muse-crew 0.7.10 → 0.7.12

package/workflows/crew-dispatch.js

@@ -103,18 +103,44 @@ phase("dispatch");
 // the schema demanded a top-level object with ready_tasks. The LLM resolved
 // the contradiction non-deterministically: some ticks validated, some
 // burned retries and blocked the whole dispatcher.)
-const boardReturn = await agent(
-  "Read the dispatch state.\n" +
-  "Run in shell and return the stdout as a raw string:\n" + crewCmd("get-dispatch-state", {}) + "\n" +
-  "Return the command's stdout JSON as a plain string, byte-for-byte, unmodified. " +
-  "Do NOT parse the JSON — your return value must be the raw stdout string, never an object. " +
-  "Do not select fields, do not rewrite, summarize, paraphrase, or reformat anything. " +
-  "The result has keys ready_tasks, projects, config, counts — ferry the string exactly as received.",
-  {
-    key: "read-board",
-    label: "Reading board state"
+// Bounded in-tick retry (canary 2026-09-15): the read-board agent
+// occasionally ferries malformed JSON (observed twice: a literal
+// `.replace()` code fragment appended to the dispatch-state string).
+// A malformed ferry used to abort the whole dispatcher tick; retrying
+// the identical dispatcher call recovered both times. Retry the read
+// in-tick instead of burning a tick — two attempts with distinct
+// replay keys, fail closed after that.
+var boardReturn = null;
+var boardReadError = null;
+for (var readAttempt = 1; readAttempt <= 2; readAttempt++) {
+  var readBoardKey = readAttempt === 1 ? "read-board" : "read-board-r" + readAttempt;
+  try {
+    boardReturn = await agent(
+      "Read the dispatch state.\n" +
+      "Run in shell and return the stdout as a raw string:\n" + crewCmd("get-dispatch-state", {}) + "\n" +
+      "Return the command's stdout JSON as a plain string, byte-for-byte, unmodified. " +
+      "Do NOT parse the JSON — your return value must be the raw stdout string, never an object. " +
+      "Do not select fields, do not rewrite, summarize, paraphrase, or reformat anything. " +
+      "The result has keys ready_tasks, projects, config, counts — ferry the string exactly as received.",
+      {
+        key: readBoardKey,
+        label: "Reading board state" + (readAttempt > 1 ? " (attempt " + readAttempt + ")" : "")
+      }
+    );
+    // Parse eagerly inside the attempt so a malformed ferry retries
+    // instead of aborting the tick. (function declarations hoist.)
+    parseBoardJson(boardReturn);
+    boardReadError = null;
+    break;
+  } catch (readErr) {
+    boardReadError = readErr;
+    log("WARNING: read-board attempt " + readAttempt + " failed (" + (readErr && readErr.message ? readErr.message : readErr) + ")" + (readAttempt < 2 ? " — retrying" : " — attempts exhausted"));
+    boardReturn = null;
   }
-);
+}
+if (boardReturn === null) {
+  throw new Error("read-board: all 2 attempts failed to ferry parseable board JSON: " + (boardReadError && boardReadError.message ? boardReadError.message : boardReadError));
+}
 
 // Deterministic board parse — the read-board agent is told to return the
 // CLI stdout as a raw string; the workflow parses it here. Fails closed:
@@ -210,6 +236,10 @@ function projectTaskRecord(t) {
     blocked: t.blocked,
     deps: Array.isArray(t.deps) ? t.deps.slice() : [],
     priority: t.priority,
+    // next_phase is the one-shot recovery routing set by recover-task. It is
+    // ferried verbatim here and consumed (validated, routed, cleared) by the
+    // deterministic eligibility logic below — never interpreted by the LLM.
+    next_phase: (typeof t.next_phase === "string" && t.next_phase.length > 0) ? t.next_phase : null,
     latest_session: ls ? { id: ls.id, status: ls.status, step: ls.step, notes: ls.notes } : null,
     retry: {
       consecutive_failures: (typeof cf === "number" && cf >= 0) ? cf : 0,
@@ -424,7 +454,7 @@ function buildPlaytestDescription(journey, personaName, personaFile, maxFilings)
     "\n" +
     "Rules:\n" +
     "- Walk the journey end to end from a user's perspective, wearing the persona.\n" +
-    "- Use the normal QA surface for this project (visual inspection via artifact_inspect for artifact projects; read-only dashboard API checks otherwise).\n" +
+    "- Use the normal QA surface for this project (the published artifact's own user interface for artifact projects; read-only dashboard API checks otherwise). Note: there is currently no agent-callable visual-inspection tool (artifact_inspect was removed by the platform 2026-09-14; artifact.inspect is malfunction diagnosis, not a substitute) — judge only what you can observe directly.\n" +
     "- There is no code change under test: skip change-verification steps (provenance / publish checks) and judge only what you observe.\n" +
     "- " + filingRule + "\n" +
     "- File via the existing createtask action with filed_by \"hazel\": set workflow to \"standard\" when the fix is clear feature work, or omit workflow (untriaged) when unsure — triage routes untriaged filings to bugfix/feature as usual.\n"
@@ -497,6 +527,30 @@ for (var t = 0; t < allTasks.length; t++) {
   var workflow = task.workflow || "standard";
   var steps = WORKFLOWS[workflow] || WORKFLOWS.standard;
 
