@tangle-network/agent-eval 0.182.0 → 0.183.0

package/dist/exporters-Df7TgHFv.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"exporters-Df7TgHFv.js","names":[],"sources":["../src/rollout/exporters.ts"],"sourcesContent":["/**\n * Pure exporters over `tangle.rollout.v1` lines → the training-data shapes\n * the improvement loops feed:\n *   - SFT chat JSONL           (clean trainable successes, {messages, metadata})\n *   - reward rows              (every scored line, success or failure, with steps)\n *   - Prime Intellect verifiers RolloutOutput (prompt/completion split + reward)\n *   - OpenAI RFT items         (prompt turns + verdict reference fields)\n *\n * All exporters are pure functions of the lines — filtering (never train on\n * holdout, reward thresholds, the realness gate) happens HERE, on inline\n * labels, no joins.\n *\n * Every exporter takes `MintedRolloutLine[]`, not `RolloutLine[]`: the reward\n * on a minted line has been checked against the anti-Goodhart invariant, and\n * the brand is what stops a hand-built object literal claiming a positive\n * reward on a gamed run from being handed to an exporter that copies it\n * verbatim into training data.\n */\n\nimport type {\n  ChatMessage,\n  MintedRolloutLine,\n  RolloutLine,\n  RolloutSplit,\n  RolloutStep,\n  ToolDef,\n} from './schema'\nimport { assertRewardGate } from './schema'\n\n// ---------------------------------------------------------------------------\n// The realness claims every emitted row carries.\n// ---------------------------------------------------------------------------\n\n/**\n * The gate's two claims, which travel TOGETHER on every emitted row.\n *\n * `realness_gated` alone is ambiguous, and the ambiguity is exploitable:\n * `false` reads as \"we screened it and nothing fired\", so a producer that has no\n * screen at all emitted rows indistinguishable from screened-clean ones, and\n * every consumer of the published dataset read them as clean. The second field\n * is what separates the two claims, and it only removes the ambiguity if it\n * reaches the WIRE — for a round it existed on `RolloutOutcome` and on no\n * exported row shape at all, which left the published rows exactly as ambiguous\n * as before.\n *\n * So there is one helper and every row shape spreads it. A row that states one\n * claim without the other is not constructible by copying the pattern, and\n * `exporters.test.ts` walks every emitted shape to prove none does.\n */\nexport interface RealnessLabels {\n  /** The screen's VERDICT: the run faked its success signal. */\n  realness_gated: boolean\n  /**\n   * Whether a screen RAN at all. `true` = it ran, so `realness_gated` is its\n   * verdict. `false` = the producer declares it has none. `null` = not stated\n   * (pre-unification producers), which is \"unknown\" and never \"clean\".\n   */\n  realness_screened: boolean | null\n}\n\nfunction realnessLabels(line: MintedRolloutLine): RealnessLabels {\n  return {\n    realness_gated: line.outcome.realness_gated === true,\n    // `?? null` rather than `?? false`: absent means the producer never stated\n    // it, and `false` is a load-bearing claim (\"no screen exists\") that would be\n    // an invention here.\n    realness_screened: line.outcome.realness_screened ?? null,\n  }\n}\n\n// ---------------------------------------------------------------------------\n// (a) SFT chat JSONL\n// ---------------------------------------------------------------------------\n\nexport interface TrainingExportOptions {\n  /** Include held-out evaluation data in training output. Default false. */\n  allowHeldOutTrainingData?: boolean\n  /** Require reward to be strictly greater than this value. Default 0. */\n  minimumQualityExclusive?: number\n}\n\n/**\n * What a signed-signal exporter (verifiers, RFT) does with lines that are not\n * clean trainable successes — realness-gated lines above all.