@tangle-network/agent-runtime 0.196.0 → 0.198.1

package/dist/{delegate-D7sonS78.js → delegate-Bez07Dlh.js}

@@ -1,5 +1,5 @@
 import { i as ConfigError } from "./errors-DodWX-cb.js";
-import { n as supervise } from "./supervise-D9aNi8_f.js";
+import { n as supervise } from "./supervise-KSe4JEcg.js";
 import { agentProfileSchema } from "@tangle-network/agent-interface";
 import { computeFindingId, makeFinding } from "@tangle-network/agent-eval";
 //#region src/runtime/supervise/authoring.ts
@@ -207,4 +207,4 @@ async function delegate(intent, opts) {
 //#endregion
 export { defaultProfileRichnessThresholds as a, assessAuthoredProfile as i, delegate as n, profileRichnessFinding as o, asAuthoredProfile as r, supervisorInstructions as s, defaultDelegateBudget as t };
 
-//# sourceMappingURL=delegate-D7sonS78.js.map
+//# sourceMappingURL=delegate-Bez07Dlh.js.map

package/dist/{delegate-D7sonS78.js.map → delegate-Bez07Dlh.js.map}

@@ -1 +1 @@
-{"version":3,"file":"delegate-D7sonS78.js","names":[],"sources":["../src/runtime/supervise/authoring.ts","../src/runtime/supervise/delegate.ts"],"sourcesContent":["/**\n *\n * The supervisor's intelligence is AUTHORING the agents it spawns — not pressing buttons.\n *\n * Every agent here is three things: instructions (system prompt), tools, and a model — its\n * `AgentProfile`. The supervisor's job is to WRITE those profiles: read the task, decompose it,\n * and for each sub-task author a tailored worker recipe. `supervisorInstructions` is the how-to the\n * supervisor reads; canonical Runtime executors materialize the resulting profile.\n *\n * The skill is the single OPTIMIZABLE surface: edit it → the supervisor designs better agents.\n * That is the self-improvement lever (the prompt/skill lever), not the execution plumbing.\n *\n * @experimental\n */\n\nimport { type AnalystFinding, computeFindingId, makeFinding } from '@tangle-network/agent-eval'\nimport {\n  type AgentProfile,\n  type AgentProfilePrompt,\n  agentProfileSchema,\n} from '@tangle-network/agent-interface'\n\n/** What the supervisor AUTHORS per sub-task: one complete canonical profile whose name and\n *  task-specific system prompt are present. Every other `AgentProfile` axis is preserved exactly. */\nexport type AuthoredProfile = AgentProfile & {\n  readonly name: string\n  readonly prompt: AgentProfilePrompt & { readonly systemPrompt: string }\n}\n\n/** Narrow an untyped `spawn_worker` profile argument to an `AuthoredProfile`, or null if the\n *  supervisor failed to author one (empty/placeholder profile — a skill violation worth catching). */\nexport function asAuthoredProfile(raw: unknown): AuthoredProfile | null {\n  const parsed = agentProfileSchema.safeParse(raw)\n  if (!parsed.success) return null\n  const systemPrompt = parsed.data.prompt?.systemPrompt\n  if (typeof systemPrompt !== 'string' || systemPrompt.trim().length === 0) return null\n  return {\n    ...parsed.data,\n    name:\n      typeof parsed.data.name === 'string' && parsed.data.name.length > 0\n        ? parsed.data.name\n        : 'worker',\n    prompt: { ...parsed.data.prompt, systemPrompt },\n  }\n}\n\n/** The supervisor skill: an explicit profile-authoring instruction, never an implicit Runtime\n * policy. Editing this text changes how a profile designs the descendants it spawns. */\nexport function supervisorInstructions(opts?: { goal?: string }): string {\n  return [\n    'Your delegation craft is AUTHORING: a spawned worker is exactly as good as the profile you write.',\n    '',\n    'For the task you are given:',\n    '1. DECOMPOSE it into the smallest set of sub-tasks a single focused worker can each deliver.',\n    '2. For EACH sub-task, AUTHOR a worker by calling spawn_worker with a COMPLETE `profile`:',\n    '   • name and description: who this specialist is and why it exists.',\n    '   • prompt.systemPrompt: rich instructions for THIS sub-task — exact output, process, evidence, and what \"done\" means.',\n    '   • model.default, model.reasoningEffort, and harness: choose the execution system deliberately when the task benefits from it.',\n    '   • tools, mcp, resources.skills/files/instructions, hooks, subagents, permissions, and modes: grant or attach every capability the worker needs; omit an axis only when it is intentionally unnecessary.',\n    '   • tools.agent_runtime_coordination_spawn_worker: true ONLY when this child should author and drive descendants. Add only the other agent_runtime_coordination_<verb> tools it will call, such as await_event or steer_agent.',\n    '   • A child with spawn_worker MUST carry this complete profile-authoring instruction as an immutable resources.skills entry with resources.failOnError: true, so it can author its own descendants from the same contract.',\n    '   • metadata may describe the work, but it never grants recursion or selects a Runtime execution path.',\n    '   NEVER spawn a worker with an empty profile. The quality of the worker IS the quality of the profile you write.',\n    \"3. await_event (kinds:['settled']) to collect each worker. Its result says valid:true only if the deployable check passed.\",\n    '4. If a worker did NOT deliver, AUTHOR A NEW profile whose prompt.systemPrompt names the SPECIFIC failure and how to fix it — never just retry the same profile.',\n    \"5. read_journal to re-read YOUR OWN record before you decide the next move: every spawn you made, every settle, every question and answer, every steer, every analyst finding — oldest first, this node only, including what you did before a restart. Use it to see what you already tried instead of trying it again. It is paged: pass the returned nextRow as the next call's sinceRow, narrow with kinds, and raise limit/maxBytes only as far as you will actually read. A truncated:true page means a bound cut it short — keep paging before you conclude you have read everything.\",\n    '6. AUTHOR YOUR OWN LENS when the questions you can already ask of a settled trace do not cover the failure you are chasing: define_analyst takes an id, a description, an area, the question in your own words, the instructions for answering it with trace evidence, and the smallest toolGroup that can answer it (model is the seat it runs on; omit it for the run default). It is DATA, never code. Then run_analyst it on any settled worker like a lens the run shipped with, and read the finding. list_analysts shows what you have. Define a lens when you need a different question asked — not a second copy of a question already on the menu.',\n    '7. EVERY refusal you get back carries a `reason` naming the exact unmet condition. Read it and change that condition — a spawn refused for max-live-workers needs an await_event, an invalid-profile needs the named field fixed, a submit_result refused because the check THREW is a broken check to report, not a result to resubmit. Never repeat a call that was refused without changing what it was refused for.',\n    '8. ask_parent ONLY when you genuinely cannot decide, and then READ ITS OUTCOME. \"queued-for-parent\" means an inbox above you now holds the question. \"no-parent\" means no inbox above you is configured to receive it: the question is still on the run record for anyone watching, but nothing will route an answer back to you, so do not block. Decide it with answer_question, or answer_question with deferReason to record that it stays open, and carry on — a blocking question left undecided also refuses your stop.',\n    '9. Stop (reply with no tool call) once the work is delivered.',\n    ...