diff --git a/.agents/agents/backend-engineer.md b/.agents/agents/backend-engineer.md
index 3204865..47776fd 100644
--- a/.agents/agents/backend-engineer.md
+++ b/.agents/agents/backend-engineer.md
@@ -40,7 +40,7 @@ Router (HTTP) → Service (Business Logic) → Repository (Data Access) → Mode
 ## Rules
 
 1. Stay in scope — only work on assigned backend tasks
-2. Write tests for all new code
+2. Write tests for all new code; honor the plan task's `test_approach` — for `tdd`, demonstrate RED before the change and record a `TDD_EVIDENCE` block (test command, RED, GREEN) in the result file
 3. Follow Repository → Service → Router pattern (no business logic in routes)
 4. Validate all inputs with the project's validation library
 5. Parameterized queries only (no string interpolation in SQL)
diff --git a/.agents/agents/debug-investigator.md b/.agents/agents/debug-investigator.md
index 7bcd41e..e3cde15 100644
--- a/.agents/agents/debug-investigator.md
+++ b/.agents/agents/debug-investigator.md
@@ -46,7 +46,7 @@ CHARTER_CHECK:
 1. Stay in scope — only work on assigned debug tasks
 2. Fix root cause, not symptoms
 3. Minimal changes only — no refactoring during bugfix; route refactoring needs to refactor-engineer
-4. Every fix gets a regression test
+4. Every fix gets a regression test; run it before the fix where feasible and record RED (failing output) → GREEN (post-fix pass) in the bug report
 5. Search for similar patterns after fixing
 6. Document out-of-scope findings for other agents
 7. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.agents/agents/frontend-engineer.md b/.agents/agents/frontend-engineer.md
index 134bdf7..2489dac 100644
--- a/.agents/agents/frontend-engineer.md
+++ b/.agents/agents/frontend-engineer.md
@@ -47,6 +47,6 @@ FSD-lite: root `src/` + feature `src/features/*/`
 5. TailwindCSS v4 for styling, design tokens 1:1 mapping
 6. Library defaults (greenfield; existing project choices win): luxon (dates), ahooks (hooks), es-toolkit (utils), jotai (client state), TanStack Query (server state)
 7. Absolute imports with `@/`
-8. Write tests for custom logic (>90% coverage target)
+8. Write tests for custom logic (>90% coverage target); honor the plan task's `test_approach` — for `tdd`, demonstrate RED before the change and record a `TDD_EVIDENCE` block (test command, RED, GREEN) in the result file
 9. Document out-of-scope dependencies for other agents
 10. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.agents/agents/mobile-engineer.md b/.agents/agents/mobile-engineer.md
index 5419fd7..31bba4d 100644
--- a/.agents/agents/mobile-engineer.md
+++ b/.agents/agents/mobile-engineer.md
@@ -46,7 +46,7 @@ Clean Architecture: domain → data → presentation (Swift native: App/Core/Fea
 5. Transport client with interceptors (Dio / axios / generated Client) + repository-layer response cache, offline-first architecture
 6. Secrets in secure storage only — never plain prefs or MMKV
 7. 60fps target performance
-8. Write widget/component tests and integration tests
+8. Write widget/component tests and integration tests; honor the plan task's `test_approach` — for `tdd`, demonstrate RED before the change and record a `TDD_EVIDENCE` block (test command, RED, GREEN) in the result file
 9. ARB-based localization: edit ARB source files only, never generated localization code
 10. Document out-of-scope dependencies for other agents
 11. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.agents/agents/pm-planner.md b/.agents/agents/pm-planner.md
index 87674c8..609d651 100644
--- a/.agents/agents/pm-planner.md
+++ b/.agents/agents/pm-planner.md
@@ -50,12 +50,13 @@ Each task must include:
 - `priority`: execution tier — 1 = independent (runs first), 2 = depends on tier 1, etc. (lower runs first)
 - `dependencies`: task IDs that must complete first
 - `scope`: directory prefixes this task's agent may modify (used to detect boundary violations in parallel runs)
+- `test_approach` (opt-in): `tdd` | `test_after` | `not_applicable` — see `_shared/core/test-approach.md`. `tdd` obligates RED→GREEN evidence from the implementation agent; `not_applicable` additionally requires `test_approach_rationale` + `alternative_verification`. Never assign `tdd` to refactor tasks (characterization tests instead)
 
 ## Rules
 
 1. Stay in scope — planning only, no code implementation
 2. API-first design
 3. Minimize dependencies for maximum parallelism
-4. Security and testing are part of every task (not separate)
+4. Security and testing are part of every task (not separate); assign per-task `test_approach` (`tdd|test_after|not_applicable`) where a test strategy matters — `not_applicable` requires rationale + alternative verification, and no approach waives the >= 80% coverage gate
 5. Each task completable by a single agent
 6. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.agents/hooks/core/keyword-detector.ts b/.agents/hooks/core/keyword-detector.ts
index e336814..08094f6 100644
--- a/.agents/hooks/core/keyword-detector.ts
+++ b/.agents/hooks/core/keyword-detector.ts
@@ -357,20 +357,35 @@ export function escapeRegex(s: string): string {
 
 /**
  * Merge a language-keyed keyword/pattern bank into a single flat list:
- * universal ("*") + English (the universal default) + the configured
- * language's own entries (skipped when lang === "en" to avoid duplicates).
- * Shared by buildPatterns and buildRawPatterns — both keyword banks and
- * pattern banks use this exact `Record<string, string[]>` shape.
+ * universal ("*") + English + EVERY other language's entries, deduped
+ * case-insensitively. Same rationale as RC4 (buildInformationalPatterns):
+ * users prompt in whichever language they think in — `language` in
+ * oma-config.yaml controls the RESPONSE language, not the prompt language —
+ * so gating by config language silently disabled e.g. every Korean trigger
+ * for `language: en` projects. A keyword written in language X can only
+ * match a prompt that contains X-script text (current banks are en/ko/ja/zh;
+ * if a Latin-script bank like es/fr is ever added, phrase distinctiveness is
+ * the gate instead), so merging all languages cannot fire on unrelated
+ * prompts. Shared by buildPatterns and buildRawPatterns — both keyword banks
+ * and pattern banks use this exact `Record<string, string[]>` shape.
  */
-export function collectLangEntries(
-  bank: Record<string, string[]>,
-  lang: string,
-): string[] {
-  return [
+export function collectLangEntries(bank: Record<string, string[]>): string[] {
+  const ordered = [
     ...(bank["*"] ?? []),
     ...(bank.en ?? []),
-    ...(lang !== "en" ? (bank[lang] ?? []) : []),
+    ...Object.entries(bank)
+      .filter(([key]) => key !== "*" && key !== "en")
+      .flatMap(([, entries]) => entries),
   ];
+  const seen = new Set<string>();
+  const out: string[] = [];
+  for (const entry of ordered) {
+    const key = entry.toLowerCase();
+    if (seen.has(key)) continue;
+    seen.add(key);
+    out.push(entry);
+  }
+  return out;
 }
 
 /**
@@ -390,7 +405,7 @@ export function buildPatternEntries(
   lang: string,
   cjkScripts: string[],
 ): KeywordPatternEntry[] {
-  return collectLangEntries(keywords, lang).map((kw) => {
+  return collectLangEntries(keywords).map((kw) => {
     const escaped = escapeRegex(kw).replace(/\s+/g, "\\s+");
     const regex =
       cjkScripts.includes(lang) || /[^\p{ASCII}]/u.test(kw)
@@ -421,11 +436,10 @@ export interface RawPatternEntry {
 
 export function buildRawPatternEntries(
   patterns: Record<string, string[]> | undefined,
-  lang: string,
 ): RawPatternEntry[] {
   if (!patterns) return [];
   const compiled: RawPatternEntry[] = [];
-  for (const raw of collectLangEntries(patterns, lang)) {
+  for (const raw of collectLangEntries(patterns)) {
     try {
       compiled.push({ regex: new RegExp(raw, "iu"), source: raw });
     } catch {
@@ -443,9 +457,8 @@ export function buildRawPatternEntries(
  */
 export function buildRawPatterns(
   patterns: Record<string, string[]> | undefined,
-  lang: string,
 ): RegExp[] {
-  return buildRawPatternEntries(patterns, lang).map((e) => e.regex);
+  return buildRawPatternEntries(patterns).map((e) => e.regex);
 }
 
 export function buildInformationalPatterns(config: TriggerConfig): RegExp[] {
@@ -707,6 +720,7 @@ function activateMode(
   projectDir: string,
   workflow: string,
   sessionId: string,
+  omaSid?: string | null,
 ): void {
   // Never persist a workflow under the unresolved-session fallback id: such a
   // file cannot be isolated per session and would cross-contaminate any later
@@ -718,6 +732,7 @@ function activateMode(
     sessionId,
     activatedAt: new Date().toISOString(),
     reinforcementCount: 0,
+    ...(omaSid ? { omaSid } : {}),
   };
   writeFileSync(
     join(getStateDir(projectDir), `${workflow}-state-${sessionId}.json`),
@@ -770,11 +785,12 @@ export const DEACTIVATION_PHRASES: Record<string, string[]> = {
   pl: ["workflow zakończony", "workflow ukończony"],
 };
 
-export function isDeactivationRequest(prompt: string, lang: string): boolean {
-  const phrases = [
-    ...(DEACTIVATION_PHRASES.en ?? []),
-    ...(lang !== "en" ? (DEACTIVATION_PHRASES[lang] ?? []) : []),
-  ];
+export function isDeactivationRequest(prompt: string): boolean {
+  // All languages merged, never gated by config language (same rationale as
+  // collectLangEntries): a user prompting in Korean must be able to say
+  // "워크플로우 완료" even when `language: en`. A phrase only matches a prompt
+  // actually written in that language, so merging cannot misfire.
+  const phrases = Object.values(DEACTIVATION_PHRASES).flat();
   const normalized = normalizeForMatching(prompt);
   return phrases.some((phrase) =>
     normalized.includes(normalizeForMatching(phrase)),
@@ -925,7 +941,7 @@ export async function run(
   const lang = detectLanguage(projectDir);
 
   // Check for deactivation request before workflow detection
-  if (isDeactivationRequest(prompt, lang)) {
+  if (isDeactivationRequest(prompt)) {
     deactivateAllPersistentModes(projectDir, sessionId);
     // Grok's resume context lives in a session-start file, not L1 stdout — clear it.
     if (vendor === "grok") clearGrokContext(projectDir);
@@ -1025,10 +1041,7 @@ export async function run(
     )) {
       considerMatch(regex, keyword);
     }
-    for (const { regex, source } of buildRawPatternEntries(
-      def.patterns,
-      lang,
-    )) {
+    for (const { regex, source } of buildRawPatternEntries(def.patterns)) {
       considerMatch(regex, source);
     }
   }
@@ -1038,10 +1051,17 @@ export async function run(
 
   const { workflow } = winner;
 
+  // Activate the L1 session first so its sid can be recorded in the
+  // persistent-mode state file (the Stop hook emits gate events under it).
+  const omaSid = await activateL1WorkflowSession(
+    projectDir,
+    workflow,
+    vendor,
+    sessionId,
+  );
   if (winner.persistent) {
-    activateMode(projectDir, workflow, sessionId);
+    activateMode(projectDir, workflow, sessionId, omaSid);
   }
-  await activateL1WorkflowSession(projectDir, workflow, vendor, sessionId);
   const updatedState = recordKwTrigger(kwState, workflow);
   saveKwState(projectDir, updatedState);
 
diff --git a/.agents/hooks/core/persistent-mode.ts b/.agents/hooks/core/persistent-mode.ts
index e20dda6..698a0c3 100644
--- a/.agents/hooks/core/persistent-mode.ts
+++ b/.agents/hooks/core/persistent-mode.ts
@@ -13,6 +13,7 @@
  * exit 2 = block stop
  */
 
+import { spawnSync } from "node:child_process";
 import {
   existsSync,
   readdirSync,
@@ -40,15 +41,123 @@ import { getProjectDir } from "./vendor-detect.ts";
 const MAX_REINFORCEMENTS = 5;
 const STALE_HOURS = 2;
 
-function detectLanguage(projectDir: string): string {
-  const prefsPath = join(projectDir, ".agents", "oma-config.yaml");
-  if (!existsSync(prefsPath)) return "en";
+// ── Goal contract: deterministic stop gate + wall-clock budget ─
+// (design-prime-agent-adoption Track B — no-exec-of-agent-writable-strings)
+
+/**
+ * The only gate values the Stop hook will ever execute. Each maps to a
+ * package.json script of the same name, run as an argv array WITHOUT a shell.
+ * The gate value lives in an agent-writable state file; executing anything
+ * outside this allowlist would be an arbitrary-command path that bypasses the
+ * PreToolUse permission layer. Never widen this to free-form strings.
+ */
+const GATE_KEYWORDS = new Set(["typecheck", "test", "lint"]);
+
+/** Hard cap on a gate run; SIGKILL after this. Keeps Stop-hook latency bounded. */
+const GATE_TIMEOUT_MS = 60_000;
+
+/** Tail of gate output carried back into the block reason. */
+const GATE_OUTPUT_TAIL_CHARS = 2_000;
+
+/**
+ * Resolve an allowlisted gate keyword to a package-runner argv, or null when
+ * the keyword is not allowlisted, package.json is absent, or it defines no
+ * script of that name. Pure node:fs — no shell, no third-party imports.
+ */
+export function resolveGateArgv(
+  gateKeyword: string,
+  projectDir: string,
+): string[] | null {
+  if (!GATE_KEYWORDS.has(gateKeyword)) return null;
+  const pkgPath = join(projectDir, "package.json");
+  if (!existsSync(pkgPath)) return null;
   try {
-    const content = readFileSync(prefsPath, "utf-8");
-    const match = content.match(/^language:\s*(\S+)/m);
-    return match?.[1] ?? "en";
+    const pkg = JSON.parse(readFileSync(pkgPath, "utf-8")) as {
+      scripts?: Record<string, unknown>;
+    };
+    if (typeof pkg.scripts?.[gateKeyword] !== "string") return null;
   } catch {
-    return "en";
+    return null;
+  }
+  if (
+    existsSync(join(projectDir, "bun.lock")) ||
+    existsSync(join(projectDir, "bun.lockb"))
+  ) {
+    return ["bun", "run", gateKeyword];
+  }
+  if (existsSync(join(projectDir, "pnpm-lock.yaml"))) {
+    return ["pnpm", "run", gateKeyword];
+  }
+  if (existsSync(join(projectDir, "yarn.lock"))) {
+    return ["yarn", gateKeyword];
+  }
+  return ["npm", "run", gateKeyword];
+}
+
+export interface GateRunResult {
+  passed: boolean;
+  timedOut: boolean;
+  outputTail: string;
+}
+
+/** Run a resolved gate argv with a hard timeout. No shell involved. */
+export function runGateCommand(
+  argv: string[],
+  projectDir: string,
+): GateRunResult {
+  const [command, ...args] = argv;
+  const result = spawnSync(command as string, args, {
+    cwd: projectDir,
+    encoding: "utf-8",
+    timeout: GATE_TIMEOUT_MS,
+    killSignal: "SIGKILL",
+    maxBuffer: 8 * 1024 * 1024,
+  });
+  const timedOut =
+    (result.error as NodeJS.ErrnoException | undefined)?.code === "ETIMEDOUT" ||
+    result.signal === "SIGKILL";
+  const combined = `${result.stdout ?? ""}\n${result.stderr ?? ""}`.trim();
+  return {
+    passed: result.status === 0 && !result.error,
+    timedOut,
+    outputTail: combined.slice(-GATE_OUTPUT_TAIL_CHARS),
+  };
+}
+
+/** True when the goal's wall-clock budget (from activatedAt) is exhausted. */
+export function isBudgetExhausted(state: ModeState): boolean {
+  const minutes = state.goal?.budget?.wallClockMinutes;
+  if (
+    typeof minutes !== "number" ||
+    !Number.isFinite(minutes) ||
+    minutes <= 0
+  ) {
+    return false;
+  }
+  const elapsedMs = Date.now() - new Date(state.activatedAt).getTime();
+  return elapsedMs >= minutes * 60_000;
+}
+
+/**
+ * Emit a gate event onto the L1 trail recorded at activation. Best-effort:
+ * older state files carry no omaSid, and event emission must never break the
+ * Stop decision itself.
+ */
+async function emitGateEvent(
+  projectDir: string,
+  state: ModeState,
+  kind: "gate.passed" | "gate.failed",
+  payload: Record<string, unknown>,
+): Promise<void> {
+  if (!state.omaSid) return;
+  try {
+    const { emitEvent } = await import("./state-emit.ts");
+    await emitEvent(projectDir, state.omaSid, {
+      kind,
+      payload: { workflow: state.workflow, ...payload },
+    });
+  } catch {
+    // best-effort — never let event I/O change the stop decision
   }
 }
 
