@qvac/harness 0.2.7 → 0.2.9

package/dist/prebuilds-manifest.json

@@ -272,6 +272,14 @@
       "win32-arm64",
       "win32-x64"
     ],
+    "bare-tty/node_modules/bare-signals": [
+      "darwin-arm64",
+      "darwin-x64",
+      "linux-arm64",
+      "linux-x64",
+      "win32-arm64",
+      "win32-x64"
+    ],
     "bare-tty": [
       "darwin-arm64",
       "darwin-x64",

package/dist-lib/index.d.ts

@@ -93,6 +93,7 @@ export interface CompletionContext {
     toolConfigs?: Record<string, Record<string, unknown>>;
     generationParams?: LlmGenerationParams;
     responseFormat?: ResponseFormat;
+    captureInferenceCalls?: boolean;
     grantedDirs?: string[];
     grantedDrives?: GrantedDrive[];
     driveOp?: DriveOpFn;
@@ -239,7 +240,7 @@ export declare class Harness extends ReadyResource {
         agent?: Agent;
         signal?: BareAbortSignal;
     }): Promise<void>;
-    completion({ model, agent, messages, cacheKey, thinkingDefault, toolConfigs, generationParams, responseFormat, grantedDirs, grantedDrives, driveOp, grantedWriteDirs, approve, attach, fetchAttachment, ragSearch, progress, askUser, signal }: CompletionContext): AsyncGenerator<CompletionEvent>;
+    completion({ model, agent, messages, cacheKey, thinkingDefault, toolConfigs, generationParams, responseFormat, captureInferenceCalls, grantedDirs, grantedDrives, driveOp, grantedWriteDirs, approve, attach, fetchAttachment, ragSearch, progress, askUser, signal }: CompletionContext): AsyncGenerator<CompletionEvent>;
     private _makeRoomForCompletion;
     private _evictWedgedModel;
     private _evictAfterLeases;

