@qvac/harness 0.2.8 → 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

@@ -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/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

@@ -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/connect.js

@@ -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/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/serve.js

@@ -414,6 +414,7 @@ function completion(harness) {
                 thinkingDefault: frame.thinkingDefault ?? undefined,
                 generationParams: frame.generationParams ?? undefined,
                 responseFormat: frame.responseFormat ?? undefined,
+                captureInferenceCalls: frame.captureInferenceCalls ?? undefined,
                 messages: (frame.messages ?? []).map(({ role, content, attachments }) => ({
                     role,
                     content,

package/dist-lib/lib/wire.js

@@ -32,6 +32,20 @@ export function toWire(event) {
             const { type, ...metrics } = event;
             return { type, metrics };
         }
+        case 'inference-call':
+            return { type: 'inference-call', payload: event.call };
+        case 'inference-sent':
+            return {
+                type: 'inference-sent',
+                payload: {
+                    round: event.round,
+                    droppedTurns: event.droppedTurns,
+                    messages: event.messages,
+                    tokens: event.tokens,
+                    outcome: event.outcome,
+                    reason: event.reason
+                }
+            };
         case 'error':
             return { type: 'error', message: event.message };
         case 'aborted':
@@ -73,6 +87,20 @@ export function fromWire(event) {
             };
         case 'metrics':
             return { type: 'metrics', ...event.metrics };
+        case 'inference-call':
+            return { type: 'inference-call', call: event.payload };
+        case 'inference-sent': {
+            const sent = (event.payload ?? {});
+            return {
+                type: 'inference-sent',
+                messages: sent.messages,
+                round: Number(sent.round ?? 0),
+                droppedTurns: Number(sent.droppedTurns ?? 0),
+                tokens: sent.tokens,
+                outcome: sent.outcome || 'completed',
+                reason: sent.reason || undefined
+            };
+        }
         case 'error':
             return { type: 'error', message: event.message ?? 'harness error' };
         case 'aborted':

package/dist-lib/schema/harness.js

@@ -107,7 +107,8 @@ export function registerHarnessTypes(ns) {
             { name: 'grantedDirs', type: 'string', array: true },
             { name: 'grantedDrives', type: '@harness/granted-drive', array: true },
             { name: 'payload', type: 'json' },
-            { name: 'grantedWriteDirs', type: 'string', array: true }
+            { name: 'grantedWriteDirs', type: 'string', array: true },
+            { name: 'captureInferenceCalls', type: 'bool' }
         ]
     });
     ns.register({
@@ -128,7 +129,9 @@ export function registerHarnessTypes(ns) {
             'question-ask',
             'done',
             'tool-activity',
-            'drive-ask'
+            'drive-ask',
+            'inference-call',
+            'inference-sent'
         ],
         strings: true
     });

package/dist-lib/spec/hrpc/messages.js

@@ -1,12 +1,12 @@
 // This file is autogenerated by the hyperschema compiler
-// Schema Version: 8
+// Schema Version: 9
 /* eslint-disable camelcase */
 /* eslint-disable quotes */
 /* eslint-disable space-before-function-paren */
 
 import { c } from 'hyperschema/runtime'
 
-const VERSION = 8
+const VERSION = 9
 
 // eslint-disable-next-line no-unused-vars
 let version = VERSION
@@ -478,7 +478,8 @@ const encoding11 = {
       ((version >= 5 && m.grantedDirs) ? 2097152 : 0) |
       ((version >= 5 && m.grantedDrives) ? 4194304 : 0) |
       ((version >= 5 && m.payload) ? 8388608 : 0) |
-      ((version >= 5 && m.grantedWriteDirs) ? 16777216 : 0)
+      ((version >= 5 && m.grantedWriteDirs) ? 16777216 : 0) |
+      ((version >= 9 && m.captureInferenceCalls) ? 33554432 : 0)
 
     encoding10.preencode(state, m.type)
     c.uint.preencode(state, flags)
@@ -533,7 +534,8 @@ const encoding11 = {
       ((version >= 5 && m.grantedDirs) ? 2097152 : 0) |
       ((version >= 5 && m.grantedDrives) ? 4194304 : 0) |
       ((version >= 5 && m.payload) ? 8388608 : 0) |
-      ((version >= 5 && m.grantedWriteDirs) ? 16777216 : 0)
+      ((version >= 5 && m.grantedWriteDirs) ? 16777216 : 0) |
+      ((version >= 9 && m.captureInferenceCalls) ? 33554432 : 0)
 