\n *\n *   - 'exclude'       — the default, the same fail-closed policy as every\n *                       other training export: positive, completed,\n *                       non-gated rows on a trainable split.\n *   - 'zero-and-flag' — keep them, at their non-positive (or null) reward,\n *                       with `RealnessLabels` on the row. The dataset release\n *                       sets this per `FORMAT_GATE_DISPOSITION`: in these\n *                       formats the reward is a signed learning signal, so a\n *                       gamed trajectory at reward 0 is a correct negative,\n *                       and dropping it would bias the negative population\n *                       toward honest failures and leave a trainer no example\n *                       of gaming being penalized. The split policy is NOT\n *                       relaxed: held-out lines still need the named opt-in.\n *\n * SFT deliberately has no such option — an SFT row is an imitation target and\n * a gamed trajectory must never appear in one at any weight.\n */\nexport type GatedLineDisposition = 'exclude' | 'zero-and-flag'\n\nexport interface SignedSignalExportOptions extends TrainingExportOptions {\n  /** Disposition for non-trainable lines. Default 'exclude'. */\n  gatedLines?: GatedLineDisposition\n}\n\nexport type SftExportOptions = TrainingExportOptions\n\nexport interface SftRow {\n  messages: ChatMessage[]\n  metadata: {\n    rollout_id: string\n    run_id: string\n    candidate_id: string | null\n    instance_id: string\n    reward: number\n  } & RealnessLabels\n}\n\n/**\n * Supervised fine-tune rows: the completed conversation of each qualifying\n * line. Fail-closed filters: trainable split only (never holdout/canary),\n * reward strictly above `minimumQualityExclusive` (default 0), realness-gated\n * lines never qualify, gap lines carry no trainable content, and\n * copied-context turns are dropped from the transcript (Harbor ATIF RFC 0001\n * rule 7 — see `ChatMessage.is_copied_context`).\n *\n * `realness_gated` is therefore always `false` on an emitted row. It is carried\n * anyway: an SFT row is a pure imitation target, so the row states its realness\n * claims instead of making the reader know the format's policy, and carrying\n * both flags on all four shapes is what lets the release accounting measure\n * every config with one rule rather than skipping the one whose row shape\n * happened to omit the field.\n */\nexport function toSftRows(lines: MintedRolloutLine[], options: SftExportOptions = {}): SftRow[] {\n  for (const line of lines) assertRewardGate(line, 'SFT export')\n  return lines\n    .filter((line) => isTrainingLineEligible(line, options) && line.messages.length > 0)\n    .map((line) => ({\n      // Dropped, not kept-and-masked: a copied-context turn was authored by\n      // another agent, and an SFT trainer has no notion of \"present but not a\n      // target\" — every message it is handed is something to imitate.\n      messages: line.messages.filter((message) => message.is_copied_context !== true),\n      metadata: {\n        rollout_id: line.rollout_id,\n        run_id: line.run_id,\n        candidate_id: line.candidate_id ?? null,\n        instance_id: line.task.instance_id,\n        reward: line.outcome.reward as number,\n        ...realnessLabels(line),\n      },\n    }))\n    .filter((row) => row.messages.length > 0)\n}\n\n// ---------------------------------------------------------------------------\n// (b) Reward rows\n// ---------------------------------------------------------------------------\n\nexport interface RewardRow {\n  /** First user turn — the task prompt. */\n  prompt: string\n  steps: RolloutStep[]\n  reward: number\n  metadata: {\n    rollout_id: string\n    run_id: string\n    candidate_id: string | null\n    instance_id: string\n    split: RolloutSplit\n    /**\n     * The anti-Goodhart verdict AND whether a screen produced it, carried on the\n     * row rather than left implicit in a reward of 0. Zeroing alone is lossy: it\n     * makes a run that FAKED its success indistinguishable from one that\n     * honestly failed, so a consumer can neither drop the gamed population nor\n     * mine it (a labeled gamed trajectory is the training signal for a gaming\n     * detector). `realness_gated` alone is lossy the same way in the other\n     * direction — see `RealnessLabels`.