The reader names the read GitHub did not answer, and the snapshot its part, by codes of their own (#688's dry run)

Merged
DO
8d ago
+346-54

The dry run of "Epic: slice 1, first approval. A Release on next reviews a pull request to this repo, and the machine user's approval stands" (#182) stopped on #688. Both of the Factory's review Runs, sf-run 42710190930 and 42711108661, stopped in harness.read with facts_unreadable, before the Engine, so no model was called. That code is the one for every read the reader makes, so the Run could not say which read GitHub did not answer. This change makes the Harness name it. The Factory Builder approved it on 2026-11-09, for v0.1.2. It opens labelled factory:hands-off.

#693's files show in this diff until #693 lands. By the Factory Builder's ruling, this branch merges #693's queued head (f920ca6d, record 0178) as well as main, so that record 0179 follows 0178 with no gap and this pull request queues straight behind #693. Once #693 merges, the squash leaves this change's own files alone. If #693 is ejected or changed, this one holds.

What

WhereWhat
harness/shared.pystaged(stage) tells an Unanswered the stage it came from. unread(err) gives its code: <stage>_unreadable, a snapshot's snapshot_<part>_unreadable, or facts_unreadable for a read no stage named. Each is a code of the Harness's own, never GitHub's text. Each matches outcome/1's reason pattern ^[a-z_]+$, which is why the part joins with _ and not :
harness/read.pyEach read stops the step with its own code: release_refs, pass_record, run_records, lens_records, opened_records, snapshot (with its part), pull, change or issue_texts. A pull request that moved while it was read stops pull_moved, kept apart
adapters/facts.pyThe snapshot names the part GitHub did not answer: pull, base, push, diff, paths, timeline, reviews, writer, labels, checks, acceptance, threads, edits, unchanged or schema. The reads move into _snapshot unchanged, line for line, so no existing mutant moved. A Moved is given no part
harness/reread.py, harness/apply_step.py, harness/approve.pyThe read again names run_records, lens_records, opened_records, snapshot and pause. Apply and the approving act record that code as the outcome's reason, where they recorded facts_unreadable
Teststest_the_reader_fails_closed gains one case per code, each a stand-in answering that read with an Unanswered, a moved pull request among them; test_a_snapshot_names_the_part_github_did_not_answer covers every part; test_the_item_is_read_again covers each stage of the read again; the apply and approving-act tests cover a named stage
tools/mutants.py18 new mutants (dropping each part and stage, swapping two, losing the fallback), and four repointed at the moved text
harness/AGENTS.mdThe reader's rule row adds "names a read not answered" (the file is at its 40 lines)
harness/approve.pyIts docstring names the stage's code

Record 0179 decides the codes, their stages and parts, and amends 0110's row "An issue's text, live", whose short read stopped as facts_unreadable and now stops as issue_texts_unreadable. Nothing reads a reason beyond the outcome's kind, so no Decision moves.

Which read it likely is

Replayed through gh api as the Maintainer, every read the reader makes for Run 2 answered 200: 34 requests, reads completed whole. So the read that fails is one the reader App's token is refused, or a check of the transport's own. The installation is ruled out: trials/P-39.reach.json has the reader's GET /installation/repositories answering this repo.

These reads are ruled out by trials recorded as combsmith-sf-reader:

ReadTrialAnswer
collaborators/{u}/permission, activityP-47, on the canary200 Ɨ10
GraphQL lastEditedAt, userContentEditsP-58.reader200
GraphQL mergeQueueEntry, reviewThreads, closingIssuesReferences, timelineItemsP-8 and P-32, on this repo200, no errors

No trial yet has the reader make these, the suspects left:

  • issues/{n}/timeline and pulls/{n}/reviews (REST);
  • branches/{base};
  • the two compares;
  • the run lists of run.yml and approve.yml, with their outcome and applied artifact downloads;
  • the poller's runs and the pass artifact.

A Moved twice in a row is the remaining alternative. With this change, the next Run names which read it is, as <read>_unreadable or pull_moved.

Verification

  • CI's verify is the answer, per the Maintainer's standing change of 2026-11-09.
  • make checks: ok.
  • Mutants, run alone on an export: the six changed modules' 141 at 81bc528. 140 were killed; the survivor, "a read no stage named is named one", was fixed in 95572c0 and killed there, along with the reader fallback's repointed mutant. The approving act's mutant, repointed at its one test in 08be5d2, is killed.

Reviews (each fix in a later commit on this branch):

ReviewerHeadUnread partsBlockingMajorMinorInfo
agy5dc6cc9none0100
muse5dc6cc9none0100
claude-review5dc6cc9none0020
claude-security5dc6cc9none0000
agy81bc528none0000
muse81bc528none0100
claude-security81bc528none0000
claude-review95572c0none0011
claude-reviewf4faa7bnone0010
FindingSeverityWhereWhatDisposition
agy@5dc6cc9 #1, muse@5dc6cc9 #1, claude-review@5dc6cc9 #1major, major, minoradapters/facts.pyThe writer, labels, second push and schema parts were held by no testfixed in 81bc528: each is tested, and each has a mutant
claude-review@5dc6cc9 #2minorharness/read.pyThe step's facts_unreadable fallback could not be reached, and no test held itfixed in 81bc528 and 95572c0: an accounts_of case reaches it, and it takes its code from shared.unread
muse@81bc528 #1majortools/mutants.py"a read no stage named is named one" survived its named testfixed in 95572c0: the reader's fallback goes through shared.unread; killed
claude-review@95572c0 #1minordocs/decisions/0110-…0110 still named facts_unreadable for a short issue readfixed in f4faa7b: record 0177 amends it
claude-review@95572c0 #2infoharness/approve.pyIts docstring named facts_unreadable for every readfixed in f4faa7b
claude-review@f4faa7b #1minordocs/decisions/0177-…The last boundary named apply's test and mutant, not the approving act'sfixed in 08be5d2

šŸ¤– Generated with Claude Code