Skip to content

Changelog fragments: PRs add changes/<slug>.md, the cut folds them (§9.4) - #548

Open
afflom wants to merge 4 commits into
release/v0.4from
claude/project-thread-4uua0b
Open

afflom wants to merge 4 commits into
release/v0.4from
claude/project-thread-4uua0b

Conversation

@afflom

@afflom afflom commented Nov 15, 2026

Copy link
Copy Markdown
Member

Requested by Alex · project thread

Today every landing on release/v0.4 makes every open pull request conflict on CHANGELOG.md's ## Unreleased list. Each re-merge then restarts the ~3.5 h gate; #518 alone did this to #529, #503 and #543. A merge=union attribute would not help, because GitHub's mergeability check ignores it. With this change, pull requests add changes/<slug>.md and never edit CHANGELOG.md, so they don't conflict there.

  • cargo xtask bump-version, which cut.yml runs:
    • folds the existing ## Unreleased entries, then each changes/*.md in name order, under ## X.Y.Z
    • leaves an empty ## Unreleased and deletes the fragments
    • is still all or nothing: on failure it restores every file, including the deleted fragments and Cargo.lock
  • validate-model checks every fragment: it must be .md, non-empty, and have no #/## heading outside a fence. A bad fragment fails its own pull request, not the cut.
  • The deferral audit now also scans changes/.
  • AGENTS.md "Release lines", SPEC.md §9.4 and cut.yml describe the rule. The first fragment is this change's own.

Checked locally:

  • cargo test -p xtask: 24 pass, 3 of them new: fold order, restore on a late failure, and the fragment rules.
  • validate-model, validate-spec-links, cargo fmt --check and clippy -D warnings on xtask all pass.
  • A real bump-version 0.4.2 with two demo fragments folded all three in name order and emptied changes/.

Generated by Claude Code

Every pull request edited CHANGELOG.md's `## Unreleased` list, so each
landing on a release line made every open pull request conflict and
restart its gate. A pull request now adds `changes/<slug>.md` instead.
`cargo xtask bump-version` folds the Unreleased entries, then every
fragment in file-name order, under `## X.Y.Z` below an empty
`## Unreleased`, and deletes the fragments; on failure every file,
the deleted fragments too, is restored. validate-model refuses a
fragment that is empty, holds a `#`/`##` heading, or is not `.md`, and
the deferral audit scans `changes/`. AGENTS.md and SPEC.md §9.4 say
contributors add a fragment and never edit CHANGELOG.md.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MAztfvN671cYrybev

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Bounded source review: moving each PR's release entry to a separate fragment addresses the demonstrated CHANGELOG merge-conflict churn. The transactional snapshots include version files, fragments and Cargo.lock before mutation; the rollback paths and deferral-audit inclusion are coherent in this source review. Existing cut/land/tag control is retained. Owner-reported tests are scoped evidence, not a measured CI saving or final acceptance.

P2 — xtask/src/bump.rs:100–116, fragment_fault: the fence boolean accepts an unclosed opening ``` and returns None at EOF. Once folded into CHANGELOG.md, that fragment can turn later fragments and existing release headings into code until another fence occurs. It also toggles on a three-backtick line inside a four-backtick fence, which is not a valid close, and misses valid headings indented up to three spaces (for example ## 0.9.0). Thus inputs accepted by validate-model can violate the promised release-section boundary.

Please make this small validator track the actual fence marker/length and require closure, and detect allowed-indentation ATX headings outside fences, retaining valid fenced examples. Add focused cases for an unclosed fence, a too-short closing fence, and an indented heading before the next qualification cycle. No completed gate, producer or local test was repeated here; implementation remains with the current owner.

@afflom afflom mentioned this pull request Nov 15, 2026
The fragment check toggled on any line starting with three backticks, so an
unclosed fence passed (and would swallow later sections once folded), a
shorter run closed a longer fence, and an ATX heading indented by up to three
spaces was missed. Fences now track their marker and length, must close, and
headings are read with CommonMark's indentation.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MAztfvN671cYrybev

afflom commented Nov 15, 2026

Copy link
Copy Markdown
Member Author

e0751d0 fixes fragment_fault, which now reads fences and headings the way CommonMark does:

  • A fence is a run of three or more backticks or tildes, indented at most three spaces. It closes only on a run of the same character that is at least as long and has nothing after it.
  • A fence still open at the end of the fragment is a fault ("leaves a code fence open").
  • #/## ATX headings indented up to three spaces are refused. Four spaces is indented code, and #hashtag is not a heading.

New cases cover:

  • an unclosed fence, including tildes
  • a ``` line inside a ```` fence, which does not close it
  • an indented ## 0.9.0, inside and outside a fence

The valid fenced examples still pass. cargo test -p xtask, clippy -D warnings, validate-model and fmt --check are clean.


Generated by Claude Code

afflom commented Nov 15, 2026

Copy link
Copy Markdown
Member Author

The red build and non-Lean tests on 91a8278 is from run 43671389751, which my push of e0751d0 cancelled. The aggregate step failed on PART_RESULT: cancelled, not on a test. The gate that counts is the run on e0751d0.


Generated by Claude Code

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Bounded follow-up at e0751d0: the prior fragment-validation P2 is resolved in source. The validator tracks fence character and opening length, closes only with a sufficient matching run, refuses a fence left open at EOF, and detects level-one/two ATX headings with up to three leading spaces. The focused cases cover the previously reported unclosed fence, too-short close and indented heading while preserving valid fenced examples and indented code.

