claudetools/session-logs/2026-06/2026-06-08-mike-harness-optimization-selfcheck.md

Session Log — 2026-06-08 — Harness Optimization Completion + GrepAI/Wiki + Self-Check Smoke Tests

User

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

Session Summary

Continued and completed the ClaudeTools harness-optimization spec (specs/claudetools-harness-optimization/) that was mid-flight at the start of the session. Picked up at P1 Task 5 (skill-registry description trims, already applied but uncommitted) and drove the spec to functional completion across P1/P2/P3 plus follow-on work. The through-line goal was Mike's stated priority: accuracy + completeness under context pressure, with token efficiency secondary.

Finished the P1 context diet: committed the 8 trimmed skill descriptions (registry injection ~3320 -> ~2123 tokens, ~36% cut, keyword triggers preserved, frontmatter validated); created the now-phoenix.sh helper (fixed UTC-7 epoch math, replacing the unreliable TZ=America/Phoenix date that silently returns UTC on Git-Bash); confirmed post-bot-alert.sh already builds JSON via jq -nc --arg (no change needed); thinned the /save and /sync command bodies to point at sync.sh as the single source instead of re-documenting its internals (Task 7). Implemented P3 Task 10 knowledge-tiering as a FORWARD convention: new session logs go in session-logs/YYYY-MM/ month folders (recall = scoped grep, no monolithic index), existing flat logs left untouched since grep covers both layouts and a mass move would break wiki source_logs refs. Verified the nested path doesn't break the Phase 0 log-detection grep or wiki slug derivation. Bumped VERSION to 1.4.0, updated CHANGELOG, committed (68ad1dbd), broadcast to the fleet via coord + filed a backstop todo.

Answered Mike's side question on whether GrepAI is still earning its keep given the wiki. Pulled real data (57 wiki articles vs 343 logs + 538 Rust + 7742 TS/TSX indexed; watcher Ready; GrepAI mentioned in 18 of 343 logs) and concluded: not redundant — its irreplaceable value is code call-graph tracing over ~8.3k files the prose wiki can't see — but it was mis-positioned. The real bug was a contradiction: CORE said "wiki first" while EXTENDED + a standard + CODING_GUIDELINES all said "GrepAI first for any context lookup." Mike chose "Fix positioning." Reconciled all four sources to one hierarchy: wiki (known entity) -> GrepAI (code/discovery/un-compiled detail) -> raw file reads. Committed e8a689b0 + d4d24b5a.

Built Task 12 (the self-check half): added a harness category to /self-check enforcing the 1.4.0 invariants (VERSION >= min_version; skill-registry description budget; global deploy targets ~/.claude/skills+commands populated — the Mac-wipe failure; guard wired into sync.sh; core scripts parse via bash -n; now-phoenix emits a valid date). All read-only/non-invasive. Added tunables to the self-check baseline manifest harness block. Verified 9/9 PASS; negative-tested the budget WARN. VERSION 1.4.1, committed edcbc5f7.

Finished with the Task 3 leftover: the command-restates-standard lint. Added a consistency category to /self-check that checks each manifest-declared command<->standard pair still carries its defer-to-SSOT pointer (catches the Syncro-timers drift mode: standard said "always timer" while /syncro said "outlier only"). Deterministic core + a model-delegated semantic contradiction pass, mirroring check_memory. Seeded with the syncro-billing pair. Verified PASS + negative-tested. VERSION 1.4.2, committed e180a463. Filed a coord todo (f1c11d0d) to revisit the guard-FATAL promotion on/after 2026-06-22 once the warn window is clean fleet-wide.

Key Decisions

• Task 10 = forward convention, not mass migration. New logs go in session-logs/YYYY-MM/; the ~283 existing flat logs stay put. Rationale: grep recall covers both layouts, and mass-moving would break wiki source_logs references for low context-pressure payoff. Verified nested paths don't break the Phase 0 grep or slug derivation (the month segment lands after session-logs/, so <project>/<slug> capture is unaffected).
• GrepAI kept but demoted, not deprecated. Its code call-graph tracing over ~8.3k files is something the 57-article prose wiki fundamentally cannot do. The problem was positioning (4 sources said "GrepAI first" contradicting CORE's "wiki first"), not the tool. New order: wiki -> GrepAI (code/discovery) -> raw reads.
• Honest about missing telemetry. Declined to claim a hard "GrepAI saves X tokens" number — there is no query telemetry today. Offered to instrument it if the number ever matters, rather than asserting.
• Self-check additions are strictly read-only. No commits/pushes/dry-run-that-mutates; the smoke tests use bash -n (parse only) + a clock read. Keeps the script's read-only census contract intact.
• Registry budget set at post-trim size + ~20% headroom (10500 chars). Current total 8686 chars (~2171 tokens) across 18 skills. A WARN means a description regrew — the metric that directly guards Mike's stated concern (Task 5 bloating back).
• command<->standard lint = deterministic pointer check + delegated semantic pass. A fully-general contradiction detector is semantic (model work), so the deterministic half only verifies the defer-to-SSOT pointer survives; the judgement of actual contradiction is delegated in SKILL.md, exactly mirroring the existing memory contradiction pattern.
• Patch-version bumps per additive change (1.4.0 -> 1.4.1 -> 1.4.2). Each fleet-visible self-check enhancement got its own VERSION bump + CHANGELOG entry; min_version stayed 1.4.0 so machines >= 1.4.0 pass.
• Guard-FATAL promotion deferred via durable todo, not memory. Warn window opened today; filed coord todo f1c11d0d dated 2026-06-22 so it surfaces durably rather than relying on recall.

