@namzu/sdk 51.0.0 → 52.0.0

package/dist/runtime/query/iteration/index.js

@@ -1,7 +1,7 @@
 import { SpanStatusCode } from '@opentelemetry/api';
 import { extractFromAssistantMessage, extractFromUserMessage, } from '../../../compaction/extractor.js';
 import { estimateMessageTokens } from '../../../compaction/token-estimate.js';
-import { AUTO_CONTINUATION_USER_MESSAGE } from '../../../constants/continuation.js';
+import { AUTO_CONTINUATION_USER_MESSAGE, REASONING_CUT_USER_MESSAGE, } from '../../../constants/continuation.js';
 import { CHARS_PER_TOKEN } from '../../../constants/limits.js';
 import { DEFAULT_STRUCTURED_OUTPUT_RETRIES, STRUCTURED_OUTPUT_REPROMPT, } from '../../../constants/tools/index.js';
 import { withStreamedMessageIdentity } from '../../../manager/session/streamed-message-identity.js';
@@ -15,7 +15,7 @@ import { getTracer } from '../../../telemetry/runtime-accessors.js';
 import { STRUCTURED_OUTPUT_TOOL_NAME } from '../../../tools/builtins/structuredOutput.js';
 import { NamzuError } from '../../../types/errors/index.js';
 import { createAssistantMessage, createRuntimeContextMessage, createSystemMessage, } from '../../../types/message/index.js';
-import { classifyProviderError } from '../../../types/provider/errors.js';
+import { classifyProviderError, readOutputOverflow } from '../../../types/provider/errors.js';
 import { toErrorMessage } from '../../../utils/error.js';
 import { stableDigest } from '../../../utils/hash.js';
 import { generateMessageId } from '../../../utils/id.js';
@@ -31,6 +31,7 @@ import { parseNativeCandidate } from './native-output.js';
 import { drainArrivedTaskCompletions, holdForOutstandingWork, settleOutstandingWork, } from './outstanding-work.js';
 import { runAdvisoryPhase } from './phases/advisory.js';
 import { runIterationCheckpoint } from './phases/checkpoint.js';
+import { compactionBreaker } from './phases/compaction-breaker.js';
 import { activeContextWindow, measureContext, relieveOverflow, runCompactionCheck, } from './phases/compaction.js';
 import { runHandoffPause } from './phases/handoff.js';
 import { runPlanGate } from './phases/plan.js';
@@ -60,6 +61,10 @@ class AnswerReviewFailure extends NamzuError {
  * on the thing that actually went wrong.
  */
 const DEFAULT_ANSWER_REVIEW_LIMIT = 3;
+/** Left unused when an output cap is shrunk to fit the window (reference: 1000). */
+const OUTPUT_CAP_MARGIN_TOKENS = 1000;
+/** A shrunk output cap below this is not worth a call (reference floor: 3000). */
+const MIN_OUTPUT_CAP_TOKENS = 3000;
 // Ending a turn changes the available actions, not the strength of its evidence.
 // Use the same standard for warning closure and empty-completion recovery.
 const CLOSING_RESPONSE_GUIDANCE = 'Give a concise response using only what the available evidence supports. Attribute unverified statements to their source instead of presenting them as observed facts. If evidence is missing or conflicting, state what cannot be established. Do not claim unfinished work is complete. Do not request any more tool calls.';
@@ -113,6 +118,67 @@ export class IterationOrchestrator {
      * sets it.
      */
     stopDeferredForOutstandingWork = false;
+    /** Nudges sent since the model last finished a call without hitting its output cap. */
+    outputContinuations = 0;
+    /**
+     * `turn_end` (and, for a child, `subagent_stop`) fired when the model
+     * finished and no hook asked for more. `finalizeTurn` then does not fire
+     * them a second time.
+     */
+    turnEndFired = false;
+    /** A hook already kept this turn going once: it is asked again with `stop_hook_active`. */
+    stopHookActive = false;
+    get turnEndHooksFired() {
+        return this.turnEndFired;
+    }
+    /**
+     * Fire `turn_end` (and `subagent_stop` for a child) at the point the model
+     * has finished, and read the answer.
+     *
+     * A hook that answers `skip` is refusing to let the turn end: its reasons
+     * are returned, joined, as feedback for the model. A hook that asks for the
+     * turn to halt wins over any such feedback (`takeHalt`). With nothing asked,
+     * the events have fired for this turn and the finalizer does not repeat them.