@@ -218,8 +327,7 @@ export async function run(
   // text (parity with the standalone main() path). Without this, persistent
   // mode could not be deactivated via the central `oma hook` dispatch.
   if (input.responseText) {
-    const lang = detectLanguage(projectDir);
-    if (isDeactivationRequest(input.responseText, lang)) {
+    if (isDeactivationRequest(input.responseText)) {
       deactivateAllForSession(projectDir, sessionId);
       return null;
     }
@@ -236,16 +344,75 @@ export async function run(
       continue;
     }
 
-    incrementReinforcement(projectDir, workflow, sessionId, state);
+    // (1) Wall-clock budget: exhausted → honest partial stop. A machine
+    // verdict, not model discretion — the stop is allowed and the exhaustion
+    // is recorded on the L1 trail.
+    if (isBudgetExhausted(state)) {
+      deactivate(projectDir, workflow, sessionId);
+      await emitGateEvent(projectDir, state, "gate.failed", {
+        gate: "budget",
+        summary: `wall-clock budget (${state.goal?.budget?.wallClockMinutes}m) exhausted for /${workflow}; stopping with partial status`,
+      });
+      continue;
+    }
 
     const stateFile = `.agents/state/${workflow}-state-${sessionId}.json`;
+
+    // (2) Deterministic completion gate. Only allowlisted keywords resolve to
+    // a runnable argv; anything else (including free-form shell strings an
+    // agent may have written into the state file) is NEVER executed and falls
+    // through to the plain reinforcement block below.
+    const gateKeyword = state.goal?.completion?.gate;
+    let ignoredGateNote = "";
+    if (gateKeyword) {
+      const argv = resolveGateArgv(gateKeyword, projectDir);
+      if (argv) {
+        const gate = runGateCommand(argv, projectDir);
+        if (gate.passed) {
+          // The gate is the mechanical proof of completion: allow the stop.
+          deactivate(projectDir, workflow, sessionId);
+          await emitGateEvent(projectDir, state, "gate.passed", {
+            gate: gateKeyword,
+            summary: `stop gate '${gateKeyword}' passed for /${workflow}`,
+          });
+          continue;
+        }
+        // Failure and timeout both count toward MAX_REINFORCEMENTS so a
+        // permanently red gate cannot block stops forever.
+        incrementReinforcement(projectDir, workflow, sessionId, state);
+        await emitGateEvent(projectDir, state, "gate.failed", {
+          gate: gateKeyword,
+          timedOut: gate.timedOut,
+          summary: `stop gate '${gateKeyword}' ${gate.timedOut ? "timed out" : "failed"} for /${workflow}`,
+        });
+        const reason = [
+          `[OMA PERSISTENT MODE: ${workflow.toUpperCase()}]`,
+          `Stop gate '${gateKeyword}' ${gate.timedOut ? `timed out after ${GATE_TIMEOUT_MS / 1000}s` : "FAILED"} (reinforcement ${state.reinforcementCount}/${MAX_REINFORCEMENTS}).`,
+          `Fix the failures below, then finish the workflow — the stop is allowed only when the gate passes.`,
+          gate.outputTail
+            ? `--- gate output (tail) ---\n${gate.outputTail}`
+            : "",
+          `To abandon instead: delete ${stateFile} or say "workflow done".`,
+        ]
+          .filter(Boolean)
+          .join("\n");
+        return { type: "block", reason };
+      }
+      ignoredGateNote = `Note: configured stop gate ${JSON.stringify(gateKeyword)} is not an allowed keyword (typecheck|test|lint) or has no matching package.json script — it was NOT executed.`;
+    }
+
+    incrementReinforcement(projectDir, workflow, sessionId, state);
+
     const reason = [
       `[OMA PERSISTENT MODE: ${workflow.toUpperCase()}]`,
       `The /${workflow} workflow is still active (reinforcement ${state.reinforcementCount}/${MAX_REINFORCEMENTS}).`,
       `Continue executing the workflow. If all tasks are genuinely complete:`,
       `  1. Delete the state file: Bash \`rm ${stateFile}\``,
       `  2. Or ask the user to say "워크플로우 완료" / "workflow done"`,
-    ].join("\n");
+      ignoredGateNote,
+    ]
+      .filter(Boolean)
+      .join("\n");
 
     return { type: "block", reason };
   }
@@ -267,7 +434,6 @@ async function main() {
   const vendor = detectVendor(input);
   const projectDir = getProjectDir(vendor, input);
   const sessionId = getSessionId(input);
-  const lang = detectLanguage(projectDir);
 
   // Check all text fields in stdin for deactivation phrases.
   // The assistant may have included "workflow done" in its response,
@@ -284,7 +450,7 @@ async function main() {
     .filter((v): v is string => typeof v === "string")
     .join(" ");
 
-  if (textToCheck && isDeactivationRequest(textToCheck, lang)) {
+  if (textToCheck && isDeactivationRequest(textToCheck)) {
     // Deactivate all persistent workflows for this session (shared helper).
     deactivateAllForSession(projectDir, sessionId);
     process.exit(0);
diff --git a/.agents/hooks/core/serena-primer.ts b/.agents/hooks/core/serena-primer.ts
index df2aabc..07eaca2 100644
--- a/.agents/hooks/core/serena-primer.ts
+++ b/.agents/hooks/core/serena-primer.ts
@@ -117,6 +117,7 @@ export function primerContext(): string {
     "- Code discovery / reading: `get_symbols_overview`, `find_symbol`, `find_referencing_symbols`, `search_for_pattern`.",
     "- Code edits: `replace_symbol_body`, `insert_after_symbol`, `insert_before_symbol`, `replace_content`.",
     "- Native grep/glob: only for initial filename/path discovery. Do not fall back to grep + Read for code navigation just because Serena's tools aren't loaded yet — load them.",
+    '- Result size: omit `max_answer_chars` on Serena tools (uses the configured default, typically 150000). Never pass small caps like `3000` on broad searches. If a call returns "The answer is too long (N characters)", retry with `max_answer_chars` > N or narrow path/glob — do not keep the low cap.',
     "- Exception — MCP timeout: if a Serena MCP call times out or hangs (seen mainly in OpenCode Desktop's long-lived sidecar), stop retrying MCP for this session: use native search/read for code, and access `.serena/memories/` files directly (or `serena memories read|write` when Serena CLI ≥ 1.5 is installed) for memory work. A full app relaunch restores Serena MCP.",
   ].join("\n");
 }
diff --git a/.agents/hooks/core/skill-injector.ts b/.agents/hooks/core/skill-injector.ts
index 66c8139..06fe400 100644
--- a/.agents/hooks/core/skill-injector.ts
+++ b/.agents/hooks/core/skill-injector.ts
@@ -188,10 +188,16 @@ export function matchSkills(
     const jsonEntry = config.skills?.[skill.name];
     if (!jsonEntry) continue;
 
+    // All languages merged, never gated by config language: users prompt in
+    // whichever language they think in (`language` controls the RESPONSE
+    // language). A keyword written in language X can only match a prompt
+    // containing X-script text, so merging cannot fire on unrelated prompts.
     const jsonTriggers = [
       ...(jsonEntry.keywords["*"] ?? []),
       ...(jsonEntry.keywords.en ?? []),
-      ...(lang !== "en" ? (jsonEntry.keywords[lang] ?? []) : []),
+      ...Object.entries(jsonEntry.keywords)
+        .filter(([key]) => key !== "*" && key !== "en")
+        .flatMap(([, entries]) => entries),
     ];
 
     const seen = new Set<string>();
diff --git a/.agents/hooks/core/triggers.json b/.agents/hooks/core/triggers.json
index 8e5bff8..9aee86c 100644
--- a/.agents/hooks/core/triggers.json
+++ b/.agents/hooks/core/triggers.json
@@ -1211,12 +1211,10 @@
           "랄프",
           "멈추지마",
           "멈추지 말고",
-          "끝까지",
           "완료될때까지",
           "될때까지 해",
           "끝날때까지",
           "다 끝내",
-          "다 해",
           "전부 완료",
           "끝까지 해",
           "중단하지마",
diff --git a/.agents/hooks/core/types.ts b/.agents/hooks/core/types.ts
index a58ffda..219a0e1 100644
--- a/.agents/hooks/core/types.ts
+++ b/.agents/hooks/core/types.ts
@@ -35,11 +35,45 @@ export interface RawHookInput {
   stopReason?: string;
 }
 
+/**
+ * Optional goal contract for a persistent workflow (design-prime-agent-adoption
+ * Track B). Written by `oma goal:set`; read by the persistent-mode Stop hook.
+ */
+export interface ModeGoal {
+  /** Human description of the objective. Informational only. */
+  description?: string;
+  budget?: {
+    /**
+     * Wall-clock budget in minutes, measured from `activatedAt`. When
+     * exceeded the Stop hook deactivates the workflow and allows an honest
+     * partial stop (machine verdict, not model discretion).
+     */
+    wallClockMinutes?: number;
+  };
+  completion?: {
+    /**
+     * Deterministic stop gate. MUST be an allowlist keyword ("typecheck" |
+     * "test" | "lint") that maps to an existing package.json script; the hook
+     * runs it as an argv array with no shell. Free-form strings are NEVER
+     * executed — this value lives in an agent-writable state file, so
+     * executing it verbatim would be an arbitrary-command-execution path
+     * that bypasses the PreToolUse permission layer.
+     */
+    gate?: string;
+  };
+}
+
 export interface ModeState {
   workflow: string;
   sessionId: string;
   activatedAt: string;
   reinforcementCount: number;
+  /**
+   * L1 session id (`oma-…`) recorded at activation so the Stop hook can emit
+   * gate.passed / gate.failed events onto the same events.jsonl trail.
+   */
+  omaSid?: string;
+  goal?: ModeGoal;
 }
 
 // ---------------------------------------------------------------------------
diff --git a/.agents/oma-config.yaml b/.agents/oma-config.yaml
index 68564bb..c9f28bd 100644
--- a/.agents/oma-config.yaml
+++ b/.agents/oma-config.yaml
@@ -46,3 +46,8 @@ scm:
     - "*.example"
     - "*.sample"
     - "*.template"
+
+# Added by oma update — new config keys (template defaults; edit freely)
+agents:
+  eval:
+    model: anthropic/claude-sonnet-4-6
diff --git a/.agents/skills/_shared/core/context-budget.md b/.agents/skills/_shared/core/context-budget.md
index 206f2ba..b1aadb6 100644
--- a/.agents/skills/_shared/core/context-budget.md
+++ b/.agents/skills/_shared/core/context-budget.md
@@ -11,6 +11,13 @@ Follow this guide to use context efficiently.
 2. **No duplicate reads**: Do not re-read files already read
 3. **Lazy resource loading**: Load resources only when needed
 4. **Maintain records**: Note read files and symbols in progress
+5. **Run functions over data, don't read data into context**: when a scene
+   processes bulk data (harvest results, logs, transcripts, large JSON), do the
+   processing through a deterministic tool/CLI stage (`CALL_TOOL`) and bring
+   back only a summary plus the artifact path. Streaming raw data through the
+   context spends tokens reading what a program could have computed — the
+   `oma market` pipe stages (harvest → score → fuse → cluster stay in JSON;
+   only the rendered brief path returns) are the reference pattern.
 
 ---
 
diff --git a/.agents/skills/_shared/core/test-approach.md b/.agents/skills/_shared/core/test-approach.md
new file mode 100644
index 0000000..41ca3f7
--- /dev/null
+++ b/.agents/skills/_shared/core/test-approach.md
@@ -0,0 +1,53 @@
+# Per-Task Test Approach & TDD Evidence
+
+Opt-in, per-task test strategy carried in the PM plan (`plan-{sessionId}.json`).
+PM assigns it (see `oma-pm/resources/execution-protocol.md` Step 3); implementation
+agents honor it; QA and Ultrawork gates verify evidence **only** for tasks marked `tdd`.
+
+## Approaches
+
+| `test_approach` | Meaning | Agent obligation |
+|---|---|---|
+| `tdd` | Deterministic, high-risk behavior (validation, authorization, state transitions, calculations, error handling) | Write and run the focused test **before** the production change (RED), make the minimal change (GREEN), refactor only if needed. Record a `TDD_EVIDENCE` block in the result file. |
+| `test_after` | Automated tests required, but a useful isolated RED state is impractical | Write tests with/after the implementation. No evidence block required. |
+| `not_applicable` | Automated tests inappropriate | Perform the plan's `alternative_verification` and report its outcome. Plan must carry `test_approach_rationale`. |
+
+Tasks without a `test_approach` field behave as today (tests per the agent's
+normal protocol). Refactor tasks never use `tdd` — they keep the
+characterization-test safety net (`oma-refactor`).
+
+## Coverage non-waiver rule
+
+No `test_approach` value relaxes the global unit-test coverage gate
+(**>= 80%**, QA checklist / SHIP_GATE). `not_applicable` code still counts
+toward the aggregate; excluding it requires a declarative entry in the
+project's coverage config with justification — never a silent drop.
+
+## TDD_EVIDENCE block format
+
+Append to the agent's result file (`result-{agent}.md`), one entry per `tdd` task:
+
+```
+TDD_EVIDENCE:
+- task: task-2
+  test_command: bun test src/services/discount.test.ts
+  red: "expected 400, received 200" (before implementation)
+  green: 12 pass, 0 fail (after implementation)
+```
+
+Requirements (enforced by `oma verify <agent>` → "TDD Evidence" check):
+
+1. Block starts with the literal marker `TDD_EVIDENCE:`
+2. Every `tdd` task id from the plan appears in the block
+3. At least one `red:` entry (the observed failure before the change) and one
+   `green:` entry (the passing result after the change)
+
+A task may opt out with `tdd_evidence_required: false` in the plan (e.g., the
+RED state is demonstrated in a linked CI run instead); the rationale belongs in
+the task description.
+
+## Debug parity
+
+Debug regression tests follow the same discipline where feasible: run the
+regression test before applying the fix, record the failing output (RED) and
+the post-fix pass (GREEN) in the bug report / result file.
diff --git a/.agents/skills/_version.json b/.agents/skills/_version.json
index 0070ef0..a33d2fa 100644
--- a/.agents/skills/_version.json
+++ b/.agents/skills/_version.json
@@ -1,6 +1,6 @@
 {
-  "version": "11.1.1",
+  "version": "11.10.3",
   "schemaVersion": 2,
   "mode": "project",
-  "installedAt": "2026-07-27T12:13:46.129Z"
+  "installedAt": "2026-08-10T10:10:15.956Z"
 }
diff --git a/.agents/skills/oma-backend/resources/execution-protocol.md b/.agents/skills/oma-backend/resources/execution-protocol.md
index c338b64..a477993 100644
--- a/.agents/skills/oma-backend/resources/execution-protocol.md
+++ b/.agents/skills/oma-backend/resources/execution-protocol.md
@@ -28,6 +28,7 @@ Follow these steps in order (adjust depth by difficulty).
 - Identify security requirements (auth, validation, rate limiting)
 
