@mastra/code-sdk 1.0.3-alpha.1 → 1.1.0-alpha.2

package/dist/evals/scorers/outcome.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../../src/evals/scorers/outcome.ts"],"sourcesContent":["/**\n * MastraCode Outcome Scorer (Code-based, Agent type, always-on)\n *\n * Grades the objective outcome of a coding session by checking hard facts:\n * - Did the build/typecheck pass?\n * - Did tests pass?\n * - What was the tool error rate?\n * - Did the agent get stuck in a loop?\n * - Was there a regression (build passed earlier, failed later)?\n * - Did the agent use ask_user unnecessarily? (autonomy)\n *\n * This scorer is always-on (no sampling) because it's cheap — it only\n * reads exit codes and error states from tool results already in the messages.\n *\n * Scoring philosophy (from Anthropic's eval guidance):\n * \"Grade outcomes, not paths\" — we check what happened, not how.\n * Partial credit: a session that passes build but fails tests scores\n * higher than one that fails build entirely.\n *\n * Dimensions that don't apply to a session (e.g. build for a docs-only task)\n * are excluded from the weighted average, so sessions aren't penalized for\n * irrelevant criteria.\n */\n\nimport type { MastraDBMessage } from '@mastra/core/agent';\nimport { createScorer, filterRun } from '@mastra/core/evals';\n\nimport { isBuildCommand, isTestCommand, getExitCode } from './classify-command.js';\nimport type { ExtractedToolCall } from './extract-tools.js';\nimport { extractToolCalls } from './extract-tools.js';\n\n// ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n// CONFIGURATION — Adjust weights and thresholds here\n// ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n\n/**\n * Dimension weights. Used to compute the weighted average of applicable dimensions.\n * Dimensions that don't apply (e.g. no build ran) are excluded from the average.\n */\nconst WEIGHTS = {\n  build: 0.3,\n  tests: 0.25,\n  toolErrors: 0.2,\n  loops: 0.1,\n  regression: 0.1,\n  autonomy: 0.05,\n} as const;\n\n/**\n * Scoring thresholds and multipliers.\n */\nconst THRESHOLDS = {\n  /** Multiplier for tool error rate penalty. Score = max(0, 1 - rate × this). */\n  toolErrorPenaltyMultiplier: 1.0,\n\n  /** How many consecutive identical calls (or same-error repeats) before it's a \"loop\". */\n  loopMinRepetitions: 3,\n\n  /** Score penalty per detected loop. */\n  loopPenaltyPerOccurrence: 0.3,\n\n  /** Score deducted per ask_user call. */\n  autonomyPenaltyPerAsk: 0.25,\n\n  /** Score when a command ran but exit code is ambiguous. */\n  ambiguousExitScore: 0.75,\n\n  /** Minimum tool calls required to produce a meaningful score. */\n  minToolCalls: 1,\n\n  /** Tool names whose errors are expected/benign (not penalized). */\n  benignErrorTools: ['search_content', 'find_files', 'lsp_inspect'] as string[],\n} as const;\n\n// ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n\ntype DimensionResult = { score: number; detail: string; applicable: boolean };\n\n/**\n * Check if the primary verb of an execute_command is a build/typecheck command.\n * Splits compound commands (&&, ;) and checks the first segment.\n */\n/** Extract command string from tool call args, then delegate to shared classifier. */\nfunction isBuildArgs(args: Record<string, unknown>): boolean {\n  return isBuildCommand(String(args.command ?? ''));\n}\n\nfunction isTestArgs(args: Record<string, unknown>): boolean {\n  return isTestCommand(String(args.command ?? ''));\n}\n\n/**\n * Score build/typecheck outcome (final state).\n */\nfunction scoreBuild(execResults: ExtractedToolCall[]): DimensionResult {\n  const builds = execResults.filter(r => r.toolName === 'execute_command' && isBuildArgs(r.args));\n  if (builds.length === 0) return { score: 0, detail: 'No build/typecheck ran', applicable: false };\n\n  const last = builds[builds.length - 1]!;\n  if (last.isError) return { score: 0, detail: 'Build/typecheck tool errored', applicable: true };\n\n  const exitCode = getExitCode(last.result);\n  if (exitCode === 0) return { score: 1, detail: 'Build/typecheck passed', applicable: true };\n  if (exitCode !== null && exitCode !== 0)\n    return { score: 0, detail: `Build failed (exit ${exitCode})`, applicable: true };\n\n  // Fallback: check result text for TS errors\n  const text = typeof last.result === 'string' ? last.result : JSON.stringify(last.result ?? '');\n  if (/error TS\\d+|Cannot find module|is not assignable/i.test(text)) {\n    return { score: 0, detail: 'Build failed (TypeScript errors in output)', applicable: true };\n  }\n\n  return { score: THRESHOLDS.ambiguousExitScore, detail: 'Build ran, outcome unclear', applicable: true };\n}\n\n/**\n * Score test outcome (final state).\n */\nfunction scoreTests(execResults: ExtractedToolCall[]): DimensionResult {\n  const tests = execResults.filter(r => r.toolName === 'execute_command' && isTestArgs(r.args));\n  if (tests.length === 0) return { score: 0, detail: 'No tests ran', applicable: false };\n\n  const last = tests[tests.length - 1]!;\n  if (last.isError) return { score: 0, detail: 'Test command errored', applicable: true };\n\n  const exitCode = getExitCode(last.result);\n  if (exitCode === 0) return { score: 1, detail: 'Tests passed', applicable: true };\n  if (exitCode !== null && exitCode !== 0)\n    return { score: 0, detail: `Tests failed (exit ${exitCode})`, applicable: true };\n\n  const text = typeof last.result === 'string' ? last.result : JSON.stringify(last.result ?? '');\n  if (/\\d+ (?:tests? )?passed|✓|PASS/i.test(text) && !/\\bfail(?:ed|ure)?\\b|\\berror\\b/i.test(text)) {\n    return { score: 1, detail: 'Tests passed (inferred)', applicable: true };\n  }\n  if (/\\bFAIL\\b|failed|✗|✘/i.test(text)) {\n    return { score: 0, detail: 'Tests failed (inferred)', applicable: true };\n  }\n\n  return { score: THRESHOLDS.ambiguousExitScore, detail: 'Tests ran, outcome unclear', applicable: true };\n}\n\n/**\n * Score tool error rate, excluding benign errors from expected-to-fail tools.\n */\nfunction scoreToolErrors(results: ExtractedToolCall[]): DimensionResult {\n  if (results.length === 0) return { score: 1, detail: 'No tool calls', applicable: false };\n\n  // Only count errors from non-benign tools\n  const nonBenign = results.filter(r => !THRESHOLDS.benignErrorTools.includes(r.toolName));\n  if (nonBenign.length === 0) return { score: 1, detail: 'All tool calls are benign-error tools', applicable: true };\n\n  const errors = nonBenign.filter(r => r.isError);\n  const rate = errors.length / nonBenign.length;\n\n  if (rate === 0) return { score: 1, detail: 'No tool errors', applicable: true };\n\n  const score = Math.max(0, 1 - rate * THRESHOLDS.toolErrorPenaltyMultiplier);\n  return {\n    score,\n    detail: `${errors.length}/${nonBenign.length} tools errored (${(rate * 100).toFixed(0)}%)`,\n    applicable: true,\n  };\n}\n\n/**\n * Detect stuck loops: same tool + same args appearing 3+ times consecutively,\n * or same tool + same args erroring 3+ times total.