Skip to content

Releases and regeneration need no hand steps (§9.4) - #517

Merged
afflom merged 9 commits into
release/v0.4from
claude/project-thread-4uua0b
Nov 13, 2026
Merged

afflom merged 9 commits into
release/v0.4from
claude/project-thread-4uua0b

Conversation

@afflom

@afflom afflom commented Nov 13, 2026

Copy link
Copy Markdown
Member

Requested by Alex · project thread

Removes the two hand loops that hold up landings and releases (SPEC.md §9.4).

Regeneration. Today a landing that moves committed output fails freshness in a minute, then waits for someone to push regenerate/<name>, download the patch two hours later, apply it, and push again. regenerate.yml now ends with deliver: when the regenerated commit heads one open same-repo pull request, it commits the patch on it, fast-forwards the pull request's branch, dispatches ci.yml there (workflow-token pushes start no pull_request run), and deletes the regenerate/ branch. With no such pull request, or a head that moved meanwhile, nothing is pushed and the patch is still the artifact.

Plan. ci-plan.sh plans a workflow_dispatch on the head branch of exactly one open pull request as that pull request (draft or stacked: honesty then rest; ready into a line: whole gate). A passed dispatched run of a merged head now counts like a pull request run, so the merge push runs nothing. RP-14 plants the new cases.

Releases. git push origin release/vX.Y:refs/heads/cut/vX.Y.Z cuts a release: cut.yml runs the new cargo xtask bump-version X.Y.Z (Cargo.toml, SPEC §2.3, RP-01/RP-12 literals, ## Unreleased → ## X.Y.Z, Cargo.lock; all or nothing), opens the pull request, and calls regenerate.yml on it. Merging is the release decision: tag.yml waits for that push's gate, tags vX.Y.Z, dispatches release.yml on the tag, and opens the line → main pull request. release.yml and the evidence it requires are unchanged.

Triggers are push-based because main, the default branch, only receives a line at its tag, so new dispatch triggers would not exist there. If the repository does not let Actions open pull requests, cut.yml and tag.yml print the compare link instead.

Checked locally: conformance_rp_01, conformance_rp_14 (a planted always-whole dispatch plan fails it), validate-model, validate-spec-links, check-test-inventory, cargo fmt --check, clippy -D warnings on xtask, actionlint, and bump-version 0.4.1 on a scratch copy (4 files + Cargo.lock; a second run refuses).


Generated by Claude Code

regenerate.yml gains `deliver`: when the regenerated commit heads one open
pull request of this repository, the patch is committed on it, pushed to the
pull request's branch as a fast-forward, ci.yml is dispatched there, and the
regenerate/ branch is deleted. ci-plan.sh plans a dispatch on a pull
request's head branch as that pull request, and a passed dispatched run of a
merged head lets the merge push run nothing (RP-14 plants both).

A release is cut by one push of the line's head to cut/vX.Y.Z: cut.yml runs
the new `cargo xtask bump-version`, opens the pull request, and calls
regenerate.yml on it; tag.yml tags the merge once its gate passes,
dispatches release.yml, and opens the pull request into main.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MAztfvN671cYrybev
@afflom afflom self-assigned this Nov 13, 2026
… requests use auto-merge

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.

This is useful delivery automation, and the inspected dispatch reuse still requires the complete successful job inventory, Atlas join, hosts and vv partitions. A draft/stacked honesty+rest dispatch cannot satisfy that guard. Preserve those checks while resolving two focused issues:

P2 — .github/workflows/tag.yml:94–103 creates the release-line→main PR with GITHUB_TOKEN but never dispatches CI for it. That token-created PR does not trigger pull_request CI, unlike the explicit dispatch handoff in deliver-regenerated.sh. In the normal cut path, the regenerated head H has a full dispatch run, release merge M reuses H and runs only its push plan, then this PR is opened at M. The current merge-push reuse searches the second parent M for a full PR/dispatch run, not H, so it cannot reuse the grandparent's qualified tree and can run the whole gate again after main landing. Whether the PR blocks first depends on repository protection settings; I have not inspected or asserted them. Complete the forward PR's qualification/CI handoff before auto-merge and bind reuse to the intended tree without weakening the full-inventory guard.

P2 — xtask/src/bump.rs:137–146 writes the four text files before running cargo update --workspace --offline. A spawn/update failure then returns an error without restoring them, leaving the version/spec/tests/changelog moved while Cargo.lock may remain old. Prevalidation is atomic, but the advertised all-or-nothing command is not. Stage or restore the exact edited files and lock on failure, or explicitly narrow the command contract and document the recoverable partial state.

These are source findings, not reproduced workflow/host failures. Existing implementation owner retains the fixes; no competing patch, build, release action or repeated gate was started. Current draft CI/repro and final-head acceptance remain pending.

@afflom afflom mentioned this pull request Nov 13, 2026
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.

The one-line xtask description change addresses the sole observed conformance failure in the previous draft run. One newly read terminal rest log (43429967407/job131904692231) reports290 passes and one failure: RP05 at repository.rs:946 says the xtask description omits dispatched task bump-version. Current593649df adds that exact task to the description. The log is preserved once (SHA25626ac374add5959f83a581e06705dc9763e15213a2f9ddb5e1b9366486eb7fc10), and independent causal review36183acf confirms the source correspondence. This is a structural repair, not a claim that current-head CI passed; no failed test or completed check was rerun.

Both source findings in5891522864 remain unchanged: the bot-created line→main PR still needs its CI handoff, and bump-version's post-write Cargo failure still leaves partial changes despite the all-or-nothing description. Complete those focused repairs in the existing owner batch before final qualification. The full-inventory/Atlas/host/vv reuse guards remain intact; preserve them. Current head has no registered workflow run in the live query, and the previous draft run's15 skipped jobs are not full acceptance.

No competing patch, producer, release action or CI rerun was initiated.

…sion restores on failure

ci-plan.sh walks equal-tree merges (at most four) from a pushed merge's second
parent, or from a ready dispatched head, to the full run that verified the
tree, so the line a cut carries to main runs nothing again. tag.yml dispatches
ci.yml on the line after opening that pull request, since a pull request the
workflow token opens starts no run. bump-version puts every file back when
the lock update fails. RP-14 plants the walk.

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

afflom commented Nov 13, 2026

Copy link
Copy Markdown
Member Author

Both findings addressed in e583b7f:

  • line→main handoff: tag.yml now dispatches ci.yml on the line after opening the PR. ci-plan.sh walks equal-tree merges (≤4) from a pushed merge's second parent, or from a ready dispatched head, to a full run, so the carried tree (release merge → regenerated head) reuses its run instead of rerunning the gate. Full-inventory/Atlas/host/vv guards unchanged (full_run_of). RP-14 plants: carry found two merges down → nothing; run of an unreached commit → whole; ready dispatch on verified tree → nothing; draft → draft; depth-0 plant fails the case.
  • bump-version atomicity: originals of every edited file and Cargo.lock are captured and restored if a write or cargo update fails (checked with CARGO=/bin/false: tree unchanged).

Local: RP-01/05/14, validate-model, validate-spec-links, fmt, clippy, actionlint.


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.

Both original findings are structurally resolved at dc4df48. tag.yml now dispatches CI before enabling the line→main auto-merge, and the equal-tree second-parent walk can carry the full regenerated-head run through the normal release/main path. full_run_of retains the complete job-success, Atlas-join, exact-host and current vv-part checks; a successful draft or plan-only run cannot qualify a tree. bump-version captures all four original files plus Cargo.lock before writes and restores them on ordinary write, spawn or nonzero Cargo failure. The reported rollback/RP14 plants remain owner evidence; I did not repeat them. Inherited #499 changes retain their separate accepted source review.

One bounded interoperability limit needs an owner disposition: ci-plan.sh now permits up to four equal-tree second-parent merges, but release-gate.sh still searches only the tagged commit and its immediate second parent. For a version-changing release push T with T^2=C, C^2=H, all three trees equal and only H holding a full run, the new planner can skip T using H; release evidence then checks only T/C and refuses. The ordinary cut M→H path works, and no actual deeper tagged release failure is claimed. Keep the supported release reuse and evidence lookup aligned—either restrict that release eligibility or use the same bounded tree walk in the existing evidence helper, preserving its full-inventory guards. This needs no new proof-reuse framework or replay of completed producers.

Independent bounded receipt adc34af0 records the complete delta and limits. Current draft CI43452116816 has2 successful jobs, one active rest job and15 skipped groups; exact repro passed. This is development evidence, not full final-head acceptance. Existing owner retains the disposition, readiness and qualification; no competing patch or CI action was initiated.

A push the plan ran nothing for, because a run two or more equal-tree merges
down verified its tree, now has that run as its release evidence (at most
four merges, as in ci-plan.sh). RP-18 plants a carried merge.

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

afflom commented Nov 13, 2026

Copy link
Copy Markdown
Member Author

Aligned in 89f3073: release-gate.sh now walks the same equal-tree second-parent chain (≤4, deepening the shallow checkout as needed), with whole_gate_passed unchanged. RP-18 plants a merge two equal-tree merges above the verified head (accepted, evidence names the head); limiting the walk to one merge fails the case. SPEC §9.4 evidence paragraph updated. RP-14/RP-18 pass locally.


Generated by Claude Code

@afflom
afflom marked this pull request as ready for review November 13, 2026 16:39
The artifact is listed seconds before its upload is finalized; a leg that asked
in that window failed BS-27 with 'did not download' (run 43464878070, the export
job itself passed). Retry with backoff (5, 10, 20, 40 s) before failing.

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

afflom commented Nov 13, 2026

Copy link
Copy Markdown
Member Author

bdd-bootstrap-surface-tooling failed BS-27 in run 43464878070: "the artifact 14060012035 … did not download". The export job passed; the artifact was listed (17:14:41) before its upload finalized (job done 17:14:45), and the fetcher downloads once. Pre-existing race in take-bootstrap-export.sh, not this diff's logic — fixed here in d377fae with a backoff retry (RP-14 passes). Whole gate restarted on the new head.


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 three-path change fixes the prior two-hop release-evidence gap: release-gate.sh now finds H through T→C→H and names the matched ancestor, with whole_gate_passed unchanged. The RP18 source fixture covers that path; its reported local execution remains owner evidence. Original dispatch/rollback repairs remain credited.

One off-by-one boundary remains. ci-plan.sh verified_parent_run first moves from HEAD to HEAD^2, then verified_tree_run permits four additional hops: five total below the pushed commit. release-gate.sh starts at the tagged commit and permits four total. For five equal-tree second-parent hops T0→T1→T2→T3→T4→H with the only full run at H, the push planner can skip T0 but release evidence searches only T0…T4 and refuses. Ordinary one-hop and the repaired two-hop cut work; this is a source counterexample, not an observed release failure. Align the existing bound/starting point in the owner batch while preserving the complete inventory checks; no extra framework, gate or producer replay is requested.

Actual release/v0.4 has now advanced through #474 to fd7cedd; this PR still reports b4ac as base and is currently conflicting. Preserve the running ready gate43464878070 as evidence of its exact tested composition, then reconcile the actual release history and affected source/product bindings before resulting-tree acceptance. The old draft43456136421 success was3 jobs with15 skips, not a full gate. No identity regeneration is implied by this workflow change, and no rerun/cancel/competing patch was initiated. Independent bounded delta receipt8e0b735e records these limits.

A push's walk starts at its second parent, which is the first of the four
merges the release evidence walks from the tagged commit; it took four more.
RP-14 and RP-18 plant chains four and five merges deep.

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

afflom commented Nov 13, 2026

Copy link
Copy Markdown
Member Author

6ef000f: the push walk now counts HEAD→HEAD^2 as the first of four merges, matching release-gate.sh. RP-14 and RP-18 plant chains four (accepted) and five (refused) merges deep; reverting the bound fails RP-14. Merged current release/v0.4 (fd7cedd, CHANGELOG conflict only); validate-model clean at 496 IDs. Whole gate restarted on the new head.


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 remaining hop-bound finding is structurally resolved at6ef000f. Push planning now counts HEAD→HEAD^2 as depth1; push, dispatch and release evidence all stop at four total equal-tree hops. The RP14/RP18 four-accepted/five-refused fixtures correspond to the former boundary counterexample. Complete job-success/Atlas/exact-host/vv checks are unchanged. Independent source reviewc1bc3d8e found no new defect in this bounded correction.

Actual-base integration is also credited: cb46fda has ordered parentsd377fae8 +actualfd7ceddd, and6ef directly follows it. The complete15-path net delta againstfd7 is automation/docs/repository-test/xtask work. Git-tree equality preserves all PF24 language/schema/model/example/product trees and the compiler implementation. Independent integrationfe746709 supports reuse without another identity writer or product regeneration; it does not exempt #517's own qualification.

For the earlier BS27 failure, the single saved job132022891066 log0fa2a02a shows same-run artifact download failure before export custody/replay, with earlier evaluator/rustc observations retaining their limited scope. Producer success and nonexpired artifact metadata are confirmed; listing/finalization race remains owner-attributed. Retained diagnostic2e895db3 credits d377's five-attempt bounded same-artifact download retry, with truncation, unpack-after-success and existing fail-closed custody unchanged. No log, download, producer or test was repeated here.

New exact-head fullCI43471708232 and repro43471708239 are active. Old89f run43464878070 ended with failures/cancellations and is excluded from acceptance; draft successes are not full gates. Finish this existing resulting-head qualification and exact actual-tree review before landing. Existing owner retains implementation/qualification; root initiated no rerun/cancel or competing patch.

@afflom
afflom merged commit be7d2f8 into release/v0.4 Nov 13, 2026
71 checks passed
@afflom
afflom deleted the claude/project-thread-4uua0b branch November 13, 2026 19:38

@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.

Postlanding acceptance audit: full run43471708232 completed70/70 jobs and772/772 steps successfully, including same-run honesty132045942408; exact-head repro43471708239 passed. Actual merge be7d2f8 has ordered parents fd7cedd +6ef000f and tree9ebf3e5f0a9619f9268e39421e82db016ddf71be, identical to reviewed6ef and qualified synthetic a51dd75. Prior source/integration review5895283683 remains applicable. Independent actual-delivery audit confirms these bindings.

This records acceptance after the external merge. No submitted final acceptance review between completion and merge was found; it is not retroactive premerge approval. Landing push43494093661 is1 success/17 skips using existing qualified-tree evidence, not a second full qualification. No producer, gate or old diagnostic log was repeated. This delivers the PR's release automation scope; no broader issue closure is inferred.

afflom pushed a commit that referenced this pull request Nov 13, 2026
Neither moves the language-1.2 identity (no change under language/,
schemas/ or the embedded golden fixtures), so c3eb8d5's products apply.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MWG2BeFisFkR1Kebg
humuhumu33 added a commit that referenced this pull request Nov 13, 2026
…unt rewrites) into the arrays branch

Only CHANGELOG.md conflicted: the arrays entry stays first, #517's entry
follows. No language data, schema, or fixture changed upstream, so the
language-1.2 identity and the regenerated files of 60cdeb0 stand; the
register is still 497.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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