A change that touches only paths a person decides waits under its own code, path_needs_person (#746, record 0207)

Merged
DO
4d ago
+373-23

A change that touches only paths a person decides now waits under its own code, path_needs_person, and the feed asks a person under the path's own question. Until now it waited as model_lens_input_needs_person, though no Lens reads such a path: record 0200 said so and left the name for a change of its own. Closes #746.

Model-Lens inputs touched: none. tools/hooks/protected.py and config/hooks.json are as they were, so the change goes by the Factory's route.

Protected edits, pre-approved by the Maintainer 2026-11-08, ruled by the Factory Builder (2026-11-13):

  • config/feed.json: one row, the question asked for the new reason.
  • sim/world.py: the simulator's Kernel is given the list. The Builder approved one line; it is three, an import, the list and a word in the comment above it, because the list is now built by one function shared with the Controller (below).

Size. 52,711 characters and 616 lines by git diff; 56,629 characters as the Kernel counts a diff, with its lines numbered, against a budget of 60,000.

The record is 0207, the number the Factory Builder gave. Records 0203 to 0206 are on main since 2026-11-13, and main is merged in at 7fdf73b, so the numbers have no gap and make checks passes.

What changes

WhereWhat
kernel/derive.pyThe Kernel reads a second list, person_only_paths. A change whose every touched path, of the two lists, is on it waits as path_needs_person; one that touches a model Lens's input as well keeps model_lens_input_needs_person. The evidence names every touched path of either list, the Lenses' inputs first, each once. A list that is not there, or is misshapen, is protected_paths_unreadable; one that names none is read
kernel/decide.pypath_needs_person needs a person's eyes
controller/configuration.pyThe Controller reads config/hooks.json once for its three lists, and person_only builds the new one from that reading: each person-only path that no model Lens reads, once. The Controller, the simulator and the tests' Kernel all take it from there
config/feed.jsonThe row for the new reason: question a_path_needs_you, choices approve or close
examples/decision.path-needs-person.jsonThe Decision's example, which a test holds to the Kernel's own answer
docs/decisions/0207-…What is decided, and each boundary's test and mutants

Who ends the wait is unchanged: a person with write access other than the author and the last pusher, as 0052 and 0200 have it. In the merge queue without one the change is taken out.

A path on both lists. None is today. Were a path listed both as a model Lens's input and as a person's, it is left off person_only_paths, so its change waits as a Lens's input does. That is the Factory Builder's ruling: a model Lens's input is the stricter name and is never masked.

From when. From the first Release cut after this merges; until then the Release on next names the wait as 0200 has it.

What is not held, and what is left alone

  • What keeps a model Lens's input from being named a person's path is controller/configuration.py, its test and its mutants. The file is neither a protected path nor a model Lens's input, so a later change to it alone needs no person's approval. Only the name and the question asked could be lost that way, never the wait or who ends it. The Factory Builder accepted it as built on 2026-11-13; whether the file should be one a person must approve is a separate decision, and the record's row "What is still not held" names it. (The security pass's note.)
  • The skill is not changed. .claude/skills/cut-a-release/SKILL.md names model_lens_input_needs_person for a change to a model Lens's input, which is still true; it says nothing of a person-only path. Its section is held word for word by a digest.

Checks it adds

  • tests/test_decide.py: the person-only wait under its own code, of any class and the Maintainer's own, in the queue, by a rename, by a pattern; the mixed change; the two lists given apart; a list that cannot be read; the feed's row and the example.
  • tests/test_controller.py: the Controller gives the list of the checkout it runs from; a path a model Lens reads is never given as a person's alone.
  • tests/test_tools.py: the tests' Kernel and the simulator's are given the same list.
  • tests/test_transitions.py: facts for the new reason.
  • 28 mutants, each named in the record's boundary rows.

Checks

CheckResult
make checksOK at 7fdf73b, main merged in
tests.test_decide, tests.test_controller.Feed, tests.test_tools.TheList, tests.test_tools.Examples, tests.test_transitionsOK
Mutants of every file the change touches, run alone on an export689 of 690 killed at d752db8; the one was the rename mutant's own syntax, fixed since. The person-only, Lens's-input and Controller mutants again at 3a2d100, after the last change to code: 102 of 102. At 7fdf73b, after the merge of main, they run again; the result is a comment here
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)91733canot run (quota), owed––––
muse91733canone1000
claude-review91733canone0011
claude-security91733canone0011
astrad752db8none0110

Astra read the head after the others' fixes. Every finding acted on was a gap in a test or a mutant, so none read again.

FindingWhatDisposition
muse #1 (blocking)Record 0207 leaves a gap in the numbers, and make checks failsgone: the four before it merged, and main is merged in
security #1, claude-review #1No test gave a model Lens's input beside a person-only path that model_lens_inputs leaves off; three mutants of the line that joins the two survived, one of them masking the Lens's inputfixed: one case with the lists given apart, and the three mutants
claude-review #2 (info)A rename out of a person-only path had a test and no mutantfixed: the mutant
security #2 (info)The rule that a Lens's input is never masked rests on a file no person need approveaccepted as built by the Factory Builder; named in the record's row "What is still not held"
astra #1 (major)No test held that a person-only path is matched as a pattern: p == path survivedfixed: a directory, a star and their other sides, with the list given apart, and the mutant
astra #2 (minor)The rename mutant lacked its newline, and was a syntax error rather than a change of behaviourfixed

Review threads, six, all resolved

ThreadDisposition
gemini: wrap the condition in parentheses, not with a backslashnot changed: kernel/derive.py continues a long condition this way three times on main
cubic: the Controller reads config/hooks.json once for each listfixed in 3a2d100: one reading for all three
cubic: overlapping patterns could leave a model Lens's file a person'snot so, and shown in 3a2d100: the Kernel matches a change against both lists, and a test holds a person-only path inside a Lens's input, and around one
cubic: expose the person-only half from tools/hooks/protected.pynot changed, by the Factory Builder's ruling: that file is a model Lens's input
codex: the record leaves a gap in the numbersgone, as said at the top
cubic: the record's row gave the wrong precedence where a path is on both listsreworded in 93f63f8: the rule was right; the row now says which pair of lists it speaks of

🤖 Generated with Claude Code