(opts?.goal ? ['', `The goal: ${opts.goal}`] : []),\n  ].join('\\n')\n}\n\n// ── Profile-richness gate ────────────────────────────────────────────────────\n//\n// The supervisor's product is the worker PROFILE it authors. The failure mode the existing\n// gates miss: `asAuthoredProfile` / `local-harness` only reject a FULLY EMPTY system prompt —\n// a two-sentence stub passes. `assessAuthoredProfile` OBSERVES the authored artifact (it reads\n// no judge verdict, so it steers cleanly past `assertTraceDerivedFindings`) and flags THIN:\n// a short/few-line system prompt, OR no tools, OR no skills, OR no MCP when the task needs one.\n// It emits a real `AnalystFinding` so it rides the SAME coordination bus the driver pulls via\n// `await_event({kinds:['finding']})` — the supervisor can self-correct and re-author richer.\n\n/** Thresholds below which a system prompt is treated as a thin stub. Tunable per call. */\nexport interface ProfileRichnessThresholds {\n  /** A prompt shorter than this many characters is thin (default 600). */\n  readonly minSystemPromptChars: number\n  /** A prompt with fewer than this many non-blank lines is thin (default 6). */\n  readonly minSystemPromptLines: number\n}\n\n/** Default thresholds for `ProfileRichnessThresholds` — 600 chars / 6 lines minimum system prompt. */\nexport const defaultProfileRichnessThresholds: ProfileRichnessThresholds = {\n  minSystemPromptChars: 600,\n  minSystemPromptLines: 6,\n}\n\n/** Per-field verdict on one authored profile — the raw material the bench renders + scores. */\nexport interface ProfileRichness {\n  readonly name: string\n  /** The resolved system prompt (canonical `prompt.systemPrompt`, the sandbox `prompt.system`\n   *  convention, or a bare-string prompt — whichever the author used). */\n  readonly systemPrompt: string\n  readonly systemPromptChars: number\n  readonly systemPromptLines: number\n  readonly sentenceCount: number\n  readonly hasDescription: boolean\n  readonly hasTools: boolean\n  readonly hasSkills: boolean\n  readonly hasMcp: boolean\n  readonly hasSubagents: boolean\n  /** 0..1 — fraction of richness signals present (prompt-depth + the four levers). */\n  readonly richness: number\n  /** True when the supervisor authored a stub instead of a real profile. */\n  readonly thin: boolean\n  /** The specific reasons it is thin (empty when rich) — used in the finding's action. */\n  readonly reasons: string[]\n}\n\n/** Read the system prompt from any authored shape: canonical `prompt.systemPrompt`, the sandbox\n *  `prompt.system` convention, or a bare-string `prompt`. */\nfunction resolveSystemPrompt(profile: AgentProfile): string {\n  const pr = (profile as { prompt?: unknown }).prompt\n  if (typeof pr === 'string') return pr\n  if (pr && typeof pr === 'object') {\n    const o = pr as { systemPrompt?: unknown; system?: unknown }\n    if (typeof o.systemPrompt === 'string') return o.systemPrompt\n    if (typeof o.system === 'string') return o.system\n  }\n  return ''\n}\n\n/** OBSERVE one authored `AgentProfile` and score its richness (no judge verdict is read). The task\n *  context (`needsMcp`) lets a domain say \"this work needs a data/tool MCP\" so a missing MCP counts. */\nexport function assessAuthoredProfile(\n  profile: AgentProfile,\n  opts?: { needsMcp?: boolean; thresholds?: Partial<ProfileRichnessThresholds> },\n): ProfileRichness {\n  const th = { ...defaultProfileRichnessThresholds, ...(opts?.thresholds ?? {}) }\n  const systemPrompt = resolveSystemPrompt(profile)\n  const trimmed = systemPrompt.trim()\n  const systemPromptChars = trimmed.length\n  const systemPromptLines = trimmed\n    ? trimmed.split('\\n').filter((l) => l.trim().length > 0).length\n    : 0\n  const sentenceCount = trimmed\n    ? (trimmed.match(/[.!?](\\s|$)/g) ?? []).length || (trimmed ? 1 : 0)\n    : 0\n  const hasDescription =\n    typeof profile.description === 'string' && profile.description.trim().length > 0\n  const tools = (profile as { tools?: Record<string, unknown> }).tools\n  const hasTools = !!tools && Object.keys(tools).length > 0\n  const skills = (profile.resources as { skills?: unknown[] } | undefined)?.skills\n  const hasSkills = Array.isArray(skills) && skills.length > 0\n  const mcp = (profile as { mcp?: Record<string, unknown> }).mcp\n  const hasMcp = !!mcp && Object.keys(mcp).length > 0\n  const subagents = (profile as { subagents?: Record<string, unknown> }).subagents\n  const hasSubagents = !!subagents && Object.keys(subagents).length > 0\n\n  const reasons: string[] = []\n  const promptThin =\n    systemPromptChars < th.minSystemPromptChars || systemPromptLines < th.minSystemPromptLines\n  if (promptThin)\n    reasons.push(\n      `system prompt is thin (${systemPromptChars} chars, ${systemPromptLines} lines; need ≥${th.minSystemPromptChars} chars and ≥${th.minSystemPromptLines} lines)`,\n    )\n  if (!hasTools)\n    reasons.push('no tools granted (a worker can only act through the tools you grant it)')\n  if (!hasSkills) reasons.push('no skills attached (no reusable how-to notes injected)')\n  if (opts?.needsMcp && !hasMcp) reasons.push('no MCP server, but the task needs data/tool access')\n\n  // Richness = fraction of signals present. Prompt-depth is one signal; the four levers are the rest.\n  const signals = [!promptThin, hasTools, hasSkills, hasDescription, opts?.needsMcp ? hasMcp : true]\n  const richness = signals.filter(Boolean).length / signals.length\n  // THIN ⟺ the prompt is a stub OR the worker has no levers at all (no tools AND no skills AND no mcp).\n  const thin = promptThin || (!hasTools && !hasSkills && !hasMcp)\n\n  return {\n    name: profile.name ?? 'worker',\n    systemPrompt,\n    systemPromptChars,\n    systemPromptLines,\n    sentenceCount,\n    hasDescription,\n    hasTools,\n    hasSkills,\n    hasMcp,\n    hasSubagents,\n    richness,\n    thin,\n    reasons,\n  }\n}\n\n/** Turn a {@link ProfileRichness} verdict into a bus-routable `AnalystFinding` (area `profile-quality`).\n *  Severity scales with thinness; the recommended action names the MISSING lever so the supervisor can\n *  re-author. `subject` = the worker name so per-worker findings diff cleanly across re-authors. */\nexport function profileRichnessFinding(\n  richness: ProfileRichness,\n  opts?: { analystId?: string; runId?: string },\n): AnalystFinding {\n  const analyst_id = opts?.analystId ?? 