oh-my-codex 0.18.7 → 0.18.9

package/src/scripts/codex-native-hook.ts

@@ -48,6 +48,7 @@ import {
   type SkillActiveState,
 } from "../hooks/keyword-detector.js";
 import { buildDeepInterviewConfigInstruction } from "../hooks/deep-interview-config-instruction.js";
+import { readTeamModeConfig } from "../config/team-mode.js";
 import {
   detectNativeStopStallPattern,
   loadAutoNudgeConfig,
@@ -82,6 +83,7 @@ import {
   onSessionStart as buildWikiSessionStartContext,
 } from "../wiki/lifecycle.js";
 import { readAutoresearchCompletionStatus, readAutoresearchModeStateForActiveDecision } from "../autoresearch/skill-validation.js";
+import { normalizeAutopilotPhase } from "../autopilot/fsm.js";
 import { readRunState } from "../runtime/run-state.js";
 import { evaluateRalphCompletionAuditEvidence, isRalphCompletePhase } from "../ralph/completion-audit.js";
 import { getRunContinuationSnapshot, shouldContinueRun } from "../runtime/run-loop.js";
@@ -189,6 +191,25 @@ function resolveHudReconcileSessionId(
   return canonicalSessionId || sessionIdForState || undefined;
 }
 
+function resolveHudReconcileSessionIds(
+  currentSessionState: SessionState | null,
+  canonicalSessionId: string | null,
+  sessionIdForState: string | null,
+  nativeSessionId: string | null,
+): string[] {
+  const ownerOmxSessionId = safeString(currentSessionState?.owner_omx_session_id).trim();
+  return uniqueNonEmpty([
+    resolveHudReconcileSessionId(currentSessionState, canonicalSessionId, sessionIdForState),
+    canonicalSessionId ?? undefined,
+    sessionIdForState ?? undefined,
+    nativeSessionId ?? undefined,
+    safeString(currentSessionState?.session_id),
+    safeString(currentSessionState?.native_session_id),
+    OMX_OWNER_SESSION_ID_PATTERN.test(ownerOmxSessionId) ? ownerOmxSessionId : undefined,
+    safeString(currentSessionState?.owner_codex_session_id),
+  ]);
+}
+
 function safeContextSnippet(value: unknown, maxLength = 300): string {
   const text = safeString(value).replace(/\s+/g, " ").trim();
   if (text.length <= maxLength) return text;
@@ -1730,6 +1751,7 @@ function buildNativeOutsideTmuxTeamPromptBlockState(
   threadId?: string,
   turnId?: string,
 ): SkillActiveState | null {
+  if (!readTeamModeConfig(cwd).enabled) return null;
   const match = detectPrimaryKeyword(prompt);
   if (match?.skill !== "team") return null;
 
@@ -1765,11 +1787,14 @@ function buildSkillStateCliInstruction(mode: string, statePath: string): string
 
 function buildAutopilotPromptActivationNote(
   skillState?: SkillActiveState | null,
-  options: { markedQuestionAnswer?: boolean } = {},
+  options: { markedQuestionAnswer?: boolean; cwd?: string } = {},
 ): string | null {
   if (skillState?.initialized_mode !== "autopilot") return null;
+  const teamHandoff = readTeamModeConfig(options.cwd).enabled
+    ? " (+ $team if needed)"
+    : "";
   return [
-    "Autopilot protocol: the durable default chain is $deep-interview -> $ralplan -> $ultragoal (+ $team if needed) -> $code-review -> $ultraqa (deep-interview -> ralplan -> ultragoal -> code-review -> ultraqa).",
+    `Autopilot protocol: the durable default chain is $deep-interview -> $ralplan -> $ultragoal${teamHandoff} -> $code-review -> $ultraqa (deep-interview -> ralplan -> ultragoal -> code-review -> ultraqa).`,
     "Start/resume at current_phase=deep-interview unless the task is clear and bounded; if deep-interview is intentionally skipped, persist and state an explicit deep_interview_gate.skip_reason before moving to ralplan.",
     "Deep-interview is a structured question chain, not a one-question gate: after an omx question answer, re-score ambiguity against the active threshold, treat max_rounds as a cap, and crystallize once ambiguity is at or below threshold and readiness gates pass.",
     options.markedQuestionAnswer
@@ -1782,6 +1807,12 @@ function buildAutopilotPromptActivationNote(
   ].filter(Boolean).join(" ");
 }
 
+function formatExecutionHandoffList(cwd: string): string {
+  return readTeamModeConfig(cwd).enabled
+    ? "`$ultragoal`, `$team`, or `$ralph`"
+    : "`$ultragoal` or `$ralph`";
+}
+
 function buildAdditionalContextMessage(
   prompt: string,
   skillState?: SkillActiveState | null,
@@ -1790,8 +1821,9 @@ function buildAdditionalContextMessage(
 ): string | null {
   if (!prompt) return null;
   const promptPriorityMessage = buildPromptPriorityMessage(prompt);
-  const matches = detectKeywords(prompt);
-  const match = detectPrimaryKeyword(prompt);
+  const teamMode = readTeamModeConfig(cwd);
+  const matches = detectKeywords(prompt).filter((entry) => teamMode.enabled || entry.skill !== "team");
+  const match = matches[0] ?? null;
   if (!match) {
     const continuedSkill = safeString(skillState?.skill).trim();
     if (!continuedSkill) return promptPriorityMessage;
@@ -1800,7 +1832,7 @@ function buildAdditionalContextMessage(
       : null;
     const deepInterviewConfigPromptActivationNote = buildDeepInterviewConfigInstruction(cwd, skillState);
     const markedQuestionAnswer = /^\s*\[omx question answered\]/i.test(prompt);
-    const autopilotPromptActivationNote = buildAutopilotPromptActivationNote(skillState, { markedQuestionAnswer });
+    const autopilotPromptActivationNote = buildAutopilotPromptActivationNote(skillState, { markedQuestionAnswer, cwd });
     return [
       `OMX native UserPromptSubmit continued active workflow skill "${continuedSkill}".`,
       promptPriorityMessage,
@@ -1826,7 +1858,7 @@ function buildAdditionalContextMessage(
       ? buildDeepInterviewQuestionBridgeInstruction(cwd, payload)
       : null;
     const deepInterviewConfigPromptActivationNote = buildDeepInterviewConfigInstruction(cwd, skillState);
-    const autopilotPromptActivationNote = buildAutopilotPromptActivationNote(skillState, { markedQuestionAnswer: true });
+    const autopilotPromptActivationNote = buildAutopilotPromptActivationNote(skillState, { markedQuestionAnswer: true, cwd });
     return [
       `OMX native UserPromptSubmit continued active workflow skill "${continuedSkill}"; workflow-like tokens inside the marked omx question answer are treated as answer text, not a new workflow activation.`,
       promptPriorityMessage,
@@ -1859,7 +1891,7 @@ function buildAdditionalContextMessage(
   const ultragoalPromptActivationNote = match.skill === "ultragoal"
     ? "Ultragoal protocol: use `omx ultragoal create-goals` / `complete-goals` / `checkpoint` for `.omx/ultragoal` artifacts, then use Codex goal model tools only from the active agent handoff (`get_goal`, `create_goal`, `update_goal`) and never overwrite a different active Codex goal. Ultragoal does not call `/goal clear`; for multiple sequential ultragoal runs in one Codex session/thread, manually clear the completed Codex goal in the UI before creating the next aggregate goal."
     : null;
-  const autopilotPromptActivationNote = buildAutopilotPromptActivationNote(skillState);
+  const autopilotPromptActivationNote = buildAutopilotPromptActivationNote(skillState, { cwd });
   const combinedTransitionMessage = (() => {
     if (!skillState?.transition_message) return null;
     if (matches.length <= 1 || activeSkills.length <= 1) return skillState.transition_message;
@@ -2464,8 +2496,11 @@ function readPayloadTurnId(payload: CodexHookPayload): string {
 async function resolveInternalSessionIdForPayload(
   cwd: string,
   payloadSessionId: string,
+  stateDir?: string,
 ): Promise<string> {
-  const currentSession = await readUsableSessionState(cwd);
+  const currentSession = stateDir
+    ? await readUsableSessionStateFromStateDir(cwd, stateDir)
+    : await readUsableSessionState(cwd);
   const canonicalSessionId = safeString(currentSession?.session_id).trim();
   if (!canonicalSessionId) return payloadSessionId;
 
@@ -2476,6 +2511,22 @@ async function resolveInternalSessionIdForPayload(
   return payloadSessionId;
 }
 
+async function readUsableSessionStateFromStateDir(
+  cwd: string,
+  stateDir: string,
+): Promise<SessionState | null> {
+  const sessionPath = join(stateDir, "session.json");
+  if (!existsSync(sessionPath)) return null;
+
+  try {
+    const content = await readFile(sessionPath, "utf-8");
+    const state = JSON.parse(content) as SessionState;
+    return isSessionStateUsable(state, cwd) ? state : null;
+  } catch {
+    return null;
+  }
+}
+
 async function readStopSessionPinnedState(
   fileName: string,
   cwd: string,
@@ -2514,7 +2565,8 @@ const PLANNING_MODE_IMPLEMENTATION_TOOL_NAMES = new Set([
 const DEEP_INTERVIEW_IMPLEMENTATION_TOOL_NAMES = PLANNING_MODE_IMPLEMENTATION_TOOL_NAMES;
 
 const RALPLAN_EXECUTION_HANDOFF_SKILLS = new Set([
-  "autopilot",
+  // Autopilot is intentionally excluded: it supervises planning phases such as
+  // ralplan/replan and is not by itself an execution authorization.
   "autoresearch",
   "ralph",
   "team",
@@ -2541,12 +2593,12 @@ function isActiveRalplanPhase(state: Record<string, unknown> | null): boolean {
   return true;
 }
 
-function isActiveAutopilotRalplanPhase(state: Record<string, unknown> | null): boolean {
-  if (!state || state.active !== true) return false;
-  const mode = safeString(state.mode).trim();
-  if (mode && mode !== "autopilot") return false;
-  const phase = safeString(state.current_phase ?? state.currentPhase).trim().toLowerCase();
-  return phase === "ralplan";
+function isAutopilotRalplanLikePhase(phase: string): boolean {
+  return normalizeAutopilotPhase(phase) === "ralplan";
+}
+
+function canAutopilotSkillMirrorSupplyRalplanPhase(phase: string): boolean {
+  return phase === "" || normalizeAutopilotPhase(phase) === "ralplan";
 }
 
 function hasExplicitExecutionHandoffSkill(
@@ -2649,7 +2701,7 @@ async function readActiveDeepInterviewStateForPreToolUse(
   const canonicalState = sessionId
     ? await readVisibleSkillActiveStateForStateDir(stateDir, sessionId)
     : await readSkillActiveState(join(stateDir, SKILL_ACTIVE_STATE_FILE));
-  if (!canonicalState) return modeState;
+  if (!canonicalState) return null;
   const hasActiveDeepInterviewSkill = listActiveSkills(canonicalState).some((entry) => (
     entry.skill === "deep-interview"
     && matchesSkillStopContext(entry, canonicalState, sessionId, threadId)
@@ -2671,7 +2723,7 @@ async function readActiveRalplanStateForPreToolUse(
     : await readSkillActiveState(join(stateDir, SKILL_ACTIVE_STATE_FILE));
   if (isActiveRalplanPhase(modeState) && modeState && modeStateMatchesSkillStopContext(modeState, cwd, sessionId)) {
     if (hasExplicitExecutionHandoffSkill(canonicalState, sessionId, threadId)) return null;
-    if (!canonicalState) return modeState;
+    if (!canonicalState) return null;
     const hasActiveRalplanSkill = listActiveSkills(canonicalState).some((entry) => (
       entry.skill === "ralplan"
       && matchesSkillStopContext(entry, canonicalState, sessionId, threadId)
@@ -2682,15 +2734,26 @@ async function readActiveRalplanStateForPreToolUse(
   const autopilotState = sessionId
     ? await readStopSessionPinnedState("autopilot-state.json", cwd, sessionId, stateDir)
     : await readJsonIfExists(join(stateDir, "autopilot-state.json"));
-  if (!isActiveAutopilotRalplanPhase(autopilotState) || !autopilotState) return null;
+  if (!autopilotState || autopilotState.active !== true) return null;
+  const autopilotMode = safeString(autopilotState.mode).trim();
+  if (autopilotMode && autopilotMode !== "autopilot") return null;
   if (!modeStateMatchesSkillStopContext(autopilotState, cwd, sessionId)) return null;
   const terminalAutopilotRunState = await readCanonicalTerminalRunStateForStop(cwd, sessionId, "autopilot");
   if (terminalAutopilotRunState) return null;
-  if (!canonicalState) return autopilotState;
+  if (!canonicalState) return null;
   const hasActiveAutopilotSkill = listActiveSkills(canonicalState).some((entry) => (
     entry.skill === "autopilot"
     && matchesSkillStopContext(entry, canonicalState, sessionId, threadId)
   ));
+  if (!hasActiveAutopilotSkill) return null;
+  const autopilotStatePhase = safeString(autopilotState.current_phase ?? autopilotState.currentPhase).trim().toLowerCase();
+  if (!canAutopilotSkillMirrorSupplyRalplanPhase(autopilotStatePhase)) return null;
+  const hasRalplanScopedAutopilotSkill = listActiveSkills(canonicalState).some((entry) => (
+    entry.skill === "autopilot"
+    && isAutopilotRalplanLikePhase(safeString(entry.phase).trim().toLowerCase())
+    && matchesSkillStopContext(entry, canonicalState, sessionId, threadId)
+  ));
+  if (!isAutopilotRalplanLikePhase(autopilotStatePhase) && !hasRalplanScopedAutopilotSkill) return null;
   return hasActiveAutopilotSkill ? autopilotState : null;
 }
 
@@ -2705,8 +2768,9 @@ async function buildRalplanPreToolUseBoundaryOutput(
   payload: CodexHookPayload,
   cwd: string,
   stateDir: string,
+  resolvedSessionId?: string,
 ): Promise<Record<string, unknown> | null> {
-  const sessionId = readPayloadSessionId(payload);
+  const sessionId = safeString(resolvedSessionId ?? readPayloadSessionId(payload)).trim();
   const threadId = readPayloadThreadId(payload);
   const activeState = await readActiveRalplanStateForPreToolUse(cwd, stateDir, sessionId, threadId);
   if (!activeState) return null;
@@ -2726,13 +2790,21 @@ async function buildRalplanPreToolUseBoundaryOutput(
   if (!blocked) return null;
 
   const phase = formatPhase(activeState.current_phase ?? activeState.currentPhase, "planning");
+  const activeMode = safeString(activeState.mode).trim().toLowerCase();
+  const planningModeLabel = activeMode === "autopilot" ? "Autopilot planning" : "Ralplan";
+  const planningModeDescription = activeMode === "autopilot"
+    ? "Autopilot is supervising a planning phase"
+    : "Ralplan is consensus-planning mode";
   return {
     decision: "block",
-    reason: `Ralplan is active (phase: ${phase}); implementation/write tools are blocked until an explicit execution handoff workflow is activated.`,
+    reason: `${planningModeLabel} is active (phase: ${phase}); implementation/write tools are blocked until an explicit execution handoff workflow is activated.`,
     hookSpecificOutput: {
       hookEventName: "PreToolUse",
       additionalContext:
-        "Ralplan is consensus-planning mode. Write only planning artifacts under `.omx/context/`, `.omx/plans/`, `.omx/specs/`, or required `.omx/state/` files. Do not edit implementation files or run implementation-focused writes from ralplan. To execute, first process an explicit handoff such as `$ultragoal`, `$team`, or `$ralph`, which must emit terminal ralplan state before implementation begins.",
+        `${planningModeDescription}. `
+        + "Write only planning artifacts under `.omx/context/`, `.omx/plans/`, `.omx/specs/`, or required `.omx/state/` files. "
+        + "Do not edit implementation files or run implementation-focused writes from planning phases. "
+        + `To execute, first process an explicit handoff such as ${formatExecutionHandoffList(cwd)}, which must emit terminal planning state before implementation begins.`,
     },
   };
 }
@@ -2741,8 +2813,9 @@ async function buildDeepInterviewPreToolUseBoundaryOutput(
   payload: CodexHookPayload,
   cwd: string,
   stateDir: string,
+  resolvedSessionId?: string,
 ): Promise<Record<string, unknown> | null> {
-  const sessionId = readPayloadSessionId(payload);
+  const sessionId = safeString(resolvedSessionId ?? readPayloadSessionId(payload)).trim();
   const threadId = readPayloadThreadId(payload);
   const activeState = await readActiveDeepInterviewStateForPreToolUse(cwd, stateDir, sessionId, threadId);
   if (!activeState) return null;
@@ -2768,7 +2841,7 @@ async function buildDeepInterviewPreToolUseBoundaryOutput(
     hookSpecificOutput: {
       hookEventName: "PreToolUse",
       additionalContext:
-        "Deep-interview is requirements/spec mode. Treat detailed user answers as interview/spec material, not implicit implementation authorization. You may write only deep-interview artifacts under `.omx/context/`, `.omx/interviews/`, `.omx/specs/`, or required `.omx/state/` files. To implement, first ask for or process an explicit transition such as `$ralplan`, `$autopilot`, `$ralph`, `$team`, or `$ultragoal`.",
+        `Deep-interview is requirements/spec mode. Treat detailed user answers as interview/spec material, not implicit implementation authorization. You may write only deep-interview artifacts under \`.omx/context/\`, \`.omx/interviews/\`, \`.omx/specs/\`, or required \`.omx/state/\` files. To implement, first ask for or process an explicit transition such as \`$ralplan\`, \`$autopilot\`, ${formatExecutionHandoffList(cwd)}.`,
     },
   };
 }
@@ -3027,6 +3100,7 @@ async function reconcileStaleRootSkillActiveStateForStop(
 function buildRalplanContinuationStatus(
   blocker: { phase: string; latestPlanPath?: string; planningComplete?: boolean; runOutcome?: string },
   activeSubagentCount: number,
+  cwd: string,
 ): { reason: string; systemMessage: string; stopReasonSuffix: string } {
   const phase = blocker.phase || "planning";
   const artifact = blocker.latestPlanPath
@@ -3062,7 +3136,7 @@ function buildRalplanContinuationStatus(
   }
 
