@theokit/sdk 5.11.1 → 5.12.0

package/dist/providers.js

@@ -1,11 +1,11 @@
-export { getProviderProfile, listProviders } from './chunk-I2ORT6QD.js';
-import './chunk-SVCPJRUB.js';
+export { getProviderProfile, listProviders } from './chunk-NKQVLWDF.js';
+import './chunk-5HSMH2ZZ.js';
 import './chunk-AGACYRJ2.js';
-import './chunk-TX56N7FR.js';
-import './chunk-5Q4RTLMP.js';
-import './chunk-H3CVJSFC.js';
+import './chunk-3FSBE4PS.js';
+import './chunk-MS3TQSXU.js';
+import './chunk-5IUDICRK.js';
 import './chunk-Q7GYMXVC.js';
-import './chunk-B5MHRQJZ.js';
-import './chunk-CZJ6Q7CW.js';
+import './chunk-TNFOU6SB.js';
+import './chunk-3OY6H3I7.js';
 //# sourceMappingURL=providers.js.map
 //# sourceMappingURL=providers.js.map

package/dist/registry-5TDAF7RK.js

@@ -0,0 +1,9 @@
+export { __getSubscribersCountForTests, __resetTaskRegistryForTests, cancel, configure, get, list, submit, subscribe } from './chunk-HWMGR3HA.js';
+import './chunk-K2BQQ445.js';
+import './chunk-AHKS3DKI.js';
+import './chunk-MLGETTYX.js';
+import './chunk-M4U24U6G.js';
+import './chunk-TNFOU6SB.js';
+import './chunk-3OY6H3I7.js';
+//# sourceMappingURL=registry-5TDAF7RK.js.map
+//# sourceMappingURL=registry-5TDAF7RK.js.map

package/dist/{registry-5FK42CRV.js.map → registry-5TDAF7RK.js.map}

@@ -1 +1 @@
-{"version":3,"sources":[],"names":[],"mappings":"","file":"registry-5FK42CRV.js"}
+{"version":3,"sources":[],"names":[],"mappings":"","file":"registry-5TDAF7RK.js"}

package/dist/registry-W7TSCLQO.cjs

@@ -0,0 +1,46 @@
+'use strict';
+
+var chunkL7JPXZ2C_cjs = require('./chunk-L7JPXZ2C.cjs');
+require('./chunk-BUUUWQMB.cjs');
+require('./chunk-FLDMPVUC.cjs');
+require('./chunk-BWYC7RZG.cjs');
+require('./chunk-ZOFRNNZD.cjs');
+require('./chunk-CEGU5AVD.cjs');
+require('./chunk-RNXXYF22.cjs');
+
+
+
+Object.defineProperty(exports, "__getSubscribersCountForTests", {
+  enumerable: true,
+  get: function () { return chunkL7JPXZ2C_cjs.__getSubscribersCountForTests; }
+});
+Object.defineProperty(exports, "__resetTaskRegistryForTests", {
+  enumerable: true,
+  get: function () { return chunkL7JPXZ2C_cjs.__resetTaskRegistryForTests; }
+});
+Object.defineProperty(exports, "cancel", {
+  enumerable: true,
+  get: function () { return chunkL7JPXZ2C_cjs.cancel; }
+});
+Object.defineProperty(exports, "configure", {
+  enumerable: true,
+  get: function () { return chunkL7JPXZ2C_cjs.configure; }
+});
+Object.defineProperty(exports, "get", {
+  enumerable: true,
+  get: function () { return chunkL7JPXZ2C_cjs.get; }
+});
+Object.defineProperty(exports, "list", {
+  enumerable: true,
+  get: function () { return chunkL7JPXZ2C_cjs.list; }
+});
+Object.defineProperty(exports, "submit", {
+  enumerable: true,
+  get: function () { return chunkL7JPXZ2C_cjs.submit; }
+});
+Object.defineProperty(exports, "subscribe", {
+  enumerable: true,
+  get: function () { return chunkL7JPXZ2C_cjs.subscribe; }
+});
+//# sourceMappingURL=registry-W7TSCLQO.cjs.map
+//# sourceMappingURL=registry-W7TSCLQO.cjs.map

package/dist/{registry-BVTNO73E.cjs.map → registry-W7TSCLQO.cjs.map}

@@ -1 +1 @@
-{"version":3,"sources":[],"names":[],"mappings":"","file":"registry-BVTNO73E.cjs"}
+{"version":3,"sources":[],"names":[],"mappings":"","file":"registry-W7TSCLQO.cjs"}

package/dist/retry.cjs

@@ -1,14 +1,14 @@
 'use strict';
 
