@theokit/sdk 4.63.3 → 5.0.0-next.1

package/dist/{run-until-U2B5FJJI.cjs → run-until-DG2KXWB2.cjs}

@@ -3,6 +3,12 @@
 var chunkOBA6YK2R_cjs = require('./chunk-OBA6YK2R.cjs');
 
 // src/internal/runtime/lifecycle/run-until.ts
+var ABORTED_MID_RUN = "aborted via AbortSignal";
+var pausedBy = (reason) => ({
+  type: "status_change",
+  status: "paused",
+  reason
+});
 async function* runUntilImpl(agent, goal, options, deps) {
   const maxTurns = options?.maxTurns ?? 20;
   const maxFails = options?.maxConsecutiveJudgeFailures ?? 3;
@@ -13,31 +19,28 @@ async function* runUntilImpl(agent, goal, options, deps) {
   let consecutiveFailures = 0;
   let tokensUsed = 0;
   let lastResponse = "";
+  const finish = (status) => ({
+    status,
+    turnsUsed: turn,
+    tokensUsed,
+    finalResponse: lastResponse || void 0
+  });
   if (isAborted()) {
-    yield {
-      type: "status_change",
-      status: "paused",
-      reason: "aborted via AbortSignal before first turn"
-    };
-    return { status: "paused", turnsUsed: 0, tokensUsed, finalResponse: void 0 };
+    yield pausedBy("aborted via AbortSignal before first turn");
+    return finish("paused");
   }
   yield { type: "status_change", status: "active", reason: "Goal started" };
   while (turn < maxTurns) {
     if (isAborted()) {
-      yield { type: "status_change", status: "paused", reason: "aborted via AbortSignal" };
-      return {
-        status: "paused",
-        turnsUsed: turn,
-        tokensUsed,
-        finalResponse: lastResponse || void 0
-      };
+      yield pausedBy(ABORTED_MID_RUN);
+      return finish("paused");
     }
     turn += 1;
     yield { type: "turn_start", turn, goal };
     const continuationPrompt = turn === 1 ? goal : composeContinuation(goal, lastResponse);
     const run = await agent.send(continuationPrompt);
     const cancelOnAbort = () => {
-      void run.cancel?.();
+      void run.cancel();
     };
     if (signal !== void 0) {
       if (signal.aborted) cancelOnAbort();
@@ -49,13 +52,8 @@ async function* runUntilImpl(agent, goal, options, deps) {
     const turnTokens = result.usage?.totalTokens;
     if (typeof turnTokens === "number") tokensUsed += turnTokens;
     if (isAborted()) {
-      yield { type: "status_change", status: "paused", reason: "aborted via AbortSignal" };
-      return {
-        status: "paused",
-        turnsUsed: turn,
-        tokensUsed,
-        finalResponse: lastResponse || void 0
-      };
+      yield pausedBy(ABORTED_MID_RUN);
+      return finish("paused");
     }
     yield { type: "agent_response", turn, content: lastResponse };
     const judgeOpts = {};
@@ -80,33 +78,18 @@ async function* runUntilImpl(agent, goal, options, deps) {
           status: "failed",
           reason: `judge model too unreliable (${consecutiveFailures} parse failures in a row)`
         };
-        return {
-          status: "failed",
-          turnsUsed: turn,
-          tokensUsed,
-          finalResponse: lastResponse || void 0
-        };
+        return finish("failed");
       }
     } else {
       consecutiveFailures = 0;
     }
     if (judgment.verdict === "done") {
       yield { type: "status_change", status: "completed", reason: judgment.reason };
-      return {
-        status: "completed",
-        turnsUsed: turn,
-        tokensUsed,
-        finalResponse: lastResponse || void 0
-      };
+      return finish("completed");
     }
     if (judgment.verdict === "blocked") {
       yield { type: "status_change", status: "blocked", reason: judgment.reason };
-      return {
-        status: "blocked",
-        turnsUsed: turn,
-        tokensUsed,
-        finalResponse: lastResponse || void 0
-      };
+      return finish("blocked");
     }
     if (judgment.verdict === "skipped") {
       yield {
@@ -114,12 +97,7 @@ async function* runUntilImpl(agent, goal, options, deps) {
         status: "completed",
         reason: `skipped: ${judgment.reason}`
       };
-      return {
-        status: "completed",
-        turnsUsed: turn,
-        tokensUsed,
-        finalResponse: lastResponse || void 0
-      };
+      return finish("completed");
     }
     if (tokenBudget !== void 0 && tokensUsed >= tokenBudget) {
       yield {
@@ -127,12 +105,7 @@ async function* runUntilImpl(agent, goal, options, deps) {
         status: "budget_limited",
         reason: `token budget (${tokenBudget}) reached: ${tokensUsed} used`
       };
-      return {
-        status: "budget_limited",
-        turnsUsed: turn,
-        tokensUsed,
-        finalResponse: lastResponse || void 0
-      };
+      return finish("budget_limited");
     }
     yield { type: "continuation", turn, prompt: continuationPrompt };
   }
@@ -141,12 +114,7 @@ async function* runUntilImpl(agent, goal, options, deps) {
     status: "failed",
     reason: `max turns (${maxTurns}) exhausted`
   };
-  return {
-    status: "failed",
-    turnsUsed: turn,
-    tokensUsed,
-    finalResponse: lastResponse || void 0
-  };
+  return finish("failed");
 }
 function composeContinuation(goal, lastResponse) {
   return [
@@ -177,5 +145,5 @@ ${lastResponse.slice(-1e3)}`
 