 ## Step 3: Implement
+- **Honor the task's `test_approach`** (see `../../_shared/core/test-approach.md`): for `tdd` tasks, write and run the focused test first (record the RED failure), make the minimal change (GREEN), then continue; for `tdd` the test comes before item 3 below
 - Create/modify files in this order:
   1. Database models + migrations
   2. Validation schemas (request/response)
@@ -41,6 +42,7 @@ Follow these steps in order (adjust depth by difficulty).
 - Run `resources/checklist.md` items
 - Run `../../_shared/core/common-checklist.md` items
 - Ensure all tests pass
+- For `tdd` tasks, append the `TDD_EVIDENCE` block (test command, RED, GREEN) to the result file per `../../_shared/core/test-approach.md`
 - Confirm OpenAPI docs are complete
 
 ## On Error
diff --git a/.agents/skills/oma-debug/resources/execution-protocol.md b/.agents/skills/oma-debug/resources/execution-protocol.md
index 258391e..a4dc6ee 100644
--- a/.agents/skills/oma-debug/resources/execution-protocol.md
+++ b/.agents/skills/oma-debug/resources/execution-protocol.md
@@ -33,11 +33,12 @@ Follow these steps in order (adjust depth by difficulty).
 - Check `resources/common-patterns.md` for known patterns
 
 ## Step 3: Fix & Test
-- Apply minimal fix that addresses the root cause
 - Write a regression test that:
   - Fails without the fix
   - Passes with the fix
   - Covers the specific edge case
+- **Run the regression test before applying the fix where feasible**: record the failing output (RED), apply the minimal fix, record the pass (GREEN) — include both in the bug report / result file (see `../../_shared/core/test-approach.md` §Debug parity)
+- Apply minimal fix that addresses the root cause
 - Check for similar patterns elsewhere: `search_for_pattern("same_bug_pattern")`
 - If found, fix proactively or report them
 
diff --git a/.agents/skills/oma-frontend/SKILL.md b/.agents/skills/oma-frontend/SKILL.md
index 22d5095..05d905e 100644
--- a/.agents/skills/oma-frontend/SKILL.md
+++ b/.agents/skills/oma-frontend/SKILL.md
@@ -129,7 +129,8 @@ Then run the project's frontend verification commands, typically lint, typecheck
 5. Run the execution checklist before handoff and include relevant verification results.
 6. **Self-describing file names**: every new file follows the File Naming convention in `../../rules/frontend.md` §Naming Conventions — domain + role readable from the basename alone (`order-summary-card.tsx`, `use-order-polling.ts`, `cart.atoms.ts`). Grab-bag names (`utils.ts`, `helpers.ts`, `misc.ts`) and version suffixes (`*-v2`, `*-final`) are banned.
 7. **Next.js 16 `proxy.ts` is mandatory; `middleware.ts` is BANNED**: this project is Next.js 16+. `middleware.ts` is NOT "deprecated"; it is forbidden, touch it and you die. The canonical request-proxy / auth-gate file is `proxy.ts` (root or `src/`) exporting a `proxy` function. NEVER create, recommend, suggest, or "restore" `middleware.ts`. NEVER flag `proxy.ts` as dead code, unused, or not-wired. Any such finding is a fatal self-error: retract it immediately and write `proxy.ts`.
-8. **Angular projects follow `resources/angular-rules.md`**: standalone components + `OnPush` + signals-first, `inject()` DI, lazy routes, new control flow. **Any non-trivial RxJS pipeline MUST ship with a marble test (`TestScheduler` from `rxjs/testing`)** — a stream without a marble test fails review. React/Next.js-specific rules (shadcn workflow, `proxy.ts`, Libraries table below) do not apply in Angular projects.
+8. **`next/link` defaults to `prefetch={false}`**: every `<Link>` MUST pass `prefetch={false}` unless there is a stated reason not to. Next.js's default prefetching fires a request per link entering the viewport, which hammers container CPU/memory and origin bandwidth on list-heavy or nav-heavy pages. Opt back in (`prefetch` omitted, or `prefetch` / `prefetch="unstable_forceStale"`) ONLY for a small, deliberate set of high-intent targets (primary CTA, next step in a funnel), and note the reason inline. A `<Link>` without an explicit prefetch decision fails review.
+9. **Angular projects follow `resources/angular-rules.md`**: standalone components + `OnPush` + signals-first, `inject()` DI, lazy routes, new control flow. **Any non-trivial RxJS pipeline MUST ship with a marble test (`TestScheduler` from `rxjs/testing`)** — a stream without a marble test fails review. React/Next.js-specific rules (shadcn workflow, `proxy.ts`, Libraries table below) do not apply in Angular projects.
 
 ### Libraries
 
diff --git a/.agents/skills/oma-frontend/resources/checklist.md b/.agents/skills/oma-frontend/resources/checklist.md
index abfb82c..dbe518e 100644
--- a/.agents/skills/oma-frontend/resources/checklist.md
+++ b/.agents/skills/oma-frontend/resources/checklist.md
@@ -8,6 +8,7 @@ Run through every item before submitting your work.
 - [ ] Do NOT flag `src/proxy.ts` as dead code or recommend renaming to `middleware.ts`; `proxy.ts` is the canonical Next.js 16+ convention
 <!-- oma-docs:ignore-end -->
 - [ ] Config flags use the `Proxy` form (e.g. `skipProxyUrlNormalize`), not the legacy `Middleware` form
+- [ ] Every `<Link>` passes `prefetch={false}`; any prefetching link is a deliberate high-intent target with an inline reason
 
 ## TypeScript
 - [ ] Strict mode, no `any` types
diff --git a/.agents/skills/oma-frontend/resources/execution-protocol.md b/.agents/skills/oma-frontend/resources/execution-protocol.md
index c118190..4f39a3e 100644
--- a/.agents/skills/oma-frontend/resources/execution-protocol.md
+++ b/.agents/skills/oma-frontend/resources/execution-protocol.md
@@ -29,6 +29,7 @@ Follow these steps in order (adjust depth by difficulty).
 - Plan responsive breakpoints and accessibility requirements
 
 ## Step 3: Implement
+- **Honor the task's `test_approach`** (see `../../_shared/core/test-approach.md`): for `tdd` tasks, write and run the focused test first (record the RED failure), make the minimal change (GREEN), then continue
 - Create/modify files in this order:
   1. TypeScript types/interfaces
   2. API client hooks (orval-generated from OpenAPI when available; hand-written TanStack Query otherwise)
@@ -42,6 +43,7 @@ Follow these steps in order (adjust depth by difficulty).
 - Run `resources/checklist.md` items
 - Run `../../_shared/core/common-checklist.md` items
 - Check TypeScript strict mode: no errors
+- For `tdd` tasks, append the `TDD_EVIDENCE` block (test command, RED, GREEN) to the result file per `../../_shared/core/test-approach.md`
 - Verify responsive design at 320px, 768px, 1024px, 1440px
 - Test keyboard navigation and screen reader compatibility
 