\n */\nfunction scoreStuckLoops(results: ExtractedToolCall[]): DimensionResult {\n  if (results.length < 3) return { score: 1, detail: 'Too few calls for loop detection', applicable: true };\n\n  const fingerprint = (r: ExtractedToolCall) => `${r.toolName}:${JSON.stringify(r.args)}`;\n\n  // Check consecutive identical calls\n  let maxConsecutive = 1;\n  let currentRun = 1;\n  for (let i = 1; i < results.length; i++) {\n    if (fingerprint(results[i]!) === fingerprint(results[i - 1]!)) {\n      currentRun++;\n      maxConsecutive = Math.max(maxConsecutive, currentRun);\n    } else {\n      currentRun = 1;\n    }\n  }\n\n  // Check repeated error fingerprints\n  const errorCounts = new Map<string, number>();\n  for (const r of results) {\n    if (!r.isError) continue;\n    const fp = fingerprint(r);\n    errorCounts.set(fp, (errorCounts.get(fp) ?? 0) + 1);\n  }\n  const maxErrorRepeat = Math.max(0, ...[...errorCounts.values()]);\n\n  const minReps = THRESHOLDS.loopMinRepetitions;\n  const loopSeverity = Math.max(\n    maxConsecutive >= minReps ? maxConsecutive - (minReps - 1) : 0,\n    maxErrorRepeat >= minReps ? maxErrorRepeat - (minReps - 1) : 0,\n  );\n\n  if (loopSeverity === 0) return { score: 1, detail: 'No stuck loops', applicable: true };\n\n  const score = Math.max(0, 1 - loopSeverity * THRESHOLDS.loopPenaltyPerOccurrence);\n  const details: string[] = [];\n  if (maxConsecutive >= minReps) details.push(`${maxConsecutive} consecutive identical calls`);\n  if (maxErrorRepeat >= minReps) details.push(`same error repeated ${maxErrorRepeat}x`);\n  return { score, detail: details.join('; '), applicable: true };\n}\n\n/**\n * Detect regressions: build/test passed at some point, then failed later.\n * Gives partial credit if the agent recovered (final state passes).\n */\nfunction scoreRegression(execResults: ExtractedToolCall[]): DimensionResult {\n  const regressions: { label: string; recovered: boolean }[] = [];\n\n  for (const [label, filter] of [\n    ['Build', isBuildArgs],\n    ['Tests', isTestArgs],\n  ] as const) {\n    const cmds = execResults.filter(r => r.toolName === 'execute_command' && filter(r.args));\n    if (cmds.length < 2) continue;\n\n    let sawPass = false;\n    let sawRegression = false;\n    for (const cmd of cmds) {\n      const exit = getExitCode(cmd.result);\n      if (exit === 0) {\n        if (sawRegression) {\n          // Recovered from regression\n          regressions.push({ label, recovered: true });\n          sawRegression = false;\n        }\n        sawPass = true;\n      }\n      if (sawPass && exit !== null && exit !== 0) {\n        sawRegression = true;\n      }\n    }\n\n    if (sawRegression) {\n      // Regression persisted at end of session\n      regressions.push({ label, recovered: false });\n    }\n  }\n\n  if (regressions.length === 0) return { score: 1, detail: 'No regressions', applicable: true };\n\n  // Persistent regressions are severe (0), recovered regressions get partial credit (0.5)\n  const persistentCount = regressions.filter(r => !r.recovered).length;\n  const recoveredCount = regressions.filter(r => r.recovered).length;\n  const score = Math.max(0, 1 - persistentCount * 0.5 - recoveredCount * 0.1);\n  const details = regressions.map(\n    r => `${r.label} ${r.recovered ? 'regressed then recovered' : 'regressed (persisted)'}`,\n  );\n  return { score, detail: details.join('; '), applicable: true };\n}\n\n/**\n * Score autonomy: penalize ask_user calls.\n * 0 calls = 1.0, each call reduces by 0.25 (up to a floor of 0).\n */\nfunction scoreAutonomy(results: ExtractedToolCall[]): DimensionResult {\n  const askUserCalls = results.filter(r => r.toolName === 'ask_user');\n  const count = askUserCalls.length;\n\n  if (count === 0) return { score: 1, detail: 'No ask_user calls (fully autonomous)', applicable: true };\n\n  const score = Math.max(0, 1 - count * THRESHOLDS.autonomyPenaltyPerAsk);\n  return { score, detail: `${count} ask_user call${count > 1 ? 's' : ''}`, applicable: true };\n}\n\n/**\n * Compute weighted average of only applicable dimensions.\n * Returns null if no dimensions are applicable.\n */\nfunction weightedAverage(dimensions: Record<string, DimensionResult>, weights: Record<string, number>): number | null {\n  let totalWeight = 0;\n  let weightedSum = 0;\n\n  for (const [key, result] of Object.entries(dimensions)) {\n    if (!result.applicable) continue;\n    const weight = weights[key] ?? 0;\n    totalWeight += weight;\n    weightedSum += result.score * weight;\n  }\n\n  if (totalWeight === 0) return null;\n  return weightedSum / totalWeight;\n}\n\nexport function createOutcomeScorer() {\n  return createScorer({\n    id: 'mastracode-outcome',\n    name: 'MastraCode Outcome',\n    description:\n      'Grades coding session outcomes: build/test pass, tool errors, stuck loops, regressions, autonomy. Always-on.',\n    type: 'agent',\n    prepareRun: filterRun({ partTypes: ['tool-invocation', 'text'] }),\n  })\n    .preprocess(async ({ run }) => {\n      const messages = (run.output ?? []) as MastraDBMessage[];\n      const allResults = extractToolCalls(messages);\n      const totalCalls = allResults.length;\n\n      // H1: Empty sessions or sessions with no tool calls → score 0\n      if (totalCalls < THRESHOLDS.minToolCalls) {\n        const empty: DimensionResult = { score: 0, detail: 'No tool calls', applicable: false };\n        return {\n          empty: true as const,\n          build: empty,\n          tests: empty,\n          toolErrors: empty,\n          loops: empty,\n          regression: empty,\n          autonomy: empty,\n          totalCalls,\n        };\n      }\n\n      const execResults = allResults.filter(r => r.toolName === 'execute_command');\n\n      return {\n        empty: false as const,\n        build: scoreBuild(execResults),\n        tests: scoreTests(execResults),\n        toolErrors: scoreToolErrors(allResults),\n        loops: scoreStuckLoops(allResults),\n        regression: scoreRegression(execResults),\n        autonomy: scoreAutonomy(allResults),\n        totalCalls,\n      };\n    })\n    .generateScore(({ results }) => {\n      const p = results.preprocessStepResult;\n      if (p.empty) return 0;\n\n      const score = weightedAverage(\n        {\n          build: p.build,\n          tests: p.tests,\n          toolErrors: p.toolErrors,\n          loops: p.loops,\n          regression: p.regression,\n          autonomy: p.autonomy,\n        },\n        WEIGHTS,\n      );\n\n      return score === null ? 0 : Math.round(score * 100) / 100;\n    })\n    .generateReason(({ results, score }) => {\n      const p = results.preprocessStepResult;\n      if (p.empty) {\n        return `Score: 0 (${p.totalCalls} tool calls — below minimum threshold for scoring)`;\n      }\n\n      const pct = (w: number) => `${(w * 100).toFixed(0)}%`;\n      const dimLine = (name: string, weight: number, r: DimensionResult) =>\n        r.applicable\n          ? `${name} (${pct(weight)}): ${r.detail} [${r.score}]`\n          : `${name}: ${r.detail} [N/A — excluded from average]`;\n\n      const parts: string[] = [`Score: ${score} (${p.totalCalls} tool calls total)`];\n      parts.push(dimLine('Build', WEIGHTS.build, p.build));\n      parts.push(dimLine('Tests', WEIGHTS.tests, p.tests));\n      parts.push(dimLine('Tool errors', WEIGHTS.toolErrors, p.toolErrors));\n      parts.push(dimLine('Loops', WEIGHTS.loops, p.loops));\n      parts.push(dimLine('Regression', WEIGHTS.regression, p.regression));\n      parts.push(dimLine('Autonomy', WEIGHTS.autonomy, p.autonomy));\n\n      return parts.join('\\n');\n    });\n}\n"],"mappings":"AAyBA,SAAS,cAAc,iBAAiB;AAExC,SAAS,gBAAgB,eAAe,mBAAmB;AAE3D,SAAS,wBAAwB;AAUjC,MAAM,UAAU;AAAA,EACd,OAAO;AAAA,EACP,OAAO;AAAA,EACP,YAAY;AAAA,EACZ,OAAO;AAAA,EACP,YAAY;AAAA,EACZ,UAAU;AACZ;AAKA,MAAM,aAAa;AAAA;AAAA,EAEjB,4BAA4B;AAAA;AAAA,EAG5B,oBAAoB;AAAA;AAAA,EAGpB,0BAA0B;AAAA;AAAA,EAG1B,uBAAuB;AAAA;AAAA,EAGvB,oBAAoB;AAAA;AAAA,EAGpB,cAAc;AAAA;AAAA,EAGd,kBAAkB,CAAC,kBAAkB,cAAc,aAAa;AAClE;AAWA,SAAS,YAAY,MAAwC;AAC3D,SAAO,eAAe,OAAO,KAAK,WAAW,EAAE,CAAC;AAClD;AAEA,SAAS,WAAW,MAAwC;AAC1D,SAAO,cAAc,OAAO,KAAK,WAAW,EAAE,CAAC;AACjD;AAKA,SAAS,WAAW,aAAmD;AACrE,QAAM,SAAS,YAAY,OAAO,OAAK,EAAE,aAAa,qBAAqB,YAAY,EAAE,IAAI,CAAC;AAC9F,MAAI,OAAO,WAAW,EAAG,QAAO,EAAE,OAAO,GAAG,QAAQ,0BAA0B,YAAY,MAAM;AAEhG,QAAM,OAAO,OAAO,OAAO,SAAS,CAAC;AACrC,MAAI,KAAK,QAAS,QAAO,EAAE,OAAO,GAAG,QAAQ,gCAAgC,YAAY,KAAK;AAE9F,QAAM,WAAW,YAAY,KAAK,MAAM;AACxC,MAAI,aAAa,EAAG,QAAO,EAAE,OAAO,GAAG,QAAQ,0BAA0B,YAAY,KAAK;AAC1F,MAAI,aAAa,QAAQ,aAAa;AACpC,WAAO,EAAE,OAAO,GAAG,QAAQ,sBAAsB,QAAQ,KAAK,YAAY,KAAK;AAGjF,QAAM,OAAO,OAAO,KAAK,WAAW,WAAW,KAAK,SAAS,KAAK,UAAU,KAAK,UAAU,EAAE;AAC7F,MAAI,oDAAoD,KAAK,IAAI,GAAG;AAClE,WAAO,EAAE,OAAO,GAAG,QAAQ,8CAA8C,YAAY,KAAK;AAAA,EAC5F;AAEA,SAAO,EAAE,OAAO,WAAW,oBAAoB,QAAQ,8BAA8B,YAAY,KAAK;AACxG;AAKA,SAAS,WAAW,aAAmD;AACrE,QAAM,QAAQ,YAAY,OAAO,OAAK,EAAE,aAAa,qBAAqB,WAAW,EAAE,IAAI,CAAC;AAC5F,MAAI,MAAM,WAAW,EAAG,QAAO,EAAE,OAAO,GAAG,QAAQ,gBAAgB,YAAY,MAAM;AAErF,QAAM,OAAO,MAAM,MAAM,SAAS,CAAC;AACnC,MAAI,KAAK,QAAS,QAAO,EAAE,OAAO,GAAG,QAAQ,wBAAwB,YAAY,KAAK;AAEtF,QAAM,WAAW,YAAY,KAAK,MAAM;AACxC,MAAI,aAAa,EAAG,QAAO,EAAE,OAAO,GAAG,QAAQ,gBAAgB,YAAY,KAAK;AAChF,MAAI,aAAa,QAAQ,aAAa;AACpC,WAAO,EAAE,OAAO,GAAG,QAAQ,sBAAsB,QAAQ,KAAK,YAAY,KAAK;AAEjF,QAAM,OAAO,OAAO,KAAK,WAAW,WAAW,KAAK,SAAS,KAAK,UAAU,KAAK,UAAU,EAAE;AAC7F,MAAI,iCAAiC,KAAK,IAAI,KAAK,CAAC,iCAAiC,KAAK,IAAI,GAAG;AAC/F,WAAO,EAAE,OAAO,GAAG,QAAQ,2BAA2B,YAAY,KAAK;AAAA,EACzE;AACA,MAAI,uBAAuB,KAAK,IAAI,GAAG;AACrC,WAAO,EAAE,OAAO,GAAG,QAAQ,2BAA2B,YAAY,KAAK;AAAA,EACzE;AAEA,SAAO,EAAE,OAAO,WAAW,oBAAoB,QAAQ,8BAA8B,YAAY,KAAK;AACxG;AAKA,SAAS,gBAAgB,SAA+C;AACtE,MAAI,QAAQ,WAAW,EAAG,QAAO,EAAE,OAAO,GAAG,QAAQ,iBAAiB,YAAY,MAAM;AAGxF,QAAM,YAAY,QAAQ,OAAO,OAAK,CAAC,WAAW,iBAAiB,SAAS,EAAE,QAAQ,CAAC;AACvF,MAAI,UAAU,WAAW,EAAG,QAAO,EAAE,OAAO,GAAG,QAAQ,yCAAyC,YAAY,KAAK;AAEjH,QAAM,SAAS,UAAU,OAAO,OAAK,EAAE,OAAO;AAC9C,QAAM,OAAO,OAAO,SAAS,UAAU;AAEvC,MAAI,SAAS,EAAG,QAAO,EAAE,OAAO,GAAG,QAAQ,kBAAkB,YAAY,KAAK;AAE9E,QAAM,QAAQ,KAAK,IAAI,GAAG,IAAI,OAAO,WAAW,0BAA0B;AAC1E,SAAO;AAAA,IACL;AAAA,IACA,QAAQ,GAAG,OAAO,MAAM,IAAI,UAAU,MAAM,oBAAoB,OAAO,KAAK,QAAQ,CAAC,CAAC;AAAA,IACtF,YAAY;AAAA,EACd;AACF;AAMA,SAAS,gBAAgB,SAA+C;AACtE,MAAI,QAAQ,SAAS,EAAG,QAAO,EAAE,OAAO,GAAG,QAAQ,oCAAoC,YAAY,KAAK;AAExG,QAAM,cAAc,CAAC,MAAyB,GAAG,EAAE,QAAQ,IAAI,KAAK,UAAU,EAAE,IAAI,CAAC;AAGrF,MAAI,iBAAiB;AACrB,MAAI,aAAa;AACjB,WAAS,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;AACvC,QAAI,YAAY,QAAQ,CAAC,CAAE,MAAM,YAAY,QAAQ,IAAI,CAAC,CAAE,GAAG;AAC7D;AACA,uBAAiB,KAAK,IAAI,gBAAgB,UAAU;AAAA,IACtD,OAAO;AACL,mBAAa;AAAA,IACf;AAAA,EACF;AAGA,QAAM,cAAc,oBAAI,IAAoB;AAC5C,aAAW,KAAK,SAAS;AACvB,QAAI,CAAC,EAAE,QAAS;AAChB,UAAM,KAAK,YAAY,CAAC;AACxB,gBAAY,IAAI,KAAK,YAAY,IAAI,EAAE,KAAK,KAAK,CAAC;AAAA,EACpD;AACA,QAAM,iBAAiB,KAAK,IAAI,GAAG,GAAG,CAAC,GAAG,YAAY,OAAO,CAAC,CAAC;AAE/D,QAAM,UAAU,WAAW;AAC3B,QAAM,eAAe,KAAK;AAAA,IACxB,kBAAkB,UAAU,kBAAkB,UAAU,KAAK;AAAA,IAC7D,kBAAkB,UAAU,kBAAkB,UAAU,KAAK;AAAA,EAC/D;AAEA,MAAI,iBAAiB,EAAG,QAAO,EAAE,OAAO,GAAG,QAAQ,kBAAkB,YAAY,KAAK;AAEtF,QAAM,QAAQ,KAAK,IAAI,GAAG,IAAI,eAAe,WAAW,wBAAwB;AAChF,QAAM,UAAoB,CAAC;AAC3B,MAAI,kBAAkB,QAAS,SAAQ,KAAK,GAAG,cAAc,8BAA8B;AAC3F,MAAI,kBAAkB,QAAS,SAAQ,KAAK,uBAAuB,cAAc,GAAG;AACpF,SAAO,EAAE,OAAO,QAAQ,QAAQ,KAAK,IAAI,GAAG,YAAY,KAAK;AAC/D;AAMA,SAAS,gBAAgB,aAAmD;AAC1E,QAAM,cAAuD,CAAC;AAE9D,aAAW,CAAC,OAAO,MAAM,KAAK;AAAA,IAC5B,CAAC,SAAS,WAAW;AAAA,IACrB,CAAC,SAAS,UAAU;AAAA,EACtB,GAAY;AACV,UAAM,OAAO,YAAY,OAAO,OAAK,EAAE,aAAa,qBAAqB,OAAO,EAAE,IAAI,CAAC;AACvF,QAAI,KAAK,SAAS,EAAG;AAErB,QAAI,UAAU;AACd,QAAI,gBAAgB;AACpB,eAAW,OAAO,MAAM;AACtB,YAAM,OAAO,YAAY,IAAI,MAAM;AACnC,UAAI,SAAS,GAAG;AACd,YAAI,eAAe;AAEjB,sBAAY,KAAK,EAAE,OAAO,WAAW,KAAK,CAAC;AAC3C,0BAAgB;AAAA,QAClB;AACA,kBAAU;AAAA,MACZ;AACA,UAAI,WAAW,SAAS,QAAQ,SAAS,GAAG;AAC1C,wBAAgB;AAAA,MAClB;AAAA,IACF;AAEA,QAAI,eAAe;AAEjB,kBAAY,KAAK,EAAE,OAAO,WAAW,MAAM,CAAC;AAAA,IAC9C;AAAA,EACF;AAEA,MAAI,YAAY,WAAW,EAAG,QAAO,EAAE,OAAO,GAAG,QAAQ,kBAAkB,YAAY,KAAK;AAG5F,QAAM,kBAAkB,YAAY,OAAO,OAAK,CAAC,EAAE,SAAS,EAAE;AAC9D,QAAM,iBAAiB,YAAY,OAAO,OAAK,EAAE,SAAS,EAAE;AAC5D,QAAM,QAAQ,KAAK,IAAI,GAAG,IAAI,kBAAkB,MAAM,iBAAiB,GAAG;AAC1E,QAAM,UAAU,YAAY;AAAA,IAC1B,OAAK,GAAG,EAAE,KAAK,IAAI,EAAE,YAAY,6BAA6B,uBAAuB;AAAA,EACvF;AACA,SAAO,EAAE,OAAO,QAAQ,QAAQ,KAAK,IAAI,GAAG,YAAY,KAAK;AAC/D;AAMA,SAAS,cAAc,SAA+C;AACpE,QAAM,eAAe,QAAQ,OAAO,OAAK,EAAE,aAAa,UAAU;AAClE,QAAM,QAAQ,aAAa;AAE3B,MAAI,UAAU,EAAG,QAAO,EAAE,OAAO,GAAG,QAAQ,wCAAwC,YAAY,KAAK;AAErG,QAAM,QAAQ,KAAK,IAAI,GAAG,IAAI,QAAQ,WAAW,qBAAqB;AACtE,SAAO,EAAE,OAAO,QAAQ,GAAG,KAAK,iBAAiB,QAAQ,IAAI,MAAM,EAAE,IAAI,YAAY,KAAK;AAC5F;AAMA,SAAS,gBAAgB,YAA6C,SAAgD;AACpH,MAAI,cAAc;AAClB,MAAI,cAAc;AAElB,aAAW,CAAC,KAAK,MAAM,KAAK,OAAO,QAAQ,UAAU,GAAG;AACtD,QAAI,CAAC,OAAO,WAAY;AACxB,UAAM,SAAS,QAAQ,GAAG,KAAK;AAC/B,mBAAe;AACf,mBAAe,OAAO,QAAQ;AAAA,EAChC;AAEA,MAAI,gBAAgB,EAAG,QAAO;AAC9B,SAAO,cAAc;AACvB;AAEO,SAAS,sBAAsB;AACpC,SAAO,aAAa;AAAA,IAClB,IAAI;AAAA,IACJ,MAAM;AAAA,IACN,aACE;AAAA,IACF,MAAM;AAAA,IACN,YAAY,UAAU,EAAE,WAAW,CAAC,mBAAmB,MAAM,EAAE,CAAC;AAAA,EAClE,CAAC,EACE,WAAW,OAAO,EAAE,IAAI,MAAM;AAC7B,UAAM,WAAY,IAAI,UAAU,CAAC;AACjC,UAAM,aAAa,iBAAiB,QAAQ;AAC5C,UAAM,aAAa,WAAW;AAG9B,QAAI,aAAa,WAAW,cAAc;AACxC,YAAM,QAAyB,EAAE,OAAO,GAAG,QAAQ,iBAAiB,YAAY,MAAM;AACtF,aAAO;AAAA,QACL,OAAO;AAAA,QACP,OAAO;AAAA,QACP,OAAO;AAAA,QACP,YAAY;AAAA,QACZ,OAAO;AAAA,QACP,YAAY;AAAA,QACZ,UAAU;AAAA,QACV;AAAA,MACF;AAAA,IACF;AAEA,UAAM,cAAc,WAAW,OAAO,OAAK,EAAE,aAAa,iBAAiB;AAE3E,WAAO;AAAA,MACL,OAAO;AAAA,MACP,OAAO,WAAW,WAAW;AAAA,MAC7B,OAAO,WAAW,WAAW;AAAA,MAC7B,YAAY,gBAAgB,UAAU;AAAA,MACtC,OAAO,gBAAgB,UAAU;AAAA,MACjC,YAAY,gBAAgB,WAAW;AAAA,MACvC,UAAU,cAAc,UAAU;AAAA,MAClC;AAAA,IACF;AAAA,EACF,CAAC,EACA,cAAc,CAAC,EAAE,QAAQ,MAAM;AAC9B,UAAM,IAAI,QAAQ;AAClB,QAAI,EAAE,MAAO,QAAO;AAEpB,UAAM,QAAQ;AAAA,MACZ;AAAA,QACE,OAAO,EAAE;AAAA,QACT,OAAO,EAAE;AAAA,QACT,YAAY,EAAE;AAAA,QACd,OAAO,EAAE;AAAA,QACT,YAAY,EAAE;AAAA,QACd,UAAU,EAAE;AAAA,MACd;AAAA,MACA;AAAA,IACF;AAEA,WAAO,UAAU,OAAO,IAAI,KAAK,MAAM,QAAQ,GAAG,IAAI;AAAA,EACxD,CAAC,EACA,eAAe,CAAC,EAAE,SAAS,MAAM,MAAM;AACtC,UAAM,IAAI,QAAQ;AAClB,QAAI,EAAE,OAAO;AACX,aAAO,aAAa,EAAE,UAAU;AAAA,IAClC;AAEA,UAAM,MAAM,CAAC,MAAc,IAAI,IAAI,KAAK,QAAQ,CAAC,CAAC;AAClD,UAAM,UAAU,CAAC,MAAc,QAAgB,MAC7C,EAAE,aACE,GAAG,IAAI,KAAK,IAAI,MAAM,CAAC,MAAM,EAAE,MAAM,KAAK,EAAE,KAAK,MACjD,GAAG,IAAI,KAAK,EAAE,MAAM;AAE1B,UAAM,QAAkB,CAAC,UAAU,KAAK,KAAK,EAAE,UAAU,oBAAoB;AAC7E,UAAM,KAAK,QAAQ,SAAS,QAAQ,OAAO,EAAE,KAAK,CAAC;AACnD,UAAM,KAAK,QAAQ,SAAS,QAAQ,OAAO,EAAE,KAAK,CAAC;AACnD,UAAM,KAAK,QAAQ,eAAe,QAAQ,YAAY,EAAE,UAAU,CAAC;AACnE,UAAM,KAAK,QAAQ,SAAS,QAAQ,OAAO,EAAE,KAAK,CAAC;AACnD,UAAM,KAAK,QAAQ,cAAc,QAAQ,YAAY,EAAE,UAAU,CAAC;AAClE,UAAM,KAAK,QAAQ,YAAY,QAAQ,UAAU,EAAE,QAAQ,CAAC;AAE5D,WAAO,MAAM,KAAK,IAAI;AAAA,EACxB,CAAC;AACL;","names":[]}
+{"version":3,"file":"outcome.js","names":[],"sources":["../../../src/evals/scorers/outcome.ts"],"sourcesContent":["/**\n * MastraCode Outcome Scorer (Code-based, Agent type, always-on)\n *\n * Grades the objective outcome of a coding session by checking hard facts:\n * - Did the build/typecheck pass?\n * - Did tests pass?\n * - What was the tool error rate?\n * - Did the agent get stuck in a loop?\n * - Was there a regression (build passed earlier, failed later)?\n * - Did the agent use ask_user unnecessarily? (autonomy)\n *\n * This scorer is always-on (no sampling) because it's cheap — it only\n * reads exit codes and error states from tool results already in the messages.\n *\n * Scoring philosophy (from Anthropic's eval guidance):\n * \"Grade outcomes, not paths\" — we check what happened, not how.\n * Partial credit: a session that passes build but fails tests scores\n * higher than one that fails build entirely.\n *\n * Dimensions that don't apply to a session (e.g. build for a docs-only task)\n * are excluded from the weighted average, so sessions aren't penalized for\n * irrelevant criteria.\n */\n\nimport type { MastraDBMessage } from '@mastra/core/agent';\nimport { createScorer, filterRun } from '@mastra/core/evals';\n\nimport { isBuildCommand, isTestCommand, getExitCode } from './classify-command.js';\nimport type { ExtractedToolCall } from './extract-tools.js';\nimport { extractToolCalls } from './extract-tools.js';\n\n// ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n// CONFIGURATION — Adjust weights and thresholds here\n// ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n\n/**\n * Dimension weights. Used to compute the weighted average of applicable dimensions.\n * Dimensions that don't apply (e.g. no build ran) are excluded from the average.\n */\nconst WEIGHTS = {\n  build: 0.3,\n  tests: 0.25,\n  toolErrors: 0.2,\n  loops: 0.1,\n  regression: 0.1,\n  autonomy: 0.05,\n} as const;\n\n/**\n * Scoring thresholds and multipliers.\n */\nconst THRESHOLDS = {\n  /** Multiplier for tool error rate penalty. Score = max(0, 1 - rate × this). */\n  toolErrorPenaltyMultiplier: 1.0,\n\n  /** How many consecutive identical calls (or same-error repeats) before it's a \"loop\". */\n  loopMinRepetitions: 3,\n\n  /** Score penalty per detected loop. */\n  loopPenaltyPerOccurrence: 0.3,\n\n  /** Score deducted per ask_user call. */\n  autonomyPenaltyPerAsk: 0.25,\n\n  /** Score when a command ran but exit code is ambiguous. */\n  ambiguousExitScore: 0.75,\n\n  /** Minimum tool calls required to produce a meaningful score. */\n  minToolCalls: 1,\n\n  /** Tool names whose errors are expected/benign (not penalized). */\n  benignErrorTools: ['search_content', 'find_files', 'lsp_inspect'] as string[],\n} as const;\n\n// ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n\ntype DimensionResult = { score: number; detail: string; applicable: boolean };\n\n/**\n * Check if the primary verb of an execute_command is a build/typecheck command.\n * Splits compound commands (&&, ;) and checks the first segment.\n */\n/** Extract command string from tool call args, then delegate to shared classifier. */\nfunction isBuildArgs(args: Record<string, unknown>): boolean {\n  return isBuildCommand(String(args.command ?? ''));\n}\n\nfunction isTestArgs(args: Record<string, unknown>): boolean {\n  return isTestCommand(String(args.command ?? ''));\n}\n\n/**\n * Score build/typecheck outcome (final state).\n */\nfunction scoreBuild(execResults: ExtractedToolCall[]): DimensionResult {\n  const builds = execResults.filter(r => r.toolName === 'execute_command' && isBuildArgs(r.args));\n  if (builds.length === 0) return { score: 0, detail: 'No build/typecheck ran', applicable: false };\n\n  const last = builds[builds.length - 1]!;\n  if (last.isError) return { score: 0, detail: 'Build/typecheck tool errored', applicable: true };\n\n  const exitCode = getExitCode(last.result);\n  if (exitCode === 0) return { score: 1, detail: 'Build/typecheck passed', applicable: true };\n  if (exitCode !== null && exitCode !== 0)\n    return { score: 0, detail: `Build failed (exit ${exitCode})`, applicable: true };\n\n  // Fallback: check result text for TS errors\n  const text = typeof last.result === 'string' ? last.result : JSON.stringify(last.result ?? '');\n  if (/error TS\\d+|Cannot find module|is not assignable/i.test(text)) {\n    return { score: 0, detail: 'Build failed (TypeScript errors in output)', applicable: true };\n  }\n\n  return { score: THRESHOLDS.ambiguousExitScore, detail: 'Build ran, outcome unclear', applicable: true };\n}\n\n/**\n * Score test outcome (final state).\n */\nfunction scoreTests(execResults: ExtractedToolCall[]): DimensionResult {\n  const tests = execResults.filter(r => r.toolName === 'execute_command' && isTestArgs(r.args));\n  if (tests.length === 0) return { score: 0, detail: 'No tests ran', applicable: false };\n\n  const last = tests[tests.length - 1]!;\n  if (last.isError) return { score: 0, detail: 'Test command errored', applicable: true };\n\n  const exitCode = getExitCode(last.result);\n  if (exitCode === 0) return { score: 1, detail: 'Tests passed', applicable: true };\n  if (exitCode !== null && exitCode !== 0)\n    return { score: 0, detail: `Tests failed (exit ${exitCode})`, applicable: true };\n\n  const text = typeof last.result === 'string' ? last.result : JSON.stringify(last.result ?? '');\n  if (/\\d+ (?:tests? )?passed|✓|PASS/i.test(text) && !/\\bfail(?:ed|ure)?\\b|\\berror\\b/i.test(text)) {\n    return { score: 1, detail: 'Tests passed (inferred)', applicable: true };\n  }\n  if (/\\bFAIL\\b|failed|✗|✘/i.test(text)) {\n    return { score: 0, detail: 'Tests failed (inferred)', applicable: true };\n  }\n\n  return { score: THRESHOLDS.ambiguousExitScore, detail: 'Tests ran, outcome unclear', applicable: true };\n}\n\n/**\n * Score tool error rate, excluding benign errors from expected-to-fail tools.\n */\nfunction scoreToolErrors(results: ExtractedToolCall[]): DimensionResult {\n  if (results.length === 0) return { score: 1, detail: 'No tool calls', applicable: false };\n\n  // Only count errors from non-benign tools\n  const nonBenign = results.filter(r => !THRESHOLDS.benignErrorTools.includes(r.toolName));\n  if (nonBenign.length === 0) return { score: 1, detail: 'All tool calls are benign-error tools', applicable: true };\n\n  const errors = nonBenign.filter(r => r.isError);\n  const rate = errors.length / nonBenign.length;\n\n  if (rate === 0) return { score: 1, detail: 'No tool errors', applicable: true };\n\n  const score = Math.max(0, 1 - rate * THRESHOLDS.toolErrorPenaltyMultiplier);\n  return {\n    score,\n    detail: `${errors.length}/${nonBenign.length} tools errored (${(rate * 100).toFixed(0)}%)`,\n    applicable: true,\n  };\n}\n\n/**\n * Detect stuck loops: same tool + same args appearing 3+ times consecutively,\n * or same tool + same args erroring 3+ times total.