@namzu/sdk 61.3.0 → 62.0.0

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

@@ -29,6 +29,8 @@ import { diffRequestContext, snapshotRequestContext, } from '../request-context.
 import { DEFAULT_MAX_REQUEST_RICH_CONTENT_BYTES, markProviderRejectedImage, projectRequestRichContent, } from '../request-rich-content.js';
 import { withoutVoidAssistantTurns } from '../request-void-turns.js';
 import { formatSteeringNote, isOperatorUserMessage } from '../steering.js';
+import { COMPLETION_AUDIT_MESSAGE, CompletionAudit } from '../turn/completion-audit.js';
+import { FORCED_STEP_MESSAGE, REASONING_CAPPED_MESSAGE, REASONING_ONLY_MESSAGE, StallMonitor, callChangedState, reasoningCapChars, } from '../turn/stall-monitor.js';
 import { effortForModel } from './effort-clamp.js';
 import { parseNativeCandidate } from './native-output.js';
 import { drainArrivedTaskCompletions, holdForOutstandingWork, settleOutstandingWork, } from './outstanding-work.js';
@@ -46,6 +48,8 @@ import { RequestRepair } from './request-repair.js';
 import { placeRequestContext } from './step-context.js';
 import { beforeStep, prepareStep, selectContextModel, stepContextMessage, workContext, } from './step-shaping.js';
 import { streamProviderTurn } from './stream-turn.js';
+/** Capped steps in a row after which the per-step cap stops applying for the turn's next step. */
+const MAX_CAPPED_STREAK = 3;
 /** A host reviewer is not the model transport, even when its cause is an HTTP failure. */
 class AnswerReviewFailure extends NamzuError {
     constructor(cause) {
@@ -193,6 +197,16 @@ export class IterationOrchestrator {
     stopDeferredForOutstandingWork = false;
     /** Nudges sent since the model last finished a call without hitting its output cap. */
     outputContinuations = 0;
+    /** Watches the turn's steps for reasoning that piles up while nothing changes. */
+    stallMonitor;
+    /** Holds the turn back once when it changed files and checked nothing afterwards. */
+    completionAudit;
+    /** The next step must call a tool, at low effort: set by the stall monitor or a capped step. */
+    forceActionNextStep = false;
+    /** Steps in a row cut at the per-step reasoning cap; past a few the cap stops applying. */
+    cappedStreak = 0;
+    /** A reply that held only reasoning was already answered with one more request. */
+    reasoningOnlyRetried = false;
     /**
      * The model a refused request moved to. Once set it replaces the model the
      * step asked for, for the rest of the turn: the request that was refused
@@ -275,6 +289,11 @@ export class IterationOrchestrator {
                 this.rememberUserMessage(createRuntimeContextMessage(text, 'steering'));
             },
         };
+        if (ctx.turnConfig.completionAudit === true)
+            this.completionAudit = new CompletionAudit();
+        if (ctx.turnConfig.reasoning?.stall !== false) {
+            this.stallMonitor = new StallMonitor(ctx.turnConfig.reasoning?.stall || undefined);
+        }
         ctx.checkpointMgr.setLatestUserMessageSource(() => this.latestUserMessage);
         ctx.checkpointMgr.setAnswerReviewAttemptsSource?.(() => this.answerReviewAttempts);
         ctx.checkpointMgr.setStructuredReviewAttemptsSource?.(() => this.structuredReviewAttempts);
@@ -871,6 +890,11 @@ export class IterationOrchestrator {
                     // died part-way has already emitted both `message_started` and
                     // `message_completed` under this id, which is the trail a
                     // reader wants most on exactly that turn.
+                    // A step the loop forces to act runs at low effort with a tool call
+                    // required, for this step only. Low is used only where the model
+                    // is known to offer it; effort is refused, not clamped.