'profile-richness'\n  const subject = richness.name\n  const claim = richness.thin\n    ? `Worker \"${richness.name}\" was authored as a THIN profile: ${richness.reasons.join('; ')}.`\n    : `Worker \"${richness.name}\" was authored as a rich profile (richness ${(richness.richness * 100).toFixed(0)}%).`\n  const severity: AnalystFinding['severity'] = richness.thin\n    ? richness.richness < 0.25\n      ? 'high'\n      : 'medium'\n    : 'info'\n  return makeFinding({\n    analyst_id,\n    severity,\n    area: 'profile-quality',\n    claim,\n    subject,\n    confidence: 0.9,\n    evidence_refs: [\n      {\n        kind: 'metric',\n        uri: `profile:${subject}`,\n        excerpt: `chars=${richness.systemPromptChars} lines=${richness.systemPromptLines} tools=${richness.hasTools} skills=${richness.hasSkills} mcp=${richness.hasMcp} richness=${richness.richness.toFixed(2)}`,\n      },\n    ],\n    ...(richness.thin\n      ? { recommended_action: `Re-author \"${richness.name}\" with: ${richness.reasons.join('; ')}.` }\n      : {}),\n    id_basis: computeFindingId({\n      analyst_id,\n      area: 'profile-quality',\n      subject,\n      claim: `richness:${richness.thin ? 'thin' : 'rich'}`,\n    }),\n  })\n}\n","/**\n *\n * `delegate` — the one generic delegation verb. You hand it an INTENT (what you want done) and it\n * hands that intent to a default AUTHORING supervisor: a router-brained supervisor whose standing\n * instruction is `supervisorInstructions()` (the authoring-agent-profiles skill). The supervisor\n * DECOMPOSES the intent and AUTHORS the worker profile it needs per sub-task — there is NO hardcoded\n * coder/researcher profile here. That is the whole point: `delegate('fix the failing test', …)` and\n * `delegate('research X and cite sources', …)` route through the SAME front door; the supervisor\n * writes a code-shaped or research-shaped worker on its own.\n *\n * It is a thin wrapper over `supervise()` — the one front door — so the conserved-budget pool, the\n * completion oracle (`deliverable`), the coordination toolbox, and equal-compute accounting all come\n * for free; nothing is hand-rolled. The result is `supervise()`'s `SupervisedResult` returned\n * UNCHANGED, so its `spentTotal` (`{ iterations, tokens, usd, ms }`) rides straight back to the\n * caller on BOTH paths — a `winner` carries the delivered worker's spend, a `no-winner` carries the\n * spend incurred before it failed. That cost channel means a `delegate()` caller always learns what\n * the delegation actually spent.\n *\n * @experimental\n */\n\nimport { ConfigError } from '../../errors'\nimport type { RouterTransportConfig } from '../router-client'\nimport type { DeliverableSpec } from './completion-gate'\nimport type { ExecutorConfig } from './runtime'\nimport { supervise } from './supervise'\nimport type { SupervisorProfile } from './supervisor-agent'\nimport type { Budget, SupervisedResult } from './types'\n\n/** The conserved pool a `delegate()` call applies when the caller does not pass its own `budget`.\n *  A modest token ceiling + a small iteration ceiling — generous enough for a few-worker decompose,\n *  bounded enough that an unsupervised intent cannot run away. Callers override via `opts.budget`. */\nexport const defaultDelegateBudget: Budget = { maxIterations: 50, maxTokens: 200_000 }\n\n/** Inputs to {@link delegate}. The intent is the first positional arg; everything here is optional\n *  with explicit execution identity, so the common call names one exact supervisor profile. */\nexport interface DelegateOptions<Out = unknown> {\n  /** The completion oracle (settled ⟺ delivered) the authored workers settle against. Strongly\n   *  recommended — without it the supervisor trusts a worker's self-report. For a code intent,\n   *  `patchDelivered()` is the canonical example; for a free-form answer, a content check. */\n  readonly deliverable?: DeliverableSpec<Out>\n  /** WHERE the authored workers run — the worker-execution backend (`router-tools` / `sandbox` /\n   *  `cli-worktree` / …). The supervisor authors the worker PROFILE; this is the substrate it runs\n   *  on. Provide this OR `makeWorkerAgent`-style wiring through `supervise()` is unavailable. */\n  readonly backend?: ExecutorConfig\n  /** The conserved compute pool for the whole delegation. Defaults to {@link defaultDelegateBudget}. */\n  readonly budget?: Budget\n  /** Exact executable authoring supervisor. Model, prompt, harness, and provider live here. */\n  readonly supervisorProfile: SupervisorProfile\n  /** Router endpoint/auth for a `cli-base` supervisor; contains no behavioral settings. */\n  readonly router: RouterTransportConfig\n  /** Restrict the run to this subset of models (forwarded to `supervise()`). */\n  readonly allowedModels?: readonly string[]\n  readonly runId?: string\n}\n\n/**\n * Delegate an INTENT to a default authoring supervisor and return its `SupervisedResult` unchanged.\n *\n * The supervisor authors + spawns whatever worker the intent needs over the conserved-budget pool;\n * `result.spentTotal` reports what the whole delegation actually cost. A `winner` result carries the\n * authored worker's delivered output; a `no-winner` result names why (never a fabricated success).\n */\nexport async function delegate<Out = unknown>(\n  intent: string,\n  opts: DelegateOptions<Out>,\n): Promise<SupervisedResult<Out>> {\n  if (typeof intent !== 'string' || intent.trim().length === 0) {\n    throw new ConfigError('delegate: `intent` must be a non-empty string')\n  }\n  return supervise(opts.supervisorProfile, intent, {\n    budget: opts.budget ?? defaultDelegateBudget,\n    ...(opts.backend ? { backend: opts.backend } : {}),\n    ...(opts.deliverable ? { deliverable: opts.deliverable as DeliverableSpec<unknown> } : {}),\n    router: opts.router,\n    ...(opts.allowedModels ? { allowedModels: opts.allowedModels } : {}),\n    ...