Skip to content

[MEDIUM] fix(verifier): repairs from the 1.16.0 sync review, and the sync lever - #5

Merged
fitziboy3-source merged 10 commits into
adam/releasefrom
adam/verifier-repairs
Nov 1, 2026
Merged

fitziboy3-source merged 10 commits into
adam/releasefrom
adam/verifier-repairs

Conversation

@fitziboy3-source

Copy link
Copy Markdown
Owner

Why

Pull request 4 merged the 1.16.0 sync into adam/release without the repairs that pull request 3's Codex review and live-gate runs produced. Without them the live gate cannot pass on the kit now installed: Claude Code 2.1.316 refuses the generated skill's write in the formal creation case, and the verifier refuses any PR that touches the fork's operator scripts as unmapped.

What changed

  • The two formal recipe cases run Claude with bypassPermissions inside their fixture, as the port's setup cases do; routed eligibility is judged by the selected harness.
  • The formal recipes read the generated skill's sync commit from its own record (labelled line or git diff base) and compare the whole token, abbreviated or full.
  • scripts/sync-port.sh does the deterministic part of the next port sync; it and the other operator scripts are registered as non-runtime paths; the cloud test asserts the installed skill matches the source; FORK.md patch 7 records the verifier's routing adaptation.

Scope

  • Verifier and operator tooling only. No shipped skill, agent or hook changes.
  • Leaves open, recorded on pull request 3: the formal creation case can pass without a real mutation, routed evidence publishes the request hash as transcriptHash, and --native-model duplicates the catalog's effort list.

Blast Radius

The verifier every fork PR runs, and the next sync. The kit on the lanes is unchanged by this PR.

Verification

  • bun run static, bun run typecheck, the verifier's tests (209 pass) and tests/cloud-setup-test.sh (all checks pass) on the head, bun 1.4.4.
  • The sync-record parser names the head in all six skills the live gate's sessions generated across three runs.
  • Codex read-only review: six rounds on pull request 3, one on this branch; see the review record comment.
  • Live gate: pending on the exact head. Runs 4 and 5 on pull request 3 passed all ten sessions with the same recipe changes on the port-only kit.

Risk tier: MEDIUM.

Where the work lives: pushed to adam/verifier-repairs on this fork, not merged, nothing deploys.

🤖 Generated with Claude Code

fitziboy3-source and others added 9 commits November 1, 2026 18:10
It reproduces this merge's deterministic resolutions on a fresh branch
and leaves only the judgment files conflicted.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
(cherry picked from commit 50b853e)
The live gate classifies every path a PR changes and refuses an unmapped
one; the operator script is not a shipped surface.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
(cherry picked from commit 1c6b00c)
The script swallowed every merge failure, so a refusal to overwrite local
edits still reached the unconditional manifest and provenance rewrites.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
(cherry picked from commit 1215fec)
…anch's policy

The verifier refused any PR touching front-door.py, cloud-setup.sh or the
cloud test as unmapped. The cloud test asserted a passive kit, which the
branch has not shipped since October 5; it now asserts the installed skill
matches the source byte for byte.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
(cherry picked from commit ba74600)
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
(cherry picked from commit d4efe19)
Claude Code 2.1.316 refuses a Write under any .claude/skills/ folder in
dontAsk mode, so the creation case could not land the generated skill and
the live gate failed. The port's setup cases already run this way for
the same reason.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
(cherry picked from commit 7bfdd73)
A Claude permission override on a shared recipe refused the routed run for
Codex too, which needs no override. The check is now a function with its
own test.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
(cherry picked from commit c19fcc6)
…git diff base

The live gate's sessions record the sync point three ways: a table cell,
a labelled line, or the base of the git diff command the skill tells the
reader to run. The check now takes the commit from that record and
compares the whole token as a prefix of the expected commit, so a stale
record fails even when prose names the new head, and a history line
naming the old commit does not fail a correct update. A test drives the
maintain assertion itself through both outcomes.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…e sync-commit line

The sync script no longer overwrites the fork's provenance files, refuses
a tag already merged, and leaves a file one side deleted for a person;
tests/sync-port-test.sh exercises it on a throwaway history and runs
with the static invariants. The two formal fixture prompts ask for one
line, "Sync commit: <full commit hash>", and the parser reads only that
label, so words before it and prose do not count and records naming
different commits count as none.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@fitziboy3-source