 exports.composeContinuation = composeContinuation;
 exports.runUntilImpl = runUntilImpl;
-//# sourceMappingURL=run-until-U2B5FJJI.cjs.map
-//# sourceMappingURL=run-until-U2B5FJJI.cjs.map
+//# sourceMappingURL=run-until-DG2KXWB2.cjs.map
+//# sourceMappingURL=run-until-DG2KXWB2.cjs.map

package/dist/run-until-DG2KXWB2.cjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/internal/runtime/lifecycle/run-until.ts"],"names":["GOAL_CONTINUATION_MARKER"],"mappings":";;;;;AA+BA,IAAM,eAAA,GAAkB,yBAAA;AAOxB,IAAM,QAAA,GAAW,CAAC,MAAA,MAA+B;AAAA,EAC/C,IAAA,EAAM,eAAA;AAAA,EACN,MAAA,EAAQ,QAAA;AAAA,EACR;AACF,CAAA,CAAA;AAGA,gBAAuB,YAAA,CACrB,KAAA,EACA,IAAA,EACA,OAAA,EACA,IAAA,EAC6C;AAC7C,EAAA,MAAM,QAAA,GAAW,SAAS,QAAA,IAAY,EAAA;AACtC,EAAA,MAAM,QAAA,GAAW,SAAS,2BAAA,IAA+B,CAAA;AACzD,EAAA,MAAM,cAAc,OAAA,EAAS,WAAA;AAC7B,EAAA,MAAM,SAAS,OAAA,EAAS,MAAA;AAKxB,EAAA,MAAM,SAAA,GAAY,MAAe,MAAA,KAAW,MAAA,IAAa,MAAA,CAAO,OAAA;AAChE,EAAA,IAAI,IAAA,GAAO,CAAA;AACX,EAAA,IAAI,mBAAA,GAAsB,CAAA;AAC1B,EAAA,IAAI,UAAA,GAAa,CAAA;AACjB,EAAA,IAAI,YAAA,GAAe,EAAA;AAWnB,EAAA,MAAM,MAAA,GAAS,CAAC,MAAA,MAA8C;AAAA,IAC5D,MAAA;AAAA,IACA,SAAA,EAAW,IAAA;AAAA,IACX,UAAA;AAAA,IACA,eAAe,YAAA,IAAgB;AAAA,GACjC,CAAA;AAGA,EAAA,IAAI,WAAU,EAAG;AACf,IAAA,MAAM,SAAS,2CAA2C,CAAA;AAE1D,IAAA,OAAO,OAAO,QAAQ,CAAA;AAAA,EACxB;AAEA,EAAA,MAAM,EAAE,IAAA,EAAM,eAAA,EAAiB,MAAA,EAAQ,QAAA,EAAU,QAAQ,cAAA,EAAe;AAExE,EAAA,OAAO,OAAO,QAAA,EAAU;AACtB,IAAA,IAAI,WAAU,EAAG;AACf,MAAA,MAAM,SAAS,eAAe,CAAA;AAC9B,MAAA,OAAO,OAAO,QAAQ,CAAA;AAAA,IACxB;AAEA,IAAA,IAAA,IAAQ,CAAA;AACR,IAAA,MAAM,EAAE,IAAA,EAAM,YAAA,EAAc,IAAA,EAAM,IAAA,EAAK;AAEvC,IAAA,MAAM,qBAAqB,IAAA,KAAS,CAAA,GAAI,IAAA,GAAO,mBAAA,CAAoB,MAAM,YAAY,CAAA;AAErF,IAAA,MAAM,GAAA,GAAM,MAAM,KAAA,CAAM,IAAA,CAAK,kBAAkB,CAAA;AAS/C,IAAA,MAAM,gBAAgB,MAAY;AAChC,MAAA,KAAK,IAAI,MAAA,EAAO;AAAA,IAClB,CAAA;AACA,IAAA,IAAI,WAAW,MAAA,EAAW;AACxB,MAAA,IAAI,MAAA,CAAO,SAAS,aAAA,EAAc;AAAA,kBACtB,gBAAA,CAAiB,OAAA,EAAS,eAAe,EAAE,IAAA,EAAM,MAAM,CAAA;AAAA,IACrE;AACA,IAAA,MAAM,MAAA,GAAS,MAAM,GAAA,CAAI,IAAA,EAAK;AAC9B,IAAA,IAAI,MAAA,KAAW,MAAA,EAAW,MAAA,CAAO,mBAAA,CAAoB,SAAS,aAAa,CAAA;AAC3E,IAAA,YAAA,GAAe,OAAO,MAAA,IAAU,EAAA;AAEhC,IAAA,MAAM,UAAA,GAAa,OAAO,KAAA,EAAO,WAAA;AACjC,IAAA,IAAI,OAAO,UAAA,KAAe,QAAA,EAAU,UAAA,IAAc,UAAA;AAElD,IAAA,IAAI,WAAU,EAAG;AACf,MAAA,MAAM,SAAS,eAAe,CAAA;AAC9B,MAAA,OAAO,OAAO,QAAQ,CAAA;AAAA,IACxB;AACA,IAAA,MAAM,EAAE,IAAA,EAAM,gBAAA,EAAkB,IAAA,EAAM,SAAS,YAAA,EAAa;AAE5D,IAAA,MAAM,YAA0B,EAAC;AACjC,IAAA,IAAI,OAAA,EAAS,UAAA,KAAe,MAAA,EAAW,SAAA,CAAU,aAAa,OAAA,CAAQ,UAAA;AACtE,IAAA,IAAI,OAAA,EAAS,WAAA,KAAgB,MAAA,EAAW,SAAA,CAAU,SAAS,OAAA,CAAQ,WAAA;AAEnE,IAAA,IAAI,OAAA,EAAS,UAAA,KAAe,MAAA,EAAW,SAAA,CAAU,aAAa,OAAA,CAAQ,UAAA;AACtE,IAAA,MAAM,QAAA,GAAyB,EAAE,IAAA,EAAM,YAAA,EAAa;AACpD,IAAA,IAAI,OAAA,EAAS,QAAA,KAAa,MAAA,EAAW,QAAA,CAAS,WAAW,OAAA,CAAQ,QAAA;AACjE,IAAA,MAAM,QAAA,GAAW,MAAM,IAAA,CAAK,KAAA,CAAM,UAAU,SAAS,CAAA;AACrD,IAAA,MAAM;AAAA,MACJ,IAAA,EAAM,eAAA;AAAA,MACN,IAAA;AAAA,MACA,SAAS,QAAA,CAAS,OAAA;AAAA,MAClB,QAAQ,QAAA,CAAS,MAAA;AAAA,MACjB,aAAa,QAAA,CAAS;AAAA,KACxB;AAEA,IAAA,IAAI,SAAS,WAAA,EAAa;AACxB,MAAA,mBAAA,IAAuB,CAAA;AACvB,MAAA,IAAI,uBAAuB,QAAA,EAAU;AACnC,QAAA,MAAM;AAAA,UACJ,IAAA,EAAM,eAAA;AAAA,UACN,MAAA,EAAQ,QAAA;AAAA,UACR,MAAA,EAAQ,+BAA+B,mBAAmB,CAAA,yBAAA;AAAA,SAC5D;AACA,QAAA,OAAO,OAAO,QAAQ,CAAA;AAAA,MACxB;AAAA,IACF,CAAA,MAAO;AACL,MAAA,mBAAA,GAAsB,CAAA;AAAA,IACxB;AAEA,IAAA,IAAI,QAAA,CAAS,YAAY,MAAA,EAAQ;AAC/B,MAAA,MAAM,EAAE,IAAA,EAAM,eAAA,EAAiB,QAAQ,WAAA,EAAa,MAAA,EAAQ,SAAS,MAAA,EAAO;AAC5E,MAAA,OAAO,OAAO,WAAW,CAAA;AAAA,IAC3B;AAKA,IAAA,IAAI,QAAA,CAAS,YAAY,SAAA,EAAW;AAClC,MAAA,MAAM,EAAE,IAAA,EAAM,eAAA,EAAiB,QAAQ,SAAA,EAAW,MAAA,EAAQ,SAAS,MAAA,EAAO;AAC1E,MAAA,OAAO,OAAO,SAAS,CAAA;AAAA,IACzB;AACA,IAAA,IAAI,QAAA,CAAS,YAAY,SAAA,EAAW;AAClC,MAAA,MAAM;AAAA,QACJ,IAAA,EAAM,eAAA;AAAA,QACN,MAAA,EAAQ,WAAA;AAAA,QACR,MAAA,EAAQ,CAAA,SAAA,EAAY,QAAA,CAAS,MAAM,CAAA;AAAA,OACrC;AACA,MAAA,OAAO,OAAO,WAAW,CAAA;AAAA,IAC3B;AAGA,IAAA,IAAI,WAAA,KAAgB,MAAA,IAAa,UAAA,IAAc,WAAA,EAAa;AAC1D,MAAA,MAAM;AAAA,QACJ,IAAA,EAAM,eAAA;AAAA,QACN,MAAA,EAAQ,gBAAA;AAAA,QACR,MAAA,EAAQ,CAAA,cAAA,EAAiB,WAAW,CAAA,WAAA,EAAc,UAAU,CAAA,KAAA;AAAA,OAC9D;AACA,MAAA,OAAO,OAAO,gBAAgB,CAAA;AAAA,IAChC;AAEA,IAAA,MAAM,EAAE,IAAA,EAAM,cAAA,EAAgB,IAAA,EAAM,QAAQ,kBAAA,EAAmB;AAAA,EACjE;AAEA,EAAA,MAAM;AAAA,IACJ,IAAA,EAAM,eAAA;AAAA,IACN,MAAA,EAAQ,QAAA;AAAA,IACR,MAAA,EAAQ,cAAc,QAAQ,CAAA,WAAA;AAAA,GAChC;AACA,EAAA,OAAO,OAAO,QAAQ,CAAA;AACxB;AASO,SAAS,mBAAA,CAAoB,MAAc,YAAA,EAA8B;AAI9E,EAAA,OAAO;AAAA,IACLA,0CAAA;AAAA,IACA,0CAAA;AAAA,IACA,EAAA;AAAA,IACA,CAAA;AAAA,EAAgB,IAAI;AAAA,YAAA,CAAA;AAAA,IACpB,EAAA;AAAA,IACA,wBAAA;AAAA,IACA,kGAAA;AAAA,IACA,mGAAA;AAAA,IACA,2BAAA;AAAA,IACA,EAAA;AAAA,IACA,qBAAA;AAAA,IACA,kGAAA;AAAA,IACA,uGAAA;AAAA,IACA,EAAA;AAAA,IACA,mBAAA;AAAA,IACA,uGAAA;AAAA,IACA,oGAAA;AAAA,IACA,EAAA;AAAA,IACA,CAAA;AAAA,EAA4B,YAAA,CAAa,KAAA,CAAM,IAAK,CAAC,CAAA;AAAA,GACvD,CAAE,KAAK,IAAI,CAAA;AACb","file":"run-until-DG2KXWB2.cjs","sourcesContent":["/**\n * Goal-driven Ralph loop (T3.2, ADRs D115-D121).\n *\n * `runUntilImpl` is an `AsyncGenerator<GoalEvent, GoalResult, void>`:\n * yields events as the loop progresses, returns the final result when\n * the goal completes, fails, or is paused. The auxiliary judge model is\n * injected via `deps.judge` so the generator stays free of the `Agent`\n * façade import.\n *\n * EC-C fix: pre-aborted signals emit only `[paused]`, not `[active, paused]`.\n * EC-D: `maxTurns: 0` is supported (vacuous yield active → failed).\n *\n * @internal\n */\n\nimport type { SDKAgent } from \"../../../types/agent.js\";\nimport type {\n  GoalEvent,\n  GoalOptions,\n  GoalResult,\n  JudgeResult,\n} from \"../../../types/goal-events.js\";\nimport type { JudgeContext, JudgeOptions } from \"../../judge/judge-call.js\";\nimport { GOAL_CONTINUATION_MARKER } from \"./goal-marker.js\";\n\n/** DI contract: the judge is the only external touchpoint. */\nexport interface RunUntilDeps {\n  judge: (ctx: JudgeContext, opts?: JudgeOptions) => Promise<JudgeResult>;\n}\n\n/** The reason both in-loop abort checks report; the pre-loop one distinguishes itself. */\nconst ABORTED_MID_RUN = \"aborted via AbortSignal\";\n\n/**\n * The pause event, built in one place. Three sites emitted the same four-line literal with only the\n * reason varying — the same duplicated-knowledge shape as `finish` below, at the event boundary\n * instead of the result boundary.\n */\nconst pausedBy = (reason: string): GoalEvent => ({\n  type: \"status_change\",\n  status: \"paused\",\n  reason,\n});\n\n// biome-ignore lint/complexity/noExcessiveCognitiveComplexity: the goal-loop interleaves turn-start, send, judge, continuation, abort-check and failure-bail — extracting the CONTROL FLOW into helpers harms that linear narrative, and this defends only that. What it never covered were the nine hand-built result literals and the three pause events, which moved to `finish` and `pausedBy` above: neither removes a branch. Measured 2026-09-01: 43 -> 31 cognitive, 150 -> 117 lines.\nexport async function* runUntilImpl(\n  agent: SDKAgent,\n  goal: string,\n  options: GoalOptions | undefined,\n  deps: RunUntilDeps,\n): AsyncGenerator<GoalEvent, GoalResult, void> {\n  const maxTurns = options?.maxTurns ?? 20;\n  const maxFails = options?.maxConsecutiveJudgeFailures ?? 3;\n  const tokenBudget = options?.tokenBudget; // M55 — undefined ⇒ unlimited\n  const signal = options?.signal;\n  // Use a function (not direct property access) so TS does not narrow\n  // `signal.aborted` to `false | undefined` after the initial check.\n  // AbortSignal.aborted is a getter that can flip true between calls.\n  // biome-ignore lint/complexity/useOptionalChain: optional-chain would re-narrow signal?.aborted to `false | undefined` after the first check, defeating the purpose of this helper.\n  const isAborted = (): boolean => signal !