Skip to content

Autonomous development for real products. Governed agents build from specs you approve, follow the rules you set, and nothing merges until another agent proves it. Your decisions, judgement and taste are the bar that work must clear. Declare your vision and let agents build it.

Merge branch 'main' into codex/sol1240-20261025

Corey Olson07f53a0f79ab
diff --cc docs/api.md
index 289c61e1,6364e719..2d730d49
--- a/docs/api.md
+++ b/docs/api.md
@@@ -685,8 -692,9 +692,11 @@@ A board-recorded item merge covers its
The JSON result includes `verification` entries with PR number, exact head, state, assurance, receipt identities, and any unverified commit IDs. GitHub statuses are shared by PRs with the same head: if an outside or unproven PR shares a commit with a proven proposal, the shared status stays red rather than lending the proposal's green status to it. Publishing observes and posts statuses; it never merges a PR or requires outside authors to use Pullboard.
+A Node test that intentionally generates CPU load declares `{ cpuLoad: true }` in its test options. `bin/run-tests.js` discovers declarations without running the module and places the selected files on a registered runner, even with local-first admission or an inherited coordinator lease. While runners are registered it waits for a compatible runner instead of using local or short capacity; browser and GUI requirements still apply. A runner failure never permits this load to fall back locally. With no runners registered it uses the ordinary local admission and prints the reason. Source must be clean and committed for remote execution. Discovery conservatively places the whole selected file, including when a name filter selects only some cases; comments and string/template examples do not declare load. Options that cannot be read statically count as declared load: computed property names, object spreads, opaque option expressions, unresolved test-function uses and unreadable test-context member uses, unresolved CommonJS registration paths, dynamic imports, executable `eval`/`Function` references and `node:vm`/`node:module` imports require runner placement. Readable direct subtests and ordinary context lifecycle hooks retain their ordinary classification. V8 identifies imported modules without executing them; comments, quoted named imports and escaped module specifiers cannot hide registration aliases. A node:test binding that the bounded reader cannot understand is conservatively placed on runners. Discovery never evaluates those expressions. Discovery inspects the reachable local executable import/re-export closure, including wrappers and barrels. It follows at most 256 files, 4 MiB and 32 edges of depth, deduplicates cycles and resolves real paths inside the repository. Missing, escaping, unsupported, unreadable or unresolved executable closure nodes require conservative runner placement; no special parser or transport names are exempt. This is placement under uncertainty, not an inference that a file actually burns CPU. A directly readable literal `cpuLoad: false` remains ordinary work when its full inspected closure is also readable.
+
+ ### Item decision queue
+
+ The full and summary board state include `decisionQueue`, the coordinator's open decisions, separate from recent `shouts` and the person's `decisions`/Needs you. Structured item questions retain the ordinary shout fields and add `item` and `options`. The state delta carries this small queue whole, so an answer removes its pending entry. Use the normal authenticated `answer` move to append an answer; a decision passed to the person requires the existing person channel.
`pullboard why <file>[:<line>] --json` uses the `why file` catalog shape; this names the file form, not a literal `file` argument. It traces commits on the repository's actual primary branch, then resolves their item trailers, served row IDs and answered decisions in one board query. The legacy `why <id>` shape is unchanged. Answered decisions carry their recorded origin, statement, rationale, decider and relayer; absent historical provenance remains `unknown`. Row and item origins are also `unknown` unless recorded, rather than inferred from a related decision. An agent may label a relayed or inherited answer but cannot use that label to mint a person-origin record. Unproven original decider identities remain unknown, with the answering relay seat recorded separately.
diff --cc docs/formats.md
index 6205e17c,d921ae12..67023092
--- a/docs/formats.md
+++ b/docs/formats.md
@@@ -706,4 -710,4 +710,5 @@@ A seat address is `pullboard:<board-id>
The activity-delivery engine, numbered at landing, adds recipient-owned `recordActivityDelivery` and `actOnActivity` moves. Interrupt ids bind a native event (`event:<id>`) or an exact claim lease (`claim:<item>:<lease-until>`). A delivery event lists only interrupts actually included in that recipient's response; an acted event records a separate explicit acknowledgement. An inbox cursor or reading a response never means acted-on. Both moves use the joined recipient as their authenticated actor, preserve earlier event history, and refuse on an older replay engine before changing its prefix. Delivery state in indexed `board_meta` keys is part of the native board and its exports, not a checkout-only cursor. `through` is the safely covered interrupt prefix, `observedThrough` is the read snapshot's high-water mark, and optional `digestThrough` is the separately delivered digest prefix; an elided interrupt must remain available after a later small interrupt is delivered.
+The machine resource database adds `waiter.runner_only` and `holder.runner_execution`, both integer booleans defaulting to zero. A runner-only gate ticket cannot consume local or short capacity while a runner is registered. The trusted SSH worker sets `runner_execution` on its own actual private gate holder; inherited admission reads that holder rather than accepting an environment-only claim of remote execution. These are local, additive scheduling metadata, not board move or relay formats.
+ The CPU-sentinel engine, numbered at landing, names an agent-worktree process after more than twenty CPU minutes, once per host, PID and start time. Current `health-scan` receipts carry `pendingCpu`, the findings whose coordinator notice was held by the existing ten-notice cap. Later scans drain that durable queue without inventing another health-failure episode. An older engine refuses while this queue is pending, preserving its history; older receipts retain their original fields. Observation never signals a process.