+     */
+    async *askStopHooks() {
+        const manager = this.ctx.pluginManager;
+        if (!manager)
+            return undefined;
+        const recorder = this.ctx.recorder;
+        const feedback = [];
+        const ask = async (event, extra) => {
+            const results = await manager.executeHooks(event, {
+                sessionId: recorder.sessionId,
+                turnId: recorder.turnId,
+                ...extra,
+                ...(this.stopHookActive ? { stopHookActive: true } : {}),
+                signal: this.ctx.abortController.signal,
+            }, this.ctx.emitEvent);
+            applyLifecycleHookResults(event, results);
+            for (const result of results) {
+                if (result.action === 'skip')
+                    feedback.push(result.reason);
+                if ((result.action === 'continue' || result.action === 'skip') && result.halt) {
+                    this.ctx.toolExecutor.requestHalt?.(result.halt.reason);
+                }
+            }
+        };
+        await ask('turn_end', {});
+        yield* this.ctx.drainPending();
+        if (this.ctx.parentSessionId !== undefined) {
+            await ask('subagent_stop', {
+                parentSessionId: this.ctx.parentSessionId,
+                ...(this.ctx.parentTurnId !== undefined ? { parentTurnId: this.ctx.parentTurnId } : {}),
+            });
+            yield* this.ctx.drainPending();
+        }
+        if (feedback.length === 0) {
+            this.turnEndFired = true;
+            return undefined;
+        }
+        this.stopHookActive = true;
+        return feedback.join('\n');
+    }
     /** One current input, independent of the compactable history array. */
     latestUserMessage;
     initialDurableDelivered = false;
@@ -274,6 +340,9 @@ export class IterationOrchestrator {
         // obvious moves left. It is cleared by a turn that actually
         // succeeded, which is the evidence that the turn is no longer stuck.
         let overflowRelieved = false;
+        // The output cap a provider told us fits beside this prompt. One request's
+        // worth: it is cleared by a call that went through.
+        let outputCapOverride;
         // A `finally` rather than a line at each exit, for the reason written
         // beside `iterSpan.end()` below: this loop leaves by eight `break`s,
         // two `return`s and a `throw`, and a rule every future edit has to
@@ -446,12 +515,16 @@ export class IterationOrchestrator {
                         applyLifecycleHookResults('iteration_start', hookResults);
                         yield* this.ctx.drainPending();
                     }
+                    if (!forceFinalize)
+                        await this.deliverReminders(iterationNum);
                     // Re-pin the working-memory block from ground truth at the primacy
                     // edge BEFORE compaction runs (so the refreshed slot is what
                     // compaction preserves). No-op when no provider is configured.
                     await refreshWorkingMemory(this.ctx);
                     await runCompactionCheck(this.ctx);
                     yield* this.ctx.drainPending();
+                    if (this.stopForCompactionBreaker())
+                        break;
                     // Cache discipline: keep the tools param byte-stable even on the
                     // forced-final iteration and forbid tool use via tool_choice
                     // 'none' instead. Dropping the tools array would invalidate the
@@ -498,6 +571,8 @@ export class IterationOrchestrator {
                         await runCompactionCheck(this.ctx);
                     // Publish context edits before entering a possibly slow provider.
                     yield* this.ctx.drainPending();
+                    if (this.stopForCompactionBreaker())
+                        break;
                     const stepAllowedTools = step.allowedTools ?? this.ctx.allowedTools;
                     const deferredMcpContext = deferredMcpHints?.forNames(stepAllowedTools) ?? '';
                     const contextText = [
@@ -660,7 +735,7 @@ export class IterationOrchestrator {
                                 messages: Object.freeze(messages.map((m) => Object.freeze({ ...m }))),
                                 toolNames: Object.freeze(llmTools.map((t) => t.function.name)),
                                 temperature: step.temperature ?? turnConfig.temperature,
-                                maxTokens: step.maxResponseTokens ?? turnConfig.maxResponseTokens,
+                                maxTokens: outputCapOverride ?? step.maxResponseTokens ?? turnConfig.maxResponseTokens,
                             }),
                         }, this.ctx.emitEvent);
                         applyLifecycleHookResults('pre_llm_call', hookResults);
@@ -726,7 +801,7 @@ export class IterationOrchestrator {
                                 ? step.toolChoice
                                 : undefined,
                         temperature: step.temperature ?? turnConfig.temperature,
-                        maxTokens: step.maxResponseTokens ?? turnConfig.maxResponseTokens,
+                        maxTokens: outputCapOverride ?? step.maxResponseTokens ?? turnConfig.maxResponseTokens,
                         cacheControl: { type: 'auto' },
                         ...(turnConfig.thinking ? { thinking: turnConfig.thinking } : {}),
                         ...(turnConfig.effort ? { effort: turnConfig.effort } : {}),
@@ -785,6 +860,9 @@ export class IterationOrchestrator {
                     // The turn went through, so the turn is not sitting on an
                     // irreducible prompt any more. Re-arm relief for the next one.
                     overflowRelieved = false;
+                    outputCapOverride = undefined;
+                    if (response.finishReason !== 'length')
+                        this.outputContinuations = 0;
                     if (this.ctx.pluginManager) {
                         const hookResults = await this.ctx.pluginManager.executeHooks('post_llm_call', {
                             sessionId: recorder.sessionId,
@@ -1055,16 +1133,43 @@ export class IterationOrchestrator {
                         //     it has no room left to continue into, and the
                         //     continuation would only send a longer prompt into
                         //     the same full window. The reply stands as it is.
