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 core/220

Corey Olson33247c492447
diff --cc docs/formats.md
index 01b5165c,545a1e5f..6c16069a
--- a/docs/formats.md
+++ b/docs/formats.md
@@@ -530,12 -530,4 +530,14 @@@ Engine 68 retains the existing validate
The engine-68 `recordSpecApprovalBaseline` move carries `{agentId, baseline: {version: 1, file, anchor, rows: [{id, text, source}]}}`. `anchor` is the current commit or `null` before the first commit; exact source rows also cover uncommitted initial requirements. Only the authenticated coordinator may establish the snapshot; a person uses the main checkout’s coordinator path. Lane commands never seed or replace it. The first snapshot is indexed under `board_meta.spec_approval_baseline` with a `spec_approval_baseline` event. Replay writes metadata only, never checkout files. Sender attribution is authenticated normally, older engines refuse the new move, and later bootstrap attempts cannot replace or expand the original snapshot. No local sidecar establishes shared approval authority. The metadata snapshot must agree with its single original coordinator event; a partial restore with either half missing or changed is invalid. A malformed stored marker remains in place with an invalid-baseline diagnostic; it is not treated as an absent migration baseline or reseeded by later commands. Board imports remain trusted recovery operations; agreement of these records is a consistency check, not cryptographic authentication of an imported history.
+## Offline flow replay
+
+`pullboard flow replay --json` reads only event ids, kinds, times, item ids and seat ids from the board. The result includes `history: { version: 1, events: [...] }`; each event has only `event_id`, `event_kind`, `event_at`, `item_id` and `event_by`. No title, description, message, path, commit or event-detail JSON is fetched or exported. Save that history object as JSON and pass it with `--history <file>` to replay the same input on another board.
+
+The default policy, `{"order":"recorded"}`, keeps recorded stage starts, durations and seat concurrency, reproducing complete measured cycles. `--policy <file>` accepts an explicit item-id order (a partial list prioritizes those ids), plus optional positive integer `settings.buildWip`, `settings.reviewWip` and `settings.landingWip`. These are hypothetical replay limits, not changes to the live board or machine. Unspecified limits come from observed concurrent stage intervals; recorded seat identities and their observed concurrency remain fixed. A different order or limit schedules ready work without preemption, retaining item arrivals and every measured build/review/landing interval. Recorded policy has exact observed start constraints; alternate policy has none of the observed queue delays.
+
+`replay` reports landing order, each first-claim-to-landed time, and stage quartiles in minutes using the existing flow-window convention (fast/slow quarters use the sorted sample at floor(n/4) and floor(3n/4), with the middle pair averaged for an even median). These counterfactual predictions are not timings measured under an alternate live queue. Landing intervals are the recorded accept-to-merged spans, including observed waiting, not an estimate of CPU work. Incomplete, abandoned or missing-boundary cycles appear in `unmeasured`; they never become zero-duration work. Rejected complete attempts retain their actual builder/reviewer seats and durations. No wall clock or random values enter the replay result.
+
+Native `event_id` order is canonical, even when an input array is shuffled. A history whose recorded timestamps decrease in that order is refused explicitly: the bench never reorders lifecycle events by a clock, clamps timestamps, or presents contradictory clock evidence as an exact reproduction.
++
+ Engine 93 adds item-addressed shouts. A shout move's optional `item` names an existing item; its event belongs to that item and includes the message text. Its canonical `shout_to` is `item:<id>`, read by the item's current builder and reviewer and the coordinator. Inbox, resume and seat watch use the same audience; an explicit seat or lane destination with `--item` still uses the item's audience. An item decision is answerable by its current participants or coordinator; replies stay in its thread and audience. The coordinator can still pass it to the person through the existing decision flow. A registered seat's shout to `all` is accepted and stored at reserved address `coordinator+person`, read only by coordination audiences. The plus sign cannot occur in valid lane names or their derived seat ids. Coordinator and person broadcasts retain `all` and reach everyone; recorded older-engine `all` rows retain their original global meaning without a retrospective sender filter. Main-checkout watch remains the person's board-wide observer. A joined seat's default watch filters shouts to its own address, lane, items and `all`, while ordinary lifecycle events remain visible.