(opts.runId ? { runId: opts.runId } : {}),\n  }) as Promise<SupervisedResult<Out>>\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;AA+BA,SAAgB,kBAAkB,KAAsC;CACtE,MAAM,SAAS,mBAAmB,UAAU,GAAG;CAC/C,IAAI,CAAC,OAAO,SAAS,OAAO;CAC5B,MAAM,eAAe,OAAO,KAAK,QAAQ;CACzC,IAAI,OAAO,iBAAiB,YAAY,aAAa,KAAK,CAAC,CAAC,WAAW,GAAG,OAAO;CACjF,OAAO;EACL,GAAG,OAAO;EACV,MACE,OAAO,OAAO,KAAK,SAAS,YAAY,OAAO,KAAK,KAAK,SAAS,IAC9D,OAAO,KAAK,OACZ;EACN,QAAQ;GAAE,GAAG,OAAO,KAAK;GAAQ;EAAa;CAChD;AACF;;;AAIA,SAAgB,uBAAuB,MAAkC;CACvE,OAAO;EACL;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,GAAI,MAAM,OAAO,CAAC,IAAI,aAAa,KAAK,MAAM,IAAI,CAAC;CACrD,CAAC,CAAC,KAAK,IAAI;AACb;;AAqBA,MAAa,mCAA8D;CACzE,sBAAsB;CACtB,sBAAsB;AACxB;;;AA0BA,SAAS,oBAAoB,SAA+B;CAC1D,MAAM,KAAM,QAAiC;CAC7C,IAAI,OAAO,OAAO,UAAU,OAAO;CACnC,IAAI,MAAM,OAAO,OAAO,UAAU;EAChC,MAAM,IAAI;EACV,IAAI,OAAO,EAAE,iBAAiB,UAAU,OAAO,EAAE;EACjD,IAAI,OAAO,EAAE,WAAW,UAAU,OAAO,EAAE;CAC7C;CACA,OAAO;AACT;;;AAIA,SAAgB,sBACd,SACA,MACiB;CACjB,MAAM,KAAK;EAAE,GAAG;EAAkC,GAAI,MAAM,cAAc,CAAC;CAAG;CAC9E,MAAM,eAAe,oBAAoB,OAAO;CAChD,MAAM,UAAU,aAAa,KAAK;CAClC,MAAM,oBAAoB,QAAQ;CAClC,MAAM,oBAAoB,UACtB,QAAQ,MAAM,IAAI,CAAC,CAAC,QAAQ,MAAM,EAAE,KAAK,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,SACvD;CACJ,MAAM,gBAAgB,WACjB,QAAQ,MAAM,cAAc,KAAK,CAAC,EAAA,CAAG,WAAW,UAAU,IAAI,KAC/D;CACJ,MAAM,iBACJ,OAAO,QAAQ,gBAAgB,YAAY,QAAQ,YAAY,KAAK,CAAC,CAAC,SAAS;CACjF,MAAM,QAAS,QAAgD;CAC/D,MAAM,WAAW,CAAC,CAAC,SAAS,OAAO,KAAK,KAAK,CAAC,CAAC,SAAS;CACxD,MAAM,SAAU,QAAQ,WAAkD;CAC1E,MAAM,YAAY,MAAM,QAAQ,MAAM,KAAK,OAAO,SAAS;CAC3D,MAAM,MAAO,QAA8C;CAC3D,MAAM,SAAS,CAAC,CAAC,OAAO,OAAO,KAAK,GAAG,CAAC,CAAC,SAAS;CAClD,MAAM,YAAa,QAAoD;CACvE,MAAM,eAAe,CAAC,CAAC,aAAa,OAAO,KAAK,SAAS,CAAC,CAAC,SAAS;CAEpE,MAAM,UAAoB,CAAC;CAC3B,MAAM,aACJ,oBAAoB,GAAG,wBAAwB,oBAAoB,GAAG;CACxE,IAAI,YACF,QAAQ,KACN,0BAA0B,kBAAkB,UAAU,kBAAkB,gBAAgB,GAAG,qBAAqB,cAAc,GAAG,qBAAqB,QACxJ;CACF,IAAI,CAAC,UACH,QAAQ,KAAK,yEAAyE;CACxF,IAAI,CAAC,WAAW,QAAQ,KAAK,wDAAwD;CACrF,IAAI,MAAM,YAAY,CAAC,QAAQ,QAAQ,KAAK,oDAAoD;CAGhG,MAAM,UAAU;EAAC,CAAC;EAAY;EAAU;EAAW;EAAgB,MAAM,WAAW,SAAS;CAAI;CACjG,MAAM,WAAW,QAAQ,OAAO,OAAO,CAAC,CAAC,SAAS,QAAQ;CAE1D,MAAM,OAAO,cAAe,CAAC,YAAY,CAAC,aAAa,CAAC;CAExD,OAAO;EACL,MAAM,QAAQ,QAAQ;EACtB;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;CACF;AACF;;;;AAKA,SAAgB,uBACd,UACA,MACgB;CAChB,MAAM,aAAa,MAAM,aAAa;CACtC,MAAM,UAAU,SAAS;CACzB,MAAM,QAAQ,SAAS,OACnB,WAAW,SAAS,KAAK,oCAAoC,SAAS,QAAQ,KAAK,IAAI,EAAE,KACzF,WAAW,SAAS,KAAK,8CAA8C,SAAS,WAAW,IAAA,CAAK,QAAQ,CAAC,EAAE;CAM/G,OAAO,YAAY;EACjB;EACA,UAP2C,SAAS,OAClD,SAAS,WAAW,MAClB,SACA,WACF;EAIF,MAAM;EACN;EACA;EACA,YAAY;EACZ,eAAe,CACb;GACE,MAAM;GACN,KAAK,WAAW;GAChB,SAAS,SAAS,SAAS,kBAAkB,SAAS,SAAS,kBAAkB,SAAS,SAAS,SAAS,UAAU,SAAS,UAAU,OAAO,SAAS,OAAO,YAAY,SAAS,SAAS,QAAQ,CAAC;EACzM,CACF;EACA,GAAI,SAAS,OACT,EAAE,oBAAoB,cAAc,SAAS,KAAK,UAAU,SAAS,QAAQ,KAAK,IAAI,EAAE,GAAG,IAC3F,CAAC;EACL,UAAU,iBAAiB;GACzB;GACA,MAAM;GACN;GACA,OAAO,YAAY,SAAS,OAAO,SAAS;EAC9C,CAAC;CACH,CAAC;AACH;;;;;;;;;;;;;;;;;;;;;;;;;;AC5MA,MAAa,wBAAgC;CAAE,eAAe;CAAI,WAAW;AAAQ;;;;;;;;AA+BrF,eAAsB,SACpB,QACA,MACgC;CAChC,IAAI,OAAO,WAAW,YAAY,OAAO,KAAK,CAAC,CAAC,WAAW,GACzD,MAAM,IAAI,YAAY,+CAA+C;CAEvE,OAAO,UAAU,KAAK,mBAAmB,QAAQ;EAC/C,QAAQ,KAAK,UAAU;EACvB,GAAI,KAAK,UAAU,EAAE,SAAS,KAAK,QAAQ,IAAI,CAAC;EAChD,GAAI,KAAK,cAAc,EAAE,aAAa,KAAK,YAAwC,IAAI,CAAC;EACxF,QAAQ,KAAK;EACb,GAAI,KAAK,gBAAgB,EAAE,eAAe,KAAK,cAAc,IAAI,CAAC;EAClE,GAAI,KAAK,QAAQ,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;CAC5C,CAAC;AACH"}
+{"version":3,"file":"delegate-Bez07Dlh.js","names":[],"sources":["../src/runtime/supervise/authoring.ts","../src/runtime/supervise/delegate.ts"],"sourcesContent":["/**\n *\n * The supervisor's intelligence is AUTHORING the agents it spawns — not pressing buttons.\n *\n * Every agent here is three things: instructions (system prompt), tools, and a model — its\n * `AgentProfile`. The supervisor's job is to WRITE those profiles: read the task, decompose it,\n * and for each sub-task author a tailored worker recipe. `supervisorInstructions` is the how-to the\n * supervisor reads; canonical Runtime executors materialize the resulting profile.\n *\n * The skill is the single OPTIMIZABLE surface: edit it → the supervisor designs better agents.\n * That is the self-improvement lever (the prompt/skill lever), not the execution plumbing.\n *\n * @experimental\n */\n\nimport { type AnalystFinding, computeFindingId, makeFinding } from '@tangle-network/agent-eval'\nimport {\n  type AgentProfile,\n  type AgentProfilePrompt,\n  agentProfileSchema,\n} from '@tangle-network/agent-interface'\n\n/** What the supervisor AUTHORS per sub-task: one complete canonical profile whose name and\n *  task-specific system prompt are present. Every other `AgentProfile` axis is preserved exactly. */\nexport type AuthoredProfile = AgentProfile & {\n  readonly name: string\n  readonly prompt: AgentProfilePrompt & { readonly systemPrompt: string }\n}\n\n/** Narrow an untyped `spawn_worker` profile argument to an `AuthoredProfile`, or null if the\n *  supervisor failed to author one (empty/placeholder profile — a skill violation worth catching). */\nexport function asAuthoredProfile(raw: unknown): AuthoredProfile | null {\n  const parsed = agentProfileSchema.safeParse(raw)\n  if (!parsed.success) return null\n  const systemPrompt = parsed.data.prompt?.systemPrompt\n  if (typeof systemPrompt !== 'string' || systemPrompt.trim().length === 0) return null\n  return {\n    ...parsed.data,\n    name:\n      typeof parsed.data.name === 'string' && parsed.data.name.length > 0\n        ? parsed.data.name\n        : 'worker',\n    prompt: { ...parsed.data.prompt, systemPrompt },\n  }\n}\n\n/** The supervisor skill: an explicit profile-authoring instruction, never an implicit Runtime\n * policy. Editing this text changes how a profile designs the descendants it spawns. */\nexport function supervisorInstructions(opts?: { goal?: string }): string {\n  return [\n    'Your delegation craft is AUTHORING: a spawned worker is exactly as good as the profile you write.',\n    '',\n    'For the task you are given:',\n    '1. DECOMPOSE it into the smallest set of sub-tasks a single focused worker can each deliver.',\n    '2. For EACH sub-task, AUTHOR a worker by calling spawn_worker with a COMPLETE `profile`:',\n    '   • name and description: who this specialist is and why it exists.',\n    '   • prompt.systemPrompt: rich instructions for THIS sub-task — exact output, process, evidence, and what \"done\" means.',\n    '   • model.default, model.reasoningEffort, and harness: choose the execution system deliberately when the task benefits from it.',\n    '   • tools, mcp, resources.skills/files/instructions, hooks, subagents, permissions, and modes: grant or attach every capability the worker needs; omit an axis only when it is intentionally unnecessary.',\n    '   • tools.agent_runtime_coordination_spawn_worker: true ONLY when this child should author and drive descendants. Add only the other agent_runtime_coordination_<verb> tools it will call, such as await_event or steer_agent.',\n    '   • A child with spawn_worker MUST carry this complete profile-authoring instruction as an immutable resources.skills entry with resources.failOnError: true, so it can author its own descendants from the same contract.',\n    '   • metadata may describe the work, but it never grants recursion or selects a Runtime execution path.',\n    '   NEVER spawn a worker with an empty profile. The quality of the worker IS the quality of the profile you write.',\n    \"3. await_event (kinds:['settled']) to collect each worker. Its result says valid:true only if the deployable check passed.