pi-crew 0.11.9 → 0.11.11

package/src/runtime/task-packet.ts

@@ -184,25 +184,39 @@ export function validateTaskPacket(packet: TaskPacket): TaskPacketValidationResu
 }
 
 /**
- * Structured handoff template for task completion reports.
- * Distilled from ECC dmux-workflows pattern — workers use this format
- * so verifiers and downstream consumers can parse output predictably.
+ * Structured handoff template for task completion reports (U10).
+ * Restructured to the 6 PINNED sections of durable's SUMMARIZATION_PROMPT
+ * (pi-durable compaction.ts:68-103 — Goal / Constraints & Preferences /
+ * Progress (Done·In Progress·Blocked) / Key Decisions / Next Steps /
+ * Critical Context), ported per spec U10 so downstream workers and the
+ * leader can parse output predictably. The R3-2 `### Provenance` identity
+ * footer is retained as the final section — it is an orthogonal pinned
+ * contract (worker PI_* self-report, see task-handoff-provenance.test.ts),
+ * not one of the 6 content sections.
  */
 export const HANDOFF_TEMPLATE = [
 	"## Handoff",
 	"",
-	"### Summary",
-	"<!-- 2-3 sentences describing what was done -->",
+	"### Goal",
+	"<!-- One line: the goal this task served, from the Task Packet objective -->",
 	"",
-	"### Files Changed",
-	"<!-- List each file changed with brief description -->",
-	"<!-- - path/to/file.ts: description -->",
+	"### Constraints & Preferences",
+	"<!-- Constraints that shaped the work: branch/commit policy, file ownership, style, risk tiers -->",
 	"",
-	"### Tests / Verification",
-	"<!-- What tests pass? What was manually verified? -->",
+	"### Progress",
+	"<!-- Report the state of the work in three groups, each as bullet lists:",
+	"Done: completed items, with file paths",
+	"In Progress: items underway when this task ended",
+	"Blocked: blockers, with the exact reason and owner -->",
 	"",
-	"### Follow-ups",
-	"<!-- Any remaining issues or next steps -->",
+	"### Key Decisions",
+	"<!-- Decisions made and why, with file:line anchors where useful -->",
+	"",
+	"### Next Steps",
+	"<!-- Remaining work, follow-ups, and the recommended next action -->",
+	"",
+	"### Critical Context",
+	"<!-- Verification evidence (commands + outcomes), changed/read files, risks the next worker must know -->",
 	"",
 	"### Provenance",
 	"<!-- Self-report your runtime identity (pre-set in your bash tool env, one line): PI_SESSION_ID, PI_SESSION_FILE, PI_MODEL, PI_PROVIDER, PI_REASONING_LEVEL — e.g. `session=<id> model=<provider>/<model> reasoning=<level>` -->",
@@ -292,6 +306,13 @@ function parseMixedContent(section: string): string[] {
  * Parse structured handoff data from agent output text.
  * Looks for the "## Handoff" heading and extracts subsections.
  * Returns empty arrays for sections not found.
+ *
+ * U10 note: targets the LEGACY section names (Summary / Files Changed /
+ * Tests / Verification / Follow-ups) for backward compatibility with
+ * outputs persisted before the 6-section template; the current
+ * HANDOFF_TEMPLATE (Goal / Constraints & Preferences / Progress /
+ * Key Decisions / Next Steps / Critical Context) has no dedicated parser
+ * yet — see spec U10 follow-ups.
  */
 export function parseHandoffFromOutput(output: string): ParsedHandoff {
 	if (!output || typeof output !== "string") {

package/src/runtime/task-runner/prompt-builder.ts

@@ -49,7 +49,7 @@ export function coordinationBridgeInstructions(task: TeamTaskState, opts?: { inc
 		"- Do not resolve cross-worker conflicts silently. Escalate via mailbox/result with: file/symbol, conflicting task if known, proposed owner, and safest next step.",
 		"- If nudged, answer with current status, blocker, or smallest next step.",
 		"- Treat inherited/dependency context as reference-only; do not continue the parent conversation directly.",
-		"- Completion handoff should include: DONE/FAILED, summary, changed/read files, verification evidence, and remaining risks.",
+		"- Completion handoff must use the structured 6-section template (Goal / Constraints & Preferences / Progress / Key Decisions / Next Steps / Critical Context): mark DONE/FAILED and split Done·In Progress·Blocked under Progress, list changed/read files and verification evidence under Critical Context, and surface remaining risks as Next Steps.",
 	].join("\n");
 }
 
