@tangle-network/agent-runtime 0.225.6 → 0.228.0

package/dist/{delegate-Dpz_acWb.js → delegate-OUy-TxtQ.js}

@@ -1,5 +1,5 @@
 import { i as ConfigError } from "./errors-DodWX-cb.js";
-import { n as supervise } from "./supervise-CkqrhpOD.js";
+import { n as supervise } from "./supervise-UgmGtrlX.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-Dpz_acWb.js.map
+//# sourceMappingURL=delegate-OUy-TxtQ.js.map

package/dist/{delegate-Dpz_acWb.js.map → delegate-OUy-TxtQ.js.map}

@@ -1 +1 @@
-{"version":3,"file":"delegate-Dpz_acWb.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-OUy-TxtQ.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 +1,39 @@
-import { Xa as SuperviseOptions, eo as supervise, wo as SupervisorProfile } from "./index-ClXIBpKe.js";
+import { Xa as SuperviseOptions, eo as supervise, wo as SupervisorProfile } from "./index-DAIij88v.js";
 import { l as RuntimeHooks, o as RuntimeHookEvent, r as RuntimeDecisionPoint } from "./runtime-hooks-Bj6wJHlH.js";
-import { Ir as ProviderModelExecutionEvidence, Pr as ProfileMaterializationReceipt, cr as BudgetViolation, dr as ExecutionBindingReceipt, ei as Spend, gi as WorkerTraceEvidence, ii as SupervisedResult, jr as NodeId, ni as SpendGap } from "./stream-agent-turn-wlsHQiUv.js";
+import { Ir as ProviderModelExecutionEvidence, Kr as RootStreamReceipt, Pr as ProfileMaterializationReceipt, _i as WorkerTraceEvidence, ai as SupervisedResult, cr as BudgetViolation, dr as ExecutionBindingReceipt, jr as NodeId, ri as SpendGap, ti as Spend, xr as ExecutorProgressEvent } from "./stream-agent-turn-Dy-QleSb.js";
+//#region src/runtime/supervise/root-stream.d.ts
+/** The root stream: one JSONL line per progress event the root's executor observed. */
+declare const ROOT_STREAM_FILE = "root-stream.jsonl";
+/** One line of `root-stream.jsonl`. */
+type RootStreamRecord = {
+  /** 1-based position in the file, continuing across drive attempts and across processes. */
+  readonly seq: number;
+  /** ISO instant the line was appended, from the run's own clock. */
+  readonly at: string;
+  /** The 1-based drive attempt of the root that produced it: a driver retry or re-prompt
+   *  re-enters the harness and continues the same file with the next attempt number. */
+  readonly attempt: number;
+} & ({
+  readonly event: ExecutorProgressEvent;
+} | {
+  /** The event could not be written as JSON; its kind and the reason stand in so the gap
+   *  is recorded in the file itself instead of silently narrowing the stream. */
+  readonly dropped: {
+    readonly kind: ExecutorProgressEvent['kind'];
+    readonly reason: string;
+  };
+});
+/**
+ * The receipt for the root stream a run directory holds, recomputed from the file's bytes, or
+ * `undefined` when the directory holds none. This is what a run that never settled — a root that
+ * died mid-turn — gets on its failure record, and it equals what `close()` returned for a run
+ * that did.
+ */
+declare function readRootStreamReceipt(runDir: string): Promise<RootStreamReceipt | undefined>;
+/** Every committed line of the root stream, in order, or `undefined` when there is no file. A
+ *  torn final line from a process that died mid-write is not a record and is left out. */
+declare function readRootStream(runDir: string): Promise<RootStreamRecord[] | undefined>;
+//#endregion
 //#region src/durable/chat-engine.d.ts
 /**
  * `handleChatTurn` is a framework-neutral chat-turn HTTP orchestrator.
@@ -485,6 +518,9 @@ interface DurableFailureRecord {
     readonly name: string;
     readonly message: string;
   };
+  /** The root's retained provider stream at the time of the throw, when the directory holds one:
+   *  a root that died mid-turn keeps what it had streamed, and this names it. */
+  readonly rootStream?: RootStreamReceipt;
 }
 /** The directory already holds a settle record, so the run it records must not be re-entered. */
 declare class SettledRunDirectoryError extends Error {
@@ -595,5 +631,5 @@ interface DurableSupervisionDiscovery {
  */
 declare function discoverDurableSupervisionRun(runDir: string): Promise<DurableSupervisionDiscovery>;
 //#endregion
-export { type ChatStreamEvent, type ChatTurnHooks, type ChatTurnIdentity, type ChatTurnProducer, type ChatTurnResult, type DurableCoordinationStreamIdentity, type 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 RunDirectoryHolderLiveness, type RunDirectoryLock, type RunDirectoryLockHolder, RunDirectoryLockedError, SETTLE_RECORD_FILE, SettledRunDirectoryError, SupervisePursuitError, type SupervisePursuitOptions, type SupervisedPursuitResult, acquireRunDirectoryLock, createFileObserverHooks, deriveExecutionId, discoverDurableSupervisionRun, handleChatTurn, observerRecordDigest, projectPursuit, readFailureRecord, readRunDirectoryLock, readSettleRecord, runDirectoryHolderIsLive, settleRecordJson, 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, ROOT_STREAM_FILE, RUN_DIRECTORY_LOCK_FILE, type RootStreamReceipt, type RootStreamRecord, type RunChatTurnInput, type RunDirectoryHolderLiveness, type RunDirectoryLock, type RunDirectoryLockHolder, RunDirectoryLockedError, SETTLE_RECORD_FILE, SettledRunDirectoryError, SupervisePursuitError, type SupervisePursuitOptions, type SupervisedPursuitResult, acquireRunDirectoryLock, createFileObserverHooks, deriveExecutionId, discoverDurableSupervisionRun, handleChatTurn, observerRecordDigest, projectPursuit, readFailureRecord, readRootStream, readRootStreamReceipt, readRunDirectoryLock, readSettleRecord, runDirectoryHolderIsLive, settleRecordJson, supervisePursuit, verifyObserverRecords };
 //# sourceMappingURL=durable.d.ts.map