== undefined && signal.aborted;\n  let turn = 0;\n  let consecutiveFailures = 0;\n  let tokensUsed = 0; // M55 — the observed sum (0 when usage is absent — fail-open)\n  let lastResponse = \"\";\n  /**\n   * The one place a `GoalResult` is built. Every exit went through the same four-field literal with\n   * only the status varying, nine times, each restating the convention that an empty final response\n   * collapses to `undefined` — duplicated KNOWLEDGE, so adding a field to `GoalResult` was a\n   * nine-site edit where missing one produces a result wrong on a single branch.\n   *\n   * It is a closure over `turn`, `tokensUsed` and `lastResponse` rather than a module function\n   * precisely so the call sites stay one word long: the biome-ignore below argues that the loop's\n   * narrative must read linearly, and this removes no branch and moves no decision.\n   */\n  const finish = (status: GoalResult[\"status\"]): GoalResult => ({\n    status,\n    turnsUsed: turn,\n    tokensUsed,\n    finalResponse: lastResponse || undefined,\n  });\n\n  // EC-C: signal already aborted BEFORE first event → emit only [paused].\n  if (isAborted()) {\n    yield pausedBy(\"aborted via AbortSignal before first turn\");\n    // `turn` is 0 and `lastResponse` is \"\" at this point, so this is the literal it replaces.\n    return finish(\"paused\");\n  }\n\n  yield { type: \"status_change\", status: \"active\", reason: \"Goal started\" };\n\n  while (turn < maxTurns) {\n    if (isAborted()) {\n      yield pausedBy(ABORTED_MID_RUN);\n      return finish(\"paused\");\n    }\n\n    turn += 1;\n    yield { type: \"turn_start\", turn, goal };\n\n    const continuationPrompt = turn === 1 ? goal : composeContinuation(goal, lastResponse);\n\n    const run = await agent.send(continuationPrompt);\n    // M55 review HIGH — thread the abort INTO the in-flight run: without this, Esc mid-turn is\n    // cosmetic (the turn keeps mutating the workspace until it finishes on its own). `run.cancel()`\n    // aborts the stream + in-flight tool calls (types/run.ts contract).\n    // Called directly, not through a hand-written shape. `SDKAgent.send()` returns `Promise<Run>`\n    // and `Run.cancel(): Promise<void>` is declared in `types/run.ts` — the cast re-declared, less\n    // precisely, what the compiler already knew, and `cancel?.()` through it meant a rename would\n    // keep compiling while Esc mid-turn silently went back to being cosmetic. Which is the exact\n    // defect the comment above says this was added to fix.\n    const cancelOnAbort = (): void => {\n      void run.cancel();\n    };\n    if (signal !== undefined) {\n      if (signal.aborted) cancelOnAbort();\n      else signal.addEventListener(\"abort\", cancelOnAbort, { once: true });\n    }\n    const result = await run.wait();\n    if (signal !== undefined) signal.removeEventListener(\"abort\", cancelOnAbort);\n    lastResponse = result.result ?? \"\";\n    // M55 — token accounting fails open: it only sums when the run reports usage.\n    const turnTokens = result.usage?.totalTokens;\n    if (typeof turnTokens === \"number\") tokensUsed += turnTokens;\n    // Re-check aborted right after the turn lands: a cancelled turn must NOT spend a judge call.\n    if (isAborted()) {\n      yield pausedBy(ABORTED_MID_RUN);\n      return finish(\"paused\");\n    }\n    yield { type: \"agent_response\", turn, content: lastResponse };\n\n    const judgeOpts: JudgeOptions = {};\n    if (options?.judgeModel !== undefined) judgeOpts.judgeModel = options.judgeModel;\n    if (options?.judgeApiKey !== undefined) judgeOpts.apiKey = options.judgeApiKey;\n    // M80 — forwards the driven agent's model so the judge can derive when `judgeModel` is omitted.\n    if (options?.agentModel !== undefined) judgeOpts.agentModel = options.agentModel;\n    const judgeCtx: JudgeContext = { goal, lastResponse };\n    if (options?.subgoals !