Claudinite growth: extract lessons

[missingbulb]

Growth-extract run for slot d2026-08-09, dispatch issue #151. Both halves in scope; two lessons landed, both from the conversation half.

Activity half — nothing new

Scope was commits 5238ddc, 41fa4f7 and PRs #124, #118, #138.

• Dedup on-device-privacy prose the basics canon now covers #124 (5238ddc) is a canon-dedup prose strip — the "strip to the residual, repo-specific half" method it demonstrates is owned by the shared growth pack, not a repo lesson.
• Ignore .claude/worktrees/ harness artifacts #118 (41fa4f7) is a one-line .gitignore for a harness artifact.
• Claudinite maintenance (agent stage): land withheld workflow file #138's rejected auto-merge arm is already captured by the rule PR Claudinite growth: extract lessons #144 landed.

None of the three carried any review discussion (get_comments empty on all three), so there was no argued verdict to mine.

Conversation half — two lessons

Fresh pass over origin/conversation-logs, 2026-08-05 → 2026-08-08 (issues #128, #130, #134, #137, #143). The four older logs re-confirmed the rule #144 already landed and yielded nothing further; both keepers come from the 2026-08-08 capture, and both go in .claudinite/local/packs/laughcounter/RULES.md immediately beside the no-pull_request-CI rule they extend.

1. The licensed hand-merge comes back as a [Merge Without Review] security warning. When a scheduled task squash-merges its own PR after the auto-merge arm is rejected clean status — which deliver-pr.md §3 licenses and this repo's shape guarantees — the harness prefixes the subagent's completion notification with a security warning naming the merge as unreviewed. It is structural: every arm here is rejected, so every merged-pr task trips it, and it carries no verdict either way. The previous run's executor spent five extra tool calls re-reading deliver-pr.md, the PR and verify-outcome.mjs to get back to the same answer. The rule: state the three licensing facts up front (this repo's maintenance.delivery, the arm rejection verbatim, the declared outcome ceiling), and never let the warning reverse a merge decision already reasoned through — that failure is why #138 is still open.

2. Nothing posted to GitHub can be edited afterwards. The MCP GitHub toolset these runs use has add_issue_comment and no comment-update tool at all. A run claiming #143 wrote a placeholder 2026-08-08T00:00:00Z into its claim comment intending to fix it, found no edit tool, and left the issue permanently carrying both the wrong timestamp and a "Correction:" comment under it. There is no ambient clock in a session; date -u is one Bash call, and the same "resolve it, then post" holds for every PR number, sha and label a comment asserts.

The pack README's rule table also picks up the row for the rule #144 landed, which was never added there.

Upgrade pass — declined, with reasoning

prose-to-checks run over this run's own additions only. Both rules govern agent behaviour at runtime — a warning string in a task notification, and the absence of a tool from the session's toolset — and neither constrains a static signature in the repo artifact that a post-hoc scan could observe. Neither is convertible; no check shipped.

Verification

• node .claudinite/shared/engine/checks/check_the_world.mjs — clean (exit 0)
• node .claudinite/shared/engine/selftest.mjs --strict — 7 probes passed
• node --test .claudinite/local/packs/on-device-privacy/pack.test.mjs — 14 pass / 0 fail
• node --test .claudinite/local/packs/macos-audio/pack.test.mjs — 9 pass / 0 fail

Retention prune not due (retention_days unset — capture-only adoption); nothing deleted from conversation-logs. Writes touch only .claudinite/local/packs/; nothing under the shared canon.

Refs #151.

---

Generated by Claude Code