-var chunk72CPLZPC_cjs = require('./chunk-72CPLZPC.cjs');
-require('./chunk-UNFBHV46.cjs');
-require('./chunk-6LHQPOMI.cjs');
+var chunk7INPXUJI_cjs = require('./chunk-7INPXUJI.cjs');
+require('./chunk-CEGU5AVD.cjs');
+require('./chunk-RNXXYF22.cjs');
 
 
 
 Object.defineProperty(exports, "Retry", {
   enumerable: true,
-  get: function () { return chunk72CPLZPC_cjs.Retry; }
+  get: function () { return chunk7INPXUJI_cjs.Retry; }
 });
 //# sourceMappingURL=retry.cjs.map
 //# sourceMappingURL=retry.cjs.map

package/dist/retry.js

@@ -1,5 +1,5 @@
-export { Retry } from './chunk-7F7XE6CM.js';
-import './chunk-B5MHRQJZ.js';
-import './chunk-CZJ6Q7CW.js';
+export { Retry } from './chunk-2S4HQCDF.js';
+import './chunk-TNFOU6SB.js';
+import './chunk-3OY6H3I7.js';
 //# sourceMappingURL=retry.js.map
 //# sourceMappingURL=retry.js.map

package/dist/{run-until-DG2KXWB2.cjs → run-until-4KHVF7YC.cjs}

@@ -46,8 +46,12 @@ async function* runUntilImpl(agent, goal, options, deps) {
       if (signal.aborted) cancelOnAbort();
       else signal.addEventListener("abort", cancelOnAbort, { once: true });
     }
-    const result = await run.wait();
-    if (signal !== void 0) signal.removeEventListener("abort", cancelOnAbort);
+    let result;
+    try {
+      result = await run.wait();
+    } finally {
+      if (signal !== void 0) signal.removeEventListener("abort", cancelOnAbort);
+    }
     lastResponse = result.result ?? "";
     const turnTokens = result.usage?.totalTokens;
     if (typeof turnTokens === "number") tokensUsed += turnTokens;
@@ -145,5 +149,5 @@ ${lastResponse.slice(-1e3)}`
 
 exports.composeContinuation = composeContinuation;
 exports.runUntilImpl = runUntilImpl;
-//# sourceMappingURL=run-until-DG2KXWB2.cjs.map
-//# sourceMappingURL=run-until-DG2KXWB2.cjs.map
+//# sourceMappingURL=run-until-4KHVF7YC.cjs.map
+//# sourceMappingURL=run-until-4KHVF7YC.cjs.map

package/dist/run-until-4KHVF7YC.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;AAGA,IAAA,IAAI,MAAA;AACJ,IAAA,IAAI;AACF,MAAA,MAAA,GAAS,MAAM,IAAI,IAAA,EAAK;AAAA,IAC1B,CAAA,SAAE;AACA,MAAA,IAAI,MAAA,KAAW,MAAA,EAAW,MAAA,CAAO,mBAAA,CAAoB,SAAS,aAAa,CAAA;AAAA,IAC7E;AACA,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-4KHVF7YC.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    // B-498 — in a `finally`: under `throwOnError` a failed turn's `wait()` rejects, and the\n    // listener must not outlive the turn on the caller's signal.\n    let result: Awaited<ReturnType<typeof run.wait>>;\n    try {\n      result = await run.wait();\n    } finally {\n      if (signal !== undefined) signal.removeEventListener(\"abort\", cancelOnAbort);\n    }\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/{run-until-6D6TQ4JN.js → run-until-G6BW2DWM.js}