Copy link
Copy Markdown
Owner Author

Live evidence: Open Pstack

  • Candidate SHA: e12dbe3341a02cba4c71bdba4853725b409643da
  • Base SHA: 1d9fc25db06218ec13f06d16e578db551bbc7e13
  • Result: PASSED — named features verified
  • Self-test: completed in both harnesses
  • Planned sessions: 2 (limit 12)
  • Publisher revision: e12dbe3341a02cba4c71bdba4853725b409643da
  • Evidence-set SHA-256: 8eef7f4284a4fd156b73fb4ba0b34c50681967f17251165f6b0179915668849e
  • Installation (claude): CLI 2.1.316 (Claude Code), plugin 1.16.0-adam.20261101, candidate e12dbe3341a02cba4c71bdba4853725b409643da, tree 80d41f374266fa9d38c1f097ea87fcae190587617029e26d4f7cc8c598613739.
  • Installation (codex): CLI codex-cli 0.162.1, plugin 1.16.0-adam.20261101, candidate e12dbe3341a02cba4c71bdba4853725b409643da, tree 80d41f374266fa9d38c1f097ea87fcae190587617029e26d4f7cc8c598613739.
  • Observation (claude / project-skill): surface claude -p stream-json with the candidate plugin-dir; action recipe project-skill: doctor; result passed: doctor:exit-0 doctor:skill-loaded doctor:doctor-ran doctor:doctor-canonical-pass; 4 assertions machine-checked; transcript SHA-256 99de36b1a2d1863726ae8446eee079fa734491241d9d3bf5922e12c9b26d9fc6; artifact-set SHA-256 2036a0fa8fcc966236ff2270b0c859f246313b5e9e066a852e58bc52d5c722f3.
  • Observation (codex / project-skill): surface codex exec --json with the run-owned candidate installation; action recipe project-skill: doctor; result passed: doctor:exit-0 doctor:skill-loaded doctor:doctor-ran doctor:doctor-canonical-pass; 4 assertions machine-checked; transcript SHA-256 da3d53d2586f8504a8c61d45467ffcc5368e5222ba3fad163e24b9cea1300404; artifact-set SHA-256 915302ff0717ccf7f1d767d850ba5519cfb2de08a2d6f82c7914169ac274ab27.
  • Policy: every new head requires a fresh run; this verifier does not change PR state, queue, or merge.

… own git identity

"Sync commit: <hash>^" named a different commit than it declared and
passed. The record's value must now be a bare hash, optionally quoted,
followed by nothing but punctuation or prose. The sync-port test sets
user.name and user.email inside its throwaway repository, so the merge
under test does not depend on the runner's git identity.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@fitziboy3-source

Copy link
Copy Markdown
Owner Author

Live evidence: Open Pstack

  • Candidate SHA: 8775340519d81079072caf592154757e8c689b4f
  • Base SHA: 1d9fc25db06218ec13f06d16e578db551bbc7e13
  • Result: PASSED — named features verified
  • Self-test: completed in both harnesses
  • Planned sessions: 2 (limit 12)
  • Publisher revision: 8775340519d81079072caf592154757e8c689b4f
  • Evidence-set SHA-256: 99e77828946fef04199eecfded0f720799fe6ed0b5e42e5d8a8bc13c3f4c8557
  • Installation (claude): CLI 2.1.316 (Claude Code), plugin 1.16.0-adam.20261101, candidate 8775340519d81079072caf592154757e8c689b4f, tree 80d41f374266fa9d38c1f097ea87fcae190587617029e26d4f7cc8c598613739.
  • Installation (codex): CLI codex-cli 0.162.1, plugin 1.16.0-adam.20261101, candidate 8775340519d81079072caf592154757e8c689b4f, tree 80d41f374266fa9d38c1f097ea87fcae190587617029e26d4f7cc8c598613739.
  • Observation (claude / project-skill): surface claude -p stream-json with the candidate plugin-dir; action recipe project-skill: doctor; result passed: doctor:exit-0 doctor:skill-loaded doctor:doctor-ran doctor:doctor-canonical-pass; 4 assertions machine-checked; transcript SHA-256 19bd89b4ff018da0801d24ee1b7217d8c0ea3f37591f76316ddb25349e6f0609; artifact-set SHA-256 c803d29aeb2725441f28082b336c4d70de260a6c8d4f6fcae2e23e6f08f1744c.
  • Observation (codex / project-skill): surface codex exec --json with the run-owned candidate installation; action recipe project-skill: doctor; result passed: doctor:exit-0 doctor:skill-loaded doctor:doctor-ran doctor:doctor-canonical-pass; 4 assertions machine-checked; transcript SHA-256 5196e61e897b62f7296763c0bcef916fd0a72aebbd41139d2c33da47bf39d9b0; artifact-set SHA-256 84b84164bcf172640824e2178ed026bc63967fc7783be0a6834daf49260383d9.
  • Policy: every new head requires a fresh run; this verifier does not change PR state, queue, or merge.