+                        const cutDuringReasoning = !hasContent &&
+                            (response.message.toolCalls?.length ?? 0) === 0 &&
+                            (response.message.reasoning?.length ?? 0) > 0;
                         if (!forceFinalize &&
                             response.finishReason === 'length' &&
                             response.finishDetail !== 'context_window' &&
-                            hasContent) {
-                            this.ctx.log.info('LLM hit max_tokens mid-text — auto-continuing', {
+                            (hasContent || cutDuringReasoning)) {
+                            // A host that opted in gets a cap on the nudges: a model that
+                            // overruns its output limit every time would otherwise spend the
+                            // whole turn on it. Past the cap the turn stops, and says why.
+                            const cap = this.ctx.maxOutputContinuations;
+                            if (cap !== undefined && this.outputContinuations >= cap) {
+                                this.ctx.log.warn('LLM kept hitting max_tokens — stopping the turn', {
+                                    [NAMZU.TURN_ID]: recorder.turnId,
+                                    [NAMZU.ITERATION]: iterationNum,
+                                    'namzu.runtime.output_continuations': this.outputContinuations,
+                                });
+                                recorder.setLastError(`The reply was cut off at the output limit ${cap + 1} times in a row. Ask for a smaller piece of the work, or raise the output limit.`);
+                                recorder.setStopReason('output_limit');
+                                await this.ctx.emitEvent({
+                                    type: 'iteration_completed',
+                                    turnId: recorder.turnId,
+                                    iteration: iterationNum,
+                                    hasToolCalls: false,
+                                });
+                                yield* this.ctx.drainPending();
+                                break;
+                            }
+                            this.outputContinuations++;
+                            this.ctx.log.info(cutDuringReasoning
+                                ? 'LLM hit max_tokens while reasoning — nudging to answer briefly'
+                                : 'LLM hit max_tokens mid-text — auto-continuing', {
                                 [NAMZU.TURN_ID]: recorder.turnId,
                                 [NAMZU.ITERATION]: iterationNum,
                                 [GENAI.USAGE_OUTPUT_TOKENS]: response.usage.completionTokens,
                             });
-                            recorder.pushMessage(createRuntimeContextMessage(AUTO_CONTINUATION_USER_MESSAGE, 'auto-continuation'));
+                            recorder.pushMessage(createRuntimeContextMessage(cutDuringReasoning ? REASONING_CUT_USER_MESSAGE : AUTO_CONTINUATION_USER_MESSAGE, 'auto-continuation'));
                             await this.ctx.emitEvent({
                                 type: 'iteration_completed',
                                 turnId: recorder.turnId,
@@ -1206,6 +1311,22 @@ export class IterationOrchestrator {
                         // also settle as `end_turn`, and there the deferred predicate
                         // is not why the turn ended: those decided the answer
                         // themselves.
+                        // The model is done and has no tool calls. A `turn_end` (or
+                        // `subagent_stop`) hook may refuse to let it be: its reason goes
+                        // back to the model and the turn goes on.
+                        if (!forceFinalize && closingStopReason === undefined) {
+                            const feedback = yield* this.askStopHooks();
+                            const stoppedByHook = this.ctx.toolExecutor.takeHalt?.();
+                            if (stoppedByHook !== undefined) {
+                                recorder.setLastError(`Stopped by a hook: ${stoppedByHook}`);
+                                recorder.setStopReason('stop_condition');
+                                break;
+                            }
+                            if (feedback !== undefined) {
+                                recorder.pushMessage(createRuntimeContextMessage(`A hook asked for more before this turn ends:\n${feedback}`, 'hook-feedback'));
+                                continue;
+                            }
+                        }
                         recorder.setStopReason(closingStopReason ??
                             (stopWasDeferredForOutstandingWork ? 'stop_condition' : 'end_turn'));
                         break;
@@ -1240,6 +1361,27 @@ export class IterationOrchestrator {
                     if ((yield* runHandoffPause(this.ctx, iterationNum, reviewOutcome.results)) === 'stop') {
                         return;
                     }
+                    // A hook asked for the turn to stop once this step's calls were
+                    // answered. Every `tool_use` has its `tool_result` by now, so the
+                    // transcript is whole; what is left is to end the turn and say why.
