claudetools/session-logs/2026-05-26-guru-kali-attribution-hardening.md

# Session Log — Work Attribution Hardening

## User
- **User:** Mike Swanson (mike)
- **Machine:** GURU-KALI
- **Role:** admin
- **Session span:** 2026-05-26, afternoon–evening MST (ending ~18:47 MST)

## Session Summary

Routine `/sync` rounds (clean fast-forwards; one larger 21-commit pull bringing GuruScan, the coord-todos system, `/wiki-compile`, and Lone Star/QuantumWMS wiki seeds) led into an `identity.json` update mandated by a coord message from the Mac: added `"claudetools_root": "/home/guru/claudetools"` (detected via `git rev-parse --show-toplevel`) and replied to the originating session to confirm.

Mike then raised the core problem: recurring misattribution of work to the wrong user/hostname, often via stale machine names. Initial instinct was to treat this as a git-authorship/"attribution rules" gap and build enforcement (mailmap/history rewrite). Mike course-corrected twice — it is not a git/gitea problem, it is (1) stale/owner-less `MEMORY.md` items corrupting reasoning-time inference, and (2) a shortcoming in the `/save`+`/sync` commands. Forensics confirmed his read precisely: across 892 commits, **zero** had an author disagreeing with the named machine's owner; across 180 session logs, **zero** User-block mismatches. Git history and logs were already clean. The only misattribution vector was inference fed by stale memory (an "ACG-5070 workstation" note with no owner; Howard's note saying "desktop hostname TBD") plus attribution being a soft model instruction rather than a deterministic step.

Implemented a three-part fix and ran it through the Code Review Agent (which found one HIGH and one MEDIUM bug — both fixed and re-verified). Result: attribution is now read deterministically from `identity.json`/`users.json`/git authorship, never inferred; every machine in memory is bound to an owner; and the 5070 box's rename lineage is recorded as Mike's.

## Key Decisions

- **No git history rewrite.** Confirmed git authorship is clean (0 mismatches/892 commits), and the repo is shared + rebase-synced across ~6 machines, so a filter-repo rewrite would be destructive and unjustified. Mailmap was also dropped — there is no display-attribution problem to paper over.
- **Attribution is read, never inferred** — codified as the governing rule. Sources of truth: `identity.json` (who's at the keyboard), `users.json` (machine registry), git authorship. Never hostname patterns, the `userEmail` hint, or memory.
- **Deterministic User block** via a dedicated script (`whoami-block.sh`) that `/save` runs and pastes verbatim — removes the model's freedom to fill the block from inference.
- **git config reconciled to identity.json on every sync** (local config, both repos), so commit authorship cannot drift going forward.
- **Confirmed with Mike:** `OC-5070` / `acg-guru-5070` / `ACG-5070` / `GURU-5070` are one physical machine, renamed over time, all Mike's. Recorded as such.
- **Namespaced this log by machine** (`-guru-kali-attribution-hardening`) instead of appending to the shared `2026-05-26-session.md`, to keep attribution unambiguous — consistent with the work itself.

## Problems Encountered

- **Self-inflicted inference, live:** initial analysis asserted "OC-5070 = old GURU-5070" and "azcomputerguru = ambiguous" as fact — the exact failure mode under repair. Corrected to verifying against `users.json`/asking Mike rather than inferring.
- **Code review found a HIGH bug:** on a present-but-malformed `identity.json`, `USER_DISPLAY` fell back to the `"unknown"` sentinel and would have been written as the git author (clobbering correct config in both repos). Fixed by guarding both reconcile call sites to skip when `USER_DISPLAY == "unknown"` and warn instead. Verified: existing config preserved.
- **MEDIUM bug:** `whoami-block.sh` dumped a Python traceback and emitted an empty block on malformed `identity.json`. Fixed with a try/except fallback block + `exit 0`. Verified clean.
- **Could not re-review via SendMessage** (tool unavailable in this environment); fixes were the review agent's own prescribed changes and were each verified empirically instead of re-spawning a full review agent.

