A Pause the port could not read is named by which read it was and why (#688's dry runs 5 to 7, record 0191)

Closed
DO
6d ago
+404-40

This makes the approving act say which of the Pause's reads failed, and how. It is hardening, not a fix: the fault it would have named is already found and mended. The Factory Builder asked for it on 2026-11-10. It opens labelled factory:hands-off.

Why

Slice 1's dry runs 5, 6 and 7 on "AGENTS.md: say what is built; drop the model claim no check holds" (#688) each stopped in the approving act:

Dry runReleaseCodeWhat it told us
5v0.1.5facts_unreadableSome read failed
6v0.1.5facts_unreadableThe same, after the Maintainer added Variables: read to the machine user's token
7v0.1.6pause_unreadableThe Pause, by record 0187's naming; not which of its two reads, or how

The cause was the machine user's token: its Variables grant was waiting for the org's approval. Once the Maintainer approved it, dry run 8 passed the approve step, and donpetry-bot approved #688. A code that said "GitHub refused the list of the org's variables" would have pointed at the grant at dry run 5. Which of the two reads GitHub refused was never recorded, so that part is not known.

What changes

  • adapters/pause.py: explained hands back the same read as read, and beside it why each scope it could not read was unreadable, by a code of the port's own:

    PartWhen
    org_refusedGitHub answered the repo's list of the org's variables 403 or 404
    org_unnamedThat list was read whole and does not name FACTORY_PAUSED
    repo_refusedGitHub answered the repo's own variable 403
    org, repoAnything else of that scope's read

    A 403 that is a rate limit's, answered with no request left or a time to wait, is anything else, not a refusal; so is a refusal on the last request left, which the port cannot tell from one. A 404 is never a limit's. read is now explained(...)[0].

  • harness/reread.py: item takes name_pause. Asked, it raises a Pause the port could not read as a read of the pause stage, its part the first unreadable scope's, the org's before the repo's. It raises after the run records are found unread, and never while a scope pauses.

  • harness/approve_step.py asks by name, so the act records pause_<part>_unreadable through unread in harness/shared.py, as it records a snapshot's part.

  • Record 0191 amends 0179's row "The parts" and 0187's row "A part not as its schema holds it", for the Pause.

What does not change

  • The Controller: read hands back what it did, so the pass, the brake before a dispatch and the health file read the Pause as before, and pause_unreadable is still their one code.
  • Apply: it does not ask by name, and names its Pause in its own order.
  • A 404 on the repo's own variable is still a variable GitHub holds none of. The port cannot tell a 404 that hides the variable from one that says there is none.
  • The outcome's kind: only the approving act's reason is finer, so no Decision moves.
  • In the approving act, pause_unreadable is left for a Pause the port read whose value is no Pause, and for a fault of the read's own code.

One order does change, and the record says so: a Pause the port could not read is now raised by the read again, so the act names it before it holds the run records and the snapshot to their schemas, as it does any read that raised.

Checks it adds

  • tests/test_effects.py::test_a_pause_not_read_says_which_read_and_why: each part on both sides of its boundary, the order of the parts, and that the read beside them is read's.
  • tests/test_steps.py::test_the_item_is_read_again: asked by name, each part is raised at the pause stage; not asked, nothing is raised; a scope that pauses, a value that is no Pause and a Pause read clear raise nothing; unread run records are named first.
  • tests/test_steps.py::test_the_approver_holds_to_its_item and ::test_an_approving_run_finds_its_act_by_its_run_name_and_approves: the step asks by name, and ends pause_org_refused_unreadable on a refused list.
  • tests/test_steps.py::test_apply_reads_with_reads: apply asks for no part.
  • tests/test_approve.py::test_each_fault_of_the_act_is_classed_as_the_design_classes_it: the five codes.
  • Forty-three new mutants, each killed by a test its record row names, and seven older mutants of the two files reworded to the text as it now stands.

Checks

CheckResult
make checksOK
The seven test modules around the changeOK, 332 tests
Mutants of the six touched files, run alone on an export146 of 146 killed, at ff8e4f4
Full make verifyNot run locally, by the Maintainer's standing change of 2026-11-09; CI's verify is the answer
ReviewerHeadUnread partsBlockingMajorMinorInfo
agy (Carbon)aa0e10dpart 6 of 6: its run timed out0011
agy (Carbon)ec6c7a5none0420
museaa0e10dnone1100
astraec6c7a5none0100
astradd12cb2none0000
claude-reviewaa0e10dnone0021
claude-securityaa0e10dnone0000

The head is ff8e4f4. After aa0e10d the code changed twice, both in how the port tells a rate limit's 403 from a refusal: once for the Claude review's finding, read by Carbon and Astra at ec6c7a5, and once for Astra's, read clean by Astra at dd12cb2. Muse and the two Claude reads did not read those two changes. ff8e4f4 differs from dd12cb2 in cases, mutants and the record's boundary rows alone, so no reader read again after it.

FindingWhatDisposition
muse@aa0e10d #1 (blocking)The record was left as NNNN, so make checks failedfixed: the record is 0191
muse@aa0e10d #2 (major), claude-review@aa0e10d #1 (minor)The part of a list whose next page contradicts its count was held by no testfixed: cases wanting org, and the mutant "a next page that contradicts the count is named as a list that does not name the Pause"
claude-review@aa0e10d #2 (minor)A 403 that is GitHub's rate limit was named a refusal by its status alonefixed: a 403 with no request left or a time to wait is org or repo; the record's row says so
claude-review@aa0e10d #3 (info)unread's docstring said a part is a snapshot's alonefixed
agy@aa0e10d #1 (minor)No case told the order of two repos' parts from the parts' own orderfixed: two cases, and the mutant "the Pause's parts are told in an order of their own"
agy@aa0e10d #2 (info)A comment cited the placeholder NNNNfixed with the number
astra@ec6c7a5 #1 (major), agy@ec6c7a5 #5 (minor)The rate-limit exception also took an org list's 404 for a limitfixed: the exception is a 403's alone; three cases and two mutants
agy@ec6c7a5 #1 (major)The short-page case was masked by its countfixed: its count is 2 with one entry
agy@ec6c7a5 #2 (major)No case had the Pause named on a page whose next link contradicts the countfixed: two cases
agy@ec6c7a5 #3 (major)No mutant held the port's other unreadable returnsfixed: eight mutants, each killed by the cases already there and two added
agy@ec6c7a5 #4 (major)No mutant held the read again's own call of the portfixed: two mutants
agy@ec6c7a5 #6 (minor)No mutant of harness/approve.pyno change: only its docstring changed, and the code is formed by unread in harness/shared.py, which the mutant "a part is named for a snapshot's read alone" holds

šŸ¤– Generated with Claude Code