== undefined) judgeCtx.subgoals = options.subgoals;\n    const judgment = await deps.judge(judgeCtx, judgeOpts);\n    yield {\n      type: \"judge_verdict\",\n      turn,\n      verdict: judgment.verdict,\n      reason: judgment.reason,\n      parseFailed: judgment.parseFailed,\n    };\n\n    if (judgment.parseFailed) {\n      consecutiveFailures += 1;\n      if (consecutiveFailures >= maxFails) {\n        yield {\n          type: \"status_change\",\n          status: \"failed\",\n          reason: `judge model too unreliable (${consecutiveFailures} parse failures in a row)`,\n        };\n        return finish(\"failed\");\n      }\n    } else {\n      consecutiveFailures = 0;\n    }\n\n    if (judgment.verdict === \"done\") {\n      yield { type: \"status_change\", status: \"completed\", reason: judgment.reason };\n      return finish(\"completed\");\n    }\n    // M80 — the missing arm. Without it, a judge recognizing impossibility could only say\n    // \"continue\", and the loop repeated the same turn until it blew the budget — reporting `failed` on\n    // a limit rather than `blocked` on impossibility. Two distinct causes with the same\n    // visible outcome, and the consumer had to reconcile after the loop to tell them apart.\n    if (judgment.verdict === \"blocked\") {\n      yield { type: \"status_change\", status: \"blocked\", reason: judgment.reason };\n      return finish(\"blocked\");\n    }\n    if (judgment.verdict === \"skipped\") {\n      yield {\n        type: \"status_change\",\n        status: \"completed\",\n        reason: `skipped: ${judgment.reason}`,\n      };\n      return finish(\"completed\");\n    }\n\n    // M55 — token-budget check AFTER the turn (Codex: gentle wind-down, stops the loop).\n    if (tokenBudget !== undefined && tokensUsed >= tokenBudget) {\n      yield {\n        type: \"status_change\",\n        status: \"budget_limited\",\n        reason: `token budget (${tokenBudget}) reached: ${tokensUsed} used`,\n      };\n      return finish(\"budget_limited\");\n    }\n\n    yield { type: \"continuation\", turn, prompt: continuationPrompt };\n  }\n\n  yield {\n    type: \"status_change\",\n    status: \"failed\",\n    reason: `max turns (${maxTurns}) exhausted`,\n  };\n  return finish(\"failed\");\n}\n\n/**\n * Build the continuation user message for turns 2+. Keeps last response\n * preview short so we don't bloat the prompt cache.\n *\n * @internal\n */\n\nexport function composeContinuation(goal: string, lastResponse: string): string {\n  // M55 — Codex-faithful continuation (ext/goal templates/goals/continuation.md): keep the FULL\n  // objective intact, work from current-state evidence, and audit completion requirement-by-requirement\n  // before declaring done. Improves the quality of what the judge then evaluates.\n  return [\n    GOAL_CONTINUATION_MARKER,\n    \"Continue working toward the active goal.\",\n    \"\",\n    `<objective>\\n${goal}\\n</objective>`,\n    \"\",\n    \"Continuation behavior:\",\n    \"- This goal persists across turns. Keep the full objective intact; if it cannot be finished now,\",\n    \"  make concrete progress toward the real requested end state and do not redefine success around a\",\n    \"  smaller or easier task.\",\n    \"\",\n    \"Work from evidence:\",\n    \"- Use the current worktree and external state as authoritative. Inspect the current state before\",\n    \"  relying on it. Improve, replace, or remove existing work as needed to satisfy the actual objective.\",\n    \"\",\n    \"Completion audit:\",\n    \"- Before deciding the goal is achieved, treat completion as unproven and verify it against the actual\",\n    \"  current state, requirement by requirement. Treat uncertain or indirect evidence as not achieved.\",\n    \"\",\n    `Your last response was:\\n${lastResponse.slice(-1000)}`,\n  ].join(\"\\n\");\n}\n"]}