     encoding10.encode(state, m.type)
     c.uint.encode(state, flags)
@@ -592,7 +594,8 @@ const encoding11 = {
       grantedDirs: (version >= 5 && (flags & 2097152) !== 0) ? encoding11_22.decode(state) : null,
       grantedDrives: (version >= 5 && (flags & 4194304) !== 0) ? encoding11_23.decode(state) : null,
       payload: (version >= 5 && (flags & 8388608) !== 0) ? c.json.decode(state) : null,
-      grantedWriteDirs: (version >= 5 && (flags & 16777216) !== 0) ? encoding11_25.decode(state) : null
+      grantedWriteDirs: (version >= 5 && (flags & 16777216) !== 0) ? encoding11_25.decode(state) : null,
+      captureInferenceCalls: (version >= 9 && (flags & 33554432) !== 0)
     }
   }
 }
@@ -613,13 +616,15 @@ const encoding12_enum = {
   "question-ask": 'question-ask',
   done: 'done',
   "tool-activity": 'tool-activity',
-  "drive-ask": 'drive-ask'
+  "drive-ask": 'drive-ask',
+  "inference-call": 'inference-call',
+  "inference-sent": 'inference-sent'
 }
 
 // @harness/completion-event-type enum
 const encoding12 = {
   preencode (state, m) {
-    state.end++ // max enum is 16 so always one byte
+    state.end++ // max enum is 18 so always one byte
   },
   encode (state, m) {
     switch (m) {
@@ -671,6 +676,12 @@ const encoding12 = {
       case 'drive-ask':
         c.uint.encode(state, 16)
         break
+      case 'inference-call':
+        c.uint.encode(state, 17)
+        break
+      case 'inference-sent':
+        c.uint.encode(state, 18)
+        break
       default:
         throw new Error('Unknown enum')
     }
@@ -709,6 +720,10 @@ const encoding12 = {
         return 'tool-activity'
       case 16:
         return 'drive-ask'
+      case 17:
+        return 'inference-call'
+      case 18:
+        return 'inference-sent'
       default: return null
     }
   }

package/dist-lib/spec/hyperschema/index.js

@@ -1,12 +1,12 @@
 // This file is autogenerated by the hyperschema compiler
-// Schema Version: 8
+// Schema Version: 9
 /* eslint-disable camelcase */
 /* eslint-disable quotes */
 /* eslint-disable space-before-function-paren */
 
 import { c } from 'hyperschema/runtime'
 
-const VERSION = 8
+const VERSION = 9
 
 // eslint-disable-next-line no-unused-vars
 let version = VERSION
@@ -478,7 +478,8 @@ const encoding11 = {
       ((version >= 5 && m.grantedDirs) ? 2097152 : 0) |
       ((version >= 5 && m.grantedDrives) ? 4194304 : 0) |
       ((version >= 5 && m.payload) ? 8388608 : 0) |
-      ((version >= 5 && m.grantedWriteDirs) ? 16777216 : 0)
+      ((version >= 5 && m.grantedWriteDirs) ? 16777216 : 0) |
+      ((version >= 9 && m.captureInferenceCalls) ? 33554432 : 0)
 
     encoding10.preencode(state, m.type)
     c.uint.preencode(state, flags)
@@ -533,7 +534,8 @@ const encoding11 = {
       ((version >= 5 && m.grantedDirs) ? 2097152 : 0) |
       ((version >= 5 && m.grantedDrives) ? 4194304 : 0) |
       ((version >= 5 && m.payload) ? 8388608 : 0) |
-      ((version >= 5 && m.grantedWriteDirs) ? 16777216 : 0)
+      ((version >= 5 && m.grantedWriteDirs) ? 16777216 : 0) |
+      ((version >= 9 && m.captureInferenceCalls) ? 33554432 : 0)
 
     encoding10.encode(state, m.type)
     c.uint.encode(state, flags)
@@ -592,7 +594,8 @@ const encoding11 = {
       grantedDirs: (version >= 5 && (flags & 2097152) !== 0) ? encoding11_22.decode(state) : null,
       grantedDrives: (version >= 5 && (flags & 4194304) !== 0) ? encoding11_23.decode(state) : null,
       payload: (version >= 5 && (flags & 8388608) !== 0) ? c.json.decode(state) : null,
-      grantedWriteDirs: (version >= 5 && (flags & 16777216) !== 0) ? encoding11_25.decode(state) : null
+      grantedWriteDirs: (version >= 5 && (flags & 16777216) !== 0) ? encoding11_25.decode(state) : null,
+      captureInferenceCalls: (version >= 9 && (flags & 33554432) !== 0)
     }
   }
 }
@@ -613,13 +616,15 @@ const encoding12_enum = {
   "question-ask": 'question-ask',
   done: 'done',
   "tool-activity": 'tool-activity',
-  "drive-ask": 'drive-ask'
+  "drive-ask": 'drive-ask',
+  "inference-call": 'inference-call',
+  "inference-sent": 'inference-sent'
 }
 