+                    const hookHalt = this.ctx.toolExecutor.takeHalt?.();
+                    if (hookHalt !== undefined) {
+                        this.ctx.log.info('A hook stopped the turn', {
+                            [NAMZU.TURN_ID]: recorder.turnId,
+                            [NAMZU.ITERATION]: iterationNum,
+                            'namzu.hook.reason': hookHalt,
+                        });
+                        recorder.setLastError(`Stopped by a hook: ${hookHalt}`);
+                        recorder.setStopReason('stop_condition');
+                        await this.ctx.emitEvent({
+                            type: 'iteration_completed',
+                            turnId: recorder.turnId,
+                            iteration: iterationNum,
+                            hasToolCalls: true,
+                        });
+                        yield* this.ctx.drainPending();
+                        break;
+                    }
                     if (reviewOutcome.decision === 'rejected') {
                         continue;
                     }
@@ -1541,6 +1683,29 @@ export class IterationOrchestrator {
                     // actually shed something. A second overflow after a successful
                     // compaction means the prompt is irreducible, and looping on it
                     // would burn the budget to arrive at the same error.
+                    // The prompt fit and only the requested output did not: ask for what
+                    // room is left (less a margin) and send it again, once. A reply that
+                    // would be too small to be useful, or a cap that is not smaller than
+                    // the one already tried, is not retried: the same failure would follow.
+                    const outputOverflow = err instanceof AnswerReviewFailure ? undefined : readOutputOverflow(err);
+                    if (outputOverflow) {
+                        const room = Math.max(0, outputOverflow.contextLimit - outputOverflow.inputTokens - OUTPUT_CAP_MARGIN_TOKENS);
+                        if (room >= MIN_OUTPUT_CAP_TOKENS &&
+                            (outputCapOverride === undefined || room < outputCapOverride)) {
+                            outputCapOverride = room;
+                            this.ctx.log.info('Output cap does not fit beside the prompt — retrying with less', {
+                                [NAMZU.TURN_ID]: recorder.turnId,
+                                [NAMZU.ITERATION]: iterationNum,
+                                'namzu.runtime.input_tokens': outputOverflow.inputTokens,
+                                'namzu.runtime.context_limit': outputOverflow.contextLimit,
+                                'namzu.runtime.adjusted_max_tokens': room,
+                            });
+                            if (iterationActivity) {
+                                this.ctx.activityStore.complete(iterationActivity.id);
+                            }
+                            continue;
+                        }
+                    }
                     if (!overflowRelieved &&
                         !(err instanceof AnswerReviewFailure) &&
                         classifyProviderError(err, this.ctx.provider.id).code === 'context_length_exceeded') {
@@ -1639,6 +1804,41 @@ export class IterationOrchestrator {
             extractFromUserMessage(manager, message.content, !manager.getState().task);
         }
     }
+    /**
+     * Ask the host's reminder source (`QueryParams.reminders`) what the model
+     * should be told before this request, and append each text as a hidden
+     * user-role message. The start of a request is a provider-valid boundary:
+     * the history ends in the operator's message or in a finished tool batch.
+     * A source that throws has nothing to say; it never fails the turn.
+     */
+    async deliverReminders(iteration) {
+        const source = this.ctx.reminders;
+        if (!source)
+            return;
+        const { recorder } = this.ctx;
+        let texts;
+        try {
+            texts = await source({
+                sessionId: recorder.sessionId,
+                turnId: recorder.turnId,
+                iteration,
+                messages: recorder.messages,
+            });
+        }
+        catch (error) {
+            this.ctx.log.warn('A reminder source threw; the request goes without its reminders', {
+                [NAMZU.TURN_ID]: recorder.turnId,
+                [NAMZU.ITERATION]: iteration,
+                'exception.message': error instanceof Error ? error.message : String(error),
+            });
+            return;
+        }
+        for (const text of texts ?? []) {
+            if (text.trim().length === 0)
+                continue;
+            recorder.pushMessage(createRuntimeContextMessage(text, 'reminder'));
+        }
+    }
     async deliverInbound() {
         const queued = this.ctx.inboundMessages?.() ?? [];
         for (const message of queued) {
@@ -1756,6 +1956,15 @@ export class IterationOrchestrator {
         recorder.setStopReason('structured_output_failed');
         return true;
     }
+    /** The compaction circuit breaker asked for the turn to stop (opt-in; see `compactionConfig.circuitBreaker`). */
+    stopForCompactionBreaker() {
+        const reason = compactionBreaker(this.ctx)?.stop;
+        if (reason === undefined)
+            return false;
+        this.ctx.recorder.setLastError(reason);
+        this.ctx.recorder.setStopReason('error');
+        return true;
+    }
     async recordToolStructuredFailure(results, iteration) {
         if (!this.needsStructuredOutput() ||
             this.ctx.structuredOutput?.mode === 'native' ||