Each step's token holds every read its code can reach: apply asks for the Pause, and a sweep walks every step (#688's third dry run)
Slice 1's third dry run, on "AGENTS.md: say what is built; drop the model claim no check holds" (#688) at v0.1.3, got past the reader and the model. Then it stopped in apply as pause_unreadable, twice, on sf-runs 42755831262 and 42756232595. harness/apply.py's READS asked for no actions_variables, which the read again's Pause takes, though config/apps.json gives it to the review App. The Factory Builder asked for this fix and for a sweep, so that the next dry run does not stop at the next such gap. It opens labelled factory:hands-off.
#705's files show in this diff until #705 lands. By the Factory Builder's ruling, this branch merges #705's queued head (16d40156, record 0183) as well as main, so that record 0184 follows 0183 with no gap and this pull request queues straight behind #705. If #705 is ejected or changed, this one holds.
What
| Where | What |
|---|---|
harness/apply.py | READS adds read:pass.pause |
tests/test_token_needs.py | For each step that mints a token (the reader, apply's read again, the queue App's act, the pass, and the pass's writes step adapters/effects.py), the test walks the step's code statically. It goes from the step's entry, through every function and class of adapters/, the Harness and the Controller the step calls, to each function that builds a request of GitHub's itself (48 of them). Each is classed by the permissions and levels its requests take, with how many requests it builds. The step's token, minted from config/apps.json as the step mints it, must hold each permission at its level. A function the walk reaches that builds a request no table names fails the test. So does one more request in a function already classed. The walk follows dotted imports too. The writes step's mint is observed at identity.from_key; the other steps' mints are pinned as asked by their own step tests. The approving act mints no App's token, and its 23 requests, through the machine user's token, are pinned |
tests/test_apply.py | apply's granted permissions now include actions_variables: read |
docs/decisions/0184-⦠| The record |
tools/mutants.py | Mutants for the reader, apply and the pass, each dropping a need its reads take, plus the effects step's existing Pause mutant, now also killed here |
What the sweep found, beyond apply's Pause:
- The queue App's act reads the pull request (
queue.pull,threads,queued) with nopull_requestsgrant, and the App holds none. - The queue App is not live (
accounts.queueis null). By the Factory Builder's ruling, that gap is filed as "queue_step mints no pull_requests read, and the queue App holds none" (#711), not fixed here. - The test holds it as owed, exactly, while
accounts.queueis null. It fails once the queue App is wired without the fix. - Every other live step's token covers every request its code can reach.
Verification
make checks: ok.tests.test_token_needs,tests.test_apply,tests.test_steps,tests.test_effects: ok, 101 tests.- Mutants run alone on an export: 148 of 148 at
14a52d5, all ofharness/apply.py's and the sweep's. The sweep's four were killed again at4fd56af. - 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) | 14a52d5 | none | 0 | 1 | 1 | 0 |
| muse | 14a52d5 | none | 0 | 2 | 1 | 0 |
| claude-review | 14a52d5 | none | 0 | 0 | 3 | 0 |
| claude-security | 14a52d5 | none | 0 | 0 | 2 | 2 |
| claude-review | 3e4667b | none | 0 | 0 | 1 | 1 |
| claude-security | 3e4667b | none | 0 | 0 | 1 | 1 |
| astra | 3e4667b | none | 0 | 3 | 0 | 0 |
| Finding | What | Disposition |
|---|---|---|
| muse@14a52d5 #2, claude-review@14a52d5 #2, claude-security@14a52d5 #1 | The pass's writes step, adapters/effects.py, was not walked | fixed in 3e4667b: a step, with its mutant |
| muse@14a52d5 #1 (major) | A read added inside an adapters function was invisible | fixed in 3e4667b: the walk goes into adapters/, to every function that builds a request |
| claude-review@14a52d5 #1, agy@14a52d5 #2 | A raw request outside adapters/, a call through an imported class, or a submodule import, was skipped | fixed in 3e4667b and 4fd56af |
| agy@14a52d5 #1 (major), claude-security@14a52d5 #3 | A write was held at a read's level | fixed in 3e4667b: levels are held; queue.resolve keeps config/apps.json's contents: write |
| claude-review@14a52d5 #3, muse@14a52d5 #3, claude-security@14a52d5 #4 | The no-mint test looked for from_key alone | fixed in 3e4667b, and pinned in 4fd56af to the act's whole set of requests |
| claude-security@14a52d5 #2 | facts' checks read takes actions, for each suite's workflow run | fixed in 3e4667b |
| claude-review@3e4667b #1 | The activity reads take contents, as decision 0048 says | fixed in 4fd56af |
| claude-review@3e4667b #2 (info) | A duplicate of the existing effects Pause mutant | fixed in 4fd56af: the existing one names this test too |
| claude-security@3e4667b #1, #2 | seen._check_runs reads the pull request; the approving act's requests unpinned | fixed in 4fd56af |
| astra@3e4667b #1, #2, #3 (major) | A dotted import was skipped; a new request inside a classed function was unseen; the writes step's mint was worked out, not observed | fixed in 4fd56af: dotted imports followed, with a boundary test; each function's request count pinned; the mint observed |
š¤ Generated with Claude Code