@orquentra/codebuddy 2.14.0 → 2.15.1

package/dist/chunk-NNJSAYHY.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/llm/local.ts"],"sourcesContent":["import { createHash } from \"node:crypto\";\nimport { createWriteStream } from \"node:fs\";\nimport { mkdir, rename, rm, stat, statfs, utimes, writeFile } from \"node:fs/promises\";\nimport { availableParallelism, freemem, totalmem } from \"node:os\";\nimport { dirname, join } from \"node:path\";\nimport { Readable } from \"node:stream\";\nimport { pipeline } from \"node:stream/promises\";\nimport { z } from \"zod\";\nimport { modelCacheDir } from \"../providers/local.js\";\nimport type { ChatJsonRequest, LlmResult } from \"./client.js\";\n\n/**\n * Local text generation (Qwen3 only: 0.6B default, 1.7B \"better\") for the small-model features (intent cards, index\n * facts, query expansion, and optionally rerank): a GGUF model run by\n * node-llama-cpp (llama.cpp bindings; prebuilt for macOS arm64/x64, Linux\n * x64/arm64 and Windows x64; Metal/CUDA/Vulkan when present, CPU otherwise).\n * See ARCHITECTURE.md for why node-llama-cpp and not Cactus.\n *\n * Installed by setup (`codebuddy`, skipped with `--no-local-llm` or\n * `CODEBUDDY_LOCAL_LLM=0`) or `codebuddy models add local-llm`: the runtime with\n * npm, the weights checksum-verified into the shared model cache. Without\n * either, every caller gets `disabled` and keeps 2.14.0 behaviour. Nothing\n * here makes a network call except `downloadLocalModel`.\n *\n * Generation runs on llama.cpp's own native threads (capped at cores − 1),\n * so the MCP event loop keeps answering while a card is written; background\n * passes also yield to foreground work through `waitForForeground`.\n */\n\nexport type LocalModelSize = \"default\" | \"better\";\n\nexport type LocalModelSpec = {\n  size: LocalModelSize;\n  id: string;\n  repo: string;\n  revision: string;\n  file: string;\n  bytes: number;\n  sha256: string;\n  /** Rough resident memory while loaded, for the free-RAM check. */\n  ramBytes: number;\n  /** Qwen3 thinks unless told not to; the JSON grammar alone does not stop it. */\n  noThink: boolean;\n};\n\nconst GB = 1024 ** 3;\n\n/** Pinned revisions and checksums (Hugging Face LFS sha256). */\nexport const LOCAL_MODELS: Record<LocalModelSize, LocalModelSpec> = {\n  default: {\n    size: \"default\",\n    id: \"qwen3-0.6b-q4_k_m\",\n    repo: \"unsloth/Qwen3-0.6B-GGUF\",\n    revision: \"50968a4468ef4233ed78cd7c3de230dd1d61a56b\",\n    file: \"Qwen3-0.6B-Q4_K_M.gguf\",\n    bytes: 396_705_472,\n    sha256: \"ac2d97712095a558e31573f62f466a3f9d93990898b0ec79d7c974c1780d524a\",\n    ramBytes: 1 * GB,\n    noThink: true,\n  },\n  better: {\n    size: \"better\",\n    id: \"qwen3-1.7b-q4_k_m\",\n    repo: \"unsloth/Qwen3-1.7B-GGUF\",\n    revision: \"d7f544eead698dbd1f15126ef60b45a1e1933222\",\n    file: \"Qwen3-1.7B-Q4_K_M.gguf\",\n    bytes: 1_107_409_472,\n    sha256: \"b139949c5bd74937ad8ed8c8cf3d9ffb1e99c866c823204dc42c0d91fa181897\",\n    ramBytes: 2.2 * GB,\n    noThink: true,\n  },\n};\n\n/** Exact runtime version the adapter was tested against. */\nexport const LLAMA_RUNTIME = \"node-llama-cpp@3.22.3\";\n\nexport function localModelPath(spec: LocalModelSpec): string {\n  return join(modelCacheDir(), \"gguf\", spec.file);\n}\n\n/** The installed model of `preferred` size, else any installed one (default first). */\nexport async function installedLocalModel(\n  preferred?: LocalModelSize,\n): Promise<LocalModelSpec | null> {\n  if (process.env.CODEBUDDY_LOCAL_LLM === \"0\") return null;\n  const order: LocalModelSize[] = [...