package/dist-lib/index.js

@@ -2,65 +2,65 @@ import fs from 'fs';
 import { platform, totalmem } from 'os';
 import process from 'process';
 import ReadyResource from 'ready-resource';
-import { probeGpuStats } from "./lib/hardware/gpu-probe.js";
-import { probeCapabilityProfile } from "./lib/hardware/capability-probe.js";
-import { closePythonWorker } from "./lib/tools/exec/adapters/python/worker/client.js";
-import { completion } from "./lib/capabilities/completion.js";
-import { download } from "./lib/capabilities/download.js";
-import { createEmbedLane } from "./lib/capabilities/embed.js";
-import { deleteAllKvCaches, deleteKvCacheKey } from "./lib/capabilities/kv-cache.js";
-import { pcmForTranscription } from "./lib/capabilities/audio-resample.js";
-import { buildWavBuffer } from "./lib/wav-encode.js";
-import { createVoiceLane, packSamples, speakOnce, speakStreamOnce, transcribeOnce, transcribeStreaming, SUPERTONIC_SAMPLE_RATE } from "./lib/capabilities/voice.js";
-import { pushQueue } from "./lib/queue.js";
-import { AbortedError, raceAbort } from "./lib/helpers/race-abort.js";
-import { audioGenDownloadMetadata, customProjectionMap, defaultProjectionMap, defaultTtsFor, defaultVoiceModel, diffusionDownloadMetadata, infraModelsFor, modelNameToCapabilityProfile, musicModelSet, setCustomModels } from "./lib/models/model-catalog.js";
-import { createModelStatusCache, invalidateModelStatusCache, listModels } from "./lib/models/model-list.js";
-import { deleteModelFiles, summarizeDelete } from "./lib/models/model-delete.js";
-import { createEmbedPool, createLoadTracker, createModelPool, createVoicePool, awaitModelCapacity, evictIdle, evictModelWithin, isResident, leaseCount, loadedModels, modelKey, releaseModel, resolveLoadOptions, progressState } from "./lib/models/model-pool.js";
-import { resolveMainGpu } from "./lib/models/gpu-selection.js";
-import { openclCacheDir } from "./lib/models/opencl-cache.js";
-import { assertEmbedDevice, createEmbedConfig } from "./lib/models/embed-config.js";
-import { createGpuFaultState, isSdkGpuDeviceLostError, isSdkModelWedgedError } from "./lib/models/sdk-faults.js";
-import { createSdkLoader } from "./lib/sdk-loader.js";
-import { createEditImageTool } from "./lib/tools/edit-image.js";
-import { toolCatalog } from "./lib/tools/catalog.js";
-import { createDiffusionRuntime, createGenerateImageTool } from "./lib/tools/generate-image.js";
-import { createGenerateMusicTool, createMusicRuntime } from "./lib/tools/generate-music.js";
-import { availableSkills, detectPlatform, loadSkillCatalog, resolveSkillStatuses, skillSupportsPlatform, versionProbeSkills } from "./lib/tools/skills/catalog.js";
-import { ensureBundledSkillsDir } from "./lib/tools/skills/bundled-dir.js";
-import { ensureSkillsOverlayDir } from "./lib/tools/skills/overlay-dir.js";
-import { copyBaseForEdit, deleteOverlaySkill, listManagedSkills, readSkillFile, writeSkillFile } from "./lib/tools/skills/overlay-crud.js";
-import { watchSkillsOverlay } from "./lib/tools/skills/overlay-watch.js";
-import { hlog } from "./lib/log.js";
+import { probeGpuStats } from './lib/hardware/gpu-probe.js';
+import { probeCapabilityProfile } from './lib/hardware/capability-probe.js';
+import { closePythonWorker } from './lib/tools/exec/adapters/python/worker/client.js';
+import { completion } from './lib/capabilities/completion.js';
+import { download } from './lib/capabilities/download.js';
+import { createEmbedLane } from './lib/capabilities/embed.js';
+import { deleteAllKvCaches, deleteKvCacheKey } from './lib/capabilities/kv-cache.js';
+import { pcmForTranscription } from './lib/capabilities/audio-resample.js';
+import { buildWavBuffer } from './lib/wav-encode.js';
+import { createVoiceLane, packSamples, speakOnce, speakStreamOnce, transcribeOnce, transcribeStreaming, SUPERTONIC_SAMPLE_RATE } from './lib/capabilities/voice.js';
+import { pushQueue } from './lib/queue.js';
+import { AbortedError, raceAbort } from './lib/helpers/race-abort.js';
+import { audioGenDownloadMetadata, customProjectionMap, defaultProjectionMap, defaultTtsFor, defaultVoiceModel, diffusionDownloadMetadata, infraModelsFor, modelNameToCapabilityProfile, musicModelSet, setCustomModels } from './lib/models/model-catalog.js';
+import { createModelStatusCache, invalidateModelStatusCache, listModels } from './lib/models/model-list.js';
+import { deleteModelFiles, summarizeDelete } from './lib/models/model-delete.js';
+import { createEmbedPool, createLoadTracker, createModelPool, createVoicePool, awaitModelCapacity, evictIdle, evictModelWithin, isResident, leaseCount, loadedModels, modelKey, releaseModel, resolveLoadOptions, progressState } from './lib/models/model-pool.js';
+import { resolveMainGpu } from './lib/models/gpu-selection.js';
+import { openclCacheDir } from './lib/models/opencl-cache.js';
+import { assertEmbedDevice, createEmbedConfig } from './lib/models/embed-config.js';
+import { createGpuFaultState, isSdkGpuDeviceLostError, isSdkModelWedgedError } from './lib/models/sdk-faults.js';
+import { createSdkLoader } from './lib/sdk-loader.js';
+import { createEditImageTool } from './lib/tools/edit-image.js';
+import { toolCatalog } from './lib/tools/catalog.js';
+import { createDiffusionRuntime, createGenerateImageTool } from './lib/tools/generate-image.js';
+import { createGenerateMusicTool, createMusicRuntime } from './lib/tools/generate-music.js';
+import { availableSkills, detectPlatform, loadSkillCatalog, resolveSkillStatuses, skillSupportsPlatform, versionProbeSkills } from './lib/tools/skills/catalog.js';
+import { ensureBundledSkillsDir } from './lib/tools/skills/bundled-dir.js';
