@tablation/crew 0.78.0 → 0.79.0

package/dist/cli.js

@@ -230,6 +230,13 @@ function extractTurnTokens(line) {
   if (!usage2) return void 0;
   return (usage2.input_tokens ?? 0) + (usage2.output_tokens ?? 0) + (usage2.cache_read_input_tokens ?? 0) + (usage2.cache_creation_input_tokens ?? 0);
 }
+function extractPrefixTokens(line) {
+  const obj = line;
+  if (!obj || obj.type !== "assistant") return void 0;
+  const usage2 = obj.message?.usage;
+  if (!usage2) return void 0;
+  return (usage2.input_tokens ?? 0) + (usage2.cache_creation_input_tokens ?? 0);
+}
 function extractResult(line) {
   const obj = line;
   if (!obj || obj.type !== "result") return void 0;
@@ -580,6 +587,7 @@ var claudeAdapter = {
   mapStreamLine,
   extractResult,
   extractTurnTokens,
+  extractPrefixTokens,
   inspectMcp(o) {
     const s = inspectClaudeMcp(o);
     return s.state === "no-claude" ? { state: "no-binary" } : s;
@@ -5114,6 +5122,8 @@ async function reportAgentRun(target, run2) {
     cacheReadTokens: run2.cacheReadTokens,
     cacheWriteTokens: run2.cacheWriteTokens,
     costUsd: run2.costUsd,
+    numTurns: run2.numTurns,
+    prefixTokens: run2.prefixTokens,
     promptVersion: run2.promptVersion,
     promptSha: run2.promptSha
   };
@@ -5271,24 +5281,31 @@ function ticketMentions(text) {
 }
 var TicketAttributionTracker = class {
   #tokensByTicket = /* @__PURE__ */ new Map();
+  #turns = 0;
   #current;
   /** Tracker record uuid → ticket key, learned when a write names both (a claim resolves the key from the last mention). */
   #keyByRecord = /* @__PURE__ */ new Map();
   #lastBy = {};
-  /** The ticket key most recently named in the transcript, or undefined before any. */
+  /**
+   * The ticket key most recently named in the transcript, or undefined before any.
+   * Only the FIRST key in a turn counts: a closing summary names its own
+   * ticket first and then lists others ("Other tickets", "Blocked by", "See
+   * also"), and those trailing keys are not what the run worked (CREW-2598).
+   */
   get current() {
     return this.#current;
   }
   /** Record the mentions in `text` without crediting tokens (a claim needs "which ticket" before the turn's usage line arrives). */
   note(text) {
     const mentions = ticketMentions(text);
-    if (mentions.length > 0) this.#current = mentions[mentions.length - 1];
+    if (mentions.length > 0) this.#current = mentions[0];
   }
   /** One assistant turn: its combined visible text (thinking + text blocks) and its own token cost. */
   add(text, tokens) {
     const mentions = ticketMentions(text);
-    if (mentions.length > 0) this.#current = mentions[mentions.length - 1];
+    if (mentions.length > 0) this.#current = mentions[0];
     if (this.#current === void 0 || tokens <= 0) return;
+    this.#turns++;
     this.#tokensByTicket.set(this.#current, (this.#tokensByTicket.get(this.#current) ?? 0) + tokens);
   }
   /**
@@ -5310,16 +5327,31 @@ var TicketAttributionTracker = class {
   /**
    * How the attribution is decided: a hand-off write beats a claim, a claim
    * beats a comment (so a follow-up ticket filed or commented on while
-   * working never displaces the claimed one), and any write beats the
-   * mention-token heuristic, which is only the fallback for runs that made
-   * no ticket-targeted write. The pre-run hint is last.
+   * working never displaces the claimed one), and any write beats everything
+   * below. Next, `hintMoved` — the post-run read-back showed the hinted
+   * ticket changed status or assignee during the run — is a hand-off signal
+   * for a run that wrote through MCP tools crew cannot see. Then the
+   * mention-token heuristic, which with a hint may override it only when a
+   * different ticket holds a strict majority of the credited tokens (a single