diff --git a/.agents/skills/oma-frontend/resources/snippets.md b/.agents/skills/oma-frontend/resources/snippets.md
index 4930d41..7e0402d 100644
--- a/.agents/skills/oma-frontend/resources/snippets.md
+++ b/.agents/skills/oma-frontend/resources/snippets.md
@@ -8,8 +8,13 @@ Copy-paste ready patterns. Use these as starting points, adapt to the specific t
 
 ```tsx
 // Internal nav: <Link>, never <a href="/...">
+// prefetch={false} is the DEFAULT — viewport prefetching eats container CPU/RAM.
 import Link from "next/link";
-<Link href="/gallery" className="...">View gallery</Link>
+<Link href="/gallery" prefetch={false} className="...">View gallery</Link>
+
+// Opt back in only for a few high-intent targets, and say why:
+// primary funnel CTA — prefetch is intentional
+<Link href="/checkout" className="...">Checkout</Link>
 
 // Custom font: next/font, never <link rel="stylesheet">
 import { Inter } from "next/font/google";
diff --git a/.agents/skills/oma-mobile/resources/execution-protocol.md b/.agents/skills/oma-mobile/resources/execution-protocol.md
index 0b18cfb..3834a21 100644
--- a/.agents/skills/oma-mobile/resources/execution-protocol.md
+++ b/.agents/skills/oma-mobile/resources/execution-protocol.md
@@ -33,6 +33,7 @@ Follow these steps in order (adjust depth by difficulty).
 - Note platform differences (iOS HIG vs Material Design 3)
 
 ## Step 3: Implement
+- **Honor the task's `test_approach`** (see `../../_shared/core/test-approach.md`): for `tdd` tasks, write and run the focused test first (record the RED failure), make the minimal change (GREEN), then continue
 - Create/modify files in this order (Flutter shown; Swift maps to Core → Features → Tests, RN to api → queries/mutations → store → ui → navigation → tests):
   1. Domain: entities and repository interfaces
   2. Data: models, API clients (Dio / axios / generated Client), repository implementations
@@ -45,6 +46,7 @@ Follow these steps in order (adjust depth by difficulty).
 ## Step 4: Verify
 - Run `resources/checklist.md` items
 - Run `../../_shared/core/common-checklist.md` items
+- For `tdd` tasks, append the `TDD_EVIDENCE` block (test command, RED, GREEN) to the result file per `../../_shared/core/test-approach.md`
 - Test on both iOS and Android (or emulators)
 - Verify 60fps performance (no jank)
 - Check dark mode support
diff --git a/.agents/skills/oma-pm/resources/examples.md b/.agents/skills/oma-pm/resources/examples.md
index 5380c09..f3671b8 100644
--- a/.agents/skills/oma-pm/resources/examples.md
+++ b/.agents/skills/oma-pm/resources/examples.md
@@ -24,6 +24,9 @@
       "dependencies": [],
       "estimated_complexity": "high",
       "scope": ["src/api/auth/"],
+      "test_approach": "tdd",
+      "test_scope": ["unit", "integration"],
+      "test_approach_rationale": "Deterministic authorization rules with clear inputs and outputs.",
       "acceptance_criteria": [
         "POST /api/auth/register with email + password",
         "POST /api/auth/login returns access + refresh tokens",
diff --git a/.agents/skills/oma-pm/resources/execution-protocol.md b/.agents/skills/oma-pm/resources/execution-protocol.md
index 29341a7..7ffa575 100644
--- a/.agents/skills/oma-pm/resources/execution-protocol.md
+++ b/.agents/skills/oma-pm/resources/execution-protocol.md
@@ -41,6 +41,13 @@ Follow these steps in order (adjust depth by difficulty).
 - Each task has: agent, title, description, acceptance criteria, priority, dependencies, **scope**
 - `agent`: one of the orchestrator-dispatchable domains — `backend`, `frontend`, `mobile`, `db`, `qa`, `debug`, `pm`, `architecture`, `refactor`, `tf-infra`, `docs` (see the agent mapping table in `.agents/workflows/orchestrate.md`)
 - `scope`: array of directory prefixes this agent is allowed to modify (e.g., `["src/api/", "migrations/"]`). Used by `verify` to detect cross-agent boundary violations in parallel execution.
+- **Test approach (opt-in, per task)**: set `test_approach` where a test strategy matters
+  - `tdd`: deterministic, high-risk behavior (validation, authorization, state transitions, calculations, error handling). Implementation agent must record RED→GREEN evidence (see `TDD_EVIDENCE` block below).
+  - `test_after`: automated tests required, but a useful isolated RED state is impractical
+  - `not_applicable`: automated tests inappropriate — **must** fill `test_approach_rationale` and `alternative_verification` (documented manual/alternative check)
+  - Do **not** mark `tdd` for: documentation, pure styling, generated code, IaC plans, behavior-preserving refactors (refactor tasks keep their characterization-test safety net), or inherently nondeterministic integrations
+  - `test_scope`: which layers the tests cover (e.g., `["unit", "integration"]`)
+  - No `test_approach` value ever waives the global unit-test coverage gate (>= 80%); `not_applicable` code still counts toward the aggregate unless declaratively excluded in coverage config with justification
 - Minimize dependencies for maximum parallel execution
 - Priority tiers: 1 = independent (run first), 2 = depends on tier 1, etc.
   - The numeric tier is the **canonical** `priority` value in plan JSON (what the orchestrator fans out on).
@@ -52,6 +59,7 @@ Follow these steps in order (adjust depth by difficulty).
 ## Step 4: Validate Plan
 - Check: Can each task be done independently given its dependencies?
 - Check: Are acceptance criteria measurable and testable?
+- Check: Is `test_approach` valid where set (`tdd|test_after|not_applicable`), with rationale + alternative verification for every `not_applicable`? (`oma verify pm` enforces this contract)
 - Check: Is security considered from the start (not deferred)?
 - Check: Are API contracts defined before frontend/mobile tasks?
 - Check: Are major risks, owners, and approval points explicit when needed?
diff --git a/.agents/skills/oma-pm/resources/task-template.json b/.agents/skills/oma-pm/resources/task-template.json
index f85f3d2..d20d4bb 100644
--- a/.agents/skills/oma-pm/resources/task-template.json
+++ b/.agents/skills/oma-pm/resources/task-template.json
@@ -28,7 +28,12 @@
       "estimated_complexity": "low|medium|high|very-high",
       "acceptance_criteria": [],
       "artifacts_expected": [],
-      "scope": []
+      "scope": [],
+      "test_approach": "tdd|test_after|not_applicable",
+      "test_scope": [],
+      "tdd_evidence_required": true,
+      "test_approach_rationale": "",
+      "alternative_verification": ""
     }
   ],
   "api_contracts": [
diff --git a/.agents/skills/oma-qa/resources/checklist.md b/.agents/skills/oma-qa/resources/checklist.md
index 475477b..fc28254 100644
--- a/.agents/skills/oma-qa/resources/checklist.md
+++ b/.agents/skills/oma-qa/resources/checklist.md
@@ -127,7 +127,8 @@
 ## Testing Checklist
 
 ### Unit Tests
-- [ ] Test coverage > 80%
+- [ ] Test coverage >= 80% (never waived by any task's `test_approach`; `not_applicable` code counts toward the aggregate unless declaratively excluded in coverage config with justification)
+- [ ] Tasks marked `test_approach: tdd` have a `TDD_EVIDENCE` block in the implementation result (focused test command, RED failure, GREEN pass) — see `../../_shared/core/test-approach.md`; do not require this evidence for `test_after` / `not_applicable` tasks
 - [ ] All business logic functions tested
 - [ ] Edge cases covered
 - [ ] Error handling tested
@@ -270,7 +271,7 @@
 - [ ] No data loss scenarios
 
 ### Important (Should Pass)
-- [ ] Test coverage > 80%
+- [ ] Test coverage >= 80%
 - [ ] Accessibility WCAG 2.2 AA
 - [ ] Code quality metrics met
 - [ ] Documentation complete
diff --git a/.agents/skills/oma-skill-creator/SKILL.md b/.agents/skills/oma-skill-creator/SKILL.md
index 010ad43..c12d556 100644
--- a/.agents/skills/oma-skill-creator/SKILL.md
+++ b/.agents/skills/oma-skill-creator/SKILL.md
@@ -159,6 +159,7 @@ Create, revise, and validate OMA skills using the SSL-lite Markdown structure de
 11. Do not create extra README, changelog, or installation docs inside a skill.
 12. Do not overwrite unrelated user edits.
 13. Enforce the three utility-predictive content dimensions — failure mechanism encoding, actionable specificity, high-risk action blacklist — per the Utility Content Checks in `resources/validation-checklist.md` (SkillLens, arXiv:2605.23899).
+14. Bulk-data scenes must run functions over data, not read data into context: route the processing through a deterministic tool/CLI stage (`CALL_TOOL`) and return only a summary plus the artifact path. Do not design scenes that stream raw harvested/parsed data through the model's context (see `../_shared/core/context-budget.md`, Core Principle 5).
 
 ## References
 - SSL-lite template: `resources/ssl-lite-template.md`
diff --git a/.agents/skills/oma-video/SKILL.md b/.agents/skills/oma-video/SKILL.md
index 7f016c4..7bd9b6f 100644
--- a/.agents/skills/oma-video/SKILL.md
+++ b/.agents/skills/oma-video/SKILL.md
@@ -203,7 +203,7 @@ Before invoking `oma video generate`, the calling agent runs this checklist. **I
 - [ ] **Locale**: narration + caption language (default from config; translated via oma-translator when non-source).
 - [ ] **Captions**: `tiktok` (centered, static windowed cues), `lower-third`, or `none`.
 - [ ] **Duration**: target seconds (<= 180) or `auto` (derived from the script).
-- [ ] **Voice / music**: voice profile or `none`; music `upbeat` / `calm` / `none`. **The default voice is `none` → a silent video with estimated caption timing.** Pass `--voice <profile>` (a Voicebox profile) whenever narration is expected. Music mixing is currently deferred (recorded + warned, not mixed).
+- [ ] **Voice / music**: voice profile or `none`; music `upbeat` / `calm` / `none`. **The default voice is `none` → a silent video with estimated caption timing.** Pass `--voice <profile>` (a Voicebox profile) whenever narration is expected. Music is rendered offline by Strudel and mixed at −18 dB; it needs a one-time `oma video doctor --install-strudel` and degrades to no music without it.
 
 **Amplification shortcut.** For a one-line brief (e.g. "shorts about Jeju coffee"), do not pop a questionnaire if the request is genuinely simple. Instead **amplify inline and show the user** the inferred plan before invoking:
 
@@ -247,7 +247,7 @@ oma video generate "<brief>" [--mode shorts|explainer|demo] \
                              [--aspect 9:16|16:9|1:1|auto] [--locale <lang>] \
                              [--captions tiktok|lower-third|none] \
                              [--visual auto|generate|stock|aigc|slide] \
-                             [--voice <profile>|none] [--music upbeat|calm|none] \
+                             [--voice <profile>|none] [--music upbeat|calm|cinematic|lofi|piano|none] \
                              [--duration <sec>|auto] [--compositor remotion|mpt] \
                              [--capture <path>] \
                              [--source file|web] [--url <url>] [--device <name>] \
@@ -298,7 +298,13 @@ Other skills call `oma video generate --format json` and parse the JSON envelope
 - **Narration is one wav**: oma-voice joins every scene line into a single `audio/narration-01.wav`, referenced by `render-spec.audio.narration`. There are no per-scene `narration-NN.wav` files.
 - **Timing**: per-line offsets live in `timing.json` (voicebox-stt -> estimated; the `tts-native` and `whisper-cpp` source values are reserved but not yet wired — `TODO(oma-deferred): whisper-cpp`); scene boundaries and caption cues are derived from it.
 - **Captions**: key-free `.srt` (+ `.vtt`) built from `timing.json`; `render-spec.captions.file` points at the `.srt`. The compositor renders **static windowed cues** — the cue active at the current frame, CSS-wrapped (no per-word animation).
-- **Music**: deferred (`TODO(oma-deferred): music`) — no music asset source is wired yet, so a requested `--music` mode is recorded in `script.json` and surfaced as a warning, but no music is mixed. When implemented, it mixes under narration at `render-spec.audio.musicGainDb` (default −18 dB).
+- **Music**: `--music <preset>` renders a BGM bed with **Strudel** and mixes it under narration at `render-spec.audio.musicGainDb` (default −18 dB). The bed is generated offline (headless Chrome + `OfflineAudioContext`), so it needs no key, no network, and no audio device — a 30s bed renders in well under a second.
+  - **Presets**: `calm` (sustained pad + arpeggio), `upbeat` (bright plucks), `cinematic` (drone build to a lead), `lofi` (warm chords, swung ticks), `piano` (neoclassical arpeggio). Each preset picks its key and mode from the run `seed`, so the same preset sounds different run to run without a second pattern.
+  - **Artifacts** in the run dir: `music/bgm.wav` (mixed by the compositor), `music/bgm.mp3` (preview), `music/bgm-raw.wav` (pre-master), and `music/pattern.strudel` — the source that produced them, editable and re-renderable by hand.
+  - **Level**: every bed is normalised to −14 LUFS with a static gain before a peak limiter, so `musicGainDb` means the same thing for every preset. Normalisation is deliberately *not* `loudnorm`'s one-pass mode, which flattens the arrangement arc.
+  - **Opt-in install**: `@strudel/*` is AGPL-3.0-or-later while the oma CLI is MIT, so the deps are never bundled and never installed implicitly. Run `oma video doctor --install-strudel` once. The CLI never imports Strudel — it spawns `resources/strudel/render.mjs` as a subprocess, the same boundary the Remotion / Playwright projects use.
+  - **Fallback**: a missing install, a missing Chrome, or a failed render degrades to *no music* with a warning. The run still succeeds and `audio.music` stays unset (never a dangling `staticFile()` ref).
+  - **Determinism**: the built-in beds are oscillator-only (sine / triangle / square / sawtooth), which render byte-identically on replay. Noise sounds (`white` / `pink` / `brown`) draw from `Math.random()` and would break that, so the templates avoid them.
 
 ## References
 
diff --git a/.agents/skills/oma-video/resources/execution-protocol.md b/.agents/skills/oma-video/resources/execution-protocol.md
index bafb9c1..e739574 100644
--- a/.agents/skills/oma-video/resources/execution-protocol.md
+++ b/.agents/skills/oma-video/resources/execution-protocol.md
@@ -15,7 +15,7 @@ plan when the brief is a one-liner.
    - `mode` ∈ {`shorts`, `explainer`, `demo`}.
    - `aspect` ∈ {`9:16`, `16:9`, `1:1`, `auto`} (`auto` snaps to the mode default: shorts -> 9:16, explainer/demo -> 16:9).
    - `captions` ∈ {`tiktok`, `lower-third`, `none`}; `visual` ∈ {`auto`, `generate`, `stock`, `aigc`, `slide`}.
-   - `music` ∈ {`upbeat`, `calm`, `none`}; `compositor` ∈ {`remotion`, `mpt`}.
+   - `music` ∈ {`upbeat`, `calm`, `cinematic`, `lofi`, `piano`, `none`}; `compositor` ∈ {`remotion`, `mpt`}.
    - `duration` ≤ `limits.max_duration_sec` (180); resulting `scenes` ≤ `limits.max_scenes` (40).
    - `out` is inside `$PWD` unless `--allow-external-out`.
    - For `demo`: `--capture` (if given) exists, is absolute + `$PWD`-guarded, and is a valid video format.
diff --git a/.agents/skills/oma-video/resources/prompt-tips.md b/.agents/skills/oma-video/resources/prompt-tips.md
index 69b53b1..6bc6c33 100644
--- a/.agents/skills/oma-video/resources/prompt-tips.md
+++ b/.agents/skills/oma-video/resources/prompt-tips.md
@@ -53,7 +53,7 @@ Scene/backdrop → Subject → Details → Constraints
 - Anchor the **arc**: hook → body → payoff. A short without a hook gets swiped past.
 - Match **aspect to mode** (9:16 shorts, 16:9 explainer/demo) or use `auto`.
 - Keep narration **per-scene and short** so caption pages and scene boundaries align.
-- Pick **music** that matches pacing (`upbeat` for shorts, `calm` for explainer) — note music mixing is currently deferred: the choice is recorded in `script.json` and warned, but not yet audible.
+- Pick **music** that matches pacing (`upbeat` for shorts, `calm` for explainer) — Strudel renders the bed offline and mixes it at −18 dB under narration. Needs a one-time `oma video doctor --install-strudel`; without it the run still succeeds, just silent.
 
 ## Don'ts
 
diff --git a/.agents/skills/oma-video/resources/script-schema.md b/.agents/skills/oma-video/resources/script-schema.md
index 86c515d..b927f03 100644
--- a/.agents/skills/oma-video/resources/script-schema.md
