heavy-duty/ceremony

Fork 2

feat: fail shape drift on the introducing PR #163

Merged

codex-bot-andresmgsl merged 3 commits from build/159-changelog-shape-guard into main

2026-07-24 16:56:58 +00:00

codex-bot-andresmgsl commented

2026-07-24 16:13:22 +00:00

(Migrated from github.com)

Closes #159

Summary

Move fragment-set shape validation into one shared library predicate, enforce it in the PR-time guard and release-time assembler, and anchor fragments to the newest published changelog section.

Acceptance criteria

A flat published changelog with one grouped fragment beside flat fragments fails changelog-armed.
A flat published changelog with an all-grouped fragment set fails changelog-armed.
Mixed fragments retain the assembler diagnosis and share it with changelog-armed.
Grouped/grouped, flat/flat, no-published-section, and empty-directory cases pass.
Shape logic exists in one shared function and changelog_assemble delegates to it.
bash test/run.sh is green.

Worklog

Centralize cross-fragment and published-section shape validation in changelog_shape_problem.
Preserve changelog_assemble mixed-shape diagnoses through the shared predicate.
Add direct library tests for both anchor directions and vacuous cases.
Enforce the shared predicate in fragment-mode changelog-armed.
Enforce the shared predicate in bin/changelog-assemble.
Add action and executable contract cases for every required pass/fail path.
Add the flat changelog.d/159.md fragment.
Run the full test, shellcheck, and actionlint gates.

Closes #159 ## Summary Move fragment-set shape validation into one shared library predicate, enforce it in the PR-time guard and release-time assembler, and anchor fragments to the newest published changelog section. ## Acceptance criteria - [x] A flat published changelog with one grouped fragment beside flat fragments fails `changelog-armed`. - [x] A flat published changelog with an all-grouped fragment set fails `changelog-armed`. - [x] Mixed fragments retain the assembler diagnosis and share it with `changelog-armed`. - [x] Grouped/grouped, flat/flat, no-published-section, and empty-directory cases pass. - [x] Shape logic exists in one shared function and `changelog_assemble` delegates to it. - [x] `bash test/run.sh` is green. ## Worklog - [x] Centralize cross-fragment and published-section shape validation in `changelog_shape_problem`. - [x] Preserve `changelog_assemble` mixed-shape diagnoses through the shared predicate. - [x] Add direct library tests for both anchor directions and vacuous cases. - [x] Enforce the shared predicate in fragment-mode `changelog-armed`. - [x] Enforce the shared predicate in `bin/changelog-assemble`. - [x] Add action and executable contract cases for every required pass/fail path. - [x] Add the flat `changelog.d/159.md` fragment. - [x] Run the full test, shellcheck, and actionlint gates.

grok-bot-andresmgsl commented

2026-07-24 16:20:44 +00:00

(Migrated from github.com)

🔎 reviewing head eb25b38c14

🔎 reviewing head eb25b38c142f05f88d7f30e6574a145f0f9499e1

claude-bot-andresmgsl commented

2026-07-24 16:20:53 +00:00

(Migrated from github.com)

🔎 reviewing head eb25b38c14

🔎 reviewing head eb25b38c142f05f88d7f30e6574a145f0f9499e1

grok-bot-andresmgsl (Migrated from github.com) approved these changes 2026-07-24 16:22:40 +00:00

grok-bot-andresmgsl (Migrated from github.com) left a comment

Verdict: approve

Reviewed head eb25b38c142f05f88d7f30e6574a145f0f9499e1 against #159's acceptance criteria and the #157 D3 shape contract.

Spec

Criterion	Result
Flat published + one grouped fragment beside flat fails `changelog-armed`	Met — action calls `changelog_shape_problem`; test `fragments-dev-mixed`
Flat published + all-grouped set fails	Met — anchor rule against newest published section; test `fragments-dev-all-grouped-over-flat`
Mixed set fails assemble with same diagnosis, shared with armed	Met — mixed diagnosis text is identical through `changelog_shape_problem`; both callers print it
Grouped/grouped, flat/flat, no published, empty directory pass	Met — library + armed cases
Shape logic in exactly one shared function; assemble delegates	Met — `changelog_shape_problem`; `changelog_assemble` passes empty changelog so it only reuses cross-fragment (anchor stays in bin, per the issue — lib does not grow a path arg)
`bash test/run.sh` green	Met — 18/18 files green in a detached worktree at this head

Code

Cross-fragment detection moved out of changelog_assemble without changing the refusal text; single-file mix still diagnosed.
Anchor uses the first non-Unreleased ## section and changelog_section body for ### presence — matches family extractors.
changelog-armed runs the shared predicate after the per-fragment loop (dev-tree PR path); bin/changelog-assemble runs it with the real changelog path before assemble so release-time also gets the anchor.
changelog.d/159.md is a flat bullet, consistent with this repo's published flat shape.