+   * credited turn, like a lone end-of-run summary naming another ticket,
+   * must not). A run with no hint
+   * keeps the plain winner. The pre-run hint is last.
    */
-  decide(hint) {
+  decide(hint, opts = {}) {
     for (const by of ["handoff", "claim", "comment"]) {
       const ticket = this.#lastBy[by];
       if (ticket) return { ticket, by };
     }
-    if (this.#tokensByTicket.size > 0) return { ticket: this.winner(hint), by: "mentions" };
+    if (hint && opts.hintMoved) return { ticket: hint, by: "readback" };
+    if (this.#tokensByTicket.size > 0) {
+      if (!hint) return { ticket: this.winner(), by: "mentions" };
+      const winner = this.winner(hint);
+      if (winner === hint) return { ticket: hint, by: "mentions" };
+      let total = 0;
+      for (const n of this.#tokensByTicket.values()) total += n;
+      const share = (this.#tokensByTicket.get(winner) ?? 0) / total;
+      return share > 0.5 && this.#turns > 1 ? { ticket: winner, by: "mentions" } : { ticket: hint, by: "hint" };
+    }
     return { ticket: hint, by: "hint" };
   }
   /**
@@ -7099,6 +7131,28 @@ async function finishRunWorktree(dir, cfg, branch) {
 // src/handoff-guard.ts
 var HANDOFF_BOUNCED_MARKER = calloutHeader("WARNING", "Hand-off Bounced");
 var pushedEventBody = (sha, remote, branch) => `pushed ${sha} to ${remote}/${branch}`;
+function acceptanceLines(description) {
+  if (typeof description !== "string") return [];
+  const out = [];
+  let inside = false;
+  for (const raw of description.split("\n")) {
+    const heading = /^#{1,6}\s+(.*?)\s*$/.exec(raw);
+    if (heading) {
+      inside = /^acceptance\b/i.test(heading[1]);
+      continue;
+    }
+    if (!inside) continue;
+    const bullet = /^\s*(?:[-*+]|\d+[.)])\s+(?:\[[ xX]\]\s+)?(.+?)\s*$/.exec(raw);
+    if (bullet) out.push(bullet[1]);
+  }
+  return out;
+}
+var flat = (s) => s.replace(/[`*_>]/g, "").replace(/\s+/g, " ").trim().toLowerCase();
+function unquotedAcceptance(lines, comment) {
+  const body = flat(comment);
+  return lines.filter((l) => !body.includes(flat(l)));
+}
+var ACCEPTANCE_BOUNCED_MARKER = calloutHeader("WARNING", "Acceptance Checklist Missing");
 function remoteTip(dir, remote, branch) {
   const out = gitOk(dir, ["ls-remote", "--heads", remote, `refs/heads/${branch}`]);
   if (out === null) return null;
@@ -7174,6 +7228,27 @@ async function takenOverSince(i, ticketRow) {
     return !!by && by !== i.shipName;
   });
 }
+async function missingAcceptance(i, t) {
+  const lines = acceptanceLines(t.description);
+  if (lines.length === 0 || !i.writer.comments) return null;
+  const cc = i.contract.comments;
+  const mine = (await i.writer.comments()).map((c) => c).filter((r) => r[cc.parent] === t.id && r[cc.author] === i.seatId).sort((a, b) => String(a[cc.createdAt] ?? "").localeCompare(String(b[cc.createdAt] ?? "")));
+  const last = mine.at(-1);
+  const missing = unquotedAcceptance(lines, last ? String(last[cc.body] ?? "") : "");
+  return missing.length > 0 ? missing : null;
+}
+function acceptanceBounceComment(missing, shipName2) {
+  return `${ACCEPTANCE_BOUNCED_MARKER}
+
+**Hand-off bounced: the hand-off comment does not quote every line of the ticket's Acceptance section.**
+
+${shipName2 ? `Ship: ${shipName2}
+
+` : ""}Not quoted verbatim:
+${missing.map((l) => `- ${l}`).join("\n")}
+
+Set back to \`in_progress\`. Post a comment that lists every Acceptance line verbatim, each followed by the test, command or manual check that exercised it in this run and its result (a line you could not exercise is listed with the reason, never omitted), then set \`fixed\` again. The branch is already pushed.`;
+}
 async function enforceHandoff(i) {
   try {
     const t = await i.writer.ticket(i.ticketId);