+                    const forcedAction = this.forceActionNextStep && !forceFinalize && llmTools.length > 0;
+                    this.forceActionNextStep = false;
                     stepMessageId = generateMessageId();
                     const requestedMember = this.ctx.servingMember?.() ?? {
                         index: 0,
@@ -895,11 +919,22 @@ export class IterationOrchestrator {
                             });
                         }
                     }
+                    // A forced action step runs at low effort, where the model is known
+                    // to offer it; effort is refused, not clamped.
+                    const offeredEfforts = this.ctx.provider.reasoningEffortLevelsFor?.(stepModel);
+                    const stepEffort = forcedAction && requestEffort && (offeredEfforts?.includes('low') ?? false)
+                        ? 'low'
+                        : requestEffort;
+                    // A provider that refuses a forced tool choice (some models do, and
+                    // so does manual extended thinking) is not sent one: the step keeps
+                    // its low effort and asks in words, on this request only.
+                    const forcedChoiceRefused = forcedAction &&
+                        this.ctx.provider.acceptsForcedToolChoiceFor?.(stepModel, turnConfig.thinking) === false;
                     const operatorInputAtDispatch = this.latestUserMessage;
                     const latestReviewUserMessage = (this.ctx.reviewAnswer || this.ctx.structuredOutput?.review) && operatorInputAtDispatch
                         ? structuredClone(operatorInputAtDispatch)
                         : undefined;
-                    const { response, messageId, requestMessages } = yield* streamProviderTurn(this.ctx.provider, {
+                    const { response, messageId, requestMessages, reasoningCapped = false, } = yield* streamProviderTurn(this.ctx.provider, {
                         model: stepModel,
                         ...(this.ctx.structuredOutput?.mode === 'native'
                             ? {
@@ -914,7 +949,9 @@ export class IterationOrchestrator {
                             }
                             : {}),
                         providerRoute: requestedRoute,
-                        messages,
+                        messages: forcedChoiceRefused
+                            ? [...messages, createRuntimeContextMessage(FORCED_STEP_MESSAGE, 'progress-nudge')]
+                            : messages,
                         tools: llmTools.length > 0 ? llmTools : undefined,
                         ...(enforceToolInputSchema ? { enforceToolInputSchema } : {}),
                         // The forced-final turn wins: a step that asked to force a
@@ -925,13 +962,15 @@ export class IterationOrchestrator {
                         toolChoice: forceFinalize && llmTools.length > 0
                             ? 'none'
                             : llmTools.length > 0
-                                ? step.toolChoice
+                                ? forcedAction && !forcedChoiceRefused
+                                    ? (step.toolChoice ?? 'required')
+                                    : step.toolChoice
                                 : undefined,
                         temperature: step.temperature ?? turnConfig.temperature,
                         maxTokens: outputCapOverride ?? step.maxResponseTokens ?? turnConfig.maxResponseTokens,
                         cacheControl: { type: 'auto' },
                         ...(turnConfig.thinking ? { thinking: turnConfig.thinking } : {}),
-                        ...(requestEffort ? { effort: requestEffort } : {}),
+                        ...(stepEffort ? { effort: stepEffort } : {}),
                         ...(!forceFinalize && turnConfig.webSearch
                             ? { webSearch: turnConfig.webSearch }
                             : {}),
@@ -943,8 +982,15 @@ export class IterationOrchestrator {
                         onAccepted: (identity) => this.acceptProviderRejectedImage(identity),
                         textOnly: this.textOnlyRecovery(stepModel),
                         repair: this.requestRepair,
-                    }, Boolean(this.ctx.reviewAnswer || this.ctx.structuredOutput?.review || this.ctx.advisoryCtx), recorder.sessionId);
+                    }, Boolean(this.ctx.reviewAnswer || this.ctx.structuredOutput?.review || this.ctx.advisoryCtx), recorder.sessionId, 
+                    // Applies only to a step that can act. Stops after a few capped
+                    // steps in a row, so a model that cannot act within the allowance
+                    // still gets to finish.