+++ b/.agents/skills/oma-video/resources/script-schema.md
@@ -15,7 +15,7 @@ required** — omitting it is the most common authoring failure.
 | `locale` | string (min 1) | ✅ | narration/caption language tag, e.g. `en`, `ko` |
 | `title` | string (min 1) | ✅ | also drives the output filename slug: `<mode>-<slug>.mp4` |
 | `scenes` | array (min 1, ≤ 40) | ✅ | see per-scene fields below |
-| `music` | `upbeat` \| `calm` \| `none` | ✅ | recorded only — music mixing is deferred (`TODO(oma-deferred): music`) |
+| `music` | `upbeat` \| `calm` \| `cinematic` \| `lofi` \| `piano` \| `none` | ✅ | drives the Strudel BGM bed; mixed at −18 dB under narration |
 | `brand` | object | — | free-form; defaults to `{}` |
 
 ### Per-scene fields (`scenes[]`)
@@ -77,5 +77,5 @@ required** — omitting it is the most common authoring failure.
 - [ ] `aspect` is concrete (`9:16` / `16:9` / `1:1`) — never `auto`.
 - [ ] Scene count ≤ 40, total `durationSec` ≤ 180.
 - [ ] Every scene has a `visual.kind` from the enum; `still` scenes carry an English `visual.prompt`.
-- [ ] `music` is set (use `"none"` unless the user asked — mixing is deferred either way).
+- [ ] `music` is set (use `"none"` unless the user asked for a bed).
 - [ ] Validate cheaply before rendering: pass the file via `--script` with `--dry-run` first.
diff --git a/.agents/skills/oma-video/resources/strudel/package.json b/.agents/skills/oma-video/resources/strudel/package.json
new file mode 100644
index 0000000..000fce4
--- /dev/null
+++ b/.agents/skills/oma-video/resources/strudel/package.json
@@ -0,0 +1,11 @@
+{
+  "name": "oma-video-strudel",
+  "version": "0.0.0",
+  "private": true,
+  "type": "module",
+  "description": "Vendored Strudel BGM renderer for oma-video. Deps are installed on demand by `oma video doctor --install-strudel`. @strudel/* is AGPL-3.0-or-later and is NEVER imported by the MIT-licensed oma CLI — render.mjs runs as a subprocess (same boundary as the Remotion / Playwright projects).",
+  "license": "MIT",
+  "dependencies": {
+    "@strudel/web": "^1.3.0"
+  }
+}
diff --git a/.agents/skills/oma-video/resources/strudel/page.html b/.agents/skills/oma-video/resources/strudel/page.html
new file mode 100644
index 0000000..afe58bf
--- /dev/null
+++ b/.agents/skills/oma-video/resources/strudel/page.html
@@ -0,0 +1,87 @@
+<!doctype html>
+<html>
+  <head>
+    <meta charset="utf-8" />
+    <title>oma-video strudel offline render</title>
+  </head>
+  <body>
+    <!-- Served by render.mjs from the resolved @strudel/web dist. The IIFE
+         bundle exposes everything on `window.strudel` + `window.initStrudel`. -->
+    <script src="/strudel.js"></script>
+    <script>
+      window.__omaLog = [];
+      window.addEventListener('error', (e) =>
+        window.__omaLog.push('error: ' + e.message),
+      );
+      window.addEventListener('unhandledrejection', (e) =>
+        window.__omaLog.push('rejection: ' + (e.reason?.message ?? e.reason)),
+      );
+
+      /**
+       * Render `code` to a WAV and stash it on window as base64.
+       *
+       * `strudel.renderPatternAudio()` is a real OfflineAudioContext render
+       * (faster than realtime, no audio device involved) but it delivers the
+       * finished wav as a browser DOWNLOAD. Rather than plumb a download
+       * directory through CDP, we intercept URL.createObjectURL to capture the
+       * Blob in-page and hand the bytes back over the debugger protocol.
+       *
+       * Returns the base64 length; the driver pulls window.__omaWav in chunks.
+       */
+      window.__omaRender = async function (code, seconds, cps, sampleRate) {
+        await window.initStrudel();
+
+        // repl.evaluate(code, autostart=false) transpiles + evaluates without
+        // starting the realtime scheduler, and returns the Pattern itself.
+        const evaluated = await window.strudel.evaluate(code, false);
+        const pattern = evaluated?.pattern ?? evaluated;
+        if (!pattern || typeof pattern.queryArc !== 'function') {
+          throw new Error(
+            'strudel: pattern code did not evaluate to a Pattern (got ' +
+              typeof pattern +
+              ')',
+          );
+        }
+
+        let captured = null;
+        const realCreate = URL.createObjectURL.bind(URL);
+        const realRevoke = URL.revokeObjectURL.bind(URL);
+        URL.createObjectURL = (blob) => {
+          captured = blob;
+          return 'blob:oma-captured';
+        };
+        URL.revokeObjectURL = () => {};
+        try {
+          await window.strudel.renderPatternAudio(
+            pattern,
+            cps,
+            0,
+            seconds * cps,
+            sampleRate,
+            128,
+            false,
+            'oma-bgm',
+          );
+        } finally {
+          URL.createObjectURL = realCreate;
+          URL.revokeObjectURL = realRevoke;
+        }
+        if (!captured) {
+          throw new Error('strudel: renderPatternAudio produced no audio blob');
+        }
+
+        const bytes = new Uint8Array(await captured.arrayBuffer());
+        let binary = '';
+        const CHUNK = 0x8000; // stay under the String.fromCharCode arg limit
+        for (let i = 0; i < bytes.length; i += CHUNK) {
+          binary += String.fromCharCode.apply(
+            null,
+            bytes.subarray(i, i + CHUNK),
+          );
+        }
+        window.__omaWav = btoa(binary);
+        return window.__omaWav.length;
+      };
+    </script>
+  </body>
+</html>
diff --git a/.agents/skills/oma-video/resources/strudel/render.mjs b/.agents/skills/oma-video/resources/strudel/render.mjs
new file mode 100644
index 0000000..e9a67c6
--- /dev/null
+++ b/.agents/skills/oma-video/resources/strudel/render.mjs
@@ -0,0 +1,295 @@
+#!/usr/bin/env node
+// Strudel BGM renderer — boundary-safe subprocess entrypoint.
+//
+// oma-video NEVER imports Strudel. @strudel/* is AGPL-3.0-or-later while the
+// oma CLI is MIT, so the TypeScript provider (`providers/music-strudel.ts` +
+// `internal/strudel-project.ts`) only ever LOCATES this project on disk and
+// spawns *this* script. Same boundary the Remotion / Playwright projects use.
+//
+// MECHANISM: `@strudel/webaudio.renderPatternAudio()` renders a pattern through
+// an OfflineAudioContext — faster than realtime, no audio device, no autoplay
+// gesture. A 30s bed renders in well under a second. The bundle only runs in a
+// browser, so we drive a headless Chrome over raw CDP (Node's global WebSocket;
+// zero npm deps here beyond @strudel/web itself) and serve the bundle from a
+// loopback-only HTTP server.
+//
+// DETERMINISM: oscillator sounds (sine/triangle/square/sawtooth) render
+// byte-identically across runs. Noise-based sounds (white/pink/brown) do NOT —
+// superdough fills those buffers from Math.random(). Patterns that must be
+// reproducible should stay oscillator-only.
+//
+// Contract — flags in, ONE JSON result line out (stdout's LAST line):
+//   IN  --pattern-file <path>  file holding the strudel pattern code (required)
+//       --out <abs wav path>   output wav, inside a run dir (required)
+//       --chrome <path>        Chrome/Chromium executable (required)
+//       --seconds <float>      bed length in seconds (required)
+//       --cps <float>          cycles per second (default 0.5)
+//       --sample-rate <int>    default 44100
+//       --timeout <ms>         hard ceiling for the whole render (default 120000)
+//   OUT {"ok":true,"output":"<abs wav>","seconds":<f>,"bytes":<n>,"sha256":"<hex>"}
+//    |  {"ok":false,"error":"<message>","code":"<reason>"}
+//
+// Exit code is 0 on success, 1 on failure; the JSON result line is authoritative.
+import { spawn } from "node:child_process";
+import { createHash } from "node:crypto";
+import { existsSync, readFileSync } from "node:fs";
+import { mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
+import { createServer } from "node:http";
+import { tmpdir } from "node:os";
+import path from "node:path";
+import process from "node:process";
+import { fileURLToPath } from "node:url";
+
+const HERE = path.dirname(fileURLToPath(import.meta.url));
+const DIST = path.join(HERE, "node_modules", "@strudel", "web", "dist");
+
+const MIME = {
+  ".js": "text/javascript",
+  ".mjs": "text/javascript",
+  ".html": "text/html",
+  ".json": "application/json",
+  ".wasm": "application/wasm",
+};
+
+let server;
+let chrome;
+let userDataDir;
+let hardTimer;
+
+function fail(message, code = "render_failed") {
+  process.stdout.write(`${JSON.stringify({ ok: false, error: message, code })}\n`);
+  cleanup().finally(() => process.exit(1));
+}
+
+async function cleanup() {
+  if (hardTimer) clearTimeout(hardTimer);
+  try {
+    chrome?.kill("SIGTERM");
+  } catch {}
+  try {
+    server?.close();
+  } catch {}
+  if (userDataDir) {
+    await rm(userDataDir, { recursive: true, force: true }).catch(() => {});
+  }
+}
+
+function parseArgs(argv) {
+  const out = {};
+  for (let i = 0; i < argv.length; i++) {
+    const token = argv[i];
+    if (!token.startsWith("--")) continue;
+    const eq = token.indexOf("=");
+    if (eq !== -1) {
+      out[token.slice(2, eq)] = token.slice(eq + 1);
+    } else {
+      out[token.slice(2)] = argv[i + 1]?.startsWith("--") ? "" : argv[++i];
+    }
+  }
+  return out;
+}
+
+const args = parseArgs(process.argv.slice(2));
+const patternFile = args["pattern-file"];
+const outPath = args.out;
+const chromePath = args.chrome;
+const seconds = Number(args.seconds);
+const cps = Number(args.cps ?? 0.5);
+const sampleRate = Number(args["sample-rate"] ?? 44100);
+const timeoutMs = Number(args.timeout ?? 120000);
+
+if (!patternFile || !outPath || !chromePath) {
+  fail("missing required flag: --pattern-file / --out / --chrome", "bad_args");
+} else if (!Number.isFinite(seconds) || seconds <= 0) {
+  fail(`--seconds must be a positive number (got ${args.seconds})`, "bad_args");
+} else if (!existsSync(DIST)) {
+  fail(
+    `@strudel/web is not installed in ${HERE} — run \`oma video doctor --install-strudel\``,
+    "not_installed",
+  );
+} else if (!existsSync(chromePath)) {
+  fail(`chrome executable not found: ${chromePath}`, "no_chrome");
+} else {
+  main().catch((err) => fail(err?.message ?? String(err)));
+}
+
+async function main() {
+  hardTimer = setTimeout(
+    () => fail(`render exceeded ${timeoutMs}ms`, "timeout"),
+    timeoutMs,
+  );
+  hardTimer.unref?.();
+
+  const code = await readFile(patternFile, "utf8");
+
+  // 1. Loopback-only static server for the page + the strudel bundle. Paths are
+  //    confined to this dir and the resolved dist (no traversal).
+  server = createServer(async (req, res) => {
+    const pathname = new URL(req.url, "http://127.0.0.1").pathname;
+    let file;
+    if (pathname === "/" || pathname === "/page.html") {
+      file = path.join(HERE, "page.html");
+    } else if (pathname === "/strudel.js") {
+      file = path.join(DIST, "index.js");
+    } else {
+      const resolved = path.resolve(DIST, `.${pathname}`);
+      file = resolved.startsWith(DIST) ? resolved : null;
+    }
+    if (!file) {
+      res.writeHead(403).end("forbidden");
+      return;
+    }
+    try {
+      const body = await readFile(file);
+      res.writeHead(200, {
+        "content-type": MIME[path.extname(file)] ?? "application/octet-stream",
+        "cross-origin-opener-policy": "same-origin",
+        "cross-origin-embedder-policy": "require-corp",
+        "cross-origin-resource-policy": "cross-origin",
+      });
+      res.end(body);
+    } catch {
+      res.writeHead(404).end("not found");
+    }
+  });
+  await new Promise((resolve) => server.listen(0, "127.0.0.1", resolve));
+  const httpPort = server.address().port;
+
+  // 2. Headless Chrome. No audio device is needed — OfflineAudioContext renders
+  //    into a buffer, so this works on a bare CI box.
+  userDataDir = await mkdtemp(path.join(tmpdir(), "oma-strudel-"));
+  chrome = spawn(
+    chromePath,
+    [
+      "--headless=new",
+      "--disable-gpu",
+      "--no-first-run",
+      "--no-default-browser-check",
+      "--disable-dev-shm-usage",
+      `--user-data-dir=${userDataDir}`,
+      "--remote-debugging-port=0",
+      "about:blank",
+    ],
+    { stdio: ["ignore", "ignore", "pipe"] },
+  );
+  chrome.on("error", (err) => fail(`chrome failed to start: ${err.message}`, "no_chrome"));
+
+  const devtoolsPort = await waitForDevtoolsPort(
+    path.join(userDataDir, "DevToolsActivePort"),
+  );
+
+  // 3. Drive it over raw CDP.
+  const targets = await (
+    await fetch(`http://127.0.0.1:${devtoolsPort}/json/list`)
+  ).json();
+  const page = targets.find((t) => t.type === "page");
+  if (!page) throw new Error("chrome exposed no page target");
+
+  const cdp = await connect(page.webSocketDebuggerUrl);
+  await cdp.send("Runtime.enable");
+  await cdp.send("Page.enable");
+  await cdp.send("Page.navigate", {
+    url: `http://127.0.0.1:${httpPort}/page.html`,
+  });
+  await cdp.waitForLoad();
+
+  const base64Length = await cdp.evaluate(
+    `window.__omaRender(${JSON.stringify(code)}, ${seconds}, ${cps}, ${sampleRate})`,
+  );
+  if (!base64Length || base64Length <= 0) {
+    throw new Error("strudel produced an empty buffer");
+  }
+
+  // 4. Pull the wav back in slices so no single CDP message gets huge.
+  const CHUNK = 4 * 1024 * 1024;
+  let b64 = "";
+  for (let offset = 0; offset < base64Length; offset += CHUNK) {
+    b64 += await cdp.evaluate(
+      `window.__omaWav.slice(${offset}, ${offset + CHUNK})`,
+    );
+  }
+  const wav = Buffer.from(b64, "base64");
+  await writeFile(outPath, wav);
+
+  process.stdout.write(
+    `${JSON.stringify({
+      ok: true,
+      output: outPath,
+      seconds,
+      bytes: wav.length,
+      sha256: createHash("sha256").update(wav).digest("hex"),
+    })}\n`,
+  );
+  await cleanup();
+  process.exit(0);
+}
+
+/** Chrome writes its chosen debugger port here once it is listening. */
+async function waitForDevtoolsPort(portFile) {
+  for (let i = 0; i < 200; i++) {
+    if (existsSync(portFile)) {
+      const [line] = readFileSync(portFile, "utf8").split("\n");
+      const port = Number(line);
+      if (Number.isFinite(port) && port > 0) return port;
+    }
+    await new Promise((resolve) => setTimeout(resolve, 50));
+  }
+  throw new Error("chrome never reported a devtools port");
+}
+
+/** Minimal CDP client over Node's global WebSocket. */
+async function connect(wsUrl) {
+  const ws = new WebSocket(wsUrl);
+  await new Promise((resolve, reject) => {
+    ws.onopen = resolve;
+    ws.onerror = () => reject(new Error("cdp websocket failed to open"));
+  });
+
+  let nextId = 0;
+  const pending = new Map();
+  let loaded = false;
+  const loadWaiters = [];
+
+  ws.onmessage = (event) => {
+    const msg = JSON.parse(event.data);
+    if (msg.id && pending.has(msg.id)) {
+      const { resolve, reject } = pending.get(msg.id);
+      pending.delete(msg.id);
+      if (msg.error) reject(new Error(JSON.stringify(msg.error)));
+      else resolve(msg.result);
+      return;
+    }
+    if (msg.method === "Page.loadEventFired") {
+      loaded = true;
+      while (loadWaiters.length) loadWaiters.shift()();
+    }
+  };
+
+  const send = (method, params = {}) =>
+    new Promise((resolve, reject) => {
+      const id = ++nextId;
+      pending.set(id, { resolve, reject });
+      ws.send(JSON.stringify({ id, method, params }));
+    });
+
+  return {
+    send,
+    waitForLoad: () =>
+      loaded ? Promise.resolve() : new Promise((resolve) => loadWaiters.push(resolve)),
+    async evaluate(expression) {
+      const res = await send("Runtime.evaluate", {
+        expression,
+        awaitPromise: true,
+        returnByValue: true,
+      });
+      if (res.exceptionDetails) {
+        const detail =
+          res.exceptionDetails.exception?.description ??
+          res.exceptionDetails.text ??
+          "unknown page error";
+        throw new Error(detail.split("\n")[0]);
+      }
+      return res.result.value;
+    },
+  };
+}
diff --git a/.agents/workflows/orchestrate.md b/.agents/workflows/orchestrate.md
index 313ca4a..9755045 100644
--- a/.agents/workflows/orchestrate.md
+++ b/.agents/workflows/orchestrate.md
@@ -37,12 +37,26 @@ The detected runtime vendor and each agent's target vendor determine how agents
 