@@ -7189,6 +7264,26 @@ async function enforceHandoff(i) {
       i.log.warn(`${t.issue_id} is at ${t.status} but the push was not checked: ${v.why}`, log);
       return { kind: "unverified", why: v.why };
     }
+    if (v.kind === "pushed" && !claimed) {
+      const missing = await missingAcceptance(i, t);
+      if (missing) {
+        const heldNow = i.hold ? t[i.hold.column] : void 0;
+        if (i.hold && heldNow && heldNow !== i.hold.shipId || await takenOverSince(i, t.id)) {
+          i.log.warn(`${t.issue_id} is held by another ship now \u2014 the acceptance check is stale, left alone`, log);
+          return { kind: "unverified", why: "ticket is held by another ship" };
+        }
+        await i.writer.updateTicket(t.id, {
+          status: i.contract.statuses.building,
+          ...t.assignee_id ? {} : { assignee_id: i.seatId },
+          ...i.hold && !heldNow ? { [i.hold.column]: i.hold.shipId } : {}
+        });
+        await i.writer.postEvent(t.id, acceptanceBounceComment(missing, i.shipName), i.seatId, "bounced").catch(() => {
+          i.log.warn("bounced, but the note failed to post", log);
+        });
+        i.log.warn(`${t.issue_id} bounced back to ${i.contract.statuses.building}: hand-off comment lacks the acceptance checklist`, log);
+        return { kind: "bounced" };
+      }
+    }
     if (v.kind === "pushed") {
       await i.writer.postEvent(t.id, pushedEventBody(v.sha, v.remote, v.branch), i.seatId, "pushed");
       i.log.emit(`${t.issue_id} pushed at ${v.sha.slice(0, 7)} \u2014 handed on to QA`, log);
@@ -8467,6 +8562,8 @@ var RunObserver = class {
   /** Hand-off writes and their outcomes, to tell a run whose hand-off never landed (CREW-2363). */
   handoff = new HandoffTranscriptTracker();
   result;
+  /** Prefix tokens of the first assistant call that carried usage; kept once. */
+  prefixTokens;
   /** Set once the stream carries the harness's final result (turn count or cost). */
   sawResult = false;
   muted = false;
@@ -8593,6 +8690,7 @@ var RunObserver = class {
         }
       }
     }
+    if (this.prefixTokens === void 0 && parsed !== void 0) this.prefixTokens = this.adapter.extractPrefixTokens?.(parsed);
     const turnTokens = parsed === void 0 ? void 0 : this.adapter.extractTurnTokens(parsed);
     if (turnTokens !== void 0) this.attribution.add(turnText, turnTokens);
   }
@@ -8616,8 +8714,15 @@ function writeRunMeta(id, startedAt, liveTicket, extra) {
   } catch {
   }
 }
-async function finishRun(id, obs, emit, startedAt, end, agentLogKey) {
-  const { ticket: attributedTicket, by: attributedBy } = obs.attribution.decide(obs.liveTicket);
+async function finishRun(id, obs, emit, startedAt, end, agentLogKey, hintMoved) {
+  let moved = false;
+  if (hintMoved) {
+    try {
+      moved = await hintMoved();
+    } catch {
+    }
+  }
+  const { ticket: attributedTicket, by: attributedBy } = obs.attribution.decide(obs.liveTicket, { hintMoved: moved });
   const identity = { toolchain: id.toolchain, harness: id.harness, provider: id.provider };
   const { result } = obs;
   const pending = end.interrupted ? null : obs.handoff.pending();
@@ -8649,6 +8754,7 @@ async function finishRun(id, obs, emit, startedAt, end, agentLogKey) {
       ...pending ? { handoff: "pending" } : {},
       sessionId: result?.sessionId ?? null,
       numTurns: result?.numTurns ?? null,
+      prefixTokens: obs.prefixTokens ?? null,
       totalCostUsd: result?.costUsd ?? null
     });
   } catch {
@@ -8661,6 +8767,8 @@ async function finishRun(id, obs, emit, startedAt, end, agentLogKey) {
       ...pending ? { handoff: "pending" } : {},
       ...end.interrupted ? { interrupted: end.interrupted } : {},
       ...identity,
+      attributedTicket: attributedTicket ?? null,
+      attributedBy,