Verification

bash test/run.sh — all green.
shellcheck -x on lib/changelog.sh, changelog-armed.sh, bin/changelog-assemble — clean.
Real tree: changelog-armed agrees; changelog_shape_problem CHANGELOG.md changelog.d silent.
Manual probes: mixed assemble/shape diagnoses match; all-grouped-over-flat and same-file mix refuse as specified.
CI test / self-guards / action exercises green on this head. release-exercise / fixture-chain failed at "Set up job" with HTTP 429 downloading actions/checkout — infrastructure, not this diff.

No blocking findings. Ship when the rest of the panel converges.

## Verdict: approve Reviewed head `eb25b38c142f05f88d7f30e6574a145f0f9499e1` against #159's acceptance criteria and the #157 D3 shape contract. ### Spec | Criterion | Result | |---|---| | Flat published + one grouped fragment beside flat fails `changelog-armed` | Met — action calls `changelog_shape_problem`; test `fragments-dev-mixed` | | Flat published + all-grouped set fails | Met — anchor rule against newest published section; test `fragments-dev-all-grouped-over-flat` | | Mixed set fails assemble with same diagnosis, shared with armed | Met — mixed diagnosis text is identical through `changelog_shape_problem`; both callers print it | | Grouped/grouped, flat/flat, no published, empty directory pass | Met — library + armed cases | | Shape logic in exactly one shared function; assemble delegates | Met — `changelog_shape_problem`; `changelog_assemble` passes empty changelog so it only reuses cross-fragment (anchor stays in bin, per the issue — lib does not grow a path arg) | | `bash test/run.sh` green | Met — 18/18 files green in a detached worktree at this head | ### Code - Cross-fragment detection moved out of `changelog_assemble` without changing the refusal text; single-file mix still diagnosed. - Anchor uses the first non-`Unreleased` `##` section and `changelog_section` body for `### ` presence — matches family extractors. - `changelog-armed` runs the shared predicate after the per-fragment loop (dev-tree PR path); `bin/changelog-assemble` runs it with the real changelog path before assemble so release-time also gets the anchor. - `changelog.d/159.md` is a flat bullet, consistent with this repo's published flat shape. ### Verification - `bash test/run.sh` — all green. - `shellcheck -x` on `lib/changelog.sh`, `changelog-armed.sh`, `bin/changelog-assemble` — clean. - Real tree: `changelog-armed` agrees; `changelog_shape_problem CHANGELOG.md changelog.d` silent. - Manual probes: mixed assemble/shape diagnoses match; all-grouped-over-flat and same-file mix refuse as specified. - CI `test` / `self-guards` / action exercises green on this head. `release-exercise / fixture-chain` failed at "Set up job" with HTTP 429 downloading `actions/checkout` — infrastructure, not this diff. No blocking findings. Ship when the rest of the panel converges.

kimi-bot-andresmgsl commented

2026-07-24 16:24:36 +00:00

(Migrated from github.com)

🔎 reviewing head eb25b38c14

🔎 reviewing head eb25b38c142f05f88d7f30e6574a145f0f9499e1

kimi-bot-andresmgsl (Migrated from github.com) approved these changes 2026-07-24 16:25:08 +00:00

kimi-bot-andresmgsl (Migrated from github.com) left a comment

Approve — head eb25b38c142f05f88d7f30e6574a145f0f9499e1, reviewed against #159's acceptance criteria in a detached worktree.

Every criterion verified by execution, not just reading:

#143 case red: flat published section + one grouped fragment beside a flat one → changelog-armed exits 1 with fragment 'changelog.d/115.md' is grouped but fragment 'changelog.d/114.md' is not — a repo is one shape or the other. Reproduced on a constructed tree.
All-grouped over flat published red: exits 1 naming the fragment and the source of the expectation — newest published section '1.2.3' in 'CHANGELOG.md' is flat. Reproduced.
Mixed set, same diagnosis everywhere: bin/changelog-assemble refuses with byte-identical text to origin/main's changelog_assemble string (grep + direct run of the main-branch lib), and changelog-armed emits the same shared diagnosis. Verified.
Anchor rule at release time: an all-grouped set over a flat published section makes bin/changelog-assemble refuse before assembling; the changelog file is left untouched. Verified.
Pass cases: flat/flat reproduced by hand (exits 0); grouped/grouped, no-published-section, and empty-directory cases are covered by the new direct changelog_shape_problem tests, which pass locally.
One definition: the refusal + diagnosis live only in changelog_shape_problem; changelog_assemble delegates with an empty changelog arg (anchor rule vacuous, cross-fragment rule intact). The re-scan of grouped_in in changelog_assemble is assembly-mode selection, not a second copy of the rule — no diagnosis text is duplicated.
Fragment: changelog.d/159.md present and flat, matching this repo's newest published section (0.2.0, flat), so the guard does not refuse its own PR; no grouped fragments remain in changelog.d/ at this head.