(preferred ? [preferred] : []), \"default\", \"better\"];\n  for (const size of new Set(order)) {\n    const spec = LOCAL_MODELS[size];\n    const info = await stat(localModelPath(spec)).catch(() => null);\n    if (info?.isFile() && info.size === spec.bytes) return spec;\n  }\n  return null;\n}\n\n/** Whether the node-llama-cpp runtime can be imported from this install. */\nexport async function localRuntimeAvailable(): Promise<boolean> {\n  try {\n    await importRuntime();\n    return true;\n  } catch {\n    return false;\n  }\n}\n\nexport type DownloadCheck = { ok: true } | { ok: false; reason: string };\n\n/**\n * Whether setup should install the model on its own. A low-end machine (under\n * 4 GB of RAM or a single core) would spend most of a session carding at a few\n * tokens per second; it keeps keyword + vector discovery, and\n * `codebuddy models add local-llm` still works there by hand.\n */\nexport function machineCanCard(): DownloadCheck {\n  if (totalmem() < 4 * GB) {\n    return { ok: false, reason: `${formatBytes(totalmem())} RAM (setup installs it from 4 GB)` };\n  }\n  if (availableParallelism() < 2) return { ok: false, reason: \"a single CPU core\" };\n  return { ok: true };\n}\n\n/** Free disk (model cache volume) and RAM before a download. */\nexport async function checkLocalModelResources(spec: LocalModelSpec): Promise<DownloadCheck> {\n  const directory = modelCacheDir();\n  await mkdir(directory, { recursive: true });\n  const fs = await statfs(directory).catch(() => null);\n  if (fs && fs.bavail * fs.bsize < spec.bytes * 1.1) {\n    return {\n      ok: false,\n      reason: `needs ${formatBytes(spec.bytes * 1.1)} free in ${directory}, ${formatBytes(fs.bavail * fs.bsize)} available`,\n    };\n  }\n  if (totalmem() < spec.ramBytes * 2) {\n    return {\n      ok: false,\n      reason: `needs about ${formatBytes(spec.ramBytes)} of RAM while loaded; this machine has ${formatBytes(totalmem())} in total`,\n    };\n  }\n  return { ok: true };\n}\n\n/**\n * Download the pinned GGUF into the model cache, verifying size and sha256\n * before it is moved into place (a partial or corrupt file is never used).\n */\nexport async function downloadLocalModel(\n  spec: LocalModelSpec,\n  onProgress?: (received: number, total: number) => void,\n  fetchImpl: typeof fetch = fetch,\n): Promise<string> {\n  const target = localModelPath(spec);\n  await mkdir(dirname(target), { recursive: true });\n  const partial = `${target}.partial`;\n  const url = `https://huggingface.co/${spec.repo}/resolve/${spec.revision}/${spec.file}`;\n  const response = await fetchImpl(url);\n  if (!response.ok || !response.body) {\n    throw new Error(`download failed: HTTP ${response.status} for ${url}`);\n  }\n  const hash = createHash(\"sha256\");\n  let received = 0;\n  const source = Readable.fromWeb(response.body as import(\"node:stream/web\").ReadableStream);\n  source.on(\"data\", (chunk: Buffer) => {\n    hash.update(chunk);\n    received += chunk.length;\n    onProgress?.(received, spec.bytes);\n  });\n  try {\n    await pipeline(source, createWriteStream(partial));\n    const digest = hash.digest(\"hex\");\n    if (received !== spec.bytes || digest !== spec.sha256) {\n      throw new Error(\n        `checksum mismatch for ${spec.file} (got ${received} bytes, sha256 ${digest.slice(0, 12)}…); not installed`,\n      );\n    }\n    await rename(partial, target);\n  } finally {\n    await rm(partial, { force: true });\n  }\n  return target;\n}\n\nexport async function removeLocalModels(): Promise<string[]> {\n  const removed: string[] = [];\n  for (const spec of Object.values(LOCAL_MODELS)) {\n    const path = localModelPath(spec);\n    if (await stat(path).catch(() => null)) {\n      await rm(path, { force: true });\n      removed.push(spec.id);\n    }\n  }\n  return removed;\n}\n\n// ── Generation ──────────────────────────────────────────────────────────\n\n/** Minimal structural types for the dynamically imported runtime. */\ntype Llama = {\n  loadModel(options: { modelPath: string }): Promise<LlamaModel>;\n  createGrammarForJsonSchema(schema: unknown): Promise<unknown>;\n  getGrammarFor(type: \"json\"): Promise<unknown>;\n};\ntype LlamaModel = {\n  createContext(options: { contextSize: number; threads: number }): Promise<LlamaContext>;\n  dispose(): Promise<void>;\n};\ntype LlamaContext = { getSequence(): unknown; dispose(): Promise<void> };\ntype ChatSession = {\n  prompt(\n    text: string,\n    options: {\n      grammar?: unknown;\n      maxTokens: number;\n      temperature: number;\n      budgets?: { thoughtTokens: number };\n      signal?: AbortSignal;\n      stopOnAbortSignal?: boolean;\n    },\n  ): Promise<string>;\n  resetChatHistory(): void;\n  setChatHistory?