A push past a Lens's review asks it again while its thread gates, and a thread its later review dropped is the Kernel's to resolve (#833, record 0212)
Closes #833.
What was wrong
Since the reviewer's account was listed as a Lens, each blocking or major Finding it posts holds its pull request until a person resolves the thread. Two things kept it there.
- Nothing reviewed again. On "Trial P-66, a queued pull request: once the base it waits on moves into conflict with it, the queue takes it out unmerged (#812)" (#847) the reviewer asked for a file that reached the head by a merge of
main. That push left the diff as it was, so the content id did not change, the Kernel read the review as done, and the pull request stayedaddressing,findings_openfor two hours. A push that changes the diff was already reviewed again. - Nothing closed the thread. A later review by the same Lens already released a pull request held by a thread that was gone (decision 0075), and closed an expired Run's thread (decision 0125). It said nothing of a thread that was still there.
What this changes
- A push that keeps the diff asks the review again. While a required model Lens's own thread gates and still stands, a push after that Lens's last review of this content began answers
verifying,awaiting_review. The Run ceiling bounds it. - The port says which thread a later review dropped.
adapters/applied.pymarks a Lens's threaddroppedwhere every Run that recorded it is one a later review of that Lens supersedes, by decision 0075's own matching: the thread's id in the Runs' records, never a comment's text.facts/1gains the optional word. - The Kernel reads it only for a newer head.
kernel/derive.pytakes the word only where the thread opened before the last push and the run records hold a review by that Lens that succeeded on this content and began after that push. Then the thread is the Kernel'sresolve_threadact,adjudicating,resolve_due, as an expired Run's thread is, and the queue App's resolution or a person's closes it. - What is never closed this way: a security Finding's thread, one whose category the Kernel does not know, and one its Lens last ruled rejected. The Kernel also never names a security Finding's thread in the act, whatever ruling the facts carry.
kernel/gate.pyandharness/are untouched. The gate already closes a Lens's thread the queue App resolved once it carries thesupersededmark, which decision 0125 put there.- Record 0212 rules it, with the slice-1 rule for carrying the act out by hand, by the Factory Builder's ruling of 2026-11-13; L-17 and the design each gain a sentence.
What it does not do
- The thread still has to be resolved on GitHub. Main's ruleset requires every thread resolved, and the queue refuses an enqueue with one open (decision 0089). The queue App that carries out
resolve_threadis not built ("A narrow queue App, combsmith-sf-queue, enqueues and dequeues with contents and pull-request write alone (P-2)", #381), so the pass records the actnot_built. Until it exists a session may resolve the thread by hand, only where the Controller's check on the current head saysresolve_duenaming it, never a security Finding's or an expired Run's, and with a comment naming the pass. - No check holds the hand's part. A test holds what the act may name. Nothing notices a session that resolves a thread the act did not name: "No check holds the hand's part of resolving a Lens's thread in slice 1: nothing notices a resolution the Kernel's act did not name" (#854). Every reader raised this; the record and L-17 say it plainly.
- A later review's silence is no ruling. The review is given none of its Lens's open Findings: "A re-review is given none of its Lens's open Findings, so a thread closes on the silence of one more review" (#856), which also holds
kernel/gate.py's docstring, a protected file's. - The simulator and the explorer cover neither rule: "The simulator and the transitions explorer never push a head that leaves the diff as it was, and never say a thread dropped" (#855). Their every push is new content.
sim/world.pyis protected and the change is at the review budget. - It reaches the live pass only at the next cut.
Checked at this desk
At ae636efa, the head: tests.test_decide, tests.test_adapters, tests.test_transitions, tests.test_controller, tests.test_harness, tests.test_steps, tests.test_reconcile and tests.test_effects, 609 tests; make checks, every check ok but the record's number, which waits on record 0211; the change's mutants and the ones on the rules it touches, 34 of 34 killed; no mutant stale; gitleaks, no leaks. The simulator on 40 seeds, at f675f05a: 46 of 49 pass, and the three that fail are the seeds' coverage tests, which fail the same way on main at 40 seeds and pass here on the full set. 58,598 characters as the Kernel counts, from a local diff. CI verifies.
Reviews
Two rounds. The first read 3e9204b2; the change was reworked for it, and Kimi and Astra read the rework, f675f05a. No reader has read ae636efa, the head, which answers the second round: two Run-time guards, one more case in each test and two mutants swapped.