@fitziboy3-source

Copy link
Copy Markdown
Owner Author

Review record (MEDIUM tier)

Who reviewed. Codex (gpt-6-astra), read-only through cross-review-run: six rounds on pull request 3 and three on this one. Each live gate ran as a fresh verifier agent. Claude Fable built the change.

Pull request 3, round 1 (the fork's whole delta over port v1.16.0): REPAIR FIRST, seven findings. Four were in scope and are fixed here: the sync script swallowed every merge failure; the fork's operator scripts were unmapped in the verifier registry; the cloud test asserted a passive kit; FORK.md omitted the verifier's routing adaptation. Three predate this work (added in pull request 2 on October 19) and stay deferred, recorded on pull request 3 because Issues are disabled on this repository: the formal creation case can pass without a real mutation, routed evidence publishes the request hash as transcriptHash, and --native-model duplicates the catalog's effort list.

Rounds 2 to 6 on pull request 3, rounds 1 to 3 here. One or two findings each, each fixed with a test, until round 3 here returned READY on 8775340. The thread was the generated skill's sync commit: Claude Code 2.1.316 refused the generated skill's write (the formal cases now run with bypassPermissions inside their fixture, as the port's setup cases do, and routed eligibility is judged by the selected harness); the sessions write the abbreviated hash git prints; a loose match accepted prose, a labelled-field parser rejected Codex's git diff form, and banning the old hash was a tactical patch. The settled design: both fixture prompts ask for one line, Sync commit: <full commit hash>; the parser reads only that label, requires a bare hash (a revision such as <hash>^ counts as none), treats records naming different commits as none, and a test drives the maintain assertion through a failing and a passing skill. The sync script no longer overwrites the fork's provenance files, refuses an already-merged tag, leaves a file one side deleted for a person, and has its own test on a throwaway history, run from bun run static.

Live gate. On pull request 3, runs 4 and 5 passed all ten sessions in both harnesses (hook:startup, skill-invocation:poteto-mode, both formal skills, project-skill) with the same recipe repairs on the port-only kit. On this pull request the verifier's scope rule allows only project-skill, because no shipped kit file changes; runs 8 and 9 passed it in both harnesses, the last on the exact head 8775340 (live-gate status set). The recipe prompt that mandates the sync-commit line has no live exercise yet; the next pull request that touches a formal skill will run it.

Risk tier. MEDIUM. The affected surface is the verifier every fork PR runs and the next sync. The kit on the lanes is unchanged by this PR.

Where the work lives: pushed to adam/verifier-repairs, not merged, nothing deploys.

@fitziboy3-source
fitziboy3-source merged commit 64ed336 into adam/release Nov 1, 2026
3 checks passed
@fitziboy3-source

Copy link
Copy Markdown
Owner Author

Merged and rolled out

Where the work lives: merged into adam/release as 64ed336 on Adam's word, with the review record and the live gate (8775340) pinned to the reviewed head. The fork's main now mirrors the port's main at b4c5e3d.

VERIFIED after merge: all five lanes (~/.claude, ~/.claude-2, ~/.claude-3, ~/.codex, ~/.codex-2) install pstack@open-pstack at 1.16.0-adam.20261101, and each installed tree matches adam/release's plugins/pstack file for file (678 files, digest f31a64cee569). The plugin payload is the same git tree as the one PR 4 installed (c9205a73472b), so the three lanes PR 4 updated did not change; ~/.claude-2 and ~/.codex-2 came up from the October build in this rollout. Both Codex marketplace snapshots are at 64ed336. No deployment outside this Mac.

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.

1 participant