The approving act finds its Decision by the run-name's Intent, and its token may read a compare of two commits (#688's fourth dry run)
Slice 1's fourth dry run, on "AGENTS.md: say what is built; drop the model claim no check holds" (#688) at v0.1.4, reached converged/approval_due for the first time (pass 42869707264). Both approving Runs then ended superseded/not_the_decision (approve.yml 42870315808 and 42870857986). The Factory Builder asked for the fix, for a walk of the approving act's whole path on the live record's shape, and for anything else that walk found in the same pull request. It opens labelled factory:hands-off. No record: no rule changes.
What
| Where | What |
|---|---|
harness/read.py | decided compared the run-name's Intent, issue-approval, with approve.APPROVAL, the Decision's action issue_approval. Every approving Run was read as an admission, found none, and stopped. It now compares with kernel.terms.ISSUE_APPROVAL. The defect is as old as the approving act (#334); the approval branch of decided had no test and no mutant |
adapters/effects.py | The walk's find, and dry run 5's stop had it not been found. Approving._read refused any path holding two dots. The facts snapshot reads its diff, its files and a review's content by GET /repos/{repo}/compare/{base}...{head}, so the act's read again would have ended environment/snapshot_diff_unreadable on every Run. Two dots are now let through when the whole path is a compare of two full commit ids in the item's repo, and in nothing else. Decision 0055's rule stands as written: the token reads the item's repo, and no path climbs out |
tests/test_steps.py | The step runs through the real approve.job on a pass record that holds an admission and the act, by the name the pass dispatched, to the one write. Then each part of the run-name is moved in turn (Run id, content, Role, item, Intent): each finds no Decision and writes nothing. So does an act that names no content. The run step's reader, given the approving run-name, builds no Task: not_the_decision |
tests/test_adapters.py | The facts snapshot is read whole through Approving, from the canary's recorded sequence, and equals the plain read |
tests/test_approve.py | The compare's boundary cases: a branch's name; either commit's short or capital form, or forty characters that are no commit's; two dots or four; two dots and another character; a line's end; a path or a query after it; another repo's compare; a climb before it; and a token with no item |
tools/mutants.py | Seventeen mutants: the Intent comparison; each clause of the approval branch, and an act with no content; the reader's stop on an approving run-name; the compare refused; the compare's shape loosened nine ways; two dots unguarded. One existing mutant follows the guard's new shape |
The walk
A desk script wrapped a read-only gh api transport in effects.Approving for #688 and ran harness/reread.py's item, the schema checks, effects.approval_of and decide/still_wanted, as approve.job does. With the fix: 52 requests, none refused by Approving, every one answered 200; the facts pass facts/1; the content is the one decided; no approval is recovered; and on the pass's facts the Kernel still wants the approval. On the facts read now it holds, hands_off, since #688 was labelled again at 18:02Z.
Not verified, and not mine to verify: the walk ran on the Maintainer's gh login, not the machine user's token. The live Runs proved that token reads the runs and artifacts. Its reads of the Pause variables (actions_variables: read), and its one write (pull_requests: write), are untried.
About the stop code: approve_step already prints it, through harness/steps.run, as every step does (not_the_decision, in the job's log at 17:58:21Z). Only GitHub's annotation shows the exit code alone. Nothing is changed there.
Verification
make checks: ok.tests.test_approve,tests.test_steps,tests.test_adapters,tests.test_token_needs,tests.test_effects,tests.test_harness: ok, 278 tests.- Both new tests failed before the fixes: the step with
not_the_decision, the snapshot withUnansweredatpull.diff. - Mutants, run alone on an export: 92 of 92 at
923ca07, all ofharness/read.py's and of the approver's transport's; the compare's 12 of 12 again at0070092. None stale. - CI's
verifyis the answer, per the Maintainer's standing change of 2026-11-09.
| Reviewer | Head | Unread parts | Blocking | Major | Minor | Info |
|---|---|---|---|---|---|---|
| agy (Argon) | 7a2a103 | none | 0 | 0 | 0 | 0 |
| muse | 7a2a103 | none | 1 | 0 | 0 | 0 |
| claude-security | 7a2a103 | none | 0 | 0 | 6 | 3 |
| astra | 923ca07 | none | 0 | 1 | 0 | 0 |
| claude-review | 7a2a103 | none | 0 | 1 | 2 | 1 |
| Finding | What | Disposition |
|---|---|---|
| muse@7a2a103 #1 (blocking), claude-review@7a2a103 #1 (major), claude-security@7a2a103 #1 | The mutant "the approver's reads may climb out of the repo" named text the file no longer held: stale | fixed in 923ca07 |
| claude-security@7a2a103 #2, claude-review@7a2a103 #2 | An act issuing the approval with no content_id, which the schema allows, raised in decided in place of stopping | fixed in 923ca07: read with .get, with its case and mutant |
| claude-security@7a2a103 #3, claude-review@7a2a103 #3 | The reader's stop on an approving run-name is live for the first time, with no test or mutant | fixed in 923ca07: both |
| claude-security@7a2a103 #4, #5, #6 | The head's clause of the compare, its escaped dots and its whole match were unheld | fixed in 923ca07: a refused case and a mutant for each |
| claude-security@7a2a103 #7 (info) | apply_step now gets the act for an approving run-name; apply.admitted still refuses it, no_decision | no change: refused as before, one step later |
| claude-security@7a2a103 #8 (info) | Not in this diff: SHA.match with $ in adapters/pull.py and adapters/timeline.py takes a trailing newline | follow-up, reported to the Factory Builder; Approving refuses such a compare here |
| claude-security@7a2a103 #9 (info) | adapters/candidate.py's compare carries ?per_page=1, which Approving refuses | no change: the approving act does not reach it |
| claude-review@7a2a103 #4 (info) | The step test patches passes.find and reread.item; only the facts snapshot runs through a real Approving | no change: the walk on #688 ran every read of reread.item through Approving, 52 requests, none refused; the reviewer traced the rest by hand and found none refused |
| astra@923ca07 #1 (major, a test gap) | Either commit's clause loosened to any forty lowercase letters and digits survived | fixed in 0070092: a refused case and a mutant on each side. No re-read, by the Factory Builder's ruling on test gaps |
š¤ Generated with Claude Code