@@ -44,8 +44,12 @@ async function* runUntilImpl(agent, goal, options, deps) {
       if (signal.aborted) cancelOnAbort();
       else signal.addEventListener("abort", cancelOnAbort, { once: true });
     }
-    const result = await run.wait();
-    if (signal !== void 0) signal.removeEventListener("abort", cancelOnAbort);
+    let result;
+    try {
+      result = await run.wait();
+    } finally {
+      if (signal !== void 0) signal.removeEventListener("abort", cancelOnAbort);
+    }
     lastResponse = result.result ?? "";
     const turnTokens = result.usage?.totalTokens;
     if (typeof turnTokens === "number") tokensUsed += turnTokens;
@@ -142,5 +146,5 @@ ${lastResponse.slice(-1e3)}`
 }
 
 export { composeContinuation, runUntilImpl };
-//# sourceMappingURL=run-until-6D6TQ4JN.js.map
-//# sourceMappingURL=run-until-6D6TQ4JN.js.map
+//# sourceMappingURL=run-until-G6BW2DWM.js.map
+//# sourceMappingURL=run-until-G6BW2DWM.js.map

package/dist/run-until-G6BW2DWM.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/internal/runtime/lifecycle/run-until.ts"],"names":[],"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;AAGA,IAAA,IAAI,MAAA;AACJ,IAAA,IAAI;AACF,MAAA,MAAA,GAAS,MAAM,IAAI,IAAA,EAAK;AAAA,IAC1B,CAAA,SAAE;AACA,MAAA,IAAI,MAAA,KAAW,MAAA,EAAW,MAAA,CAAO,mBAAA,CAAoB,SAAS,aAAa,CAAA;AAAA,IAC7E;AACA,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,IACL,wBAAA;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-G6BW2DWM.js","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    // B-498 — in a `finally`: under `throwOnError` a failed turn's `wait()` rejects, and the\n    // listener must not outlive the turn on the caller's signal.\n    let result: Awaited<ReturnType<typeof run.wait>>;\n    try {\n      result = await run.wait();\n    } finally {\n      if (signal !== undefined) signal.removeEventListener(\"abort\", cancelOnAbort);\n    }\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/index.cjs

@@ -1,9 +1,9 @@
 'use strict';
 