\n     */\n  } & RealnessLabels\n}\n\n/**\n * Reward-labeled rows for completed, positive-quality training runs.\n */\nexport function toRewardRows(\n  lines: MintedRolloutLine[],\n  options: TrainingExportOptions = {},\n): RewardRow[] {\n  for (const line of lines) assertRewardGate(line, 'reward-row export')\n  return lines\n    .filter(\n      (line) =>\n        isTrainingLineEligible(line, options) &&\n        line.messages.some(\n          (message) =>\n            message.role === 'user' &&\n            typeof message.content === 'string' &&\n            message.content.length > 0,\n        ),\n    )\n    .map((line) => ({\n      prompt: line.messages.find((m) => m.role === 'user')?.content ?? '',\n      steps: line.steps ?? [],\n      reward: line.outcome.reward as number,\n      metadata: {\n        rollout_id: line.rollout_id,\n        run_id: line.run_id,\n        candidate_id: line.candidate_id ?? null,\n        instance_id: line.task.instance_id,\n        split: line.task.split,\n        ...realnessLabels(line),\n      },\n    }))\n}\n\n// ---------------------------------------------------------------------------\n// (c) Prime Intellect verifiers RolloutOutput\n// ---------------------------------------------------------------------------\n\nexport interface VerifiersTokenUsage {\n  input_tokens: number | null\n  output_tokens: number | null\n  reasoning_tokens: number | null\n  cache_read_tokens: number | null\n  cache_write_tokens: number | null\n}\n\nexport interface VerifiersRolloutOutput {\n  /** Messages through the last turn BEFORE the first assistant turn. */\n  prompt: ChatMessage[]\n  /** The first assistant turn onward — what the policy produced. */\n  completion: ChatMessage[]\n  reward: number | null\n  metrics: Record<string, unknown>\n  tool_defs: ToolDef[]\n  token_usage: VerifiersTokenUsage\n  info: {\n    task: RolloutLine['task']\n    policy: RolloutLine['policy']\n    rollout_id: string\n    run_id: string\n    experiment_id: string | null\n    candidate_id: string | null\n    generation: number | null\n    candidate_index: number | null\n    role: RolloutLine['role']\n  } & RealnessLabels\n}\n\n/** Index of the first assistant turn; messages.length when none exists. */\nfunction firstAssistantIndex(messages: ChatMessage[]): number {\n  const index = messages.findIndex((m) => m.role === 'assistant')\n  return index === -1 ? messages.length : index\n}\n\nexport function toVerifiersRolloutOutput(line: MintedRolloutLine): VerifiersRolloutOutput {\n  assertRewardGate(line, 'verifiers export')\n  const split = firstAssistantIndex(line.messages)\n  return {\n    prompt: line.messages.slice(0, split),\n    completion: line.messages.slice(split),\n    reward: line.outcome.reward,\n    metrics: line.outcome.metrics,\n    tool_defs: line.tool_defs,\n    token_usage: {\n      input_tokens: line.cost.tokens_in,\n      output_tokens: line.cost.tokens_out,\n      reasoning_tokens: line.cost.tokens_reasoning,\n      cache_read_tokens: line.cost.cache_read,\n      cache_write_tokens: line.cost.cache_write,\n    },\n    info: {\n      task: line.task,\n      policy: line.policy,\n      rollout_id: line.rollout_id,\n      run_id: line.run_id,\n      experiment_id: line.experiment_id ?? null,\n      candidate_id: line.candidate_id ?? null,\n      generation: line.generation,\n      candidate_index: line.candidate_index,\n      role: line.role,\n      ...realnessLabels(line),\n    },\n  }\n}\n\nexport function toVerifiersRolloutOutputs(\n  lines: MintedRolloutLine[],\n  options: SignedSignalExportOptions = {},\n): VerifiersRolloutOutput[] {\n  if (options.gatedLines === 'zero-and-flag') {\n    return lines\n      .filter((line) => isSplitEligible(line, options) && line.messages.length > 0)\n      .map(toVerifiersRolloutOutput)\n  }\n  return lines\n    .filter(\n      (line) =>\n        isTrainingLineEligible(line, options) &&\n        firstAssistantIndex(line.messages) > 0 &&\n        firstAssistantIndex(line.messages) < line.messages.length,\n    )\n    .map(toVerifiersRolloutOutput)\n}\n\n// ---------------------------------------------------------------------------\n// (d) OpenAI RFT items\n// ---------------------------------------------------------------------------\n\nexport interface RftItem {\n  /** Prompt turns only — the graded completion is re-sampled during RFT. */\n  messages: ChatMessage[]\n  /** Verdict/label fields the grader references as item.reference.* */\n  reference: {\n    reward: number | null\n    reward_source: string | null\n    verdict: unknown\n    instance_id: string\n    suite: string\n    split: RolloutSplit\n    rollout_id: string\n  } & RealnessLabels\n}\n\nexport function toRftItem(line: MintedRolloutLine): RftItem {\n  assertRewardGate(line, 'RFT export')\n  const split = firstAssistantIndex(line.messages)\n  return {\n    messages: line.messages.slice(0, split),\n    reference: {\n      reward: line.outcome.reward,\n      reward_source: line.outcome.reward_source,\n      verdict: line.outcome.verdict,\n      instance_id: line.task.instance_id,\n      suite: line.task.suite,\n      split: line.task.split,\n      rollout_id: line.rollout_id,\n      ...realnessLabels(line),\n    },\n  }\n}\n\n/** RFT needs a real prompt: lines whose transcript starts with prompt turns. */\nexport function toRftItems(\n  lines: MintedRolloutLine[],\n  options: SignedSignalExportOptions = {},\n): RftItem[] {\n  return lines\n    .filter(\n      (line) =>\n        (options.gatedLines === 'zero-and-flag'\n          ? isSplitEligible(line, options)\n          : isTrainingLineEligible(line, options)) &&\n        line.messages.length > 0 &&\n        firstAssistantIndex(line.messages) > 0,\n    )\n    .map(toRftItem)\n}\n\n// ---------------------------------------------------------------------------\n// Serialization — one JSON object per line, the interchange format for\n// every export. `tangle.rollout.v1` lines and export rows alike.\n// ---------------------------------------------------------------------------\n\nexport function toJsonl(rows: ReadonlyArray<unknown>): string {\n  return rows.map((r) => JSON.stringify(r)).join('\\n') + (rows.length ? '\\n' : '')\n}\n\n/**\n * The split half of the training policy alone: `search` is trainable, held-out\n * needs the named opt-in, `dev` and `canary` never ship. This is the ONE check\n * `'zero-and-flag'` does not relax — a gated line is shipped as a labeled\n * negative, not as a licence to train on evaluation data.\n *\n * Exported as the single implementation of that rule: `rl/exporters` applies\n * it on its line paths too, so the two waists cannot drift on which splits are\n * trainable.\n */\nexport function isSplitEligible(\n  line: RolloutLine,\n  options: Pick<TrainingExportOptions, 'allowHeldOutTrainingData'>,\n): boolean {\n  if (line.task.split === 'search') return true\n  return line.task.split === 'holdout' && options.allowHeldOutTrainingData === true\n}\n\nfunction isTrainingLineEligible(\n  line: RolloutLine,\n  options: TrainingExportOptions,\n): line is RolloutLine & { outcome: RolloutLine['outcome'] & { reward: number } } {\n  const minimumQualityExclusive = options.minimumQualityExclusive ?? 0\n  if (!Number.isFinite(minimumQualityExclusive)) {\n    throw new Error('minimumQualityExclusive must be finite')\n  }\n\n  const reward = line.outcome.reward\n  if (reward === null) return false\n  if (!Number.isFinite(reward)) {\n    throw new Error(`training reward for rollout \"${line.rollout_id}\" must be finite`)\n  }\n  if (reward <= minimumQualityExclusive) return false\n  if (!line.outcome.is_completed || line.outcome.is_truncated || line.outcome.error !