An act's read again of its item asks when that item's Runs began, all of them past the cap (#724, record 0199)
Part 3 of 4 for #724. It does not close the issue. Part 1 merged as #751 and part 2 as #759; the bound on the pass's asking merged as #771 (record 0196).
It changes harness/reread.py. Whether that is a person's to approve is the Factory's rule to say, not this body's.
What changes
The read again before an act now asks when its item's Runs began, as the pass will, so that both readings carry began_at before the Kernel reads it in part 4.
harness/reread.py: hands the records portbegan={}andon, the item its Run is on. Nothing is kept: a Run keeps nothing from one read to the next.harness/read.pyis not changed: the read that builds a Task hands the Kernel no run record, so it asks no job. A test and a mutant hold that. An earlier head of this pull request had it ask too; that bought nothing and could refuse a Task over jobs nothing there reads.- The port's
on: a read of one item asks that item's Runs alone. Another item's Runs are not asked, say nothing of a beginning, and do not make the read less than whole. - Past the cap: a read of one item asks all of that item's Runs, however many. Under the cap of 50, an item with more Runs than that would never be read whole by an act again.
- It fails closed: a Run of the item whose jobs cannot be read still leaves the read not whole, and the act defers.
A record of its own
The uncapped read of one item contradicts record 0193's row "The cap", and 0193 foretold that both of the Harness's reads would ask. A Boundaries row decides nothing, so this part adds a record that amends those two rows. It is record 0199, the number the Factory Builder gave on 2026-11-12; main ends at 0198.
The exception to the cap, and leaving harness/read.py out, were my design, made to answer reviewers' findings. Neither was part of the Factory Builder's ruling of 2026-11-11, which named both reads and put them in a pull request of their own.
Said plainly
- The bound on one read: one request for each Run of the item that the list gives, at most the 1,000 GitHub serves of a workflow's runs.
- The type is not held apart: the test tells one item from another by its repo and by its number. An issue and a pull request share no number.
- Beside the pass's bound: main's
of(record 0196) bounds the pass's asking to the open pull requests' Runs.onis a read of one item and is handed withoutof; the port keeps both.
Checks
make checks: OK atd487f87e, which holds main ate0ac3ffb.tests.test_adapters,tests.test_steps,tests.test_controller: OK atd487f87e(240 tests). No other module was run at the desk: CI's verify is the answer.- This part's mutants, filtered: 17 of 17 killed at
d487f87e. No mutant is stale. - gitleaks: no leaks.
- Size as the Kernel counts a diff: 22,415 characters, 225 lines.
Reviews
Carbon, Muse, Astra, the Claude review and the security pass, each of this part alone against main. Every finding is fixed or answered.
| Reviewer | Head | Unread parts | Blocking | Major | Minor | Info |
|---|---|---|---|---|---|---|
| agy | 5f33725 | none | 0 | 0 | 0 | 0 |
| muse | 5f33725 | none | 0 | 1 | 0 | 0 |
| astra | d7992a9 | none | 0 | 1 | 0 | 0 |
| claude-review | d7992a9 | none | 0 | 0 | 2 | 2 |
| claude-security | d7992a9 | none | 0 | 0 | 1 | 2 |
| Finding | Severity | Criterion | Where | What | Disposition |
|---|---|---|---|---|---|
| muse#1 | major | PR-3 | tests/test_adapters.py:1415 | An on-read whose own item's jobs cannot be read is not held to leave the read not whole | fixed in 7271fb1a: the item's third Run's jobs answer 404, past the cap, and the read is not whole, with the mutant "a read of one item is whole whatever jobs of its Runs cannot be read" and a row in the record |
| astra#1 | major | PR-5 | tests/test_adapters.py:1424 | The uncapped-item boundary test does not cover completed Runs | fixed in 5f337256: the test feeds the item's Runs in progress and completed, with the mutant "one item's completed Runs are asked no further than the cap" |
| claude-review#1 | minor | none | docs/decisions/0193-when-a-run-began-is-read-from-its-jobs-trusted-only-once-a-job-is-in-progress-or-ran-and-kept-by-the-pass-records.md:29 | The record's Decision still says no more than 50 Runs are asked in one read; the code now asks one item's without a cap, and only a Boundaries row says so | fixed in 5f337256: the exception has its own record, which amends 0193's row on the cap for a read of one item; this change no longer touches 0193 |
| claude-review#2 | minor | PR-3 | tests/test_adapters.py:1424 | That one item's Runs are all asked past the cap is held only for Runs in progress: a mutant that counts the item's completed Runs toward the cap survives | fixed in 5f337256, as astra#1; another repo's Run in progress is fed too, with the mutant "another item's Run in progress leaves a read of one item not whole" |
| claude-review#3 | info | none | the change | Reviewer's note, no defect of the change: this read first unpacked the head over an older scratch folder named copy, before moving to a folder of its own | no defect of the change: three of the four Claude reads shared one scratch folder for about two minutes; each redid its results in a folder of its own, and no worktree was written |
| claude-review#4 | info | PR-3 | adapters/records.py:341 | That item's alone is held by its number only: a mutant that compares the number and not the repo or the type survives | fixed in 5f337256 for the repo and the number, each with a mutant; the type is not held apart, since an issue and a pull request share no number, and the record's row says repo and number |
| claude-security#1 | minor | PR-5 | adapters/records.py:352 | "All of them, past the cap" is held for one item's Runs in progress only: counting its completed Runs toward the cap survives | fixed in 5f337256, as astra#1 |
| claude-security#2 | info | none | the change | Not a defect of the change: this reader overwrote the shared scratchpad directory copy at about 16:40-16:41 CST; results taken from it then should be re-taken | no defect of the change: the shared scratch folder, as claude-review#3 |
| claude-security#3 | info | PR-5 | adapters/records.py:341 | "That item's alone" is held by number only: an on matched by number, whatever the repo or type, survives | fixed in 5f337256, as claude-review#4 |
Carbon and Muse read 5f337256, which holds the fixes for Astra's and the Claude reads' findings. No reader of mine has read 68149089, which holds the fix for Muse's finding and takes the reader's asking out, nor d487f87e, which merges main (the port now keeps main's of beside on) and renumbers the record.
On GitHub, at 7271fb1a:
| Reviewer | Where | What | Disposition |
|---|---|---|---|
| cubic, P2 | harness/read.py:149 | The reader's new read can come back not whole while the snapshot is told the records were whole | answered in 68149089: the reader no longer asks a job, so this pull request adds no way for its list to be not whole. The literal is main's, from before this change; a list read short already holds the Task to nothing recorded, by what the Lenses' records port is handed |
| cubic, P2 | adapters/records.py:341 | on scopes the jobs alone; another item's unreadable outcome can still leave the read not whole | declined: that is main's reading of outcomes, which this pull request does not change, and it fails closed. Scoping outcomes to one item is its own change |
| cubic, P3 | the record, line 3 | The Status points at a pull request it does not name | fixed in 68149089: the Status names the issue alone |
| Gemini, medium | adapters/records.py:352 | asked += 0 if on is not None else 1 should be an if | declined: equivalent, and three mutants are written on that line |
| CodeAnt, nitpick | adapters/records.py:352 | With on, every Run of the item is asked in turn, and a long history could pass the step's time limit | no change: true, and the record's row "What bounds it" says the bound is the list. A step that times out makes no act, so it fails closed. The most Runs any open pull request had on 2026-11-11 was 12 |
Codex found nothing. CodeRabbit: skipped, auto reviews are off on this repo.
š¤ Generated with Claude Code