- Kimi: 2 findings, 1 taken. Both on its second read; its first was clean.
- Astra: 2 findings on each read, all four taken.
- Claude review and Claude security read
3e9204b2. A second Claude security read, off675f05a, had not answered when this opened: what it finds I answer in a comment. - Carbon: not run (quota). Owed.
- Muse: not run (quota). Owed.
First round, of 3e9204b2
| Reviewer | Head | Unread parts | Blocking | Major | Minor | Info |
|---|---|---|---|---|---|---|
| kimi | 3e9204b | none | 0 | 0 | 0 | 0 |
| astra | 3e9204b | none | 0 | 2 | 0 | 0 |
| claude-review | 3e9204b | none | 0 | 1 | 5 | 2 |
| claude-security | 3e9204b | none | 0 | 1 | 6 | 0 |
| Finding | Severity | Criterion | Where | What | Disposition |
|---|---|---|---|---|---|
| astra#1 | major | PR-5 | docs/lessons.md:104 | The named check does not hold the hand-resolution security restriction | fixed in f675f05a: the Kernel never names a security Finding's thread in the act, whatever ruling the facts carry, and the invariant test tries security with an accepted ruling. What a hand does with the act is held by no check: issue 854, named in the record |
| astra#2 | major | PR-5 | kernel/derive.py:125 | The dropped-thread path lacks an unreadable-review-time boundary mutant | fixed in f675f05a: a review whose start is no time bears nothing out, tried in the test, with the mutant "a review at no time that can be read bears a dropped thread out" |
| claude-review#1 | major | PR-5 | docs/lessons.md:104 | L-17's new rule on a session's hand act names a test that holds what the Kernel names, not what the session does | so: the record and L-17 now say a test holds what the act may name and no check holds the hand's part, which is issue 854 |
| claude-review#2 | minor | none | docs/architecture/software-factory-design.md:552 | The design's new sentence says a dropped thread asks the review again; the code asks for a thread that still stands | fixed in f675f05a: only a category the Kernel knows, and not security, is read as dropped; resolved by the queue App, such a thread is thread_unreadable as on main, tried for a null category and an unknown word |
| claude-review#3 | minor | none | docs/lessons.md:102 | L-17's Rule line, and gate.py's closed() docstring, still say a Lens's thread closes only in the three older ways | fixed in f675f05a for all three: a run list that is none, a record that is none and a start that is no time are each tried; the last has a mutant, the first two are held by the test alone, to keep the change inside the review budget |
| claude-review#4 | minor | PR-3 | kernel/derive.py:121 | A dropped thread whose category cannot be read is taken as no security Finding, so the queue App's resolution closes it | not in this change: issue 855 holds the simulator, the explorer, tests/support.py and the example, and the record's row "The fake" says so. sim/world.py and tests/test_simulator.py are protected, and the change is at the review budget |
| claude-review#5 | minor | PR-3 | kernel/derive.py:121 | Three guards of dropped_by_review are held by no test: with each removed the Kernel raises where it now defers | fixed in f675f05a: the design's sentence names the thread a review of a newer head no longer records, and no longer says such a thread asks a review |
| claude-review#6 | minor | none | schemas/facts.schema.json:358 | Neither new rule is under the seeded simulation or the transitions explorer, and the new fact is in the schema alone | fixed in f675f05a for L-17, which says the queue App's resolution also closes such a thread; kernel/gate.py's docstring is a protected file's and is left, in issue 856 |
| claude-review#7 | info | none | docs/decisions/0212-a-push-past-a-lenss-review-asks-it-again-while-its-thread-gates-and-a-thread-its-later-review-dropped-is-the-kernels-to-resolve.md:1 | Record number 0212 is also the number on the pushed head of the open pull request 816 | not this change's: _wanted_run admits the first required Lens whatever the evidence names. The two-Lens case is no longer in the test, to keep the change inside the review budget; the configuration requires one. Noted in issue 855 |
| claude-review#8 | info | PR-3 | tests/test_decide.py:1170 | The two-Lens case asserts the evidence only: the Decision admits the other Lens, and the item runs to the Run ceiling | so: the sealed ledger's record, pull request 816, moves to 0220 by the Factory Builder's ruling, renumbered on its branch and pushed after this opens |