package/dist/sandbox/bwrap.d.cts

@@ -98,5 +98,15 @@ export declare const realProbes: BwrapProbes;
  * promoted to a positive, which would require the expensive probe back.
  */
 export declare function detectBwrapMemoized(probes?: BwrapProbes): BwrapDetection;
-/** TEST seam — clears the memo. Production never calls it (see m71-cost-per-turn#ADR-1). */
+/**
+ * Clears the `detectBwrapMemoized` memo.
+ *
+ * Kept PUBLIC, unlike the two warn-latch resets that moved to the `__…ForTests` convention, because
+ * this one has a use outside a test that the others do not: a long-lived host that probed before
+ * `bwrap` was installed is memoized on the wrong answer forever, and this is the only way to make it
+ * re-probe. It is the companion to `detectBwrapMemoized` and is documented as such rather than as a
+ * seam. Production inside this package never calls it (see m71-cost-per-turn#ADR-1).
+ *
+ * @public
+ */
 export declare function resetBwrapMemo(): void;

package/dist/sandbox/bwrap.d.ts

@@ -98,5 +98,15 @@ export declare const realProbes: BwrapProbes;
  * promoted to a positive, which would require the expensive probe back.
  */
 export declare function detectBwrapMemoized(probes?: BwrapProbes): BwrapDetection;
-/** TEST seam — clears the memo. Production never calls it (see m71-cost-per-turn#ADR-1). */
+/**
+ * Clears the `detectBwrapMemoized` memo.
+ *
+ * Kept PUBLIC, unlike the two warn-latch resets that moved to the `__…ForTests` convention, because
+ * this one has a use outside a test that the others do not: a long-lived host that probed before
+ * `bwrap` was installed is memoized on the wrong answer forever, and this is the only way to make it
+ * re-probe. It is the companion to `detectBwrapMemoized` and is documented as such rather than as a
+ * seam. Production inside this package never calls it (see m71-cost-per-turn#ADR-1).
+ *
+ * @public
+ */
 export declare function resetBwrapMemo(): void;

package/dist/sandbox/index.cjs

@@ -1,10 +1,9 @@
 'use strict';
 
-var chunkJ24VJOH3_cjs = require('../chunk-J24VJOH3.cjs');
+var chunkUADQJBHR_cjs = require('../chunk-UADQJBHR.cjs');
 require('../chunk-2ZPEDVLM.cjs');
-var chunk4ZHTIJNE_cjs = require('../chunk-4ZHTIJNE.cjs');
-var chunkXL7Q7HB2_cjs = require('../chunk-XL7Q7HB2.cjs');
-require('../chunk-EG4QM7ZJ.cjs');
+var chunkJ7J7J2GN_cjs = require('../chunk-J7J7J2GN.cjs');
+var chunk6LHQPOMI_cjs = require('../chunk-6LHQPOMI.cjs');
 var child_process = require('child_process');
 var fs = require('fs');
 var path = require('path');
@@ -210,16 +209,13 @@ function buildSeccompFilter(opts) {
 }
 