-var chunk7NQ7HJ3E_cjs = require('../chunk-7NQ7HJ3E.cjs');
+var chunkBRWI4HLK_cjs = require('../chunk-BRWI4HLK.cjs');
 require('../chunk-2ZPEDVLM.cjs');
-var chunkUNFBHV46_cjs = require('../chunk-UNFBHV46.cjs');
-var chunk6LHQPOMI_cjs = require('../chunk-6LHQPOMI.cjs');
+var chunkCEGU5AVD_cjs = require('../chunk-CEGU5AVD.cjs');
+var chunkRNXXYF22_cjs = require('../chunk-RNXXYF22.cjs');
 var child_process = require('child_process');
 var fs = require('fs');
 var path = require('path');
@@ -214,8 +214,8 @@ function wrapCommandForSandbox(mode, opts, command) {
   if (argv === null) return null;
   const bin = opts.bin ?? "bwrap";
   const seccompArgv = opts.seccompPath !== void 0 ? ["--seccomp", "3"] : [];
-  const base = `${chunk7NQ7HJ3E_cjs.shellEscapePosix(bin)} ${[...argv.slice(0, -1), ...seccompArgv, "--"].map(chunk7NQ7HJ3E_cjs.shellEscapePosix).join(" ")} /bin/sh -c ${chunk7NQ7HJ3E_cjs.shellEscapePosix(command)}`;
-  return opts.seccompPath !== void 0 ? `${base} 3< ${chunk7NQ7HJ3E_cjs.shellEscapePosix(opts.seccompPath)}` : base;
+  const base = `${chunkBRWI4HLK_cjs.shellEscapePosix(bin)} ${[...argv.slice(0, -1), ...seccompArgv, "--"].map(chunkBRWI4HLK_cjs.shellEscapePosix).join(" ")} /bin/sh -c ${chunkBRWI4HLK_cjs.shellEscapePosix(command)}`;
+  return opts.seccompPath !== void 0 ? `${base} 3< ${chunkBRWI4HLK_cjs.shellEscapePosix(opts.seccompPath)}` : base;
 }
 var ENV_ALLOWLIST = [
   "PATH",
@@ -236,7 +236,7 @@ function allowlistedEnv(source = process.env) {
   }
   return out;
 }
-var LinuxSandbox = class extends chunk7NQ7HJ3E_cjs.LocalSandbox {
+var LinuxSandbox = class extends chunkBRWI4HLK_cjs.LocalSandbox {
   mode;
   network;
   cwd;
@@ -286,7 +286,7 @@ var LinuxSandbox = class extends chunk7NQ7HJ3E_cjs.LocalSandbox {
     const target = path2.startsWith("/") ? path2 : `${this.cwd}/${path2}`;
     const b64 = Buffer.from(content).toString("base64");
     const r = await this.execute(
-      `mkdir -p ${chunk7NQ7HJ3E_cjs.shellEscapePosix(path.dirname(target))} && printf %s ${chunk7NQ7HJ3E_cjs.shellEscapePosix(b64)} | base64 -d > ${chunk7NQ7HJ3E_cjs.shellEscapePosix(target)}`
+      `mkdir -p ${chunkBRWI4HLK_cjs.shellEscapePosix(path.dirname(target))} && printf %s ${chunkBRWI4HLK_cjs.shellEscapePosix(b64)} | base64 -d > ${chunkBRWI4HLK_cjs.shellEscapePosix(target)}`
     );
     if (r.exitCode !== 0) {
       throw new Error(
@@ -330,7 +330,7 @@ function seccompPathForArch(arch, warn) {
 var seccompFilterPath;
 function restrictedSeccompPath() {
   if (seccompFilterPath !== void 0) return seccompFilterPath ?? void 0;
-  const path2 = seccompPathForArch(process.arch, (m) => console.warn(chunk6LHQPOMI_cjs.redactSecrets(m)));
+  const path2 = seccompPathForArch(process.arch, (m) => console.warn(chunkRNXXYF22_cjs.redactSecrets(m)));
   seccompFilterPath = path2 ?? null;
   return path2;
 }
@@ -350,17 +350,17 @@ function resolveSandboxPosture(opts) {
 }
 function createSandboxBackend(opts) {
   const config = { workDir: opts.workDir, timeoutMs: opts.timeoutMs };
-  if (opts.mode === "danger-full-access") return new chunk7NQ7HJ3E_cjs.LocalSandbox(config);
+  if (opts.mode === "danger-full-access") return new chunkBRWI4HLK_cjs.LocalSandbox(config);
   const detection = (opts.detect ?? detectBwrapMemoized)();
   if (!detection.ok) {
     if (!warnedUnavailable) {
       warnedUnavailable = true;
-      const warn = opts.warn ?? ((m) => console.warn(chunk6LHQPOMI_cjs.redactSecrets(m)));
+      const warn = opts.warn ?? ((m) => console.warn(chunkRNXXYF22_cjs.redactSecrets(m)));
       warn(
         `[sandbox] OS-level enforcement unavailable (${detection.reason}) \u2014 falling back to tool-level gating only (sandbox_mode=${opts.mode}).`
       );
     }
-    return new chunk7NQ7HJ3E_cjs.LocalSandbox(config);
+    return new chunkBRWI4HLK_cjs.LocalSandbox(config);
   }
   return new LinuxSandbox(config, { mode: opts.mode, network: opts.network, bin: detection.bin });
 }
@@ -377,7 +377,7 @@ function decideConfinement(opts) {
   if (detection.ok) return detection.bin;
   if (!interactiveWarned) {
     interactiveWarned = true;
-    const warn = opts.warn ?? ((m) => console.warn(chunk6LHQPOMI_cjs.redactSecrets(m)));
+    const warn = opts.warn ?? ((m) => console.warn(chunkRNXXYF22_cjs.redactSecrets(m)));
     warn(
       `[sandbox] OS-level enforcement unavailable (${detection.reason}) \u2014 interactive session runs WITHOUT kernel confinement (sandbox_mode=${opts.mode}).`
     );
@@ -403,7 +403,7 @@ function interactiveWrapCommand(opts) {
 }
 