 ## Step 1: Load or Create Plan
 
+### 1a. Load
+
 Look for a plan file:
 
 1. Check `.agents/results/plan-{sessionId}.json` (current session's plan).
 2. If not found: find the most recent `.agents/results/plan-*.json` file.
-3. If none exist: ask the user to run `/plan` first, or ask them to describe the tasks to execute.
-- **Do NOT proceed without a plan.**
+3. A plan is **usable** only when every task carries an agent assignment, a priority tier, its dependencies, and acceptance criteria. A plan missing any of these is not execution-ready — fall through to 1b rather than fanning out against it.
+
+### 1b. Create (no usable plan)
+
+A missing plan is not a stop condition. `/orchestrate` creates the plan itself instead of handing the request back to the user:
+
+1. Generate the session ID now (format: `session-YYYYMMDD-HHMMSS`). Step 2 reuses this id verbatim — do not generate a second one.
+2. Read and follow `.agents/workflows/plan.md` step by step, passing this session ID as its `{sessionId}` so the artifact lands at `.agents/results/plan-{sessionId}.json`.
+3. **Do NOT skip `plan.md` Step 6 (Review Plan with User).** It is this run's approval gate — the Step 3 fan-out is authorized by it. Delegation never removes a user gate.
+4. Once the plan is saved and approved, load it and continue to Step 2 with the same session ID.
+
+Stop and report only when the plan cannot be produced: the user declines to plan, or `plan.md` blocks because the request is too underspecified to decompose.
+
+- **Do NOT spawn agents without a usable plan.**
 
 ---
 
@@ -66,7 +80,7 @@ Look for a plan file:
    └──────────┴───────────────────┘
    ```
 
-3. Generate session ID (format: `session-YYYYMMDD-HHMMSS`).
+3. Session ID: reuse the id generated in Step 1b when the plan was created in this run; otherwise generate one now (format: `session-YYYYMMDD-HHMMSS`).
 4. **Domain gate**: for each planned task, classify it into `domain_tags` by matching against the `Intent signature` block of each installed `.agents/skills/oma-*/SKILL.md`, and derive `exposed_skill_set` (skills whose name is in `domain_tags`). If fewer than 2 skills match confidently, fall back to the full installed set and mark `exposure_fallback: true`. See `.agents/skills/oma-orchestrator/SKILL.md` (PHASE 1.5) for the full rules.
 5. Use memory write tool to create `orchestrator-session.md` and `task-board.md` in the memory base path. Record `Exposed Skills` and `Exposure Fallback` per task in `task-board.md`.
 6. Set session status to RUNNING.
@@ -148,6 +162,7 @@ Also use memory read tool to poll `progress-{agent}[-{sessionId}].md` for logic
 
 - Use memory edit tool to update `task-board.md` with turn counts and status changes.
 - Watch for: completion, failures, crashes.
+- A `no-artifact` status (or `oma agent:spawn` exit code 3) means the vendor exited 0 but wrote no result artifact under the workspace — a silent misdirected write. Treat it as a failed spawn: do NOT collect it as completed; re-dispatch (natively if the external vendor is unreliable) and check the session trail for the `blocker.raised` event.
 
 ### Context Anxiety Check (per polling cycle)
 
diff --git a/.agents/workflows/ultrawork.md b/.agents/workflows/ultrawork.md
index 35a35a2..64f89d4 100644
--- a/.agents/workflows/ultrawork.md
+++ b/.agents/workflows/ultrawork.md
@@ -59,6 +59,9 @@ Reviewers are read-only evaluators. Implementation and refactor **actions** (Pha
 10. Record session start using memory write tool:
    - Create `session-ultrawork.md` in the memory base path
    - Include: session start time, session ID, user request summary, workflow version (ultrawork)
+11. (Recommended) Attach a mechanical stop gate when the project has a cheap deterministic check:
+   - `oma goal:set --gate typecheck` (allowlist: `typecheck` | `test` | `lint`; maps to the package.json script)
+   - While set, the Stop hook allows the session to end only when the gate passes; failures return the output tail. Add `--budget-minutes <n>` to bound unattended runs with an honest partial stop.
 
 ---
 
diff --git a/.agents/workflows/ultrawork/resources/phase-gates.md b/.agents/workflows/ultrawork/resources/phase-gates.md
index feb9467..816ea4c 100644
--- a/.agents/workflows/ultrawork/resources/phase-gates.md
+++ b/.agents/workflows/ultrawork/resources/phase-gates.md
@@ -13,6 +13,7 @@ The "Owner" of each gate coordinates the phase and records the verdict; it does
 
 ### Criteria
 - [ ] Plan documented with acceptance criteria
+- [ ] Where set, `test_approach` is valid (`tdd|test_after|not_applicable`); every `not_applicable` carries `test_approach_rationale` + `alternative_verification`; refactor tasks are never `tdd` (see `_shared/core/test-approach.md`)
 - [ ] Assumptions explicitly listed
 - [ ] Alternatives considered for architecture decisions (min 2)
 - [ ] Over-engineering review completed
@@ -36,6 +37,7 @@ Revise plan, do not proceed to IMPL
 ### Criteria
 - [ ] Code compiles/builds successfully
 - [ ] Tests pass
+- [ ] Tasks marked `test_approach: tdd` have a `TDD_EVIDENCE` block (focused test command, RED failure, GREEN pass) in the result — checked **only** for `tdd` tasks; `oma verify <agent>` automates this
 - [ ] Only planned files modified
 - [ ] No unrequested features added
 - [ ] Diff reviewed for scope creep
@@ -98,7 +100,7 @@ Address issues, re-verify
 ### Criteria
 - [ ] Lint passes
 - [ ] Type check passes
-- [ ] Test coverage >= 80%
+- [ ] Test coverage >= 80% (hard floor — no task's `test_approach`, including `not_applicable`, waives or lowers it)
 - [ ] UX flows verified
 - [ ] No hardcoded secrets
 - [ ] Migrations safe
diff --git a/.agents/workflows/video.md b/.agents/workflows/video.md
index b1754a9..fbc78e7 100644
--- a/.agents/workflows/video.md
+++ b/.agents/workflows/video.md
@@ -57,7 +57,7 @@ For `demo`, also resolve the **source**: a recorded file or Cap → `--source fi
   | stock video | Pexels (`PEXELS_API_KEY`) | oma-image stills + Ken Burns | `TODO(oma-deferred): pexels` |
   | AIGC video | Pixelle-MCP + RunningHub (`RUNNINGHUB_API_KEY`) | oma-image stills | `TODO(oma-deferred): pixelle` |
   | caption timing | voicebox-stt (MCP `voicebox_transcribe` → REST) | estimate | `TODO(oma-deferred): whisper-cpp` |
-  | music mixing | (not wired — recorded + warned only) | render without music | `TODO(oma-deferred): music` |
+  | music mixing | Strudel offline render (`oma video doctor --install-strudel`) | render without music | — |
   | premium TTS | (not needed — oma-voice is local) | — | — |
 
 - **Pixelle AIGC is a community MCP**: off by default, requires one-time explicit user consent plus a source review before connecting, and is always cost-gated on RunningHub credits.
@@ -113,7 +113,7 @@ The agent writes the script — this is the start of the determinism boundary. D
    ```bash
    oma video generate "<brief>" --mode <mode> --aspect <aspect> --locale <lang> \
      --captions <tiktok|lower-third|none> --visual <auto|generate|stock|aigc|slide> \
-     --voice <profile|none> --music <upbeat|calm|none> --duration <sec|auto> \
+     --voice <profile|none> --music <upbeat|calm|cinematic|lofi|piano|none> --duration <sec|auto> \
      --compositor <remotion|mpt> --seed <n> \
      --script <path-to-agent-authored-script.json> --dry-run --format json
    ```
@@ -193,7 +193,7 @@ Review the finished video against the brief and the quality bars. Iterate by re-
    - Narration audio is present (or intentionally silent) and aligns to scenes.
    - Captions are synced to `timing.json`, within the safe area, and legible (static windowed cues, CSS-wrapped, Pretendard, design rule 2).
    - Visuals match each scene's intent; no placeholder leakage unless the run intentionally used the fallback.
-   - Aspect / dimensions are correct for the mode; branding applied as requested. (Music mixing is deferred — a requested music mode only produces a warning, never audio.)
+   - Aspect / dimensions are correct for the mode; branding applied as requested. (A requested music mode yields `music/bgm.wav` mixed at −18 dB, or a fallback warning and a silent render when Strudel is not installed.)
 2. **Route each defect to its stage:**
    - script/narration/scene-count → **Step 3** (re-author script).
    - audio/timing → **Step 4** voice track (check oma-voice, re-synthesize).
diff --git a/.claude/agents/backend-engineer.md b/.claude/agents/backend-engineer.md
index 7a4221d..cad333b 100644
--- a/.claude/agents/backend-engineer.md
+++ b/.claude/agents/backend-engineer.md
@@ -46,7 +46,7 @@ Router (HTTP) → Service (Business Logic) → Repository (Data Access) → Mode
 ## Rules
 
 1. Stay in scope — only work on assigned backend tasks
-2. Write tests for all new code
+2. Write tests for all new code; honor the plan task's `test_approach` — for `tdd`, demonstrate RED before the change and record a `TDD_EVIDENCE` block (test command, RED, GREEN) in the result file
 3. Follow Repository → Service → Router pattern (no business logic in routes)
 4. Validate all inputs with the project's validation library
 5. Parameterized queries only (no string interpolation in SQL)
diff --git a/.claude/agents/debug-investigator.md b/.claude/agents/debug-investigator.md
index 0e0bc5d..a18e1a4 100644
--- a/.claude/agents/debug-investigator.md
+++ b/.claude/agents/debug-investigator.md
@@ -53,7 +53,7 @@ CHARTER_CHECK:
 1. Stay in scope — only work on assigned debug tasks
 2. Fix root cause, not symptoms
 3. Minimal changes only — no refactoring during bugfix; route refactoring needs to refactor-engineer
-4. Every fix gets a regression test
+4. Every fix gets a regression test; run it before the fix where feasible and record RED (failing output) → GREEN (post-fix pass) in the bug report
 5. Search for similar patterns after fixing
 6. Document out-of-scope findings for other agents
 7. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.claude/agents/frontend-engineer.md b/.claude/agents/frontend-engineer.md
index cc83c56..9a29107 100644
--- a/.claude/agents/frontend-engineer.md
+++ b/.claude/agents/frontend-engineer.md
@@ -54,6 +54,6 @@ FSD-lite: root `src/` + feature `src/features/*/`
 5. TailwindCSS v4 for styling, design tokens 1:1 mapping
 6. Library defaults (greenfield; existing project choices win): luxon (dates), ahooks (hooks), es-toolkit (utils), jotai (client state), TanStack Query (server state)
 7. Absolute imports with `@/`
-8. Write tests for custom logic (>90% coverage target)
+8. Write tests for custom logic (>90% coverage target); honor the plan task's `test_approach` — for `tdd`, demonstrate RED before the change and record a `TDD_EVIDENCE` block (test command, RED, GREEN) in the result file
 9. Document out-of-scope dependencies for other agents
 10. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.claude/agents/mobile-engineer.md b/.claude/agents/mobile-engineer.md
index ac734b1..dc85dba 100644
--- a/.claude/agents/mobile-engineer.md
+++ b/.claude/agents/mobile-engineer.md
@@ -53,7 +53,7 @@ Clean Architecture: domain → data → presentation (Swift native: App/Core/Fea
 5. Transport client with interceptors (Dio / axios / generated Client) + repository-layer response cache, offline-first architecture
 6. Secrets in secure storage only — never plain prefs or MMKV
 7. 60fps target performance
-8. Write widget/component tests and integration tests
+8. Write widget/component tests and integration tests; honor the plan task's `test_approach` — for `tdd`, demonstrate RED before the change and record a `TDD_EVIDENCE` block (test command, RED, GREEN) in the result file
 9. ARB-based localization: edit ARB source files only, never generated localization code
 10. Document out-of-scope dependencies for other agents
 11. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.claude/agents/pm-planner.md b/.claude/agents/pm-planner.md
index 5dd3096..ba66e5f 100644
--- a/.claude/agents/pm-planner.md
+++ b/.claude/agents/pm-planner.md
@@ -56,12 +56,13 @@ Each task must include:
 - `priority`: execution tier — 1 = independent (runs first), 2 = depends on tier 1, etc. (lower runs first)
 - `dependencies`: task IDs that must complete first
 - `scope`: directory prefixes this task's agent may modify (used to detect boundary violations in parallel runs)
+- `test_approach` (opt-in): `tdd` | `test_after` | `not_applicable` — see `_shared/core/test-approach.md`. `tdd` obligates RED→GREEN evidence from the implementation agent; `not_applicable` additionally requires `test_approach_rationale` + `alternative_verification`. Never assign `tdd` to refactor tasks (characterization tests instead)
 
 ## Rules
 
 1. Stay in scope — planning only, no code implementation
 2. API-first design
 3. Minimize dependencies for maximum parallelism
-4. Security and testing are part of every task (not separate)
+4. Security and testing are part of every task (not separate); assign per-task `test_approach` (`tdd|test_after|not_applicable`) where a test strategy matters — `not_applicable` requires rationale + alternative verification, and no approach waives the >= 80% coverage gate
 5. Each task completable by a single agent
 6. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.codex/agents/backend-engineer.toml b/.codex/agents/backend-engineer.toml
index d0f944b..dcd3f4d 100644
--- a/.codex/agents/backend-engineer.toml
+++ b/.codex/agents/backend-engineer.toml
@@ -40,7 +40,7 @@ Router (HTTP) → Service (Business Logic) → Repository (Data Access) → Mode
 ## Rules
 
 1. Stay in scope — only work on assigned backend tasks
-2. Write tests for all new code
+2. Write tests for all new code; honor the plan task's `test_approach` — for `tdd`, demonstrate RED before the change and record a `TDD_EVIDENCE` block (test command, RED, GREEN) in the result file
 3. Follow Repository → Service → Router pattern (no business logic in routes)
 4. Validate all inputs with the project's validation library
 5. Parameterized queries only (no string interpolation in SQL)
diff --git a/.codex/agents/debug-investigator.toml b/.codex/agents/debug-investigator.toml
index cad7527..4b56814 100644
--- a/.codex/agents/debug-investigator.toml
+++ b/.codex/agents/debug-investigator.toml
@@ -46,7 +46,7 @@ CHARTER_CHECK:
 1. Stay in scope — only work on assigned debug tasks
 2. Fix root cause, not symptoms
 3. Minimal changes only — no refactoring during bugfix; route refactoring needs to refactor-engineer
-4. Every fix gets a regression test
+4. Every fix gets a regression test; run it before the fix where feasible and record RED (failing output) → GREEN (post-fix pass) in the bug report
 5. Search for similar patterns after fixing
 6. Document out-of-scope findings for other agents
 7. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.codex/agents/frontend-engineer.toml b/.codex/agents/frontend-engineer.toml
index e5816ad..d277d57 100644
--- a/.codex/agents/frontend-engineer.toml
+++ b/.codex/agents/frontend-engineer.toml
@@ -46,7 +46,7 @@ FSD-lite: root `src/` + feature `src/features/*/`
 5. TailwindCSS v4 for styling, design tokens 1:1 mapping
 6. Library defaults (greenfield; existing project choices win): luxon (dates), ahooks (hooks), es-toolkit (utils), jotai (client state), TanStack Query (server state)
 7. Absolute imports with `@/`
-8. Write tests for custom logic (>90% coverage target)
+8. Write tests for custom logic (>90% coverage target); honor the plan task's `test_approach` — for `tdd`, demonstrate RED before the change and record a `TDD_EVIDENCE` block (test command, RED, GREEN) in the result file
 9. Document out-of-scope dependencies for other agents
 10. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
 """
diff --git a/.codex/agents/mobile-engineer.toml b/.codex/agents/mobile-engineer.toml
index 5cb6cc8..82f66f1 100644
--- a/.codex/agents/mobile-engineer.toml
+++ b/.codex/agents/mobile-engineer.toml
@@ -46,7 +46,7 @@ Clean Architecture: domain → data → presentation (Swift native: App/Core/Fea
 5. Transport client with interceptors (Dio / axios / generated Client) + repository-layer response cache, offline-first architecture
 6. Secrets in secure storage only — never plain prefs or MMKV
 7. 60fps target performance
-8. Write widget/component tests and integration tests
+8. Write widget/component tests and integration tests; honor the plan task's `test_approach` — for `tdd`, demonstrate RED before the change and record a `TDD_EVIDENCE` block (test command, RED, GREEN) in the result file
