@theokit/sdk 5.10.0 → 5.11.0

package/dist/{chunk-EXI7EHDN.js → chunk-7Y23B7XX.js}

@@ -46,7 +46,7 @@ function persistTurnToTranscript(store, loc, sessionId, turn, onCompact) {
       const count = (recordCounts.get(key) ?? 0) + 1;
       recordCounts.set(key, count);
       if (turn.autoCompact !== void 0) {
-        const { autoCompactIfNeeded } = await import('./compact-session-PVAWTKKL.js');
+        const { autoCompactIfNeeded } = await import('./compact-session-F2UDGZCN.js');
         const fired = await autoCompactIfNeeded({
           store,
           loc,
@@ -105,5 +105,5 @@ function enqueueSessionWrite(cwd, agentId, fn) {
 }
 
 export { appendSessionMessage, clearSession, discardSession, enqueueSessionWrite, flushSessionWrites, getSessionMessages, hydrateSession, persistTurnToTranscript };
-//# sourceMappingURL=chunk-EXI7EHDN.js.map
-//# sourceMappingURL=chunk-EXI7EHDN.js.map
+//# sourceMappingURL=chunk-7Y23B7XX.js.map
+//# sourceMappingURL=chunk-7Y23B7XX.js.map

package/dist/{chunk-EXI7EHDN.js.map → chunk-7Y23B7XX.js.map}

@@ -1 +1 @@
-{"version":3,"sources":["../src/internal/session/agent-session.ts"],"names":[],"mappings":";;;;;AAyCA,IAAM,aAAA,uBAAoB,GAAA,EAA2B;AACrD,IAAM,YAAA,uBAAmB,GAAA,EAAoB;AAStC,SAAS,oBAAA,CAAqB,SAAiB,OAAA,EAA+B;AACnF,EAAA,MAAM,QAAA,GAAW,QAAA,CAAS,GAAA,CAAI,OAAO,KAAK,EAAC;AAC3C,EAAA,QAAA,CAAS,KAAK,OAAO,CAAA;AAIrB,EAAA,QAAA,CAAS,OAAO,OAAO,CAAA;AACvB,EAAA,QAAA,CAAS,GAAA,CAAI,SAAS,QAAQ,CAAA;AAC9B,EAAA,cAAA,EAAe;AACjB;AAWO,IAAM,mBAAA,GAAsB,EAAA;AAEnC,SAAS,cAAA,GAAuB;AAC9B,EAAA,OAAO,QAAA,CAAS,OAAO,mBAAA,EAAqB;AAC1C,IAAA,MAAM,MAAA,GAAS,QAAA,CAAS,IAAA,EAAK,CAAE,MAAK,CAAE,KAAA;AACtC,IAAA,IAAI,WAAW,MAAA,EAAW;AAC1B,IAAA,QAAA,CAAS,OAAO,MAAM,CAAA;AACtB,IAAA,iBAAA,CAAkB,MAAM,CAAA;AAAA,EAC1B;AACF;AAeA,SAAS,kBAAkB,OAAA,EAAuB;AAChD,EAAA,MAAM,MAAA,GAAS,cAAc,EAAA,EAAI,OAAO,EAAE,KAAA,CAAM,CAAA,GAAI,OAAA,CAAQ,MAAA,GAAS,CAAC,CAAA;AACtE,EAAA,KAAA,MAAW,CAAA,IAAK,CAAC,GAAG,YAAY,CAAA,EAAG,IAAI,CAAA,CAAE,QAAA,CAAS,MAAM,CAAA,EAAG,YAAA,CAAa,MAAA,CAAO,CAAC,CAAA;AAChF,EAAA,KAAA,MAAW,CAAA,IAAK,CAAC,GAAG,aAAA,CAAc,MAAM,CAAA,EAAG,IAAI,CAAA,CAAE,QAAA,CAAS,MAAM,CAAA,EAAG,aAAA,CAAc,OAAO,CAAC,CAAA;AACzF,EAAA,KAAA,MAAW,CAAA,IAAK,CAAC,GAAG,YAAA,CAAa,MAAM,CAAA,EAAG,IAAI,CAAA,CAAE,QAAA,CAAS,MAAM,CAAA,EAAG,YAAA,CAAa,OAAO,CAAC,CAAA;AACzF;AAaO,SAAS,cAAA,CAAe,KAAa,OAAA,EAAyB;AACnE,EAAA,MAAM,GAAA,GAAM,aAAA,CAAc,GAAA,EAAK,OAAO,CAAA;AACtC,EAAA,IAAI,OAAA,GAAU,CAAA;AAUd,EAAA,IAAI,YAAA,CAAa,MAAA,CAAO,GAAG,CAAA,EAAG,OAAA,EAAA;AAC9B,EAAA,IAAI,aAAA,CAAc,MAAA,CAAO,GAAG,CAAA,EAAG,OAAA,EAAA;AAC/B,EAAA,IAAI,YAAA,CAAa,MAAA,CAAO,GAAG,CAAA,EAAG,OAAA,EAAA;AAC9B,EAAA,OAAO,OAAA;AACT;AAEO,SAAS,mBAAmB,OAAA,EAAmC;AACpE,EAAA,OAAO,QAAA,CAAS,GAAA,CAAI,OAAO,CAAA,IAAK,EAAC;AACnC;AAWO,SAAS,uBAAA,CACd,KAAA,EACA,GAAA,EACA,SAAA,EACA,MACA,SAAA,EACM;AACN,EAAA,MAAM,GAAA,GAAM,aAAA,CAAc,GAAA,CAAI,GAAA,EAAK,IAAI,OAAO,CAAA;AAM9C,EAAA,MAAM,OAAA,GAAA,CAAW,cAAc,GAAA,CAAI,GAAG,KAAK,OAAA,CAAQ,OAAA,EAAQ,EAAG,IAAA,CAAK,YAAY;AAC7E,IAAA,IAAI;AACF,MAAA,MAAM,WAAA,CAAY,KAAA,EAAO,GAAA,EAAK,SAAA,EAAW,IAAI,CAAA;AAC7C,MAAA,MAAM,KAAA,GAAA,CAAS,YAAA,CAAa,GAAA,CAAI,GAAG,KAAK,CAAA,IAAK,CAAA;AAC7C,MAAA,YAAA,CAAa,GAAA,CAAI,KAAK,KAAK,CAAA;AAI3B,MAAA,IAAI,IAAA,CAAK,gBAAgB,KAAA,CAAA,EAAW;AAClC,QAAA,MAAM,EAAE,mBAAA,EAAoB,GAAI,MAAM,OAAO,+BAAsB,CAAA;AACnE,QAAA,MAAM,KAAA,GAAQ,MAAM,mBAAA,CAAoB;AAAA,UACtC,KAAA;AAAA,UACA,GAAA;AAAA,UACA,SAAA;AAAA,UACA,UAAA,EAAY,KAAK,WAAA,CAAY,UAAA;AAAA,UAC7B,aAAA,EAAe,KAAK,WAAA,CAAY,aAAA;AAAA,UAChC,SAAA,EAAW,KAAA;AAAA,UACX,SAAA,EAAW,KAAK,WAAA,CAAY;AAAA,SAC7B,CAAA;AACD,QAAA,IAAI,OAAO,SAAA,IAAY;AAAA,MACzB;AAAA,IACF,SAAS,KAAA,EAAO;AACd,MAAA,MAAM,MAAM,KAAA,YAAiB,KAAA,GAAQ,KAAA,CAAM,OAAA,GAAU,OAAO,KAAK,CAAA;AACjE,MAAA,IAAA,CAAK,CAAA,+CAAA,EAAkD,GAAA,CAAI,OAAO,CAAA,GAAA,EAAM,GAAG;AAAA,CAAI,CAAA;AAAA,IACjF;AAAA,EACF,CAAC,CAAA;AACD,EAAA,aAAA,CAAc,GAAA;AAAA,IACZ,GAAA;AAAA,IACA,OAAA,CAAQ,IAAA;AAAA,MACN,MAAM,MAAA;AAAA,MACN,MAAM;AAAA;AACR,GACF;AACF;AAQA,eAAsB,cAAA,CACpB,SACA,GAAA,EACe;AACf,EAAA,MAAM,GAAA,GAAM,aAAA,CAAc,GAAA,CAAI,GAAA,EAAK,OAAO,CAAA;AAC1C,EAAA,IAAI,YAAA,CAAa,GAAA,CAAI,GAAG,CAAA,EAAG;AAC3B,EAAA,YAAA,CAAa,IAAI,GAAG,CAAA;AAEpB,EAAA,MAAM,SAAA,GAAY,MAAM,mBAAA,CAAoB,GAAA,CAAI,OAAO,OAAO,CAAA;AAC9D,EAAA,IAAI,SAAA,CAAU,WAAW,CAAA,EAAG;AAM5B,EAAA,QAAA,CAAS,GAAA,CAAI,SAAS,SAAS,CAAA;AACjC;AAQA,eAAsB,kBAAA,GAAoC;AACxD,EAAA,OAAO,aAAA,CAAc,OAAO,CAAA,EAAG;AAC7B,IAAA,MAAM,GAAA,GAAM,KAAA,CAAM,IAAA,CAAK,aAAA,CAAc,QAAQ,CAAA;AAC7C,IAAA,aAAA,CAAc,KAAA,EAAM;AACpB,IAAA,MAAM,OAAA,CAAQ,IAAI,GAAG,CAAA;AAAA,EACvB;AACF;AAkBO,SAAS,aAAa,OAAA,EAAuB;AAClD,EAAA,QAAA,CAAS,OAAO,OAAO,CAAA;AACzB;AAOO,SAAS,mBAAA,CACd,GAAA,EACA,OAAA,EACA,EAAA,EACY;AACZ,EAAA,MAAM,GAAA,GAAM,aAAA,CAAc,GAAA,EAAK,OAAO,CAAA;AACtC,EAAA,MAAM,QAAQ,aAAA,CAAc,GAAA,CAAI,GAAG,CAAA,IAAK,QAAQ,OAAA,EAAQ;AACxD,EAAA,MAAM,MAAA,GAAS,KAAA,CAAM,IAAA,CAAK,EAAE,CAAA;AAC5B,EAAA,aAAA,CAAc,GAAA;AAAA,IACZ,GAAA;AAAA,IACA,MAAA,CAAO,IAAA;AAAA,MACL,MAAM,MAAA;AAAA,MACN,MAAM;AAAA;AACR,GACF;AACA,EAAA,OAAO,MAAA;AACT","file":"chunk-EXI7EHDN.js","sourcesContent":["import type { SessionStore } from \"../../types/session-store.js\";\nimport {\n  type PersistTurnInput,\n  persistTurn,\n  readSessionMessages,\n  type TranscriptLocation,\n} from \"./agent-session-store.js\";\n\n/**\n * Per-agent conversation history kept across runs (and across `Agent.resume()`\n * within the same process). Lets the fixture responder recall prior facts when\n * the user asks a follow-up question.