\",\n    '4. If a worker did NOT deliver, AUTHOR A NEW profile whose prompt.systemPrompt names the SPECIFIC failure and how to fix it — never just retry the same profile.',\n    \"5. read_journal to re-read YOUR OWN record before you decide the next move: every spawn you made, every settle, every question and answer, every steer, every analyst finding — oldest first, this node only, including what you did before a restart. Use it to see what you already tried instead of trying it again. It is paged: pass the returned nextRow as the next call's sinceRow, narrow with kinds, and raise limit/maxBytes only as far as you will actually read. A truncated:true page means a bound cut it short — keep paging before you conclude you have read everything.\",\n    '6. AUTHOR YOUR OWN LENS when the questions you can already ask of a settled trace do not cover the failure you are chasing: define_analyst takes an id, a description, an area, the question in your own words, the instructions for answering it with trace evidence, and the smallest toolGroup that can answer it (model is the seat it runs on; omit it for the run default). It is DATA, never code. Then run_analyst it on any settled worker like a lens the run shipped with, and read the finding. list_analysts shows what you have. Define a lens when you need a different question asked — not a second copy of a question already on the menu.',\n    '7. EVERY refusal you get back carries a `reason` naming the exact unmet condition. Read it and change that condition — a spawn refused for max-live-workers needs an await_event, an invalid-profile needs the named field fixed, a submit_result refused because the check THREW is a broken check to report, not a result to resubmit. Never repeat a call that was refused without changing what it was refused for.',\n    '8. ask_parent ONLY when you genuinely cannot decide, and then READ ITS OUTCOME. \"queued-for-parent\" means an inbox above you now holds the question. \"no-parent\" means no inbox above you is configured to receive it: the question is still on the run record for anyone watching, but nothing will route an answer back to you, so do not block. Decide it with answer_question, or answer_question with deferReason to record that it stays open, and carry on — a blocking question left undecided also refuses your stop.',\n    '9. Stop (reply with no tool call) once the work is delivered.',\n    ...(opts?.goal ? ['', `The goal: ${opts.goal}`] : []),\n  ].join('\\n')\n}\n\n// ── Profile-richness gate ────────────────────────────────────────────────────\n//\n// The supervisor's product is the worker PROFILE it authors. The failure mode the existing\n// gates miss: `asAuthoredProfile` / `local-harness` only reject a FULLY EMPTY system prompt —\n// a two-sentence stub passes. `assessAuthoredProfile` OBSERVES the authored artifact (it reads\n// no judge verdict, so it steers cleanly past `assertTraceDerivedFindings`) and flags THIN:\n// a short/few-line system prompt, OR no tools, OR no skills, OR no MCP when the task needs one.\n// It emits a real `AnalystFinding` so it rides the SAME coordination bus the driver pulls via\n// `await_event({kinds:['finding']})` — the supervisor can self-correct and re-author richer.\n\n/** Thresholds below which a system prompt is treated as a thin stub. Tunable per call. */\nexport interface ProfileRichnessThresholds {\n  /** A prompt shorter than this many characters is thin (default 600). */\n  readonly minSystemPromptChars: number\n  /** A prompt with fewer than this many non-blank lines is thin (default 6). */\n  readonly minSystemPromptLines: number\n}\n\n/** Default thresholds for `ProfileRichnessThresholds` — 600 chars / 6 lines minimum system prompt. */\nexport const defaultProfileRichnessThresholds: ProfileRichnessThresholds = {\n  minSystemPromptChars: 600,\n  minSystemPromptLines: 6,\n}\n\n/** Per-field verdict on one authored profile — the raw material the bench renders + scores. */\nexport interface ProfileRichness {\n  readonly name: string\n  /** The resolved system prompt (canonical `prompt.systemPrompt`, the sandbox `prompt.system`\n   *  convention, or a bare-string prompt — whichever the author used). */\n  readonly systemPrompt: string\n  readonly systemPromptChars: number\n  readonly systemPromptLines: number\n  readonly sentenceCount: number\n  readonly hasDescription: boolean\n  readonly hasTools: boolean\n  readonly hasSkills: boolean\n  readonly hasMcp: boolean\n  readonly hasSubagents: boolean\n  /** 0..1 — fraction of richness signals present (prompt-depth + the four levers). */\n  readonly richness: number\n  /** True when the supervisor authored a stub instead of a real profile. */\n  readonly thin: boolean\n  /** The specific reasons it is thin (empty when rich) — used in the finding's action. */\n  readonly reasons: string[]\n}\n\n/** Read the system prompt from any authored shape: canonical `prompt.systemPrompt`, the sandbox\n *  `prompt.system` convention, or a bare-string `prompt`. */\nfunction resolveSystemPrompt(profile: AgentProfile): string {\n  const pr = (profile as { prompt?: unknown }).prompt\n  if (typeof pr === 'string') return pr\n  if (pr && typeof pr === 'object') {\n    const o = pr as { systemPrompt?: unknown; system?: unknown }\n    if (typeof o.systemPrompt === 'string') return o.systemPrompt\n    if (typeof o.system === 'string') return o.system\n  }\n  return ''\n}\n\n/** OBSERVE one authored `AgentProfile` and score its richness (no judge verdict is read). The task\n *  context (`needsMcp`) lets a domain say \"this work needs a data/tool MCP\" so a missing MCP counts. */\nexport function assessAuthoredProfile(\n  profile: AgentProfile,\n  opts?: { needsMcp?: boolean; thresholds?: Partial<ProfileRichnessThresholds> },\n): ProfileRichness {\n  const th = { ...defaultProfileRichnessThresholds, ...(opts?.thresholds ?? {}) }\n  const systemPrompt = resolveSystemPrompt(profile)\n  const trimmed = systemPrompt.trim()\n  const systemPromptChars = trimmed.length\n  const systemPromptLines = trimmed\n    ? trimmed.split('\\n').filter((l) => l.trim().length > 0).length\n    : 0\n  const sentenceCount = trimmed\n    ? (trimmed.match(/[.!?](\\s|$)/g) ?? []).length || (trimmed ? 1 : 0)\n    : 0\n  const hasDescription =\n    typeof profile.description === 'string' && profile.description.trim().length > 0\n  const tools = (profile as { tools?: Record<string, unknown> }).tools\n  const hasTools = !!tools && Object.keys(tools).length > 0\n  const skills = (profile.resources as { skills?: unknown[] } | undefined)?.skills\n  const hasSkills = Array.isArray(skills) && skills.length > 0\n  const mcp = (profile as { mcp?: Record<string, unknown> }).mcp\n  const hasMcp = !!mcp && Object.keys(mcp).length > 0\n  const subagents = (profile as { subagents?: Record<string, unknown> }).subagents\n  const hasSubagents = !!subagents && Object.keys(subagents).length > 0\n\n  const reasons: string[] = []\n  const promptThin =\n    systemPromptChars < th.minSystemPromptChars || systemPromptLines < th.minSystemPromptLines\n  if (promptThin)\n    reasons.push(\n      `system prompt is thin (${systemPromptChars} chars, ${systemPromptLines} lines; need ≥${th.minSystemPromptChars} chars and ≥${th.minSystemPromptLines} lines)`,\n    )\n  if (!hasTools)\n    reasons.push('no tools granted (a worker can only act through the tools you grant it)')\n  if (!hasSkills) reasons.push('no skills attached (no reusable how-to notes injected)')\n  if (opts?.needsMcp && !hasMcp) reasons.push('no MCP server, but the task needs data/tool access')\n\n  // Richness = fraction of signals present. Prompt-depth is one signal; the four levers are the rest.