== null) {\n    return false\n  }\n  if (line.outcome.realness_gated === true) return false\n  return isSplitEligible(line, options)\n}\n"],"mappings":";;AA4DA,SAAS,eAAe,MAAyC;CAC/D,OAAO;EACL,gBAAgB,KAAK,QAAQ,mBAAmB;EAIhD,mBAAmB,KAAK,QAAQ,qBAAqB;CACvD;AACF;;;;;;;;;;;;;;;;AAoEA,SAAgB,UAAU,OAA4B,UAA4B,CAAC,GAAa;CAC9F,KAAK,MAAM,QAAQ,OAAO,iBAAiB,MAAM,YAAY;CAC7D,OAAO,MACJ,QAAQ,SAAS,uBAAuB,MAAM,OAAO,KAAK,KAAK,SAAS,SAAS,CAAC,CAAC,CACnF,KAAK,UAAU;EAId,UAAU,KAAK,SAAS,QAAQ,YAAY,QAAQ,sBAAsB,IAAI;EAC9E,UAAU;GACR,YAAY,KAAK;GACjB,QAAQ,KAAK;GACb,cAAc,KAAK,gBAAgB;GACnC,aAAa,KAAK,KAAK;GACvB,QAAQ,KAAK,QAAQ;GACrB,GAAG,eAAe,IAAI;EACxB;CACF,EAAE,CAAC,CACF,QAAQ,QAAQ,IAAI,SAAS,SAAS,CAAC;AAC5C;;;;AAgCA,SAAgB,aACd,OACA,UAAiC,CAAC,GACrB;CACb,KAAK,MAAM,QAAQ,OAAO,iBAAiB,MAAM,mBAAmB;CACpE,OAAO,MACJ,QACE,SACC,uBAAuB,MAAM,OAAO,KACpC,KAAK,SAAS,MACX,YACC,QAAQ,SAAS,UACjB,OAAO,QAAQ,YAAY,YAC3B,QAAQ,QAAQ,SAAS,CAC7B,CACJ,CAAC,CACA,KAAK,UAAU;EACd,QAAQ,KAAK,SAAS,MAAM,MAAM,EAAE,SAAS,MAAM,CAAC,EAAE,WAAW;EACjE,OAAO,KAAK,SAAS,CAAC;EACtB,QAAQ,KAAK,QAAQ;EACrB,UAAU;GACR,YAAY,KAAK;GACjB,QAAQ,KAAK;GACb,cAAc,KAAK,gBAAgB;GACnC,aAAa,KAAK,KAAK;GACvB,OAAO,KAAK,KAAK;GACjB,GAAG,eAAe,IAAI;EACxB;CACF,EAAE;AACN;;AAqCA,SAAS,oBAAoB,UAAiC;CAC5D,MAAM,QAAQ,SAAS,WAAW,MAAM,EAAE,SAAS,WAAW;CAC9D,OAAO,UAAU,KAAK,SAAS,SAAS;AAC1C;AAEA,SAAgB,yBAAyB,MAAiD;CACxF,iBAAiB,MAAM,kBAAkB;CACzC,MAAM,QAAQ,oBAAoB,KAAK,QAAQ;CAC/C,OAAO;EACL,QAAQ,KAAK,SAAS,MAAM,GAAG,KAAK;EACpC,YAAY,KAAK,SAAS,MAAM,KAAK;EACrC,QAAQ,KAAK,QAAQ;EACrB,SAAS,KAAK,QAAQ;EACtB,WAAW,KAAK;EAChB,aAAa;GACX,cAAc,KAAK,KAAK;GACxB,eAAe,KAAK,KAAK;GACzB,kBAAkB,KAAK,KAAK;GAC5B,mBAAmB,KAAK,KAAK;GAC7B,oBAAoB,KAAK,KAAK;EAChC;EACA,MAAM;GACJ,MAAM,KAAK;GACX,QAAQ,KAAK;GACb,YAAY,KAAK;GACjB,QAAQ,KAAK;GACb,eAAe,KAAK,iBAAiB;GACrC,cAAc,KAAK,gBAAgB;GACnC,YAAY,KAAK;GACjB,iBAAiB,KAAK;GACtB,MAAM,KAAK;GACX,GAAG,eAAe,IAAI;EACxB;CACF;AACF;AAEA,SAAgB,0BACd,OACA,UAAqC,CAAC,GACZ;CAC1B,IAAI,QAAQ,eAAe,iBACzB,OAAO,MACJ,QAAQ,SAAS,gBAAgB,MAAM,OAAO,KAAK,KAAK,SAAS,SAAS,CAAC,CAAC,CAC5E,IAAI,wBAAwB;CAEjC,OAAO,MACJ,QACE,SACC,uBAAuB,MAAM,OAAO,KACpC,oBAAoB,KAAK,QAAQ,IAAI,KACrC,oBAAoB,KAAK,QAAQ,IAAI,KAAK,SAAS,MACvD,CAAC,CACA,IAAI,wBAAwB;AACjC;AAqBA,SAAgB,UAAU,MAAkC;CAC1D,iBAAiB,MAAM,YAAY;CACnC,MAAM,QAAQ,oBAAoB,KAAK,QAAQ;CAC/C,OAAO;EACL,UAAU,KAAK,SAAS,MAAM,GAAG,KAAK;EACtC,WAAW;GACT,QAAQ,KAAK,QAAQ;GACrB,eAAe,KAAK,QAAQ;GAC5B,SAAS,KAAK,QAAQ;GACtB,aAAa,KAAK,KAAK;GACvB,OAAO,KAAK,KAAK;GACjB,OAAO,KAAK,KAAK;GACjB,YAAY,KAAK;GACjB,GAAG,eAAe,IAAI;EACxB;CACF;AACF;;AAGA,SAAgB,WACd,OACA,UAAqC,CAAC,GAC3B;CACX,OAAO,MACJ,QACE,UACE,QAAQ,eAAe,kBACpB,gBAAgB,MAAM,OAAO,IAC7B,uBAAuB,MAAM,OAAO,MACxC,KAAK,SAAS,SAAS,KACvB,oBAAoB,KAAK,QAAQ,IAAI,CACzC,CAAC,CACA,IAAI,SAAS;AAClB;AAOA,SAAgB,QAAQ,MAAsC;CAC5D,OAAO,KAAK,KAAK,MAAM,KAAK,UAAU,CAAC,CAAC,CAAC,CAAC,KAAK,IAAI,KAAK,KAAK,SAAS,OAAO;AAC/E;;;;;;;;;;;AAYA,SAAgB,gBACd,MACA,SACS;CACT,IAAI,KAAK,KAAK,UAAU,UAAU,OAAO;CACzC,OAAO,KAAK,KAAK,UAAU,aAAa,QAAQ,6BAA6B;AAC/E;AAEA,SAAS,uBACP,MACA,SACgF;CAChF,MAAM,0BAA0B,QAAQ,2BAA2B;CACnE,IAAI,CAAC,OAAO,SAAS,uBAAuB,GAC1C,MAAM,IAAI,MAAM,wCAAwC;CAG1D,MAAM,SAAS,KAAK,QAAQ;CAC5B,IAAI,WAAW,MAAM,OAAO;CAC5B,IAAI,CAAC,OAAO,SAAS,MAAM,GACzB,MAAM,IAAI,MAAM,gCAAgC,KAAK,WAAW,iBAAiB;CAEnF,IAAI,UAAU,yBAAyB,OAAO;CAC9C,IAAI,CAAC,KAAK,QAAQ,gBAAgB,KAAK,QAAQ,gBAAgB,KAAK,QAAQ,UAAU,MACpF,OAAO;CAET,IAAI,KAAK,QAAQ,mBAAmB,MAAM,OAAO;CACjD,OAAO,gBAAgB,MAAM,OAAO;AACtC"}