@@ -303,9 +303,19 @@ export function promptSkillMode(): "index" | "full" {
 	return getCrewEnv("PI_CREW_PROMPT_SKILLS") === "full" ? "full" : "index";
 }
 
-/** Estimated tokens (chars/4 — the in-tree heuristic; no tokenizer dep). */
-export function estimateTokens(chars: number): number {
-	return Math.round(chars / 4);
+/** Estimated tokens (chars/4 — the in-tree heuristic; no tokenizer dep).
+ *  U10 measured-token anchor: when a REAL measurement is available (message_end
+ *  usage / task result usage persisted on `TeamTaskState.usage`, threaded into
+ *  the dependency context by `aggregateUsage`), the effective estimate is
+ *  `max(heuristic, measured)` — a real measurement can only TIGHTEN the
+ *  estimate, never loosen it, so a budget sized in effective tokens is never
+ *  exceeded when measured usage is larger than the heuristic (spec U10
+ *  acceptance). Callers without a measurement keep the plain heuristic.
+ *  Mirrors `measureHandoffTokens` in src/runtime/task-output-context.ts
+ *  (duplicated, not imported — see the sync note there). */
+export function estimateTokens(chars: number, measuredTokens?: number): number {
+	const heuristic = Math.round(chars / 4);
+	return measuredTokens === undefined ? heuristic : Math.max(heuristic, measuredTokens);
 }
 
 export async function renderTaskPrompt(
@@ -380,7 +390,7 @@ export async function renderTaskPrompt(
 		"Task:",
 		sanitizeTaskText(step.task.replaceAll("{goal}", manifest.goal)),
 		"",
-		"When your task is complete, structure your final output using this handoff template:",
+		"When your task is complete, structure your final output using this handoff template (Goal / Constraints & Preferences / Progress / Key Decisions / Next Steps / Critical Context — U10 pinned sections):",
 		HANDOFF_TEMPLATE,
 	].join("\n");
 

package/src/state/stores/state-store.ts

@@ -491,6 +491,11 @@ export function createRunManifest(params: {
 				`saveManifestAndTasksAtomicSync: manifestWritten=${result.manifestWritten}, tasksWritten=${result.tasksWritten}${cause}`,
 			);
 	}
+	// U2 (2026-10-10): kept SYNC — run birth anchor with an EXPLICIT seq:1.
+	// createRunManifest is a sync API whose callers may immediately append more
+	// events in the same tick (e.g. updateRunStatus(run.running)); a delayed
+	// run.created would let those reserve seq 1 from the sidecar first —
+	// duplicate seq + inverted audit order. Critical-path same-tick durability.
 	appendEvent(paths.eventsPath, {
 		type: "run.created",
 		runId: paths.runId,
@@ -910,6 +915,10 @@ export function updateRunStatus(
 	// allowTerminalExit for its legitimate cancelled→running transition.
 	const saved = saveRunManifest(updated, { allowTerminalExit: options.allowTerminalExit });
 	if (saved.status !== status) {
+		// U2 (2026-10-10): kept SYNC (both updateRunStatus events) — public sync
+		// API; same-poll status readers and the 2026-09-10 WI-2.2/M2 remediation
+		// (b6eba80f) require same-tick event visibility, and callers may append
+		// more events in this same tick that must order around run.<status>.
 		appendEvent(saved.eventsPath, {
 			type: "run.terminal_preserved",
 			runId: saved.runId,
@@ -929,6 +938,8 @@ export function updateRunStatus(
 			/* non-critical */
 		}
 	}
+	// U2 (2026-10-10): kept SYNC — see the run.terminal_preserved note above
+	// (same-tick visibility for same-poll status readers; b6eba80f).
 	appendEvent(updated.eventsPath, {
 		type: `run.${status}`,
 		runId: updated.runId,