package/dist/durable.js

@@ -1,7 +1,7 @@
-import { Ii as zeroSpend, Ir as isNoEntError, Lr as parseCommittedJsonLines, Rr as prepareJsonlAppend, Si as addSpend, wi as cloneSpend, zr as writeAllBytes } from "./supervisor-DCQxXRR8.js";
+import { Ii as zeroSpend, Ir as isNoEntError, Lr as parseCommittedJsonLines, Rr as prepareJsonlAppend, Si as addSpend, wi as cloneSpend, zr as writeAllBytes } from "./supervisor-DBnQuTyJ.js";
 import { a as withPursuitContext, t as composeRuntimeHooks } from "./runtime-hooks-tXpAarhW.js";
 import { i as writeAtomicDurableFile, r as publishExclusiveDurableFile } from "./durable-file-D24y9zg7.js";
-import { n as supervise } from "./supervise-CkqrhpOD.js";
+import { g as readRootStreamReceipt, h as readRootStream, m as ROOT_STREAM_FILE, n as supervise } from "./supervise-UgmGtrlX.js";
 import { canonicalCandidateJson } from "@tangle-network/agent-interface";
 import { createHash } from "node:crypto";
 import { execFile } from "node:child_process";
@@ -1132,7 +1132,7 @@ async function readFailureRecord(runDir) {
 		throw new Error(`supervisePursuit: ${path} is not valid JSON`, { cause });
 	}
 	const record = parsed;
-	if (typeof record !== "object" || record === null || typeof record.runId !== "string" || typeof record.pursuitId !== "string" || typeof record.at !== "string" || typeof record.error?.name !== "string" || typeof record.error.message !== "string") throw new Error(`supervisePursuit: ${path} is not a failure record`);
+	if (typeof record !== "object" || record === null || typeof record.runId !== "string" || typeof record.pursuitId !== "string" || typeof record.at !== "string" || typeof record.error?.name !== "string" || typeof record.error.message !== "string" || record.rootStream !== void 0 && (typeof record.rootStream.ref !== "string" || typeof record.rootStream.events !== "number")) throw new Error(`supervisePursuit: ${path} is not a failure record`);
 	return parsed;
 }
 //#endregion
@@ -1210,6 +1210,7 @@ async function supervisePursuit(profile, task, opts) {
 			}
 			if (observerError !== void 0 || pursuit === void 0) throw new Error("supervisePursuit: Runtime failed and durable observer completeness could not be proven", { cause: new AggregateError(observerError === void 0 ? [error] : [error, observerError]) });
 			try {
+				const rootStream = await readRootStreamReceipt(runDir);
 				await writeFailureRecord(runDir, {
 					runId,
 					pursuitId,
@@ -1217,7 +1218,8 @@ async function supervisePursuit(profile, task, opts) {
 					error: {
 						name: errorName(error),
 						message: errorMessage(error)
-					}
+					},
+					...rootStream === void 0 ? {} : { rootStream }
 				});
 			} catch (recordError) {
 				throw new Error(`supervisePursuit: Runtime failed and the failure record ${failurePath} could not be written`, { cause: new AggregateError([error, recordError]) });
@@ -1338,6 +1340,6 @@ function compareText(left, right) {
 	return left < right ? -1 : left > right ? 1 : 0;
 }
 //#endregion
-export { FAILURE_RECORD_FILE, FileObserverJournal, RUN_DIRECTORY_LOCK_FILE, RunDirectoryLockedError, SETTLE_RECORD_FILE, SettledRunDirectoryError, SupervisePursuitError, acquireRunDirectoryLock, createFileObserverHooks, deriveExecutionId, discoverDurableSupervisionRun, handleChatTurn, observerRecordDigest, projectPursuit, readFailureRecord, readRunDirectoryLock, readSettleRecord, runDirectoryHolderIsLive, settleRecordJson, supervisePursuit, verifyObserverRecords };
+export { FAILURE_RECORD_FILE, FileObserverJournal, ROOT_STREAM_FILE, RUN_DIRECTORY_LOCK_FILE, RunDirectoryLockedError, SETTLE_RECORD_FILE, SettledRunDirectoryError, SupervisePursuitError, acquireRunDirectoryLock, createFileObserverHooks, deriveExecutionId, discoverDurableSupervisionRun, handleChatTurn, observerRecordDigest, projectPursuit, readFailureRecord, readRootStream, readRootStreamReceipt, readRunDirectoryLock, readSettleRecord, runDirectoryHolderIsLive, settleRecordJson, supervisePursuit, verifyObserverRecords };
 
 //# sourceMappingURL=durable.js.map