   const completeHint = blocker.planningComplete
-    ? " The planning artifacts are present; if consensus is approved, emit terminal ralplan complete/approved handoff state and stop planning. Implementation must wait for an explicit $ultragoal, $team, or $ralph handoff."
+    ? ` The planning artifacts are present; if consensus is approved, emit terminal ralplan complete/approved handoff state and stop planning. Implementation must wait for an explicit ${formatExecutionHandoffList(cwd).replaceAll("`", "")} handoff.`
     : "";
 
   return {
@@ -3282,6 +3356,11 @@ async function maybeBuildOrdinaryStopNoProgressOutput(
   stateDir: string,
   canonicalSessionId?: string,
 ): Promise<Record<string, unknown> | null> {
+  const lastAssistantMessage = safeString(
+    payload.last_assistant_message ?? payload.lastAssistantMessage,
+  ).trim();
+  if (!lastAssistantMessage) return null;
+
   const statePath = join(stateDir, NATIVE_STOP_STATE_FILE);
   const state = await readJsonIfExists(statePath) ?? {};
   const sessions = safeObject(state.sessions);
@@ -3313,9 +3392,6 @@ async function maybeBuildOrdinaryStopNoProgressOutput(
   await mkdir(stateDir, { recursive: true });
   await writeFile(statePath, JSON.stringify({ ...state, sessions }, null, 2));
 
-  const stopHookActive = payload.stop_hook_active === true || payload.stopHookActive === true;
-  if (!stopHookActive) return null;
-
   const maxRepeats = parseBoundedPositiveInteger(
     process.env.OMX_NATIVE_STOP_NO_PROGRESS_MAX_REPEATS,
     ORDINARY_STOP_NO_PROGRESS_DEFAULT_MAX_REPEATS,
@@ -3520,7 +3596,7 @@ async function buildSkillStopOutput(
   const activeSubagentCount = subagentSummary?.activeSubagentThreadIds.length ?? 0;
 
   if (blocker.skill === "ralplan") {
-    const status = buildRalplanContinuationStatus(blocker, activeSubagentCount);
+    const status = buildRalplanContinuationStatus(blocker, activeSubagentCount, cwd);
     return {
       decision: "block",
       reason: status.reason,
@@ -3978,7 +4054,9 @@ export async function dispatchCodexNativeHook(
   // Native hooks must use the same authoritative runtime state root as HUD/MCP
   // when boxed/team roots are active; do not bypass it with cwd/.omx/state.
   const stateDir = getBaseStateDir(cwd);
-  await mkdir(stateDir, { recursive: true });
+  if (hookEventName !== "Stop") {
+    await mkdir(stateDir, { recursive: true });
+  }
 
   const omxEventName = mapCodexHookEventToOmxEvent(hookEventName);
   let skillState: SkillActiveState | null = null;
@@ -4179,16 +4257,23 @@ export async function dispatchCodexNativeHook(
         triageAdditionalContext = null;
       }
     }
+    const skipHudReconcileForDoctorSmoke = process.env.OMX_NATIVE_HOOK_DOCTOR_SMOKE === "1";
     const skipHudReconcileForTeamWorkerPane = !isSubagentPromptSubmit
       && await isConfirmedTeamWorkerPromptSubmitPane(cwd).catch(() => false);
-    if (!skipHudReconcileForTeamWorkerPane) {
+    if (!skipHudReconcileForDoctorSmoke && !skipHudReconcileForTeamWorkerPane) {
       const reconcileHudForPromptSubmitFn = options.reconcileHudForPromptSubmitFn ?? reconcileHudForPromptSubmit;
       const hudSessionId = resolveHudReconcileSessionId(
         currentSessionState,
         canonicalSessionId,
         sessionIdForState,
       );
-      await reconcileHudForPromptSubmitFn(cwd, { sessionId: hudSessionId }).catch(() => {});
+      const hudSessionIds = resolveHudReconcileSessionIds(
+        currentSessionState,
+        canonicalSessionId,
+        sessionIdForState,
+        nativeSessionId,
+      );
+      await reconcileHudForPromptSubmitFn(cwd, { sessionId: hudSessionId, sessionIds: hudSessionIds }).catch(() => {});
     }
   }
 
