@buildaharness/aielia 0.3.7 → 0.3.9

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sources":["../src/plan-builder.ts","../src/scripted-llm-client.ts"],"sourcesContent":["import type { ILLMClient, TokenUsage } from '@buildaharness/runtime'\nimport type { DecomposedTaskSpec } from './decomposition-classifier.js'\nimport type { PlanTemplate } from './plan-templates/index.js'\nimport { parseModelJson } from './model-json.js'\n\nexport interface Plan {\n  templateName: string\n  successCriteria: string\n  tasks: DecomposedTaskSpec[]\n  /**\n   * Why this approach, not just what the success criteria are — optional for now since neither\n   * buildPlanFromTemplate's template-personalization call nor any other current producer of a\n   * Plan asks the LLM for one; plan-store.ts's createPlanRecord defaults it to '' when absent.\n   * A future from-scratch drafting path (plan mode's P3) is expected to populate it.\n   */\n  rationale?: string\n}\n\nconst PLAN_SCHEMA = {\n  type: 'object',\n  properties: {\n    tasks: {\n      type: 'array',\n      items: {\n        type: 'object',\n        properties: {\n          id: { type: 'string' },\n          description: { type: 'string' },\n          depends_on: { type: 'array', items: { type: 'string' } },\n        },\n        required: ['id', 'description', 'depends_on'],\n      },\n    },\n  },\n  required: ['tasks'],\n}\n\nfunction buildSystemPrompt(template: PlanTemplate): string {\n  const skeleton = template.tasks\n    .map((t) => `- id: ${t.id}; title: ${t.title}; depends_on: [${t.depends_on.join(', ')}]`)\n    .join('\\n')\n  return (\n    `You are adapting a \"${template.name}\" plan template to a specific user request. ` +\n    `Here is the template's task skeleton — keep the exact same ids and depends_on structure, ` +\n    `one output task per skeleton task, but personalize each description to the actual request:\\n${skeleton}\\n\\n` +\n    `Success criteria for this kind of plan: ${template.success_criteria}\\n\\n` +\n    'Phrase each `description` starting with the concrete subject or object it acts on (e.g. \"the login tests: ' +\n    'rerun after the config fix\" rather than \"rerun the login tests after the config fix\"), so later comparisons ' +\n    \"against this task's completion/failure beliefs share matching vocabulary. \" +\n    'Respond with JSON only, no prose: {\"tasks\":[{\"id\": string, \"description\": string, \"depends_on\": ' +\n    'string[]}]}. `id` and `depends_on` values must exactly match the skeleton above.'\n  )\n}\n\n// Only the shape the LLM actually returns (id/description/depends_on) — riskLevel isn't asked for\n// here at all (see buildPlanFromTemplate's doc comment for why) and gets attached afterward from\n// the template's own already-curated risk_level, so this is deliberately narrower than\n// DecomposedTaskSpec itself.\ninterface RawPlanTask {\n  id: string\n  description: string\n  depends_on: string[]\n}\n\nfunction isRawPlanTask(value: unknown): value is RawPlanTask {\n  if (typeof value !== 'object' || value === null) return false\n  const v = value as Record<string, unknown>\n  return (\n    typeof v.id === 'string' &&\n    typeof v.description === 'string' &&\n    Array.isArray(v.depends_on) &&\n    v.depends_on.every((d) => typeof d === 'string')\n  )\n}\n\n/**\n * Spends one real LLM call personalizing `template`'s task skeleton to `message` —\n * only called once classifyTurnIntent has already named a matchedPlanTemplate, so an\n * ordinary turn never pays for this. Same \"malformed/incomplete JSON is the expected\n * failure mode, not the edge case\" fallback classifyTurnIntent itself uses: any parse\n * failure or a response with fewer than 2 usable tasks returns null, meaning \"fall back to\n * the caller's own ad hoc decomposition for this turn\" rather than throwing.\n *\n * Each output task's `riskLevel` is attached from the matching skeleton task's own\n * `risk_level` (curated by whoever wrote the template's JSON) by id, not asked of the LLM —\n * the template already has this, more accurately than an LLM re-deriving it from a\n * personalized one-line description would, and it costs zero extra output tokens. Falls back\n * to 'LOW' only for the pathological case of a task id the LLM invented that doesn't match any\n * skeleton task (the id/depends_on-must-match-skeleton instruction above means this should not\n * happen in practice).\n */\nexport async function buildPlanFromTemplate(\n  llmClient: ILLMClient,\n  message: string,\n  template: PlanTemplate,\n  model?: string,\n  onUsage?: (usage: TokenUsage) => void,\n): Promise<Plan | null> {\n  try {\n    const response = await llmClient.callChatStructured(\n      [\n        { role: 'system', content: buildSystemPrompt(template) },\n        { role: 'user', content: message },\n      ],\n      undefined,\n      { model, onUsage, structuredOutput: { schema: PLAN_SCHEMA } },\n    )\n    const parsed = parseModelJson(response.content) as { tasks?: unknown }\n    if (!Array.isArray(parsed.tasks)) return null\n    const rawTasks = parsed.tasks.filter(isRawPlanTask)\n    if (rawTasks.length <= 1) return null\n    const riskByTemplateId = new Map(template.tasks.map((t) => [t.id, t.risk_level]))\n    const tasks: DecomposedTaskSpec[] = rawTasks.map((t) => ({ ...t, riskLevel: riskByTemplateId.get(t.id) ?? 