oh-my-knowledge 0.48.0 → 0.49.0

package/dist/eval-workflows/run-evaluation.js

@@ -1,19 +1,21 @@
 import { resolve } from 'node:path';
 import { DEFAULT_OUTPUT_DIR, persistReport } from '../eval-core/evaluation-reporting.js';
-import { createExecutor, DEFAULT_MODEL, JUDGE_MODEL } from '../executors/index.js';
+import { createExecutor } from '../executors/index.js';
 import { discoverBatchSkills } from '../inputs/skill-loader.js';
 import { confidenceInterval, tTest, effectSize } from '../eval-core/statistics.js';
 import { executeBatchEvaluationRuns, buildBatchVariantSpecs } from './batch-evaluation-workflow.js';
 import { buildDryRunTaskReport, prepareEvaluationRun, } from './evaluation-preparation.js';
 import { executeEvaluationPipeline } from './evaluation-pipeline.js';
+import { checkResumeCompatibility } from '../eval-core/resume-compatibility.js';
 import { findSaturationPoint } from '../analysis/saturation.js';
 import { bootstrapMeanCI, DEFAULT_BOOTSTRAP_ALPHA, DEFAULT_BOOTSTRAP_SAMPLES } from '../eval-core/bootstrap.js';
-export async function runEvaluation({ samplesPath, skillDir, variantSpecs = [], model = DEFAULT_MODEL, outputDir = DEFAULT_OUTPUT_DIR, project, owner, tags, noJudge = false, dryRun = false, concurrency = 1, timeoutMs, noCache = false, executorName = 'claude', jobStore = null, persistJob = true, onProgress = null, skipConnectivity = false, skipDoctor = false, lang = 'zh', mcpConfig, verbose = false, retry = 0, resume, runId, layeredStats = false, repeat, holdoutRatio, batch, judgeRepeat, judgeModels, bootstrap, bootstrapSamples, lengthDebias, budget, strictBaseline, effort, noDiagnostic, }) {
+import { ownRecordValue, setOwnRecordValue } from '../shared/record-count.js';
+export async function runEvaluation({ samplesPath, skillDir, variantSpecs = [], model, outputDir = DEFAULT_OUTPUT_DIR, project, owner, tags, noJudge = false, dryRun = false, concurrency = 1, timeoutMs, noCache = false, executorName, jobStore = null, persistJob = true, onProgress = null, skipConnectivity = false, skipDoctor = false, lang = 'zh', mcpConfig, verbose = false, retry = 0, resume, runId, layeredStats = false, repeat, holdoutRatio, batch, judgeRepeat, judgeModels, bootstrap, bootstrapSamples, lengthDebias, budget, strictBaseline, effort, noDiagnostic, }) {
     // Unified judgeModels → derive single-judge fields for downstream pipeline / grading
     // (which still operate on string `judgeModel` + `judgeExecutorName` fields per call).
     const effectiveJudgeModels = judgeModels && judgeModels.length > 0
         ? judgeModels
-        : [{ executor: executorName, model: JUDGE_MODEL }];
+        : [{ executor: executorName, model }];
     const judgeModel = effectiveJudgeModels[0].model;
     const judgeExecutorName = effectiveJudgeModels[0].executor;
     const { samples, artifacts: resolvedArtifacts, tasks, variantNames, requires, samplesBaseDir, samplesSourceFiles } = await prepareEvaluationRun({
@@ -66,8 +68,6 @@ export async function runEvaluation({ samplesPath, skillDir, variantSpecs = [],
             }
         }
     }
-    // --resume 时自动跳过 LLM 连通性检测(原 run 已经验过, 重跑是浪费 LLM 调用)
-    const effectiveSkipConnectivity = resume ? true : skipConnectivity;
     if (dryRun) {
         // Emit power warnings during dry-run too — this is exactly when users
         // preview the run, the right moment to flag "you might be wasting it".
