@theokit/sdk 4.17.0 → 4.18.0

package/CHANGELOG.md

@@ -1,5 +1,11 @@
 # Changelog
 
+## 4.17.1
+
+### Patch Changes
+
+- fix(session): hydration REPLACES the cache from disk (source of truth) instead of skipping when non-empty — after an invalidation (compact/inject), an in-flight turn repopulating the cache with one message used to pin the parent to a 1-message context until restart (M51 review F4; race test added).
+
 ## 4.17.0
 
 ### Minor Changes

package/dist/{cron-B_vVYc-_.d.ts → cron-DOW3f-Xu.d.ts}

@@ -598,7 +598,7 @@ type GoalEvent = {
     prompt: string;
 } | {
     type: "status_change";
-    status: "active" | "paused" | "completed" | "failed";
+    status: "active" | "paused" | "completed" | "failed" | "budget_limited" | "blocked";
     reason: string;
 };
 /**
@@ -608,8 +608,10 @@ type GoalEvent = {
  * @public
  */
 interface GoalResult {
-    status: "completed" | "failed" | "paused";
+    status: "completed" | "failed" | "paused" | "budget_limited" | "blocked";
     turnsUsed: number;
+    /** M55 — tokens somados ao longo do loop (0 quando `usage` esteve ausente — fail-open). */
+    tokensUsed: number;
     finalResponse: string | undefined;
 }
 /**
@@ -625,8 +627,14 @@ type RunUntilIterator = AsyncGenerator<GoalEvent, GoalResult, void>;
  * @public
  */
 interface GoalOptions {
-    /** Hard cap on iterations. Default `20`. */
+    /** Hard cap on iterations (safety net against runaway). Default `20`. */
     maxTurns?: number;
+    /**
+     * M55 — token budget (Codex ext/goal parity, tool.rs:454-465). Soma `run.wait().usage.totalTokens`
+     * por turno; ao cruzar, o loop para com status `budget_limited`. Omitido ⇒ ilimitado (só maxTurns).
+     * `usage` ausente nunca estoura o budget (fail-open).
+     */
+    tokenBudget?: number;
     /** Bail after N consecutive judge parse failures. Default `3` (ADR D121). */
     maxConsecutiveJudgeFailures?: number;
     /** Judge model identifier. Default `"openai/gpt-4o-mini"` (ADR D119). */

package/dist/{cron-Dj8FAU5V.d.cts → cron-DvP-E6Xx.d.cts}

@@ -598,7 +598,7 @@ type GoalEvent = {
     prompt: string;
 } | {
     type: "status_change";
-    status: "active" | "paused" | "completed" | "failed";
+    status: "active" | "paused" | "completed" | "failed" | "budget_limited" | "blocked";
     reason: string;
 };
 /**
@@ -608,8 +608,10 @@ type GoalEvent = {
  * @public
  */
 interface GoalResult {
-    status: "completed" | "failed" | "paused";
+    status: "completed" | "failed" | "paused" | "budget_limited" | "blocked";
     turnsUsed: number;
+    /** M55 — tokens somados ao longo do loop (0 quando `usage` esteve ausente — fail-open). */
+    tokensUsed: number;
     finalResponse: string | undefined;
 }
 /**
@@ -625,8 +627,14 @@ type RunUntilIterator = AsyncGenerator<GoalEvent, GoalResult, void>;
  * @public
  */
 interface GoalOptions {
-    /** Hard cap on iterations. Default `20`. */
+    /** Hard cap on iterations (safety net against runaway). Default `20`. */
     maxTurns?: number;
+    /**
+     * M55 — token budget (Codex ext/goal parity, tool.rs:454-465). Soma `run.wait().usage.totalTokens`
+     * por turno; ao cruzar, o loop para com status `budget_limited`. Omitido ⇒ ilimitado (só maxTurns).
+     * `usage` ausente nunca estoura o budget (fail-open).
+     */
+    tokenBudget?: number;
     /** Bail after N consecutive judge parse failures. Default `3` (ADR D121). */
     maxConsecutiveJudgeFailures?: number;
     /** Judge model identifier. Default `"openai/gpt-4o-mini"` (ADR D119). */