\n *\n * SE40 — disk persistence IS the native Claude-shaped `.jsonl` transcript. The\n * in-memory cache holds the narrowed `SessionMessage[]` (user/assistant text) so\n * a send can read `priorMessages` synchronously; the whole rich turn (user +\n * assistant + tool blocks) lands on disk once per send via {@link persistTurnToTranscript}\n * (fed by `run.conversation()`). Hydration on resume reconstructs the session by\n * walking the transcript DAG (`readSessionMessages`).\n *\n * @internal\n */\n\n// `SessionMessage` lives in `./session-types.ts` (leaf types file). Re-exported\n// for back-compat with downstream importers that pulled it from here.\nexport type { SessionMessage } from \"./types.js\";\n\nimport type { SessionMessage } from \"./types.js\";\n\n// M75 — the cache lives in a leaf; see session-cache.ts for the reason (cycle broken by extraction).\nexport {\n  hydratedKeys,\n  invalidateSessionCache,\n  sessions,\n  transcriptKey,\n} from \"./session-cache.js\";\n\nimport { diag } from \"../diagnostics.js\";\nimport { globalSingleton } from \"../global-singleton.js\";\nimport { hydratedKeys, sessions, transcriptKey } from \"./session-cache.js\";\n\nconst pendingWrites = new Map<string, Promise<void>>();\nconst recordCounts = new Map<string, number>();\n\n/**\n * Append a session message to the in-memory cache only. Disk persistence for the\n * whole turn happens once per send via {@link persistTurnToTranscript}; the cache\n * feeds `priorMessages` / `onBeforeSend.previousMessageCount` synchronously.\n *\n * @internal\n */\nexport function appendSessionMessage(agentId: string, message: SessionMessage): void {\n  const existing = sessions.get(agentId) ?? [];\n  existing.push(message);\n  // `delete` + `set` reinserts at the END: a JS `Map` preserves insertion order, so the first key\n  // is always the least recently touched. That is the entire LRU, with no new structure (parsimony\n  // rungs 2/5 — the ordering we need is already a language guarantee).\n  sessions.delete(agentId);\n  sessions.set(agentId, existing);\n  enforceCeiling();\n}\n\n/**\n * Ceiling on sessions kept in memory.\n *\n * The runtime only reads the **active** session; the rest is pure cache, rebuildable from the\n * on-disk transcript. 32 is deliberately generous — the primary removal path is the explicit\n * `discardSession()` at the end of the agent's life, and this ceiling is a safety net against a\n * long-lived process running hundreds of sessions (plan risk #2: a tight ceiling could evict a\n * session still referenced by an in-flight async path).\n */\nexport const MAX_CACHED_SESSIONS = 32;\n\nfunction enforceCeiling(): void {\n  while (sessions.size > MAX_CACHED_SESSIONS) {\n    const oldest = sessions.keys().next().value;\n    if (oldest === undefined) return;\n    sessions.delete(oldest);\n    forgetBookkeeping(oldest);\n  }\n}\n\n/**\n * Erases the bookkeeping for an `agentId` across EVERY `cwd` it appears in.\n *\n * The three maps are keyed by `transcriptKey(cwd, agentId)`; `sessions` is the only one keyed by the\n * raw `agentId`. The ceiling's first version deleted from all three using `sessions`' key — that is,\n * deleted nothing — and left `hydratedKeys` **orphaned**. Since `hydrateSession` returns early when\n * the marker is present, an evicted session came back **empty** instead of rehydrating from disk:\n * silent amnesia, and a regression new to M95, because before it nothing evicted.\n *\n * The ceiling only knows the `agentId`, not the `cwd`, so it scans by suffix — which is the format\n * `transcriptKey` produces. A scan rather than an index because these maps hold tens of entries, not\n * thousands: a reverse index here would be new structure for a problem that does not exist.\n */\nfunction forgetBookkeeping(agentId: string): void {\n  const suffix = transcriptKey(\"\", agentId).slice(0 - agentId.length - 2);\n  for (const k of [...hydratedKeys]) if (k.endsWith(suffix)) hydratedKeys.delete(k);\n  for (const k of [...pendingWrites.keys()]) if (k.endsWith(suffix)) pendingWrites.delete(k);\n  for (const k of [...recordCounts.keys()]) if (k.endsWith(suffix)) recordCounts.delete(k);\n}\n\n/**\n * Erases the agent's module bookkeeping and returns how many entries were removed.\n *\n * M95 — `invalidateSessionCache` used to clear **two** of the four maps (`sessions`, `hydratedKeys`);\n * `pendingWrites` and `recordCounts` were never touched by id, so they grew for the life of the\n * process. Neither is large per entry — the leak is in count, not volume — but a cache with no owner\n * for removal is a cache that only grows.\n *\n * Returns the count so the caller can prove the removal; a second discard returns 0, which is what\n * makes the idempotency test possible without exposing the maps.\n */\nexport function discardSession(cwd: string, agentId: string): number {\n  const key = transcriptKey(cwd, agentId);\n  let removed = 0;\n  // `sessions` is NOT erased here — and the distinction is measured, not aesthetic. It holds the\n  // readable conversation, and there is a legitimate reader AFTER dispose: the golden\n  // `two-concurrent-sends-serialize` calls `getSessionMessages(agentId)` after `agent.dispose()`.\n  // Erasing it here returned an empty list and broke two goldens. What bounds it is the LRU ceiling\n  // above; the three below are pure bookkeeping, with no post-dispose reader.\n  //\n  // The key is `transcriptKey(cwd, agentId)` in ALL THREE — not the raw `agentId`. The first version\n  // erased two of them by `agentId` and therefore **never erased anything**; the test did not catch\n  // it because it only asserted that the SECOND call returns 0, which is true either way.\n  if (hydratedKeys.delete(key)) removed++;\n  if (pendingWrites.delete(key)) removed++;\n  if (recordCounts.delete(key)) removed++;\n  return removed;\n}\n\nexport function getSessionMessages(agentId: string): SessionMessage[] {\n  return sessions.get(agentId) ?? [];\n}\n\n/**\n * Persist a full conversation turn (user + assistant + tool blocks) to the native\n * transcript. Chained per-(agent, transcript) so on-disk order matches send order,\n * and fire-and-forget so `send()` is not blocked by disk I/O. Every\n * M50 — when the caller supplies `turn.autoCompact`, size-driven auto-compaction (usage real vs\n * the model's context window) runs in this same chain, surfaced via the optional `onCompact` observer.\n *\n * @internal\n */\nexport function persistTurnToTranscript(\n  store: SessionStore,\n  loc: TranscriptLocation,\n  sessionId: string,\n  turn: PersistTurnInput,\n  onCompact?: () => void,\n): void {\n  const key = transcriptKey(loc.cwd, loc.agentId);\n  // PRE-EXISTING debt, exposed when M75 fixed the Biome config that used to abort before\n  // sweeping these files (a nested root under refactor/). It is not new code and was not touched\n  // by M75; refactoring SDK internals without review would trade a visible problem for a diff\n  // risky. Tracked in usetheodev/theokit-sdk#151.\n  // biome-ignore lint/complexity/noExcessiveCognitiveComplexity: see the reason just above\n  const chained = (pendingWrites.get(key) ?? Promise.resolve()).then(async () => {\n    try {\n      await persistTurn(store, loc, sessionId, turn);\n      const count = (recordCounts.get(key) ?? 0) + 1;\n      recordCounts.set(key, count);\n      // M50 — the 50-turn no-summary boundary stub is GONE (it silently amnesia'd the session).\n      // Auto-compaction is now size-driven with a real summary: see `maybeAutoCompact` below,\n      // invoked in this same write chain by the post-run lifecycle when usage is known.\n      if (turn.autoCompact !