 9. ARB-based localization: edit ARB source files only, never generated localization code
 10. Document out-of-scope dependencies for other agents
 11. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.codex/agents/pm-planner.toml b/.codex/agents/pm-planner.toml
index 9a64b73..a58d2a4 100644
--- a/.codex/agents/pm-planner.toml
+++ b/.codex/agents/pm-planner.toml
@@ -50,13 +50,14 @@ Each task must include:
 - `priority`: execution tier — 1 = independent (runs first), 2 = depends on tier 1, etc. (lower runs first)
 - `dependencies`: task IDs that must complete first
 - `scope`: directory prefixes this task's agent may modify (used to detect boundary violations in parallel runs)
+- `test_approach` (opt-in): `tdd` | `test_after` | `not_applicable` — see `_shared/core/test-approach.md`. `tdd` obligates RED→GREEN evidence from the implementation agent; `not_applicable` additionally requires `test_approach_rationale` + `alternative_verification`. Never assign `tdd` to refactor tasks (characterization tests instead)
 
 ## Rules
 
 1. Stay in scope — planning only, no code implementation
 2. API-first design
 3. Minimize dependencies for maximum parallelism
-4. Security and testing are part of every task (not separate)
+4. Security and testing are part of every task (not separate); assign per-task `test_approach` (`tdd|test_after|not_applicable`) where a test strategy matters — `not_applicable` requires rationale + alternative verification, and no approach waives the >= 80% coverage gate
 5. Each task completable by a single agent
 6. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
 """
diff --git a/.cursor/agents/backend-engineer.md b/.cursor/agents/backend-engineer.md
index 2fefd28..8596eb0 100644
--- a/.cursor/agents/backend-engineer.md
+++ b/.cursor/agents/backend-engineer.md
@@ -45,7 +45,7 @@ Router (HTTP) → Service (Business Logic) → Repository (Data Access) → Mode
 ## Rules
 
 1. Stay in scope — only work on assigned backend tasks
-2. Write tests for all new code
+2. Write tests for all new code; honor the plan task's `test_approach` — for `tdd`, demonstrate RED before the change and record a `TDD_EVIDENCE` block (test command, RED, GREEN) in the result file
 3. Follow Repository → Service → Router pattern (no business logic in routes)
 4. Validate all inputs with the project's validation library
 5. Parameterized queries only (no string interpolation in SQL)
diff --git a/.cursor/agents/debug-investigator.md b/.cursor/agents/debug-investigator.md
index 3b96a2a..4d27b0b 100644
--- a/.cursor/agents/debug-investigator.md
+++ b/.cursor/agents/debug-investigator.md
@@ -51,7 +51,7 @@ CHARTER_CHECK:
 1. Stay in scope — only work on assigned debug tasks
 2. Fix root cause, not symptoms
 3. Minimal changes only — no refactoring during bugfix; route refactoring needs to refactor-engineer
-4. Every fix gets a regression test
+4. Every fix gets a regression test; run it before the fix where feasible and record RED (failing output) → GREEN (post-fix pass) in the bug report
 5. Search for similar patterns after fixing
 6. Document out-of-scope findings for other agents
 7. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.cursor/agents/frontend-engineer.md b/.cursor/agents/frontend-engineer.md
index f3413cd..07a8874 100644
--- a/.cursor/agents/frontend-engineer.md
+++ b/.cursor/agents/frontend-engineer.md
@@ -53,6 +53,6 @@ FSD-lite: root `src/` + feature `src/features/*/`
 5. TailwindCSS v4 for styling, design tokens 1:1 mapping
 6. Library defaults (greenfield; existing project choices win): luxon (dates), ahooks (hooks), es-toolkit (utils), jotai (client state), TanStack Query (server state)
 7. Absolute imports with `@/`
-8. Write tests for custom logic (>90% coverage target)
+8. Write tests for custom logic (>90% coverage target); honor the plan task's `test_approach` — for `tdd`, demonstrate RED before the change and record a `TDD_EVIDENCE` block (test command, RED, GREEN) in the result file
 9. Document out-of-scope dependencies for other agents
 10. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.cursor/agents/mobile-engineer.md b/.cursor/agents/mobile-engineer.md
index c439583..4cfbd95 100644
--- a/.cursor/agents/mobile-engineer.md
+++ b/.cursor/agents/mobile-engineer.md
@@ -52,7 +52,7 @@ Clean Architecture: domain → data → presentation (Swift native: App/Core/Fea
 5. Transport client with interceptors (Dio / axios / generated Client) + repository-layer response cache, offline-first architecture
 6. Secrets in secure storage only — never plain prefs or MMKV
 7. 60fps target performance
-8. Write widget/component tests and integration tests
+8. Write widget/component tests and integration tests; honor the plan task's `test_approach` — for `tdd`, demonstrate RED before the change and record a `TDD_EVIDENCE` block (test command, RED, GREEN) in the result file
 9. ARB-based localization: edit ARB source files only, never generated localization code
 10. Document out-of-scope dependencies for other agents
 11. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.cursor/agents/pm-planner.md b/.cursor/agents/pm-planner.md
index c145402..1c6cfc7 100644
--- a/.cursor/agents/pm-planner.md
+++ b/.cursor/agents/pm-planner.md
@@ -55,12 +55,13 @@ Each task must include:
 - `priority`: execution tier — 1 = independent (runs first), 2 = depends on tier 1, etc. (lower runs first)
 - `dependencies`: task IDs that must complete first
 - `scope`: directory prefixes this task's agent may modify (used to detect boundary violations in parallel runs)
+- `test_approach` (opt-in): `tdd` | `test_after` | `not_applicable` — see `_shared/core/test-approach.md`. `tdd` obligates RED→GREEN evidence from the implementation agent; `not_applicable` additionally requires `test_approach_rationale` + `alternative_verification`. Never assign `tdd` to refactor tasks (characterization tests instead)
 
 ## Rules
 
 1. Stay in scope — planning only, no code implementation
 2. API-first design
 3. Minimize dependencies for maximum parallelism
-4. Security and testing are part of every task (not separate)
+4. Security and testing are part of every task (not separate); assign per-task `test_approach` (`tdd|test_after|not_applicable`) where a test strategy matters — `not_applicable` requires rationale + alternative verification, and no approach waives the >= 80% coverage gate
 5. Each task completable by a single agent
 6. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.gitignore b/.gitignore
index 3b645f6..f4db8f3 100644
--- a/.gitignore
+++ b/.gitignore
@@ -98,3 +98,5 @@ docs/generated/
 .agents/backup/
 
 docs/plans/
+
+.qwen/tmp/
diff --git a/.opencode/agents/backend-engineer.md b/.opencode/agents/backend-engineer.md
index 5100d78..7c71bba 100644
--- a/.opencode/agents/backend-engineer.md
+++ b/.opencode/agents/backend-engineer.md
@@ -42,7 +42,7 @@ Router (HTTP) → Service (Business Logic) → Repository (Data Access) → Mode
 ## Rules
 
 1. Stay in scope — only work on assigned backend tasks
-2. Write tests for all new code
+2. Write tests for all new code; honor the plan task's `test_approach` — for `tdd`, demonstrate RED before the change and record a `TDD_EVIDENCE` block (test command, RED, GREEN) in the result file
 3. Follow Repository → Service → Router pattern (no business logic in routes)
 4. Validate all inputs with the project's validation library
 5. Parameterized queries only (no string interpolation in SQL)
diff --git a/.opencode/agents/debug-investigator.md b/.opencode/agents/debug-investigator.md
index fe2f2e7..e777357 100644
--- a/.opencode/agents/debug-investigator.md
+++ b/.opencode/agents/debug-investigator.md
@@ -49,7 +49,7 @@ CHARTER_CHECK:
 1. Stay in scope — only work on assigned debug tasks
 2. Fix root cause, not symptoms
 3. Minimal changes only — no refactoring during bugfix; route refactoring needs to refactor-engineer
-4. Every fix gets a regression test
+4. Every fix gets a regression test; run it before the fix where feasible and record RED (failing output) → GREEN (post-fix pass) in the bug report
 5. Search for similar patterns after fixing
 6. Document out-of-scope findings for other agents
 7. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.opencode/agents/frontend-engineer.md b/.opencode/agents/frontend-engineer.md
index d3adc79..395b8bf 100644
--- a/.opencode/agents/frontend-engineer.md
+++ b/.opencode/agents/frontend-engineer.md
@@ -49,6 +49,6 @@ FSD-lite: root `src/` + feature `src/features/*/`
 5. TailwindCSS v4 for styling, design tokens 1:1 mapping
 6. Library defaults (greenfield; existing project choices win): luxon (dates), ahooks (hooks), es-toolkit (utils), jotai (client state), TanStack Query (server state)
 7. Absolute imports with `@/`
-8. Write tests for custom logic (>90% coverage target)
+8. Write tests for custom logic (>90% coverage target); honor the plan task's `test_approach` — for `tdd`, demonstrate RED before the change and record a `TDD_EVIDENCE` block (test command, RED, GREEN) in the result file
 9. Document out-of-scope dependencies for other agents
 10. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.opencode/agents/mobile-engineer.md b/.opencode/agents/mobile-engineer.md
index c8d8ac9..bb4360d 100644
--- a/.opencode/agents/mobile-engineer.md
+++ b/.opencode/agents/mobile-engineer.md
@@ -49,7 +49,7 @@ Clean Architecture: domain → data → presentation (Swift native: App/Core/Fea
 5. Transport client with interceptors (Dio / axios / generated Client) + repository-layer response cache, offline-first architecture
 6. Secrets in secure storage only — never plain prefs or MMKV
 7. 60fps target performance
-8. Write widget/component tests and integration tests
+8. Write widget/component tests and integration tests; honor the plan task's `test_approach` — for `tdd`, demonstrate RED before the change and record a `TDD_EVIDENCE` block (test command, RED, GREEN) in the result file
 9. ARB-based localization: edit ARB source files only, never generated localization code
 10. Document out-of-scope dependencies for other agents
 11. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.opencode/agents/pm-planner.md b/.opencode/agents/pm-planner.md
index 4284400..871de02 100644
--- a/.opencode/agents/pm-planner.md
+++ b/.opencode/agents/pm-planner.md
@@ -52,12 +52,13 @@ Each task must include:
 - `priority`: execution tier — 1 = independent (runs first), 2 = depends on tier 1, etc. (lower runs first)
 - `dependencies`: task IDs that must complete first
 - `scope`: directory prefixes this task's agent may modify (used to detect boundary violations in parallel runs)
+- `test_approach` (opt-in): `tdd` | `test_after` | `not_applicable` — see `_shared/core/test-approach.md`. `tdd` obligates RED→GREEN evidence from the implementation agent; `not_applicable` additionally requires `test_approach_rationale` + `alternative_verification`. Never assign `tdd` to refactor tasks (characterization tests instead)
 
 ## Rules
 
 1. Stay in scope — planning only, no code implementation
 2. API-first design
 3. Minimize dependencies for maximum parallelism
-4. Security and testing are part of every task (not separate)
+4. Security and testing are part of every task (not separate); assign per-task `test_approach` (`tdd|test_after|not_applicable`) where a test strategy matters — `not_applicable` requires rationale + alternative verification, and no approach waives the >= 80% coverage gate
 5. Each task completable by a single agent
 6. Never modify `.agents/` files (SSOT) — run outputs under `.agents/results/` and `.agents/state/memories/` are the only exceptions
diff --git a/.opencode/plugins/oma/keyword-detector.ts b/.opencode/plugins/oma/keyword-detector.ts
index e336814..08094f6 100644
--- a/.opencode/plugins/oma/keyword-detector.ts
+++ b/.opencode/plugins/oma/keyword-detector.ts
@@ -357,20 +357,35 @@ export function escapeRegex(s: string): string {
 
 /**
  * Merge a language-keyed keyword/pattern bank into a single flat list:
- * universal ("*") + English (the universal default) + the configured
- * language's own entries (skipped when lang === "en" to avoid duplicates).
- * Shared by buildPatterns and buildRawPatterns — both keyword banks and
- * pattern banks use this exact `Record<string, string[]>` shape.
+ * universal ("*") + English + EVERY other language's entries, deduped
+ * case-insensitively. Same rationale as RC4 (buildInformationalPatterns):
+ * users prompt in whichever language they think in — `language` in
+ * oma-config.yaml controls the RESPONSE language, not the prompt language —
+ * so gating by config language silently disabled e.g. every Korean trigger
+ * for `language: en` projects. A keyword written in language X can only
+ * match a prompt that contains X-script text (current banks are en/ko/ja/zh;
+ * if a Latin-script bank like es/fr is ever added, phrase distinctiveness is
+ * the gate instead), so merging all languages cannot fire on unrelated
+ * prompts. Shared by buildPatterns and buildRawPatterns — both keyword banks
+ * and pattern banks use this exact `Record<string, string[]>` shape.
  */
-export function collectLangEntries(
-  bank: Record<string, string[]>,
-  lang: string,
-): string[] {
-  return [
+export function collectLangEntries(bank: Record<string, string[]>): string[] {
+  const ordered = [
     ...(bank["*"] ?? []),
     ...(bank.en ?? []),
-    ...(lang !== "en" ? (bank[lang] ?? []) : []),
+    ...Object.entries(bank)
+      .filter(([key]) => key !== "*" && key !== "en")
+      .flatMap(([, entries]) => entries),
   ];
+  const seen = new Set<string>();
+  const out: string[] = [];
+  for (const entry of ordered) {
+    const key = entry.toLowerCase();
+    if (seen.has(key)) continue;
+    seen.add(key);
+    out.push(entry);
+  }
+  return out;
 }
 
 /**
@@ -390,7 +405,7 @@ export function buildPatternEntries(
   lang: string,
   cjkScripts: string[],
 ): KeywordPatternEntry[] {
-  return collectLangEntries(keywords, lang).map((kw) => {
+  return collectLangEntries(keywords).map((kw) => {
     const escaped = escapeRegex(kw).replace(/\s+/g, "\\s+");
     const regex =
       cjkScripts.includes(lang) || /[^\p{ASCII}]/u.test(kw)
@@ -421,11 +436,10 @@ export interface RawPatternEntry {
 
 export function buildRawPatternEntries(
   patterns: Record<string, string[]> | undefined,
-  lang: string,
 ): RawPatternEntry[] {
   if (!patterns) return [];
   const compiled: RawPatternEntry[] = [];
-  for (const raw of collectLangEntries(patterns, lang)) {
+  for (const raw of collectLangEntries(patterns)) {
     try {
       compiled.push({ regex: new RegExp(raw, "iu"), source: raw });
     } catch {
@@ -443,9 +457,8 @@ export function buildRawPatternEntries(
  */
 export function buildRawPatterns(
   patterns: Record<string, string[]> | undefined,
-  lang: string,
 ): RegExp[] {
-  return buildRawPatternEntries(patterns, lang).map((e) => e.regex);
+  return buildRawPatternEntries(patterns).map((e) => e.regex);
 }
 
 export function buildInformationalPatterns(config: TriggerConfig): RegExp[] {
@@ -707,6 +720,7 @@ function activateMode(
   projectDir: string,
   workflow: string,
   sessionId: string,
+  omaSid?: string | null,
 ): void {
   // Never persist a workflow under the unresolved-session fallback id: such a
   // file cannot be isolated per session and would cross-contaminate any later
@@ -718,6 +732,7 @@ function activateMode(
     sessionId,
     activatedAt: new Date().toISOString(),
     reinforcementCount: 0,
+    ...(omaSid ? { omaSid } : {}),
   };
   writeFileSync(
     join(getStateDir(projectDir), `${workflow}-state-${sessionId}.json`),
@@ -770,11 +785,12 @@ export const DEACTIVATION_PHRASES: Record<string, string[]> = {
   pl: ["workflow zakończony", "workflow ukończony"],
 };
 
-export function isDeactivationRequest(prompt: string, lang: string): boolean {
-  const phrases = [
-    ...(DEACTIVATION_PHRASES.en ?? []),
-    ...(lang !== "en" ? (DEACTIVATION_PHRASES[lang] ?? []) : []),
-  ];
+export function isDeactivationRequest(prompt: string): boolean {
+  // All languages merged, never gated by config language (same rationale as
+  // collectLangEntries): a user prompting in Korean must be able to say
+  // "워크플로우 완료" even when `language: en`. A phrase only matches a prompt
