Claudinite tracker: Growth Extract

[missingbulb]

Standing log for the Claudinite fleet growth-extract-new-instructions daily routine on this repo — it reviews the recent activity window and folds any durable, reusable lesson into the project's own instruction docs via an auto-merging PR. Per-run history lives in dated comments below; a run that finds nothing adds nothing and logs nothing. Its open/closed state carries no meaning — only the log does; never open/close/reopen it for signalling.

Method: .claudinite/shared/packs/grow_with_claudinite/extract.md.

(Supersedes the earlier untitled tracker #21, which used the pre-standard title "Growth: extract new instructions (daily)". New entries land here under the standard title.)

[missingbulb]

2026-07-23 — Added 1 lesson.

• Where: dev/procedures/ci-release.md — new section "Signing & notarization is an optional, secret-gated lane (macOS Gatekeeper: fix install docs + add optional signing/notarization #4)".
• What: the CI-side gotchas from the Developer ID signing + Apple notarization work (macOS Gatekeeper: fix install docs + add optional signing/notarization #4), which were only recorded in the DMG workflow YAML comments — notarization's Hardened-Runtime → mic-entitlement coupling (com.apple.security.device.audio-input), ad-hoc-can't-be-notarized, the ephemeral signing keychain needing to be added to the searchable list (its sed pipe being why ci: run DMG workflow steps with pipefail #27 added shell: bash/pipefail), keeping the lane secret-gated with an ad-hoc fallback, and macOS 15 Sequoia removing the right-click → Open Gatekeeper bypass.
• Landed via: PR Claudinite growth: extract lessons #28 (squash-merged to main).
• Deliberately skipped: the ci: run DMG workflow steps with pipefail #27 pipefail lesson itself — already a canon blocking check (gha/run-pipefail) + git-github-advanced skill prose, so a local copy would duplicate the canon.

---

Generated by Claude Code

[missingbulb]

2026-07-26 — Nothing added. No PR opened.

Window (dispatch #38): commits 03b9bd0, 3b7bfaf, 24e341d, ded4a03, 90a2a71, beda6cb; PR #34; issues #33, #23, #26, #25.

Everything durable in the window was already captured, in-window or by the canon:

• beda6cb (fix(mac): release the mic on every path — exit, signals, sleep races + keep-awake toggle #22, mic-release / sleep races / keep-awake) — the same commit wrote dev/procedures/mac-audio-lifecycle.md, which already carries the release-the-mic-on-every-path invariant, the no-signal-handlers exit gap, the restartGeneration guard on deferred work, and the config-change reconcile rule. The one thing not in that doc — the keep-awake power assertion riding refreshTitle() as the listening-transition chokepoint — is already a comment at its own call site (mac/Sources/LaughCounter/AppDelegate.swift:304), i.e. the file-local rule is satisfied; central prose would only duplicate it.
• ded4a03 (macOS Gatekeeper: fix install docs + add optional signing/notarization #4) + 24e341d (ci: run DMG workflow steps with pipefail #27) + 3b7bfaf (Claudinite growth: extract lessons #28) — captured by the 2026-07-23 run into dev/procedures/ci-release.md (Signing & notarization is an optional, secret-gated lane). Issue [needs-human] Enable Developer ID signing + notarization for the DMG #26 is the human enrollment checklist for that same lane and adds no new gotcha; pipefail stays canon-owned (gha/run-pipefail).
• 90a2a71 (Claudinite: flip to the vendored mount (canon phase 2) #24) + 03b9bd0 (Cut over to the Claudinite per-repo task scheduler #32) — Claudinite plumbing. The hashed-cron-minute / converge-wiring.mjs shape is canon (basics/scheduled-tasks.md + the scheduler-workflow-shape check), and the "prove the check is live by breaking its input" discipline is canon too (basics/skills/writing-tests). Local prose would restate a check's own failure message.
• PR Drop vendored-canon path from CLAUDE.md orientation header #34 (open) — the claudinite-isolation hit on CLAUDE.md:4 is a canon-side rule/description discrepancy, already raised as Implement per-project maintenance scheduling (owner-orchestrated rollout) Claudinite#394 and already enforced by a check. Not a LaughCounter lesson.
• Claudinite check: missing pip-install hint in robust.py's optional-import re-raise #23 (advisory python-optional-import-install-hint, still an open judgment call), Re-paste the Claudinite environment Setup script #25 (one-off out-of-repo Setup-script re-paste), Label seed: ready-for-agent (executor Routine wiring) #33 (label-seed issue kept open so the CCR trigger picker keeps offering ready-for-agent — fleet wiring, not this repo's territory).