(history: unknown): void;\n  dispose(options?: { disposeSequence?: boolean }): void;\n};\ntype Runtime = {\n  getLlama(options?: { gpu?: \"auto\" | false }): Promise<Llama>;\n  LlamaChatSession: new (options: {\n    contextSequence: unknown;\n    autoDisposeSequence?: boolean;\n    systemPrompt?: string;\n  }) => ChatSession;\n};\n\ntype Loaded = {\n  spec: LocalModelSpec;\n  llama: Llama;\n  model: LlamaModel;\n  context: LlamaContext;\n  sequence: unknown;\n  runtime: Runtime;\n};\n\nlet loading: Promise<Loaded> | null = null;\nlet loaded: Loaded | null = null;\n/** One generation at a time: a single context sequence is shared. */\nlet queue: Promise<unknown> = Promise.resolve();\nconst grammars = new Map<unknown, Promise<unknown>>();\n\nasync function importRuntime(): Promise<Runtime> {\n  // A variable specifier: not a dependency, never bundled, resolved at run time.\n  const specifier = process.env.CODEBUDDY_LLAMA_RUNTIME ?? \"node-llama-cpp\";\n  return (await import(specifier)) as Runtime;\n}\n\n/** Context large enough for the bounded card/fact prompts (~1k tokens in, ≤ 400 out). */\nconst CONTEXT_TOKENS = 4096;\n\n/** Physical cores minus one (logical on SMT machines: a ceiling, not an exact count). */\nexport function generationThreads(): number {\n  const configured = Number(process.env.CODEBUDDY_LLM_THREADS);\n  if (Number.isInteger(configured) && configured > 0) return configured;\n  return Math.max(1, availableParallelism() - 1);\n}\n\nasync function load(spec: LocalModelSpec): Promise<Loaded> {\n  if (loaded?.spec.id === spec.id) return loaded;\n  if (!loading) {\n    loading = (async () => {\n      const runtime = await importRuntime();\n      const llama = await runtime.getLlama({ gpu: \"auto\" });\n      const model = await llama.loadModel({ modelPath: localModelPath(spec) });\n      const context = await model.createContext({\n        contextSize: CONTEXT_TOKENS,\n        threads: generationThreads(),\n      });\n      loaded = { spec, llama, model, context, sequence: context.getSequence(), runtime };\n      return loaded;\n    })().finally(() => {\n      loading = null;\n    });\n  }\n  return loading;\n}\n\n/**\n * The model leaves memory this long after its last generation, so a finished\n * card pass does not keep ~1 GB resident in `codebuddy serve` for the rest of\n * the session. The next pass (or a warm-up) loads it again in under a second.\n */\nexport const IDLE_UNLOAD_MS = 60_000;\nlet idleTimer: NodeJS.Timeout | null = null;\nlet running = 0;\n\nfunction scheduleUnload(): void {\n  if (idleTimer) clearTimeout(idleTimer);\n  idleTimer = setTimeout(() => void unloadLocalModel(), IDLE_UNLOAD_MS);\n  // Never keeps the process alive on its own.\n  idleTimer.unref();\n}\n\n/** Free the model and its context now (idle timer, `serve` shutdown). Never throws. */\nexport async function unloadLocalModel(): Promise<void> {\n  if (idleTimer) clearTimeout(idleTimer);\n  idleTimer = null;\n  const current = loaded;\n  if (!current || running > 0) return;\n  loaded = null;\n  grammars.clear();\n  await current.context.dispose().catch(() => {});\n  await current.model.dispose().catch(() => {});\n}\n\n/** True once the model is resident in this process (query expansion only runs then). */\nexport function isLocalModelWarm(): boolean {\n  return loaded !== null;\n}\n\n/** Load the model now (background warm-up); resolves false when it cannot load. */\nexport async function warmLocalModel(spec: LocalModelSpec): Promise<boolean> {\n  try {\n    await load(spec);\n    return true;\n  } catch {\n    return false;\n  }\n}\n\n/**\n * `chatJson` for the local backend: JSON constrained by a grammar built from\n * the request's zod schema (plain JSON grammar if the schema cannot be\n * converted), then validated with zod. Never throws.