+  // actually written in that language, so merging cannot misfire.
+  const phrases = Object.values(DEACTIVATION_PHRASES).flat();
   const normalized = normalizeForMatching(prompt);
   return phrases.some((phrase) =>
     normalized.includes(normalizeForMatching(phrase)),
@@ -925,7 +941,7 @@ export async function run(
   const lang = detectLanguage(projectDir);
 
   // Check for deactivation request before workflow detection
-  if (isDeactivationRequest(prompt, lang)) {
+  if (isDeactivationRequest(prompt)) {
     deactivateAllPersistentModes(projectDir, sessionId);
     // Grok's resume context lives in a session-start file, not L1 stdout — clear it.
     if (vendor === "grok") clearGrokContext(projectDir);
@@ -1025,10 +1041,7 @@ export async function run(
     )) {
       considerMatch(regex, keyword);
     }
-    for (const { regex, source } of buildRawPatternEntries(
-      def.patterns,
-      lang,
-    )) {
+    for (const { regex, source } of buildRawPatternEntries(def.patterns)) {
       considerMatch(regex, source);
     }
   }
@@ -1038,10 +1051,17 @@ export async function run(
 
   const { workflow } = winner;
 
+  // Activate the L1 session first so its sid can be recorded in the
+  // persistent-mode state file (the Stop hook emits gate events under it).
+  const omaSid = await activateL1WorkflowSession(
+    projectDir,
+    workflow,
+    vendor,
+    sessionId,
+  );
   if (winner.persistent) {
-    activateMode(projectDir, workflow, sessionId);
+    activateMode(projectDir, workflow, sessionId, omaSid);
   }
-  await activateL1WorkflowSession(projectDir, workflow, vendor, sessionId);
   const updatedState = recordKwTrigger(kwState, workflow);
   saveKwState(projectDir, updatedState);
 
diff --git a/.opencode/plugins/oma/persistent-mode.ts b/.opencode/plugins/oma/persistent-mode.ts
index e20dda6..698a0c3 100644
--- a/.opencode/plugins/oma/persistent-mode.ts
+++ b/.opencode/plugins/oma/persistent-mode.ts
@@ -13,6 +13,7 @@
  * exit 2 = block stop
  */
 
+import { spawnSync } from "node:child_process";
 import {
   existsSync,
   readdirSync,
@@ -40,15 +41,123 @@ import { getProjectDir } from "./vendor-detect.ts";
 const MAX_REINFORCEMENTS = 5;
 const STALE_HOURS = 2;
 
-function detectLanguage(projectDir: string): string {
-  const prefsPath = join(projectDir, ".agents", "oma-config.yaml");
-  if (!existsSync(prefsPath)) return "en";
+// ── Goal contract: deterministic stop gate + wall-clock budget ─
+// (design-prime-agent-adoption Track B — no-exec-of-agent-writable-strings)
+
+/**
+ * The only gate values the Stop hook will ever execute. Each maps to a
+ * package.json script of the same name, run as an argv array WITHOUT a shell.
+ * The gate value lives in an agent-writable state file; executing anything
+ * outside this allowlist would be an arbitrary-command path that bypasses the
+ * PreToolUse permission layer. Never widen this to free-form strings.
+ */
+const GATE_KEYWORDS = new Set(["typecheck", "test", "lint"]);
+
+/** Hard cap on a gate run; SIGKILL after this. Keeps Stop-hook latency bounded. */
+const GATE_TIMEOUT_MS = 60_000;
+
+/** Tail of gate output carried back into the block reason. */
+const GATE_OUTPUT_TAIL_CHARS = 2_000;
+
+/**
+ * Resolve an allowlisted gate keyword to a package-runner argv, or null when
+ * the keyword is not allowlisted, package.json is absent, or it defines no
+ * script of that name. Pure node:fs — no shell, no third-party imports.
+ */
+export function resolveGateArgv(
+  gateKeyword: string,
+  projectDir: string,
+): string[] | null {
+  if (!GATE_KEYWORDS.has(gateKeyword)) return null;
+  const pkgPath = join(projectDir, "package.json");
+  if (!existsSync(pkgPath)) return null;
   try {
-    const content = readFileSync(prefsPath, "utf-8");
-    const match = content.match(/^language:\s*(\S+)/m);
-    return match?.[1] ?? "en";
+    const pkg = JSON.parse(readFileSync(pkgPath, "utf-8")) as {
+      scripts?: Record<string, unknown>;
+    };
+    if (typeof pkg.scripts?.[gateKeyword] !== "string") return null;
   } catch {
-    return "en";
+    return null;
+  }
+  if (
+    existsSync(join(projectDir, "bun.lock")) ||
+    existsSync(join(projectDir, "bun.lockb"))
+  ) {
+    return ["bun", "run", gateKeyword];
+  }
+  if (existsSync(join(projectDir, "pnpm-lock.yaml"))) {
+    return ["pnpm", "run", gateKeyword];
+  }
+  if (existsSync(join(projectDir, "yarn.lock"))) {
+    return ["yarn", gateKeyword];
+  }
+  return ["npm", "run", gateKeyword];
+}
+
+export interface GateRunResult {
+  passed: boolean;
+  timedOut: boolean;
+  outputTail: string;
+}
+
+/** Run a resolved gate argv with a hard timeout. No shell involved. */
+export function runGateCommand(
+  argv: string[],
+  projectDir: string,
+): GateRunResult {
+  const [command, ...args] = argv;
+  const result = spawnSync(command as string, args, {
+    cwd: projectDir,
+    encoding: "utf-8",
+    timeout: GATE_TIMEOUT_MS,
+    killSignal: "SIGKILL",
+    maxBuffer: 8 * 1024 * 1024,
+  });
+  const timedOut =
+    (result.error as NodeJS.ErrnoException | undefined)?.code === "ETIMEDOUT" ||
+    result.signal === "SIGKILL";
+  const combined = `${result.stdout ?? ""}\n${result.stderr ?? ""}`.trim();
+  return {
+    passed: result.status === 0 && !result.error,
+    timedOut,
+    outputTail: combined.slice(-GATE_OUTPUT_TAIL_CHARS),
+  };
+}
+
+/** True when the goal's wall-clock budget (from activatedAt) is exhausted. */
+export function isBudgetExhausted(state: ModeState): boolean {
+  const minutes = state.goal?.budget?.wallClockMinutes;
+  if (
+    typeof minutes !== "number" ||
+    !Number.isFinite(minutes) ||
+    minutes <= 0
+  ) {
+    return false;
+  }
+  const elapsedMs = Date.now() - new Date(state.activatedAt).getTime();
+  return elapsedMs >= minutes * 60_000;
+}
+
+/**
+ * Emit a gate event onto the L1 trail recorded at activation. Best-effort:
+ * older state files carry no omaSid, and event emission must never break the
+ * Stop decision itself.
+ */
+async function emitGateEvent(
+  projectDir: string,
+  state: ModeState,
+  kind: "gate.passed" | "gate.failed",
+  payload: Record<string, unknown>,
+): Promise<void> {
+  if (!state.omaSid) return;
+  try {
+    const { emitEvent } = await import("./state-emit.ts");
+    await emitEvent(projectDir, state.omaSid, {
+      kind,
+      payload: { workflow: state.workflow, ...payload },
+    });
+  } catch {
+    // best-effort — never let event I/O change the stop decision
   }
 }
 
@@ -218,8 +327,7 @@ export async function run(
   // text (parity with the standalone main() path). Without this, persistent
   // mode could not be deactivated via the central `oma hook` dispatch.
   if (input.responseText) {
-    const lang = detectLanguage(projectDir);
-    if (isDeactivationRequest(input.responseText, lang)) {
+    if (isDeactivationRequest(input.responseText)) {
       deactivateAllForSession(projectDir, sessionId);
       return null;
     }
@@ -236,16 +344,75 @@ export async function run(
       continue;
     }
 
-    incrementReinforcement(projectDir, workflow, sessionId, state);
+    // (1) Wall-clock budget: exhausted → honest partial stop. A machine
+    // verdict, not model discretion — the stop is allowed and the exhaustion
+    // is recorded on the L1 trail.
+    if (isBudgetExhausted(state)) {
+      deactivate(projectDir, workflow, sessionId);
+      await emitGateEvent(projectDir, state, "gate.failed", {
+        gate: "budget",
+        summary: `wall-clock budget (${state.goal?.budget?.wallClockMinutes}m) exhausted for /${workflow}; stopping with partial status`,
+      });
+      continue;
+    }
 
     const stateFile = `.agents/state/${workflow}-state-${sessionId}.json`;
+
+    // (2) Deterministic completion gate. Only allowlisted keywords resolve to
+    // a runnable argv; anything else (including free-form shell strings an
+    // agent may have written into the state file) is NEVER executed and falls
+    // through to the plain reinforcement block below.
+    const gateKeyword = state.goal?.completion?.gate;
+    let ignoredGateNote = "";
+    if (gateKeyword) {
+      const argv = resolveGateArgv(gateKeyword, projectDir);
+      if (argv) {
+        const gate = runGateCommand(argv, projectDir);
+        if (gate.passed) {
+          // The gate is the mechanical proof of completion: allow the stop.
+          deactivate(projectDir, workflow, sessionId);
+          await emitGateEvent(projectDir, state, "gate.passed", {
+            gate: gateKeyword,
+            summary: `stop gate '${gateKeyword}' passed for /${workflow}`,
+          });
+          continue;
+        }
+        // Failure and timeout both count toward MAX_REINFORCEMENTS so a
+        // permanently red gate cannot block stops forever.
+        incrementReinforcement(projectDir, workflow, sessionId, state);
+        await emitGateEvent(projectDir, state, "gate.failed", {
+          gate: gateKeyword,
+          timedOut: gate.timedOut,
+          summary: `stop gate '${gateKeyword}' ${gate.timedOut ? "timed out" : "failed"} for /${workflow}`,
+        });
+        const reason = [
+          `[OMA PERSISTENT MODE: ${workflow.toUpperCase()}]`,
+          `Stop gate '${gateKeyword}' ${gate.timedOut ? `timed out after ${GATE_TIMEOUT_MS / 1000}s` : "FAILED"} (reinforcement ${state.reinforcementCount}/${MAX_REINFORCEMENTS}).`,
+          `Fix the failures below, then finish the workflow — the stop is allowed only when the gate passes.`,
+          gate.outputTail
+            ? `--- gate output (tail) ---\n${gate.outputTail}`
+            : "",
+          `To abandon instead: delete ${stateFile} or say "workflow done".`,
+        ]
+          .filter(Boolean)
+          .join("\n");
+        return { type: "block", reason };
+      }
+      ignoredGateNote = `Note: configured stop gate ${JSON.stringify(gateKeyword)} is not an allowed keyword (typecheck|test|lint) or has no matching package.json script — it was NOT executed.`;
+    }
+
+    incrementReinforcement(projectDir, workflow, sessionId, state);
+
     const reason = [
       `[OMA PERSISTENT MODE: ${workflow.toUpperCase()}]`,
       `The /${workflow} workflow is still active (reinforcement ${state.reinforcementCount}/${MAX_REINFORCEMENTS}).`,
       `Continue executing the workflow. If all tasks are genuinely complete:`,
       `  1. Delete the state file: Bash \`rm ${stateFile}\``,
       `  2. Or ask the user to say "워크플로우 완료" / "workflow done"`,
-    ].join("\n");
+      ignoredGateNote,
+    ]
+      .filter(Boolean)
+      .join("\n");
 
     return { type: "block", reason };
   }
@@ -267,7 +434,6 @@ async function main() {
   const vendor = detectVendor(input);
   const projectDir = getProjectDir(vendor, input);
   const sessionId = getSessionId(input);
-  const lang = detectLanguage(projectDir);
 
   // Check all text fields in stdin for deactivation phrases.
   // The assistant may have included "workflow done" in its response,
@@ -284,7 +450,7 @@ async function main() {
     .filter((v): v is string => typeof v === "string")
     .join(" ");
 
-  if (textToCheck && isDeactivationRequest(textToCheck, lang)) {
+  if (textToCheck && isDeactivationRequest(textToCheck)) {
     // Deactivate all persistent workflows for this session (shared helper).
     deactivateAllForSession(projectDir, sessionId);
     process.exit(0);
diff --git a/.opencode/plugins/oma/serena-primer.ts b/.opencode/plugins/oma/serena-primer.ts
index df2aabc..07eaca2 100644
--- a/.opencode/plugins/oma/serena-primer.ts
+++ b/.opencode/plugins/oma/serena-primer.ts
@@ -117,6 +117,7 @@ export function primerContext(): string {
     "- Code discovery / reading: `get_symbols_overview`, `find_symbol`, `find_referencing_symbols`, `search_for_pattern`.",
     "- Code edits: `replace_symbol_body`, `insert_after_symbol`, `insert_before_symbol`, `replace_content`.",
     "- Native grep/glob: only for initial filename/path discovery. Do not fall back to grep + Read for code navigation just because Serena's tools aren't loaded yet — load them.",
+    '- Result size: omit `max_answer_chars` on Serena tools (uses the configured default, typically 150000). Never pass small caps like `3000` on broad searches. If a call returns "The answer is too long (N characters)", retry with `max_answer_chars` > N or narrow path/glob — do not keep the low cap.',
     "- Exception — MCP timeout: if a Serena MCP call times out or hangs (seen mainly in OpenCode Desktop's long-lived sidecar), stop retrying MCP for this session: use native search/read for code, and access `.serena/memories/` files directly (or `serena memories read|write` when Serena CLI ≥ 1.5 is installed) for memory work. A full app relaunch restores Serena MCP.",
   ].join("\n");
 }
diff --git a/.opencode/plugins/oma/skill-injector.ts b/.opencode/plugins/oma/skill-injector.ts
index 66c8139..06fe400 100644
--- a/.opencode/plugins/oma/skill-injector.ts
+++ b/.opencode/plugins/oma/skill-injector.ts
@@ -188,10 +188,16 @@ export function matchSkills(
     const jsonEntry = config.skills?.[skill.name];
     if (!jsonEntry) continue;
 
+    // All languages merged, never gated by config language: users prompt in
+    // whichever language they think in (`language` controls the RESPONSE
+    // language). A keyword written in language X can only match a prompt
+    // containing X-script text, so merging cannot fire on unrelated prompts.
     const jsonTriggers = [
       ...(jsonEntry.keywords["*"] ?? []),
       ...(jsonEntry.keywords.en ?? []),
-      ...(lang !== "en" ? (jsonEntry.keywords[lang] ?? []) : []),
+      ...Object.entries(jsonEntry.keywords)
+        .filter(([key]) => key !== "*" && key !== "en")
+        .flatMap(([, entries]) => entries),
     ];
 
     const seen = new Set<string>();
diff --git a/.opencode/plugins/oma/triggers.json b/.opencode/plugins/oma/triggers.json
index 8e5bff8..9aee86c 100644
--- a/.opencode/plugins/oma/triggers.json
+++ b/.opencode/plugins/oma/triggers.json
@@ -1211,12 +1211,10 @@
           "랄프",
           "멈추지마",
           "멈추지 말고",
-          "끝까지",
           "완료될때까지",
           "될때까지 해",
           "끝날때까지",
           "다 끝내",
-          "다 해",
           "전부 완료",
           "끝까지 해",
           "중단하지마",
diff --git a/.opencode/plugins/oma/types.ts b/.opencode/plugins/oma/types.ts
index a58ffda..219a0e1 100644
--- a/.opencode/plugins/oma/types.ts
+++ b/.opencode/plugins/oma/types.ts
@@ -35,11 +35,45 @@ export interface RawHookInput {
   stopReason?: string;
 }
 
+/**
+ * Optional goal contract for a persistent workflow (design-prime-agent-adoption
+ * Track B). Written by `oma goal:set`; read by the persistent-mode Stop hook.
+ */
+export interface ModeGoal {
+  /** Human description of the objective. Informational only. */
+  description?: string;
+  budget?: {
+    /**
+     * Wall-clock budget in minutes, measured from `activatedAt`. When
+     * exceeded the Stop hook deactivates the workflow and allows an honest
+     * partial stop (machine verdict, not model discretion).
+     */
+    wallClockMinutes?: number;
+  };
+  completion?: {
+    /**
+     * Deterministic stop gate. MUST be an allowlist keyword ("typecheck" |
+     * "test" | "lint") that maps to an existing package.json script; the hook
+     * runs it as an argv array with no shell. Free-form strings are NEVER
+     * executed — this value lives in an agent-writable state file, so
+     * executing it verbatim would be an arbitrary-command-execution path
+     * that bypasses the PreToolUse permission layer.
+     */
+    gate?: string;
+  };
+}
+
 export interface ModeState {
   workflow: string;
   sessionId: string;
   activatedAt: string;
   reinforcementCount: number;
+  /**
+   * L1 session id (`oma-…`) recorded at activation so the Stop hook can emit
+   * gate.passed / gate.failed events onto the same events.jsonl trail.
+   */
+  omaSid?: string;
+  goal?: ModeGoal;
 }
 
 // ---------------------------------------------------------------------------
diff --git a/.qwen/settings.json b/.qwen/settings.json
index b332f70..ba6a426 100644
--- a/.qwen/settings.json
+++ b/.qwen/settings.json
@@ -72,5 +72,10 @@
   "$version": 4,
   "hooksConfig": {
     "enabled": true
+  },
+  "model": {
+    "generationConfig": {
+      "timeout": 300000
+    }
   }
 }