+import { ensureSkillsOverlayDir } from './lib/tools/skills/overlay-dir.js';
+import { copyBaseForEdit, deleteOverlaySkill, listManagedSkills, readSkillFile, writeSkillFile } from './lib/tools/skills/overlay-crud.js';
+import { watchSkillsOverlay } from './lib/tools/skills/overlay-watch.js';
+import { hlog } from './lib/log.js';
 import { harness } from './spec/index.js';
-import { installResolutionDetector, resolutionDetail } from "./lib/resolution-detector.js";
-import { probeBinVersions } from "./lib/tools/skills/bin-check.js";
-import { bareRuntime } from "./lib/tools/exec.js";
-import { runCommand } from "./lib/tools/exec/runner.js";
-import { createLocalRoots } from "./lib/tools/fs/local.js";
-import { createRemoteRoots } from "./lib/tools/fs/remote.js";
-import { resetCredentialHealth as resetCredentialHealthAfterSnapshot } from "./lib/tools/skills/credential-health.js";
-import { isDeviceSharedCredentialKey, scopeKey } from "./lib/tools/skills/scope-key.js";
-import { settingsScopeKey } from "./lib/tools/skills/skill-settings.js";
-export * from "./lib/capabilities/completion.js";
-export { attachmentTurn, loadedSkillsFrom, skillLoadedTurn } from "./lib/capabilities/tool-turns.js";
-export * from "./lib/capabilities/download.js";
-export * from "./lib/models/model-catalog.js";
-export * from "./lib/models/model-list.js";
-export { harnessChild, spawnHarness } from "./lib/spawn.js";
-export { settingsDimension, settingsSteps } from "./lib/tools/generate-image.js";
-export * from "./lib/tools/skills/catalog.js";
-export { requiredSettings, settingEnvName } from "./lib/tools/skills/skill-settings.js";
-export { TOOLSETS, toolsetEnabled } from "./lib/tools/toolsets.js";
-export { markCredentialRejected, markCredentialValid, recordCredentialHealth, resetAllCredentialHealth, resetCredentialHealth } from "./lib/tools/skills/credential-health.js";
-export { readSetup as normalizeSkillSetup } from "./lib/tools/skills/manifest-parse.js";
-export { isAvailableState, SKILL_STATE } from "./lib/tools/skills/skill-state.js";
-export { connectHarness, HARNESS_DIED, RemoteHarness } from "./lib/connect.js";
-export { serveHarness } from "./lib/serve.js";
-export { hlog, setHarnessLog, setHarnessLogSink } from "./lib/log.js";
-export { harnessHooks } from "./lib/hooks/registry.js";
-export { EgressGuardHook, SkillCatalogGuardHook } from "./lib/hooks/built-in/index.js";
+import { installResolutionDetector, resolutionDetail } from './lib/resolution-detector.js';
+import { probeBinVersions } from './lib/tools/skills/bin-check.js';
+import { bareRuntime } from './lib/tools/exec.js';
+import { runCommand } from './lib/tools/exec/runner.js';
+import { createLocalRoots } from './lib/tools/fs/local.js';
+import { createRemoteRoots } from './lib/tools/fs/remote.js';
+import { resetCredentialHealth as resetCredentialHealthAfterSnapshot } from './lib/tools/skills/credential-health.js';
+import { isDeviceSharedCredentialKey, scopeKey } from './lib/tools/skills/scope-key.js';
+import { settingsScopeKey } from './lib/tools/skills/skill-settings.js';
+export * from './lib/capabilities/completion.js';
+export { attachmentTurn, loadedSkillsFrom, skillLoadedTurn } from './lib/capabilities/tool-turns.js';
+export * from './lib/capabilities/download.js';
+export * from './lib/models/model-catalog.js';
+export * from './lib/models/model-list.js';
+export { harnessChild, spawnHarness } from './lib/spawn.js';
+export { settingsDimension, settingsSteps } from './lib/tools/generate-image.js';
+export * from './lib/tools/skills/catalog.js';
+export { requiredSettings, settingEnvName } from './lib/tools/skills/skill-settings.js';
+export { TOOLSETS, toolsetEnabled } from './lib/tools/toolsets.js';
+export { markCredentialRejected, markCredentialValid, recordCredentialHealth, resetAllCredentialHealth, resetCredentialHealth } from './lib/tools/skills/credential-health.js';
+export { readSetup as normalizeSkillSetup } from './lib/tools/skills/manifest-parse.js';
+export { isAvailableState, SKILL_STATE } from './lib/tools/skills/skill-state.js';
+export { connectHarness, HARNESS_DIED, RemoteHarness } from './lib/connect.js';
+export { serveHarness } from './lib/serve.js';
+export { hlog, setHarnessLog, setHarnessLogSink } from './lib/log.js';
+export { harnessHooks } from './lib/hooks/registry.js';
+export { EgressGuardHook, SkillCatalogGuardHook } from './lib/hooks/built-in/index.js';
 const HarnessState = harness.getEnum('@harness/state');
 const BIN_VERSION_PROBE_TTL_MS = 45_000;
 const TRANSCRIBE_STREAM_MIN_BYTES = 16_000 * 2 * 60;
@@ -943,7 +943,7 @@ export class Harness extends ReadyResource {
             releaseModel(pool, key);
         }
     }
-    async *completion({ model, agent, messages, cacheKey, thinkingDefault, toolConfigs, generationParams, responseFormat, grantedDirs, grantedDrives, driveOp, grantedWriteDirs, approve, attach, fetchAttachment, ragSearch, progress, askUser, signal }) {
+    async *completion({ model, agent, messages, cacheKey, thinkingDefault, toolConfigs, generationParams, responseFormat, captureInferenceCalls, grantedDirs, grantedDrives, driveOp, grantedWriteDirs, approve, attach, fetchAttachment, ragSearch, progress, askUser, signal }) {
         await this.ready();
         if (this.suspended)
             throw new Error('harness is suspended');
@@ -1016,6 +1016,7 @@ export class Harness extends ReadyResource {
                 thinkingDefault,
                 generationParams,
                 responseFormat,
+                captureInferenceCalls,
                 messages: turns,
                 tools: this._tools,
                 skillCatalog: await this._completionSkills(agent?.id ?? null, model),