 // src/sandbox/linux-sandbox.ts
-function shellQuote(s) {
-  return `'${s.replaceAll("'", `'\\''`)}'`;
-}
 function wrapCommandForSandbox(mode, opts, command) {
   const argv = buildBwrapArgv(mode, { cwd: opts.cwd, network: opts.network, env: opts.env });
   if (argv === null) return null;
   const bin = opts.bin ?? "bwrap";
   const seccompArgv = opts.seccompPath !== void 0 ? ["--seccomp", "3"] : [];
-  const base = `${shellQuote(bin)} ${[...argv.slice(0, -1), ...seccompArgv, "--"].map(shellQuote).join(" ")} /bin/sh -c ${shellQuote(command)}`;
-  return opts.seccompPath !== void 0 ? `${base} 3< ${shellQuote(opts.seccompPath)}` : base;
+  const base = `${chunkUADQJBHR_cjs.shellEscapePosix(bin)} ${[...argv.slice(0, -1), ...seccompArgv, "--"].map(chunkUADQJBHR_cjs.shellEscapePosix).join(" ")} /bin/sh -c ${chunkUADQJBHR_cjs.shellEscapePosix(command)}`;
+  return opts.seccompPath !== void 0 ? `${base} 3< ${chunkUADQJBHR_cjs.shellEscapePosix(opts.seccompPath)}` : base;
 }
 var ENV_ALLOWLIST = [
   "PATH",
@@ -240,7 +236,7 @@ function allowlistedEnv(source = process.env) {
   }
   return out;
 }
-var LinuxSandbox = class extends chunkJ24VJOH3_cjs.LocalSandbox {
+var LinuxSandbox = class extends chunkUADQJBHR_cjs.LocalSandbox {
   mode;
   network;
   cwd;
@@ -290,7 +286,7 @@ var LinuxSandbox = class extends chunkJ24VJOH3_cjs.LocalSandbox {
     const target = path2.startsWith("/") ? path2 : `${this.cwd}/${path2}`;
     const b64 = Buffer.from(content).toString("base64");
     const r = await this.execute(
-      `mkdir -p ${shellQuote(path.dirname(target))} && printf %s ${shellQuote(b64)} | base64 -d > ${shellQuote(target)}`
+      `mkdir -p ${chunkUADQJBHR_cjs.shellEscapePosix(path.dirname(target))} && printf %s ${chunkUADQJBHR_cjs.shellEscapePosix(b64)} | base64 -d > ${chunkUADQJBHR_cjs.shellEscapePosix(target)}`
     );
     if (r.exitCode !== 0) {
       throw new Error(
@@ -334,12 +330,12 @@ function seccompPathForArch(arch, warn) {
 var seccompFilterPath;
 function restrictedSeccompPath() {
   if (seccompFilterPath !== void 0) return seccompFilterPath ?? void 0;
-  const path2 = seccompPathForArch(process.arch, (m) => console.warn(chunkXL7Q7HB2_cjs.redactSecrets(m)));
+  const path2 = seccompPathForArch(process.arch, (m) => console.warn(chunk6LHQPOMI_cjs.redactSecrets(m)));
   seccompFilterPath = path2 ?? null;
   return path2;
 }
 var warnedUnavailable = false;
-function resetSandboxWarnLatch() {
+function __resetSandboxWarnLatchForTests() {
   warnedUnavailable = false;
 }
 function resolveSandboxPosture(opts) {
@@ -354,22 +350,22 @@ function resolveSandboxPosture(opts) {
 }
 function createSandboxBackend(opts) {
   const config = { workDir: opts.workDir, timeoutMs: opts.timeoutMs };
-  if (opts.mode === "danger-full-access") return new chunkJ24VJOH3_cjs.LocalSandbox(config);
+  if (opts.mode === "danger-full-access") return new chunkUADQJBHR_cjs.LocalSandbox(config);
   const detection = (opts.detect ?? detectBwrapMemoized)();
   if (!detection.ok) {
     if (!warnedUnavailable) {
       warnedUnavailable = true;
-      const warn = opts.warn ?? ((m) => console.warn(chunkXL7Q7HB2_cjs.redactSecrets(m)));
+      const warn = opts.warn ?? ((m) => console.warn(chunk6LHQPOMI_cjs.redactSecrets(m)));
       warn(
         `[sandbox] OS-level enforcement unavailable (${detection.reason}) \u2014 falling back to tool-level gating only (sandbox_mode=${opts.mode}).`
       );
     }
-    return new chunkJ24VJOH3_cjs.LocalSandbox(config);
+    return new chunkUADQJBHR_cjs.LocalSandbox(config);
   }
   return new LinuxSandbox(config, { mode: opts.mode, network: opts.network, bin: detection.bin });
 }
 var interactiveWarned = false;
-function resetInteractiveWarnLatch() {
+function __resetInteractiveWarnLatchForTests() {
   interactiveWarned = false;
 }
 function seccompPathForNetwork(networkOpen) {
@@ -381,7 +377,7 @@ function decideConfinement(opts) {
   if (detection.ok) return detection.bin;
   if (!interactiveWarned) {
     interactiveWarned = true;
-    const warn = opts.warn ?? ((m) => console.warn(chunkXL7Q7HB2_cjs.redactSecrets(m)));
+    const warn = opts.warn ?? ((m) => console.warn(chunk6LHQPOMI_cjs.redactSecrets(m)));
     warn(
       `[sandbox] OS-level enforcement unavailable (${detection.reason}) \u2014 interactive session runs WITHOUT kernel confinement (sandbox_mode=${opts.mode}).`
     );
@@ -407,7 +403,7 @@ function interactiveWrapCommand(opts) {
 }
 