'LOW' }))\n    return { templateName: template.name, successCriteria: template.success_criteria, tasks }\n  } catch {\n    return null\n  }\n}\n","import type {\n  ChatMessage,\n  ChatOptions,\n  ILLMClient,\n  LLMStructuredResponse,\n  ToolDefinition,\n} from '@buildaharness/runtime'\nimport { classifyRiskLexical } from './risk-classifier.js'\n\n/**\n * A deterministic, network-free {@link ILLMClient} for tests and demos — the one publicly\n * supported way to drive a real `PersonalAssistant` turn end to end without a proxy, an API key,\n * or a live model. Built for the internal plan phase B1 (the browser-e2e lane\n * that unblocks the `ASSISTANT_ONE_LOOP` default-flip), and reusable by chat-ui's `demo-seed.ts`.\n *\n * It mirrors the shape of the internal `ScriptedToolLLMClient` that `assistant.test.ts` has used\n * for a long time: every turn spends exactly one mandatory `classifyTurnIntent` call up front,\n * which this client recognises (by a distinctive phrase from that classifier's system prompt) and\n * answers on its own from a lightweight, deterministic default — so `responses` only has to\n * script the *tool loop / one-loop proposer's* own calls, not that bookkeeping call.\n *\n * `responses` entries are consumed in order, one per non-classifier, non-side-call\n * `callChatStructured` call:\n *   - a plain `string` → `{ content: string }` (a final answer, no tool calls)\n *   - an `LLMStructuredResponse` → use as-is (set `toolCalls` to simulate the model calling a tool)\n *\n * A handful of other calls are auxiliary semantic escalations (the change reviewer, the failure\n * matcher, the LLM injection classifier) that a tool-using turn can trigger without any script\n * author asking for them — with lexicalMode's default now `disabled`, the lexical checks these\n * sit on top of never short-circuit them, so they fire on every qualifying call, not just ones a\n * scenario is specifically testing. Answered from a safe, inert default (see SIDE_CALL_MARKERS)\n * so they never silently consume a `responses` slot meant for the real tool loop — the exact\n * silent-consumption bug that broke chat-ui's W5 proxy e2e test when the default flipped. Override\n * or add to this list via `sideResponses` for a test that specifically wants one of these calls to\n * answer something else (e.g. a genuine injection flag).\n *\n * `callChat`/`callChatSync` (the no-tools reply path, and post-approval synthesis) stream\n * `streamChunks` (default: `['']`).\n */\nexport interface ScriptedLLMClientScript {\n  /** Ordered structured responses for the tool loop / one-loop proposer. A bare string is shorthand for a final answer with no tool calls. */\n  responses?: Array<string | LLMStructuredResponse>\n  /** Chunks streamed from `callChat`/`callChatSync`. Defaults to `['']`. */\n  streamChunks?: string[]\n  /**\n   * Per-user-message override of the auto-answered `classifyTurnIntent` JSON fields (e.g.\n   * `{ isTrivial: true }`, `{ riskLevel: 'HIGH' }`, `{ decomposedTasks: [...] }`). Return\n   * `undefined` to keep the derived default for that message.\n   */\n  classify?: (userMessage: string) => Record<string, unknown> | undefined\n  /**\n   * Overrides/extends SIDE_CALL_MARKERS' default answers for a system-prompt substring — checked\n   * before the built-in defaults, so a test can make the change reviewer/failure matcher/injection\n   * classifier answer something other than \"nothing to see here\" without it consuming a\n   * `responses` slot either way. `[marker, rawJsonContent]` pairs, same shape as the built-ins.\n   */\n  sideResponses?: Array<[string, string]>\n}\n\n/**\n * Known auxiliary semantic-escalation system prompts and their safe, inert default answer — see\n * ScriptedLLMClientScript's doc comment. Kept in sync by hand with each check's own system prompt\n * (review-checker.ts, failure-mode-matcher.ts, trust-tagging.ts, grounding-check.ts, task-completion-check.ts, constraint-check.ts, semantic-compaction.ts, source-reliability.ts, semantic-hypotheses.ts, episodic-digest.ts, memory-consolidation.ts, memory-reviewer.ts) since there's no shared constant\n * to import without a circular dependency back into this package's own public surface.\n */\nexport const SIDE_CALL_MARKERS: Array<[string, string]> = [\n  ['You check whether a proposed action genuinely conflicts', '{\"conflict\":false}'],\n  ['You match a set of observed symptoms against a curated library', 'null'],\n  ['You are a security classifier analyzing untrusted external content', '{\"flagged\":false}'],\n  ['You check whether an assistant\\'s reply is faithful to the raw results', '{\"verdict\":\"grounded\"}'],\n  ['You judge whether an assistant actually accomplished a task', '{\"done\":true}'],\n  ['You condense the earlier part of a conversation', '{\"summary\":\"\"}'],\n  ['You judge whether an assistant\\'s reply violates constraints', '{\"violations\":[]}'],\n  ['You weigh the reliability of the sources an assistant read', '{\"assessments\":[],\"weighed\":true}'],\n  ['You propose competing explanations for an underdetermined request', '{\"hypotheses\":[]}'],\n  ['You decide which explanations a set of new observations rules out', '{\"contradicted\":[]}'],\n  ['You write a session handoff digest', '{\"digest\":null}'],\n  ['You consolidate a small store of remembered facts about a user', '{\"proposals\":[]}'],\n  ['You review what a user said across several turns', '{\"ops\":[]}'],\n  ['You check proposed memory entries against what a user actually said', '{\"verdicts\":[]}'],\n]\n\n/** The phrase that opens `turn-intent-classifier.ts`'s system prompt — stable, and how every scripted client tells that mandatory call apart from a real tool-loop call. */\nconst TURN_INTENT_MARKER = ' independent judgments'\n\nfunction isTurnIntentRequest(messages: ChatMessage[]): boolean {\n  return messages.some((m) => m.role === 'system' && m.content.includes(TURN_INTENT_MARKER))\n}\n\n/**\n * The deterministic stand-in for what a real model would return for `classifyTurnIntent` —\n * risk comes from the same lexical `classifyRisk` the real fallback path uses; everything else\n * is inert (not trivial, not a reminder, no decomposition, no plan template) unless `override`\n * says otherwise. `isTrivial` defaults to `false` so a scripted turn actually exercises the\n * harness loop (the interesting case for one-loop parity tests); pass `{ isTrivial: true }` via\n * `classify` for the triviality fast path.\n */\nfunction deriveTurnIntentJSON(messages: ChatMessage[], override?: Record<string, unknown>): string {\n  const userContent = messages.find((m) => m.role === 'user')?.content ?? ''\n  const risk = classifyRiskLexical(userContent)\n  const isReminderRequest = risk.reason.includes('reminder')\n  const base = {\n    riskLevel: risk.riskLevel,\n    riskReason: risk.reason,\n    isTrivial: false,\n    decomposedTasks: [],\n    isReminderRequest,\n    isBulkReminderRequest: isReminderRequest && risk.requiresApproval,\n    isAbandonRequest: false,\n    isPlanQuestion: false,\n    isUnderdetermined: false,\n    matchedPlanTemplate: null,\n    needsMultiStepPlan: false,\n    statesDurableFacts: [],\n  }\n  return JSON.stringify({ ...base, ...override })\n}\n\nclass ScriptedLLMClient implements ILLMClient {\n  /** Non-classifier `callChatStructured` calls made so far — the index into `responses`. */\n  toolCalls = 0\n  private responseIndex = 0\n\n  constructor(private readonly script: ScriptedLLMClientScript) {}\n\n  private get streamChunks(): string[] {\n    return this.script.streamChunks && this.script.streamChunks.length > 0 ? this.script.streamChunks : ['']\n  }\n\n  async *callChat(_messages: ChatMessage[], _options?: ChatOptions): AsyncIterable<string> {\n    for (const chunk of this.streamChunks) yield chunk\n  }\n\n  async callChatSync(messages: ChatMessage[], options?: ChatOptions): Promise<string> {\n    const chunks: string[] = []\n    for await (const chunk of this.callChat(messages, options)) chunks.push(chunk)\n    return chunks.join('')\n  }\n\n  async callChatStructured(messages: ChatMessage[], _tools?: ToolDefinition[], _options?: ChatOptions): Promise<LLMStructuredResponse> {\n    if (isTurnIntentRequest(messages)) {\n      const userContent = messages.find((m) => m.role === 'user')?.content ?? ''\n      return { content: deriveTurnIntentJSON(messages, this.script.classify?.(userContent)) }\n    }\n    const system = messages.find((m) => m.role === 'system')?.content ?? ''\n    const side = [...(this.script.sideResponses ?? []), ...SIDE_CALL_MARKERS].find(([marker]) => system.includes(marker))\n    if (side) return { content: side[1] }\n    this.toolCalls++\n    const responses = this.script.responses ?? []\n    if (this.responseIndex >= responses.length) {\n      throw new Error(\n        `createScriptedLLMClient: no scripted response for tool-loop call #${this.responseIndex + 1} (scripted ${responses.length})`,\n      )\n    }\n    const next = responses[this.responseIndex++]\n    return typeof next === 'string' ? { content: next } : next\n  }\n}\n\n/**\n * Builds a scripted {@link ILLMClient}. See {@link ScriptedLLMClientScript}. The returned value\n * is a plain `ILLMClient` — pass it straight to `PersonalAssistant.create({ llmClient })`.\n */\nexport function createScriptedLLMClient(script: ScriptedLLMClientScript = {}): ILLMClient {\n  return new ScriptedLLMClient(script)\n}\n"],"names":["t","v","m"],"mappings":";;;;;;AAkBA,MAAM,cAAc;AAAA,EAClB,MAAM;AAAA,EACN,YAAY;AAAA,IACV,OAAO;AAAA,MACL,MAAM;AAAA,MACN,OAAO;AAAA,QACL,MAAM;AAAA,QACN,YAAY;AAAA,UACV,IAAI,EAAE,MAAM,SAAA;AAAA,UACZ,aAAa,EAAE,MAAM,SAAA;AAAA,UACrB,YAAY,EAAE,MAAM,SAAS,OAAO,EAAE,MAAM,WAAS;AAAA,QAAE;AAAA,QAEzD,UAAU,CAAC,MAAM,eAAe,YAAY;AAAA,MAAA;AAAA,IAC9C;AAAA,EACF;AAAA,EAEF,UAAU,CAAC,OAAO;AACpB;AAEA,SAAS,kBAAkB,UAAgC;AACzD,QAAM,WAAW,SAAS,MACvB,IAAI,CAACA,OAAM,SAASA,GAAE,EAAE,YAAYA,GAAE,KAAK,kBAAkBA,GAAE,WAAW,KAAK,IAAI,CAAC,GAAG,EACvF,KAAK,IAAI;AACZ,SACE,uBAAuB,SAAS,IAAI;AAAA,EAE2D,QAAQ;AAAA;AAAA,0CAC5D,SAAS,gBAAgB;AAAA;AAAA;AAOxE;AAYA,SAAS,cAAc,OAAsC;AAC3D,MAAI,OAAO,UAAU,YAAY,UAAU,KAAM,QAAO;AACxD,QAAMC,KAAI;AACV,SACE,OAAOA,GAAE,OAAO,YAChB,OAAOA,GAAE,gBAAgB,YACzB,MAAM,QAAQA,GAAE,UAAU,KAC1BA,GAAE,WAAW,MAAM,CAAC,MAAM,OAAO,MAAM,QAAQ;AAEnD;AAkBA,eAAsB,sBACpB,WACA,SACA,UACA,OACA,SACsB;AACtB,MAAI;AACF,UAAM,WAAW,MAAM,UAAU;AAAA,MAC/B;AAAA,QACE,EAAE,MAAM,UAAU,SAAS,kBAAkB,QAAQ,EAAA;AAAA,QACrD,EAAE,MAAM,QAAQ,SAAS,QAAA;AAAA,MAAQ;AAAA,MAEnC;AAAA,MACA,EAAE,OAAO,SAAS,kBAAkB,EAAE,QAAQ,cAAY;AAAA,IAAE;AAE9D,UAAM,SAAS,eAAe,SAAS,OAAO;AAC9C,QAAI,CAAC,MAAM,QAAQ,OAAO,KAAK,EAAG,QAAO;AACzC,UAAM,WAAW,OAAO,MAAM,OAAO,aAAa;AAClD,QAAI,SAAS,UAAU,EAAG,QAAO;AACjC,UAAM,mBAAmB,IAAI,IAAI,SAAS,MAAM,IAAI,CAACD,OAAM,CAACA,GAAE,IAAIA,GAAE,UAAU,CAAC,CAAC;AAChF,UAAM,QAA8B,SAAS,IAAI,CAACA,QAAO,EAAE,GAAGA,IAAG,WAAW,iBAAiB,IAAIA,GAAE,EAAE,KAAK,QAAQ;AAClH,WAAO,EAAE,cAAc,SAAS,MAAM,iBAAiB,SAAS,kBAAkB,MAAA;AAAA,EACpF,QAAQ;AACN,WAAO;AAAA,EACT;AACF;ACpDO,MAAM,oBAA6C;AAAA,EACxD,CAAC,2DAA2D,oBAAoB;AAAA,EAChF,CAAC,kEAAkE,MAAM;AAAA,EACzE,CAAC,sEAAsE,mBAAmB;AAAA,EAC1F,CAAC,yEAA0E,wBAAwB;AAAA,EACnG,CAAC,+DAA+D,eAAe;AAAA,EAC/E,CAAC,mDAAmD,gBAAgB;AAAA,EACpE,CAAC,+DAAgE,mBAAmB;AAAA,EACpF,CAAC,8DAA8D,mCAAmC;AAAA,EAClG,CAAC,qEAAqE,mBAAmB;AAAA,EACzF,CAAC,qEAAqE,qBAAqB;AAAA,EAC3F,CAAC,sCAAsC,iBAAiB;AAAA,EACxD,CAAC,kEAAkE,kBAAkB;AAAA,EACrF,CAAC,oDAAoD,YAAY;AAAA,EACjE,CAAC,uEAAuE,iBAAiB;AAC3F;AAGA,MAAM,qBAAqB;AAE3B,SAAS,oBAAoB,UAAkC;AAC7D,SAAO,SAAS,KAAK,CAACE,OAAMA,GAAE,SAAS,YAAYA,GAAE,QAAQ,SAAS,kBAAkB,CAAC;AAC3F;AAUA,SAAS,qBAAqB,UAAyB,UAA4C;;AACjG,QAAM,gBAAc,cAAS,KAAK,CAACA,OAAMA,GAAE,SAAS,MAAM,MAAtC,mBAAyC,YAAW;AACxE,QAAM,OAAO,oBAAoB,WAAW;AAC5C,QAAM,oBAAoB,KAAK,OAAO,SAAS,UAAU;AACzD,QAAM,OAAO;AAAA,IACX,WAAW,KAAK;AAAA,IAChB,YAAY,KAAK;AAAA,IACjB,WAAW;AAAA,IACX,iBAAiB,CAAA;AAAA,IACjB;AAAA,IACA,uBAAuB,qBAAqB,KAAK;AAAA,IACjD,kBAAkB;AAAA,IAClB,gBAAgB;AAAA,IAChB,mBAAmB;AAAA,IACnB,qBAAqB;AAAA,IACrB,oBAAoB;AAAA,IACpB,oBAAoB,CAAA;AAAA,EAAC;AAEvB,SAAO,KAAK,UAAU,EAAE,GAAG,MAAM,GAAG,UAAU;AAChD;AAEA,MAAM,kBAAwC;AAAA,EAK5C,YAA6B,QAAiC;AAH9D;AAAA,qCAAY;AACJ,yCAAgB;AAEK,SAAA,SAAA;AAAA,EAAkC;AAAA,EAE/D,IAAY,eAAyB;AACnC,WAAO,KAAK,OAAO,gBAAgB,KAAK,OAAO,aAAa,SAAS,IAAI,KAAK,OAAO,eAAe,CAAC,EAAE;AAAA,EACzG;AAAA,EAEA,OAAO,SAAS,WAA0B,UAA+C;AACvF,eAAW,SAAS,KAAK,aAAc,OAAM;AAAA,EAC/C;AAAA,EAEA,MAAM,aAAa,UAAyB,SAAwC;AAClF,UAAM,SAAmB,CAAA;AACzB,qBAAiB,SAAS,KAAK,SAAS,UAAU,OAAO,EAAG,QAAO,KAAK,KAAK;AAC7E,WAAO,OAAO,KAAK,EAAE;AAAA,EACvB;AAAA,EAEA,MAAM,mBAAmB,UAAyB,QAA2B,UAAwD;;AACnI,QAAI,oBAAoB,QAAQ,GAAG;AACjC,YAAM,gBAAc,cAAS,KAAK,CAACA,OAAMA,GAAE,SAAS,MAAM,MAAtC,mBAAyC,YAAW;AACxE,aAAO,EAAE,SAAS,qBAAqB,WAAU,gBAAK,QAAO,aAAZ,4BAAuB,YAAY,EAAA;AAAA,IACtF;AACA,UAAM,WAAS,cAAS,KAAK,CAACA,OAAMA,GAAE,SAAS,QAAQ,MAAxC,mBAA2C,YAAW;AACrE,UAAM,OAAO,CAAC,GAAI,KAAK,OAAO,iBAAiB,CAAA,GAAK,GAAG,iBAAiB,EAAE,KAAK,CAAC,CAAC,MAAM,MAAM,OAAO,SAAS,MAAM,CAAC;AACpH,QAAI,KAAM,QAAO,EAAE,SAAS,KAAK,CAAC,EAAA;AAClC,SAAK;AACL,UAAM,YAAY,KAAK,OAAO,aAAa,CAAA;AAC3C,QAAI,KAAK,iBAAiB,UAAU,QAAQ;AAC1C,YAAM,IAAI;AAAA,QACR,qEAAqE,KAAK,gBAAgB,CAAC,cAAc,UAAU,MAAM;AAAA,MAAA;AAAA,IAE7H;AACA,UAAM,OAAO,UAAU,KAAK,eAAe;AAC3C,WAAO,OAAO,SAAS,WAAW,EAAE,SAAS,SAAS;AAAA,EACxD;AACF;AAMO,SAAS,wBAAwB,SAAkC,IAAgB;AACxF,SAAO,IAAI,kBAAkB,MAAM;AACrC;"}
+{"version":3,"file":"index.js","sources":["../src/plan-builder.ts","../src/scripted-llm-client.ts"],"sourcesContent":["import type { ILLMClient, TokenUsage } from '@buildaharness/runtime'\nimport type { DecomposedTaskSpec } from './decomposition-classifier.js'\nimport type { PlanTemplate } from './plan-templates/index.js'\nimport { parseModelJson } from './model-json.js'\n\nexport interface Plan {\n  templateName: string\n  successCriteria: string\n  tasks: DecomposedTaskSpec[]\n  /**\n   * Why this approach, not just what the success criteria are — optional for now since neither\n   * buildPlanFromTemplate's template-personalization call nor any other current producer of a\n   * Plan asks the LLM for one; plan-store.ts's createPlanRecord defaults it to '' when absent.\n   * A future from-scratch drafting path (plan mode's P3) is expected to populate it.\n   */\n  rationale?: string\n}\n\nconst PLAN_SCHEMA = {\n  type: 'object',\n  properties: {\n    tasks: {\n      type: 'array',\n      items: {\n        type: 'object',\n        properties: {\n          id: { type: 'string' },\n          description: { type: 'string' },\n          depends_on: { type: 'array', items: { type: 'string' } },\n        },\n        required: ['id', 'description', 'depends_on'],\n      },\n    },\n  },\n  required: ['tasks'],\n}\n\nfunction buildSystemPrompt(template: PlanTemplate): string {\n  const skeleton = template.tasks\n    .map((t) => `- id: ${t.id}; title: ${t.title}; depends_on: [${t.depends_on.join(', ')}]`)\n    .join('\\n')\n  return (\n    `You are adapting a \"${template.name}\" plan template to a specific user request. ` +\n    `Here is the template's task skeleton — keep the exact same ids and depends_on structure, ` +\n    `one output task per skeleton task, but personalize each description to the actual request:\\n${skeleton}\\n\\n` +\n    `Success criteria for this kind of plan: ${template.success_criteria}\\n\\n` +\n    'Phrase each `description` starting with the concrete subject or object it acts on (e.g. \"the login tests: ' +\n    'rerun after the config fix\" rather than \"rerun the login tests after the config fix\"), so later comparisons ' +\n    \"against this task's completion/failure beliefs share matching vocabulary. \" +\n    'Respond with JSON only, no prose: {\"tasks\":[{\"id\": string, \"description\": string, \"depends_on\": ' +\n    'string[]}]}. `id` and `depends_on` values must exactly match the skeleton above.'