\n  const signals = [!promptThin, hasTools, hasSkills, hasDescription, opts?.needsMcp ? hasMcp : true]\n  const richness = signals.filter(Boolean).length / signals.length\n  // THIN ⟺ the prompt is a stub OR the worker has no levers at all (no tools AND no skills AND no mcp).\n  const thin = promptThin || (!hasTools && !hasSkills && !hasMcp)\n\n  return {\n    name: profile.name ?? 'worker',\n    systemPrompt,\n    systemPromptChars,\n    systemPromptLines,\n    sentenceCount,\n    hasDescription,\n    hasTools,\n    hasSkills,\n    hasMcp,\n    hasSubagents,\n    richness,\n    thin,\n    reasons,\n  }\n}\n\n/** Turn a {@link ProfileRichness} verdict into a bus-routable `AnalystFinding` (area `profile-quality`).\n *  Severity scales with thinness; the recommended action names the MISSING lever so the supervisor can\n *  re-author. `subject` = the worker name so per-worker findings diff cleanly across re-authors. */\nexport function profileRichnessFinding(\n  richness: ProfileRichness,\n  opts?: { analystId?: string; runId?: string },\n): AnalystFinding {\n  const analyst_id = opts?.analystId ?? 'profile-richness'\n  const subject = richness.name\n  const claim = richness.thin\n    ? `Worker \"${richness.name}\" was authored as a THIN profile: ${richness.reasons.join('; ')}.`\n    : `Worker \"${richness.name}\" was authored as a rich profile (richness ${(richness.richness * 100).toFixed(0)}%).`\n  const severity: AnalystFinding['severity'] = richness.thin\n    ? richness.richness < 0.25\n      ? 'high'\n      : 'medium'\n    : 'info'\n  return makeFinding({\n    analyst_id,\n    severity,\n    area: 'profile-quality',\n    claim,\n    subject,\n    confidence: 0.9,\n    evidence_refs: [\n      {\n        kind: 'metric',\n        uri: `profile:${subject}`,\n        excerpt: `chars=${richness.systemPromptChars} lines=${richness.systemPromptLines} tools=${richness.hasTools} skills=${richness.hasSkills} mcp=${richness.hasMcp} richness=${richness.richness.toFixed(2)}`,\n      },\n    ],\n    ...(richness.thin\n      ? { recommended_action: `Re-author \"${richness.name}\" with: ${richness.reasons.join('; ')}.` }\n      : {}),\n    id_basis: computeFindingId({\n      analyst_id,\n      area: 'profile-quality',\n      subject,\n      claim: `richness:${richness.thin ? 'thin' : 'rich'}`,\n    }),\n  })\n}\n","/**\n *\n * `delegate` — the one generic delegation verb. You hand it an INTENT (what you want done) and it\n * hands that intent to a default AUTHORING supervisor: a router-brained supervisor whose standing\n * instruction is `supervisorInstructions()` (the authoring-agent-profiles skill). The supervisor\n * DECOMPOSES the intent and AUTHORS the worker profile it needs per sub-task — there is NO hardcoded\n * coder/researcher profile here. That is the whole point: `delegate('fix the failing test', …)` and\n * `delegate('research X and cite sources', …)` route through the SAME front door; the supervisor\n * writes a code-shaped or research-shaped worker on its own.\n *\n * It is a thin wrapper over `supervise()` — the one front door — so the conserved-budget pool, the\n * completion oracle (`deliverable`), the coordination toolbox, and equal-compute accounting all come\n * for free; nothing is hand-rolled. The result is `supervise()`'s `SupervisedResult` returned\n * UNCHANGED, so its `spentTotal` (`{ iterations, tokens, usd, ms }`) rides straight back to the\n * caller on BOTH paths — a `winner` carries the delivered worker's spend, a `no-winner` carries the\n * spend incurred before it failed. That cost channel means a `delegate()` caller always learns what\n * the delegation actually spent.\n *\n * @experimental\n */\n\nimport { ConfigError } from '../../errors'\nimport type { RouterTransportConfig } from '../router-client'\nimport type { DeliverableSpec } from './completion-gate'\nimport type { ExecutorConfig } from './runtime'\nimport { supervise } from './supervise'\nimport type { SupervisorProfile } from './supervisor-agent'\nimport type { Budget, SupervisedResult } from './types'\n\n/** The conserved pool a `delegate()` call applies when the caller does not pass its own `budget`.\n *  A modest token ceiling + a small iteration ceiling — generous enough for a few-worker decompose,\n *  bounded enough that an unsupervised intent cannot run away. Callers override via `opts.budget`. */\nexport const defaultDelegateBudget: Budget = { maxIterations: 50, maxTokens: 200_000 }\n\n/** Inputs to {@link delegate}. The intent is the first positional arg; everything here is optional\n *  with explicit execution identity, so the common call names one exact supervisor profile. */\nexport interface DelegateOptions<Out = unknown> {\n  /** The completion oracle (settled ⟺ delivered) the authored workers settle against. Strongly\n   *  recommended — without it the supervisor trusts a worker's self-report. For a code intent,\n   *  `patchDelivered()` is the canonical example; for a free-form answer, a content check. */\n  readonly deliverable?: DeliverableSpec<Out>\n  /** WHERE the authored workers run — the worker-execution backend (`router-tools` / `sandbox` /\n   *  `cli-worktree` / …). The supervisor authors the worker PROFILE; this is the substrate it runs\n   *  on. Provide this OR `makeWorkerAgent`-style wiring through `supervise()` is unavailable. */\n  readonly backend?: ExecutorConfig\n  /** The conserved compute pool for the whole delegation. Defaults to {@link defaultDelegateBudget}. */\n  readonly budget?: Budget\n  /** Exact executable authoring supervisor. Model, prompt, harness, and provider live here. */\n  readonly supervisorProfile: SupervisorProfile\n  /** Router endpoint/auth for a `cli-base` supervisor; contains no behavioral settings. */\n  readonly router: RouterTransportConfig\n  /** Restrict the run to this subset of models (forwarded to `supervise()`). */\n  readonly allowedModels?: readonly string[]\n  readonly runId?: string\n}\n\n/**\n * Delegate an INTENT to a default authoring supervisor and return its `SupervisedResult` unchanged.\n *\n * The supervisor authors + spawns whatever worker the intent needs over the conserved-budget pool;\n * `result.spentTotal` reports what the whole delegation actually cost. A `winner` result carries the\n * authored worker's delivered output; a `no-winner` result names why (never a fabricated success).