package/dist-lib/lib/audio-transcode-bare.js

@@ -1,6 +1,6 @@
 import ffmpeg from 'bare-ffmpeg';
-import { hlog } from "./log.js";
-import { buildWavBuffer, TARGET_SAMPLE_RATE } from "./wav-encode.js";
+import { hlog } from './log.js';
+import { buildWavBuffer, TARGET_SAMPLE_RATE } from './wav-encode.js';
 function bufferReadIO(input) {
     let offset = 0;
     return new ffmpeg.IOContext(Buffer.alloc(4096), {

package/dist-lib/lib/capabilities/completion.d.ts

@@ -221,8 +221,40 @@ export type MetricsEvent = {
     ctxSize?: number | undefined;
     backends?: RunBackendUse[] | undefined;
     usage?: UsageBreakdown | undefined;
+    stopReason?: string | undefined;
 };
-export type CompletionEvent = TextEvent | ToolCallEvent | ToolResultEvent | ToolActivityEvent | ErrorEvent | AbortedEvent | MetricsEvent;
+export interface InferenceCall {
+    round: number;
+    droppedTurns: number;
+    modelId: string;
+    messages: Message[];
+    tools: SdkTool[] | undefined;
+    generationParams: LlmGenerationParams | undefined;
+    responseFormat: ResponseFormat | undefined;
+    toolDialect: ToolDialect | undefined;
+    cacheKey: string | undefined;
+    promptPrefix: string;
+}
+export type InferenceCallEvent = {
+    type: 'inference-call';
+    call: InferenceCall;
+};
+export type InferenceSentEvent = {
+    type: 'inference-sent';
+    round: number;
+    droppedTurns: number;
+    messages: Message[] | undefined;
+    tokens: InferenceTokens | undefined;
+    outcome: 'completed' | 'overflowed' | 'failed';
+    reason: string | undefined;
+};
+export interface InferenceTokens {
+    prompt?: number | undefined;
+    generated?: number | undefined;
+    cached?: number | undefined;
+    timeToFirstToken?: number | undefined;
+}
+export type CompletionEvent = TextEvent | ToolCallEvent | ToolResultEvent | ToolActivityEvent | ErrorEvent | AbortedEvent | MetricsEvent | InferenceCallEvent | InferenceSentEvent;
 export interface CompletionSdkResult {
     requestId?: string;
     tokenStream: AsyncIterable<string>;
@@ -283,6 +315,7 @@ interface CompletionArgs {
     toolConfigs?: Record<string, Record<string, unknown>>;
     fsRoots?: readonly DriveRoot[];
     fsWriteRoots?: readonly DriveRoot[];
+    captureInferenceCalls?: boolean;
 }
 export declare const CONTEXT_OVERFLOW_MESSAGE = "This message is too large for the model context window.";
 export declare function trustedSourcesFrom(history: readonly Message[]): TrustedSource[];
@@ -290,5 +323,5 @@ export declare function toSdkTool(tool: ToolDefinition): SdkTool;
 export type CompletionSkillEntry = SkillCatalogEntry & {
     readonly status?: SkillState;
 };
-export declare function completion(sdk: CompletionSdk, { modelId, acquireRound, model, agent, ctxSize, cacheKey, thinkingDefault, generationParams: generationParamsOverride, responseFormat, messages: storedMessages, tools: allTools, skillCatalog, approve, attach, fetchAttachment, ragSearch, progress, askUser, signal, toolConfigs, fsRoots, fsWriteRoots }: CompletionArgs): AsyncGenerator<CompletionEvent>;
+export declare function completion(sdk: CompletionSdk, { modelId, acquireRound, model, agent, ctxSize, cacheKey, thinkingDefault, generationParams: generationParamsOverride, responseFormat, messages: storedMessages, tools: allTools, skillCatalog, approve, attach, fetchAttachment, ragSearch, progress, askUser, signal, toolConfigs, fsRoots, fsWriteRoots, captureInferenceCalls }: CompletionArgs): AsyncGenerator<CompletionEvent>;
 export {};