Note for the owner, not a lesson from this window: this repo has no .claudinite/local/packs/ — its local lessons live in dev/procedures/, so the growth lifecycle's dedup/promote stages, which read only the local-packs subtree, can't see them. Adopting the pack structure (generate-project-instructions) would be a separate, deliberate change.

---

Generated by Claude Code

[missingbulb]

2026-07-30 run (dispatch #81) — no new lesson added.

Reviewed the window's 5 substantive commits and their merged PRs (#69, #77, #75, #73, #59; none had review discussion beyond the PR body), plus touched PR #78 and issue #79. Every commit in the window had already captured its own lesson at merge time, in the location this project's own dev/procedures/CLAUDE.md designates for it:

• on-device-privacy: convert "metadata persists, audio doesn't" to a check #69 (827f08d) — added the no-audio-persistence check directly to .claudinite/local/packs/on-device-privacy/ (rule + RULES.md + README table + red-first fixtures).
• Detect an absent mic via inputFormat, and keep checking instead of giving up #77 (7a20588) and Fix the v0.3.0 launch crash: attach the input node before prepare() #73 (61df205) and Fix the installTap crash on every audio device change, and reactivate the counter on return from standby #59 (7b5d417) — each appended its gotcha straight into dev/procedures/mac-audio-lifecycle.md, the doc this repo's own CLAUDE.md names as owning audio-lifecycle lessons.
• Show the app version in the menu header #75 (dd2c952) — the version-in-menu-header rationale is a settled implementation choice with nothing left to say but "why it looks like that"; that's already an in-code comment on versionLabel plus a mac/README.md user-facing note, exactly where task-doc guidance says it belongs (never local-pack prose).

PR #78 (open, unmerged) states its own lessons are already captured in mac-audio-lifecycle.md on its branch; out of scope until it merges. Issue #79 is the bug report driving #78, no separate action.

Checked against existing .claudinite/local/packs/laughcounter/RULES.md and on-device-privacy/RULES.md — nothing in the window is undocumented. No PR opened.

---

Generated by Claude Code

[missingbulb]

2026-07-31 — growth extract. Window: commits 9027934, 8152166, 00f3c86; PRs #84, #87, #78; issues #88, #23. Dispatched by #94.

Added — 2 lessons, both prose in .claudinite/local/packs/laughcounter/RULES.md (PR #98, auto-merge armed):

• "Not statically checkable" is a verdict about the tree's shape at the time, not about the rule — re-ask it after the tree changes. From on-device-privacy: convert "one deletable directory" to a check #84: the 2026-07-29 sweep left single-storage-directory's rule as prose because proving it looked like it needed data-flow tracking; that objection was sound for the many-entry-point Python reference and dissolved the moment that tree went away.
• Author a check as a positive whitelist over an enumerated API surface, not a ban list. Also on-device-privacy: convert "one deletable directory" to a check #84: matching the allowed literal catches an API case that doesn't exist yet, and makes a non-literal argument — the one thing that would defeat the check — a finding too.

README.md's enforcement table and the pack's ruleRoutingGuidance updated alongside. Neither lesson converts to a check (both are in-flight authoring judgment with no static signature).

Deduped out: everything from the mac-audio work (#78, #87, #88) — listening never meant audio arrived, HAL read vs. build-an-engine, systemUptime vs Date(), the per-client CADefaultDeviceAggregate, command -v swift not testing for a toolchain, and the explicit did-not-settle section — is already captured in dev/procedures/mac-audio-lifecycle.md. The version-bump-or-the-Release-silently-refreshes gotcha is already in mac/README.md. #23 is a moot advisory (its file left with the Python tree).

Green offline before opening: on-device-privacy/pack.test.mjs 11/11; run-active-pack-rules.mjs, check_the_world.mjs, check_the_work.mjs all exit 0.

---

Generated by Claude Code

[missingbulb]

Addendum to the 2026-07-31 entry above: arming auto-merge on #98 was refused — the repo has no pull-request-triggered CI (both DMG workflows are push + paths: mac/**, which this pack-only diff doesn't touch), so GitHub reported the PR already mergeable and auto-merge had nothing to wait on. Per the task's no-CI fallback ("GitHub lands it as soon as it's mergeable") it was squash-merged directly instead — a186acf, no human review, same outcome auto-merge would have produced immediately.

---

Generated by Claude Code

[missingbulb]

2026-08-01 — window: commits 4ce1d97, b6bfe51, 5f20c3a; PRs #101, #100, #99 (merged), #96 (touched).