\n */\nexport async function localChatJson<T>(\n  spec: LocalModelSpec,\n  request: ChatJsonRequest<T>,\n): Promise<LlmResult<T>> {\n  const started = Date.now();\n  const fail = (reason: \"timeout\" | \"invalid\" | \"disabled\", detail: string): LlmResult<T> => ({\n    ok: false,\n    reason,\n    detail: detail.slice(0, 300),\n    latencyMs: Date.now() - started,\n  });\n  let model: Loaded;\n  try {\n    model = await withDeadline(load(spec), request.deadlineMs);\n  } catch (error) {\n    return (error as Error).name === \"TimeoutError\"\n      ? fail(\"timeout\", \"model still loading\")\n      : fail(\"disabled\", `local model unavailable: ${(error as Error).message}`);\n  }\n  const run = async (): Promise<LlmResult<T>> => {\n    running++;\n    try {\n      return await generate();\n    } finally {\n      running--;\n      scheduleUnload();\n    }\n  };\n  const generate = async (): Promise<LlmResult<T>> => {\n    // The idle unload may have run while this call waited in the queue.\n    if (!loaded) model = await load(spec);\n    const remaining = request.deadlineMs - (Date.now() - started);\n    if (remaining <= 0) return fail(\"timeout\", `no answer within ${request.deadlineMs} ms`);\n    const grammar = await grammarFor(model.llama, request.jsonSchema ?? request.schema);\n    const session = new model.runtime.LlamaChatSession({\n      contextSequence: model.sequence,\n      autoDisposeSequence: false,\n      systemPrompt: request.system,\n    });\n    const controller = new AbortController();\n    const timer = setTimeout(() => controller.abort(), remaining);\n    try {\n      const text = await session.prompt(`${spec.noThink ? \"/no_think\\n\" : \"\"}${request.user}`, {\n        grammar,\n        maxTokens: request.maxTokens,\n        temperature: 0,\n        ...(spec.noThink ? { budgets: { thoughtTokens: 0 } } : {}),\n        signal: controller.signal,\n        stopOnAbortSignal: true,\n      });\n      if (controller.signal.aborted) {\n        return fail(\"timeout\", `no answer within ${request.deadlineMs} ms`);\n      }\n      let value: unknown;\n      try {\n        value = JSON.parse(text);\n      } catch {\n        return fail(\"invalid\", `output is not JSON: ${text.slice(0, 120)}`);\n      }\n      const parsed = request.schema.safeParse(value);\n      if (!parsed.success)\n        return fail(\"invalid\", `output failed validation: ${text.slice(0, 120)}`);\n      return {\n        ok: true,\n        value: parsed.data,\n        model: spec.id,\n        latencyMs: Date.now() - started,\n        tokens: Math.ceil((request.system.length + request.user.length + text.length) / 4),\n        cached: false,\n      };\n    } catch (error) {\n      return controller.signal.aborted\n        ? fail(\"timeout\", `no answer within ${request.deadlineMs} ms`)\n        : fail(\"invalid\", String((error as Error)?.message ?? error));\n    } finally {\n      clearTimeout(timer);\n      session.dispose({ disposeSequence: false });\n    }\n  };\n  const result = queue.then(run, run);\n  queue = result.catch(() => undefined);\n  return result;\n}\n\nasync function grammarFor(llama: Llama, schema: z.ZodType | object): Promise<unknown> {\n  let grammar = grammars.get(schema);\n  if (!grammar) {\n    grammar = (async () => {\n      try {\n        const json =\n          schema instanceof z.ZodType\n            ? z.toJSONSchema(schema, { io: \"input\", unrepresentable: \"any\" })\n            : schema;\n        return await llama.createGrammarForJsonSchema(stripSchemaMeta(json));\n      } catch {\n        return llama.getGrammarFor(\"json\");\n      }\n    })();\n    grammars.set(schema, grammar);\n  }\n  return grammar;\n}\n\n/** llama.cpp's schema grammar wants plain structure: no `$schema`, no defaults. */\nfunction stripSchemaMeta(value: unknown): unknown {\n  if (Array.isArray(value)) return value.map(stripSchemaMeta);\n  if (!value || typeof value !