| claude-security#1 | major | PR-5 | docs/lessons.md:104 | The slice-1 rule for a session resolving a thread by hand names a test that holds what the act names, not what a session does | so: the record and L-17 now say no check holds the hand's part; issue 854 asks for one, and notes that the comment renews the Run ceiling |
| claude-security#2 | minor | PR-3 | adapters/applied.py:194 | A thread its Lens ruled rejected after the review that dropped it is still said dropped, and the Kernel resolves it | fixed in f675f05a: a thread is read as dropped only where it opened before the last push and the review began after that push, so a Run that failed and is tried again on the same head drops nothing. That a re-review is given none of its Lens's open Findings is issue 856, named in the record |
| claude-security#3 | minor | none | docs/decisions/0212-a-push-past-a-lenss-review-asks-it-again-while-its-thread-gates-and-a-thread-its-later-review-dropped-is-the-kernels-to-resolve.md:11 | A thread is closed on the silence of one more independent sample, with no push and no change: a failed Run's thread is dropped by its own retry | fixed in f675f05a, the safe way: a thread whose Lens's latest ruling is rejected is never read as dropped, whichever came later |
| claude-security#4 | minor | PR-5 | docs/decisions/0212-a-push-past-a-lenss-review-asks-it-again-while-its-thread-gates-and-a-thread-its-later-review-dropped-is-the-kernels-to-resolve.md:30 | The record's last boundary says the act names only an accepted or a dropped thread of no security Finding; it also names an expired Run's thread, whose category cannot be read | fixed in f675f05a: as claude-review#2 |
| claude-security#5 | minor | PR-3 | kernel/derive.py:121 | A dropped thread whose category cannot be read is closed by the queue App's resolution, where a security Finding's is not | fixed in f675f05a: a push at no time that can be read asks no review and bears nothing out, tried in both rules, each with a mutant. A push dated later than every Run still asks a review a pass, to the Run ceiling |
| claude-security#6 | minor | PR-3 | kernel/derive.py:162 | pushed_past_review compares last_push_at without reading it as a time: one that is none raises, and one later than every Run admits a review at each pass | fixed in f675f05a: the record says a hand never resolves an expired Run's thread, whose category cannot be read, and the invariant test tries the expired case |
| claude-security#7 | minor | none | sim/world.py:941 | The simulator's world never says a thread dropped, and its own check still holds resolve_thread to an accepted ruling | not in this change: issue 855, as claude-review#4 |
Second round, of f675f05a
| Reviewer | Head | Unread parts | Blocking | Major | Minor | Info |
|---|---|---|---|---|---|---|
| kimi | f675f05 | none | 0 | 2 | 0 | 0 |
| astra | f675f05 | none | 0 | 2 | 0 | 0 |
| Finding | Severity | Criterion | Where | What | Disposition |
|---|---|---|---|---|---|
| kimi#1 | major | PR-5 | docs/decisions/0212-a-push-past-a-lenss-review-asks-it-again-while-its-thread-gates-and-a-thread-its-later-review-dropped-is-the-kernels-to-resolve.md:16 | The new manual-resolve permission explicitly leaves its write preconditions unchecked | so: the Factory Builder ruled the slice-1 rule and the test that holds what the act may name. The record and L-17 say no check holds the hand's part, and issue 854 asks for one. Not taken as a change here |
| kimi#2 | major | PR-3 | kernel/derive.py:125 | The completed-Run condition for a dropped thread has no killing test or mutant | fixed in 73828919: a review still running bears nothing out, tried in the test, with the mutant "a review still running bears a dropped thread out" |
| astra#1 | major | PR-3 | kernel/derive.py:165 | Unreadable Run start times are compared as valid timestamps | fixed in ae636efa: a Run whose start is no time answers no push, tried in the test, with the mutant "a review at no time that can be read answers the push". A start that is none is the run records', which the Kernel sorts before this, as on main |
| astra#2 | major | PR-5 | tests/test_decide.py:1304 | The resolve-thread oracle accepts unreadable facts on an accepted ruling | fixed in ae636efa: the matrix asserts thread_unreadable for a word that is no yes or no, whatever the ruling. The mutant the finding names was run by hand and fails that test; it is not added to the list, to stay inside the review budget |
š¤ Generated with Claude Code