@@ -75,7 +75,10 @@ export async function runEvaluation({ samplesPath, skillDir, variantSpecs = [],
         for (const w of buildPowerWarnings(samples.length, repeat ?? 1, lang)) {
             process.stderr.write(`${w}\n`);
         }
-        for (const w of buildIsolationWarnings(resolvedArtifacts, strictBaseline)) {
+        for (const w of buildIsolationWarnings(resolvedArtifacts, strictBaseline, {
+            executorName,
+            lang,
+        })) {
             process.stderr.write(`${w}\n`);
         }
         return {
@@ -101,27 +104,47 @@ export async function runEvaluation({ samplesPath, skillDir, variantSpecs = [],
         const store = createOverlayReportStore(resolve(outputDir || DEFAULT_OUTPUT_DIR), globalReportsDir());
         const existing = await store.get(resume);
         if (existing?.kind === 'evaluation') {
-            existingResults = {};
-            for (const entry of existing.results || []) {
-                existingResults[entry.sample_id] = entry.variants;
-            }
-            if (onProgress) {
-                const count = Object.values(existingResults).reduce((sum, v) => sum + Object.values(v).filter((r) => r.ok).length, 0);
-                process.stderr.write(`\n📂 resumed ${count} completed results from report ${resume}\n`);
-            }
-            // Skill isolation 一致性校验:resumed report 的 meta.skillIsolation
-            // 跟当前 strict-baseline 状态不一致时 stderr warn 不阻塞,让用户判断
-            // 是否要 --no-cache 强制重跑(避免混合污染 / 干净 entries)。
-            const reportIsolation = existing.meta?.skillIsolation;
-            const expectedIsolated = strictBaseline !== false; // default true
-            if (!reportIsolation && expectedIsolated) {
-                process.stderr.write(`\n⚠️  resumed report ${resume} 无 meta.skillIsolation 字段,baseline 可能被 ~/.claude/skills/ 污染。\n`
-                    + `   当前 --strict-baseline 默认开启,新 entries 会被隔离 → 与现有 entries 不可比。\n`
-                    + `   建议:加 --no-cache 强制重跑,或显式 --no-strict-baseline 接受混合数据(慎用)。\n`);
+            const compatibility = checkResumeCompatibility(existing, {
+                variants: variantNames,
+                model,
+                executorName,
+                effort,
+                noJudge,
+                judgeModels: effectiveJudgeModels,
+                judgeRepeat,
+                lengthDebias,
+                budget,
+                timeoutMs,
+                retry,
+                noDiagnostic,
+                skillDir,
+                samples,
+                samplesBaseDir,
+                tasks,
+                artifacts: resolvedArtifacts,
+            });
+            if (compatibility.compatible) {
+                const sourceBySample = new Map(existing.results.map((entry) => [entry.sample_id, entry.variants]));
+                const resumedResults = Object.create(null);
+                for (const task of tasks) {
+                    const result = sourceBySample.get(task.sample_id)?.[task.variant];
+                    if (!result?.ok)
+                        continue;
+                    const variants = ownRecordValue(resumedResults, task.sample_id)
+                        ?? setOwnRecordValue(resumedResults, task.sample_id, {});
+                    setOwnRecordValue(variants, task.variant, result);
+                }
+                existingResults = resumedResults;
+                const count = Object.values(resumedResults).reduce((sum, variants) => sum + Object.keys(variants).length, 0);
+                process.stderr.write(lang === 'zh'
+                    ? `\n已从报告 ${resume} 恢复 ${count} 条兼容的成功结果。\n`
+                    : `\nResumed ${count} compatible successful result(s) from report ${resume}.\n`);
             }
-            else if (reportIsolation && !expectedIsolated) {
-                process.stderr.write(`\n⚠️  resumed report ${resume} 已 strict-isolated(meta.skillIsolation 存在),但本次 --no-strict-baseline。\n`
-                    + `   新 entries 不隔离 → 与现有 entries 不可比。建议恢复默认 strict-baseline。\n`);
+            else {
+                const fields = compatibility.mismatches.join(', ');
+                process.stderr.write(lang === 'zh'
+                    ? `\n报告 ${resume} 与当前评测契约不兼容，将从头运行。差异字段：${fields}。\n`
+                    : `\nReport ${resume} is incompatible with the current evaluation contract; starting from scratch. Mismatched fields: ${fields}.\n`);
             }
         }
         else if (existing?.kind === 'batch-evaluation') {
@@ -131,6 +154,10 @@ export async function runEvaluation({ samplesPath, skillDir, variantSpecs = [],
             process.stderr.write(`\n⚠️  report ${resume} not found, starting from scratch\n`);
         }
     }
+    // 只有通过完整契约校验、真正接受恢复结果时，才能沿用原 run 的连通性结论。
+    const effectiveSkipConnectivity = existingResults !== undefined
+        ? true
+        : skipConnectivity;
     const executor = createExecutor(executorName);
     const judgeExecutor = createExecutor(judgeExecutorName || executorName);
     return executeEvaluationPipeline({
@@ -201,7 +228,7 @@ const LAYER_EXTRACTORS = {
 };
 function buildMetricStats(runs, variant, extractor) {
     const scores = runs
-        .map((run) => extractor(run.summary?.[variant]))
+        .map((run) => extractor(ownRecordValue(run.summary, variant)))
         .filter((x) => typeof x === 'number');
     if (scores.length === 0)
         return null;
@@ -236,8 +263,8 @@ function buildSaturationData(runs, bootstrapSamples = DEFAULT_BOOTSTRAP_SAMPLES,
     if (variants.length === 0)
         return undefined;
     // Per-variant: cumulative composite scores after each repeat.
-    const cumulativeByVariant = {};
-    const tracesByVariant = {};
+    const cumulativeByVariant = Object.create(null);
+    const tracesByVariant = Object.create(null);
     const checkpointSampleCounts = [];
     const acc = Object.fromEntries(variants.map((v) => [v, []]));
     for (let runIdx = 0; runIdx < runs.length; runIdx++) {
@@ -245,47 +272,48 @@ function buildSaturationData(runs, bootstrapSamples = DEFAULT_BOOTSTRAP_SAMPLES,
         for (const variant of variants) {
             const newScores = [];
             for (const entry of run.results ?? []) {
-                const v = entry.variants?.[variant];
+                const v = ownRecordValue(entry.variants, variant);
                 if (!v || typeof v.compositeScore !== 'number' || v.compositeScore <= 0)
                     continue;
                 newScores.push(v.compositeScore);
             }
-            acc[variant] = acc[variant].concat(newScores);
+            setOwnRecordValue(acc, variant, (ownRecordValue(acc, variant) ?? []).concat(newScores));
         }
         // Snapshot cumulative state for this checkpoint.
-        const checkpointN = acc[variants[0]]?.length ?? 0;
+        const checkpointN = ownRecordValue(acc, variants[0])?.length ?? 0;
         checkpointSampleCounts.push(checkpointN);
         for (const variant of variants) {
-            if (!cumulativeByVariant[variant])
-                cumulativeByVariant[variant] = [];
-            cumulativeByVariant[variant].push([...acc[variant]]);
+            const cumulative = ownRecordValue(cumulativeByVariant, variant)
+                ?? setOwnRecordValue(cumulativeByVariant, variant, []);
+            cumulative.push([...(ownRecordValue(acc, variant) ?? [])]);
             // Per-checkpoint trace: bootstrap CI on cumulative scores.
-            const ci = bootstrapMeanCI(acc[variant], DEFAULT_BOOTSTRAP_ALPHA, bootstrapSamples, seed);
-            if (!tracesByVariant[variant])
-                tracesByVariant[variant] = [];
-            tracesByVariant[variant].push({
-                n: acc[variant].length,
+            const scores = ownRecordValue(acc, variant) ?? [];
+            const ci = bootstrapMeanCI(scores, DEFAULT_BOOTSTRAP_ALPHA, bootstrapSamples, seed);
+            const traces = ownRecordValue(tracesByVariant, variant)
+                ?? setOwnRecordValue(tracesByVariant, variant, []);
+            traces.push({
+                n: scores.length,
                 mean: ci.estimate,
                 ciLow: ci.low,
                 ciHigh: ci.high,
             });
         }
     }
-    const verdicts = {};
+    const verdicts = Object.create(null);
     if (runs.length >= 5) {
         for (const variant of variants) {
-            const cumulative = cumulativeByVariant[variant];
+            const cumulative = ownRecordValue(cumulativeByVariant, variant);
             if (!cumulative)
                 continue;
             const r = findSaturationPoint(cumulative, 'bootstrap-ci-width', undefined, undefined, bootstrapSamples, seed);
-            verdicts[variant] = {
+            setOwnRecordValue(verdicts, variant, {
                 saturated: r.saturated,
                 atN: r.atN,
                 confidence: r.confidence,
                 method: r.method,
                 threshold: r.threshold,
                 reason: r.reason,
-            };
+            });
         }
     }
     return {
@@ -299,7 +327,7 @@ export function buildVarianceData(runs, bootstrapSamples = DEFAULT_BOOTSTRAP_SAM
         return null;
     }
     const variants = runs[0].meta.variants || [];
-    const perVariant = {};
+    const perVariant = Object.create(null);
     for (const variant of variants) {
         // Composite lives on the legacy flat fields.
         const composite = buildMetricStats(runs, variant, COMPOSITE_EXTRACTOR);
@@ -319,17 +347,17 @@ export function buildVarianceData(runs, bootstrapSamples = DEFAULT_BOOTSTRAP_SAM
             if (layerStats)
                 byLayer[key] = layerStats;
         }
-        perVariant[variant] = {
+        setOwnRecordValue(perVariant, variant, {
             ...composite,
             ...(Object.keys(byMetric).length > 0 ? { byMetric } : {}),
             ...(Object.keys(byLayer).length > 0 ? { byLayer } : {}),
-        };
+        });
     }
     const comparisons = [];
     for (let i = 0; i < variants.length; i++) {
         for (let j = i + 1; j < variants.length; j++) {
-            const vA = perVariant[variants[i]];
-            const vB = perVariant[variants[j]];
+            const vA = ownRecordValue(perVariant, variants[i]);
+            const vB = ownRecordValue(perVariant, variants[j]);
             if (!vA || !vB)
                 continue;
             const compositeComp = buildComparisonMetric(vA.scores, vB.scores, vA.mean, vB.mean);
@@ -367,12 +395,12 @@ export function buildVarianceData(runs, bootstrapSamples = DEFAULT_BOOTSTRAP_SAM
         ...(saturation ? { saturation } : {}),
     };
 }
-export async function runBatchEvaluation({ skillDir, model = DEFAULT_MODEL, outputDir = DEFAULT_OUTPUT_DIR, project, owner, tags, noJudge = false, dryRun = false, concurrency = 1, timeoutMs, executorName = 'claude', jobStore = null, persistJob = true, onProgress = null, onSkillProgress = null, skipConnectivity = false, skipDoctor = false, lang = 'zh', mcpConfig, verbose = false, repeat, holdoutRatio, judgeRepeat, judgeModels, bootstrap, bootstrapSamples, lengthDebias, budget, noCache = false, strictBaseline, variantAllowedSkills, }) {
+export async function runBatchEvaluation({ skillDir, model, outputDir = DEFAULT_OUTPUT_DIR, project, owner, tags, noJudge = false, dryRun = false, concurrency = 1, timeoutMs, executorName, jobStore = null, persistJob = true, onProgress = null, onSkillProgress = null, skipConnectivity = false, skipDoctor = false, lang = 'zh', mcpConfig, verbose = false, repeat, holdoutRatio, judgeRepeat, judgeModels, bootstrap, bootstrapSamples, lengthDebias, budget, noCache = false, strictBaseline, variantAllowedSkills, }) {
     // Same unified judge derivation as runEvaluation (downstream pipeline / report build
     // still uses single judgeModel + judgeExecutorName per call).