== \"object\") return value;\n  const out: Record<string, unknown> = {};\n  for (const [key, child] of Object.entries(value)) {\n    if (key === \"$schema\" || key === \"default\" || key === \"additionalProperties\") continue;\n    out[key] = stripSchemaMeta(child);\n  }\n  return out;\n}\n\nfunction withDeadline<T>(promise: Promise<T>, ms: number): Promise<T> {\n  return new Promise((resolve, reject) => {\n    const timer = setTimeout(\n      () => reject(Object.assign(new Error(\"deadline\"), { name: \"TimeoutError\" })),\n      ms,\n    );\n    promise.then(\n      (value) => {\n        clearTimeout(timer);\n        resolve(value);\n      },\n      (error) => {\n        clearTimeout(timer);\n        reject(error);\n      },\n    );\n  });\n}\n\n// ── Foreground yield ────────────────────────────────────────────────────\n\n/** A prompt hook or MCP tool call touched this within the window: background work waits. */\nconst FOREGROUND_WINDOW_MS = 2_000;\nconst FOREGROUND_MAX_WAIT_MS = 15_000;\n\nfunction foregroundPath(root: string): string {\n  return join(root, \".codebuddy\", \"cache\", \"foreground\");\n}\n\n/** Record that a hook or tool call is running (cheap: one utimes). Never throws. */\nexport async function markForeground(root: string): Promise<void> {\n  const path = foregroundPath(root);\n  const now = new Date();\n  try {\n    await utimes(path, now, now);\n  } catch {\n    await mkdir(dirname(path), { recursive: true }).catch(() => {});\n    await writeFile(path, \"\").catch(() => {});\n  }\n}\n\n/** Wait (bounded) until no foreground work was marked within the window. */\nexport async function waitForForeground(root: string, signal?: AbortSignal): Promise<void> {\n  const until = Date.now() + FOREGROUND_MAX_WAIT_MS;\n  while (Date.now() < until && !signal?.aborted) {\n    const info = await stat(foregroundPath(root)).catch(() => null);\n    if (!info || Date.now() - info.mtimeMs > FOREGROUND_WINDOW_MS) return;\n    await new Promise((resolve) => setTimeout(resolve, 250));\n  }\n}\n\n// ── Status ──────────────────────────────────────────────────────────────\n\nexport type LocalModelStatus = {\n  runtime: boolean;\n  installed: LocalModelSpec | null;\n  path: string | null;\n};\n\nexport async function localModelStatus(preferred?: LocalModelSize): Promise<LocalModelStatus> {\n  const installed = await installedLocalModel(preferred);\n  return {\n    runtime: await localRuntimeAvailable(),\n    installed,\n    path: installed ? localModelPath(installed) : null,\n  };\n}\n\nexport function formatBytes(bytes: number): string {\n  return bytes >= GB ? `${(bytes / GB).toFixed(1)} GB` : `${Math.round(bytes / 1024 ** 2)} MB`;\n}\n\n/** Free RAM now (for the `models add` warning; the hard check uses total RAM). */\nexport function freeMemory(): number {\n  return freemem();\n}\n"],"mappings":";;;;;AAAA,SAAS,kBAAkB;AAC3B,SAAS,yBAAyB;AAClC,SAAS,OAAO,QAAQ,IAAI,MAAM,QAAQ,QAAQ,iBAAiB;AACnE,SAAS,sBAAsB,SAAS,gBAAgB;AACxD,SAAS,SAAS,YAAY;AAC9B,SAAS,gBAAgB;AACzB,SAAS,gBAAgB;AACzB,SAAS,SAAS;AAsClB,IAAM,KAAK,QAAQ;AAGZ,IAAM,eAAuD;AAAA,EAClE,SAAS;AAAA,IACP,MAAM;AAAA,IACN,IAAI;AAAA,IACJ,MAAM;AAAA,IACN,UAAU;AAAA,IACV,MAAM;AAAA,IACN,OAAO;AAAA,IACP,QAAQ;AAAA,IACR,UAAU,IAAI;AAAA,IACd,SAAS;AAAA,EACX;AAAA,EACA,QAAQ;AAAA,IACN,MAAM;AAAA,IACN,IAAI;AAAA,IACJ,MAAM;AAAA,IACN,UAAU;AAAA,IACV,MAAM;AAAA,IACN,OAAO;AAAA,IACP,QAAQ;AAAA,IACR,UAAU,MAAM;AAAA,IAChB,SAAS;AAAA,EACX;AACF;AAGO,IAAM,gBAAgB;AAEtB,SAAS,eAAe,MAA8B;AAC3D,SAAO,KAAK,cAAc,GAAG,QAAQ,KAAK,IAAI;AAChD;AAGA,eAAsB,oBACpB,WACgC;AAChC,MAAI,QAAQ,IAAI,wBAAwB,IAAK,QAAO;AACpD,QAAM,QAA0B,CAAC,GAAI,YAAY,CAAC,SAAS,IAAI,CAAC,GAAI,WAAW,QAAQ;AACvF,aAAW,QAAQ,IAAI,IAAI,KAAK,GAAG;AACjC,UAAM,OAAO,aAAa,IAAI;AAC9B,UAAM,OAAO,MAAM,KAAK,eAAe,IAAI,CAAC,EAAE,MAAM,MAAM,IAAI;AAC9D,QAAI,MAAM,OAAO,KAAK,KAAK,SAAS,KAAK,MAAO,QAAO;AAAA,EACzD;AACA,SAAO;AACT;AAGA,eAAsB,wBAA0C;AAC9D,MAAI;AACF,UAAM,cAAc;AACpB,WAAO;AAAA,EACT,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAUO,SAAS,iBAAgC;AAC9C,MAAI,SAAS,IAAI,IAAI,IAAI;AACvB,WAAO,EAAE,IAAI,OAAO