You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Part of the epic "Epic: nine gaps in the Kernel, found in review".
Why. Nothing counts acts. An approval that never lands is issued again on every pass, and the Run ceiling does not see it. The issue Lifecycle adds more acts: the Controller accepts an issue and places a Hold by adding a label.
Done when
The Maintainer rules: how an act is recorded, since run records are for Runs, and which ceiling it counts toward.
An act that does not land is bounded, shown by a boundary test written first.
The simulator has an approval that fails to land.
Its mutant is killed.
Needs. The Maintainer's ruling, asked first.
Design. §3.6, recovery keys; I-8; §4.1, the issue Lifecycle, Acceptance.
Working ruling (Solution Architect, delegated). The Maintainer may overturn it in review.
Question. How is an act recorded, and what bounds it? An act is a change the Kernel decides with no model: an approval, an enqueue, a dequeue, or a label the Controller adds.
Options.
The approval is a Run with no Engine, because only apply holds the machine user's token. Its run record limits attempts at an approval that does not land. Every act that does land is counted from GitHub's timeline as acts_since_writer, against a new limit max_acts_per_item.
Every act is a Run. This needs a runner job for each label and queue move, moves the Controller's key out of the poller, and reverses the Maintainer's ruling that acceptance is "an act of the Controller, not a Run" (§4.1).
Count from the timeline only. An approval that never lands is still issued again on every pass, so the gap stays.
The Kernel reads its own pass records. That makes the Factory a store it writes and reads, which breaks decision 0002.
Ruling: 1.
It reads only what GitHub keeps (§3.9) and leaves the credential table in §4.9 unchanged.
It makes FR-17 and I-8 hold as a sum. Every automated action either happens inside a Run, bounded by max_runs_per_item, or is an act, bounded by max_acts_per_item.
An approval Run counts toward its own attempts but not toward the Run ceiling, which stays the cost ceiling (FR-39).
Acts that stop work, such as a Hold label or a dequeue, are counted but never refused.
Acceptance criteria.
test_an_approval_that_does_not_land_is_bounded: max_attempts failed issue-approval runs whose fault is the item's give hold with reason attempts_exhausted. A writer's action grants a fresh allowance (I-28).
test_acts_are_bounded: acts_since_writer at the limit gives hold, reason act_ceiling, attention needs_you; one below the limit, the act is made. The new fact goes into the facts schema, the example, tests/support.py and sim/world.py, with a counter-example. max_acts_per_item goes into config/defaults.json and its schema, starting at 6.
A mutant for each rule. The simulator gains an approval that never lands and a queue that ejects every time, and asserts I-8.
Documents:
Design §3.3: actors_since_human becomes acts_since_writer, counted since a writer acted.
Design §3.10: the act Run.
Design §4.2: "with no model Run".
Design §4.1 and §9.9: remove the open question about counting the Controller's acts.
docs/glossary.md gains "Act".
Decision 0003 and L-03.
Confidence. Medium. I would rule otherwise, and choose option 2, if trial P-8 shows that reading the timeline is too expensive.
Needs the Maintainer? No. The numbers are placeholders.
Activity
don-petry commentedon Oct 14, 2026
#37 Kernel: how an act is recorded and bounded
Working ruling (Solution Architect, delegated). The Maintainer may overturn it in review.
Question. How is an act recorded, and what bounds it? An act is a change the Kernel decides with no model: an approval, an enqueue, a dequeue, or a label the Controller adds.
Options.
applyholds the machine user's token. Its run record limits attempts at an approval that does not land. Every act that does land is counted from GitHub's timeline asacts_since_writer, against a new limitmax_acts_per_item.Ruling: 1.
max_runs_per_item, or is an act, bounded bymax_acts_per_item.Acceptance criteria.
test_an_approval_that_does_not_land_is_bounded:max_attemptsfailedissue-approvalruns whose fault is the item's giveholdwith reasonattempts_exhausted. A writer's action grants a fresh allowance (I-28).test_acts_are_bounded:acts_since_writerat the limit giveshold, reasonact_ceiling, attentionneeds_you; one below the limit, the act is made. The new fact goes into the facts schema, the example,tests/support.pyandsim/world.py, with a counter-example.max_acts_per_itemgoes intoconfig/defaults.jsonand its schema, starting at 6.actors_since_humanbecomesacts_since_writer, counted since a writer acted.docs/glossary.mdgains "Act".Confidence. Medium. I would rule otherwise, and choose option 2, if trial P-8 shows that reading the timeline is too expensive.
Needs the Maintainer? No. The numbers are placeholders.