What I could not run locally and what I relied on instead: this box has no node, no jq, no shellcheck, no actionlint (mawk-only awk). The package-json version-source cases and the ruling/issueflow-reconcile suites fail here for that reason alone; the PR touches none of that surface, and the changelog-family suites (changelog, changelog-armed, changelog-assemble, changelog-assembled) pass locally in full. For the rest I relied on CI: the test, self-guards, action-exercise, docs-sync-exercise, and both step-replay jobs are green on this head.

One non-blocking note: release-exercise / fixture-chain is red, but the log shows a GitHub-side 429 downloading actions/checkout — the job never ran a line of PR code. A re-run should clear it; no code change owed.

**Approve** — head `eb25b38c142f05f88d7f30e6574a145f0f9499e1`, reviewed against #159's acceptance criteria in a detached worktree. Every criterion verified by execution, not just reading: - **#143 case red**: flat published section + one grouped fragment beside a flat one → `changelog-armed` exits 1 with `fragment 'changelog.d/115.md' is grouped but fragment 'changelog.d/114.md' is not — a repo is one shape or the other`. Reproduced on a constructed tree. - **All-grouped over flat published red**: exits 1 naming the fragment *and* the source of the expectation — `newest published section '1.2.3' in 'CHANGELOG.md' is flat`. Reproduced. - **Mixed set, same diagnosis everywhere**: `bin/changelog-assemble` refuses with byte-identical text to `origin/main`'s `changelog_assemble` string (grep + direct run of the main-branch lib), and `changelog-armed` emits the same shared diagnosis. Verified. - **Anchor rule at release time**: an all-grouped set over a flat published section makes `bin/changelog-assemble` refuse *before* assembling; the changelog file is left untouched. Verified. - **Pass cases**: flat/flat reproduced by hand (exits 0); grouped/grouped, no-published-section, and empty-directory cases are covered by the new direct `changelog_shape_problem` tests, which pass locally. - **One definition**: the refusal + diagnosis live only in `changelog_shape_problem`; `changelog_assemble` delegates with an empty changelog arg (anchor rule vacuous, cross-fragment rule intact). The re-scan of `grouped_in` in `changelog_assemble` is assembly-mode selection, not a second copy of the rule — no diagnosis text is duplicated. - **Fragment**: `changelog.d/159.md` present and flat, matching this repo's newest published section (0.2.0, flat), so the guard does not refuse its own PR; no grouped fragments remain in `changelog.d/` at this head. What I could not run locally and what I relied on instead: this box has **no `node`, no `jq`, no `shellcheck`, no `actionlint`** (mawk-only awk). The package-json version-source cases and the `ruling`/`issueflow-reconcile` suites fail here for that reason alone; the PR touches none of that surface, and the changelog-family suites (`changelog`, `changelog-armed`, `changelog-assemble`, `changelog-assembled`) pass locally in full. For the rest I relied on CI: the `test`, `self-guards`, `action-exercise`, `docs-sync-exercise`, and both `step-replay` jobs are green on this head. One non-blocking note: `release-exercise / fixture-chain` is red, but the log shows a GitHub-side 429 downloading `actions/checkout` — the job never ran a line of PR code. A re-run should clear it; no code change owed.

claude-bot-andresmgsl (Migrated from github.com) approved these changes 2026-07-24 16:25:28 +00:00

claude-bot-andresmgsl (Migrated from github.com) left a comment

Approve — reviewed the whole PR at head eb25b38, in a detached worktree.

Verified, not opined:

bash test/run.sh at the head: 18 test files, all green. shellcheck-all.sh and actionlint-all.sh both clean.
One definition, verbatim diagnoses. I diffed the cross-fragment scan and its two refusal strings in changelog_shape_problem (lib/changelog.sh:159) against what changelog_assemble carried on main — byte-identical, so the mixed-set diagnoses survive exactly as #159 requires. changelog_assemble delegates (lib/changelog.sh:227) by passing an empty changelog path, which keeps the anchor rule out of the lib-level function without growing it a new argument — precisely the spec's shape. No second copy of the validation survives in either caller.
Anchor rule. Newest published section = first non-Unreleased ## heading; grouped iff its changelog_section body has a ### line. The $1 == "##" awk idiom matches the existing uses at changelog-armed.sh:85 and bin/changelog-assemble:89, so the family keeps agreeing on what a heading is. Empty set, no published section, and absent changelog are all vacuous, per spec.
Wiring. changelog-armed.sh:99 runs the check in fragment mode after the per-fragment loop, on every tree; bin/changelog-assemble:79 runs it before assembling — and therefore also under --check, which the release-replay path inherits; the added grouped-replay coverage in test/changelog-assembled.test.sh caught exactly that interaction.
Acceptance criteria, each against a test I ran: one grouped fragment beside flat ones goes red in changelog-armed with the shared two-fragment diagnosis (the #143 case); an all-grouped set over a flat published section goes red naming fragment, section, and file (the case rule 1 structurally cannot see); a mixed set still fails the assembler with today's text and fails the guard with the same shared text; grouped/grouped, flat/flat, no-published-section, and empty-directory all pass; changelog.d/159.md is present and flat.
Self-consumption. I ran actions/changelog-armed/changelog-armed.sh on the PR's own tree: exit 0 — the all-flat changelog.d/ anchors cleanly against the flat newest published section, so the #158 ordering paid off.

CI note: release-exercise / fixture-chain is red at this head, but the job died in "Set up job" on a GitHub-side 429 rate limit downloading actions/checkout — no repo code ever ran. Its content is the same contract test/release-chain.test.sh exercises, which is green locally at this head. A plain re-run should clear it; nothing in the diff to answer for it.

Non-blocking nit (builder's discretion): test/changelog-assembled.test.sh:88 — the sed -i '/^- The shipped entry/i ### Fixed\\\n' insert leaves a literal trailing backslash in the fixture: the seeded heading is ### Fixed\, not ### Fixed. It still matches ^### , so the replay anchors exactly as intended, but the fixture reads as a typo; seeding the grouped changelog with a heredoc (as the armed tests do) would say what it means.

**Approve** — reviewed the whole PR at head eb25b38, in a detached worktree. **Verified, not opined:** - `bash test/run.sh` at the head: 18 test files, all green. `shellcheck-all.sh` and `actionlint-all.sh` both clean. - **One definition, verbatim diagnoses.** I diffed the cross-fragment scan and its two refusal strings in `changelog_shape_problem` (lib/changelog.sh:159) against what `changelog_assemble` carried on main — byte-identical, so the mixed-set diagnoses survive exactly as #159 requires. `changelog_assemble` delegates (lib/changelog.sh:227) by passing an empty changelog path, which keeps the anchor rule out of the lib-level function without growing it a new argument — precisely the spec's shape. No second copy of the validation survives in either caller. - **Anchor rule.** Newest published section = first non-`Unreleased` `## ` heading; grouped iff its `changelog_section` body has a `### ` line. The `$1 == "##"` awk idiom matches the existing uses at changelog-armed.sh:85 and bin/changelog-assemble:89, so the family keeps agreeing on what a heading is. Empty set, no published section, and absent changelog are all vacuous, per spec. - **Wiring.** changelog-armed.sh:99 runs the check in fragment mode after the per-fragment loop, on every tree; bin/changelog-assemble:79 runs it before assembling — and therefore also under `--check`, which the release-replay path inherits; the added grouped-replay coverage in test/changelog-assembled.test.sh caught exactly that interaction. - **Acceptance criteria, each against a test I ran:** one grouped fragment beside flat ones goes red in `changelog-armed` with the shared two-fragment diagnosis (the #143 case); an all-grouped set over a flat published section goes red naming fragment, section, and file (the case rule 1 structurally cannot see); a mixed set still fails the assembler with today's text and fails the guard with the same shared text; grouped/grouped, flat/flat, no-published-section, and empty-directory all pass; `changelog.d/159.md` is present and flat. - **Self-consumption.** I ran `actions/changelog-armed/changelog-armed.sh` on the PR's own tree: exit 0 — the all-flat `changelog.d/` anchors cleanly against the flat newest published section, so the #158 ordering paid off. **CI note:** `release-exercise / fixture-chain` is red at this head, but the job died in "Set up job" on a GitHub-side 429 rate limit downloading `actions/checkout` — no repo code ever ran. Its content is the same contract `test/release-chain.test.sh` exercises, which is green locally at this head. A plain re-run should clear it; nothing in the diff to answer for it. **Non-blocking nit (builder's discretion):** test/changelog-assembled.test.sh:88 — the `sed -i '/^- The shipped entry/i ### Fixed\\\n'` insert leaves a literal trailing backslash in the fixture: the seeded heading is `### Fixed\`, not `### Fixed`. It still matches `^### `, so the replay anchors exactly as intended, but the fixture reads as a typo; seeding the grouped changelog with a heredoc (as the armed tests do) would say what it means.