 // @harness/completion-event-type enum
 const encoding12 = {
   preencode (state, m) {
-    state.end++ // max enum is 16 so always one byte
+    state.end++ // max enum is 18 so always one byte
   },
   encode (state, m) {
     switch (m) {
@@ -671,6 +676,12 @@ const encoding12 = {
       case 'drive-ask':
         c.uint.encode(state, 16)
         break
+      case 'inference-call':
+        c.uint.encode(state, 17)
+        break
+      case 'inference-sent':
+        c.uint.encode(state, 18)
+        break
       default:
         throw new Error('Unknown enum')
     }
@@ -709,6 +720,10 @@ const encoding12 = {
         return 'tool-activity'
       case 16:
         return 'drive-ask'
+      case 17:
+        return 'inference-call'
+      case 18:
+        return 'inference-sent'
       default: return null
     }
   }

package/dist-lib/spec/hyperschema/schema.json

@@ -1,5 +1,5 @@
 {
-  "version": 8,
+  "version": 9,
   "schema": [
     {
       "name": "empty",
@@ -446,6 +446,11 @@
           "array": true,
           "type": "string",
           "version": 5
+        },
+        {
+          "name": "captureInferenceCalls",
+          "type": "bool",
+          "version": 9
         }
       ]
     },
@@ -517,6 +522,14 @@
         {
           "key": "drive-ask",
           "version": 5
+        },
+        {
+          "key": "inference-call",
+          "version": 9
+        },
+        {
+          "key": "inference-sent",
+          "version": 9
         }
       ],
       "strings": true

package/dist-lib/spec/hyperschema/types.d.ts

@@ -1,5 +1,5 @@
 // This file is autogenerated by hyperschema-ts
-// Schema version: 8
+// Schema version: 9
 // Do not edit manually.
 
 export interface HarnessEmpty {
@@ -86,9 +86,10 @@ export interface HarnessCompletionFrame {
   grantedDrives?: HarnessGrantedDrive[] | null
   payload?: unknown | null
   grantedWriteDirs?: string[] | null
+  captureInferenceCalls?: boolean
 }
 
-export type HarnessCompletionEventType = "content" | "thinking" | "tool-call" | "tool-result" | "approval-ask" | "metrics" | "error" | "aborted" | "attach-ask" | "fetch-ask" | "progress" | "rag-ask" | "question-ask" | "done" | "tool-activity" | "drive-ask"
+export type HarnessCompletionEventType = "content" | "thinking" | "tool-call" | "tool-result" | "approval-ask" | "metrics" | "error" | "aborted" | "attach-ask" | "fetch-ask" | "progress" | "rag-ask" | "question-ask" | "done" | "tool-activity" | "drive-ask" | "inference-call" | "inference-sent"
 
 export interface HarnessCompletionEvent {
   type: HarnessCompletionEventType

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@qvac/harness",
-  "version": "0.2.8",
+  "version": "0.2.9",
   "type": "module",
   "description": "An AI harness implementing generate(context) via @qvac/sdk. Knows nothing about workbench, hyperdb, or chunks — just inference in, events out.",
   "exports": {
@@ -555,9 +555,8 @@
   },
   "scripts": {
     "build:schema": "bare schema/build.ts",
-    "test": "npm run vendor:wheels && npm run test:node && npm run test:bare && npm run test:addons",
-    "test:node": "QVAC_NO_SKILL_BIN_PROBE=1 brittle-node test/all.ts | prettytap",
-    "test:bare": "QVAC_NO_SKILL_BIN_PROBE=1 brittle-bare test/all.ts | prettytap",
+    "test": "npm run vendor:wheels && npm run test:bare && npm run test:addons",
+    "test:bare": "QVAC_NO_SKILL_BIN_PROBE=1 prettytap -- brittle-bare test/all.ts",
     "format": "prettier . --write",
     "lint": "prettier . --check && lunte && tsc --noEmit",
     "build:bundle": "node scripts/build-bundle.mjs",
@@ -714,7 +713,7 @@
     "lunte": "^1.8.3",
     "prettier": "^3.6.2",
     "prettier-config-holepunch": "^2.0.0",
-    "prettytap": "^0.1.3",
+    "prettytap": "^0.2.0",
     "terser": "^5.51.2",
     "typescript": "^7.0.2"
   },
@@ -724,7 +723,8 @@
     "localwatch@1.4.3": "patches/localwatch@1.4.3.patch",
     "hypercore-blob-server@1.15.0": "patches/hypercore-blob-server@1.15.0.patch",
     "bare-stow@0.1.5": "patches/bare-stow@0.1.5.patch",
-    "@qvac/inference@0.19.0": "patches/@qvac%2Finference@0.19.0.patch"
+    "@qvac/inference@0.19.0": "patches/@qvac%2Finference@0.19.0.patch",
+    "brittle@4.1.0": "patches/brittle@4.1.0.patch"
   },
   "repository": {
     "type": "git",