package/dist-lib/lib/capabilities/completion.js

@@ -1,34 +1,34 @@
 import process from 'process';
-import { contextCandidates } from "../context-window.js";
-import { errorCode, errorMessage } from "../error-fields.js";
-import { bufferBurst } from "../helpers/buffer-burst.js";
-import { endsTurn, stopAtControlToken, stripControlTokens, truncateAtControlToken } from "../helpers/control-tokens.js";
-import { createMarkupStreamFilter } from "../helpers/markup-stream-filter.js";
-import { AbortedError, raceAbort } from "../helpers/race-abort.js";
-import { createRunOutcomeTracker, degenerateAnswer, unnamedWrites, writeReport } from "../helpers/run-outcome.js";
-import { splitThinking } from "../helpers/thinking.js";
-import { clipRawCall, normalizeToolArgs, recoverMarkupToolCalls, recoverPrintedSkillLoad, recoverSkillNamedCall, stripToolCallMarkup, stripToolResponseBlocks, toolCallMarkup } from "../helpers/tool-call-markup.js";
-import { INTENT_PARAM, INTENT_PARAM_DESCRIPTION, takeToolIntent } from "../helpers/tool-intent.js";
-import { HookPipeline } from "../hooks/pipeline.js";
-import { fallbackToolLabel } from "../helpers/tool-label.js";
-import { hlog } from "../log.js";
-import { resolveLlmGenerationParams } from "../models/llm-presets.js";
-import { customToolDialect } from "../models/model-catalog.js";
-import { isContextOverflowError, isSdkModelWedgedError } from "../models/sdk-faults.js";
-import { annotateLastUserTime, skillsPromptSection, systemPromptParts } from "../prompt.js";
-import { isMetadataMethod } from "../tools/mcp/utils.js";
-import { unquoteArg } from "../tools/model-args.js";
-import { SKILL_TOOL_NAME, WEB_SEARCH_TOOL_NAME } from "./tool-names.js";
-import { promptUsageParts, usageBreakdown } from "./usage-breakdown.js";
-import { collectToolGrants, enabledTools, isSelected, skillIsActive, skillsAreSelected } from "./tool-selection.js";
-import { guardToolCall } from "../tools/outbound-guard.js";
-import { isAvailableState } from "../tools/skills/skill-state-constants.js";
-import { resolveImplicitSkill, resolveSkillMention, skillPreloadMessage } from "../tools/skills/mention.js";
-import { permittedTools } from "../tools/toolsets.js";
-import { promptPrefixHash } from "./prompt-prefix.js";
-import { logToolPrompt } from "./tool-prompt-log.js";
-import { DIRECT_CALL_ID, loadedSkillFilesFrom, loadedSkillsFrom, TOOL_CONTINUATION_NUDGE, toolResultTurn, withoutSkillLoadTurns } from "./tool-turns.js";
-import { projectVisionHistory, visionHistoryCacheKey } from "./vision-history.js";
+import { contextCandidates } from '../context-window.js';
+import { errorCode, errorMessage } from '../error-fields.js';
+import { bufferBurst } from '../helpers/buffer-burst.js';
+import { endsTurn, stopAtControlToken, stripControlTokens, truncateAtControlToken } from '../helpers/control-tokens.js';
+import { createMarkupStreamFilter } from '../helpers/markup-stream-filter.js';
+import { AbortedError, raceAbort } from '../helpers/race-abort.js';
+import { createRunOutcomeTracker, degenerateAnswer, unnamedWrites, writeReport } from '../helpers/run-outcome.js';
+import { splitThinking } from '../helpers/thinking.js';
+import { clipRawCall, normalizeToolArgs, recoverMarkupToolCalls, recoverPrintedSkillLoad, recoverSkillNamedCall, stripToolCallMarkup, stripToolResponseBlocks, toolCallMarkup } from '../helpers/tool-call-markup.js';
+import { INTENT_PARAM, INTENT_PARAM_DESCRIPTION, takeToolIntent } from '../helpers/tool-intent.js';
+import { HookPipeline } from '../hooks/pipeline.js';
+import { fallbackToolLabel } from '../helpers/tool-label.js';
+import { hlog } from '../log.js';
+import { resolveLlmGenerationParams } from '../models/llm-presets.js';
+import { customToolDialect } from '../models/model-catalog.js';
+import { isContextOverflowError, isSdkModelWedgedError } from '../models/sdk-faults.js';
+import { annotateLastUserTime, skillsPromptSection, systemPromptParts } from '../prompt.js';
+import { isMetadataMethod } from '../tools/mcp/utils.js';
+import { unquoteArg } from '../tools/model-args.js';
+import { SKILL_TOOL_NAME, WEB_SEARCH_TOOL_NAME } from './tool-names.js';
+import { promptUsageParts, usageBreakdown } from './usage-breakdown.js';
+import { collectToolGrants, enabledTools, isSelected, skillIsActive, skillsAreSelected } from './tool-selection.js';
+import { guardToolCall } from '../tools/outbound-guard.js';
+import { isAvailableState } from '../tools/skills/skill-state-constants.js';
+import { resolveImplicitSkill, resolveSkillMention, skillPreloadMessage } from '../tools/skills/mention.js';
+import { permittedTools } from '../tools/toolsets.js';
+import { promptPrefixHash } from './prompt-prefix.js';
+import { logToolPrompt } from './tool-prompt-log.js';
+import { DIRECT_CALL_ID, loadedSkillFilesFrom, loadedSkillsFrom, TOOL_CONTINUATION_NUDGE, toolResultTurn, withoutSkillLoadTurns } from './tool-turns.js';
+import { projectVisionHistory, visionHistoryCacheKey } from './vision-history.js';
 export function imageSettingsFrom(agent) {
     if (!agent)
         return undefined;
@@ -157,7 +157,8 @@ function preloadLastUser(messages, name, body) {
     }
     return messages;
 }
-function collectStats(state, stats) {
+function collectStats(state, stats, stopReason) {
+    state.stopReason = stopReason;
     if (!stats)
         return;
     if (state.ttft === undefined && stats.timeToFirstToken)
@@ -191,7 +192,8 @@ function metricsEvent(state, usage) {
         cacheTokens: state.cacheTokens,
         ...(state.ctxSize ? { ctxSize: state.ctxSize } : {}),
         ...(state.backends.length ? { backends: state.backends } : {}),
-        ...(usage ? { usage } : {})
+        ...(usage ? { usage } : {}),
+        ...(state.stopReason ? { stopReason: state.stopReason } : {})
     };
 }
 const TOOL_DENIED_MESSAGE = 'The user declined this call. Do not retry it or call it with different arguments to get ' +
@@ -463,6 +465,21 @@ async function* runRoundTokens(sdk, result, state, signal, startInThink = false)
         await events.return(null);
     }
 }
+function sentMessages(stats) {
+    const sent = stats?.sentMessages;
+    return Array.isArray(sent) ? sent : null;
+}
+function roundTokens(stats) {
+    if (!stats)
+        return undefined;
+    const tokens = {
+        prompt: stats.promptTokens,
+        generated: stats.generatedTokens,
+        cached: stats.cacheTokens,
+        timeToFirstToken: stats.timeToFirstToken
+    };
+    return Object.values(tokens).some((value) => value !== undefined) ? tokens : undefined;
+}
 const TRUNCATED_FINAL_MS = 2000;
 function truncatedFinal(final, contentText) {
     const accepted = { contentText, toolCalls: [], raw: { fullText: contentText } };
@@ -496,17 +513,35 @@ async function* runRecoveredRoundWith(sdk, args, modelId) {
