.16.3.20:3000 — used for repo inspection + GC push (per internal-API preference)
• GC relay server: 172.16.3.30:3002, proxied via NPM to connect.azcomputerguru.com
• GuruRMM server: 172.16.3.30:3001, dashboard rmm.azcomputerguru.com
• GC repo CI: .gitea/workflows/{build-and-test,test,deploy}.yml — deploy only on v*.*.* tags / manual dispatch

Commands & Outputs

Repo divergence check (local vs remote GC), shallow clone + diff -rq — confirmed local far ahead; cleaned up temp clone afterward.

GC publish (Gitea Agent):

• git push origin main → 5b7cf5f..e3e95f8 main -> main (fast-forward, 73 files changed, 15611 insertions, 5760 deletions; Cargo.lock dropped — not tracked in the authoritative copy)

Submodule conversion (Gitea Agent):

• git rm -r --cached projects/msp-tools/guru-connect + rm -rf + git submodule add -b main <url>
• git submodule status → e3e95f8 ... guru-connect (heads/main), 7701d26 ... guru-rmm (heads/main)

Pending / Incomplete Tasks

• claudetools commits are LOCAL, not pushed: 53e14da (submodule conversion) + 1fc2401/afbe5a8 (spec + pointer bump) from earlier. Push when ready.
• GC repo housekeeping: re-add Cargo.lock (dropped in the snapshot; wanted for reproducible builds).
• GC submodule URL uses the internal IP 172.16.3.20:3000; guru-rmm uses the public git.azcomputerguru.com. Off-network clones (Howard's Mac) won't resolve the internal IP — consider switching to the public hostname for parity.
• GC CI run kicked off by the publish push may be red (the snapshot may not build cleanly; Cargo.lock removed). Check the Actions run.
• Implementation of the feature itself has not started — Task 0 of the spec (commit the spec) is effectively satisfied; Tasks 1+ are not begun.

Reference Information

• Spec: projects/msp-tools/guru-connect/specs/native-remote-control/ (4 files) — now in the GC repo at e3e95f8
• ADR: projects/msp-tools/guru-rmm/docs/ARCHITECTURE_DECISIONS.md ADR-008
• RMM pointer: projects/msp-tools/guru-rmm/docs/GURU_CONNECT_INTEGRATION.md
• GC repo: azcomputerguru/guru-connect; published 5b7cf5f → e3e95f8
• guru-rmm submodule commit: 7701d26
• claudetools commits: afbe5a8 (spec), 1fc2401 (submodule ptr bump for ADR), 53e14da (GC submodule conversion)
• Roadmap context: projects/msp-tools/guru-rmm/docs/FEATURE_ROADMAP.md:635-675, docs/UI_GAPS.md:155-186
• Key GC integration files: server/src/middleware/security_headers.rs:30,37-39 (frame-ancestors), server/static/viewer.html, server/src/relay/mod.rs:187 (agent key validation), server/src/main.rs:300 (/api/version)
• Key RMM files: server/src/api/commands.rs:87-157 (command dispatch), agent/src/device_id.rs, dashboard/src/pages/AgentDetail.tsx:1893-1931