,QAAQ,GAAG,YAAY,SAAS,CAAC,CAAC,qCAAqC;AAAA,EAC7F;AACA,MAAI,qBAAqB,IAAI,EAAG,QAAO,EAAE,IAAI,OAAO,QAAQ,oBAAoB;AAChF,SAAO,EAAE,IAAI,KAAK;AACpB;AAGA,eAAsB,yBAAyB,MAA8C;AAC3F,QAAM,YAAY,cAAc;AAChC,QAAM,MAAM,WAAW,EAAE,WAAW,KAAK,CAAC;AAC1C,QAAM,KAAK,MAAM,OAAO,SAAS,EAAE,MAAM,MAAM,IAAI;AACnD,MAAI,MAAM,GAAG,SAAS,GAAG,QAAQ,KAAK,QAAQ,KAAK;AACjD,WAAO;AAAA,MACL,IAAI;AAAA,MACJ,QAAQ,SAAS,YAAY,KAAK,QAAQ,GAAG,CAAC,YAAY,SAAS,KAAK,YAAY,GAAG,SAAS,GAAG,KAAK,CAAC;AAAA,IAC3G;AAAA,EACF;AACA,MAAI,SAAS,IAAI,KAAK,WAAW,GAAG;AAClC,WAAO;AAAA,MACL,IAAI;AAAA,MACJ,QAAQ,eAAe,YAAY,KAAK,QAAQ,CAAC,0CAA0C,YAAY,SAAS,CAAC,CAAC;AAAA,IACpH;AAAA,EACF;AACA,SAAO,EAAE,IAAI,KAAK;AACpB;AAMA,eAAsB,mBACpB,MACA,YACA,YAA0B,OACT;AACjB,QAAM,SAAS,eAAe,IAAI;AAClC,QAAM,MAAM,QAAQ,MAAM,GAAG,EAAE,WAAW,KAAK,CAAC;AAChD,QAAM,UAAU,GAAG,MAAM;AACzB,QAAM,MAAM,0BAA0B,KAAK,IAAI,YAAY,KAAK,QAAQ,IAAI,KAAK,IAAI;AACrF,QAAM,WAAW,MAAM,UAAU,GAAG;AACpC,MAAI,CAAC,SAAS,MAAM,CAAC,SAAS,MAAM;AAClC,UAAM,IAAI,MAAM,yBAAyB,SAAS,MAAM,QAAQ,GAAG,EAAE;AAAA,EACvE;AACA,QAAM,OAAO,WAAW,QAAQ;AAChC,MAAI,WAAW;AACf,QAAM,SAAS,SAAS,QAAQ,SAAS,IAAgD;AACzF,SAAO,GAAG,QAAQ,CAAC,UAAkB;AACnC,SAAK,OAAO,KAAK;AACjB,gBAAY,MAAM;AAClB,iBAAa,UAAU,KAAK,KAAK;AAAA,EACnC,CAAC;AACD,MAAI;AACF,UAAM,SAAS,QAAQ,kBAAkB,OAAO,CAAC;AACjD,UAAM,SAAS,KAAK,OAAO,KAAK;AAChC,QAAI,aAAa,KAAK,SAAS,WAAW,KAAK,QAAQ;AACrD,YAAM,IAAI;AAAA,QACR,yBAAyB,KAAK,IAAI,SAAS,QAAQ,kBAAkB,OAAO,MAAM,GAAG,EAAE,CAAC;AAAA,MAC1F;AAAA,IACF;AACA,UAAM,OAAO,SAAS,MAAM;AAAA,EAC9B,UAAE;AACA,UAAM,GAAG,SAAS,EAAE,OAAO,KAAK,CAAC;AAAA,EACnC;AACA,SAAO;AACT;AAEA,eAAsB,oBAAuC;AAC3D,QAAM,UAAoB,CAAC;AAC3B,aAAW,QAAQ,OAAO,OAAO,YAAY,GAAG;AAC9C,UAAM,OAAO,eAAe,IAAI;AAChC,QAAI,MAAM,KAAK,IAAI,EAAE,MAAM,MAAM,IAAI,GAAG;AACtC,YAAM,GAAG,MAAM,EAAE,OAAO,KAAK,CAAC;AAC9B,cAAQ,KAAK,KAAK,EAAE;AAAA,IACtB;AAAA,EACF;AACA,SAAO;AACT;AAiDA,IAAI,UAAkC;AACtC,IAAI,SAAwB;AAE5B,IAAI,QAA0B,QAAQ,QAAQ;AAC9C,IAAM,WAAW,oBAAI,IAA+B;AAEpD,eAAe,gBAAkC;AAE/C,QAAM,YAAY,QAAQ,IAAI,2BAA2B;AACzD,SAAQ,MAAM,OAAO;AACvB;AAGA,IAAM,iBAAiB;AAGhB,SAAS,oBAA4B;AAC1C,QAAM,aAAa,OAAO,QAAQ,IAAI,qBAAqB;AAC3D,MAAI,OAAO,UAAU,UAAU,KAAK,aAAa,EAAG,QAAO;AAC3D,SAAO,KAAK,IAAI,GAAG,qBAAqB,IAAI,CAAC;AAC/C;AAEA,eAAe,KAAK,MAAuC;AACzD,MAAI,QAAQ,KAAK,OAAO,KAAK,GAAI,QAAO;AACxC,MAAI,CAAC,SAAS;AACZ,eAAW,YAAY;AACrB,YAAM,UAAU,MAAM,cAAc;AACpC,YAAM,QAAQ,MAAM,QAAQ,SAAS,EAAE,KAAK,OAAO,CAAC;AACpD,YAAM,QAAQ,MAAM,MAAM,UAAU,EAAE,WAAW,eAAe,IAAI,EAAE,CAAC;AACvE,YAAM,UAAU,MAAM,MAAM,cAAc;AAAA,QACxC,aAAa;AAAA,QACb,SAAS,kBAAkB;AAAA,MAC7B,CAAC;AACD,eAAS,EAAE,MAAM,OAAO,OAAO,SAAS,UAAU,QAAQ,YAAY,GAAG,QAAQ;AACjF,aAAO;AAAA,IACT,GAAG,EAAE,QAAQ,MAAM;AACjB,gBAAU;AAAA,IACZ,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAOO,IAAM,iBAAiB;AAC9B,IAAI,YAAmC;AACvC,IAAI,UAAU;AAEd,SAAS,iBAAuB;AAC9B,MAAI,UAAW,cAAa,SAAS;AACrC,cAAY,WAAW,MAAM,KAAK,iBAAiB,GAAG,cAAc;AAEpE,YAAU,MAAM;AAClB;AAGA,eAAsB,mBAAkC;AACtD,MAAI,UAAW,cAAa,SAAS;AACrC,cAAY;AACZ,QAAM,UAAU;AAChB,MAAI,CAAC,WAAW,UAAU,EAAG;AAC7B,WAAS;AACT,WAAS,MAAM;AACf,QAAM,QAAQ,QAAQ,QAAQ,EAAE,MAAM,MAAM;AAAA,EAAC,CAAC;AAC9C,QAAM,QAAQ,MAAM,QAAQ,EAAE,MAAM,MAAM;AAAA,EAAC,CAAC;AAC9C;AAGO,SAAS,mBAA4B;AAC1C,SAAO,WAAW;AACpB;AAGA,eAAsB,eAAe,MAAwC;AAC3E,MAAI;AACF,UAAM,KAAK,IAAI;AACf,WAAO;AAAA,EACT,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAOA,eAAsB,cACpB,MACA,SACuB;AACvB,QAAM,UAAU,KAAK,IAAI;AACzB,QAAM,OAAO,CAAC,QAA4C,YAAkC;AAAA,IAC1F,IAAI;AAAA,IACJ;AAAA,IACA,QAAQ,OAAO,MAAM,GAAG,GAAG;AAAA,IAC3B,WAAW,KAAK,IAAI,IAAI;AAAA,EAC1B;AACA,MAAI;AACJ,MAAI;AACF,YAAQ,MAAM,aAAa,KAAK,IAAI,GAAG,QAAQ,UAAU;AAAA,EAC3D,SAAS,OAAO;AACd,WAAQ,MAAgB,SAAS,iBAC7B,KAAK,WAAW,qBAAqB,IACrC,KAAK,YAAY,4BAA6B,MAAgB,OAAO,EAAE;AAAA,EAC7E;AACA,QAAM,MAAM,YAAmC;AAC7C;AACA,QAAI;AACF,aAAO,MAAM,SAAS;AAAA,IACxB,UAAE;AACA;AACA,qBAAe;AAAA,IACjB;AAAA,EACF;AACA,QAAM,WAAW,YAAmC;AAElD,QAAI,CAAC,OAAQ,SAAQ,MAAM,KAAK,IAAI;AACpC,UAAM,YAAY,QAAQ,cAAc,KAAK,IAAI,IAAI;AACrD,QAAI,aAAa,EAAG,QAAO,KAAK,WAAW,oBAAoB,QAAQ,UAAU,KAAK;AACtF,UAAM,UAAU,MAAM,WAAW,MAAM,OAAO,QAAQ,cAAc,QAAQ,MAAM;AAClF,UAAM,UAAU,IAAI,MAAM,QAAQ,iBAAiB;AAAA,MACjD,iBAAiB,MAAM;AAAA,MACvB,qBAAqB;AAAA,MACrB,cAAc,QAAQ;AAAA,IACxB,CAAC;AACD,UAAM,aAAa,IAAI,gBAAgB;AACvC,UAAM,QAAQ,WAAW,MAAM,WAAW,MAAM,GAAG,SAAS;AAC5D,QAAI;AACF,YAAM,OAAO,MAAM,QAAQ,OAAO,GAAG,KAAK,UAAU,gBAAgB,EAAE,GAAG,QAAQ,IAAI,IAAI;AAAA,QACvF;AAAA,QACA,WAAW,QAAQ;AAAA,QACnB,aAAa;AAAA,QACb,GAAI,KAAK,UAAU,EAAE,SAAS,EAAE,eAAe,EAAE,EAAE,IAAI,CAAC;AAAA,QACxD,QAAQ,WAAW;AAAA,QACnB,mBAAmB;AAAA,MACrB,CAAC;AACD,UAAI,WAAW,OAAO,SAAS;AAC7B,eAAO,KAAK,WAAW,oBAAoB,QAAQ,UAAU,KAAK;AAAA,MACpE;AACA,UAAI;AACJ,UAAI;AACF,gBAAQ,KAAK,MAAM,IAAI;AAAA,MACzB,QAAQ;AACN,eAAO,KAAK,WAAW,uBAAuB,KAAK,MAAM,GAAG,GAAG,CAAC,EAAE;AAAA,MACpE;AACA,YAAM,SAAS,QAAQ,OAAO,UAAU,KAAK;AAC7C,UAAI,CAAC,OAAO;AACV,eAAO,KAAK,WAAW,6BAA6B,KAAK,MAAM,GAAG,GAAG,CAAC,EAAE;AAC1E,aAAO;AAAA,QACL,IAAI;AAAA,QACJ,OAAO,OAAO;AAAA,QACd,OAAO,KAAK;AAAA,QACZ,WAAW,KAAK,IAAI,IAAI;AAAA,QACxB,QAAQ,KAAK,MAAM,QAAQ,OAAO,SAAS,QAAQ,KAAK,SAAS,KAAK,UAAU,CAAC;AAAA,QACjF,QAAQ;AAAA,MACV;AAAA,IACF,SAAS,OAAO;AACd,aAAO,WAAW,OAAO,UACrB,KAAK,WAAW,oBAAoB,QAAQ,UAAU,KAAK,IAC3D,KAAK,WAAW,OAAQ,OAAiB,WAAW,KAAK,CAAC;AAAA,IAChE,UAAE;AACA,mBAAa,KAAK;AAClB,cAAQ,QAAQ,EAAE,iBAAiB,MAAM,CAAC;AAAA,IAC5C;AAAA,EACF;AACA,QAAM,SAAS,MAAM,KAAK,KAAK,GAAG;AAClC,UAAQ,OAAO,MAAM,MAAM,MAAS;AACpC,SAAO;AACT;AAEA,eAAe,WAAW,OAAc,QAA8C;AACpF,MAAI,UAAU,SAAS,IAAI,MAAM;AACjC,MAAI,CAAC,SAAS;AACZ,eAAW,YAAY;AACrB,UAAI;AACF,cAAM,OACJ,kBAAkB,EAAE,UAChB,EAAE,aAAa,QAAQ,EAAE,IAAI,SAAS,iBAAiB,MAAM,CAAC,IAC9D;AACN,eAAO,MAAM,MAAM,2BAA2B,gBAAgB,IAAI,CAAC;AAAA,MACrE,QAAQ;AACN,eAAO,MAAM,cAAc,MAAM;AAAA,MACnC;AAAA,IACF,GAAG;AACH,aAAS,IAAI,QAAQ,OAAO;AAAA,EAC9B;AACA,SAAO;AACT;AAGA,SAAS,gBAAgB,OAAyB;AAChD,MAAI,MAAM,QAAQ,KAAK,EAAG,QAAO,MAAM,IAAI,eAAe;AAC1D,MAAI,CAAC,SAAS,OAAO,UAAU,SAAU,QAAO;AAChD,QAAM,MAA+B,CAAC;AACtC,aAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,KAAK,GAAG;AAChD,QAAI,QAAQ,aAAa,QAAQ,aAAa,QAAQ,uBAAwB;AAC9E,QAAI,GAAG,IAAI,gBAAgB,KAAK;AAAA,EAClC;AACA,SAAO;AACT;AAEA,SAAS,aAAgB,SAAqB,IAAwB;AACpE,SAAO,IAAI,QAAQ,CAAC,SAAS,WAAW;AACtC,UAAM,QAAQ;AAAA,MACZ,MAAM,OAAO,OAAO,OAAO,IAAI,MAAM,UAAU,GAAG,EAAE,MAAM,eAAe,CAAC,CAAC;AAAA,MAC3E;AAAA,IACF;AACA,YAAQ;AAAA,MACN,CAAC,UAAU;AACT,qBAAa,KAAK;AAClB,gBAAQ,KAAK;AAAA,MACf;AAAA,MACA,CAAC,UAAU;AACT,qBAAa,KAAK;AAClB,eAAO,KAAK;AAAA,MACd;AAAA,IACF;AAAA,EACF,CAAC;AACH;AAKA,IAAM,uBAAuB;AAC7B,IAAM,yBAAyB;AAE/B,SAAS,eAAe,MAAsB;AAC5C,SAAO,KAAK,MAAM,cAAc,SAAS,YAAY;AACvD;AAGA,eAAsB,eAAe,MAA6B;AAChE,QAAM,OAAO,eAAe,IAAI;AAChC,QAAM,MAAM,oBAAI,KAAK;AACrB,MAAI;AACF,UAAM,OAAO,MAAM,KAAK,GAAG;AAAA,EAC7B,QAAQ;AACN,UAAM,MAAM,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC,EAAE,MAAM,MAAM;AAAA,IAAC,CAAC;AAC9D,UAAM,UAAU,MAAM,EAAE,EAAE,MAAM,MAAM;AAAA,IAAC,CAAC;AAAA,EAC1C;AACF;AAGA,eAAsB,kBAAkB,MAAc,QAAqC;AACzF,QAAM,QAAQ,KAAK,IAAI,IAAI;AAC3B,SAAO,KAAK,IAAI,IAAI,SAAS,CAAC,QAAQ,SAAS;AAC7C,UAAM,OAAO,MAAM,KAAK,eAAe,IAAI,CAAC,EAAE,MAAM,MAAM,IAAI;AAC9D,QAAI,CAAC,QAAQ,KAAK,IAAI,IAAI,KAAK,UAAU,qBAAsB;AAC/D,UAAM,IAAI,QAAQ,CAAC,YAAY,WAAW,SAAS,GAAG,CAAC;AAAA,EACzD;AACF;AAUA,eAAsB,iBAAiB,WAAuD;AAC5F,QAAM,YAAY,MAAM,oBAAoB,SAAS;AACrD,SAAO;AAAA,IACL,SAAS,MAAM,sBAAsB;AAAA,IACrC;AAAA,IACA,MAAM,YAAY,eAAe,SAAS,IAAI;AAAA,EAChD;AACF;AAEO,SAAS,YAAY,OAAuB;AACjD,SAAO,SAAS,KAAK,IAAI,QAAQ,IAAI,QAAQ,CAAC,CAAC,QAAQ,GAAG,KAAK,MAAM,QAAQ,QAAQ,CAAC,CAAC;AACzF;AAGO,SAAS,aAAqB;AACnC,SAAO,QAAQ;AACjB;","names":[]}

package/dist/{chunk-XSNW6Y3X.js → chunk-OO7Q7XZF.js}

@@ -6,7 +6,7 @@ import {
   detectTestCommand,
   estimateTokens,
   findCoveringTest
-} from "./chunk-4DKYKDHI.js";
+} from "./chunk-7IPSDDO5.js";
 import {
   hookRoot
 } from "./chunk-TGFTKWOR.js";
@@ -17,17 +17,6 @@ import {
   MEMORY_CLASSES,
   memoryCandidateSchema
 } from "./chunk-DFV2SK6C.js";
-import {
-  INCIDENT_SEVERITIES,
-  MemoryFileStore,
-  admitsLexically,
-  isEligibleMemory,
-  lexicalTerms,
-  termOverlap
-} from "./chunk-7QMHIJBA.js";
-import {
-  generateEntityId
-} from "./chunk-CAEPHLHC.js";
 import {
   NON_SOURCE_PATH,
   TASK_PATH_LIMIT,
@@ -42,11 +31,22 @@ import {
   readBoundedSource,
   syncFileVectors,
   taskTerms
-} from "./chunk-YNSTIMCF.js";
+} from "./chunk-4DTO2E4Y.js";
 import {
   extractSymbolTable,
   supportsSymbols
 } from "./chunk-Y7D7PUBE.js";
+import {
+  INCIDENT_SEVERITIES,
+  MemoryFileStore,