+                    this.cappedStreak >= MAX_CAPPED_STREAK || forceFinalize || llmTools.length === 0
+                        ? undefined
+                        : reasoningCapChars(turnConfig.reasoning, stepEffort));
                     stepResponse = response;
+                    this.cappedStreak = reasoningCapped ? this.cappedStreak + 1 : 0;
                     const reviewRequest = {
                         ...(requestMessages ? { requestMessages } : {}),
                         ...(latestReviewUserMessage ? { latestUserMessage: latestReviewUserMessage } : {}),
@@ -1206,6 +1252,25 @@ export class IterationOrchestrator {
                             usageBefore,
                             costBefore,
                         });
+                        // The step spent its whole reasoning allowance before acting. Tell
+                        // the model so and make the next step an action step.
+                        if (reasoningCapped && !forceFinalize) {
+                            this.ctx.log.info('Reasoning reached the per-step cap — asking the model to act', {
+                                [NAMZU.TURN_ID]: recorder.turnId,
+                                [NAMZU.ITERATION]: iterationNum,
+                            });
+                            recorder.pushMessage(createRuntimeContextMessage(REASONING_CAPPED_MESSAGE, 'progress-nudge'));
+                            this.forceActionNextStep = true;
+                            this.observeStall(response, []);
+                            await this.ctx.emitEvent({
+                                type: 'iteration_completed',
+                                turnId: recorder.turnId,
+                                iteration: iterationNum,
+                                hasToolCalls: false,
+                            });
+                            yield* this.ctx.drainPending();
+                            continue;
+                        }
                         if (this.ctx.structuredOutput?.mode === 'native') {
                             const candidate = await parseNativeCandidate(this.ctx.structuredOutput.schema, response, this.ctx.abortController.signal);
                             let outcome;
@@ -1445,6 +1510,27 @@ export class IterationOrchestrator {
                         let closingStopReason = forceFinalize
                             ? guardResult.stopReason
                             : undefined;
+                        const reasoningOnly = !hasContent && (response.message.reasoning?.length ?? 0) > 0;
+                        if (hasContent)
+                            this.reasoningOnlyRetried = false;
+                        // A reply of reasoning alone is not an end: ask once more, plainly,
+                        // before the closing request below.
+                        if (reasoningOnly && !forceFinalize && !this.reasoningOnlyRetried) {
+                            this.reasoningOnlyRetried = true;
+                            this.ctx.log.warn('Reply held reasoning only — asking once more', {
+                                [NAMZU.TURN_ID]: recorder.turnId,
+                                [NAMZU.ITERATION]: iterationNum,
+                            });
+                            recorder.pushMessage(createRuntimeContextMessage(REASONING_ONLY_MESSAGE, 'progress-nudge'));
+                            await this.ctx.emitEvent({
+                                type: 'iteration_completed',
+                                turnId: recorder.turnId,
+                                iteration: iterationNum,
+                                hasToolCalls: false,
+                            });
+                            yield* this.ctx.drainPending();
+                            continue;
+                        }
                         if (!hasContent && !forceFinalize) {
                             this.ctx.log.warn('Empty completion detected — requesting final summary', {
                                 [NAMZU.ITERATION]: iterationNum,
@@ -1452,6 +1538,13 @@ export class IterationOrchestrator {
                             });
                             closingStopReason = await this.requestFinalResponse(this.refusalPinnedModel ?? model, 'end_turn');
                             yield* this.ctx.drainPending();
+                            // Never end silently: if even the closing request produced nothing,
+                            // say so in plain words and how to go on.
+                            if (reasoningOnly &&
+                                closingStopReason === undefined &&
+                                !([...recorder.messages].reverse().find((m) => m.role === 'assistant')?.content ?? '').length) {
+                                recorder.setLastError('The model stopped after thinking without giving an answer or taking an action, even when asked again. Say "continue" and it will pick the work up from here.');
+                            }
                         }
                         await this.ctx.emitEvent({
                             type: 'iteration_completed',
@@ -1483,6 +1576,17 @@ export class IterationOrchestrator {
                         // 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.