package/dist/index-CJ3LhKIX.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"index-CJ3LhKIX.d.ts","names":[],"sources":["../src/rollout/exporters.ts","../src/rollout/gate-checks.ts","../src/rollout/interchange/harbor.ts","../src/rollout/ledger.ts","../src/rollout/reward.ts","../src/rollout/mint.ts","../src/rollout/readers/claude-jsonl.ts","../src/rollout/readers/opencode-sqlite.ts","../src/rollout/release/gate-report.ts","../src/rollout/release/scrub.ts","../src/rollout/release/card.ts","../src/rollout/release/hf-dataset.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;UAiDiB;;EAEf;;;;;;EAMA;;UAiBe;;EAEf;;EAEA;;;;;;;;;;;;;;;;;;;;;;KAuBU;UAEK,kCAAkC;;EAEjD,aAAa;;KAGH,mBAAmB;UAEd;EACf,UAAU;EACV;IACE;IACA;IACA;IACA;IACA;MACE;;;;;;;;;;;;;;;;;iBAkBU,UAAU,OAAO,qBAAqB,UAAS,mBAAwB;UAyBtE;;EAEf;EACA,OAAO;EACP;EACA;IACE;IACA;IACA;IACA;IACA,OAAO;MAUL;;;;;iBAMU,aACd,OAAO,qBACP,UAAS,wBACR;UAgCc;EACf;EACA;EACA;EACA;EACA;;UAGe;;EAEf,QAAQ;;EAER,YAAY;EACZ;EACA,SAAS;EACT,WAAW;EACX,aAAa;EACb;IACE,MAAM;IACN,QAAQ;IACR;IACA;IACA;IACA;IACA;IACA;IACA,MAAM;MACJ;;iBASU,yBAAyB,MAAM,oBAAoB;iBA+BnD,0BACd,OAAO,qBACP,UAAS,4BACR;UAoBc;;EAEf,UAAU;;EAEV;IACE;IACA;IACA;IACA;IACA;IACA,OAAO;IACP;MACE;;iBAGU,UAAU,MAAM,oBAAoB;;iBAmBpC,WACd,OAAO,qBACP,UAAS,4BACR;iBAkBa,QAAQ,MAAM;;;;;;;;;;cCpUjB;KAOD,sBAAsB;;;;;;;;;KAUtB,qBAAqB,QAAQ,kBAAkB,SAAS;;;;;;;;;;;;;;;;;UAkBnD;EACf,SAAS;;EAET;;UAOe;EACf,IAAI;;EAEJ;;EAEA,SAAS,SAAS;;;;;;;;;;;;EAYlB,WAAW;;;;;;;;;;;;;iBA6LG,gBAAgB,SAAS,cAAc;;;;;;cA6I1C,yBAAyB,KAAK,cAAc;;;;;;;;;KAe7C;WACG;;;;WAEA;WAAyB;;;;WAEzB;WAAuB;;;KAU1B,yBAAyB,KAAK,cAAc;;;;;;;;;cAU3C;;;;;;;aAWT;eAnCiB;;aAoCjB,kBAAgB;aAMhB,2BAAyB;aAMzB,qBAAmB;;;;;;;;aAenB;;;aACA,kBAAgB;aAChB,2BAAyB;aACzB;;;;;;;;;;;aASA;;;aACA;;;aACA;;;aACA;;;;;;;;;;;aASA;;;aACA;;;aACA;;;aACA;;;;;;KAKQ,8BAA8B;;;;;;iBAO1B,WAAW,SAAS,aAAa,QAAQ;;;cCxd5C;;cAGA;KAwBD;UAEK;EACf;EACA;;UAGe;EACf;EACA;EACA,SAAS;;UAGM;EACf;EACA;;EAEA,WAAW;EACX,QAAQ;;UAGO;EACf;EACA;;EAEA;EACA,QAAQ;;UAGO;EACf;EACA,mBAAmB;EACnB,0BAA0B;EAC1B,QAAQ;;UAGO;EACf,SAAS;;UAGM;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA,QAAQ;;UAGO;;EAEf;EACA;EACA,QAAQ;EACR;EACA;EACA,kBAAkB;EAClB;EACA,aAAa;EACb,cAAc;EACd,UAAU;EACV;EACA;EACA,QAAQ;;UAGO;EACf;EACA;EACA;;EAEA,mBAAmB;EACnB,QAAQ;;UAGO;EACf;EACA;EACA;EACA;EACA;EACA,QAAQ;;UAGO;EACf;EACA;;EAEA;EACA,OAAO;EACP,OAAO;EACP;EACA,gBAAgB;EAChB;EACA,wBAAwB;EACxB,QAAQ;;;;;;;;;;;;;;;;;;iBAmXM,mBAAmB,OAAO,gBAAgB;;iBAa1C,qBAAqB,OAAO,gBAAgB;UA2O3C;;EAEf,YAAY;;;;;;;;;;;iBAwGE,qBACd,YAAY,kBACZ,UAAS,oBACR;;;;;;;;;;;;;;;iBAiCa,qBACd,gBAAgB,eAChB,OAAO,eACN;;;;iBC/6BmB,mBAAmB,cAAc,OAAO,gBAAgB;;iBAOxD,mBAAmB,cAAc,OAAO,gBAAgB;;;;;;;;;;;iBAiBxD,kBAAkB,eAAe,QAAQ;;;;;;;;;;;;;;;iBAkBzC,mBAAmB,eAAe,QAAQ;;;;;;;;KCzCpD;;KAGP,SAAS,KAAK;;iBAGH,gBAAgB,QAAQ;;;;;;;;;;;;;;iBAiBxB,mBAAmB,QAAQ,QAAQ,OAAO;;;;;;;;;;;;;;;;iBAmB1C,cACd,QAAQ,QACR,SAAQ;;KAOE;;;;;;;iBAQI,YAAY,QAAQ,QAAQ,SAAQ,kBAA8B;;;;;;;;;iBAelE,cACd,QAAQ,QACR,SAAQ;;;;;;;;;;;;iBAiBM,eAAe,QAAQ;EAAW;EAAuB;;;;;KCvE7D,mBAAmB;UAEd;EACf,QAAQ;;EAER;;EAEA,OAAO;;EAEP;;EAEA,YAAY;;UAGG;EACf,MAAM;;EAEN;;;;;;;;;;;;;;iBAkLc,kBAAkB,QAAQ;;;;;;;iBA+JpB,gBACpB,SAAS,aACT,OAAO,YACP,UAAS,qBACR,QAAQ;;;cCrYE;;iBAGG,kBAAkB;UAIjB;EACf;EACA;;;iBAIoB,sBACpB,aACA,uBACC,QAAQ;UASM;EACf;EACA;EACA;EACA;;UAGe;EACf,UAAU;EACV,OAAO;;EAEP;EACA;EACA;;UAyDe;;;;;;WAMN;;;iBAgBW,qBACpB,cACA,UAAS,8BACR,QAAQ;;;cCnHE;UAEI;EACf;EACA;EACA;EACA;;EAEA;IAAS;IAAa;;EACtB;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;;iBAYoB,eACpB,gBACC,QAAQ;;iBAiDK,gCACd,IAAI,cACJ,oBACC;;;;;;;iBA2Ba,4BAA4B,IAAI,cAAc,oBAAoB;;;;KC5EtE;cAEC,yBAAyB,OAAO,eAAe;;UAQ3C;EACf;EACA;;;;;;;;;;;;EAYA;;;;;;;;EAQA;;;;;;;;;;;;EAYA;;;UAIe;;EAEf;EACA;;UAGe;;EAEf;;EAEA;;;;;;EAMA;;EAEA;;;;;;;;;;;;;;EAcA,oBAAoB;;;;;;;;;;;;EAYpB;;EAEA;;;;;;;;;;EAUA,wBAAwB;;UAGT;;EAEf;EACA,UAAU,QAAQ,OAAO,eAAe;;;iBAI1B,gBAAgB,gBAAgB,sBAAsB;;;;;;;cAUzD;EAIX,MAAG,eAAkB,aAAW;EAShC,YAAS,eAAkB,6BAA2B;EAStD,MAAG,eAAkB,cAAY;EAOjC,MAAG,gBAAmB,wBAAsB;;;iBAgC9B,kBACd,OAAO,qBACP,eAAe,kBACd;;;;;;;;;;;;;;;;iBA+Fa,iBAAiB,QAAQ;;;UCrUxB;EACf;EACA,SAAS;;EAET,UAAU,eAAe;;;KAqEf,cAAc;iBAEV,oBAAoB;iBAIpB,eAAe,MAAM,aAAa,MAAM,cAAc;iBAKtD,UAAU,cAAc,QAAQ;;;;;;;;;iBA8BhC,iBAAiB,MAAM,mBAAmB,QAAQ,cAAc;iBAIhE,WAAW,OAAO;EAChC,OAAO;EACP,QAAQ;;;;;;;;;iBAaM,uBAAuB;;;cCrI1B;KACD,wBAAwB;;cAGvB,cAAc,OAAO;UAejB;;EAEf,OAAO;EACP,SAAS;EACT;;EAEA;EACA,aAAa;EACb;IAAY;IAAmB;;EAC/B,cAAc,QAAQ,OAAO;;;;;;;EAO7B,MAAM;;iBA4GQ,iBAAiB,QAAQ;;;UC1HxB;EACf;EACA,SAAS;EACT;;UAGe;;EAEf,OAAO,eAAe;EACtB,QAAQ;EACR;IAAY;IAAmB;;;UAGhB;EACf;EACA;EACA;EACA,OAAO;EACP,cAAc,QAAQ,OAAO;;EAE7B,MAAM;EACN;;iBAGoB,eACpB,kBACA,SAAS,eACR,QAAQ;iBAuGK,gBAAgB,cAAc;UAoB7B,8BAA8B;EAC7C;EACA;;iBAMc,wBAAwB,iBAAiB;;iBA0CnC,qBAAqB,iBAAiB"}