\n */\nfunction scoreStuckLoops(results: ExtractedToolCall[]): DimensionResult {\n  if (results.length < 3) return { score: 1, detail: 'Too few calls for loop detection', applicable: true };\n\n  const fingerprint = (r: ExtractedToolCall) => `${r.toolName}:${JSON.stringify(r.args)}`;\n\n  // Check consecutive identical calls\n  let maxConsecutive = 1;\n  let currentRun = 1;\n  for (let i = 1; i < results.length; i++) {\n    if (fingerprint(results[i]!) === fingerprint(results[i - 1]!)) {\n      currentRun++;\n      maxConsecutive = Math.max(maxConsecutive, currentRun);\n    } else {\n      currentRun = 1;\n    }\n  }\n\n  // Check repeated error fingerprints\n  const errorCounts = new Map<string, number>();\n  for (const r of results) {\n    if (!r.isError) continue;\n    const fp = fingerprint(r);\n    errorCounts.set(fp, (errorCounts.get(fp) ?? 0) + 1);\n  }\n  const maxErrorRepeat = Math.max(0, ...[...errorCounts.values()]);\n\n  const minReps = THRESHOLDS.loopMinRepetitions;\n  const loopSeverity = Math.max(\n    maxConsecutive >= minReps ? maxConsecutive - (minReps - 1) : 0,\n    maxErrorRepeat >= minReps ? maxErrorRepeat - (minReps - 1) : 0,\n  );\n\n  if (loopSeverity === 0) return { score: 1, detail: 'No stuck loops', applicable: true };\n\n  const score = Math.max(0, 1 - loopSeverity * THRESHOLDS.loopPenaltyPerOccurrence);\n  const details: string[] = [];\n  if (maxConsecutive >= minReps) details.push(`${maxConsecutive} consecutive identical calls`);\n  if (maxErrorRepeat >= minReps) details.push(`same error repeated ${maxErrorRepeat}x`);\n  return { score, detail: details.join('; '), applicable: true };\n}\n\n/**\n * Detect regressions: build/test passed at some point, then failed later.\n * Gives partial credit if the agent recovered (final state passes).\n */\nfunction scoreRegression(execResults: ExtractedToolCall[]): DimensionResult {\n  const regressions: { label: string; recovered: boolean }[] = [];\n\n  for (const [label, filter] of [\n    ['Build', isBuildArgs],\n    ['Tests', isTestArgs],\n  ] as const) {\n    const cmds = execResults.filter(r => r.toolName === 'execute_command' && filter(r.args));\n    if (cmds.length < 2) continue;\n\n    let sawPass = false;\n    let sawRegression = false;\n    for (const cmd of cmds) {\n      const exit = getExitCode(cmd.result);\n      if (exit === 0) {\n        if (sawRegression) {\n          // Recovered from regression\n          regressions.push({ label, recovered: true });\n          sawRegression = false;\n        }\n        sawPass = true;\n      }\n      if (sawPass && exit !== null && exit !== 0) {\n        sawRegression = true;\n      }\n    }\n\n    if (sawRegression) {\n      // Regression persisted at end of session\n      regressions.push({ label, recovered: false });\n    }\n  }\n\n  if (regressions.length === 0) return { score: 1, detail: 'No regressions', applicable: true };\n\n  // Persistent regressions are severe (0), recovered regressions get partial credit (0.5)\n  const persistentCount = regressions.filter(r => !r.recovered).length;\n  const recoveredCount = regressions.filter(r => r.recovered).length;\n  const score = Math.max(0, 1 - persistentCount * 0.5 - recoveredCount * 0.1);\n  const details = regressions.map(\n    r => `${r.label} ${r.recovered ? 'regressed then recovered' : 'regressed (persisted)'}`,\n  );\n  return { score, detail: details.join('; '), applicable: true };\n}\n\n/**\n * Score autonomy: penalize ask_user calls.\n * 0 calls = 1.0, each call reduces by 0.25 (up to a floor of 0).\n */\nfunction scoreAutonomy(results: ExtractedToolCall[]): DimensionResult {\n  const askUserCalls = results.filter(r => r.toolName === 'ask_user');\n  const count = askUserCalls.length;\n\n  if (count === 0) return { score: 1, detail: 'No ask_user calls (fully autonomous)', applicable: true };\n\n  const score = Math.max(0, 1 - count * THRESHOLDS.autonomyPenaltyPerAsk);\n  return { score, detail: `${count} ask_user call${count > 1 ? 's' : ''}`, applicable: true };\n}\n\n/**\n * Compute weighted average of only applicable dimensions.\n * Returns null if no dimensions are applicable.\n */\nfunction weightedAverage(dimensions: Record<string, DimensionResult>, weights: Record<string, number>): number | null {\n  let totalWeight = 0;\n  let weightedSum = 0;\n\n  for (const [key, result] of Object.entries(dimensions)) {\n    if (!result.applicable) continue;\n    const weight = weights[key] ?? 0;\n    totalWeight += weight;\n    weightedSum += result.score * weight;\n  }\n\n  if (totalWeight === 0) return null;\n  return weightedSum / totalWeight;\n}\n\nexport function createOutcomeScorer() {\n  return createScorer({\n    id: 'mastracode-outcome',\n    name: 'MastraCode Outcome',\n    description:\n      'Grades coding session outcomes: build/test pass, tool errors, stuck loops, regressions, autonomy. Always-on.',\n    type: 'agent',\n    prepareRun: filterRun({ partTypes: ['tool-invocation', 'text'] }),\n  })\n    .preprocess(async ({ run }) => {\n      const messages = (run.output ?? []) as MastraDBMessage[];\n      const allResults = extractToolCalls(messages);\n      const totalCalls = allResults.length;\n\n      // H1: Empty sessions or sessions with no tool calls → score 0\n      if (totalCalls < THRESHOLDS.minToolCalls) {\n        const empty: DimensionResult = { score: 0, detail: 'No tool calls', applicable: false };\n        return {\n          empty: true as const,\n          build: empty,\n          tests: empty,\n          toolErrors: empty,\n          loops: empty,\n          regression: empty,\n          autonomy: empty,\n          totalCalls,\n        };\n      }\n\n      const execResults = allResults.filter(r => r.toolName === 'execute_command');\n\n      return {\n        empty: false as const,\n        build: scoreBuild(execResults),\n        tests: scoreTests(execResults),\n        toolErrors: scoreToolErrors(allResults),\n        loops: scoreStuckLoops(allResults),\n        regression: scoreRegression(execResults),\n        autonomy: scoreAutonomy(allResults),\n        totalCalls,\n      };\n    })\n    .generateScore(({ results }) => {\n      const p = results.preprocessStepResult;\n      if (p.empty) return 0;\n\n      const score = weightedAverage(\n        {\n          build: p.build,\n          tests: p.tests,\n          toolErrors: p.toolErrors,\n          loops: p.loops,\n          regression: p.regression,\n          autonomy: p.autonomy,\n        },\n        WEIGHTS,\n      );\n\n      return score === null ? 