@@ -4248,8 +4333,12 @@ export async function dispatchCodexNativeHook(
       };
     }
   } else if (hookEventName === "PreToolUse") {
-    outputJson = await buildDeepInterviewPreToolUseBoundaryOutput(payload, cwd, stateDir)
-      ?? await buildRalplanPreToolUseBoundaryOutput(payload, cwd, stateDir)
+    const payloadSessionId = readPayloadSessionId(payload);
+    const preToolUseSessionId = payloadSessionId
+      ? await resolveInternalSessionIdForPayload(cwd, payloadSessionId, stateDir)
+      : "";
+    outputJson = await buildDeepInterviewPreToolUseBoundaryOutput(payload, cwd, stateDir, preToolUseSessionId)
+      ?? await buildRalplanPreToolUseBoundaryOutput(payload, cwd, stateDir, preToolUseSessionId)
       ?? buildNativePreToolUseOutput(payload);
   } else if (hookEventName === "PostToolUse") {
     if (detectMcpTransportFailure(payload)) {

package/src/scripts/demo-team-e2e.sh

@@ -89,10 +89,8 @@ OMX_TEAM_WORKER_LAUNCH_ARGS="${OMX_TEAM_WORKER_LAUNCH_ARGS:--c model_reasoning_e
 TEAM_STARTED=0
 cleanup() {
   if ((TEAM_STARTED == 1)); then
-    echo "[cleanup] shutting down team: $TEAM_NAME"
-    omx team shutdown "$TEAM_NAME" >/dev/null 2>&1 || true
-    echo "[cleanup] cleaning state for team: $TEAM_NAME"
-    omx team api cleanup --input "{\"team_name\":\"$TEAM_NAME\"}" --json >/dev/null 2>&1 || true
+    echo "[cleanup] force-cleaning demo team: $TEAM_NAME"
+    omx team api cleanup --input "{\"team_name\":\"$TEAM_NAME\",\"force\":true,\"confirm_issues\":true}" --json >/dev/null 2>&1 || true
   fi
 }
 trap cleanup EXIT
@@ -106,7 +104,13 @@ echo "OMX_TEAM_WORKER_CLI_MAP=$OMX_TEAM_WORKER_CLI_MAP"
 echo "OMX_TEAM_WORKER_LAUNCH_ARGS=$OMX_TEAM_WORKER_LAUNCH_ARGS"
 
 echo "[1/8] start team (${WORKER_COUNT} mixed workers)"
-omx team "${WORKER_COUNT}:executor" "$TEAM_TASK"
+START_OUTPUT="$(omx team "${WORKER_COUNT}:executor" "$TEAM_TASK")"
+echo "$START_OUTPUT"
+ACTUAL_TEAM_NAME="$(echo "$START_OUTPUT" | sed -nE 's/^Team started: ([^[:space:]]+)$/\1/p' | head -n 1)"
+if [[ -n "$ACTUAL_TEAM_NAME" && "$ACTUAL_TEAM_NAME" != "$TEAM_NAME" ]]; then
+  echo "TEAM_NAME_RESOLVED=$ACTUAL_TEAM_NAME"
+  TEAM_NAME="$ACTUAL_TEAM_NAME"
+fi
 TEAM_STARTED=1
 
 echo "[2/8] status"
@@ -174,8 +178,7 @@ SUMMARY_JSON="$(omx team api get-summary --input "$SUMMARY_INPUT" --json)"
 echo "$SUMMARY_JSON" | jq -e '.schema_version == "1.0" and .operation == "get-summary" and .ok == true' >/dev/null
 
 echo "[8/8] shutdown + cleanup"
-omx team shutdown "$TEAM_NAME"
-omx team api cleanup --input "{\"team_name\":\"$TEAM_NAME\"}" --json >/dev/null
+omx team api cleanup --input "{\"team_name\":\"$TEAM_NAME\",\"force\":true,\"confirm_issues\":true}" --json >/dev/null
 TEAM_STARTED=0
 
 echo "E2E demo complete."

package/src/scripts/eval/eval-parity-smoke.ts

@@ -9,7 +9,7 @@ if (build.status !== 0) {
 }
 
 const test = spawnSync('node', [
-  '--test',
+  'dist/scripts/run-test-files.js',
   'dist/autoresearch/__tests__/contracts.test.js',
   'dist/autoresearch/__tests__/runtime.test.js',
   'dist/cli/__tests__/autoresearch.test.js',

package/src/scripts/notify-hook/auto-nudge.ts

@@ -269,7 +269,9 @@ export async function syncSkillStateFromTurn(stateDir, payload) {
 
 
 export async function isDeepInterviewStateActive(stateDir, sessionId) {
-  const modeState = await readScopedJsonIfExists(stateDir, 'deep-interview-state.json', sessionId, null);
+  const modeState = typeof sessionId === 'string' && sessionId.trim()
+    ? await readScopedJsonIfExists(stateDir, 'deep-interview-state.json', sessionId, null)
+    : await readJsonIfExists(join(stateDir, 'deep-interview-state.json'), null);
   return Boolean(modeState && modeState.active === true);
 }
 

package/src/scripts/notify-hook/ralph-session-resume.ts

@@ -1,8 +1,7 @@
 import { existsSync } from 'fs';
 import { mkdir, readFile, readdir, rename, rm, stat, writeFile } from 'fs/promises';
-import { dirname, join, resolve } from 'path';
+import { dirname, join } from 'path';
 import { captureTmuxPaneFromEnv } from '../../state/mode-state-context.js';
-import { isSessionStateUsable } from '../../hooks/session.js';
 import { resolveCodexPane } from '../tmux-hook-engine.js';
 import { safeString } from './utils.js';
 