       ...result?.sessionId ? { sessionId: result.sessionId } : {},
       ...result?.numTurns !== void 0 ? { numTurns: result.numTurns } : {},
       ...result?.costUsd !== void 0 ? { totalCostUsd: result.costUsd } : {}
@@ -8679,6 +8787,9 @@ async function finishRun(id, obs, emit, startedAt, end, agentLogKey) {
       cacheReadTokens: result?.cacheReadTokens,
       cacheWriteTokens: result?.cacheWriteTokens,
       costUsd: result?.costUsd,
+      // Both come from a finished run: without the result line (interrupted) neither is trustworthy.
+      ...result?.numTurns !== void 0 ? { numTurns: result.numTurns } : {},
+      ...result?.numTurns !== void 0 && obs.prefixTokens !== void 0 ? { prefixTokens: obs.prefixTokens } : {},
       harness: identity.harness,
       provider: identity.provider,
       promptVersion: id.promptVersion,
@@ -8723,7 +8834,7 @@ async function finishRun(id, obs, emit, startedAt, end, agentLogKey) {
   }
   return { ms: ms2, ticket: attributedTicket, handoffPending: !!pending };
 }
-function spawnAgent(plan, emit) {
+function spawnAgent(plan, emit, hintMoved) {
   const adapter = adapterFor(plan.harness ?? LEGACY_TOOLCHAIN_HARNESS);
   const started = Date.now();
   const startedAt = new Date(started).toISOString();
@@ -8788,7 +8899,7 @@ function spawnAgent(plan, emit) {
       obs.end();
       await endSink(eventsSink);
       if (spawnFailed) return;
-      const { ms: ms2, handoffPending } = await finishRun(id, obs, emit, startedAt, { code: code ?? 1 }, plan.agentLog?.apiKey);
+      const { ms: ms2, handoffPending } = await finishRun(id, obs, emit, startedAt, { code: code ?? 1 }, plan.agentLog?.apiKey, hintMoved);
       resolve6({ code: code ?? 1, ms: ms2, ...handoffPending ? { handoffPending } : {} });
     });
     child.stdin?.on("error", (err) => {
@@ -16914,7 +17025,19 @@ async function runRoleAgent(role, decision, ctx) {
   }
   try {
     const runStart = (/* @__PURE__ */ new Date()).toISOString();
-    await spawnAgent(plan, emit);
+    let hintMoved;
+    if (workingId) {
+      try {
+        const before = await tracker.ticket(workingId);
+        const col = tracker.contract.columns;
+        hintMoved = async () => {
+          const after = await tracker.ticket(workingId);
+          return after[col.status] !== before[col.status] || after[col.assignee] !== before[col.assignee];
+        };
+      } catch {
+      }
+    }
+    await spawnAgent(plan, emit, hintMoved);
     await guardRunHandoff(route, ship, tracker, role, workingId, emit, runStart);
   } finally {
     if (memberId) {
@@ -19383,7 +19506,7 @@ ${changed2.map((c) => `  - ${c}
           }
           for (const e of entries) {
             process.stdout.write(
-              `${e.id}  ${e.started_at ?? "?"}  ${e.outcome ?? "?"}${e.ticket_reference ? `  ${e.ticket_reference}` : ""}${e.model ? `  ${e.model}` : ""}${e.harness || e.provider ? `  (${[e.harness, e.provider].filter(Boolean).join(" \xB7 ")})` : ""}
+              `${e.id}  ${e.started_at ?? "?"}  ${e.outcome ?? "?"}${e.ticket_reference ? `  ${e.ticket_reference}` : ""}${e.model ? `  ${e.model}` : ""}${e.harness || e.provider ? `  (${[e.harness, e.provider].filter(Boolean).join(" \xB7 ")})` : ""}${e.num_turns != null ? `  turns ${e.num_turns}` : ""}${e.prefix_tokens != null ? `  prefix ${e.prefix_tokens}` : ""}
 `
             );
           }
@@ -19401,7 +19524,9 @@ ticket:     ${entry.ticket_reference ?? "(none)"}
 model:      ${entry.model ?? "?"}
 harness:    ${entry.harness ?? "?"}
 provider:   ${entry.provider ?? "?"}
-`
+` + (entry.num_turns != null ? `turns:      ${entry.num_turns}
+` : "") + (entry.prefix_tokens != null ? `prefix:     ${entry.prefix_tokens} tokens
+` : "")
           );
           if (flag("prompt")) {
             process.stdout.write(`