     const effectiveJudgeModels = judgeModels && judgeModels.length > 0
         ? judgeModels
-        : [{ executor: executorName, model: JUDGE_MODEL }];
+        : [{ executor: executorName, model }];
     const judgeModel = effectiveJudgeModels[0].model;
     const judgeExecutorName = effectiveJudgeModels[0].executor;
     const skillEntries = discoverBatchSkills(resolve(skillDir));

package/dist/executors/anthropic-api.js

@@ -1,4 +1,5 @@
-import { asErrorLike, DEFAULT_TIMEOUT_MS, errorMessage } from './shared.js';
+import { asErrorLike, DEFAULT_TIMEOUT_MS, errorMessage, readJsonResponse, responseBodyPreview, } from './shared.js';
+import { optionalTokenCount } from '../shared/token-usage.js';
 export async function anthropicApiExecutor({ model, system, prompt, timeoutMs = DEFAULT_TIMEOUT_MS }) {
     const apiKey = process.env.ANTHROPIC_API_KEY;
     if (!apiKey)
@@ -19,17 +20,72 @@ export async function anthropicApiExecutor({ model, system, prompt, timeoutMs =
             body: JSON.stringify(reqBody),
             signal: AbortSignal.timeout(timeoutMs),
         });
-        const data = await res.json();
+        const { data, rawBody } = await readJsonResponse(res);
         const durationMs = Date.now() - start;
         if (!res.ok) {
-            return { ok: false, error: data.error?.message || `API error ${res.status}`, durationMs, durationApiMs: 0, inputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheCreationTokens: 0, costUSD: 0, output: null, stopReason: 'error', numTurns: 0 };
+            const bodyPreview = responseBodyPreview(rawBody);
+            return { ok: false, error: data?.error?.message || `API error ${res.status}${bodyPreview ? `: ${bodyPreview}` : ''}`, durationMs, durationApiMs: 0, inputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheCreationTokens: 0, tokenUsageReportedByExecutor: false, costUSD: 0, costReportedByExecutor: false, output: null, stopReason: 'error', numTurns: 0 };
+        }
+        if (!data) {
+            return { ok: false, error: 'Anthropic API returned an empty or non-JSON response', durationMs, durationApiMs: 0, inputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheCreationTokens: 0, tokenUsageReportedByExecutor: false, costUSD: 0, costReportedByExecutor: false, output: null, stopReason: 'error', numTurns: 0 };
+        }
+        const usage = data.usage;
+        const inputTokens = optionalTokenCount(usage?.input_tokens);
+        const outputTokens = optionalTokenCount(usage?.output_tokens);
+        const cacheReadTokens = usage?.cache_read_input_tokens === undefined
+            ? 0
+            : optionalTokenCount(usage.cache_read_input_tokens);
+        const cacheCreationTokens = usage?.cache_creation_input_tokens === undefined
+            ? 0
+            : optionalTokenCount(usage.cache_creation_input_tokens);
+        if (inputTokens === undefined
+            || outputTokens === undefined
+            || cacheReadTokens === undefined
+            || cacheCreationTokens === undefined) {
+            return {
+                ok: false,
+                error: 'Anthropic response contained missing or invalid token usage',
+                durationMs,
+                durationApiMs: 0,
+                inputTokens: 0,
+                outputTokens: 0,
+                cacheReadTokens: 0,
+                cacheCreationTokens: 0,
+                tokenUsageReportedByExecutor: false,
+                costUSD: 0,
+                costReportedByExecutor: false,
+                output: null,
+                stopReason: 'error',
+                numTurns: 1,
+            };
+        }
+        const output = data.content
+            ?.filter((block) => block.type === undefined || block.type === 'text')
+            .map((block) => block.text || '')
+            .join('') ?? '';
+        if (!output.trim()) {
+            return {
+                ok: false,
+                error: 'Anthropic response did not contain assistant text',
+                durationMs,
+                durationApiMs: 0,
+                inputTokens,
+                outputTokens,
+                cacheReadTokens,
+                cacheCreationTokens,
+                costUSD: 0,
+                costReportedByExecutor: false,
+                output: null,
+                stopReason: 'error',
+                numTurns: 1,
+            };
         }
-        const usage = data.usage || {};
         return {