 // src/sandbox/provision.ts
-var RepoProvisionError = class extends chunk4ZHTIJNE_cjs.TheokitAgentError {
+var RepoProvisionError = class extends chunkJ7J7J2GN_cjs.TheokitAgentError {
   constructor(instanceId, message, options = {}) {
     super(`[${instanceId}] ${message}`, {
       code: "repo_provision_failed",
@@ -421,7 +417,7 @@ var RepoProvisionError = class extends chunk4ZHTIJNE_cjs.TheokitAgentError {
 };
 var SAFE_INSTANCE_ID = /^[A-Za-z0-9][A-Za-z0-9._-]*$/;
 async function provisionRepo(sandboxOrOpts, maybeOpts) {
-  const sandbox = maybeOpts !== void 0 ? sandboxOrOpts : new chunkJ24VJOH3_cjs.LocalSandbox();
+  const sandbox = maybeOpts !== void 0 ? sandboxOrOpts : new chunkUADQJBHR_cjs.LocalSandbox();
   const opts = maybeOpts ?? sandboxOrOpts;
   const { repoUrl, ref, instanceId } = opts;
   if (!SAFE_INSTANCE_ID.test(instanceId)) {
@@ -434,20 +430,20 @@ async function provisionRepo(sandboxOrOpts, maybeOpts) {
     throw new RepoProvisionError(instanceId, `invalid ref: must not begin with '-' (got ${ref})`);
   }
   const clone = await sandbox.execute(
-    `git -c protocol.ext.allow=never clone --quiet -- ${chunkJ24VJOH3_cjs.shellEscapePosix(repoUrl)} ${chunkJ24VJOH3_cjs.shellEscapePosix(instanceId)}`
+    `git -c protocol.ext.allow=never clone --quiet -- ${chunkUADQJBHR_cjs.shellEscapePosix(repoUrl)} ${chunkUADQJBHR_cjs.shellEscapePosix(instanceId)}`
   );
   if (clone.exitCode !== 0) {
     throw new RepoProvisionError(instanceId, `clone failed: ${clone.stderr.trim()}`);
   }
   const top = await sandbox.execute(
-    `git -C ${chunkJ24VJOH3_cjs.shellEscapePosix(instanceId)} rev-parse --show-toplevel`
+    `git -C ${chunkUADQJBHR_cjs.shellEscapePosix(instanceId)} rev-parse --show-toplevel`
   );
   if (top.exitCode !== 0) {
     throw new RepoProvisionError(instanceId, `resolve repoDir failed: ${top.stderr.trim()}`);
   }
   const repoDir = top.stdout.trim();
   const checkout = await sandbox.execute(
-    `git -C ${chunkJ24VJOH3_cjs.shellEscapePosix(repoDir)} checkout --quiet ${chunkJ24VJOH3_cjs.shellEscapePosix(ref)}`
+    `git -C ${chunkUADQJBHR_cjs.shellEscapePosix(repoDir)} checkout --quiet ${chunkUADQJBHR_cjs.shellEscapePosix(ref)}`
   );
   if (checkout.exitCode !== 0) {
     throw new RepoProvisionError(instanceId, `checkout ${ref} failed: ${checkout.stderr.trim()}`);
@@ -457,23 +453,23 @@ async function provisionRepo(sandboxOrOpts, maybeOpts) {
 
 Object.defineProperty(exports, "LocalSandbox", {
   enumerable: true,
-  get: function () { return chunkJ24VJOH3_cjs.LocalSandbox; }
+  get: function () { return chunkUADQJBHR_cjs.LocalSandbox; }
 });
 Object.defineProperty(exports, "SandboxBackend", {
   enumerable: true,
-  get: function () { return chunkJ24VJOH3_cjs.SandboxBackend; }
+  get: function () { return chunkUADQJBHR_cjs.SandboxBackend; }
 });
 Object.defineProperty(exports, "SandboxNotAvailableError", {
   enumerable: true,
-  get: function () { return chunkJ24VJOH3_cjs.SandboxNotAvailableError; }
+  get: function () { return chunkUADQJBHR_cjs.SandboxNotAvailableError; }
 });
 Object.defineProperty(exports, "SandboxSecurityError", {
   enumerable: true,
-  get: function () { return chunkJ24VJOH3_cjs.SandboxSecurityError; }
+  get: function () { return chunkUADQJBHR_cjs.SandboxSecurityError; }
 });
 Object.defineProperty(exports, "resolveSandbox", {
   enumerable: true,
-  get: function () { return chunkJ24VJOH3_cjs.resolveSandbox; }
+  get: function () { return chunkUADQJBHR_cjs.resolveSandbox; }
 });
 exports.LinuxSandbox = LinuxSandbox;
 exports.RepoProvisionError = RepoProvisionError;
@@ -488,8 +484,8 @@ exports.provisionRepo = provisionRepo;
 exports.realProbeCount = realProbeCount;
 exports.realProbes = realProbes;
 exports.resetBwrapMemo = resetBwrapMemo;
-exports.resetInteractiveWarnLatch = resetInteractiveWarnLatch;
-exports.resetSandboxWarnLatch = resetSandboxWarnLatch;
+exports.resetInteractiveWarnLatch = __resetInteractiveWarnLatchForTests;
+exports.resetSandboxWarnLatch = __resetSandboxWarnLatchForTests;
 exports.resolveSandboxPosture = resolveSandboxPosture;
 exports.restrictedSeccompPath = restrictedSeccompPath;
 exports.seccompPathForArch = seccompPathForArch;