         let emitted = false;
         const state = { stopToken: null, unproven: null, rawContentText: '' };
         const cacheable = candidate.history === args.history;
+        const params = {
+            modelId,
+            history: projectVisionHistory(candidate.history),
+            stream: true,
+            tools: args.tools,
+            generationParams: args.generationParams,
+            responseFormat: args.responseFormat,
+            toolDialect: args.toolDialect,
+            kvCache: cacheable ? args.cacheKey : undefined
+        };
+        if (args.capture) {
+            yield {
+                type: 'inference-call',
+                call: {
+                    round: args.round,
+                    droppedTurns: candidate.droppedTurns,
+                    modelId,
+                    messages: params.history,
+                    tools: params.tools,
+                    generationParams: params.generationParams,
+                    responseFormat: params.responseFormat,
+                    toolDialect: params.toolDialect,
+                    cacheKey: params.kvCache,
+                    promptPrefix: args.promptPrefix
+                }
+            };
+        }
         try {
-            const result = sdk.completion({
-                modelId,
-                history: projectVisionHistory(candidate.history),
-                stream: true,
-                tools: args.tools,
-                generationParams: args.generationParams,
-                ...(args.responseFormat ? { responseFormat: args.responseFormat } : {}),
-                toolDialect: args.toolDialect,
-                ...(cacheable && args.cacheKey ? { kvCache: args.cacheKey } : {})
-            });
+            const result = sdk.completion(params);
             result.final.catch(noop);
             const startInThink = args.generationParams?.reasoning_budget === -1;
             for await (const event of runRoundTokens(sdk, result, state, args.signal, startInThink)) {
@@ -518,6 +553,17 @@ async function* runRecoveredRoundWith(sdk, args, modelId) {
             const final = state.stopToken
                 ? await truncatedFinal(result.final, state.rawContentText)
                 : await raceAbort(result.final, args.signal);
+            if (args.capture) {
+                yield {
+                    type: 'inference-sent',
+                    round: args.round,
+                    droppedTurns: candidate.droppedTurns,
+                    messages: sentMessages(final.stats) ?? undefined,
+                    tokens: roundTokens(final.stats),
+                    outcome: 'completed',
+                    reason: undefined
+                };
+            }
             const replay = unprovenReplay(state, final);
             hlog('final', {
                 stopReason: final.stopReason,
@@ -535,6 +581,17 @@ async function* runRecoveredRoundWith(sdk, args, modelId) {
             };
         }
         catch (err) {
+            if (args.capture) {
+                yield {
+                    type: 'inference-sent',
+                    round: args.round,
+                    droppedTurns: candidate.droppedTurns,
+                    messages: undefined,
+                    tokens: undefined,
+                    outcome: isContextOverflowError(err) ? 'overflowed' : 'failed',
+                    reason: err instanceof Error ? err.message : String(err)
+                };
+            }
             if (!isContextOverflowError(err))
                 throw err;
             if (emitted)
@@ -551,7 +608,7 @@ function* reportUnnamedWrites(said, landed) {
     if (unreported)
         yield { type: 'content', text: unreported };
 }
-export async function* completion(sdk, { modelId, acquireRound, model, agent, ctxSize, cacheKey, thinkingDefault, generationParams: generationParamsOverride, responseFormat, messages: storedMessages, tools: allTools = [], skillCatalog = [], approve, attach, fetchAttachment, ragSearch, progress, askUser, signal, toolConfigs, fsRoots = [], fsWriteRoots = [] }) {
+export async function* completion(sdk, { modelId, acquireRound, model, agent, ctxSize, cacheKey, thinkingDefault, generationParams: generationParamsOverride, responseFormat, messages: storedMessages, tools: allTools = [], skillCatalog = [], approve, attach, fetchAttachment, ragSearch, progress, askUser, signal, toolConfigs, fsRoots = [], fsWriteRoots = [], captureInferenceCalls = false }) {
     const pipeline = new HookPipeline();
     const agentId = agent?.id ?? null;
     const tools = permittedTools(allTools, agent?.disabledToolsets);
@@ -789,7 +846,9 @@ export async function* completion(sdk, { modelId, acquireRound, model, agent, ct
         let mute = 0;
         let narrated = false;
         let cappedRounds = 0;
-        for (let round = 0; round < MAX_TOOL_ROUNDS; round++) {
+        let roundsSent = 0;
+        let round = 0;
+        for (; round < MAX_TOOL_ROUNDS; round++) {
             if (signal?.aborted)
                 throw new AbortedError();
             if (Date.now() >= runBudget.deadlineMs) {
@@ -820,13 +879,17 @@ export async function* completion(sdk, { modelId, acquireRound, model, agent, ct
                 responseFormat,
                 toolDialect,
                 cacheKey: roundCacheKey,
-                signal
+                signal,
+                round,
+                promptPrefix,
+                capture: captureInferenceCalls
             });
+            roundsSent = round + 1;
             const final = recovered.final;
             if (!recovered.cacheable)
                 roundCacheKey = undefined;
             history = recovered.history;
-            collectStats(stats, final.stats);
+            collectStats(stats, final.stats, final.stopReason);
             hlog('round stats', { round, wallMs: Date.now() - roundStarted, ...(final.stats ?? {}) });
             const markup = stripToolResponseBlocks(truncateAtControlToken(final.raw.fullText) || recovered.unproven || '');
             let toolCalls = final.toolCalls?.length
@@ -1060,9 +1123,12 @@ export async function* completion(sdk, { modelId, acquireRound, model, agent, ct
             tools: undefined,
             generationParams,
             cacheKey: roundCacheKey,
-            signal
+            signal,
+            round: roundsSent,
+            promptPrefix: promptPrefixHash({ system: systemContent, tools: undefined }),
+            capture: captureInferenceCalls
         });
-        collectStats(stats, recovered.final.stats);
+        collectStats(stats, recovered.final.stats, recovered.final.stopReason);
         const answer = recovered.replay ? stripToolCallMarkup(recovered.replay) : '';
         if (answer)
             yield { type: 'content', text: answer };