-            ok: true, output: data.content?.map((c) => c.text || '').join('') || '', durationMs, durationApiMs: 0,
-            inputTokens: usage.input_tokens || 0, outputTokens: usage.output_tokens || 0,
-            cacheReadTokens: usage.cache_read_input_tokens || 0, cacheCreationTokens: usage.cache_creation_input_tokens || 0,
-            costUSD: 0, stopReason: data.stop_reason || 'end_turn', numTurns: 1,
+            ok: true, output, durationMs, durationApiMs: 0,
+            inputTokens, outputTokens,
+            cacheReadTokens, cacheCreationTokens,
+            costUSD: 0, costReportedByExecutor: false,
+            stopReason: data.stop_reason || 'end_turn', numTurns: 1,
         };
     }
     catch (err) {
@@ -37,6 +93,6 @@ export async function anthropicApiExecutor({ model, system, prompt, timeoutMs =
         const details = asErrorLike(err);
         const stopReason = details.name === 'TimeoutError' ? 'timeout' : 'error';
         const error = details.name === 'TimeoutError' ? `API request timed out after ${timeoutMs / 1000}s` : errorMessage(err);
-        return { ok: false, error, durationMs, durationApiMs: 0, inputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheCreationTokens: 0, costUSD: 0, output: null, stopReason, numTurns: 0 };
+        return { ok: false, error, durationMs, durationApiMs: 0, inputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheCreationTokens: 0, tokenUsageReportedByExecutor: false, costUSD: 0, costReportedByExecutor: false, output: null, stopReason, numTurns: 0 };
     }
 }

package/dist/executors/claude-cli.js

@@ -1,6 +1,6 @@
-import { extractAgentTrace, isClaudeSdkResultMessage } from './claude-sdk-trace.js';
 import { buildExecEnv, DEFAULT_TIMEOUT_MS, errorMessage, interruptedExecResult, MAX_BUFFER, spawnWithSigintPropagation, timeoutExecResult, } from './shared.js';
 import { materializeForCliConfigDir } from '../eval-core/mocks-runtime.js';
+import { buildClaudeResult, parseClaudeStreamJson } from './claude-protocol.js';
 // claude CLI 用 `--disable-slash-commands` (文档:"Disable all skills") +
 // `--disallowedTools Skill` 实现与 SDK 等价的完全隔离。非空 skill 白名单不再支持
 // (它无法真正隔离,见 claude-sdk.ts buildSdkIsolationOptions),所以 [name1, ...] 必须 throw。
@@ -19,19 +19,6 @@ function applySkillIsolationToCliArgs(args, allowedSkills) {
     // 完全隔离:双堵 main session skill 发现 + subagent Skill 工具
     args.push('--disable-slash-commands', '--disallowedTools', 'Skill');
 }
-function parseStreamJson(stdout) {
-    const messages = [];
-    for (const line of stdout.split('\n')) {
-        const trimmed = line.trim();
-        if (!trimmed)
-            continue;
-        try {
-            messages.push(JSON.parse(trimmed));
-        }
-        catch { /* skip non-JSON lines */ }
-    }
-    return messages;
-}
 export async function claudeCliExecutor({ model, system, prompt, cwd, skillDir, timeoutMs = DEFAULT_TIMEOUT_MS, allowedSkills, mocks, mocksBaseDir, mocksStrict, lean, effort }) {
     const args = ['-p', prompt, '--output-format', 'stream-json', '--verbose', '--model', model,
         // 评测必须 bypass permission,否则 Bash / Edit / Write 等工具调用会卡在交互式确认。
@@ -86,40 +73,15 @@ export async function claudeCliExecutor({ model, system, prompt, cwd, skillDir,
         child.stdin?.end();
         const { stdout } = await done;
         const durationMs = Date.now() - start;
-        const messages = parseStreamJson(stdout);
-        // 提取 result 消息
-        const resultMsgs = messages.filter(isClaudeSdkResultMessage);
-        if (resultMsgs.length === 0) {
-            const ms = captureMockStats();
-            return {
-                ok: false, error: 'no result message in stream-json output',
-                durationMs, durationApiMs: 0,
-                inputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheCreationTokens: 0,
-                costUSD: 0, output: null, stopReason: 'error', numTurns: 0,
-                ...(ms && { mockStats: ms }),
-            };
-        }
-        const last = resultMsgs[resultMsgs.length - 1];
-        const usage = last.usage || {};
-        // 提取 trace
-        const trace = extractAgentTrace(messages);
+        const parsed = parseClaudeStreamJson(stdout);
         const ms = captureMockStats();
         return {
-            ok: !last.errors?.length && last.subtype !== 'error',
-            durationMs: last.duration_ms || durationMs,
-            durationApiMs: last.duration_api_ms || 0,
-            inputTokens: usage.input_tokens || 0,
-            outputTokens: usage.output_tokens || 0,