== undefined) {\n        const { autoCompactIfNeeded } = await import(\"./compact-session.js\");\n        const fired = await autoCompactIfNeeded({\n          store,\n          loc,\n          sessionId,\n          usageTotal: turn.autoCompact.usageTotal,\n          contextWindow: turn.autoCompact.contextWindow,\n          turnCount: count,\n          summarize: turn.autoCompact.summarize,\n        });\n        if (fired) onCompact?.();\n      }\n    } catch (cause) {\n      const msg = cause instanceof Error ? cause.message : String(cause);\n      diag(`[theokit-sdk] session transcript write failed (${loc.agentId}): ${msg}\\n`);\n    }\n  });\n  pendingWrites.set(\n    key,\n    chained.then(\n      () => undefined,\n      () => undefined,\n    ),\n  );\n}\n\n/**\n * Load the persisted transcript into the in-memory cache. Idempotent per\n * (baseDir, cwd, agentId). Call once per agent lifecycle before the first read.\n *\n * @internal\n */\nexport async function hydrateSession(\n  agentId: string,\n  loc: { store: SessionStore; cwd: string },\n): Promise<void> {\n  const key = transcriptKey(loc.cwd, agentId);\n  if (hydratedKeys.has(key)) return;\n  hydratedKeys.add(key);\n\n  const persisted = await readSessionMessages(loc.store, agentId);\n  if (persisted.length === 0) return;\n  // M51 review F4 — the DISK is the source of truth at hydration time: after an invalidation\n  // (compact/inject), an in-flight turn may have repopulated the cache with a SINGLE message before\n  // this hydrate ran; the old \"skip when non-empty\" guard then pinned the parent to a 1-message\n  // context (history + injected pair lost until restart). The persist chain serializes writes, so\n  // the disk already contains that in-flight turn — replacing is always correct.\n  sessions.set(agentId, persisted);\n}\n\n/**\n * Wait for all pending transcript writes to settle. Used by tests and by the\n * agent dispose path so on-disk state matches in-memory before the caller proceeds.\n *\n * @internal\n */\nexport async function flushSessionWrites(): Promise<void> {\n  while (pendingWrites.size > 0) {\n    const all = Array.from(pendingWrites.values());\n    pendingWrites.clear();\n    await Promise.all(all);\n  }\n}\n\n/**\n * Drop the in-memory message cache for `agentId` and NOTHING else — the hydration marker in\n * `hydratedKeys` deliberately stays.\n *\n * That asymmetry is the feature, not an oversight, and it is what separates this from\n * `invalidateSessionCache`, which drops both. The single caller is the personality switch's\n * `reset: true` path (`internal/personality/switch.ts`): leaving the marker set makes\n * `hydrateSession` return early, so the next send starts from an EMPTY context instead of\n * replaying the pre-reset transcript off disk. Clearing the marker here would re-hydrate the\n * conversation the reset was asked to discard.