## Configuration Changes

**Modified:**
- `.claude/scripts/sync.sh` — added `reconcile_git_identity()`; called (guarded against the `unknown` sentinel) in the claudetools repo and the vault repo; added a stale identity.json-machine vs hostname warning.
- `.claude/commands/save.md` — User block now generated by `whoami-block.sh` (not hand-written/inferred); corrected the stale description of sync.sh staging (`git add -A` after garbled-path purge, not "by name").
- `.claude/memory/user_howard.md` — machines now ACG-TECH03L + Howard-Home, deferring to `users.json`; removed the "desktop hostname TBD" hole.
- `.claude/memory/reference_workstation_setup.md` — retitled to GURU-5070 (Mike's primary); recorded the OC-5070 → ACG-5070/acg-guru-5070 → GURU-5070 rename chain and explicit owner.
- `.claude/memory/MEMORY.md` — updated the two stale 5070 index lines and Howard's line; surfaced the new attribution rule at the top of Feedback.
- `.claude/identity.json` — added `claudetools_root: /home/guru/claudetools` (gitignored, per-machine; not synced).

**Created:**
- `.claude/scripts/whoami-block.sh` — deterministic `## User` block generator from identity.json (+users.json role); handles missing/malformed identity and missing Python.
- `.claude/memory/feedback_attribution_from_identity.md` — the keystone "attribution is read, never inferred" rule with the why and how-to-apply.

## Credentials & Secrets

None created or discovered. No secret values touched. The git config reconcile uses only the name/email already in `identity.json`.

## Infrastructure & Servers

- Coord API `http://172.16.3.30:8001/api/coord` — sent confirmation message `dfeb6f2a-2d90-4bfe-bf95-a14eec449b3d` (GURU-KALI → Mikes-MacBook-Air) re: claudetools_root.
- This machine: GURU-KALI (Linux/Kali), git config `Mike Swanson <mike@azcomputerguru.com>` (matches identity.json — reconcile is a no-op here).
- Machine→owner registry remains authoritative in `.claude/users.json`: mike = GURU-5070, Mikes-MacBook-Air, GURU-BEAST-ROG, GURU-KALI (DESKTOP-0O8A1RL retired); howard = ACG-TECH03L, Howard-Home.

## Commands & Outputs

- Forensic cross-check (read-only): `git log --all --format=... | python3` mapping machine-in-subject → owner vs author → **0** author≠owner across 302 machine-named commits; 7 distinct author identities, all reconcilable to Mike or Howard.
- Session-log scan: 180 logs, 128 with User/Machine blocks, **0** user≠machine-owner mismatches; 52 pre-protocol (unattributed, not misattributed).
- `whoami-block.sh` verified: happy path (Mike/GURU-KALI/admin), malformed identity (UNREADABLE fallback, exit 0), missing identity (UNKNOWN fallback, exit 0).
- `reconcile_git_identity` verified in scratch repo: corrects on drift, silent no-op when matching, no clobber on empty args, `unknown`-sentinel guard preserves existing config.

## Pending / Incomplete Tasks

- **Broader memory-staleness sweep** — offered, not yet done. The attribution-scoped items are fixed; a general pass over the other ~50 memory files (e.g. `feedback_bypass_permissions_setting.md`, machine-setup notes) is a separate task awaiting go-ahead.
- **Code Review re-approval** — the review verdict was "Requires Revision"; all checklist items were fixed + verified but not formally re-approved by the agent (SendMessage unavailable). Re-run a review if a formal sign-off is wanted.
- **GuruRMM BUG-005 (mac build)** — still open, awaiting Mike's product decision (ship mac agents vs defer + make the audit pipeline treat a stubbed platform as N/A).