package/dist-lib/lib/capabilities/download.js

@@ -1,6 +1,6 @@
 import { Readable } from 'streamx';
-import { errorMessage } from "../error-fields.js";
-import { resolveModelSrc } from "../models/model-source.js";
+import { errorMessage } from '../error-fields.js';
+import { resolveModelSrc } from '../models/model-source.js';
 const noop = () => { };
 export function download(sdkOrLoader, args, onSettled) {
     const { model, addon = 'llm', seed, signal, projection, companions = [], assetSizes } = args;

package/dist-lib/lib/capabilities/speech-text.js

@@ -1,5 +1,5 @@
 import { ascii } from 'asenion';
-import { CURRENCY_CODES, SPEECH_LEXICON } from "./speech-lexicon.js";
+import { CURRENCY_CODES, SPEECH_LEXICON } from './speech-lexicon.js';
 const EMOJI_MODIFIERS = /[\uFE0E\uFE0F\u200D\u{1F3FB}-\u{1F3FF}]/gu;
 const UNSPEAKABLE = /[^\nA-Za-z0-9 .,;:!?'"()\u002D_]/gu;
 export function normalizeForSpeech(text, lang = 'en') {

package/dist-lib/lib/capabilities/tool-prompt-log.js

@@ -1,4 +1,4 @@
-import { harnessLogEnabled, hlog } from "../log.js";
+import { harnessLogEnabled, hlog } from '../log.js';
 export function logToolPrompt({ promptPrefix, modelId, cacheKey, agentId, messageCount, responseFormat, tools = [], definitions, activeSkills, skillsSelected, enabledSkills, mentionedSkill, loadedSkills, toolOps }) {
     if (!harnessLogEnabled())
         return;

package/dist-lib/lib/capabilities/tool-selection.js

@@ -1,5 +1,5 @@
-import { parseToolGrant } from "../tools/skills/tool-grants.js";
-import { SKILL_TOOL_NAME, WEB_SEARCH_TOOL_NAME } from "./tool-names.js";
+import { parseToolGrant } from '../tools/skills/tool-grants.js';
+import { SKILL_TOOL_NAME, WEB_SEARCH_TOOL_NAME } from './tool-names.js';
 export function collectToolGrants(skills) {
     const grants = {
         grantedNames: new Set(),

package/dist-lib/lib/capabilities/tool-turns.js

@@ -1,4 +1,4 @@
-import { SKILL_TOOL_NAME } from "./tool-names.js";
+import { SKILL_TOOL_NAME } from './tool-names.js';
 export const DIRECT_CALL_ID = 'call-direct';
 export const TOOL_CONTINUATION_NUDGE = "[Tool results above. Continue the user's request: call the next tool, or answer.]";
 export function toolResultTurn(name, id, output) {