-            cacheReadTokens: usage.cache_read_input_tokens || 0,
-            cacheCreationTokens: usage.cache_creation_input_tokens || 0,
-            costUSD: last.total_cost_usd || 0,
-            output: last.result || '',
-            stopReason: last.subtype || 'unknown',
-            numTurns: last.num_turns || 1,
-            fullNumTurns: trace.fullNumTurns,
-            numSubAgents: trace.numSubAgents,
-            ...(trace.turns.length > 0 && { turns: trace.turns }),
-            ...(trace.toolCalls.length > 0 && { toolCalls: trace.toolCalls }),
+            ...buildClaudeResult({
+                messages: parsed.messages,
+                malformedLineCount: parsed.malformedLineCount,
+                wallClockDurationMs: durationMs,
+                source: 'claude stream-json',
+            }),
             ...(ms && { mockStats: ms }),
         };
     }
@@ -134,41 +96,16 @@ export async function claudeCliExecutor({ model, system, prompt, cwd, skillDir,
             const ms = captureMockStats();
             return { ...interruptedExecResult(durationMs), ...(ms && { mockStats: ms }) };
         }
-        // 尝试从 stdout 解析 stream-json(即使进程退出码非 0 也可能有 result)
-        const messages = parseStreamJson(details.stdout || '');
-        const resultMsgs = messages.filter(isClaudeSdkResultMessage);
-        if (resultMsgs.length > 0) {
-            const last = resultMsgs[resultMsgs.length - 1];
-            const usage = last.usage || {};
-            const trace = extractAgentTrace(messages);
-            const ms = captureMockStats();
-            return {
-                ok: false,
-                error: last.errors?.join('; ') || last.result || errorMessage(err),
-                durationMs: last.duration_ms || durationMs,
-                durationApiMs: last.duration_api_ms || 0,
-                inputTokens: usage.input_tokens || 0,
-                outputTokens: usage.output_tokens || 0,
-                cacheReadTokens: usage.cache_read_input_tokens || 0,
-                cacheCreationTokens: usage.cache_creation_input_tokens || 0,
-                costUSD: last.total_cost_usd || 0,
-                output: last.result || null,
-                stopReason: 'error',
-                numTurns: last.num_turns || 0,
-                fullNumTurns: trace.fullNumTurns,
-                numSubAgents: trace.numSubAgents,
-                ...(trace.turns.length > 0 && { turns: trace.turns }),
-                ...(trace.toolCalls.length > 0 && { toolCalls: trace.toolCalls }),
-                ...(ms && { mockStats: ms }),
-            };
-        }
+        const parsed = parseClaudeStreamJson(details.stdout || '');
         const ms = captureMockStats();
         return {
-            ok: false,
-            error: errorMessage(err),
-            durationMs, durationApiMs: 0,
-            inputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheCreationTokens: 0,
-            costUSD: 0, output: null, stopReason: 'error', numTurns: 0,
+            ...buildClaudeResult({
+                messages: parsed.messages,
+                malformedLineCount: parsed.malformedLineCount,
+                wallClockDurationMs: durationMs,
+                source: 'claude stream-json',
+                forcedError: details.stderr?.trim() || errorMessage(err),
+            }),
             ...(ms && { mockStats: ms }),
         };
     }

package/dist/executors/claude-protocol.d.ts

@@ -0,0 +1,28 @@
+import type { ExecResult } from '../types/index.js';
+import type { ClaudeSdkBaseMessage, ClaudeSdkResultMessage } from './shared.js';
+export interface ClaudeSdkMeasurements {
+    durationMs: number;
+    durationApiMs: number;
+    inputTokens: number;
+    outputTokens: number;
+    cacheReadTokens: number;
+    cacheCreationTokens: number;
+    costUSD: number;
+    numTurns: number;
+}
+export declare function normalizeClaudeSdkMeasurements(result: ClaudeSdkResultMessage): ClaudeSdkMeasurements | {
+    error: string;
+};
+export interface ClaudeStreamParseResult {
+    messages: ClaudeSdkBaseMessage[];
+    malformedLineCount: number;
+}
+export declare function parseClaudeStreamJson(stdout: string): ClaudeStreamParseResult;
+export declare function buildClaudeResult(options: {
+    messages: ClaudeSdkBaseMessage[];
+    wallClockDurationMs: number;
+    source: 'claude stream-json' | 'claude-sdk';
+    malformedLineCount?: number;
+    forcedError?: string;
+    messageTimestamps?: number[];
+}): ExecResult;