\n  )\n}\n\n// Only the shape the LLM actually returns (id/description/depends_on) — riskLevel isn't asked for\n// here at all (see buildPlanFromTemplate's doc comment for why) and gets attached afterward from\n// the template's own already-curated risk_level, so this is deliberately narrower than\n// DecomposedTaskSpec itself.\ninterface RawPlanTask {\n  id: string\n  description: string\n  depends_on: string[]\n}\n\nfunction isRawPlanTask(value: unknown): value is RawPlanTask {\n  if (typeof value !== 'object' || value === null) return false\n  const v = value as Record<string, unknown>\n  return (\n    typeof v.id === 'string' &&\n    typeof v.description === 'string' &&\n    Array.isArray(v.depends_on) &&\n    v.depends_on.every((d) => typeof d === 'string')\n  )\n}\n\n/**\n * Spends one real LLM call personalizing `template`'s task skeleton to `message` —\n * only called once classifyTurnIntent has already named a matchedPlanTemplate, so an\n * ordinary turn never pays for this. Same \"malformed/incomplete JSON is the expected\n * failure mode, not the edge case\" fallback classifyTurnIntent itself uses: any parse\n * failure or a response with fewer than 2 usable tasks returns null, meaning \"fall back to\n * the caller's own ad hoc decomposition for this turn\" rather than throwing.\n *\n * Each output task's `riskLevel` is attached from the matching skeleton task's own\n * `risk_level` (curated by whoever wrote the template's JSON) by id, not asked of the LLM —\n * the template already has this, more accurately than an LLM re-deriving it from a\n * personalized one-line description would, and it costs zero extra output tokens. Falls back\n * to 'LOW' only for the pathological case of a task id the LLM invented that doesn't match any\n * skeleton task (the id/depends_on-must-match-skeleton instruction above means this should not\n * happen in practice).\n */\nexport async function buildPlanFromTemplate(\n  llmClient: ILLMClient,\n  message: string,\n  template: PlanTemplate,\n  model?: string,\n  onUsage?: (usage: TokenUsage) => void,\n): Promise<Plan | null> {\n  try {\n    const response = await llmClient.callChatStructured(\n      [\n        { role: 'system', content: buildSystemPrompt(template) },\n        { role: 'user', content: message },\n      ],\n      undefined,\n      { model, onUsage, structuredOutput: { schema: PLAN_SCHEMA } },\n    )\n    const parsed = parseModelJson(response.content) as { tasks?: unknown }\n    if (!Array.isArray(parsed.tasks)) return null\n    const rawTasks = parsed.tasks.filter(isRawPlanTask)\n    if (rawTasks.length <= 1) return null\n    const riskByTemplateId = new Map(template.tasks.map((t) => [t.id, t.risk_level]))\n    const tasks: DecomposedTaskSpec[] = rawTasks.map((t) => ({ ...t, riskLevel: riskByTemplateId.get(t.id) ?? 'LOW' }))\n    return { templateName: template.name, successCriteria: template.success_criteria, tasks }\n  } catch {\n    return null\n  }\n}\n","import type {\n  ChatMessage,\n  ChatOptions,\n  ILLMClient,\n  LLMStructuredResponse,\n  ToolDefinition,\n} from '@buildaharness/runtime'\nimport { classifyRiskLexical } from './risk-classifier.js'\n\n/**\n * A deterministic, network-free {@link ILLMClient} for tests and demos — the one publicly\n * supported way to drive a real `PersonalAssistant` turn end to end without a proxy, an API key,\n * or a live model. Built for the internal plan phase B1 (the browser-e2e lane\n * that unblocks the `ASSISTANT_ONE_LOOP` default-flip), and reusable by chat-ui's `demo-seed.ts`.\n *\n * It mirrors the shape of the internal `ScriptedToolLLMClient` that `assistant.test.ts` has used\n * for a long time: every turn spends exactly one mandatory `classifyTurnIntent` call up front,\n * which this client recognises (by a distinctive phrase from that classifier's system prompt) and\n * answers on its own from a lightweight, deterministic default — so `responses` only has to\n * script the *tool loop / one-loop proposer's* own calls, not that bookkeeping call.\n *\n * `responses` entries are consumed in order, one per non-classifier, non-side-call\n * `callChatStructured` call:\n *   - a plain `string` → `{ content: string }` (a final answer, no tool calls)\n *   - an `LLMStructuredResponse` → use as-is (set `toolCalls` to simulate the model calling a tool)\n *\n * A handful of other calls are auxiliary semantic escalations (the change reviewer, the failure\n * matcher, the LLM injection classifier) that a tool-using turn can trigger without any script\n * author asking for them — with lexicalMode's default now `disabled`, the lexical checks these\n * sit on top of never short-circuit them, so they fire on every qualifying call, not just ones a\n * scenario is specifically testing. Answered from a safe, inert default (see SIDE_CALL_MARKERS)\n * so they never silently consume a `responses` slot meant for the real tool loop — the exact\n * silent-consumption bug that broke chat-ui's W5 proxy e2e test when the default flipped. Override\n * or add to this list via `sideResponses` for a test that specifically wants one of these calls to\n * answer something else (e.g. a genuine injection flag).\n *\n * `callChat`/`callChatSync` (the no-tools reply path, and post-approval synthesis) stream\n * `streamChunks` (default: `['']`).