package/dist-lib/lib/capabilities/usage-breakdown.js

@@ -1,4 +1,4 @@
-import { detectToolDialect } from "../models/tool-dialect.js";
+import { detectToolDialect } from '../models/tool-dialect.js';
 const QWEN35 = { prose: 4.38, schema: 4.08, toolBlockTokens: 189 };
 const DENSITY_BY_DIALECT = new Map([['qwen35', QWEN35]]);
 export function densityFor(modelName) {

package/dist-lib/lib/capabilities/voice.js

@@ -1,7 +1,7 @@
 import b4a from 'b4a';
-import { AbortedError } from "../helpers/race-abort.js";
-import { hlog } from "../log.js";
-import { normalizeForSpeech } from "./speech-text.js";
+import { AbortedError } from '../helpers/race-abort.js';
+import { hlog } from '../log.js';
+import { normalizeForSpeech } from './speech-text.js';
 export const SUPERTONIC_SAMPLE_RATE = 44100;
 export function createVoiceLane() {
     let tail = Promise.resolve();

package/dist-lib/lib/connect.js

@@ -1,18 +1,18 @@
 import EventEmitter from 'events';
 import fs from 'fs';
-import { settleOnClose } from "./rpc-session.js";
+import { settleOnClose } from './rpc-session.js';
 import { harnessClient } from '../spec/calls.js';
 import HRPC from '../spec/hrpc/index.js';
-import { unsupportedKvCacheDeletion } from "./capabilities/kv-cache.js";
-import { hlog } from "./log.js";
-import { assertEmbedDevice } from "./models/embed-config.js";
-import { errorCode, errorMessage } from "./error-fields.js";
-import { pushQueue } from "./queue.js";
-import { stageFile, sweepStaging } from "./staging.js";
-import { readInstall, readSetup } from "./tools/skills/manifest-parse.js";
-import { SKILL_STATE } from "./tools/skills/skill-state.js";
-import { asToolset } from "./tools/toolsets.js";
-import { fromDownloadWire, fromWire, unpackVectors } from "./wire.js";
+import { unsupportedKvCacheDeletion } from './capabilities/kv-cache.js';
+import { hlog } from './log.js';
+import { assertEmbedDevice } from './models/embed-config.js';
+import { errorCode, errorMessage } from './error-fields.js';
+import { pushQueue } from './queue.js';
+import { stageFile, sweepStaging } from './staging.js';
+import { readInstall, readSetup } from './tools/skills/manifest-parse.js';
+import { SKILL_STATE } from './tools/skills/skill-state.js';
+import { asToolset } from './tools/toolsets.js';
+import { fromDownloadWire, fromWire, unpackVectors } from './wire.js';
 export const HARNESS_DIED = 'HARNESS_DIED';
 export function connectHarness(transport) {
     return new RemoteHarness(transport);
@@ -112,6 +112,7 @@ export class RemoteHarness extends EventEmitter {
                 thinkingDefault: context.thinkingDefault,
                 generationParams: context.generationParams,
                 responseFormat: context.responseFormat,
+                captureInferenceCalls: context.captureInferenceCalls,
                 messages: context.messages.map(({ role, content, attachments }) => ({
                     role,
                     content,

package/dist-lib/lib/context-window.js

@@ -1,4 +1,4 @@
-import { TOOL_CONTINUATION_NUDGE } from "./capabilities/tool-turns.js";
+import { TOOL_CONTINUATION_NUDGE } from './capabilities/tool-turns.js';
 export const TOOL_RESULT_TRUNCATION_MARKER = '\n[tool result truncated for context]\n';
 function isTurnBoundary(message) {
     return message.role === 'user' && message.content !== TOOL_CONTINUATION_NUDGE;

package/dist-lib/lib/contract-types.d.ts

@@ -48,6 +48,7 @@ export interface CompletionContext {
     toolConfigs?: Record<string, Record<string, unknown>>;
     generationParams?: LlmGenerationParams;
     responseFormat?: ResponseFormat;
+    captureInferenceCalls?: boolean;
     grantedDirs?: string[];
     grantedDrives?: GrantedDrive[];
     driveOp?: DriveOpFn;

package/dist-lib/lib/docx/index.js

@@ -1,8 +1,8 @@
 import b4a from 'b4a';
 import fs from 'fs';
-import { runPython } from "../tools/exec/adapters/python/pyodide.js";
-import { planVendoredInstall, vendoredManifest, wheelsDir } from "../tools/exec/adapters/python/wheels.js";
-export { closePythonWorker as closeDocxWorker } from "../tools/exec/adapters/python/worker/client.js";
+import { runPython } from '../tools/exec/adapters/python/pyodide.js';
+import { planVendoredInstall, vendoredManifest, wheelsDir } from '../tools/exec/adapters/python/wheels.js';
+export { closePythonWorker as closeDocxWorker } from '../tools/exec/adapters/python/worker/client.js';
 const MAX_OFFICE_BYTES = 20 * 1024 * 1024;
 const PACKAGES = ['python-docx==1.2.0'];
 const EXTRACT = `from docx import Document