Repository navigation
feat: authoritative CHANGELOG.md action ledger — gate-on-write + reconcile-on-read (S2–S7) - #46
Merged
Conversation
…ledger gate (S2) Component A of the authoritative-CHANGELOG campaign (ratified plan docs/planning/changelog-authoritative-ledger-gate-plan.md, D1–D7): a write-time hard gate that forces every session to record its actions in the authoritative CHANGELOG.md ledger at close-out. - New failure mode FM KJ5HST#27 "Unrecorded action" (APPENDED; FMs 1–26 unchanged, not renumbered) + a paired Degradation-Detection row. - New first bullet in SESSION_RUNNER §3F and a rewrite of ITERATIVE_METHODOLOGY Phase 6 step 8: append one dated, source-tagged entry per action ([issue #<N>] · [BL-<N>] · [ad hoc]) before commit; remove a completed BACKLOG.md item in the same commit; self-provision the ledger from the bootstrap seed if absent (D3) — the only opt-out is a CLAUDE.md-recorded "no CHANGELOG" decision. - Keys on "authored/retained a commit OR took a non-commit action" (D2/D4), never "completed units only"; "too small to log" IS the failure mode, not an exception; anti-erosion clause ties it to FM KJ5HST#17. Learning KJ5HST#7 count reconciliation (same commit): "26 → 27 failure modes" at 8 live sites + CLAUDE.md renumber clause (1–25 → 1–26); "16 → 17 warning signs" in README. Dated history (§What's New, §Versioning, docs/planning) left verbatim. Verified via a 6-lens adversarial workflow (append-only, count-completeness, source-tag vocabulary, D2/D3/D4 compliance, cross-doc consistency, scope): 5 lenses clean, 1 confirmed defect (missed 16→17 warning-signs count) caught and fixed. Dogfood root CHANGELOG (Component E) is deferred to S6 per plan order. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Component B of the authoritative-CHANGELOG campaign (ratified plan docs/planning/changelog-authoritative-ledger-gate-plan.md, §3 Component B, D2/D3/D4) — the dependability keystone. The central finding was that a close-out write-gate ALONE is not dependable (is_dependable:false): a session that crashes before Phase 3F, an out-of-band commit, or an in-progress hand-off lands work in git with no ledger entry. Reconcile-on-read makes the ledger self-healing so it is true WHEN READ, not merely "sessions were asked to log." - SESSION_RUNNER Phase 0 step 6 now reconciles CHANGELOG.md against git log: frontier = `git log -1 --format=%H -- CHANGELOG.md`; gap = commits since with no entry; backfill each — during step 6, before the report and STOP — as its own `docs(changelog): backfill` commit. New ledger-reconcile note carries the mechanics; step 7 report surfaces the result. - Phase 1B stub gains a `CHANGELOG: pending` crash breadcrumb; the git-log gap is the always-reliable backstop (covers the 3D->3F window the marker cannot). - ITERATIVE_METHODOLOGY mirror: Pre-Flight step 4, purpose carve-out, gate, anti-pattern, Phase 1B step 1, Quality-Gate-1 question, Session Doc Template. - CLAUDE.md:80 cross-ref corrected (Phase 0 "step 5"->"step 6", now also reconciles) — Learning KJ5HST#7: fix the cited destination of the changed step. Decisions honored: D2 keys on any non-merge commit past the frontier (no "completed unit" filter); D3 self-provisions if no committed ledger exists; D4 keeps the honest division — reconcile backstops commit-bearing actions, non-commit actions (releases/tags/PR/issue/access) remain the FM KJ5HST#27 write-gate's. FM KJ5HST#17 anti-erosion clause: the backfill records history, is not the deliverable, licenses no extra scope. Verified via a 7-lens refute-default workflow. count-integrity and source-tag greppability CLEAN; 11 defects found and fixed, chiefly two ship-blockers: (1) an empty frontier (`CHANGELOG.md` never committed) collapsed `<frontier>..HEAD` to an empty range and silently no-op'd escape #4 — fixed by making committed-ledger existence the FIRST discriminator (reproduced across absent / uncommitted-seed / normal / current cases); (2) the backfill-write timing straddled the STOP gate — pinned to "during step 6, before the report and STOP." §5.2 reconciled: the backfill entry leads with a source tag (default [ad hoc]) so the audit grep still enumerates it, rather than the plan's un-greppable bracketed [backfilled] sketch. No FM / numbered gate / Phase-0 step added: FM count stays 27, warning signs 17, quality gates 12, Phase 0 = 8 steps, Pre-Flight = 7 steps — no count reconciliation owed. Dogfood root CHANGELOG (Component E) is S6; dashboard freshness (C) is S5. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…(S4) Component D of the authoritative-CHANGELOG campaign (ratified plan docs/planning/changelog-authoritative-ledger-gate-plan.md, §3 Component D + D6): the entry-format spec, the SESSION_NOTES division-of-labor, and the recompose of the starter-kit CHANGELOG seed from Keep-a-Changelog/[Unreleased] to the action-log shape. Both edited files are SEED disposition, so this reaches new adopters only; enforcement for existing adopters already lives in the synced §3F / FM KJ5HST#27 / Phase 0 text shipped by S2/S3. - starter-kit/CHANGELOG.md recomposed: authoritative-action-ledger framing; the ratified entry format (### header line = required greppable unit, three detail bullets recommended — a true superset of S2/S3's shipped single-line form, not a contradiction); the one verbatim source-tag vocabulary [issue #<N>] · [BL-<N>] · [ad hoc] (§5 must-fix #2), byte-identical to §3F/IM; the D2 keying rule (author or retain a commit OR take any non-commit action; NOT "completed units only"), an (in progress) marker for in-progress/reverted work (closes escape #3), and the sole empty-diff exemption; the CHANGELOG-vs-SESSION_NOTES division (transient handoff vs cumulative ledger, the commit SHA the only shared key, distill-not-copy); reverse- chron/prepend ordering with ## YYYY-MM time-grouping (not release-grouping). - Signal-D freshness sentinel (D6): a single greppable METHODOLOGY-SEED-SENTINEL token whose documented contract — token present AND zero real-date ### headers => fresh seed, suppress "never used" — is the anchor S5's dashboard will consume. The fresh seed carries zero real-date headers (the template uses a literal YYYY-MM-DD placeholder inside a fenced block), so it is correctly classified fresh, not stale. - starter-kit/SESSION_NOTES.md gains the bidirectional counterpart pointer (transient here; cumulative in CHANGELOG; the SHA the shared key; distill-not-copy). Decisions honored: D2 (keying + in-progress marker + empty-diff-only exemption); D6 (recompose now, stays SEED in bin/_manifest.py, sentinel consistent). Out of scope by plan order: Component C dashboard freshness is S5 (it consumes this sentinel); Component E root CHANGELOG + boundary header is S6; D7 release pointer lands with E. Loose "completed work history" captions in README/CLAUDE/BOOTSTRAP are left verbatim (the S2/S3 caption convention) — a deliberate, recorded deferral. Verified via a 6-lens refute-default workflow (vocabulary-verbatim, format-superset, Signal-D invariant, D2-keying, count/cross-doc, scope): 0 ship-blockers, 0 should-fixes; 2 nits applied — deduplicated the sentinel token so a naive S5 token-grep can't match a survivor line, and reworded the FM one-liner to mirror the mechanical keying. Deterministic gates green: vocabulary byte-identical to §3F/IM, zero real-date headers, sentinel single- occurrence, 51/51 bin/tests.sh. No FM / numbered gate / Phase step added: FM count stays 27, quality gates 12 — no count reconciliation owed. NEXT = S5 (Component C dashboard freshness). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Stop rewarding mere presence of a CHANGELOG. A new evaluate_changelog_freshness() grades whether the ledger still tracks the work, and feeds an advisory-only signal (a RISK line + at most a 1-point documentation nudge — never a hard fail; the authoritative gate lives in the session runner via FM KJ5HST#27 + Phase 0 reconcile). - Signal C: >=10 non-merge commits since the ledger was last committed. - Signal B: ledger frontier trails HEAD >21 days on an active repo. - Signal D: never-used seed (sentinel present + zero real dated entries) on a repo with real history — keyed on token AND absence of `### YYYY-MM-DD`, never token alone. - Signal F: done-marked BACKLOG.md items not migrated to the ledger. - New-adopter grace (<10 commits) keeps a freshly bootstrapped seed silent. - D3 absent-defect: absent CHANGELOG + real history is flagged only for adopters (SESSION_RUNNER.md present); non-adopters stay silent. Doc sub-score: the old flat "+2 has_changelog" splits into +1 present / +1 fresh; the 20-point cap is unchanged. Wired at the §5.5 anchor (after the metrics dict, before scores). DASHBOARD_VERSION 2.6.1 -> 2.7.0. Both twins byte-identical. Verified: 35 fixture assertions isolating each signal + grace + doc-split + D3 + negative controls; 51/51 bin/tests.sh; live collect_all on this repo and a full portfolio run rendering the risk line + version in terminal and HTML. Advisory — no principle, phase, gate, workstream, or FM change; FM count stays 27. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…g boundary (S6) Component E of the authoritative-CHANGELOG campaign (plan 1710e90, decisions D5/D7). Creates this repo's own root CHANGELOG.md — the per-action operational timeline, distinct from CLAUDE.md §Versioning's released-version narrative. - Bounded backfill (D5): 7 entries covering everything v3.0-forward, newest on top, no hole at the recent edge, no full reconstruction. All three source tags demonstrated: [issue KJ5HST#43] (v3.0 MIT relicense), [BL-4] (backlog item), [ad hoc] (releases/PRs/grooming). Every SHA/date/tag verified against git. - Self-logs this campaign with an (in progress) marker (D2 + FM KJ5HST#27 dogfood): the S6 commit records its own action. - Release entries are one-line pointers into §Versioning, never re-narrations (D7 / cite-don't-restate); CLAUDE.md §Versioning gains the reverse boundary pointer so the two ledgers cannot diverge. - Repo-scoped, not branch-scoped: fork-main-only backlog SHAs (ff5cee9, 69dad12, 72dc914) honestly annotated (fork main), per D5's explicit inclusion. - Upstream issue KJ5HST#43 cited as an absolute KJ5HST URL, never a bare #NN. Not in the sync manifest (canonical-only file). Adversarial verification (5 dimensions x refute) returned 0 confirmed defects; 51/51 bin/tests.sh. Next: S7 (D1 pre-commit hook + step added to other session-type checklists). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ion-type + campaign checklist (S7, escape KJ5HST#8) Closes escape KJ5HST#8 of the CHANGELOG-authoritative-ledger campaign: a session-type checklist that summarizes close-out silently exempts any gate it omits. Adds the CHANGELOG.md ledger step (Phase 3F, FM KJ5HST#27) to: - SESSION_RUNNER Planning Session Checklist + vertical-slice revert close-out - ITERATIVE_METHODOLOGY §Session Types: Review/Audit, Planning/Preparation, Debugging - all six campaign per-execution + per-campaign (consolidation) checklists - new Learning KJ5HST#8 codifying the rule (append-only; rows 1-7 unchanged) Vocabulary byte-identical; FM count unchanged at 27. Checkpoint commit (--no-verify): the campaign ledger entry lands with the close-out commit so the reconcile frontier tracks HEAD. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…GELOG-ledger campaign (S7, D1) Ships .githooks/pre-commit — the mechanical enforcement of FM KJ5HST#27 where a repo has no root SESSION_RUNNER to run the close-out gate (escape #5): it refuses a commit that changes tracked content unless CHANGELOG.md is co-staged. Deliberately narrow and bypassable: skips merges/rebases and absent-ledger repos, `git commit --no-verify` bypasses it, and Phase 0 reconcile-on-read (Component B) is the guarantee that backfills anything past it. Match is fixed-string (grep -qxF), not a regex. Docs: SAFEGUARDS Commit Discipline "Ledger Co-Staging Hook" subsection; BOOTSTRAP Step 10 setup paragraph + scoped the stale "does not ship its own hook script" line to blast-radius only; BOOTSTRAP First Session Checklist step 7 gains the ledger step (escape KJ5HST#8, same class). CHANGELOG.md: campaign S2-S7 marked complete. This commit is gated by the hook it ships and passes (CHANGELOG co-staged) — the dogfood. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Designates the CHANGELOG-ledger campaign (S2-S7) as minor release v3.1 (first new failure mode since v2.7): CLAUDE.md §Versioning gains the narrated v3.1 entry and current-version bump; README §What's New its public restatement; root CHANGELOG.md a one-line pointer into §Versioning (decision D7 triple-write). FM count 26 -> 27; 6 phases / 12 gates / four session types / workstreams unchanged. Tag + GitHub Release cut at merge. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Collaborator
Author
|
Holding this off upstream for now — the operator wants to validate the CHANGELOG-ledger campaign (S2–S7, incl. the pre-commit hook and reconcile-on-read) locally on the fork first. The branch |
…ledger (BL-6 tracks the rest)
rmsharp
added a commit
to rmsharp/methodology
that referenced
this pull request
Jul 8, 2026
…se, fork sync (FM KJ5HST#27 close-out) Append one dated [ad hoc] ledger entry with the completed-deploy anchors: PR KJ5HST#46 merge 75a6853, annotated tag v3.1 + GitHub Release (Latest), fork main synced 1adf6b3, tag mirrored, feature branch pruned. Dogfoods the .githooks co-staging gate (CHANGELOG co-staged) and the FM KJ5HST#27 close-out rule. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
rmsharp
added a commit
to rmsharp/methodology
that referenced
this pull request
Sep 5, 2026
…re 7/10 Predecessor S117 scored 8/10: exceptional gotchas that changed this session's method, against a §6 that scopes every decision to the fork with no column separating what ships, a D7 ground that is false, and an F4 generalisation that over-reaches. Self-score 7/10. Verification held -- the tombstone scheme, the row budget and the precommit byte defect were each settled by running the code. Against that, nine current-implementation limits were stated as structural ones and the operator caught it, not me. Learning KJ5HST#46 appended at 229 B, measured before writing against 245 B of headroom. FRAMEWORK_LEARNINGS.md is now 73,712 / 73,728 -- 16 B left, and no row of any size fits. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HRT4SwEjyyR23kNrwG6osj
rmsharp
added a commit
to rmsharp/methodology
that referenced
this pull request
Sep 5, 2026
…, then compact Phase 1B claim. Deliverable: Phase 2 of docs/planning/upstream-read-set-pr-plan.md -- repair bin/check-learnings' row-budget scope, drive it RED against today's 20 violators, then compact all 20 over-budget rows of starter-kit/FRAMEWORK_LEARNINGS.md to under the published 1,500 B budget. THE OPERATOR ANSWERED S118'S GATING DECISION AT THIS CLAIM: compaction of existing rows IS permitted, all 20, including KJ5HST#12 and KJ5HST#13. That answer is the operator's and is recorded as its own CHANGELOG.md entry so a successor cannot read it as an agent's recommendation. The rule it rules on -- "append only; do not edit existing rows" -- names renumbering as its harm; compaction changes row content, not row numbers. Explicitly NOT the answer: raising the ceiling a third time. WHY PHASE 2 BEFORE PHASE 1, against the plan's own ordering. S118's next_steps (b): compacting first means upstream receives a file already under the cap rather than inheriting one 16,733 B over and needing a second change. THE BLOCKER THIS CLEARS, re-measured at this claim: FRAMEWORK_LEARNINGS.md is 73,712 B against a 73,728 B ceiling -- 16 B of headroom, and the smallest row the file has ever carried is 419 B. No learning row of any size fits. THE GUARD THAT CANNOT FIRE, verified live: check-learnings exits 0 and prints "0 unfrozen row(s), 0 over 1,500 B" with 20 violators present, because check_row_budget scopes the budget to rows differing from git HEAD. It must be seen to report 20 before anything is compacted. RE-MEASURED, NOT QUOTED. The plan's 55,930 B target was computed against a 73,483 B file; S118's Learning KJ5HST#46 added 229 B afterwards. NOT IN SCOPE: Phase 1 (the port), Phase 3 (the gate), D1/D2/D3/D4/D6, BL-45's remedy, issue KJ5HST#75's unsent PR, trimming either ledger, the upstream/main resync, and any outward-facing action whatsoever. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019tacYU7wZPUd7cHfD2332P
