A path only a person decides is listed apart from a Lens's inputs, and a change to it needs a person's approval (record 0200)

Merged
DO
5d ago
+262-13

This makes a change to release/decision-points.json, the Maintainer's continue or stop, need a person's approval. Record 0194 said the check holds the name written in that file and not the hand, so a session could type his decision and the Factory could approve it. The Factory Builder ruled on 2026-11-11 that it be closed: "A change to the continue-or-stop file must need a person's approval, as a model-Lens input does; a model review alone isn't enough."

Model-Lens inputs and protected paths touched, on purpose: config/hooks.json, schemas/hooks.schema.json and tools/hooks/protected.py. The edits are pre-approved by the Maintainer 2026-11-08, ruled by the Factory Builder. By record 0052 the Factory never approves a change to these, so this pull request merges by the Maintainer's hand: it is kept small and alone, and is reported to the Factory Builder when it is green with its threads resolved.

Size. 31,852 characters and 422 lines by git diff; 34,596 characters as the Kernel counts a diff, with its lines numbered, against a budget of 60,000.

What changes

  • config/hooks.json gains a list, person_only: each path only a person decides. Its one row is release/decision-points.json. Every person-only path must be a protected path too, or the file is not read: so a session hook asks before an edit to it, and no Role's push may touch it.
  • tools/hooks/protected.py reads the list in its one reading of the file. What the Kernel is given as model_lens_inputs is the model Lens's own inputs and then the person-only paths, each once.
  • The Kernel changes by no line. A change to the file waits in model_lens_input_needs_person until a person with write access other than its author and its last pusher approves its head, as a change to harness/ does. In the merge queue without one it is taken out. A Role's push may not touch it.
  • schemas/hooks.schema.json describes the list. It does not require it, and a config/hooks.json with no such list names none: a pull request's base can be a commit from before this one.
  • Record 0200 says the list, why it is not rows of lens_inputs, and that the code model_lens_input_needs_person now also covers a path no Lens reads. The Factory Builder files the rename as a follow-up.

What it means for his decision

His own change to the file is not converged on his own approval, which is its author's, nor on the machine user's, which is nobody's. With no other person's approval it merges by his hand as the machine user, by record 0198, and that approval is the act that shows a person decided. A session may type the three fields on his word; it cannot bring them to main.

What is still not held, and the record says so

  • Whose approval. The rule is 0052's whole: any person with write access but the author and the last pusher. Another person with write could pass a decision typed in the Maintainer's name. Narrowing it to him alone is a change to the Kernel.
  • Where the decision is kept. A change that points tools/check_release.py at another file is judged by the model Lens alone.
  • From when. The Kernel is given the list by the Factory's own checkout, so the rule holds from the first Release cut after this merges.

Checks it adds

  • tests/test_decide.py::test_a_change_to_a_path_only_a_person_decides_waits_for_a_person: in each class, the model review first and then a person; the machine user's approval counts for none; the Maintainer's own change waits on his own and the machine user's approval and passes on another person's; in the queue it is taken out; a rename out of the path waits; with the path off the list it waits for no one; and no file beside it is caught.
  • tests/test_hooks.py::test_a_path_only_a_person_decides_is_read_with_the_lists_and_given_after_a_model_lenss_inputs: the reading, the order, each path once, nothing given where no model Lens's input is named, each path protected or the list not read, a file with no list read as before, and sixteen malformed lists refused.
  • One counter-example of the schema, a person-only path with no reason.
  • Twelve mutants, and one older mutant reworded to the text as it now stands.

Checks

CheckResult
make checksOK, at 39d28ff
tests.test_decide, tests.test_adapters, tests.test_release, the hooks' list testsOK, 462 tests, at 39d28ff
tests.test_simulator, tests.test_issue_simulator, on their fixed seedsOK, 56 tests, at the first head
Mutants of tools/hooks/protected.py, config/hooks.json and the schema, run alone on an export79 of 79 killed at 408f2b6, and the twelve of the person-only list at 40ab78a
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)d11cfcdnone of 20020
mused11cfcdnone0000
claude-reviewd11cfcdnone0011
claude-securityd11cfcdnone0012
astraceb148cnone0000
astra408f2b6none0100

Astra read the head after the others' findings were fixed. Since then the branch was rebuilt on main and its record renumbered from 0195 to 0200, with the two records it cites renumbered as they merged; nothing else changed, and no reader read again. Record 0199, another lane's, merged first, and main is merged in at 39d28ff, so the records run with no gap. Its merge is to be made under a Pause, to show the third box of #198: the Maintainer's own approval and the required check merge a fix with the Factory paused.

FindingWhatDisposition
security #1With no model Lens's input named, the person-only path kept the list non-empty, so the Kernel no longer deferredfixed: nothing is given then, and it defers as 0052 has it; a test and a mutant
agy #2, claude-review #2A path listed twice was given to the Kernel twicefixed, with a test and a mutant
agy #1What a person-only path is could be empty or no textfixed in the reading, with a mutant
claude-review #1, security #3The record said the Maintainer's own change is never converged; another person's approval converges itfixed: the record's row says what holds, "What is still not held" names it, and the test has the case
security #2The files that say where the decision is kept are not on the listnamed in the record as still not held
astra@408f2b6 #1Nothing showed a second, distinct person-only path reaching the Kernelfixed at 40ab78a: the test gives the Kernel's list for one, and a mutant that hands on only the first is killed; a gap in a test, so no reader read again

Review threads, two, both resolved

ThreadDisposition
cubic: a path on the person-only list alone is asked of nobody at the edit, and a push can touch it where no model Lens's input is namedfixed in ccb1341: a list that names a path not protected is not read; a test and a mutant
cubic: a test comment says person-only paths are still given when no model Lens is namedfixed in ccb1341: the comment said the first draft

šŸ¤– Generated with Claude Code