\n */\nexport interface ScriptedLLMClientScript {\n  /** Ordered structured responses for the tool loop / one-loop proposer. A bare string is shorthand for a final answer with no tool calls. */\n  responses?: Array<string | LLMStructuredResponse>\n  /** Chunks streamed from `callChat`/`callChatSync`. Defaults to `['']`. */\n  streamChunks?: string[]\n  /**\n   * Per-user-message override of the auto-answered `classifyTurnIntent` JSON fields (e.g.\n   * `{ isTrivial: true }`, `{ riskLevel: 'HIGH' }`, `{ decomposedTasks: [...] }`). Return\n   * `undefined` to keep the derived default for that message.\n   */\n  classify?: (userMessage: string) => Record<string, unknown> | undefined\n  /**\n   * Overrides/extends SIDE_CALL_MARKERS' default answers for a system-prompt substring — checked\n   * before the built-in defaults, so a test can make the change reviewer/failure matcher/injection\n   * classifier answer something other than \"nothing to see here\" without it consuming a\n   * `responses` slot either way. `[marker, rawJsonContent]` pairs, same shape as the built-ins.\n   */\n  sideResponses?: Array<[string, string]>\n}\n\n/**\n * Known auxiliary semantic-escalation system prompts and their safe, inert default answer — see\n * ScriptedLLMClientScript's doc comment. Kept in sync by hand with each check's own system prompt\n * (review-checker.ts, failure-mode-matcher.ts, trust-tagging.ts, grounding-check.ts, task-completion-check.ts, constraint-check.ts, semantic-compaction.ts, source-reliability.ts, semantic-hypotheses.ts, episodic-digest.ts, memory-consolidation.ts, memory-reviewer.ts) since there's no shared constant\n * to import without a circular dependency back into this package's own public surface.\n */\nexport const SIDE_CALL_MARKERS: Array<[string, string]> = [\n  ['You check whether a proposed action genuinely conflicts', '{\"conflict\":false}'],\n  ['You match a set of observed symptoms against a curated library', 'null'],\n  ['You are a security classifier analyzing untrusted external content', '{\"flagged\":false}'],\n  ['You check whether an assistant\\'s reply is faithful to the raw results', '{\"verdict\":\"grounded\"}'],\n  ['You judge whether an assistant actually accomplished a task', '{\"done\":true}'],\n  ['You decide whether a user\\'s request to an assistant with file and shell tools fully determines the', '{\"ambiguous\":false,\"question\":\"\"}'],\n  ['You condense the earlier part of a conversation', '{\"summary\":\"\"}'],\n  ['You judge whether an assistant\\'s reply violates constraints', '{\"violations\":[]}'],\n  ['You weigh the reliability of the sources an assistant read', '{\"assessments\":[],\"weighed\":true}'],\n  ['You propose competing explanations for an underdetermined request', '{\"hypotheses\":[]}'],\n  ['You decide which explanations a set of new observations rules out', '{\"contradicted\":[]}'],\n  ['You write a session handoff digest', '{\"digest\":null}'],\n  ['You consolidate a small store of remembered facts about a user', '{\"proposals\":[]}'],\n  ['You review what a user said across several turns', '{\"ops\":[]}'],\n  ['You check proposed memory entries against what a user actually said', '{\"verdicts\":[]}'],\n]\n\n/** The phrase that opens `turn-intent-classifier.ts`'s system prompt — stable, and how every scripted client tells that mandatory call apart from a real tool-loop call. */\nconst TURN_INTENT_MARKER = ' independent judgments'\n\nfunction isTurnIntentRequest(messages: ChatMessage[]): boolean {\n  return messages.some((m) => m.role === 'system' && m.content.includes(TURN_INTENT_MARKER))\n}\n\n/**\n * The deterministic stand-in for what a real model would return for `classifyTurnIntent` —\n * risk comes from the same lexical `classifyRisk` the real fallback path uses; everything else\n * is inert (not trivial, not a reminder, no decomposition, no plan template) unless `override`\n * says otherwise. `isTrivial` defaults to `false` so a scripted turn actually exercises the\n * harness loop (the interesting case for one-loop parity tests); pass `{ isTrivial: true }` via\n * `classify` for the triviality fast path.\n */\nfunction deriveTurnIntentJSON(messages: ChatMessage[], override?: Record<string, unknown>): string {\n  const userContent = messages.find((m) => m.role === 'user')?.content ?? ''\n  const risk = classifyRiskLexical(userContent)\n  const isReminderRequest = risk.reason.includes('reminder')\n  const base = {\n    riskLevel: risk.riskLevel,\n    riskReason: risk.reason,\n    isTrivial: false,\n    decomposedTasks: [],\n    isReminderRequest,\n    isBulkReminderRequest: isReminderRequest && risk.requiresApproval,\n    isAbandonRequest: false,\n    isPlanQuestion: false,\n    isUnderdetermined: false,\n    matchedPlanTemplate: null,\n    needsMultiStepPlan: false,\n    statesDurableFacts: [],\n  }\n  return JSON.stringify({ ...base, ...override })\n}\n\nclass ScriptedLLMClient implements ILLMClient {\n  /** Non-classifier `callChatStructured` calls made so far — the index into `responses`. */\n  toolCalls = 0\n  private responseIndex = 0\n\n  constructor(private readonly script: ScriptedLLMClientScript) {}\n\n  private get streamChunks(): string[] {\n    return this.script.streamChunks && this.script.streamChunks.length > 0 ? this.script.streamChunks : ['']\n  }\n\n  async *callChat(_messages: ChatMessage[], _options?: ChatOptions): AsyncIterable<string> {\n    for (const chunk of this.streamChunks) yield chunk\n  }\n\n  async callChatSync(messages: ChatMessage[], options?: ChatOptions): Promise<string> {\n    const chunks: string[] = []\n    for await (const chunk of this.callChat(messages, options)) chunks.push(chunk)\n    return chunks.join('')\n  }\n\n  async callChatStructured(messages: ChatMessage[], _tools?: ToolDefinition[], _options?: ChatOptions): Promise<LLMStructuredResponse> {\n    if (isTurnIntentRequest(messages)) {\n      const userContent = messages.find((m) => m.role === 'user')?.content ?? ''\n      return { content: deriveTurnIntentJSON(messages, this.script.classify?.