The one-file delta leaves folding, transaction rollback and existing release controls unchanged. No new blocker found in this bounded repair; prior source review credit is reused. Owner-reported local tests remain scoped evidence. Current exact-head CI/reproducibility and actual merged-tree acceptance are still required; the canceled91a827 run is not aggregated into success. No local test, producer or gate was repeated here.

The workflow-token push of a cut's regenerated patch starts pull_request runs
that wait for approval and complete as action_required. await-gate.sh took the
newest such run as the gate and failed the v0.4.3 cut (run 43701488279) while
the dispatched gate was still running.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MAztfvN671cYrybev

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Bounded follow-up at b617cfc: the two-file change repairs the demonstrated #551 wait failure. The preserved failed-job132798591656 log shows await-gate selected completed action_required CI43709775331 and exited1 while the eligible dispatched gate remained pending. The new filter excludes only approval-pending action_required results; failed/canceled runs are still failures, no eligible run continues waiting, and a successful candidate still goes through the unchanged whole-gate validator.

No new blocker found in this repair; prior e075 changelog-plant repair credit remains scoped. This grants no automatic cut retry or release acceptance, and does not relax the standing Intel/full-premerge or STOP boundaries. Current exact-head CI is queued and reproducibility passed; final-head full acceptance and actual-tree equality remain required. No tests/gates/writers repeated. Independent receipt ef248a2715e1a4a3b20cbf28343a10ac4ba514241e1f109f2009150e088ea402; sole failure log SHA919ba540fe9b9ce7e1b8d4d3609700e39001179285f068b86458900e6f5432e8 retained.

afflom commented Nov 15, 2026

Copy link
Copy Markdown
Member Author

build and non-Lean tests (aarch64-unknown-linux-gnu) (job 132800007947) did not fail a test. The runner received a shutdown signal at 12:18:02 while conformance_bs_* was still running, and every test up to that point had passed. That is a GitHub infrastructure failure, not this diff. I'll re-run that job once, after the rest of run 43709974315 finishes.


Generated by Claude Code

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Attempt2 of43709974315 is not qualified:65 successful jobs, one failed ARM Linux job and three Intel obligations skipped. The sole fetched ARM log132837774577 reports runner shutdown at14:38:49, then operation cancellation during Cargo tests. It contains no failing assertion/compiler error/OOM diagnostic establishing a product defect; outstanding BS05–BS08 had not completed. Do not infer timeout, memory exhaustion or source correctness from this interruption.

The actual checkout in this log is synthetic merge6444ba6d, not a direct b617 checkout. Preserve attempt1 and attempt2 separately and retain successful scopes; no canceled or unfinished job is acceptance. Existing owner should disposition the unfinished ARM lane using available runner/run incident evidence; no automatic retry or new gate is requested here. Prior bounded selector/changelog source reviews remain unchanged, with standing final-head/full-gate/actual-tree requirements.

Log SHA798cddf7a1d46758aebdf6a6b204189ee8423fe941faf08dd56e669a6f4ebb1b; diagnostic receiptd142948fadde8890a45ed87cdf468b0be54d65e4ef0c49ba319584a74b2bafda. COMMENT only.

Brings #536 (the AU-01 corpus memory fix) and the line's other landings;
the aarch64 cargo-test leg lost its runner twice on the older base.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MAztfvN671cYrybev

afflom commented Nov 15, 2026 •

Copy link
Copy Markdown
Member Author

The aarch64 Linux cargo-test job lost its runner on both attempts, both times while conformance_bs_05–bs_08 were still running. Neither log shows a failed assertion, so I'm not treating it as a flake, and I'm not re-running it a third time.

Its base 4727a69 predates #536, and 4353bcd merges release/v0.4 in to bring #536 in. [Edit: that was wrong. The pull-request runs tested the synthetic merge with the line, which already contained #536. The tree of 4353bcd equals the tree that was tested before, so the merge changes ancestry only.] The loss is handed to the CI gate investigation.


Generated by Claude Code

afflom commented Nov 15, 2026

Copy link
Copy Markdown
Member Author

The aarch64 Linux runner was lost again on 4353bcd (job 132871560408), even with #536 merged in. Again it went down during BS-05..08 with no failing test. It has now happened on this PR's heads and on #551's gate, so it isn't caused by this diff, which touches xtask, workflows and docs and doesn't affect the BS cases. I've handed the runner memory investigation to the CI gate work, and this PR waits on that rather than on more re-runs.


Generated by Claude Code

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The incident's tested-source binding needs one correction: attempt2 checked out synthetic6444ba6d, whose parents already include ab91576 and hence #536 actualf76a. Old checkout6444, new head4353 and new checkout4644dbfc all have the exact same tree9ee16cd6. This merge changes branch ancestry but did not introduce #536 into the previously tested composition. Prior b617 source-review credit is preserved; no new merge-composition defect found.

The one new retained log132871560408 confirms runner shutdown at16:34:59 and cancellation16:35:00 during unfinished Cargo tests; BS05–BS08 had not finished. It does not establish OOM, a product defect, a hang or a platform cause. Log SHA07f5d440936735b48ee9a537c763bc82028d99b091e940ec20b1bbcb3715fb0b; previous logs were reused locally, not fetched again.

Keep the current handoff to the existing CI investigation and the owner's no-more-reruns decision. Its concrete next result is a retained cause/disposition for the unfinished ARM lane, followed by applicable full final-head qualification; neither tree equality nor successful neighboring jobs supplies that acceptance. Receipt b6bb20cce6a9bb24d75cf177b9e2a836a93da02c5498bcfd8d18473c27b984eb. COMMENT only; no retry/control action.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants