Repository navigation
Heal repair artifact was empty on every run: derive include-hidden-files from the path roster - #7830
Merged
Merged
Conversation
…ved-closure namespace-reference-derived-closure carried a six-capability set-difference closing contract in which every row read Unavailable, while the node itself was the only dispatchable thing in the lane -- one startable row standing for three separable pieces of work with different substrates. The cut follows the contract's own triggers, which already record what each capability waits on: namespace-structural-observations 4 caps, P2aStructuralCandidateProducer7515 namespace-cross-file-provenance 1 cap, P2aReferenceDependencyProjection7515 namespace-pool-independence 1 cap, P2aPoolIndependentDependencyProjection7515 The first two are parallel -- no dependency runs between the same-file rules and the cross-file projection. Pool independence depends on cross-file provenance because a differential needs a projector to perturb. No new identity was minted for integration or for the census. namespace-reference-derived-closure keeps its durable identity and now denotes the aggregate handback (its first_slice moved; its boundary, which still covers all six, did not). namespace-ambiguity-discharge keeps its identity and its existing dependency on the closure node. The three new rows are deliberately ExecutionContractUnspecified. Each first_slice names authoring its own closing check as its first act, per v1_lane_binding_survey_note: the contract follows the witness, never precedes it. The derived closing-contract tasks for the two startable rows are the honest fail-closed state, not a gap; pool-independence gets none because it is dependency-held, and could not carry one anyway -- its differential compares an output shape its prerequisite has not yet produced. Also adds roadmap-receipt-continuity (roadmap-runtime, off the namespace spine): a receipt that was valid and is now absent refuses unless an explicit revocation names the exact node, the exact record, a reason and its disposition. Motivated by #7739, where a branch spent real effort reconstructing an acceptance record that already existed because nothing refused when it went missing. Distinct from startable_nodes_missing_closing_contract, which finds nodes that never had a check rather than accepted state that vanished. Evidence, by execution on this tree: - roadmap_authority_test: 42/42 witnesses PASS - generated_artifact_drift_test: 7/7 PASS - ROADMAP.md regenerated via main_wet on dag/tools/generated_artifact_gate.dag; the projection shows reference-derived-closure requiring the two new prerequisites, and derived closing-contract tasks appearing for exactly the two startable unbound rows. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…ion/loyal-ram-550
An empty file named 'true' was created in the worktree by a shell-quoting mishap while sending dashboard messages, then picked up and committed by the WIP auto-commit process. It is not on main and carries no content. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
briansrls
marked this pull request as ready for review
August 5, 2026 18:29
Author-committed because the GitHub App lacks workflows:write, so the heal job can detect this drift but can never push the fix. Emitted from the ci_workflow.dag change in this branch; the value is derived from author_commit_required_committed_artifact_paths(), not hard-coded. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
gunbai-bot
Bot
changed the title
The first draft cited "#7770 twice, #7791, #7835, #7857". Two errors: #7770 hit the author-commit-required class three times on 2026-08-05 (ci.yml at 04:39 and 17:42, falsifier.yml at 06:34), and #7835 was only a prospective warning, never a confirmed incident — the real fourth PR is #7772. Verified against this lane's dispatch receipts rather than recall, and the count is now stated as a floor rather than a census. A wrong enumeration inside a canonical carrier is the citation class DESIGN §3 names, so it gets the same bar as any other cited fact. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The merge brought in 52 commits of main, and the generated-artifact merge driver keeps this side's copy for generated paths, so the merged ci.yml was this branch's pre-merge bytes: it was missing main's new registered plans and stage0 modules in AUTHORED_CONFLICTS, the heal commit's --no-verify, several step ids, and a test-negation respelling. Regenerated from the authority rather than hand-merged. The output now differs from origin/main by exactly one line — the include-hidden-files key this branch adds — which is the check that the regeneration is correct rather than a plausible-looking text merge. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
# Conflicts: # dag/gunbc/plans/branch_merge_admission_model.dag
10 of 11 tasks
review 49102 is correct and this was my error. ci_workflow.dag declared needs: [build, regen, heal_generated_artifacts] with a job-level if, while the committed ci.yml carried neither -- an internally inconsistent PR that would have failed ci_yaml_parse_witness, since expected_ci_yml() serializes the model. The cause was mechanical rather than a decision: the gating work was in this worktree when the WIP auto-commit picked it up and pushed the .dag half, while the regenerated ci.yml was discarded by a local reset moments later. The two halves were split across a push boundary. The fix is removal, not regeneration. That change is not in this PR's scope and already exists as #7882, authored on main with its own witnesses and a proven RED control. Regenerating ci.yml here would have made this PR self-consistent by duplicating another PR's change, which is the worse resolution of the two. The three files are restored to cfea75a, the last head where this branch carried only the heal-artifact fix. Regen after the restore produces no ci.yml drift, which is the check that model and artifact now agree. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Contributor
Author
|
What happened. The cause was mechanical rather than a decision. That gating work was in the same worktree when the WIP auto-commit picked it up and pushed the Why removal rather than regeneration. The floor/heal gating is out of scope here and already exists as #7882, authored on The three files are restored to This PR is now the four files it was always meant to be: the derived — sent from loyal-ram-550 |
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.
Summary
The CI auto-heal job packages the corrected bytes of author-commit-required workflow files into
heal-author-commit-required/and uploads them, so an author can download the fix the GitHub App is structurally unable to push (it lacksworkflows:write). That artifact has been empty on every run since it landed.actions/upload-artifactdefaultsinclude-hidden-filestofalse, and@actions/glob'sexcludeHiddenFilesskips any entry whose basename starts with a dot before the descend step — so a dot-directory is never traversed and nothing beneath it is ever yielded. Every path this artifact can carry isauthor_commit_required_committed_artifact_paths(), which is.github/workflows/ci.ymland.github/workflows/falsifier.yml. The payload lives entirely under one hidden segment, so the upload could never publish anything.Measured, not reasoned. Run 31031996072 (PR #7857) emitted
HealAuthorCommitRequirednaming both paths — so the packaging block provably ran and populated the directory — and the very next step reportedNo files were found with the provided path: heal-author-commit-required.Priced. Because the handoff was always empty, every author re-ran regen locally instead: five separate times on 2026-08-05 alone (#7770 twice, #7791, #7835, #7857).
Why derived rather than a literal
trueci_heal_author_commit_artifact_includes_hidden_files()reads the flag off the same roster that determines the payload. Register a non-hidden author-commit path and it becomesfalseon its own; register another dotted one and it staystrue. The always-empty state is unwritable rather than re-checked (DESIGN §5, construction over validation). A literaltruewould be a second representation of a fact the roster already carries, and would go silently wrong the day the roster changes.Scope bound
This walls one step, not the class. Any upload step whose payload is entirely hidden is always-empty; the note carries a named dissolve-on for a lens deriving the flag for every upload step from its declared path population. The corpus's other two uploads (
release-bins.tgz,target/floor-component-receipt.json) carry no hidden segment and are unaffected — which is why this stayed local and invisible.Test plan
claim_batch --source-root dag --source-root src/v2 --entry dag/test/claim/ci_heal_job_witness_test.dag, four witnesses, all PASS:heal_author_commit_upload_publishes_its_hidden_payload— the delivery test. Red before this change (noinclude-hidden-fileskv existed, so the predicate returned false); green after. This is the gapheal_author_commit_upload_step_enrolledleft open: the step being present was already witnessed while the artifact it published was empty. Enrollment is not delivery.author_commit_required_paths_all_carry_a_hidden_segment— pins the premise the derivation rests on.hidden_segment_predicate_discriminates— RED control fixing both polarities on real corpus paths, including a mid-path dot-directory.heal_author_commit_upload_step_enrolled— unchanged, still green..github/workflows/ci.ymlregenerated viagenerated_artifact_gate main_wetand committed by me, since the App cannot push workflow paths — the same obligation this artifact exists to serve.Note on the
branch_merge_admission_model.dagimport fixAn earlier revision of this branch carried a restored-import fix for that module, because this PR touches
.github/workflows/ci.yml— a path outside the selectable universe — which forces compile-clean onto the whole-tree baseline rooted atwitness_layer_roots, where the module's stripped bare references stop resolving. Confirmed on this branch's own run (job 92433268875).That fix now lives on main and has been dropped from this diff. Main's version imports
gunbc.planandgunbc.plans.md_helpersonly; mine additionally importedstd.markdown, which is redundant — both of those modules importstd.markdownthemselves, so it arrives through declared edges rather than by pool coincidence. Main's version is sufficient by construction and is the single authority, so the merge takes it verbatim rather than forking a second variant of one fix.This diff is therefore back to the four files of the heal-artifact change.