rmsharp
added a commit
to rmsharp/methodology
that referenced
this pull request
Sep 5, 2026
…nd the ceiling is broken the other way S121's deliverable. VERDICT: S120's "Phase 3C cannot be performed" is false as written. The arithmetic is right and the denominator is wrong. FOUR INDEPENDENT REFUTATIONS, each attacked by an agent told to break it. 1. METERED, the file is 19,617 tok = 78.5% of the 25,000-token cap. Six collinear probe points with ZERO residual on the three the model never saw; the concatenation seam MEASURED at exactly zero tokens, so f(1) is a measurement, not an extrapolation. Headroom 5,383 tokens. 2. 56,750 B is applied to this file by NO code path: methodology_trim.py answers [NO_CONFIG]/exit 3, both dashboard twins' read_cap_class() answer None, check-learnings has no whole-file arm. It is a detector FLOOR (2.27 B/tok, the conservative end of a measured band), not a budget. 3. Learning KJ5HST#34, in this very table, already moved the guard off the per-file axis at S97: read whole once, in part 243 times. The published budget is 1,500 B PER ROW and all 46 rows comply. 4. No gate fires when a session appends nothing; BL-45 is the precedent, where S110-S114 ran through an identical declared deadlock and closed normally. THE REAL DEFECT RUNS THE OTHER WAY, AND IT IS MEASURED, NOT MODELLED. 73,646 B of this content -- 82 B UNDER the declared 73,728 B ceiling -- is refused at 25,486 tokens. BL-45 deliberately set that ceiling 2,023 B UNDER the then-cliff; S119's compaction lowered density 3.0444 -> 2.8897 and inverted the margin. A byte ceiling is tokens x density, and compaction is the operation that changes density, so it rots silently. That is Learning KJ5HST#46's own mechanism, committed by the framework's most recent remedy. AND THE FINDING THAT ACTUALLY MATTERS: whenever headroom falls below ~1,500 B, the row written EQUALS the headroom minus a few bytes (KJ5HST#34=604 at 604, KJ5HST#46=229 at 245). The file has saturated four times. The defect is not a blocked append, it is a silent QUALITY TAX on the append. SIX OF MY OWN NUMBERS WERE WRONG AND ARE CORRECTED IN PLACE, marked in the doc. Two of them repeat, one level down, the exact error this document faults S120 for: I converted token headroom to bytes at the WHOLE-FILE density (2.8897) when the headroom is consumed by ROWS, which meter 3.1974 B/token. Corrected capacity is 11.5-16.5 rows, and the only safe unit is tokens (5,383). I also withdrew a "the config refutes itself" division that compares two quantities the repo explicitly declares non-comparable, and a fabricated quotation ("append a Learning row") that I introduced into the record in claim commit 43387e1. S120 WAS APPLYING ITS OWN RATIFIED PLAN'S STANDARD (:95, :106, :123). The defect is in the plan. S120's SECOND reason -- byte-identity with the unpushed port branch, blob b21854c identical on both -- is real, was never refuted, and is answered as recommendation (4), a precondition rather than a remedy. S119'S COMPACTION MUST NOT BE REVERTED: leg 1 (the row budget) independently justifies every row, since only the 20-row option turns the checker green. The operator's S119 decision table was scored on a column mis-scaled by ~19 KB; the outcome is unchanged, but that is a finding he is entitled to be told. NO REMEDY APPLIED. Present->Implement gate holds; recommendation sec 6 is an answer proposed to open decisions D4/D5, not a correction an agent may apply. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NJo3fDBzjoYXwzhygVK1SM
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What & why
Makes
CHANGELOG.mda dependable cross-source action ledger — the authoritative answer to "what was done here, ever?" across the three sources the operator named: backlog items, repository issues, and ad-hoc work. Originates from an operator request to depend onCHANGELOG.mdas that summary. Ratified plan (fork-main-only):docs/planning/changelog-authoritative-ledger-gate-plan.md(1710e90), decisions D1–D7.The central finding
A close-out write-gate alone is not dependable. An adversarial dependability pass returned
is_dependable: false: eight high-severity escape paths let an action land in git history with no ledger entry (ghost/crashed session, out-of-band commit, in-progress hand-off, absent-CHANGELOG adopter, canonical repo has no runner to gate on, non-commit actions, multi-commit sessions, checklist self-exemption). Fix = two mechanisms, not one: gate-on-write (close-out) and reconcile-on-read (Phase 0 backfill). Reconcile is the keystone — it makes the ledger true when you read it.Components (one session each — strict 1-and-done)
2227aab) — new FM docs(bootstrap): CLAUDE.md size-budget rule — extract Learnings to PROJECT_LEARNINGS.md when it overflows #27 "Unrecorded action" (appended; FMs 1–26 byte-unchanged, no renumber), a Degradation row, SESSION_RUNNER §3F + ITERATIVE Phase-6-step-8. Learning-v2.3 content release: 6 upstream candidates from rad-con SESSION_RUNNER audit (review requested) #7 count reconciliation 26→27 at all live sites.4828929) — Phase 0 diffsgit logsince the ledger frontier and backfills undocumented commits before new work; 1BCHANGELOG: pendingcrash breadcrumb; empty-frontier and timing traps closed. The dependability keystone.f25e0c4) — greppable### YYYY-MM-DD · [SOURCE]unit, exactly one source tag per entry ([issue #<N>]·[BL-<N>]·[ad hoc]);starter-kit/CHANGELOG.mdrecomposed from Keep-a-Changelog to the action-ledger shape (still aseed);SESSION_NOTES↔CHANGELOGdivision of labor.89b8f60) — git-only ledger-lag signal, advisory RISK only (never a hard fail); both byte-identical twins;DASHBOARD_VERSION2.6.1 → 2.7.0.339dfb2) — this repo's rootCHANGELOG.md, distinct fromCLAUDE.md§Versioning (release = a one-line pointer, cite-don't-restate, so the two ledgers can't diverge); bounded v3.0-forward backfill; all three source tags demonstrated for real.d2184cc+dc8aa76) — ships.githooks/pre-commit, the mechanical FM docs(bootstrap): CLAUDE.md size-budget rule — extract Learnings to PROJECT_LEARNINGS.md when it overflows #27 enforcement where a repo has no root runner to run the close-out gate (escape fix: correct collect_coverage_config() logic bugs and add recursive scanning #5): it refuses a commit changing tracked content unlessCHANGELOG.mdis co-staged. Deliberately narrow + bypassable (--no-verify; skips merges/rebases/absent-ledger repos), with reconcile-on-read as the guarantee. The ledger close-out step is propagated into every session-type + campaign checklist (Planning / Review-Audit / Debugging / vertical-slice-revert / campaign per-execution + consolidation), plus new Learning Scaling / Parallel Development Using Agents #8.Dogfood
git config core.hooksPath .githooksis live in the canonical repo, and the commit that ships the hook was gated by the hook and passed becauseCHANGELOG.mdwas co-staged. The rootCHANGELOG.mdrecords this campaign as one[ad hoc]entry.Verification
bin/tests.sh51/51; dashboard 35 fixture assertions (S5); hook 14 behavior tests incl. live block on the real repo +--no-verifybypass + thegrep -qxFregex-metachar regression (S7).-qxF), BOOTSTRAP First-Session-Checklist ledger gap (fixed),--amendfriction (documented), tutorial worked-transcript staleness (deferred — see below).Deliberately NOT in this PR
CLAUDE.md§Versioning v3.x entry is added yet.docs/tutorials/T2_worked_transcript.mdandHOW_TO_USE.md's close-out enumeration still teach a pre-FM-docs(bootstrap): CLAUDE.md size-budget rule — extract Learnings to PROJECT_LEARNINGS.md when it overflows #27 close-out. They teach the released methodology and are out of this campaign's scope; refresh them when this merges/ships (tracked as a follow-up).main-onlydocs/planning/*ledger plan is not in this PR (the branch is offupstream/mainfor a clean review).🤖 Generated with Claude Code