\n */\nexport async function delegate<Out = unknown>(\n  intent: string,\n  opts: DelegateOptions<Out>,\n): Promise<SupervisedResult<Out>> {\n  if (typeof intent !== 'string' || intent.trim().length === 0) {\n    throw new ConfigError('delegate: `intent` must be a non-empty string')\n  }\n  return supervise(opts.supervisorProfile, intent, {\n    budget: opts.budget ?? defaultDelegateBudget,\n    ...(opts.backend ? { backend: opts.backend } : {}),\n    ...(opts.deliverable ? { deliverable: opts.deliverable as DeliverableSpec<unknown> } : {}),\n    router: opts.router,\n    ...(opts.allowedModels ? { allowedModels: opts.allowedModels } : {}),\n    ...(opts.runId ? { runId: opts.runId } : {}),\n  }) as Promise<SupervisedResult<Out>>\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;AA+BA,SAAgB,kBAAkB,KAAsC;CACtE,MAAM,SAAS,mBAAmB,UAAU,GAAG;CAC/C,IAAI,CAAC,OAAO,SAAS,OAAO;CAC5B,MAAM,eAAe,OAAO,KAAK,QAAQ;CACzC,IAAI,OAAO,iBAAiB,YAAY,aAAa,KAAK,CAAC,CAAC,WAAW,GAAG,OAAO;CACjF,OAAO;EACL,GAAG,OAAO;EACV,MACE,OAAO,OAAO,KAAK,SAAS,YAAY,OAAO,KAAK,KAAK,SAAS,IAC9D,OAAO,KAAK,OACZ;EACN,QAAQ;GAAE,GAAG,OAAO,KAAK;GAAQ;EAAa;CAChD;AACF;;;AAIA,SAAgB,uBAAuB,MAAkC;CACvE,OAAO;EACL;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,GAAI,MAAM,OAAO,CAAC,IAAI,aAAa,KAAK,MAAM,IAAI,CAAC;CACrD,CAAC,CAAC,KAAK,IAAI;AACb;;AAqBA,MAAa,mCAA8D;CACzE,sBAAsB;CACtB,sBAAsB;AACxB;;;AA0BA,SAAS,oBAAoB,SAA+B;CAC1D,MAAM,KAAM,QAAiC;CAC7C,IAAI,OAAO,OAAO,UAAU,OAAO;CACnC,IAAI,MAAM,OAAO,OAAO,UAAU;EAChC,MAAM,IAAI;EACV,IAAI,OAAO,EAAE,iBAAiB,UAAU,OAAO,EAAE;EACjD,IAAI,OAAO,EAAE,WAAW,UAAU,OAAO,EAAE;CAC7C;CACA,OAAO;AACT;;;AAIA,SAAgB,sBACd,SACA,MACiB;CACjB,MAAM,KAAK;EAAE,GAAG;EAAkC,GAAI,MAAM,cAAc,CAAC;CAAG;CAC9E,MAAM,eAAe,oBAAoB,OAAO;CAChD,MAAM,UAAU,aAAa,KAAK;CAClC,MAAM,oBAAoB,QAAQ;CAClC,MAAM,oBAAoB,UACtB,QAAQ,MAAM,IAAI,CAAC,CAAC,QAAQ,MAAM,EAAE,KAAK,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,SACvD;CACJ,MAAM,gBAAgB,WACjB,QAAQ,MAAM,cAAc,KAAK,CAAC,EAAA,CAAG,WAAW,UAAU,IAAI,KAC/D;CACJ,MAAM,iBACJ,OAAO,QAAQ,gBAAgB,YAAY,QAAQ,YAAY,KAAK,CAAC,CAAC,SAAS;CACjF,MAAM,QAAS,QAAgD;CAC/D,MAAM,WAAW,CAAC,CAAC,SAAS,OAAO,KAAK,KAAK,CAAC,CAAC,SAAS;CACxD,MAAM,SAAU,QAAQ,WAAkD;CAC1E,MAAM,YAAY,MAAM,QAAQ,MAAM,KAAK,OAAO,SAAS;CAC3D,MAAM,MAAO,QAA8C;CAC3D,MAAM,SAAS,CAAC,CAAC,OAAO,OAAO,KAAK,GAAG,CAAC,CAAC,SAAS;CAClD,MAAM,YAAa,QAAoD;CACvE,MAAM,eAAe,CAAC,CAAC,aAAa,OAAO,KAAK,SAAS,CAAC,CAAC,SAAS;CAEpE,MAAM,UAAoB,CAAC;CAC3B,MAAM,aACJ,oBAAoB,GAAG,wBAAwB,oBAAoB,GAAG;CACxE,IAAI,YACF,QAAQ,KACN,0BAA0B,kBAAkB,UAAU,kBAAkB,gBAAgB,GAAG,qBAAqB,cAAc,GAAG,qBAAqB,QACxJ;CACF,IAAI,CAAC,UACH,QAAQ,KAAK,yEAAyE;CACxF,IAAI,CAAC,WAAW,QAAQ,KAAK,wDAAwD;CACrF,IAAI,MAAM,YAAY,CAAC,QAAQ,QAAQ,KAAK,oDAAoD;CAGhG,MAAM,UAAU;EAAC,CAAC;EAAY;EAAU;EAAW;EAAgB,MAAM,WAAW,SAAS;CAAI;CACjG,MAAM,WAAW,QAAQ,OAAO,OAAO,CAAC,CAAC,SAAS,QAAQ;CAE1D,MAAM,OAAO,cAAe,CAAC,YAAY,CAAC,aAAa,CAAC;CAExD,OAAO;EACL,MAAM,QAAQ,QAAQ;EACtB;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;CACF;AACF;;;;AAKA,SAAgB,uBACd,UACA,MACgB;CAChB,MAAM,aAAa,MAAM,aAAa;CACtC,MAAM,UAAU,SAAS;CACzB,MAAM,QAAQ,SAAS,OACnB,WAAW,SAAS,KAAK,oCAAoC,SAAS,QAAQ,KAAK,IAAI,EAAE,KACzF,WAAW,SAAS,KAAK,8CAA8C,SAAS,WAAW,IAAA,CAAK,QAAQ,CAAC,EAAE;CAM/G,OAAO,YAAY;EACjB;EACA,UAP2C,SAAS,OAClD,SAAS,WAAW,MAClB,SACA,WACF;EAIF,MAAM;EACN;EACA;EACA,YAAY;EACZ,eAAe,CACb;GACE,MAAM;GACN,KAAK,WAAW;GAChB,SAAS,SAAS,SAAS,kBAAkB,SAAS,SAAS,kBAAkB,SAAS,SAAS,SAAS,UAAU,SAAS,UAAU,OAAO,SAAS,OAAO,YAAY,SAAS,SAAS,QAAQ,CAAC;EACzM,CACF;EACA,GAAI,SAAS,OACT,EAAE,oBAAoB,cAAc,SAAS,KAAK,UAAU,SAAS,QAAQ,KAAK,IAAI,EAAE,GAAG,IAC3F,CAAC;EACL,UAAU,iBAAiB;GACzB;GACA,MAAM;GACN;GACA,OAAO,YAAY,SAAS,OAAO,SAAS;EAC9C,CAAC;CACH,CAAC;AACH;;;;;;;;;;;;;;;;;;;;;;;;;;AC5MA,MAAa,wBAAgC;CAAE,eAAe;CAAI,WAAW;AAAQ;;;;;;;;AA+BrF,eAAsB,SACpB,QACA,MACgC;CAChC,IAAI,OAAO,WAAW,YAAY,OAAO,KAAK,CAAC,CAAC,WAAW,GACzD,MAAM,IAAI,YAAY,+CAA+C;CAEvE,OAAO,UAAU,KAAK,mBAAmB,QAAQ;EAC/C,QAAQ,KAAK,UAAU;EACvB,GAAI,KAAK,UAAU,EAAE,SAAS,KAAK,QAAQ,IAAI,CAAC;EAChD,GAAI,KAAK,cAAc,EAAE,aAAa,KAAK,YAAwC,IAAI,CAAC;EACxF,QAAQ,KAAK;EACb,GAAI,KAAK,gBAAgB,EAAE,eAAe,KAAK,cAAc,IAAI,CAAC;EAClE,GAAI,KAAK,QAAQ,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;CAC5C,CAAC;AACH"}

package/dist/durable.d.ts

@@ -1,6 +1,6 @@
-import { Uo as supervise, ps as SupervisorProfile, zo as SuperviseOptions } from "./index-CYOJsxSg.js";
+import { Uo as supervise, ps as SupervisorProfile, zo as SuperviseOptions } from "./index-D8LWK3Ha.js";
 import { l as RuntimeHooks, o as RuntimeHookEvent, r as RuntimeDecisionPoint } from "./runtime-hooks-Bj6wJHlH.js";
-import { E as NodeId, Y as SpendGap, j as ProviderModelExecutionEvidence, k as ProfileMaterializationReceipt, lt as WorkerTraceEvidence, o as ExecutionBindingReceipt, q as Spend } from "./types-DFLZMaeh.js";
+import { Er as NodeId, Yr as SpendGap, Zr as SupervisedResult, jr as ProviderModelExecutionEvidence, kr as ProfileMaterializationReceipt, li as WorkerTraceEvidence, or as ExecutionBindingReceipt, qr as Spend } from "./stream-agent-turn-CLOQr497.js";
 //#region src/durable/chat-engine.d.ts
 /**
  * `handleChatTurn` is a framework-neutral chat-turn HTTP orchestrator.
@@ -282,8 +282,22 @@ interface PursuitRunTotals {
    */
   readonly exclusiveByNode: Readonly<Record<string, Spend>>;
 }
+/**
+ * One attempt at one concrete Runtime run: the stretch of `agent.run` lifecycle from a `before`
+ * to the `after` or `error` that settles it. A run whose first attempt threw and whose corrected
+ * attempt settled is two rows, so `error` names only the attempt that failed.
+ *
+ * A `before` observed while an attempt is still open does not open another: it is a process
+ * that resumed the run after the previous process died without a terminal record, and the row
+ * counts it in `resumeCount`. The alternative, a row per process start, would leave the killed
+ * process's row `running` for the rest of history.
+ */
 interface PursuitRunProjection {
   readonly runId: string;
+  /** Zero-based position of this attempt among the run's attempts, in journal order. */
+  readonly attemptIndex: number;
+  /** How many times a process resumed this attempt after a predecessor died mid-run. */
+  readonly resumeCount: number;
   readonly status: PursuitStatus;
   readonly settledAt?: number;
   readonly error?: string;
@@ -380,8 +394,9 @@ interface PursuitProjection {
  *
  * Topology comes only from Runtime's canonical `agent.spawn` facts. Terminal node
  * state comes only from `agent.child`; concrete run state comes only from the root
- * `agent.run` lifecycle emitted by `supervisePursuit`. Node identity is scoped to the
- * concrete Runtime run so independent trees may both contain `root:s0` without aliasing.
+ * `agent.run` lifecycle emitted by `supervisePursuit`, one row per attempt. Node identity is
+ * scoped to the concrete Runtime run so independent trees may both contain `root:s0` without
+ * aliasing, and a node belongs to the attempt that spawned it.