+  admitsLexically,
+  isEligibleMemory,
+  lexicalTerms,
+  termOverlap
+} from "./chunk-VBDM7HC5.js";
+import {
+  generateEntityId
+} from "./chunk-EMWGLKPR.js";
 import {
   REDACTED_LINE,
   claimText,
@@ -58,21 +58,21 @@ import {
   scanFields,
   scanSensitive
 } from "./chunk-AL6VBTNA.js";
-import {
-  DEFAULT_RERANK_MODELS,
-  LOCAL_EMBEDDING_MODEL
-} from "./chunk-HPOR6YXA.js";
 import {
   loadPlanPolicy,
   loadRuntimeConfig,
   readConfigFile
-} from "./chunk-KNR2P2UY.js";
+} from "./chunk-Y4QQML24.js";
 import {
   assertNoSymlinkBetween,
   composeMarkdown,
   parseFrontMatter,
   writeAtomic
 } from "./chunk-CG5B3LYG.js";
+import {
+  DEFAULT_RERANK_MODELS,
+  LOCAL_EMBEDDING_MODEL
+} from "./chunk-HPOR6YXA.js";
 
 // src/hooks/runtime.ts
 import { mkdir as mkdir7, readFile as readFile10, rm as rm5, stat as stat4, writeFile } from "fs/promises";
@@ -1474,7 +1474,7 @@ async function buildBeforeEditContext(input) {
     ...hasExplicitTargets ? { useGit: false } : input.useGit !== void 0 ? { useGit: input.useGit } : {}
   });
   const mapPromise = loadProjectArchitectureMap(repositoryRoot, manifest);
-  const discovery = !hasExplicitTargets && input.task ? await discoverTaskPaths(repositoryRoot, input.task, manifest, mapPromise) : { paths: [], hits: [], explain: [] };
+  const discovery = !hasExplicitTargets && input.task ? await discoverTaskPaths(repositoryRoot, input.task, manifest, mapPromise) : { paths: [], hits: [], explain: [], confidence: void 0 };
   const taskPaths = discovery.paths;
   const askedTask = Boolean(input.task?.trim());
   const targetPaths = hasExplicitTargets ? resolvedPaths : taskPaths.length > 0 ? taskPaths : askedTask && input.useGit !== true ? [] : resolvedPaths;
@@ -1592,11 +1592,15 @@ async function buildBeforeEditContext(input) {
     tests,
     testCommand,
     diffs,
-    explain
+    explain,
+    ...discovery.confidence && !hasExplicitTargets ? { confidence: discovery.confidence } : {},
+    // A plain abstention: no guessed files, and the agent is told to search itself.
+    ...askedTask && !hasExplicitTargets && targetPaths.length === 0 ? { note: NO_MATCH_NOTE } : {}
   };
   const tokens = buildTokenStats(assembled, input.tokenBudget);
   return enforceContextBudget({ ...assembled, tokens });
 }
+var NO_MATCH_NOTE = "No confident file match for this task; search the repository normally (grep/glob) instead of relying on CodeBuddy targets.";
 async function recallHandoffs(repositoryRoot, input) {
   try {
     return await new WorklogStore(repositoryRoot).recall(input);
@@ -3846,7 +3850,7 @@ async function openHookSemantics(repositoryRoot) {
   const connect = async () => {
     const config = await loadRuntimeConfig(repositoryRoot, { requireDatabase: false });
     if (!config.postgresUrl) return null;
-    const { createRuntime } = await import("./operations-6Z6ZDVBP.js");
+    const { createRuntime } = await import("./operations-BPI633VN.js");
     const opening = createRuntime(config, {
       repositoryRoot,
       startWorker: false,
@@ -4176,7 +4180,7 @@ async function runStageHook(input) {
 }
 async function llmMemoryOptions(repositoryRoot) {
   try {
-    return await (await import("./memory-TGHVTBME.js")).llmCaptureOptions(repositoryRoot);
+    return await (await import("./memory-LSZKSF53.js")).llmCaptureOptions(repositoryRoot);
   } catch {
     return {};
   }
@@ -4271,7 +4275,7 @@ async function runSyncHook(repositoryRoot, options = {}) {
     await writeFile(dirty, "").catch(() => void 0);
     return null;
   }
-  const { createRuntime } = await import("./operations-6Z6ZDVBP.js");
+  const { createRuntime } = await import("./operations-BPI633VN.js");
   const runtime = await createRuntime(config, { repositoryRoot }).catch(async (error) => {
     await rm5(lock, { recursive: true, force: true });
     throw error;
@@ -4729,4 +4733,4 @@ export {
   readCaptureTurn,
   renderContextBlock
 };
-//# sourceMappingURL=chunk-XSNW6Y3X.js.map
+//# sourceMappingURL=chunk-OO7Q7XZF.js.map