 // src/sandbox/provision.ts
-var RepoProvisionError = class extends chunkUNFBHV46_cjs.TheokitAgentError {
+var RepoProvisionError = class extends chunkCEGU5AVD_cjs.TheokitAgentError {
   constructor(instanceId, message, options = {}) {
     super(`[${instanceId}] ${message}`, {
       code: "repo_provision_failed",
@@ -417,7 +417,7 @@ var RepoProvisionError = class extends chunkUNFBHV46_cjs.TheokitAgentError {
 };
 var SAFE_INSTANCE_ID = /^[A-Za-z0-9][A-Za-z0-9._-]*$/;
 async function provisionRepo(sandboxOrOpts, maybeOpts) {
-  const sandbox = maybeOpts !== void 0 ? sandboxOrOpts : new chunk7NQ7HJ3E_cjs.LocalSandbox();
+  const sandbox = maybeOpts !== void 0 ? sandboxOrOpts : new chunkBRWI4HLK_cjs.LocalSandbox();
   const opts = maybeOpts ?? sandboxOrOpts;
   const { repoUrl, ref, instanceId } = opts;
   if (!SAFE_INSTANCE_ID.test(instanceId)) {
@@ -430,20 +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 -- ${chunk7NQ7HJ3E_cjs.shellEscapePosix(repoUrl)} ${chunk7NQ7HJ3E_cjs.shellEscapePosix(instanceId)}`
+    `git -c protocol.ext.allow=never clone --quiet -- ${chunkBRWI4HLK_cjs.shellEscapePosix(repoUrl)} ${chunkBRWI4HLK_cjs.shellEscapePosix(instanceId)}`
   );
   if (clone.exitCode !== 0) {
     throw new RepoProvisionError(instanceId, `clone failed: ${clone.stderr.trim()}`);
   }
   const top = await sandbox.execute(
-    `git -C ${chunk7NQ7HJ3E_cjs.shellEscapePosix(instanceId)} rev-parse --show-toplevel`
+    `git -C ${chunkBRWI4HLK_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 ${chunk7NQ7HJ3E_cjs.shellEscapePosix(repoDir)} checkout --quiet ${chunk7NQ7HJ3E_cjs.shellEscapePosix(ref)}`
+    `git -C ${chunkBRWI4HLK_cjs.shellEscapePosix(repoDir)} checkout --quiet ${chunkBRWI4HLK_cjs.shellEscapePosix(ref)}`
   );
   if (checkout.exitCode !== 0) {
     throw new RepoProvisionError(instanceId, `checkout ${ref} failed: ${checkout.stderr.trim()}`);
@@ -453,23 +453,23 @@ async function provisionRepo(sandboxOrOpts, maybeOpts) {
 
 Object.defineProperty(exports, "LocalSandbox", {
   enumerable: true,
-  get: function () { return chunk7NQ7HJ3E_cjs.LocalSandbox; }
+  get: function () { return chunkBRWI4HLK_cjs.LocalSandbox; }
 });
 Object.defineProperty(exports, "SandboxBackend", {
   enumerable: true,
-  get: function () { return chunk7NQ7HJ3E_cjs.SandboxBackend; }
+  get: function () { return chunkBRWI4HLK_cjs.SandboxBackend; }
 });
 Object.defineProperty(exports, "SandboxNotAvailableError", {
   enumerable: true,
-  get: function () { return chunk7NQ7HJ3E_cjs.SandboxNotAvailableError; }
+  get: function () { return chunkBRWI4HLK_cjs.SandboxNotAvailableError; }
 });
 Object.defineProperty(exports, "SandboxSecurityError", {
   enumerable: true,
-  get: function () { return chunk7NQ7HJ3E_cjs.SandboxSecurityError; }
+  get: function () { return chunkBRWI4HLK_cjs.SandboxSecurityError; }
 });
 Object.defineProperty(exports, "resolveSandbox", {
   enumerable: true,
-  get: function () { return chunk7NQ7HJ3E_cjs.resolveSandbox; }
+  get: function () { return chunkBRWI4HLK_cjs.resolveSandbox; }
 });
 exports.LinuxSandbox = LinuxSandbox;
 exports.RepoProvisionError = RepoProvisionError;

package/dist/sandbox/index.js

@@ -1,8 +1,8 @@
-import { shellEscapePosix, LocalSandbox } from '../chunk-GMSRRJ53.js';
-export { LocalSandbox, SandboxBackend, SandboxNotAvailableError, SandboxSecurityError, resolveSandbox } from '../chunk-GMSRRJ53.js';
+import { shellEscapePosix, LocalSandbox } from '../chunk-EEK7S77K.js';
+export { LocalSandbox, SandboxBackend, SandboxNotAvailableError, SandboxSecurityError, resolveSandbox } from '../chunk-EEK7S77K.js';
 import '../chunk-HKCKOAIO.js';
-import { TheokitAgentError } from '../chunk-B5MHRQJZ.js';
-import { redactSecrets } from '../chunk-CZJ6Q7CW.js';
+import { TheokitAgentError } from '../chunk-TNFOU6SB.js';
+import { redactSecrets } from '../chunk-3OY6H3I7.js';
 import { execFileSync } from 'child_process';
 import { existsSync, mkdtempSync, writeFileSync, rmSync } from 'fs';
 import path, { dirname, join } from 'path';