0 : Math.round(score * 100) / 100;\n    })\n    .generateReason(({ results, score }) => {\n      const p = results.preprocessStepResult;\n      if (p.empty) {\n        return `Score: 0 (${p.totalCalls} tool calls — below minimum threshold for scoring)`;\n      }\n\n      const pct = (w: number) => `${(w * 100).toFixed(0)}%`;\n      const dimLine = (name: string, weight: number, r: DimensionResult) =>\n        r.applicable\n          ? `${name} (${pct(weight)}): ${r.detail} [${r.score}]`\n          : `${name}: ${r.detail} [N/A — excluded from average]`;\n\n      const parts: string[] = [`Score: ${score} (${p.totalCalls} tool calls total)`];\n      parts.push(dimLine('Build', WEIGHTS.build, p.build));\n      parts.push(dimLine('Tests', WEIGHTS.tests, p.tests));\n      parts.push(dimLine('Tool errors', WEIGHTS.toolErrors, p.toolErrors));\n      parts.push(dimLine('Loops', WEIGHTS.loops, p.loops));\n      parts.push(dimLine('Regression', WEIGHTS.regression, p.regression));\n      parts.push(dimLine('Autonomy', WEIGHTS.autonomy, p.autonomy));\n\n      return parts.join('\\n');\n    });\n}\n"],"mappings":";;;;;;;;AAuCA,MAAM,UAAU;CACd,OAAO;CACP,OAAO;CACP,YAAY;CACZ,OAAO;CACP,YAAY;CACZ,UAAU;AACZ;;;;AAKA,MAAM,aAAa;;CAEjB,4BAA4B;;CAG5B,oBAAoB;;CAGpB,0BAA0B;;CAG1B,uBAAuB;;CAGvB,oBAAoB;;CAGpB,cAAc;;CAGd,kBAAkB;EAAC;EAAkB;EAAc;CAAa;AAClE;;;;;;AAWA,SAAS,YAAY,MAAwC;CAC3D,OAAO,eAAe,OAAO,KAAK,WAAW,EAAE,CAAC;AAClD;AAEA,SAAS,WAAW,MAAwC;CAC1D,OAAO,cAAc,OAAO,KAAK,WAAW,EAAE,CAAC;AACjD;;;;AAKA,SAAS,WAAW,aAAmD;CACrE,MAAM,SAAS,YAAY,QAAO,MAAK,EAAE,aAAa,qBAAqB,YAAY,EAAE,IAAI,CAAC;CAC9F,IAAI,OAAO,WAAW,GAAG,OAAO;EAAE,OAAO;EAAG,QAAQ;EAA0B,YAAY;CAAM;CAEhG,MAAM,OAAO,OAAO,OAAO,SAAS;CACpC,IAAI,KAAK,SAAS,OAAO;EAAE,OAAO;EAAG,QAAQ;EAAgC,YAAY;CAAK;CAE9F,MAAM,WAAW,YAAY,KAAK,MAAM;CACxC,IAAI,aAAa,GAAG,OAAO;EAAE,OAAO;EAAG,QAAQ;EAA0B,YAAY;CAAK;CAC1F,IAAI,aAAa,QAAQ,aAAa,GACpC,OAAO;EAAE,OAAO;EAAG,QAAQ,sBAAsB,SAAS;EAAI,YAAY;CAAK;CAGjF,MAAM,OAAO,OAAO,KAAK,WAAW,WAAW,KAAK,SAAS,KAAK,UAAU,KAAK,UAAU,EAAE;CAC7F,IAAI,oDAAoD,KAAK,IAAI,GAC/D,OAAO;EAAE,OAAO;EAAG,QAAQ;EAA8C,YAAY;CAAK;CAG5F,OAAO;EAAE,OAAO,WAAW;EAAoB,QAAQ;EAA8B,YAAY;CAAK;AACxG;;;;AAKA,SAAS,WAAW,aAAmD;CACrE,MAAM,QAAQ,YAAY,QAAO,MAAK,EAAE,aAAa,qBAAqB,WAAW,EAAE,IAAI,CAAC;CAC5F,IAAI,MAAM,WAAW,GAAG,OAAO;EAAE,OAAO;EAAG,QAAQ;EAAgB,YAAY;CAAM;CAErF,MAAM,OAAO,MAAM,MAAM,SAAS;CAClC,IAAI,KAAK,SAAS,OAAO;EAAE,OAAO;EAAG,QAAQ;EAAwB,YAAY;CAAK;CAEtF,MAAM,WAAW,YAAY,KAAK,MAAM;CACxC,IAAI,aAAa,GAAG,OAAO;EAAE,OAAO;EAAG,QAAQ;EAAgB,YAAY;CAAK;CAChF,IAAI,aAAa,QAAQ,aAAa,GACpC,OAAO;EAAE,OAAO;EAAG,QAAQ,sBAAsB,SAAS;EAAI,YAAY;CAAK;CAEjF,MAAM,OAAO,OAAO,KAAK,WAAW,WAAW,KAAK,SAAS,KAAK,UAAU,KAAK,UAAU,EAAE;CAC7F,IAAI,iCAAiC,KAAK,IAAI,KAAK,CAAC,iCAAiC,KAAK,IAAI,GAC5F,OAAO;EAAE,OAAO;EAAG,QAAQ;EAA2B,YAAY;CAAK;CAEzE,IAAI,uBAAuB,KAAK,IAAI,GAClC,OAAO;EAAE,OAAO;EAAG,QAAQ;EAA2B,YAAY;CAAK;CAGzE,OAAO;EAAE,OAAO,WAAW;EAAoB,QAAQ;EAA8B,YAAY;CAAK;AACxG;;;;AAKA,SAAS,gBAAgB,SAA+C;CACtE,IAAI,QAAQ,WAAW,GAAG,OAAO;EAAE,OAAO;EAAG,QAAQ;EAAiB,YAAY;CAAM;CAGxF,MAAM,YAAY,QAAQ,QAAO,MAAK,CAAC,WAAW,iBAAiB,SAAS,EAAE,QAAQ,CAAC;CACvF,IAAI,UAAU,WAAW,GAAG,OAAO;EAAE,OAAO;EAAG,QAAQ;EAAyC,YAAY;CAAK;CAEjH,MAAM,SAAS,UAAU,QAAO,MAAK,EAAE,OAAO;CAC9C,MAAM,OAAO,OAAO,SAAS,UAAU;CAEvC,IAAI,SAAS,GAAG,OAAO;EAAE,OAAO;EAAG,QAAQ;EAAkB,YAAY;CAAK;CAG9E,OAAO;EACL,OAFY,KAAK,IAAI,GAAG,IAAI,OAAO,WAAW,0BAE1C;EACJ,QAAQ,GAAG,OAAO,OAAO,GAAG,UAAU,OAAO,mBAAmB,OAAO,IAAA,CAAK,QAAQ,CAAC,EAAE;EACvF,YAAY;CACd;AACF;;;;;AAMA,SAAS,gBAAgB,SAA+C;CACtE,IAAI,QAAQ,SAAS,GAAG,OAAO;EAAE,OAAO;EAAG,QAAQ;EAAoC,YAAY;CAAK;CAExG,MAAM,eAAe,MAAyB,GAAG,EAAE,SAAS,GAAG,KAAK,UAAU,EAAE,IAAI;CAGpF,IAAI,iBAAiB;CACrB,IAAI,aAAa;CACjB,KAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAClC,IAAI,YAAY,QAAQ,EAAG,MAAM,YAAY,QAAQ,IAAI,EAAG,GAAG;EAC7D;EACA,iBAAiB,KAAK,IAAI,gBAAgB,UAAU;CACtD,OACE,aAAa;CAKjB,MAAM,8BAAc,IAAI,IAAoB;CAC5C,KAAK,MAAM,KAAK,SAAS;EACvB,IAAI,CAAC,EAAE,SAAS;EAChB,MAAM,KAAK,YAAY,CAAC;EACxB,YAAY,IAAI,KAAK,YAAY,IAAI,EAAE,KAAK,KAAK,CAAC;CACpD;CACA,MAAM,iBAAiB,KAAK,IAAI,GAAG,GAAG,CAAC,GAAG,YAAY,OAAO,CAAC,CAAC;CAE/D,MAAM,UAAU,WAAW;CAC3B,MAAM,eAAe,KAAK,IACxB,kBAAkB,UAAU,kBAAkB,UAAU,KAAK,GAC7D,kBAAkB,UAAU,kBAAkB,UAAU,KAAK,CAC/D;CAEA,IAAI,iBAAiB,GAAG,OAAO;EAAE,OAAO;EAAG,QAAQ;EAAkB,YAAY;CAAK;CAEtF,MAAM,QAAQ,KAAK,IAAI,GAAG,IAAI,eAAe,WAAW,wBAAwB;CAChF,MAAM,UAAoB,CAAC;CAC3B,IAAI,kBAAkB,SAAS,QAAQ,KAAK,GAAG,eAAe,6BAA6B;CAC3F,IAAI,kBAAkB,SAAS,QAAQ,KAAK,uBAAuB,eAAe,EAAE;CACpF,OAAO;EAAE;EAAO,QAAQ,QAAQ,KAAK,IAAI;EAAG,YAAY;CAAK;AAC/D;;;;;AAMA,SAAS,gBAAgB,aAAmD;CAC1E,MAAM,cAAuD,CAAC;CAE9D,KAAK,MAAM,CAAC,OAAO,WAAW,CAC5B,CAAC,SAAS,WAAW,GACrB,CAAC,SAAS,UAAU,CACtB,GAAY;EACV,MAAM,OAAO,YAAY,QAAO,MAAK,EAAE,aAAa,qBAAqB,OAAO,EAAE,IAAI,CAAC;EACvF,IAAI,KAAK,SAAS,GAAG;EAErB,IAAI,UAAU;EACd,IAAI,gBAAgB;EACpB,KAAK,MAAM,OAAO,MAAM;GACtB,MAAM,OAAO,YAAY,IAAI,MAAM;GACnC,IAAI,SAAS,GAAG;IACd,IAAI,eAAe;KAEjB,YAAY,KAAK;MAAE;MAAO,WAAW;KAAK,CAAC;KAC3C,gBAAgB;IAClB;IACA,UAAU;GACZ;GACA,IAAI,WAAW,SAAS,QAAQ,SAAS,GACvC,gBAAgB;EAEpB;EAEA,IAAI,eAEF,YAAY,KAAK;GAAE;GAAO,WAAW;EAAM,CAAC;CAEhD;CAEA,IAAI,YAAY,WAAW,GAAG,OAAO;EAAE,OAAO;EAAG,QAAQ;EAAkB,YAAY;CAAK;CAG5F,MAAM,kBAAkB,YAAY,QAAO,MAAK,CAAC,EAAE,SAAS,CAAC,CAAC;CAC9D,MAAM,iBAAiB,YAAY,QAAO,MAAK,EAAE,SAAS,CAAC,CAAC;CAK5D,OAAO;EAAE,OAJK,KAAK,IAAI,GAAG,IAAI,kBAAkB,KAAM,iBAAiB,EAI1D;EAAG,QAHA,YAAY,KAC1B,MAAK,GAAG,EAAE,MAAM,GAAG,EAAE,YAAY,6BAA6B,yBAElC,CAAC,CAAC,KAAK,IAAI;EAAG,YAAY;CAAK;AAC/D;;;;;AAMA,SAAS,cAAc,SAA+C;CAEpE,MAAM,QADe,QAAQ,QAAO,MAAK,EAAE,aAAa,UAC/B,CAAC,CAAC;CAE3B,IAAI,UAAU,GAAG,OAAO;EAAE,OAAO;EAAG,QAAQ;EAAwC,YAAY;CAAK;CAGrG,OAAO;EAAE,OADK,KAAK,IAAI,GAAG,IAAI,QAAQ,WAAW,qBACpC;EAAG,QAAQ,GAAG,MAAM,gBAAgB,QAAQ,IAAI,MAAM;EAAM,YAAY;CAAK;AAC5F;;;;;AAMA,SAAS,gBAAgB,YAA6C,SAAgD;CACpH,IAAI,cAAc;CAClB,IAAI,cAAc;CAElB,KAAK,MAAM,CAAC,KAAK,WAAW,OAAO,QAAQ,UAAU,GAAG;EACtD,IAAI,CAAC,OAAO,YAAY;EACxB,MAAM,SAAS,QAAQ,QAAQ;EAC/B,eAAe;EACf,eAAe,OAAO,QAAQ;CAChC;CAEA,IAAI,gBAAgB,GAAG,OAAO;CAC9B,OAAO,cAAc;AACvB;AAEA,SAAgB,sBAAsB;CACpC,OAAO,aAAa;EAClB,IAAI;EACJ,MAAM;EACN,aACE;EACF,MAAM;EACN,YAAY,UAAU,EAAE,WAAW,CAAC,mBAAmB,MAAM,EAAE,CAAC;CAClE,CAAC,CAAC,CACC,WAAW,OAAO,EAAE,UAAU;EAE7B,MAAM,aAAa,iBADD,IAAI,UAAU,CAAC,CACW;EAC5C,MAAM,aAAa,WAAW;EAG9B,IAAI,aAAa,WAAW,cAAc;GACxC,MAAM,QAAyB;IAAE,OAAO;IAAG,QAAQ;IAAiB,YAAY;GAAM;GACtF,OAAO;IACL,OAAO;IACP,OAAO;IACP,OAAO;IACP,YAAY;IACZ,OAAO;IACP,YAAY;IACZ,UAAU;IACV;GACF;EACF;EAEA,MAAM,cAAc,WAAW,QAAO,MAAK,EAAE,aAAa,iBAAiB;EAE3E,OAAO;GACL,OAAO;GACP,OAAO,WAAW,WAAW;GAC7B,OAAO,WAAW,WAAW;GAC7B,YAAY,gBAAgB,UAAU;GACtC,OAAO,gBAAgB,UAAU;GACjC,YAAY,gBAAgB,WAAW;GACvC,UAAU,cAAc,UAAU;GAClC;EACF;CACF,CAAC,CAAC,CACD,eAAe,EAAE,cAAc;EAC9B,MAAM,IAAI,QAAQ;EAClB,IAAI,EAAE,OAAO,OAAO;EAEpB,MAAM,QAAQ,gBACZ;GACE,OAAO,EAAE;GACT,OAAO,EAAE;GACT,YAAY,EAAE;GACd,OAAO,EAAE;GACT,YAAY,EAAE;GACd,UAAU,EAAE;EACd,GACA,OACF;EAEA,OAAO,UAAU,OAAO,IAAI,KAAK,MAAM,QAAQ,GAAG,IAAI;CACxD,CAAC,CAAC,CACD,gBAAgB,EAAE,SAAS,YAAY;EACtC,MAAM,IAAI,QAAQ;EAClB,IAAI,EAAE,OACJ,OAAO,aAAa,EAAE,WAAW;EAGnC,MAAM,OAAO,MAAc,IAAI,IAAI,IAAA,CAAK,QAAQ,CAAC,EAAE;EACnD,MAAM,WAAW,MAAc,QAAgB,MAC7C,EAAE,aACE,GAAG,KAAK,IAAI,IAAI,MAAM,EAAE,KAAK,EAAE,OAAO,IAAI,EAAE,MAAM,KAClD,GAAG,KAAK,IAAI,EAAE,OAAO;EAE3B,MAAM,QAAkB,CAAC,UAAU,MAAM,IAAI,EAAE,WAAW,mBAAmB;EAC7E,MAAM,KAAK,QAAQ,SAAS,QAAQ,OAAO,EAAE,KAAK,CAAC;EACnD,MAAM,KAAK,QAAQ,SAAS,QAAQ,OAAO,EAAE,KAAK,CAAC;EACnD,MAAM,KAAK,QAAQ,eAAe,QAAQ,YAAY,EAAE,UAAU,CAAC;EACnE,MAAM,KAAK,QAAQ,SAAS,QAAQ,OAAO,EAAE,KAAK,CAAC;EACnD,MAAM,KAAK,QAAQ,cAAc,QAAQ,YAAY,EAAE,UAAU,CAAC;EAClE,MAAM,KAAK,QAAQ,YAAY,QAAQ,UAAU,EAAE,QAAQ,CAAC;EAE5D,OAAO,MAAM,KAAK,IAAI;CACxB,CAAC;AACL"}

package/dist/goal-manager.js

@@ -1,268 +1,273 @@
-import { randomUUID } from "node:crypto";
-import { getGoalActivityDurationMs } from "@mastra/core/agent";
 import { loadSettings } from "./onboarding/settings.js";
+import { randomUUID } from "crypto";
+import { getGoalActivityDurationMs } from "@mastra/core/agent";
+//#region src/goal-manager.ts
+/**
+* GoalManager — persistent cross-turn goals, backed by the Agent's native goal
+* mechanism.
+*
+* The objective lives in the durable `threadState` `'goal'` slot (via
+* `agent.setObjective`/`getObjective`/`clearObjective`/`updateObjectiveOptions`)
+* and is judged in-loop by the core goal step. This manager is a thin adapter:
+* it keeps a synchronous in-memory view of the current objective for the TUI
+* (status line, modal, keyboard shortcuts) and delegates persistence to the
+* agent. There is no standalone judge agent and no between-turn re-invocation —
+* the core goal step drives continuation and surfaces progress via `goal` stream
+* chunks.
+*/
 const DEFAULT_MAX_TURNS = 50;
 const THREAD_GOAL_KEY = "goal";
 function normalizeActiveDurationMs(value) {
-  return value !== void 0 && Number.isFinite(value) && value >= 0 ? value : 0;
-}
-class GoalManager {
-  /** Synchronous in-memory view of the active objective record (source of truth is ThreadState). */
-  record = null;
-  threadId;
-  agentId;
-  persistGoalOnNextThreadCreate = false;
-  // ---------------------------------------------------------------------------
-  // Synchronous TUI surface
-  // ---------------------------------------------------------------------------
-  getGoal() {
-    if (!this.record) return null;
-    const { judgeModelId, maxTurns } = this.effectiveSettings(this.record);
-    return {
-      id: this.record.id,
-      objective: this.record.objective,
-      status: this.record.status,
-      turnsUsed: this.record.runsUsed,
-      maxTurns,
-      judgeModelId,
-      startedAt: new Date(this.record.startedAt).toISOString(),
-      activeDurationMs: this.agentId && this.threadId ? getGoalActivityDurationMs({
-        agentId: this.agentId,
-        threadId: this.threadId,
-        objectiveId: this.record.id,
-        activeDurationMs: this.record.activeDurationMs
-      }) : normalizeActiveDurationMs(this.record.activeDurationMs)
-    };
-  }
-  isActive() {
-    return this.record?.status === "active";
-  }
-  persistOnNextThreadCreate() {
-    this.persistGoalOnNextThreadCreate = true;
-  }
-  consumePersistOnNextThreadCreate() {
-    if (!this.persistGoalOnNextThreadCreate) return false;
-    this.persistGoalOnNextThreadCreate = false;
-    return true;
-  }
-  // ---------------------------------------------------------------------------
-  // Objective lifecycle (ThreadState-backed via the agent)
-  // ---------------------------------------------------------------------------
-  /**
-   * Set a new objective. Persists to ThreadState via `agent.setObjective` and
-   * updates the in-memory view. Only the provided settings are persisted into
-   * the record; unset ones fall back to the agent's `goal` config at read time.
-   */
-  async setGoal(state, objective, judgeModelId, maxTurns = DEFAULT_MAX_TURNS) {
-    const threadId = state.session.thread.getId();
-    const agent = this.getAgent(state);
-    const now = Date.now();
-    const id = randomUUID();
-    this.threadId = threadId ?? void 0;
-    this.agentId = agent?.id;
-    if (agent && threadId) {
-      const persisted = await agent.setObjective(objective, {
-        id,
-        threadId,
-        resourceId: state.session.identity.getResourceId(),
-        ...judgeModelId ? { judgeModelId } : {},
-        maxRuns: maxTurns
-      });
-      this.record = persisted ? { ...persisted, id: persisted.id ?? id } : this.localRecord(objective, judgeModelId, maxTurns, now, id);
-    } else {
-      this.record = this.localRecord(objective, judgeModelId, maxTurns, now, id);
-    }
-    return this.getGoal();
-  }
-  /**
-   * Update the judge model / max-runs defaults. Persists into the active record
-   * (so the override is remembered in thread state) when a goal is set.
-   */
-  async updateJudgeDefaults(state, judgeModelId, maxTurns) {
-    if (!this.record) return null;
-    const threadId = state.session.thread.getId();
-    const agent = this.getAgent(state);
-    if (agent && threadId) {
-      const updated = await agent.updateObjectiveOptions({
-        threadId,
-        ...judgeModelId ? { judgeModelId } : {},
-        maxRuns: maxTurns
-      });
-      if (updated) this.record = { ...updated, id: this.record.id };
-    } else {
-      this.record = {
-        ...this.record,
-        ...judgeModelId ? { judgeModelId } : {},
-        maxRuns: maxTurns,
-        updatedAt: Date.now()
-      };
-    }
-    return this.getGoal();
-  }
-  pause(reason) {
-    if (this.record && this.record.status === "active") {
-      this.record = { ...this.record, status: "paused", pausedReason: reason, updatedAt: Date.now() };
-    }
-    return this.getGoal();
-  }
-  resume() {
-    if (this.record && this.record.status === "paused") {
-      this.record = { ...this.record, status: "active", pausedReason: void 0, updatedAt: Date.now() };
-    }
-    return this.getGoal();
-  }
-  markDone() {
-    if (this.record) {
-      this.record = { ...this.record, status: "done", updatedAt: Date.now() };
-    }
-  }
-  clear() {
-    this.record = null;
-    this.threadId = void 0;
-    this.agentId = void 0;
-    this.persistGoalOnNextThreadCreate = false;
-  }
-  /**
-   * Sync the latest objective record from ThreadState into the in-memory view.
-   * Called from the `goal` stream-chunk handler after each evaluation.
-   */
-  applyEvaluation(update) {
-    if (!this.record) return null;
-    this.record = { ...this.record, runsUsed: update.runsUsed, status: update.status, updatedAt: Date.now() };
-    return this.getGoal();
-  }
-  // ---------------------------------------------------------------------------
-  // Persistence
-  // ---------------------------------------------------------------------------
-  /**
-   * Persist the active objective to ThreadState via the agent. The objective
-   * record is the source of truth; the legacy thread-metadata key is cleared so
-   * stale state from older sessions does not resurface.
-   */
-  async saveToThread(state) {
-    const threadId = state.session.thread.getId();
-    const agent = this.getAgent(state);
-    try {
-      if (agent && threadId) {
-        if (this.record) {
-          const updated = await agent.updateObjectiveOptions({
-            threadId,
-            status: this.record.status,
-            ...this.record.pausedReason ? { pausedReason: this.record.pausedReason } : {},
-            ...this.record.judgeModelId ? { judgeModelId: this.record.judgeModelId } : {},
-            ...this.record.maxRuns !== void 0 ? { maxRuns: this.record.maxRuns } : {}
-          });
-          if (!updated) {
-            const desiredStatus = this.record.status;
-            await agent.setObjective(this.record.objective, {
-              id: this.record.id,
-              threadId,
-              resourceId: state.session.identity.getResourceId(),
-              activeDurationMs: normalizeActiveDurationMs(this.record.activeDurationMs),
-              ...this.record.judgeModelId ? { judgeModelId: this.record.judgeModelId } : {},
-              ...this.record.maxRuns !== void 0 ? { maxRuns: this.record.maxRuns } : {}
-            });
-            if (desiredStatus !== "active") {
-              await agent.updateObjectiveOptions({
-                threadId,
-                status: desiredStatus,
-                ...this.record.pausedReason ? { pausedReason: this.record.pausedReason } : {}
-              });
-            }
-          }
-        } else {
-          await agent.clearObjective({ threadId });
-        }
-      }
-      await state.session.thread.setSetting({ key: THREAD_GOAL_KEY, value: void 0 });
-    } catch {
-    }
-  }
-  /**
-   * Load the objective from ThreadState (called on thread switch). Falls back to
-   * the legacy thread-metadata goal for threads created before this migration.
-   */
-  async loadFromThread(state) {
-    this.persistGoalOnNextThreadCreate = false;
-    const threadId = state.session.thread.getId();
-    const agent = this.getAgent(state);
-    this.threadId = threadId ?? void 0;
-    this.agentId = agent?.id;
-    if (agent && threadId) {
-      try {
-        const record = await agent.getObjective({ threadId });
-        if (record) {
-          this.record = {
-            ...record,
-            id: record.id ?? randomUUID(),
-            activeDurationMs: normalizeActiveDurationMs(record.activeDurationMs)
-          };
-          return;
-        }
-      } catch {
-      }
-    }
-    this.record = null;
-  }
-  /**
-   * Legacy entry point retained for thread-switch call sites that only have the
-   * thread metadata available. Hydrates from a previously-persisted GoalState.