+  // One-shot recovery routing: recover-task sets next_phase to an explicit
+  // human-chosen phase. It takes precedence over every session-derived path
+  // below — the human's redirect wins over the automatic resume — but never
+  // over work in flight, and never on a task in a terminal state.
+  // Invalid values fail closed: logged, skipped, and NOT cleared, so a
+  // human can fix it with a corrected recover-task. Valid values bypass
+  // the retry cap by design: an explicit redirect is not an automatic
+  // retry and must not consume retry budget. The launched workflow consumes
+  // (clears) next_phase on its successful self-claim, exactly once.
+  if (task.next_phase) {
+    if (task.state !== "todo" && task.state !== "in_progress") {
+      log("Skipped \"" + task.title + "\" — next_phase \"" + task.next_phase + "\" set but state is " + task.state + "; left set for inspection");
+      continue;
+    }
+    if (latest && latest.status === "running") continue; // work in flight
+    var npIdx = steps.indexOf(task.next_phase);
+    if (npIdx < 0) {
+      log("Skipped \"" + task.title + "\" — next_phase \"" + task.next_phase + "\" not in " + workflow + " step registry; left set for a corrected recover-task");
+      continue;
+    }
+    eligible.push({ task: task, startStep: npIdx, reason: "next_phase", workflow: workflow, nextPhase: task.next_phase });
+    continue;
+  }
+
   if (task.state === "todo") {
     // Idle playtests always enter at QA — the filing carries the full
     // assignment, so Triage/Map have nothing to add.
@@ -844,7 +898,12 @@ for (var p = 0; p < toProcess.length; p++) {
     // whether the task record had no workflow (playtest filings carry explicit
     // workflow:'standard', so they correctly get false).
     resolved_workflow: iworkflow,
-    workflow_was_null: (itask.workflow == null)
+    workflow_was_null: (itask.workflow == null),
+    // One-shot recovery routing: the value the dispatcher routed on. The
+    // workflow's successful self-claim consumes (clears) it atomically via
+    // claim-task's expected_next_phase — only an exact match clears, so a
+    // newer recover-task written in the race window survives.
+    next_phase: item.nextPhase || null
   };
 
   log("Recommended " + iworkflow + " for \"" + itask.title + "\" [" + taskProject + "] at step " + nextStepName);

package/workflows/docs.js

@@ -22,6 +22,15 @@ const startStepIndex = inputs.start_step_index || 0;
 // resolution back via updatetask in the self-claim below.
 const RESOLVED_WORKFLOW = inputs.resolved_workflow || null;
 const WORKFLOW_WAS_NULL = inputs.workflow_was_null === true;
+// One-shot recovery routing: the dispatcher sets inputs.next_phase when it
+// routes this run via an explicit recover-task redirect. The value is
+// consumed (cleared) atomically by the successful self-claim below:
+// claim-task takes expected_next_phase and clears the matching next_phase in
+// the same transaction as the winning session insert, so no platform death
+// can slip between claim and consumption and replay the routing. A stale or
+// superseded routing survives — only an exact match clears.
+// what the dispatcher routed on.
+const NEXT_PHASE_ROUTED = (typeof inputs.next_phase === "string" && inputs.next_phase.length > 0) ? inputs.next_phase : null;
 
 // Config from args — backward-compatible fallbacks for manual launches
 const crewHome = inputs.crewHome || "~/workspace/.jarvis";
@@ -148,7 +157,7 @@ while (i < STEPS.length) {
     const claimResult = await agent(
       "Claim this task for the " + step.name + " step.\n" +
       "Run in shell and return the stdout verbatim:\n" + crewCmd("update-task", firstClaimUpdateArgs) + "\n" +
-      "Then run in shell and return the stdout verbatim:\n" + crewCmd("claim-task", { task_id: taskId, identity: step.identity, step: step.name, notes: step.name + " step started" }) + "\n" +
+      "Then run in shell and return the stdout verbatim:\n" + crewCmd("claim-task", { task_id: taskId, identity: step.identity, step: step.name, notes: step.name + " step started", ...(NEXT_PHASE_ROUTED ? { expected_next_phase: NEXT_PHASE_ROUTED } : {}) }) + "\n" +
       "Do not interpret the claim response. It already contains an explicit \"claimed\" field — copy it verbatim.\n" +
       "Return { claimed: <verbatim>, session_id: \"<...>\" }. If claimed is false there is no session_id; return { claimed: false, session_id: \"\" }.",
       {
@@ -393,7 +402,7 @@ function describeWorkAgentFailure(stepName, identity, attempts) {
 
   var instructions = "";
   if (step.name === "Triage") {
-    instructions = "Validate the task, check clarity, confirm the docs workflow assignment.\nReport back in plain prose — your assessment.";
+    instructions = "Validate the task against the project's repo at " + REPO_PATH + " — that exact checkout, not any other copy of the project on disk. If you run git commands, cd " + REPO_PATH + " first.\nCheck clarity, confirm the docs workflow assignment.\nReport back in plain prose — your assessment.";
   } else if (step.name === "Write") {
     instructions = "Write or revise the documentation the task asks for.\nFollow Tate's voice — clear, conversational, no jargon unless it earns its place.\nDo your work in the project repository at " + REPO_PATH + " — all doc files go there, not under the crew home." +
       (rejectionNotes ? "\n\nREWORK after review rejection. Address:\n" + rejectionNotes : "") +