## Reference Information

- Code Review Agent id (this session): `a2b8c4d44c3c72b5d`.
- Coord confirmation message id: `dfeb6f2a-2d90-4bfe-bf95-a14eec449b3d`.
- New rule memory: `.claude/memory/feedback_attribution_from_identity.md`.
- Related memory: `feedback_identity_precedence.md` (identity.json beats the userEmail hint).
- Scripts: `.claude/scripts/whoami-block.sh`, `.claude/scripts/sync.sh` (`reconcile_git_identity`).

## Update: ~19:30 MST — broader memory-staleness sweep

Audited all 60 files in `.claude/memory/` (delegated the read+cross-reference to an agent against users.json, CLAUDE.md, infra_office_network, and live coord). 44 CURRENT, 16 flagged. Triaged into safe-fix / stale-but-uncertain / needs-real-world-check and applied fixes (delegated, then verified link integrity).

**Fixed (content corrections):**
- `radio_show_no_cohost_named_tom.md` — reattributed "the 5070 Ti session" to "Mike's session on GURU-5070 (then acg-guru-5070)"; linked the attribution rule.
- `feedback_1password_service_token.md`, `feedback_complete_vault_operations_end_to_end.md` — de-hardcoded `/c/Users/guru/vault` and `D:/vault` to the identity.json `vault_path` / `vault.sh` wrapper pattern (CLAUDE.md forbids hardcoded vault paths).
- `project_dataforth_incident_2026-03-27.md` — added `[RESOLVED]` banner (CA enforced 2026-04-04).

**Reframed + caveated (machine certain, underlying fact unverifiable):**
- `reference_ix_server_ssh.md`, `feedback_d2testnas_ssh.md` — CachyOS/acg-guru-5070 → GURU-5070 (now Windows 11); kept credential facts, added `[VERIFY]` caveats since the OS reinstall may have changed SSH key state.
- `reference_pluto_build_server.md` — flagged the `guru@DESKTOP-0O8A1RL` authorized-key (retired machine) for rotation/verify.
- `feedback_ollama_tier0_routing.md` — fixed `HOWARD-HOME` → `Howard-Home`; added `[DISCREPANCY]` caveat (note's `100.92.127.64` vs CLAUDE.md's canonical GURU-BEAST-ROG `100.101.122.4`).
- `project_mac_gururmm_setup_pending.md`, `project-cascades-migration-plan.md`, `project_neptune_sbr_email_routing.md` — added `[STALE]`/`[VERIFY]` caveats on aged open-items / machine-bound plan paths.

**Deduped:** deleted `reference_neptune_access_d2testnas.md` (covered by infra_office_network) and `project_email_routing_neptune.md` (folded its unique note into `project_neptune_sbr_email_routing.md`); updated MEMORY.md index. No dangling links to deleted files.

**Needs Mike's real-world check (left caveated in-place, not guessed):**
- Ollama Tailscale fallback — is it Howard-Home @ 100.92.127.64 or GURU-BEAST-ROG @ 100.101.122.4 (CLAUDE.md)?
- Has `install-hooks.sh` run on the Mac? (if yes, delete `project_mac_gururmm_setup_pending.md`)
- Neptune 2026-03-22 overnight TODOs and Pluto key rotation — still open?

## Update: 20:00 MST — Ollama per-machine fallback + Pluto/Neptune key evaluation

Resolved the three items Mike flagged from the memory sweep.

**Session Summary**

Mike directed three follow-ups: (1) Ollama fallback is a per-machine user choice — store it in identity.json; (2) ask the Mac about the pending hooks setup; (3) evaluate the Pluto/Neptune SSH key-rotation question, noting both run the GuruRMM agent.

For Ollama, added `ollama_fallback` to GURU-KALI's identity.json (GURU-BEAST-ROG @ 100.101.122.4, Mike's choice) and rewrote the fallback resolution in three places — `feedback_ollama_tier0_routing.md`, `.claude/OLLAMA.md`, and the CLAUDE.md Ollama table — to read the per-machine endpoint from identity.json instead of hardcoding an IP. This removed the [DISCREPANCY] caveat (the memory's 100.92.127.64/Howard-Home value was the lone outlier; CLAUDE.md/OLLAMA.md already used Beast).