package/dist/metrics-Qv-cpptD.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"metrics-Qv-cpptD.js","names":[],"sources":["../src/metrics.ts"],"sourcesContent":["interface TokenPrice {\n  input: number\n  output: number\n}\n\n/** Per-1K token pricing for exact model ids. */\nexport const MODEL_PRICING: Record<string, TokenPrice> = {\n  'gpt-4o': { input: 0.0025, output: 0.01 },\n  'gpt-4o-mini': { input: 0.00015, output: 0.0006 },\n  'gpt-4-turbo': { input: 0.01, output: 0.03 },\n  'claude-sonnet-4-20250514': { input: 0.003, output: 0.015 },\n  'claude-opus-4-20250514': { input: 0.015, output: 0.075 },\n  'claude-3-haiku-20240307': { input: 0.00025, output: 0.00125 },\n  // Router response-header rates observed for the exact served model on\n  // 2026-08-15. Keep this exact row ahead of the deliberately coarse family\n  // fallback so glm-5.3 campaign costs are not understated by ~2.5x.\n  'glm-5.3': { input: 0.00168, output: 0.00528 },\n}\n\n/** Family-level pricing fallbacks (per-1K), matched against a normalized id\n *  after exact lookup misses. Ordered — first match wins. Covers the model\n *  ids actually used through the Tangle router + cli-bridge harnesses\n *  (`claude-code/sonnet`, `opencode/zai-coding-plan/glm-5.1`,\n *  `kimi-code/kimi-k2.6`, `deepseek-v4-pro`, `anthropic/claude-sonnet-4-6`, …),\n *  none of which appear in the exact table above — without this they priced\n *  to a silent $0, blanking every cost/Pareto axis downstream. */\nconst FAMILY_PRICING: Array<[RegExp, TokenPrice]> = [\n  [/claude.*opus/, { input: 0.015, output: 0.075 }],\n  [/claude.*haiku/, { input: 0.0008, output: 0.004 }],\n  [/claude.*sonnet|claude-code|claude-sonnet/, { input: 0.003, output: 0.015 }],\n  [/gpt-4o-mini/, { input: 0.00015, output: 0.0006 }],\n  [/gpt-5|gpt-4\\.1|o[134]\\b/, { input: 0.00125, output: 0.01 }],\n  [/gpt-4o|gpt-4/, { input: 0.0025, output: 0.01 }],\n  [/deepseek/, { input: 0.0003, output: 0.0011 }],\n  [/glm|zhipu|zai/, { input: 0.0006, output: 0.0022 }],\n  [/kimi|moonshot/, { input: 0.0006, output: 0.0025 }],\n  [/qwen/, { input: 0.0004, output: 0.0012 }],\n  [/gemini.*flash/, { input: 0.0001, output: 0.0004 }],\n  [/gemini/, { input: 0.00125, output: 0.005 }],\n  [/llama/, { input: 0.0002, output: 0.0006 }],\n]\n\n/** Normalize a model id for pricing: drop a `@snapshot` suffix, lowercase,\n *  and return both the complete harness/provider-prefixed id and its final\n *  served-model segment (`opencode/zai-coding-plan/glm-5.3` → full id plus\n *  `glm-5.3`). Exact rows are checked against both before family fallback. */\nfunction normalizedModelIds(model: string): string[] {\n  const id = (model.split('@')[0] ?? model).trim().toLowerCase()\n  const finalSegment = id.slice(id.lastIndexOf('/') + 1)\n  return finalSegment === id ? [id] : [id, finalSegment]\n}\n\n/** Resolve pricing for a model id: exact table, then family fallback.\n *  Returns null when the id matches nothing (caller decides — never a\n *  silent-zero masquerading as a real $0 cost). */\nexport function resolveModelPricing(model: string): TokenPrice | null {\n  if (MODEL_PRICING[model]) return MODEL_PRICING[model]\n  const ids = normalizedModelIds(model)\n  for (const id of ids) {\n    if (MODEL_PRICING[id]) return MODEL_PRICING[id]\n  }\n  for (const [pattern, price] of FAMILY_PRICING) {\n    if (ids.some((id) => pattern.test(id))) return price\n  }\n  return null\n}\n\n/** True when `model` has known pricing (exact or family). Lets cost-aware\n *  callers distinguish a real $0 from an unpriced model. */\nexport function isModelPriced(model: string): boolean {\n  return resolveModelPricing(model) !