Claudinite growth: conversation extract

[missingbulb]

Scheduled grow_with_claudinite/conversation-extract run (dispatch #764).

Queue: 4 logs on conversation-logs. The issue-717 log was already mined twice (#732, #746/#747) and re-read this pass with nothing new. The three 2026-07-26 logs (issue-734, issue-753, issue-760) are fresh — never extracted.

Landed: 2 rules, both as bullets in .claudinite/local/packs/gcec/RULES.md → Working rules. Prose is the strongest available mechanism for each: neither condition is observable from a working tree, so no local check can carry them, and neither is an activity-scoped procedure that a skill would gate.

1. A silent SessionStart-hook miss (from the issue-760 log)

That session's context held only CLAUDE.md, the date, and the owner's email — the Claudinite injection (owner preferences + pack prose) never landed, and no Stop-hook feedback arrived between turns either, so the whole hook config was inert. Consequence: the preference making "LGTM" the merge-to-main command was absent, so at 13:25:47Z the session answered lgtm with "say the word if you'd like me to merge it" instead of merging. The owner had to ask "did you load my preferences?" at 13:26:22Z; recovery, merge included, ran to 13:27:51Z.

The rule that would have prevented it was already written down — it just wasn't in context. So the durable fix isn't the rule, it's the detection + one-call recovery: the tell (only CLAUDE.md present; no Stop-hook feedback) and bash .claudinite/shared/engine/hooks/session-start-command.sh, which is exactly what that session used to recover.

Routing caveat, for the reviewer / the promote stage: this bullet lives in prose that the very hook it describes is responsible for injecting, so it is unreadable in precisely the session that needs it. The genuinely strongest placements — hook reliability upstream, or a line in CLAUDE.md (the one surface confirmed to survive the miss) — are both outside this task's write surface (.claudinite/local/packs/ only). Captured here so it is recorded and promotable, flagged rather than silently mis-placed.

2. git checkout <old-branch> -- <paths> destroys the uncommitted edits it was meant to carry (from the issue-734 log)

At 06:39:31Z the run tried to relocate finished-but-uncommitted work onto a fresh branch in one chain:

git fetch origin main --quiet && git checkout -b claude/growth-dedup/734-a1f3c2 origin/main --quiet \
  && git checkout claude/charming-einstein-rmaeom -- <two paths>

The third clause restores the paths' committed content from the old branch — i.e. the pre-edit version — so it wiped the three edits it was added to preserve. Detected at 06:39:39Z, re-applied by 06:39:54Z. The larger cost is downstream: check_the_world.mjs + npm run test:offline (308 tests) had gone green at 06:38:57Z against the now-destroyed tree, so the whole verification pass was invalidated and re-run at 06:39:55Z before the 06:40:12Z commit.

git checkout -b <new> origin/main already carries a clean working tree; the "carry the paths across" clause is the trap.

Dedupe

Both were greped against the mounted canon (.claudinite/shared/, stamp 61b90ee) and the gcec pack:

• No canon or local coverage of a silently-inert SessionStart hook, or of verifying that injection landed — the canon's inject-preferences.sh is deliberately fail-soft on a miss, and nothing tells a session how to notice.
• git-github-advanced covers adjacent-but-different ground (don't commit environment drift a stop-hook nags about; give scheduled runs unique branch names). Nothing covers the destroy-your-own-edits path-checkout. This one is portable — captured locally per the method, for promote to lift.

Considered and rejected (below the bar / already covered)

• validate-dispatch.mjs has no CLI entrypoint — two separate executor sessions (issue-734, issue-753) each hand-wrote an ad-hoc Node wrapper to run it. Real recurring cost, but it's a canon-engine gap to fix upstream, not a local rule; capturing a workaround here would outlive the fix.
• The 81-minute idle wait in the issue-753 log (07:14:04Z executor done → 08:35:27Z owner's "merge to main") and the analogous 63-minute wait already measured in the issue-717 log — same finding, already reported in Claudinite growth: conversation extract #747: reclaiming it is an auto-merge delivery-policy call for the owner, not something an agent rule can encode.
• The stale-origin/main spurious squash-merge-history finding, hit again in the issue-734 log — already captured verbatim as a gcec gotcha by Claudinite growth: conversation extract #747, and since pruned by Claudinite growth dedup: prune two gcec rules the canon now covers #754 as canon-covered. Not re-added.
• The issue-717 log's re-read — nothing beyond what Claudinite growth: conversation extract #732 and Claudinite growth: conversation extract #746 already landed.

Verification

check_the_world.mjs: 0 blocking, 11 advisory (all pre-existing, unchanged by this run). Docs-only change — no code or test surface touched. Nothing under .claudinite/shared/ (the read-only canon) was edited; the diff is one file.

Refs #764

🤖 Generated with Claude Code

https://claude.ai/code/session_01FaX36U1zAVmLqLNXLdAPWb

---

Generated by Claude Code