For the Pluto/Neptune keys, used the GuruRMM API from GURU-KALI (reachable) to dispatch a read-only PowerShell command to both agents enumerating authorized_keys comments. Findings resolved the question factually: Pluto's old `guru@DESKTOP-0O8A1RL` key is already rotated out (current keys are the build server's: `gururmm-build@gururmm-server`, `guru@gururmm-build`); Neptune has no SSH authorized_keys at all. Updated reference_pluto_build_server.md to the verified state. Sent a coord message to the Mac re: install-hooks.sh (no reply yet).

**Key Decisions**

- Ollama fallback made a per-machine identity.json field (not a fleet-wide constant) per Mike's "user can choose" directive; resolver degrades to local-only when unset, so other machines are safe until they set their own.
- Evaluated the key question via RMM read-only dispatch rather than guessing or SSH-ing — both targets run the agent, and reading authorized_keys is non-destructive.
- Did NOT add GURU-5070's pubkey to Pluto or rotate anything — Mike said "evaluate," and key writes need explicit confirmation. Surfaced it as a decision instead.

**Configuration Changes**

- `.claude/identity.json` (gitignored, not synced) — added `ollama_fallback: {host: GURU-BEAST-ROG, endpoint: http://100.101.122.4:11434}`.
- `.claude/memory/feedback_ollama_tier0_routing.md` — resolver now reads identity.json `ollama_fallback`; removed [DISCREPANCY] caveat.
- `.claude/OLLAMA.md` — Endpoints resolver + rationale rewritten to per-machine identity.json; fixed retired DESKTOP-0O8A1RL / HOWARD-HOME references.
- `.claude/CLAUDE.md` — Ollama table + fallback note now point at identity.json `ollama_fallback`.
- `.claude/memory/reference_pluto_build_server.md` — authorized-key line replaced with RMM-verified current keys; noted GURU-5070 key not yet authorized.

**Infrastructure & Servers**

- GuruRMM API: `http://172.16.3.30:3001`, auth via SOPS `infrastructure/gururmm-server.sops.yaml` (claude-api@azcomputerguru.com). Reachable from GURU-KALI.
- Agent IDs: PLUTO `5316f56f-a1b3-4ac5-97ac-71ddf6a74d2e` (172.16.3.36), NEPTUNE `7d4f823c-f23d-40b8-ae72-b83cd2ccb09d`. 60 agents enrolled total.
- Pluto authorized keys (verified 2026-05-26): `gururmm-build@gururmm-server`, `guru@gururmm-build` in both `C:\ProgramData\ssh\administrators_authorized_keys` and `Administrator\.ssh\authorized_keys`. Neptune: none.

**Commands & Outputs**

- RMM execute pattern: `POST /api/agents/:id/command` (command_type=powershell), poll `GET /api/commands/:id`. zsh gotcha: `status` is a read-only var — ran the dispatcher under `bash` and used `st` instead.
- Coord message to Mac: id `625d80a7-6689-4aef-87a6-b6aa46e26dd9`.

**Pending / Incomplete Tasks**

- Mac `install-hooks.sh` status — awaiting Mac coord reply; delete `project_mac_gururmm_setup_pending.md` once confirmed.
- Decision: add GURU-5070's pubkey to Pluto so the documented workstation-SSH workflow works? (write op, needs confirmation.)
- Decision: coord-broadcast a "set your `ollama_fallback`" note to other machines, or leave each to choose?

**Reference Information**

- Commits this session: attribution hardening `4e97e20`, memory sweep `d3f3d28`, Ollama/Pluto `8bdb919`.
- New script: `.claude/scripts/whoami-block.sh`. New rule memory: `feedback_attribution_from_identity.md`. GURU-5070-added memory: `feedback_rmm_dev_is_mike.md`.