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tablation/crew",
-  "version": "0.78.0",
+  "version": "0.79.0",
   "description": "A standing team of headless agents that picks work off a Tablation board, does it on a machine you control, and reports back.",
   "type": "module",
   "license": "MIT",

package/prompts/default/personas/lane-qa.md

@@ -94,6 +94,11 @@ pass.
    Judge by exit code and summary line, and quote the summary in your
    comment.
 7. **Check the hand-off itself.**
+   - **Acceptance checklist.** Start from the builder's checklist: their
+     comment quotes each Acceptance line with the check that exercised it.
+     Re-run the checks that matter, and when a line fails, a bounce cites that
+     line by its quoted text. A line the builder marked unexercised is yours
+     to exercise.
    - **Documentation.** If the repository's agent instructions (CLAUDE.md,
      AGENTS.md, GEMINI.md or equivalent) name a user-facing documentation
      location, a change to a user-facing surface must come with a matching

package/prompts/default/skills/crew-lane-agent.md

@@ -416,6 +416,14 @@ before going further, and never change a ticket's `report_type` yourself.
    used, and anything you could not test yourself. Anything not committed and
    pushed is gone.
 
+   **When the ticket has an Acceptance section, your last progress comment
+   quotes it.** List every line of that section verbatim, each followed by the
+   test, command or manual check that exercised it in this run and its result.
+   A line you could not exercise is listed with the reason, never omitted. The
+   runner checks your newest comment for every line and, if one is missing,
+   sets the ticket back to `in_progress` instead of passing it to QA. A ticket
+   with no Acceptance section is unaffected.
+
 ## Step 4 — you never merge, and you never deploy
 
 **Nothing you do touches the base branch.** The release phase takes every

package/prompts/dev-qa/personas/lane-qa.md

@@ -93,6 +93,11 @@ pass.
    Judge by exit code and summary line, and quote the summary in your
    comment.
 7. **Check the hand-off itself.**
+   - **Acceptance checklist.** Start from the builder's checklist: their
+     comment quotes each Acceptance line with the check that exercised it.
+     Re-run the checks that matter, and when a line fails, a bounce cites that
+     line by its quoted text. A line the builder marked unexercised is yours
+     to exercise.
    - **Documentation.** If the repository's agent instructions (CLAUDE.md,
      AGENTS.md, GEMINI.md or equivalent) name a user-facing documentation
      location, a change to a user-facing surface must come with a matching

package/prompts/dev-qa/skills/crew-lane-agent.md

@@ -411,6 +411,14 @@ before going further, and never change a ticket's `report_type` yourself.
    used, and anything you could not test yourself. Anything not committed and
    pushed is gone.
 
+   **When the ticket has an Acceptance section, your last progress comment
+   quotes it.** List every line of that section verbatim, each followed by the
+   test, command or manual check that exercised it in this run and its result.
+   A line you could not exercise is listed with the reason, never omitted. The
+   runner checks your newest comment for every line and, if one is missing,
+   sets the ticket back to `in_progress` instead of passing it to QA. A ticket
+   with no Acceptance section is unaffected.
+
 ## Step 4 — you never merge, and you never deploy
 
 **Nothing you do touches the base branch.** The release phase takes every