(userContent)) }\n    }\n    const system = messages.find((m) => m.role === 'system')?.content ?? ''\n    const side = [...(this.script.sideResponses ?? []), ...SIDE_CALL_MARKERS].find(([marker]) => system.includes(marker))\n    if (side) return { content: side[1] }\n    this.toolCalls++\n    const responses = this.script.responses ?? []\n    if (this.responseIndex >= responses.length) {\n      throw new Error(\n        `createScriptedLLMClient: no scripted response for tool-loop call #${this.responseIndex + 1} (scripted ${responses.length})`,\n      )\n    }\n    const next = responses[this.responseIndex++]\n    return typeof next === 'string' ? { content: next } : next\n  }\n}\n\n/**\n * Builds a scripted {@link ILLMClient}. See {@link ScriptedLLMClientScript}. The returned value\n * is a plain `ILLMClient` — pass it straight to `PersonalAssistant.create({ llmClient })`.\n */\nexport function createScriptedLLMClient(script: ScriptedLLMClientScript = {}): ILLMClient {\n  return new ScriptedLLMClient(script)\n}\n"],"names":["d","m"],"mappings":";;;;;;AAkBA,MAAM,cAAc;AAAA,EAClB,MAAM;AAAA,EACN,YAAY;AAAA,IACV,OAAO;AAAA,MACL,MAAM;AAAA,MACN,OAAO;AAAA,QACL,MAAM;AAAA,QACN,YAAY;AAAA,UACV,IAAI,EAAE,MAAM,SAAA;AAAA,UACZ,aAAa,EAAE,MAAM,SAAA;AAAA,UACrB,YAAY,EAAE,MAAM,SAAS,OAAO,EAAE,MAAM,WAAS;AAAA,QAAE;AAAA,QAEzD,UAAU,CAAC,MAAM,eAAe,YAAY;AAAA,MAAA;AAAA,IAC9C;AAAA,EACF;AAAA,EAEF,UAAU,CAAC,OAAO;AACpB;AAEA,SAAS,kBAAkB,UAAgC;AACzD,QAAM,WAAW,SAAS,MACvB,IAAI,CAAC,MAAM,SAAS,EAAE,EAAE,YAAY,EAAE,KAAK,kBAAkB,EAAE,WAAW,KAAK,IAAI,CAAC,GAAG,EACvF,KAAK,IAAI;AACZ,SACE,uBAAuB,SAAS,IAAI;AAAA,EAE2D,QAAQ;AAAA;AAAA,0CAC5D,SAAS,gBAAgB;AAAA;AAAA;AAOxE;AAYA,SAAS,cAAc,OAAsC;AAC3D,MAAI,OAAO,UAAU,YAAY,UAAU,KAAM,QAAO;AACxD,QAAM,IAAI;AACV,SACE,OAAO,EAAE,OAAO,YAChB,OAAO,EAAE,gBAAgB,YACzB,MAAM,QAAQ,EAAE,UAAU,KAC1B,EAAE,WAAW,MAAM,CAACA,OAAM,OAAOA,OAAM,QAAQ;AAEnD;AAkBA,eAAsB,sBACpB,WACA,SACA,UACA,OACA,SACsB;AACtB,MAAI;AACF,UAAM,WAAW,MAAM,UAAU;AAAA,MAC/B;AAAA,QACE,EAAE,MAAM,UAAU,SAAS,kBAAkB,QAAQ,EAAA;AAAA,QACrD,EAAE,MAAM,QAAQ,SAAS,QAAA;AAAA,MAAQ;AAAA,MAEnC;AAAA,MACA,EAAE,OAAO,SAAS,kBAAkB,EAAE,QAAQ,cAAY;AAAA,IAAE;AAE9D,UAAM,SAAS,eAAe,SAAS,OAAO;AAC9C,QAAI,CAAC,MAAM,QAAQ,OAAO,KAAK,EAAG,QAAO;AACzC,UAAM,WAAW,OAAO,MAAM,OAAO,aAAa;AAClD,QAAI,SAAS,UAAU,EAAG,QAAO;AACjC,UAAM,mBAAmB,IAAI,IAAI,SAAS,MAAM,IAAI,CAAC,MAAM,CAAC,EAAE,IAAI,EAAE,UAAU,CAAC,CAAC;AAChF,UAAM,QAA8B,SAAS,IAAI,CAAC,OAAO,EAAE,GAAG,GAAG,WAAW,iBAAiB,IAAI,EAAE,EAAE,KAAK,QAAQ;AAClH,WAAO,EAAE,cAAc,SAAS,MAAM,iBAAiB,SAAS,kBAAkB,MAAA;AAAA,EACpF,QAAQ;AACN,WAAO;AAAA,EACT;AACF;ACpDO,MAAM,oBAA6C;AAAA,EACxD,CAAC,2DAA2D,oBAAoB;AAAA,EAChF,CAAC,kEAAkE,MAAM;AAAA,EACzE,CAAC,sEAAsE,mBAAmB;AAAA,EAC1F,CAAC,yEAA0E,wBAAwB;AAAA,EACnG,CAAC,+DAA+D,eAAe;AAAA,EAC/E,CAAC,sGAAuG,mCAAmC;AAAA,EAC3I,CAAC,mDAAmD,gBAAgB;AAAA,EACpE,CAAC,+DAAgE,mBAAmB;AAAA,EACpF,CAAC,8DAA8D,mCAAmC;AAAA,EAClG,CAAC,qEAAqE,mBAAmB;AAAA,EACzF,CAAC,qEAAqE,qBAAqB;AAAA,EAC3F,CAAC,sCAAsC,iBAAiB;AAAA,EACxD,CAAC,kEAAkE,kBAAkB;AAAA,EACrF,CAAC,oDAAoD,YAAY;AAAA,EACjE,CAAC,uEAAuE,iBAAiB;AAC3F;AAGA,MAAM,qBAAqB;AAE3B,SAAS,oBAAoB,UAAkC;AAC7D,SAAO,SAAS,KAAK,CAACC,OAAMA,GAAE,SAAS,YAAYA,GAAE,QAAQ,SAAS,kBAAkB,CAAC;AAC3F;AAUA,SAAS,qBAAqB,UAAyB,UAA4C;;AACjG,QAAM,gBAAc,cAAS,KAAK,CAACA,OAAMA,GAAE,SAAS,MAAM,MAAtC,mBAAyC,YAAW;AACxE,QAAM,OAAO,oBAAoB,WAAW;AAC5C,QAAM,oBAAoB,KAAK,OAAO,SAAS,UAAU;AACzD,QAAM,OAAO;AAAA,IACX,WAAW,KAAK;AAAA,IAChB,YAAY,KAAK;AAAA,IACjB,WAAW;AAAA,IACX,iBAAiB,CAAA;AAAA,IACjB;AAAA,IACA,uBAAuB,qBAAqB,KAAK;AAAA,IACjD,kBAAkB;AAAA,IAClB,gBAAgB;AAAA,IAChB,mBAAmB;AAAA,IACnB,qBAAqB;AAAA,IACrB,oBAAoB;AAAA,IACpB,oBAAoB,CAAA;AAAA,EAAC;AAEvB,SAAO,KAAK,UAAU,EAAE,GAAG,MAAM,GAAG,UAAU;AAChD;AAEA,MAAM,kBAAwC;AAAA,EAK5C,YAA6B,QAAiC;AAH9D;AAAA,qCAAY;AACJ,yCAAgB;AAEK,SAAA,SAAA;AAAA,EAAkC;AAAA,EAE/D,IAAY,eAAyB;AACnC,WAAO,KAAK,OAAO,gBAAgB,KAAK,OAAO,aAAa,SAAS,IAAI,KAAK,OAAO,eAAe,CAAC,EAAE;AAAA,EACzG;AAAA,EAEA,OAAO,SAAS,WAA0B,UAA+C;AACvF,eAAW,SAAS,KAAK,aAAc,OAAM;AAAA,EAC/C;AAAA,EAEA,MAAM,aAAa,UAAyB,SAAwC;AAClF,UAAM,SAAmB,CAAA;AACzB,qBAAiB,SAAS,KAAK,SAAS,UAAU,OAAO,EAAG,QAAO,KAAK,KAAK;AAC7E,WAAO,OAAO,KAAK,EAAE;AAAA,EACvB;AAAA,EAEA,MAAM,mBAAmB,UAAyB,QAA2B,UAAwD;;AACnI,QAAI,oBAAoB,QAAQ,GAAG;AACjC,YAAM,gBAAc,cAAS,KAAK,CAACA,OAAMA,GAAE,SAAS,MAAM,MAAtC,mBAAyC,YAAW;AACxE,aAAO,EAAE,SAAS,qBAAqB,WAAU,gBAAK,QAAO,aAAZ,4BAAuB,YAAY,EAAA;AAAA,IACtF;AACA,UAAM,WAAS,cAAS,KAAK,CAACA,OAAMA,GAAE,SAAS,QAAQ,MAAxC,mBAA2C,YAAW;AACrE,UAAM,OAAO,CAAC,GAAI,KAAK,OAAO,iBAAiB,CAAA,GAAK,GAAG,iBAAiB,EAAE,KAAK,CAAC,CAAC,MAAM,MAAM,OAAO,SAAS,MAAM,CAAC;AACpH,QAAI,KAAM,QAAO,EAAE,SAAS,KAAK,CAAC,EAAA;AAClC,SAAK;AACL,UAAM,YAAY,KAAK,OAAO,aAAa,CAAA;AAC3C,QAAI,KAAK,iBAAiB,UAAU,QAAQ;AAC1C,YAAM,IAAI;AAAA,QACR,qEAAqE,KAAK,gBAAgB,CAAC,cAAc,UAAU,MAAM;AAAA,MAAA;AAAA,IAE7H;AACA,UAAM,OAAO,UAAU,KAAK,eAAe;AAC3C,WAAO,OAAO,SAAS,WAAW,EAAE,SAAS,SAAS;AAAA,EACxD;AACF;AAMO,SAAS,wBAAwB,SAAkC,IAAgB;AACxF,SAAO,IAAI,kBAAkB,MAAM;AACrC;"}