@@ -222,15 +221,10 @@ function readSessionIdFromEnvironment(env: NodeJS.ProcessEnv = process.env): str
 
 async function readCurrentOmxSessionId(stateDir: string, env: NodeJS.ProcessEnv = process.env): Promise<string> {
   const envSessionId = readSessionIdFromEnvironment(env);
-  if (envSessionId) {
-    const envScopedDir = join(stateDir, 'sessions', envSessionId);
-    if (existsSync(envScopedDir)) return envSessionId;
-  }
+  if (envSessionId) return envSessionId;
 
-  const cwd = resolve(stateDir, '..', '..');
   const session = await readJson(join(stateDir, 'session.json'));
   if (!session || typeof session !== 'object') return '';
-  if (!isSessionStateUsable(session as any, cwd)) return '';
   const sessionId = safeString(session?.session_id).trim();
   return SESSION_ID_PATTERN.test(sessionId) ? sessionId : '';
 }

package/src/scripts/notify-hook/state-io.ts

@@ -3,12 +3,10 @@
  */
 
 import { mkdir, readFile, readdir, writeFile } from 'fs/promises';
-import { dirname, join, resolve } from 'path';
-import { existsSync } from 'fs';
-import { isSessionStateUsable } from '../../hooks/session.js';
+import { dirname, join } from 'path';
+import { validateSessionId } from '../../mcp/state-paths.js';
 import { asNumber, safeString } from './utils.js';
 
-const SESSION_ID_PATTERN = /^[A-Za-z0-9_-]{1,64}$/;
 
 export { readdir };
 
@@ -25,46 +23,90 @@ function isSafeStateFileName(fileName: string): boolean {
     && !fileName.includes('\\');
 }
 
-function readSessionIdFromEnvironment(env: NodeJS.ProcessEnv = process.env): string | undefined {
-  const candidates = [env.OMX_SESSION_ID, env.CODEX_SESSION_ID, env.SESSION_ID];
-  for (const candidate of candidates) {
-    const sessionId = safeString(candidate).trim();
-    if (!SESSION_ID_PATTERN.test(sessionId)) continue;
-    return sessionId;
+interface SessionMetadata {
+  sessionId?: string;
+  nativeAliases: string[];
+}
+
+async function readSessionMetadataFromBaseStateDir(baseStateDir: string): Promise<SessionMetadata> {
+  const session = await readJsonIfExists(join(baseStateDir, 'session.json'), null);
+  let sessionId: string | undefined;
+  try {
+    sessionId = validateSessionId(session?.session_id);
+  } catch {
+    sessionId = undefined;
   }
-  return undefined;
+  const nativeAliases = [
+    session?.native_session_id,
+    session?.codex_session_id,
+    session?.previous_native_session_id,
+  ]
+    .map((value) => safeString(value).trim())
+    .filter(Boolean);
+  return { sessionId, nativeAliases: [...new Set(nativeAliases)] };
 }
 
-export async function readCurrentSessionId(baseStateDir: string): Promise<string | undefined> {
-  const envSessionId = readSessionIdFromEnvironment();
-  if (envSessionId) {
-    const envScopedDir = join(baseStateDir, 'sessions', envSessionId);
-    if (existsSync(envScopedDir)) return envSessionId;
+function readSessionIdFromEnvironment(): string | undefined {
+  for (const candidate of [process.env.OMX_SESSION_ID, process.env.CODEX_SESSION_ID, process.env.SESSION_ID]) {
+    if (typeof candidate !== 'string') continue;
+    const trimmed = candidate.trim();
+    if (!trimmed) continue;
+    try {
+      const sessionId = validateSessionId(trimmed);
+      if (sessionId) return sessionId;
+    } catch {
+      continue;
+    }
   }
+  return undefined;
+}
 
-  const cwd = resolve(baseStateDir, '..', '..');
-  const session = await readJsonIfExists(join(baseStateDir, 'session.json'), null);
-  if (!session || typeof session !== 'object') return undefined;
-  if (!isSessionStateUsable(session, cwd)) return undefined;
-  const sessionId = safeString(session?.session_id);
-  return SESSION_ID_PATTERN.test(sessionId) ? sessionId : undefined;
+function resolveCanonicalSessionId(candidate: string | undefined, metadata: SessionMetadata): string | undefined {
+  if (!candidate) return undefined;
+  return metadata.sessionId && metadata.nativeAliases.includes(candidate)
+    ? metadata.sessionId
+    : candidate;
 }
 
