Kernel: what a person placed, only a person ends
Part of the epic "Epic: nine gaps in the Kernel, found in review". Closes #61.
The pull request's facts said whether a label was present and whether a thread was resolved, but not who did it, and had no kind for an App or the machine user. The reviewer's App could remove factory:hold and resolve the Maintainer's threads, and GitHub cannot narrow that. This follows the working ruling on #61, which gathers what the security passes on #33, #34, #55 and #60 left here.
The rule
| Act | Binds only when |
|---|---|
Removing factory:hold | a person removed it; a Hold anyone else removed still binds |
Placing factory:hands-off | a person placed it; placed by anyone else it is read as a Hold (human_hold, evidence factory:hands-off), so it stops and takes a queued item out |
Placing factory:admit | a person placed it, after the latest push; otherwise outside_not_admitted |
| Resolving a person's thread | that person resolved it |
| Resolving a Lens's thread | that Lens's App, or a person with write access, resolved it |
| Ending a request for changes | a person did |
| Approving | the machine user named as the approver, with write access |
- Every actor has a kind:
human,lens,appormachine_user. A person ishumanwithcan_writeexactly true. The port gives the kind from GitHub's account type and the configuration, never from a name's spelling (I-28, L-57). The one place a spelling is read only refuses: ahumanwhose login ends in[bot], which GitHub user logins cannot contain, is unreadable. The approver read as any kind butmachine_user, another name read asmachine_user, or an agent's or dependency update's author read as anything butapp, is unreadable too. lensis a fourth kind beyond the ruling's three. A Lens's thread is closed only by that Lens's own App or a person, and telling the reviewer's App resolving its own Finding from the author agent's App resolving the threads that hold it needs a kind for an App that reports as a Lens. The port giveslensfrom the configuration's Lens Apps.- Labels are events.
label_events(who placed or removed which label, when) replaceslabelsandadmitted_at. What a person placed binds until a person removes it, and nobody else placing it again replaces the person's placement. In the same second a removal is read first, so the label stands.complete.timelinecovers the events, and hands-off binds only when they were read in full: a timeline read short, which GitHub pages oldest first, could otherwise drop a person's later removal and keep a queued item in the queue. - An actor or label event that cannot be read defers for the Maintainer (
label_unreadable,thread_unreadable,review_unreadable), and a queued item leaves the queue (#33). - The guard #34 added by name is replaced by the kind rule.
Tests
The OnlyAPersonEnds class in tests/test_decide.py holds a boundary test for each rule, written first, and each has a mutant: 203 of 203 killed. Counter-examples for an actor without its kind, a kind the Kernel does not know, a label event without who made it, and a thread resolved by a name alone. In the simulator an App takes a Hold off, adds hands-off and resolves the Maintainer's thread, and a held item never moves on except to be dequeued (test_what_an_app_does_in_a_persons_place_ends_nothing).
Documents: design §3.3, §4.7 (the Hold, Release, Hands-off and Admit rows), I-21 and I-28; L-17 and L-02; Boundaries rows in decisions 0003, 0005 and 0007 (no Decision or Why text changed); docs/standards/code.md, docs/standards/security-boundaries.md, tests/AGENTS.md. config/apps.json and its schema read the timeline in place of the labels.
Read by
A fresh-context review and a separate security pass (Claude), on the first commit; the second commit answers them. The security pass found no way for an App, the author agent, the machine user or a non-writer to release a Hold, admit an item, close a gate or approve; 46,268 random schema-valid facts never made the Kernel raise. Both found that a timeline read short could keep a removed hands-off binding and block a dequeue, now fixed. make verify passes on a git archive export (367 tests).
🤖 Generated with Claude Code
Summary by CodeRabbit
-
New Features
- Pull-request decisions now account for who placed or removed control labels, who resolved review threads, and whether a person has write access.
- Label history and actor details are included in pull-request facts, enabling decisions to reflect event order and actor identity.
- Unreadable or incomplete timeline and actor information can defer an item for review and remove it from the merge queue.
- Outside items require a person’s admission after the author’s latest push.
-
Documentation
- Updated architecture, security, and decision guidance to describe the revised review, label, and actor rules.