Added: one lesson, as prose in .claudinite/local/packs/laughcounter/RULES.md (+ a row in that pack's README.md table) — "Already covered by another check" is a claim to test, not to reason out: hand the sibling check a violating file before dropping a rule as a duplicate. From #99, where three sweeps in a row left "the app listens on no socket" as prose on the reasoning that an NWListener needs import Network and no-network-client bans that — true of one API, false of the capability (socket(2) via import Foundation, NSXPCListener/CFSocket*/NSSocketPort, an embedded HTTP-server import), and a file containing only socket(AF_INET, ...) passed check_the_world. The listener surface itself already lives at the site in no-listener.mjs, so only the dedupe discipline was captured. Prose rung, not a check: it's a judgment rule about how sweeps dedupe, not a deterministic one.

Not added: nothing from #100 / #101 — dev/procedures/mac-audio-lifecycle.md already carries the aggregate-churn ladder, three-state health reporting and the one-remedy UI rule, the witnessed-arrival settle rule, and "seed timer-maintained state at start". #96 is the machine-written usage fold.

Landed as PR #106 (check_the_world exit 0, on-device-privacy fixtures 14/14). Auto-merge could not be armed — the repo runs no PR-triggered workflows, so GitHub reported the PR already clean and refused; squash-merged directly, which is the same end state.

---

Generated by Claude Code

[missingbulb]

2026-08-02 — 3 lessons added (dispatch #113, PR #123)

All three into .claudinite/local/packs/laughcounter/ (RULES.md prose + README table row); nothing under .claudinite/shared/.

1. Ship a version the running app can state, and bump it per distinguishable build — from Show the app version in the menu, so you can tell which build is installed #74/Show the app version in the menu header #75. A merge that leaves CFBundleShortVersionString alone refreshes the same Release, so distinct builds shared a version string and "which binary is installed?" had to be answered from crash-report symbol names. Show version and build from Bundle.main, not a source constant.
2. A scheduled job that skips its work still exits 0 — from Claudinite: refresh the vendored mount to canon 7e89c0c #56. The scheduler ran green nightly while silently skipping baselining ("no vendored mount (no stamp)"). Read the skip line, not the exit code; and a bug inside the self-updating mechanism has to be fixed out of band.
3. Delete the flagged coupling before reaching for an accept — from Drop vendored-canon path from CLAUDE.md orientation header #34. An accept is for a crossing that must exist; the claudinite-isolation finding on CLAUDE.md pointed at text carrying nothing a reader could act on.

Deduped out: the whole mac window (#59, #73, #77, #78, #87, #100, #101, #108) is already in dev/procedures/mac-audio-lifecycle.md; #55/#69/#84/#99 in this pack's existing prose and the on-device-privacy checks. Conversation-logs half (2026-08-01 logs) yielded nothing new. No retention_days set, so no prune and no log deleted.

Verified: check_the_world.mjs exit 0, on-device-privacy fixtures 14/14.

Auto-merge could not be armed — GitHub rejected it because the PR is already in clean status (the change touches no path any workflow watches, so no check runs were created). The PR is left open and mergeable.

---

Generated by Claude Code

[missingbulb]

2026-08-08 — one lesson landed (dispatch #143, slot d2026-08-08)

Both halves in scope. PR #144 — merged (squash, b8c250b).

Added: one rule in .claudinite/local/packs/laughcounter/RULES.md —

Nothing in this repo runs on pull_request, so most PRs carry no checks at all — and arming auto-merge on one is rejected every time.

Verified against all three workflow files: build-macos-dmg.yml / release-macos-dmg.yml trigger on push scoped to mac/**, claudinite-scheduler.yml on schedule/workflow_dispatch; no pull_request trigger exists. The rule carries the two corollaries that actually cost time — the Action's maintenance PR is already merged by the time the agent stage starts (check merged main first, don't hunt the PR API), and "CI will catch it" is false outside mac/**.

Activity half: 5238ddc/#124 and 41fa4f7/#118 yielded nothing (a canon-dedup prose strip whose method the shared pack owns; a one-line .gitignore for a harness artifact). #138 yielded the keeper — its enable_pr_auto_merge was rejected "already in clean status" and the run escalated to needs-human instead of taking the squash merge deliver-pr.md §3 licenses. #138 is still open and unmerged — outside this run's scope, flagged for whoever picks it up.

Conversation half: fresh pass over the four logs of 2026-08-05 → 2026-08-07 (#128, #130, #134, #137). Same finding, recurring across three consecutive baselining cycles. Provenance summaries posted on all four issues.

Checks converted: none. The addition records a fact about workflow shape that steers a delivery decision, not a condition to enforce — a check firing on a new pull_request trigger would be backwards.

Retention prune: not due (retention_days unset — capture-only adoption). Nothing deleted.

Verification: check_the_world clean, selftest --strict 7/7, both local pack test suites pass.

Delivery: maintenance.delivery is auto-merge; the arm was rejected "already in clean status" — the very shape this run's rule describes — so per deliver-pr.md §3 the PR was squash-merged directly (0 check runs on the head sha, nothing pending or failing).

---

Generated by Claude Code