package/dist/llm-client-factory.d.ts

@@ -0,0 +1,13 @@
+import { ILLMClient } from '@buildaharness/runtime';
+import { AssistantConfig } from './config.js';
+export interface LlmClientFactoryOptions {
+    /** Passed to every HTTP client. The desktop app supplies Tauri's fetch, because its webview CSP blocks the global one. */
+    fetchImpl?: typeof fetch;
+    /**
+     * Builds the surface's own claude-cli client (the CLI's file/shell/web/action tools, desktop's Tauri bridge).
+     * Return undefined where `claude -p` cannot run (a plain browser tab) and the factory falls back to the proxy.
+     */
+    claudeCli: () => ILLMClient | undefined;
+}
+/** The one switch over the five backends. Only the claude-cli client and `fetch` differ per surface, and both come in through the options. */
+export declare function createLlmClient(config: AssistantConfig, { fetchImpl, claudeCli }: LlmClientFactoryOptions): ILLMClient;

package/dist/memory-service.d.ts

@@ -47,6 +47,8 @@ export interface AuditEntry {
     undoes?: number;
     /** M6: the user erased this fact from history (`/memory archive forget`); before/after were removed and the entry can no longer be undone. */
     erased?: boolean;
+    /** Export only: the images of a forgotten fact are left out of `/memory export`; the stored log still has them so `/memory undo` works. */
+    withheld?: boolean;
     /** M5: entries applied together (one consolidation proposal) share a group; undoing one undoes the group. */
     group?: string;
     /** M5: position the pre-image held in the durable list, so undo restores order exactly. */

package/dist/model-pricing.d.ts

@@ -1,19 +1,22 @@
 /**
- * Static, approximate per-model pricing — a fallback used only by the proxy backend's /cost
- * display. The proxy (packages/proxy/src/forward.ts) is a thin pass-through to Anthropic/
- * OpenAI and never computes or returns a dollar cost itself, only the raw token counts each
- * provider's response already includes. This table is NOT billing-accurate: it doesn't account
- * for prompt-caching discounts, batch pricing, or any rate change since it was last updated —
- * update the numbers below if Anthropic's published pricing changes. The Claude CLI backend
- * needs no such table; its costUsd comes from Claude's own real accounting (see
- * claude-cli-prompt.ts's parseClaudeCliOutput).
+ * Static, approximate per-model pricing — a fallback for every backend whose response carries
+ * token counts but no dollar figure. Anthropic and OpenAI responses report tokens only; OpenRouter
+ * (usage.cost) and the Claude CLI (total_cost_usd) report a real cost, which callers prefer over
+ * this table. This table is NOT billing-accurate: it ignores batch pricing, long-context tiers and
+ * any rate change since it was last updated — update the numbers below when published pricing changes.
  */
 export interface ModelPriceUsd {
     inputPerMTok: number;
     outputPerMTok: number;
+    /** Rate for input tokens served from the provider's prompt cache; absent = same as inputPerMTok. */
+    cachedInputPerMTok?: number;
 }
-/** Returns undefined (rather than a misleading number) when the model is unset or not a recognized Claude tier — callers must not silently treat "no estimate" as "free". */
+/**
+ * Returns undefined (rather than a misleading number) when the model is unset or not a recognized one — callers must not silently treat "no estimate" as "free".
+ * `cachedInputTokens` follows OpenAI's convention (a subset of inputTokens, billed at the cached rate); it is ignored for models with no cached rate.
+ */
 export declare function estimateCostUsd(model: string | undefined, usage: {
     inputTokens: number;
     outputTokens: number;
+    cachedInputTokens?: number;
 }): number | undefined;

package/dist/node-display-names.d.ts

@@ -20,7 +20,7 @@ export interface WhyChainItem {
  * Collapses a turn's raw layerActivity (one event per layer per main-loop iteration — see
  * harness-runtime.ts's onLayerActivity doc comment) down to just the layers that fired, in the
  * order they fired, merging consecutive re-fires of the same layer (back-to-back loop
- * iterations) into a single link. Shared by the CLI's `/why` and chat-ui's "Why?" panel so both
+ * iterations) into a single link. Used by run-explain.ts, which both the CLI's `/why` and chat-ui's `/why` print, so both
  * surfaces summarize a turn identically instead of drifting apart.
  */
 export declare function buildWhyChain(layerActivity: LayerActivityEvent[]): WhyChainItem[];

package/dist/project-command.d.ts

@@ -0,0 +1,16 @@
+/**
+ * `/project [name]` for every surface: no argument shows the active project, `clear` reverts to the workspace
+ * default, anything else switches to that name. Pure: the surface persists the config change and shows the message.
+ */
+export type ProjectCommand = {
+    kind: 'show';
+    message: string;
+} | {
+    kind: 'set';
+    project: string;
+    message: string;
+} | {
+    kind: 'clear';
+    message: string;
+};
+export declare function parseProjectCommand(args: readonly string[], currentProject: string | undefined): ProjectCommand;