  *
  * Usage, cost and timing are reported at the class the runtime measured them at. A missing
  * class stays ABSENT and the run names the node in `spendGaps`; nothing here converts an
@@ -389,6 +404,83 @@ interface PursuitProjection {
  */
 declare function projectPursuit(records: readonly ObserverRecord[]): PursuitProjection;
 //#endregion
+//#region src/durable/run-lock.d.ts
+/** The lock file `supervisePursuit` holds inside a run directory for the life of one call. */
+declare const RUN_DIRECTORY_LOCK_FILE = "supervise.lock";
+/** What the lock file records about its holder. */
+interface RunDirectoryLockHolder {
+  readonly pid: number;
+  /** ISO instant the holder took the lock. */
+  readonly startedAt: string;
+  readonly runId: string;
+  /**
+   * The OS's start token for `pid` when the host reports one (`readProcessStart`). A holder
+   * without one is judged by pid liveness alone, which cannot detect a reused pid.
+   */
+  readonly processStart?: string;
+}
+/** A held lock. `release()` removes the file; it is safe to call more than once. */
+interface RunDirectoryLock extends RunDirectoryLockHolder {
+  readonly path: string;
+  release(): Promise<void>;
+}
+/** The directory is held by a live process. `holder` is what that process recorded. */
+declare class RunDirectoryLockedError extends Error {
+  readonly path: string;
+  readonly holder: RunDirectoryLockHolder;
+  constructor(path: string, holder: RunDirectoryLockHolder);
+}
+/**
+ * Take `runDir/supervise.lock`, or refuse.
+ *
+ * The file is published with its full content or not at all, so a contender never reads a
+ * half-written holder. A lock whose holder is gone (its pid no longer exists, or the pid now
+ * belongs to a process with a different start token) is stale and is removed under the mutation guard;
+ * a pid this process may not signal (`EPERM`) is alive and refuses. An empty file names no
+ * holder and is reclaimed. A file with unreadable content is left in place and refused:
+ * reclaiming it could evict a live holder written by something other than this module.
+ */
+declare function acquireRunDirectoryLock(runDir: string, runId: string, now?: () => number): Promise<RunDirectoryLock>;
+/** Read the holder a lock file names, or `undefined` when no lock file names one. */
+declare function readRunDirectoryLock(runDir: string): Promise<RunDirectoryLockHolder | undefined>;
+//#endregion
+//#region src/durable/settle-record.d.ts
+/** The settle record: the returned `SupervisedResult` as canonical JSON, written once. */
+declare const SETTLE_RECORD_FILE = "result.json";
+/** The failure record: the most recent throw, replaced by a later throw. */
+declare const FAILURE_RECORD_FILE = "failure.json";
+/** What `failure.json` records about the most recent throw. */
+interface DurableFailureRecord {
+  readonly runId: string;
+  readonly pursuitId: string;
+  /** ISO instant the throw was recorded. */
+  readonly at: string;
+  readonly error: {
+    readonly name: string;
+    readonly message: string;
+  };
+}
+/** The directory already holds a settle record, so the run it records must not be re-entered. */
+declare class SettledRunDirectoryError extends Error {
+  readonly path: string;
+  /** The root of the recorded run, which is its `runId`. */
+  readonly recordedRunId: string;
+  constructor(path: string, recordedRunId: string, requestedRunId: string);
+}
+/**
+ * The exact bytes `result.json` holds for a result: its JSON value serialized as RFC 8785
+ * canonical JSON. Throws `UnrecordableSettleValueError` before any byte is written when the
+ * result carries a value JSON would misstate.
+ */
+declare function settleRecordJson(result: unknown): string;
+/**
+ * Read the settle record a run directory holds, or `undefined` when it holds none. A file that
+ * is present but is not a settle record is corruption and fails loud.
+ */
+declare function readSettleRecord(runDir: string): Promise<SupervisedResult<unknown> | undefined>;
+/** Read the most recent failure record, or `undefined` when the directory holds none. */
+declare function readFailureRecord(runDir: string): Promise<DurableFailureRecord | undefined>;
+//#endregion
 //#region src/durable/supervise-pursuit.d.ts
 interface SupervisePursuitOptions extends SuperviseOptions {
   /** Stable objective identity spanning concrete Runtime runs. */
@@ -404,12 +496,16 @@ interface SupervisedPursuitResult<Result> {
   readonly result: Result;
   readonly pursuit: PursuitProjection;
   readonly observerPath: string;
+  /** `runDir/result.json`: `result` as canonical JSON, written once at settle. */
+  readonly settlePath: string;
 }
 /** A failed Runtime execution whose complete third-person projection was retained. */
 declare class SupervisePursuitError extends Error {
   readonly pursuit: PursuitProjection;
   readonly observerPath: string;
-  constructor(cause: unknown, pursuit: PursuitProjection, observerPath: string);
+  /** `runDir/failure.json`: the record of this throw. */
+  readonly failurePath: string;
+  constructor(cause: unknown, pursuit: PursuitProjection, observerPath: string, failurePath: string);
 }
 /**
  * One-call durable pursuit execution over the canonical `supervise()` kernel.
@@ -422,6 +518,14 @@ declare class SupervisePursuitError extends Error {
  * Every concrete execution writes only inside its own `runDir`. Cross-run pursuit
  * aggregation is therefore lock-free at the observer layer: reuse `pursuitId` across
  * run directories and let Intelligence join the independently verified projections.
+ *
+ * The directory's terminal state is recorded beside `observer.jsonl`: `result.json` holds the
+ * returned result once the run settles and `failure.json` the most recent throw. A directory
+ * whose `result.json` exists refuses re-entry before any compute; a failure record alone does
+ * not, because a caller can correct its input and drive the same run again. For the life of the
+ * call the directory is held by `supervise.lock`, so a second process on the same directory
+ * refuses and names the holder instead of sharing one journal.
+ * An abandoned `supervise.lock.guard` requires removal after confirming no lock mutation is active.
  */
 declare function supervisePursuit(profile: SupervisorProfile, task: unknown, opts: SupervisePursuitOptions): Promise<SupervisedPursuitResult<Awaited<ReturnType<typeof supervise>>>>;
 //#endregion
@@ -457,5 +561,5 @@ interface DurableSupervisionDiscovery {
  */
 declare function discoverDurableSupervisionRun(runDir: string): Promise<DurableSupervisionDiscovery>;
 //#endregion
-export { type ChatStreamEvent, type ChatTurnHooks, type ChatTurnIdentity, type ChatTurnProducer, type ChatTurnResult, type DurableCoordinationStreamIdentity, type DurableSupervisionDiscovery, FileObserverJournal, type ObserverJournal, type ObserverRecord, type ObserverRecordKind, type PursuitCostProvenance, type PursuitNodeCost, type PursuitNodePlacement, type PursuitNodePlatform, type PursuitNodeProjection, type PursuitNodeTiming, type PursuitNodeUsage, type PursuitProjection, type PursuitRunProjection, type PursuitRunTotals, type PursuitStatus, type RunChatTurnInput, SupervisePursuitError, type SupervisePursuitOptions, type SupervisedPursuitResult, createFileObserverHooks, deriveExecutionId, discoverDurableSupervisionRun, handleChatTurn, observerRecordDigest, projectPursuit, supervisePursuit, verifyObserverRecords };
+export { type ChatStreamEvent, type ChatTurnHooks, type ChatTurnIdentity, type ChatTurnProducer, type ChatTurnResult, type DurableCoordinationStreamIdentity, type DurableFailureRecord, type DurableSupervisionDiscovery, FAILURE_RECORD_FILE, FileObserverJournal, type ObserverJournal, type ObserverRecord, type ObserverRecordKind, type PursuitCostProvenance, type PursuitNodeCost, type PursuitNodePlacement, type PursuitNodePlatform, type PursuitNodeProjection, type PursuitNodeTiming, type PursuitNodeUsage, type PursuitProjection, type PursuitRunProjection, type PursuitRunTotals, type PursuitStatus, RUN_DIRECTORY_LOCK_FILE, type RunChatTurnInput, type RunDirectoryLock, type RunDirectoryLockHolder, RunDirectoryLockedError, SETTLE_RECORD_FILE, SettledRunDirectoryError, SupervisePursuitError, type SupervisePursuitOptions, type SupervisedPursuitResult, acquireRunDirectoryLock, createFileObserverHooks, deriveExecutionId, discoverDurableSupervisionRun, handleChatTurn, observerRecordDigest, projectPursuit, readFailureRecord, readRunDirectoryLock, readSettleRecord, settleRecordJson, supervisePursuit, verifyObserverRecords };
 //# sourceMappingURL=durable.d.ts.map