+                        // Files were changed and nothing checked them afterwards: ask once
+                        // for a check before the turn is allowed to end.
+                        if (!forceFinalize &&
+                            closingStopReason === undefined &&
+                            this.completionAudit?.takeBlock()) {
+                            this.ctx.log.info('Files changed and nothing checked them — asking for a check', {
+                                [NAMZU.TURN_ID]: recorder.turnId,
+                            });
+                            recorder.pushMessage(createRuntimeContextMessage(COMPLETION_AUDIT_MESSAGE, 'progress-nudge'));
+                            continue;
+                        }
                         if (!forceFinalize && closingStopReason === undefined) {
                             const feedback = yield* this.askStopHooks();
                             const stoppedByHook = this.ctx.toolExecutor.takeHalt?.();
@@ -1520,6 +1624,9 @@ export class IterationOrchestrator {
                         usageBefore,
                         costBefore,
                     });
+                    this.reasoningOnlyRetried = false;
+                    this.observeStall(response, reviewOutcome.results);
+                    this.observeAudit(response, reviewOutcome.results);
                     if (reviewOutcome.decision === 'stop') {
                         return;
                     }
@@ -2645,6 +2752,62 @@ export class IterationOrchestrator {
             return undefined;
         }
     }
+    /** Feed the completion audit the calls of one finished step. */
+    observeAudit(response, results) {
+        const audit = this.completionAudit;
+        if (!audit)
+            return;
+        const calls = new Map((response.message.toolCalls ?? []).map((c) => [c.id, c]));
+        audit.observe(results.map((result) => {
+            let input;
+            try {
+                const call = calls.get(result.toolCallId);
+                input = call ? JSON.parse(call.function.arguments) : undefined;
+            }
+            catch {
+                input = undefined;
+            }
+            return { toolName: result.toolName, input, isError: result.isError };
+        }));
+    }
+    /**
+     * Feed the stall monitor one finished step and act on its verdict: a nudge
+     * is recorded as a message, a force makes the next step an action step.
+     */
+    observeStall(response, results) {
+        const monitor = this.stallMonitor;
+        if (!monitor)
+            return;
+        const calls = new Map((response.message.toolCalls ?? []).map((c) => [c.id, c]));
+        let changedState = false;
+        for (const result of results) {
+            const call = calls.get(result.toolCallId);
+            let readOnly;
+            try {
+                const input = call ? JSON.parse(call.function.arguments) : undefined;
+                readOnly = this.ctx.tools.get(result.toolName)?.isReadOnly?.(input);
+            }
+            catch {
+                readOnly = undefined;
+            }
+            if (callChangedState(result.toolName, readOnly, result.isError))
+                changedState = true;
+        }
+        const reasoningChars = (response.message.reasoning ?? []).reduce((n, block) => n + (block.text?.length ?? 0), 0);
+        const directive = monitor.observe({ reasoningChars, changedState });
+        if (directive.kind === 'nudge') {
+            this.ctx.log.info('Many steps without a change — nudging the model to act', {
+                [NAMZU.TURN_ID]: this.ctx.recorder.turnId,
+            });
+            this.ctx.recorder.pushMessage(createRuntimeContextMessage(directive.text, 'progress-nudge'));
+        }
+        else if (directive.kind === 'force') {
+            this.ctx.log.info('Still no change after the nudge — forcing an action step', {
+                [NAMZU.TURN_ID]: this.ctx.recorder.turnId,
+            });
+            this.forceActionNextStep = true;
+        }
+    }
     async requestFinalResponse(model, reason) {
         if (this.ctx.structuredOutput?.mode === 'native')
             return 'structured_output_failed';