-export async function resolveScopedStateDir(
+async function resolveBaseScopedStateDir(
   baseStateDir: string,
   explicitSessionId?: string,
 ): Promise<string> {
-  const normalizedExplicit = safeString(explicitSessionId).trim();
-  if (SESSION_ID_PATTERN.test(normalizedExplicit)) {
-    return join(baseStateDir, 'sessions', normalizedExplicit);
-  }
+  const normalizedExplicit = typeof explicitSessionId === 'string' && explicitSessionId.trim()
+    ? explicitSessionId.trim()
+    : undefined;
+  const validatedExplicit = validateSessionId(normalizedExplicit);
+  const metadata = await readSessionMetadataFromBaseStateDir(baseStateDir);
+  const sessionId = resolveCanonicalSessionId(validatedExplicit, metadata)
+    ?? resolveCanonicalSessionId(readSessionIdFromEnvironment(), metadata)
+    ?? metadata.sessionId;
+  return sessionId ? join(baseStateDir, 'sessions', sessionId) : baseStateDir;
+}
 
-  const currentSessionId = await readCurrentSessionId(baseStateDir);
-  if (currentSessionId) {
-    return join(baseStateDir, 'sessions', currentSessionId);
-  }
+async function resolveBaseScopedStateDirs(
+  baseStateDir: string,
+  explicitSessionId?: string,
+  options: { includeRootFallback?: boolean } = {},
+): Promise<string[]> {
+  const scopedDir = await resolveBaseScopedStateDir(baseStateDir, explicitSessionId);
+  return options.includeRootFallback === true && scopedDir !== baseStateDir
+    ? [scopedDir, baseStateDir]
+    : [scopedDir];
+}
+
+
+
+
+export async function readCurrentSessionId(baseStateDir: string): Promise<string | undefined> {
+  const metadata = await readSessionMetadataFromBaseStateDir(baseStateDir);
+  return resolveCanonicalSessionId(readSessionIdFromEnvironment(), metadata) ?? metadata.sessionId;
+}
 
-  return baseStateDir;
+export async function resolveScopedStateDir(
+  baseStateDir: string,
+  explicitSessionId?: string,
+): Promise<string> {
+  return resolveBaseScopedStateDir(baseStateDir, explicitSessionId);
 }
 
 export async function getScopedStateDirsForCurrentSession(
@@ -72,11 +114,7 @@ export async function getScopedStateDirsForCurrentSession(
   explicitSessionId?: string,
   options: { includeRootFallback?: boolean } = {},
 ): Promise<string[]> {
-  const scopedDir = await resolveScopedStateDir(baseStateDir, explicitSessionId);
-  if (scopedDir === baseStateDir || options.includeRootFallback !== true) {
-    return [scopedDir];
-  }
-  return [scopedDir, baseStateDir];
+  return resolveBaseScopedStateDirs(baseStateDir, explicitSessionId, options);
 }
 
 export async function getScopedStatePath(

package/src/scripts/notify-hook/team-leader-nudge.ts

@@ -25,6 +25,7 @@ import { writeTeamLeaderAttention } from '../../team/state.js';
 import { readLatestTeamProgressEvidenceMs } from '../../team/progress-evidence.js';
 import { validateSessionId } from '../../mcp/state-paths.js';
 import { TEAM_NAME_SAFE_PATTERN } from '../../team/contracts.js';
+import { isDeepInterviewStateActive } from './auto-nudge.js';
 const LEADER_PANE_MISSING_NO_INJECTION_REASON = 'leader_pane_missing_no_injection';
 const LEADER_PANE_SHELL_NO_INJECTION_REASON = 'leader_pane_shell_no_injection';
 const LEADER_PANE_SAME_CLASSIFIED_STATE_SUPPRESSED_REASON = 'pane_already_shows_same_classified_state';
@@ -571,6 +572,12 @@ export async function maybeNudgeTeamLeader({
 
   const candidateTeamNames = new Set();
   const currentSessionId = await resolveCurrentSessionId(stateDir);
+  const deepInterviewActive = currentSessionId
+    ? await isDeepInterviewStateActive(stateDir, currentSessionId).catch(() => false)
+    : await isDeepInterviewStateActive(stateDir, undefined).catch(() => false);
+  if (deepInterviewActive) {
+    return;
+  }
   try {
     const scopedDirs = await getScopedStateDirsForCurrentSession(stateDir);
     const candidateStateDirs = [...new Set([...scopedDirs, stateDir])];