== null\n}\n\nconst warnedUnpricedModels = new Set<string>()\n\n/** Estimate token count from string length (chars / 4 approximation) */\nexport function estimateTokens(text: string): number {\n  return Math.ceil(text.length / 4)\n}\n\n/** Calculate cost in USD from token counts and model. Unknown models warn\n *  once (not a silent zero) and return 0 so callers that ignore pricing keep\n *  working; cost-sensitive callers should gate on {@link isModelPriced}. */\nexport function estimateCost(inputTokens: number, outputTokens: number, model: string): number {\n  const pricing = resolveModelPricing(model)\n  if (!pricing) {\n    if (!warnedUnpricedModels.has(model)) {\n      warnedUnpricedModels.add(model)\n      console.warn(\n        `estimateCost: no pricing for model \"${model}\" — returning 0; add it to ` +\n          'MODEL_PRICING/FAMILY_PRICING (cost/Pareto axes will be blank until then)',\n      )\n    }\n    return 0\n  }\n  return (inputTokens / 1000) * pricing.input + (outputTokens / 1000) * pricing.output\n}\n\n/**\n * TokenCounter — accumulates token usage and cost across turns.\n */\nexport class TokenCounter {\n  private totalInput = 0\n  private totalOutput = 0\n  private totalCost = 0\n  private model: string\n\n  constructor(model = 'gpt-4o') {\n    this.model = model\n  }\n\n  /** Record tokens for a turn, returns per-turn cost */\n  record(inputTokens: number, outputTokens: number): number {\n    this.totalInput += inputTokens\n    this.totalOutput += outputTokens\n    const cost = estimateCost(inputTokens, outputTokens, this.model)\n    this.totalCost += cost\n    return cost\n  }\n\n  /** Estimate and record from raw text */\n  recordFromText(\n    inputText: string,\n    outputText: string,\n  ): { inputTokens: number; outputTokens: number; cost: number } {\n    const inputTokens = estimateTokens(inputText)\n    const outputTokens = estimateTokens(outputText)\n    const cost = this.record(inputTokens, outputTokens)\n    return { inputTokens, outputTokens, cost }\n  }\n\n  getTotalInput(): number {\n    return this.totalInput\n  }\n  getTotalOutput(): number {\n    return this.totalOutput\n  }\n  getTotalCost(): number {\n    return this.totalCost\n  }\n}\n"],"mappings":";;AAMA,MAAa,gBAA4C;CACvD,UAAU;EAAE,OAAO;EAAQ,QAAQ;CAAK;CACxC,eAAe;EAAE,OAAO;EAAS,QAAQ;CAAO;CAChD,eAAe;EAAE,OAAO;EAAM,QAAQ;CAAK;CAC3C,4BAA4B;EAAE,OAAO;EAAO,QAAQ;CAAM;CAC1D,0BAA0B;EAAE,OAAO;EAAO,QAAQ;CAAM;CACxD,2BAA2B;EAAE,OAAO;EAAS,QAAQ;CAAQ;CAI7D,WAAW;EAAE,OAAO;EAAS,QAAQ;CAAQ;AAC/C;;;;;;;;AASA,MAAM,iBAA8C;CAClD,CAAC,gBAAgB;EAAE,OAAO;EAAO,QAAQ;CAAM,CAAC;CAChD,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAM,CAAC;CAClD,CAAC,4CAA4C;EAAE,OAAO;EAAO,QAAQ;CAAM,CAAC;CAC5E,CAAC,eAAe;EAAE,OAAO;EAAS,QAAQ;CAAO,CAAC;CAClD,CAAC,2BAA2B;EAAE,OAAO;EAAS,QAAQ;CAAK,CAAC;CAC5D,CAAC,gBAAgB;EAAE,OAAO;EAAQ,QAAQ;CAAK,CAAC;CAChD,CAAC,YAAY;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CAC9C,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CACnD,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CACnD,CAAC,QAAQ;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CAC1C,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CACnD,CAAC,UAAU;EAAE,OAAO;EAAS,QAAQ;CAAM,CAAC;CAC5C,CAAC,SAAS;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;AAC7C;;;;;AAMA,SAAS,mBAAmB,OAAyB;CACnD,MAAM,MAAM,MAAM,MAAM,GAAG,CAAC,CAAC,MAAM,MAAA,CAAO,KAAK,CAAC,CAAC,YAAY;CAC7D,MAAM,eAAe,GAAG,MAAM,GAAG,YAAY,GAAG,IAAI,CAAC;CACrD,OAAO,iBAAiB,KAAK,CAAC,EAAE,IAAI,CAAC,IAAI,YAAY;AACvD;;;;AAKA,SAAgB,oBAAoB,OAAkC;CACpE,IAAI,cAAc,QAAQ,OAAO,cAAc;CAC/C,MAAM,MAAM,mBAAmB,KAAK;CACpC,KAAK,MAAM,MAAM,KACf,IAAI,cAAc,KAAK,OAAO,cAAc;CAE9C,KAAK,MAAM,CAAC,SAAS,UAAU,gBAC7B,IAAI,IAAI,MAAM,OAAO,QAAQ,KAAK,EAAE,CAAC,GAAG,OAAO;CAEjD,OAAO;AACT;;;AAIA,SAAgB,cAAc,OAAwB;CACpD,OAAO,oBAAoB,KAAK,MAAM;AACxC;AAEA,MAAM,uCAAuB,IAAI,IAAY;;AAG7C,SAAgB,eAAe,MAAsB;CACnD,OAAO,KAAK,KAAK,KAAK,SAAS,CAAC;AAClC;;;;AAKA,SAAgB,aAAa,aAAqB,cAAsB,OAAuB;CAC7F,MAAM,UAAU,oBAAoB,KAAK;CACzC,IAAI,CAAC,SAAS;EACZ,IAAI,CAAC,qBAAqB,IAAI,KAAK,GAAG;GACpC,qBAAqB,IAAI,KAAK;GAC9B,QAAQ,KACN,uCAAuC,MAAM,oGAE/C;EACF;EACA,OAAO;CACT;CACA,OAAQ,cAAc,MAAQ,QAAQ,QAAS,eAAe,MAAQ,QAAQ;AAChF"}