package/dist/cron.cjs

@@ -5948,9 +5948,7 @@ async function hydrateSession(agentId, loc) {
   hydratedKeys.add(key);
   const persisted = await readSessionMessages(loc.store, agentId);
   if (persisted.length === 0) return;
-  if (!sessions.has(agentId) || sessions.get(agentId)?.length === 0) {
-    sessions.set(agentId, persisted);
-  }
+  sessions.set(agentId, persisted);
 }
 async function flushSessionWrites() {
   while (pendingWrites.size > 0) {
@@ -6258,10 +6256,12 @@ __export(run_until_exports, {
 async function* runUntilImpl(agent, goal, options, deps) {
   const maxTurns = options?.maxTurns ?? 20;
   const maxFails = options?.maxConsecutiveJudgeFailures ?? 3;
+  const tokenBudget = options?.tokenBudget;
   const signal = options?.signal;
   const isAborted = () => signal !== void 0 && signal.aborted;
   let turn = 0;
   let consecutiveFailures = 0;
+  let tokensUsed = 0;
   let lastResponse = "";
   if (isAborted()) {
     yield {
@@ -6269,13 +6269,13 @@ async function* runUntilImpl(agent, goal, options, deps) {
       status: "paused",
       reason: "aborted via AbortSignal before first turn"
     };
-    return { status: "paused", turnsUsed: 0, finalResponse: void 0 };
+    return { status: "paused", turnsUsed: 0, tokensUsed, finalResponse: void 0 };
   }
   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, finalResponse: lastResponse || void 0 };
+      return { status: "paused", turnsUsed: turn, tokensUsed, finalResponse: lastResponse || void 0 };
     }
     turn += 1;
     yield { type: "turn_start", turn, goal };
@@ -6283,6 +6283,8 @@ async function* runUntilImpl(agent, goal, options, deps) {
     const run = await agent.send(continuationPrompt);
     const result = await run.wait();
     lastResponse = result.result ?? "";
+    const turnTokens = result.usage?.totalTokens;
+    if (typeof turnTokens === "number") tokensUsed += turnTokens;
     yield { type: "agent_response", turn, content: lastResponse };
     const judgeOpts = {};
     if (options?.judgeModel !== void 0) judgeOpts.judgeModel = options.judgeModel;
@@ -6308,6 +6310,7 @@ async function* runUntilImpl(agent, goal, options, deps) {
         return {
           status: "failed",
           turnsUsed: turn,
+          tokensUsed,
           finalResponse: lastResponse || void 0
         };
       }
@@ -6316,7 +6319,7 @@ async function* runUntilImpl(agent, goal, options, deps) {
     }
     if (judgment.verdict === "done") {
       yield { type: "status_change", status: "completed", reason: judgment.reason };
-      return { status: "completed", turnsUsed: turn, finalResponse: lastResponse || void 0 };
+      return { status: "completed", turnsUsed: turn, tokensUsed, finalResponse: lastResponse || void 0 };
     }
     if (judgment.verdict === "skipped") {
       yield {
@@ -6324,7 +6327,15 @@ async function* runUntilImpl(agent, goal, options, deps) {
         status: "completed",
         reason: `skipped: ${judgment.reason}`
       };
-      return { status: "completed", turnsUsed: turn, finalResponse: lastResponse || void 0 };
+      return { status: "completed", turnsUsed: turn, tokensUsed, finalResponse: lastResponse || void 0 };
+    }
+    if (tokenBudget !== void 0 && tokensUsed >= tokenBudget) {
+      yield {
+        type: "status_change",
+        status: "budget_limited",
+        reason: `token budget (${tokenBudget}) reached: ${tokensUsed} used`
+      };
+      return { status: "budget_limited", turnsUsed: turn, tokensUsed, finalResponse: lastResponse || void 0 };
     }
     yield { type: "continuation", turn, prompt: continuationPrompt };
   }
@@ -6333,13 +6344,32 @@ async function* runUntilImpl(agent, goal, options, deps) {
     status: "failed",
     reason: `max turns (${maxTurns}) exhausted`
   };
-  return { status: "failed", turnsUsed: turn, finalResponse: lastResponse || void 0 };
+  return { status: "failed", turnsUsed: turn, tokensUsed, finalResponse: lastResponse || void 0 };
 }
 function composeContinuation(goal, lastResponse) {
-  return `Continue working toward the goal: ${goal}
-
-Your last response was:
-${lastResponse.slice(0, 1e3)}`;
+  return [
+    "Continue working toward the active goal.",
+    "",
+    `<objective>
+${goal}
+</objective>`,
+    "",
+    "Continuation behavior:",
+    "- This goal persists across turns. Keep the full objective intact; if it cannot be finished now,",
+    "  make concrete progress toward the real requested end state and do not redefine success around a",
+    "  smaller or easier task.",
+    "",
+    "Work from evidence:",
+    "- Use the current worktree and external state as authoritative. Inspect the current state before",
+    "  relying on it. Improve, replace, or remove existing work as needed to satisfy the actual objective.",
+    "",
+    "Completion audit:",
+    "- Before deciding the goal is achieved, treat completion as unproven and verify it against the actual",
+    "  current state, requirement by requirement. Treat uncertain or indirect evidence as not achieved.",
+    "",
+    `Your last response was:
+${lastResponse.slice(0, 1e3)}`
+  ].join("\n");
 }
 var init_run_until = __esm({
   "src/internal/runtime/lifecycle/run-until.ts"() {
@@ -18583,7 +18613,7 @@ function localAgentRunUntil(agent, goal, options) {
         status: "paused",
         reason: "runUntil() requires an explicit goal (durable objectives removed in v4.0)"
       };
-      return { status: "paused", turnsUsed: 0, finalResponse: void 0 };
+      return { status: "paused", turnsUsed: 0, tokensUsed: 0, finalResponse: void 0 };
     }
     return yield* runUntilImpl2(agent, goal, options, await buildDeps());
   }