Problems Encountered

• guru-connect submodule gitlink showed modified. Confirmed it is a registered submodule, staged only harness files explicitly, and let sync.sh's Task-1 logic unstage the gitlink — commits carried no submodule bump (the submodule-safe sync working as designed).
• Transient duplicate-divergence WARNs in self-check. After editing repo copies of self-check/remediation-tool SKILL.md, the duplicates check flagged repo vs ~/.claude divergence. Identified as transient (resolves on next /sync Phase 5c skill redeploy), not a regression; the git post-commit WARN is pre-existing on this box. New harness+consistency checks were all-PASS.
• Cosmetic regex-in-output. The consistency PASS message initially echoed the raw must_reference regex (showing escaped backslashes). Simplified the message to "SSOT pointer present" rather than printing the regex.

Configuration Changes

Modified:

• .claude/skills/{gc-audit,human-flow,impeccable,mailprotector,memory-dream,packetdial,remediation-tool,self-check}/SKILL.md — one-line description: trims (Task 5; committed this session).
• .claude/commands/save.md — Phase 4 thinned to point at sync.sh; location table updated to session-logs/YYYY-MM/ convention (Task 7 + Task 10).
• .claude/commands/sync.md — "What this does" compressed to point at sync.sh; exit-75 semantics kept (Task 7).
• .claude/CLAUDE_EXTENDED.md — GrepAI section rewritten: wiki-first recall hierarchy.
• .claude/CODING_GUIDELINES.md — "Context Lookup — GrepAI First" -> wiki-first for known entities.
• .claude/standards/context-lookup/grepai-first.md — title + body + frontmatter reconciled to wiki-first; lookup-table rows split known-entity (wiki) vs code (GrepAI).
• .claude/skills/self-check/scripts/self-check.sh — added check_harness_smoke() + check_command_standard() + ver_ge() helper; wired both into the run list.
• .claude/skills/self-check/baseline/manifest.json — added harness block (min_version, version_file, registry_desc_budget_chars, syntax_check_scripts, guard_wired_in) + command_standard_links block.
• .claude/skills/self-check/SKILL.md — documented harness + consistency categories + the command/standard semantic pass.
• .claude/harness/VERSION — 1.3.0 -> 1.4.0 -> 1.4.1 -> 1.4.2.
• .claude/harness/CHANGELOG.md — entries for 1.4.0, 1.4.1, 1.4.2.
• specs/claudetools-harness-optimization/plan.md — rollout-status block updated: P1/P2/P3 + Task 12 marked DONE; remaining ops follow-ups listed.

Created:

• .claude/scripts/now-phoenix.sh — deterministic Phoenix-time helper (fixed UTC-7 epoch math; --iso/--date/--datetime/--epoch/--fmt).
• session-logs/2026-06/2026-06-08-mike-harness-optimization-selfcheck.md — this log (first to use the Task 10 month-folder convention).

Credentials & Secrets

None discovered or created this session.

Infrastructure & Servers

• Coord API: http://172.16.3.30:8001/api/coord (broadcast + 2 todos filed).
• Gitea origin: http://172.16.3.20:3000/azcomputerguru/claudetools.git.
• GrepAI watcher: scheduled task "GrepAI Watcher - claudetools" (State: Ready). GrepAI binary at D:/claudetools/grepai (~22 MB).

Commands & Outputs

• bash .claude/scripts/now-phoenix.sh --date -> 2026-06-08 (15:07 UTC -> 08:07 PT, fixed -7h verified).
• Self-check harness category: 9/9 PASS (registry 8671/10500 chars; ~/.claude/skills 19; commands 25; guard wired; sync/guard/now-phoenix parse clean).
• Self-check consistency category: PASS (syncro-billing standard defers to owning command).
• Negative tests: budget WARN trips at synthetic 5000-char budget; consistency WARN trips when defer-to-SSOT pointer removed.
• Registry measurement: 18 skills, 8686 description chars (~2171 tokens).

Pending / Incomplete Tasks

• Guard-FATAL promotion — time-gated; coord todo f1c11d0d for on/after 2026-06-22. Re-check .claude/harness/guard.log clean across the fleet, then flip HARNESS_GUARD_FATAL default + bump VERSION + broadcast.
• Deferred by design: Task 8 (shard big command bodies — revisit only with a real lazy-load gate); full Python port of the core harness (its own future spec).
• Low-priority/judgment-call: schedule memory-dream --apply-safe per machine; optional later migration of existing flat logs into month folders.
• Mike-side parked (from earlier in the broader session, pre-compaction): Jupiter VM DHCP-vs-static; offline classification overrides (SIF-SERVER/Server2013); 3 MSP360 console todos; 3 RMM Thoughts; FunctionRail DRY refactor; Robert Wolkin Tailscale; consolidate 3 Wolkin client slugs.

Reference Information

• Spec: specs/claudetools-harness-optimization/plan.md.
• Commits this session: 68ad1dbd (P1+P2+P3 / VERSION 1.4.0), e8a689b0 (GrepAI demote in EXTENDED), d4d24b5a (GrepAI standard+guidelines reconcile), edcbc5f7 (self-check harness smoke tests / 1.4.1), e180a463 (consistency lint / 1.4.2).
• Coord: broadcast 41e80704 (harness 1.4.0 live); todos d1d5ad60 (fleet pull/sync 1.4.0), f1c11d0d (guard-FATAL re-check 2026-06-22).
• Harness VERSION: 1.4.2. Min-version baseline in self-check manifest: 1.4.0.
• Helper: .claude/scripts/now-phoenix.sh (use --date for log dates).