-   */
-  loadFromThreadMetadata(metadata) {
-    const saved = metadata?.[THREAD_GOAL_KEY];
-    this.persistGoalOnNextThreadCreate = false;
-    this.threadId = void 0;
-    this.agentId = void 0;
-    if (saved && saved.objective && saved.status) {
-      this.record = {
-        objective: saved.objective,
-        status: saved.status,
-        runsUsed: saved.turnsUsed ?? 0,
-        activeDurationMs: normalizeActiveDurationMs(saved.activeDurationMs),
-        maxRuns: saved.maxTurns ?? DEFAULT_MAX_TURNS,
-        judgeModelId: saved.judgeModelId ?? "",
-        startedAt: saved.startedAt ? Date.parse(saved.startedAt) || Date.now() : Date.now(),
-        updatedAt: Date.now(),
-        id: saved.id ?? randomUUID()
-      };
-    } else {
-      this.record = null;
-    }
-  }
-  // ---------------------------------------------------------------------------
-  // Private
-  // ---------------------------------------------------------------------------
-  getAgent(state) {
-    try {
-      return state.controller.getCurrentAgent(state.session);
-    } catch {
-      return void 0;
-    }
-  }
-  /** Resolve effective judge model + max runs (record value → settings default). */
-  effectiveSettings(record) {
-    const settings = loadSettings();
-    return {
-      judgeModelId: record.judgeModelId ?? settings.models.goalJudgeModel ?? "",
-      maxTurns: record.maxRuns ?? settings.models.goalMaxTurns ?? DEFAULT_MAX_TURNS
-    };
-  }
-  localRecord(objective, judgeModelId, maxTurns, now, id) {
-    return {
-      objective,
-      status: "active",
-      runsUsed: 0,
-      activeDurationMs: 0,
-      maxRuns: maxTurns,
-      ...judgeModelId ? { judgeModelId } : {},
-      startedAt: now,
-      updatedAt: now,
-      id
-    };
-  }
+	return value !== void 0 && Number.isFinite(value) && value >= 0 ? value : 0;
 }
-export {
-  DEFAULT_MAX_TURNS,
-  GoalManager
+var GoalManager = class {
+	/** Synchronous in-memory view of the active objective record (source of truth is ThreadState). */
+	record = null;
+	threadId;
+	agentId;
+	persistGoalOnNextThreadCreate = false;
+	getGoal() {
+		if (!this.record) return null;
+		const { judgeModelId, maxTurns } = this.effectiveSettings(this.record);
+		return {
+			id: this.record.id,
+			objective: this.record.objective,
+			status: this.record.status,
+			turnsUsed: this.record.runsUsed,
+			maxTurns,
+			judgeModelId,
+			startedAt: new Date(this.record.startedAt).toISOString(),
+			activeDurationMs: this.agentId && this.threadId ? getGoalActivityDurationMs({
+				agentId: this.agentId,
+				threadId: this.threadId,
+				objectiveId: this.record.id,
+				activeDurationMs: this.record.activeDurationMs
+			}) : normalizeActiveDurationMs(this.record.activeDurationMs)
+		};
+	}
+	isActive() {
+		return this.record?.status === "active";
+	}
+	persistOnNextThreadCreate() {
+		this.persistGoalOnNextThreadCreate = true;
+	}
+	consumePersistOnNextThreadCreate() {
+		if (!this.persistGoalOnNextThreadCreate) return false;
+		this.persistGoalOnNextThreadCreate = false;
+		return true;
+	}
+	/**
+	* Set a new objective. Persists to ThreadState via `agent.setObjective` and
+	* updates the in-memory view. Only the provided settings are persisted into
+	* the record; unset ones fall back to the agent's `goal` config at read time.
+	*/
+	async setGoal(state, objective, judgeModelId, maxTurns = 50) {
+		const threadId = state.session.thread.getId();
+		const agent = this.getAgent(state);
+		const now = Date.now();
+		const id = randomUUID();
+		this.threadId = threadId ?? void 0;
+		this.agentId = agent?.id;
+		if (agent && threadId) {
+			const persisted = await agent.setObjective(objective, {
+				id,
+				threadId,
+				resourceId: state.session.identity.getResourceId(),
+				...judgeModelId ? { judgeModelId } : {},
+				maxRuns: maxTurns
+			});
+			this.record = persisted ? {
+				...persisted,
+				id: persisted.id ?? id
+			} : this.localRecord(objective, judgeModelId, maxTurns, now, id);
+		} else this.record = this.localRecord(objective, judgeModelId, maxTurns, now, id);
+		return this.getGoal();
+	}
+	/**
+	* Update the judge model / max-runs defaults. Persists into the active record
+	* (so the override is remembered in thread state) when a goal is set.
+	*/
+	async updateJudgeDefaults(state, judgeModelId, maxTurns) {
+		if (!this.record) return null;
+		const threadId = state.session.thread.getId();
+		const agent = this.getAgent(state);
+		if (agent && threadId) {
+			const updated = await agent.updateObjectiveOptions({
+				threadId,
+				...judgeModelId ? { judgeModelId } : {},
+				maxRuns: maxTurns
+			});
+			if (updated) this.record = {
+				...updated,
+				id: this.record.id
+			};
+		} else this.record = {
+			...this.record,
+			...judgeModelId ? { judgeModelId } : {},
+			maxRuns: maxTurns,
+			updatedAt: Date.now()
+		};
+		return this.getGoal();
+	}
+	pause(reason) {
+		if (this.record && this.record.status === "active") this.record = {
+			...this.record,
+			status: "paused",
+			pausedReason: reason,
+			updatedAt: Date.now()
+		};
+		return this.getGoal();
+	}
+	resume() {
+		if (this.record && this.record.status === "paused") this.record = {
+			...this.record,
+			status: "active",
+			pausedReason: void 0,
+			updatedAt: Date.now()
+		};
+		return this.getGoal();
+	}
+	markDone() {
+		if (this.record) this.record = {
+			...this.record,
+			status: "done",
+			updatedAt: Date.now()
+		};
+	}
+	clear() {
+		this.record = null;
+		this.threadId = void 0;
+		this.agentId = void 0;
+		this.persistGoalOnNextThreadCreate = false;
+	}
+	/**
+	* Sync the latest objective record from ThreadState into the in-memory view.
+	* Called from the `goal` stream-chunk handler after each evaluation.
+	*/
+	applyEvaluation(update) {
+		if (!this.record) return null;
+		this.record = {
+			...this.record,
+			runsUsed: update.runsUsed,
+			status: update.status,
+			updatedAt: Date.now()
+		};
+		return this.getGoal();
+	}
+	/**
+	* Persist the active objective to ThreadState via the agent. The objective
+	* record is the source of truth; the legacy thread-metadata key is cleared so
+	* stale state from older sessions does not resurface.
+	*/
+	async saveToThread(state) {
+		const threadId = state.session.thread.getId();
+		const agent = this.getAgent(state);
+		try {
+			if (agent && threadId) if (this.record) {
+				if (!await agent.updateObjectiveOptions({
+					threadId,
+					status: this.record.status,
+					...this.record.pausedReason ? { pausedReason: this.record.pausedReason } : {},
+					...this.record.judgeModelId ? { judgeModelId: this.record.judgeModelId } : {},
+					...this.record.maxRuns !== void 0 ? { maxRuns: this.record.maxRuns } : {}
+				})) {
+					const desiredStatus = this.record.status;
+					await agent.setObjective(this.record.objective, {
+						id: this.record.id,
+						threadId,
+						resourceId: state.session.identity.getResourceId(),
+						activeDurationMs: normalizeActiveDurationMs(this.record.activeDurationMs),
+						...this.record.judgeModelId ? { judgeModelId: this.record.judgeModelId } : {},
+						...this.record.maxRuns !== void 0 ? { maxRuns: this.record.maxRuns } : {}
+					});
+					if (desiredStatus !== "active") await agent.updateObjectiveOptions({
+						threadId,
+						status: desiredStatus,
+						...this.record.pausedReason ? { pausedReason: this.record.pausedReason } : {}
+					});
+				}
+			} else await agent.clearObjective({ threadId });
+			await state.session.thread.setSetting({
+				key: THREAD_GOAL_KEY,
+				value: void 0
+			});
+		} catch {}
+	}
+	/**
+	* Load the objective from ThreadState (called on thread switch). Falls back to
+	* the legacy thread-metadata goal for threads created before this migration.
+	*/
+	async loadFromThread(state) {
+		this.persistGoalOnNextThreadCreate = false;
+		const threadId = state.session.thread.getId();
+		const agent = this.getAgent(state);
+		this.threadId = threadId ?? void 0;
+		this.agentId = agent?.id;
+		if (agent && threadId) try {
+			const record = await agent.getObjective({ threadId });
+			if (record) {
+				this.record = {
+					...record,
+					id: record.id ?? randomUUID(),
+					activeDurationMs: normalizeActiveDurationMs(record.activeDurationMs)
+				};
+				return;
+			}
+		} catch {}
+		this.record = null;
+	}
+	/**
+	* Legacy entry point retained for thread-switch call sites that only have the
+	* thread metadata available. Hydrates from a previously-persisted GoalState.
+	*/
+	loadFromThreadMetadata(metadata) {
+		const saved = metadata?.[THREAD_GOAL_KEY];
+		this.persistGoalOnNextThreadCreate = false;
+		this.threadId = void 0;
+		this.agentId = void 0;
+		if (saved && saved.objective && saved.status) this.record = {
+			objective: saved.objective,
+			status: saved.status,
+			runsUsed: saved.turnsUsed ?? 0,
+			activeDurationMs: normalizeActiveDurationMs(saved.activeDurationMs),
+			maxRuns: saved.maxTurns ?? 50,
+			judgeModelId: saved.judgeModelId ?? "",
+			startedAt: saved.startedAt ? Date.parse(saved.startedAt) || Date.now() : Date.now(),
+			updatedAt: Date.now(),
+			id: saved.id ?? randomUUID()
+		};
+		else this.record = null;
+	}
+	getAgent(state) {
+		try {
+			return state.controller.getCurrentAgent(state.session);
+		} catch {
+			return;
+		}
+	}
+	/** Resolve effective judge model + max runs (record value → settings default). */
+	effectiveSettings(record) {
+		const settings = loadSettings();
+		return {
+			judgeModelId: record.judgeModelId ?? settings.models.goalJudgeModel ?? "",
+			maxTurns: record.maxRuns ?? settings.models.goalMaxTurns ?? 50
+		};
+	}
+	localRecord(objective, judgeModelId, maxTurns, now, id) {
+		return {
+			objective,
+			status: "active",
+			runsUsed: 0,
+			activeDurationMs: 0,
+			maxRuns: maxTurns,
+			...judgeModelId ? { judgeModelId } : {},
+			startedAt: now,
+			updatedAt: now,
+			id
+		};
+	}
 };
+//#endregion
+export { DEFAULT_MAX_TURNS, GoalManager };
+
 //# sourceMappingURL=goal-manager.js.map