\n *\n * The same shape read as a bug one function up: `forgetBookkeeping` documents an evicted session\n * coming back empty as \"silent amnesia\". The difference is intent — eviction wants the transcript\n * back, a reset does not. Anyone reaching for this to invalidate a cache wants\n * `invalidateSessionCache(cwd, agentId)` instead.\n */\nexport function clearSession(agentId: string): void {\n  sessions.delete(agentId);\n}\n\n/**\n * M50 review F5 — run `fn` serialized on the SAME per-(cwd,agentId) write chain the per-turn\n * persistence uses, so a manual `Agent.compact` can never interleave with an in-flight turn's\n * writes (boundary landing mid-turn would orphan the turn from the replay).\n */\nexport function enqueueSessionWrite<T>(\n  cwd: string,\n  agentId: string,\n  fn: () => Promise<T>,\n): Promise<T> {\n  const key = transcriptKey(cwd, agentId);\n  const prior = pendingWrites.get(key) ?? Promise.resolve();\n  const result = prior.then(fn);\n  pendingWrites.set(\n    key,\n    result.then(\n      () => undefined,\n      () => undefined,\n    ),\n  );\n  return result;\n}\n\n/** Test-only: drop every cached session and hydration marker. @internal */\nexport function clearAllSessions(): void {\n  sessions.clear();\n  hydratedKeys.clear();\n  recordCounts.clear();\n  // M50 review F11 — the auto-compact attempt marks live on globalThis; tests reset them here.\n  // Through the helper, with the SAME key compact-session.ts uses: two files hand-rolling one slot is\n  // how two copies of a mechanism drift apart.\n  globalSingleton(\n    \"theokit-sdk.session.auto-compact-attempts\",\n    () => new Map<string, number>(),\n  ).clear();\n}\n"]}
+{"version":3,"sources":["../src/internal/session/agent-session.ts"],"names":[],"mappings":";;;;;AAyCA,IAAM,aAAA,uBAAoB,GAAA,EAA2B;AACrD,IAAM,YAAA,uBAAmB,GAAA,EAAoB;AAStC,SAAS,oBAAA,CAAqB,SAAiB,OAAA,EAA+B;AACnF,EAAA,MAAM,QAAA,GAAW,QAAA,CAAS,GAAA,CAAI,OAAO,KAAK,EAAC;AAC3C,EAAA,QAAA,CAAS,KAAK,OAAO,CAAA;AAIrB,EAAA,QAAA,CAAS,OAAO,OAAO,CAAA;AACvB,EAAA,QAAA,CAAS,GAAA,CAAI,SAAS,QAAQ,CAAA;AAC9B,EAAA,cAAA,EAAe;AACjB;AAWO,IAAM,mBAAA,GAAsB,EAAA;AAEnC,SAAS,cAAA,GAAuB;AAC9B,EAAA,OAAO,QAAA,CAAS,OAAO,mBAAA,EAAqB;AAC1C,IAAA,MAAM,MAAA,GAAS,QAAA,CAAS,IAAA,EAAK,CAAE,MAAK,CAAE,KAAA;AACtC,IAAA,IAAI,WAAW,MAAA,EAAW;AAC1B,IAAA,QAAA,CAAS,OAAO,MAAM,CAAA;AACtB,IAAA,iBAAA,CAAkB,MAAM,CAAA;AAAA,EAC1B;AACF;AAeA,SAAS,kBAAkB,OAAA,EAAuB;AAChD,EAAA,MAAM,MAAA,GAAS,cAAc,EAAA,EAAI,OAAO,EAAE,KAAA,CAAM,CAAA,GAAI,OAAA,CAAQ,MAAA,GAAS,CAAC,CAAA;AACtE,EAAA,KAAA,MAAW,CAAA,IAAK,CAAC,GAAG,YAAY,CAAA,EAAG,IAAI,CAAA,CAAE,QAAA,CAAS,MAAM,CAAA,EAAG,YAAA,CAAa,MAAA,CAAO,CAAC,CAAA;AAChF,EAAA,KAAA,MAAW,CAAA,IAAK,CAAC,GAAG,aAAA,CAAc,MAAM,CAAA,EAAG,IAAI,CAAA,CAAE,QAAA,CAAS,MAAM,CAAA,EAAG,aAAA,CAAc,OAAO,CAAC,CAAA;AACzF,EAAA,KAAA,MAAW,CAAA,IAAK,CAAC,GAAG,YAAA,CAAa,MAAM,CAAA,EAAG,IAAI,CAAA,CAAE,QAAA,CAAS,MAAM,CAAA,EAAG,YAAA,CAAa,OAAO,CAAC,CAAA;AACzF;AAaO,SAAS,cAAA,CAAe,KAAa,OAAA,EAAyB;AACnE,EAAA,MAAM,GAAA,GAAM,aAAA,CAAc,GAAA,EAAK,OAAO,CAAA;AACtC,EAAA,IAAI,OAAA,GAAU,CAAA;AAUd,EAAA,IAAI,YAAA,CAAa,MAAA,CAAO,GAAG,CAAA,EAAG,OAAA,EAAA;AAC9B,EAAA,IAAI,aAAA,CAAc,MAAA,CAAO,GAAG,CAAA,EAAG,OAAA,EAAA;AAC/B,EAAA,IAAI,YAAA,CAAa,MAAA,CAAO,GAAG,CAAA,EAAG,OAAA,EAAA;AAC9B,EAAA,OAAO,OAAA;AACT;AAEO,SAAS,mBAAmB,OAAA,EAAmC;AACpE,EAAA,OAAO,QAAA,CAAS,GAAA,CAAI,OAAO,CAAA,IAAK,EAAC;AACnC;AAWO,SAAS,uBAAA,CACd,KAAA,EACA,GAAA,EACA,SAAA,EACA,MACA,SAAA,EACM;AACN,EAAA,MAAM,GAAA,GAAM,aAAA,CAAc,GAAA,CAAI,GAAA,EAAK,IAAI,OAAO,CAAA;AAM9C,EAAA,MAAM,OAAA,GAAA,CAAW,cAAc,GAAA,CAAI,GAAG,KAAK,OAAA,CAAQ,OAAA,EAAQ,EAAG,IAAA,CAAK,YAAY;AAC7E,IAAA,IAAI;AACF,MAAA,MAAM,WAAA,CAAY,KAAA,EAAO,GAAA,EAAK,SAAA,EAAW,IAAI,CAAA;AAC7C,MAAA,MAAM,KAAA,GAAA,CAAS,YAAA,CAAa,GAAA,CAAI,GAAG,KAAK,CAAA,IAAK,CAAA;AAC7C,MAAA,YAAA,CAAa,GAAA,CAAI,KAAK,KAAK,CAAA;AAI3B,MAAA,IAAI,IAAA,CAAK,gBAAgB,KAAA,CAAA,EAAW;AAClC,QAAA,MAAM,EAAE,mBAAA,EAAoB,GAAI,MAAM,OAAO,+BAAsB,CAAA;AACnE,QAAA,MAAM,KAAA,GAAQ,MAAM,mBAAA,CAAoB;AAAA,UACtC,KAAA;AAAA,UACA,GAAA;AAAA,UACA,SAAA;AAAA,UACA,UAAA,EAAY,KAAK,WAAA,CAAY,UAAA;AAAA,UAC7B,aAAA,EAAe,KAAK,WAAA,CAAY,aAAA;AAAA,UAChC,SAAA,EAAW,KAAA;AAAA,UACX,SAAA,EAAW,KAAK,WAAA,CAAY;AAAA,SAC7B,CAAA;AACD,QAAA,IAAI,OAAO,SAAA,IAAY;AAAA,MACzB;AAAA,IACF,SAAS,KAAA,EAAO;AACd,MAAA,MAAM,MAAM,KAAA,YAAiB,KAAA,GAAQ,KAAA,CAAM,OAAA,GAAU,OAAO,KAAK,CAAA;AACjE,MAAA,IAAA,CAAK,CAAA,+CAAA,EAAkD,GAAA,CAAI,OAAO,CAAA,GAAA,EAAM,GAAG;AAAA,CAAI,CAAA;AAAA,IACjF;AAAA,EACF,CAAC,CAAA;AACD,EAAA,aAAA,CAAc,GAAA;AAAA,IACZ,GAAA;AAAA,IACA,OAAA,CAAQ,IAAA;AAAA,MACN,MAAM,MAAA;AAAA,MACN,MAAM;AAAA;AACR,GACF;AACF;AAQA,eAAsB,cAAA,CACpB,SACA,GAAA,EACe;AACf,EAAA,MAAM,GAAA,GAAM,aAAA,CAAc,GAAA,CAAI,GAAA,EAAK,OAAO,CAAA;AAC1C,EAAA,IAAI,YAAA,CAAa,GAAA,CAAI,GAAG,CAAA,EAAG;AAC3B,EAAA,YAAA,CAAa,IAAI,GAAG,CAAA;AAEpB,EAAA,MAAM,SAAA,GAAY,MAAM,mBAAA,CAAoB,GAAA,CAAI,OAAO,OAAO,CAAA;AAC9D,EAAA,IAAI,SAAA,CAAU,WAAW,CAAA,EAAG;AAM5B,EAAA,QAAA,CAAS,GAAA,CAAI,SAAS,SAAS,CAAA;AACjC;AAQA,eAAsB,kBAAA,GAAoC;AACxD,EAAA,OAAO,aAAA,CAAc,OAAO,CAAA,EAAG;AAC7B,IAAA,MAAM,GAAA,GAAM,KAAA,CAAM,IAAA,CAAK,aAAA,CAAc,QAAQ,CAAA;AAC7C,IAAA,aAAA,CAAc,KAAA,EAAM;AACpB,IAAA,MAAM,OAAA,CAAQ,IAAI,GAAG,CAAA;AAAA,EACvB;AACF;AAkBO,SAAS,aAAa,OAAA,EAAuB;AAClD,EAAA,QAAA,CAAS,OAAO,OAAO,CAAA;AACzB;AAOO,SAAS,mBAAA,CACd,GAAA,EACA,OAAA,EACA,EAAA,EACY;AACZ,EAAA,MAAM,GAAA,GAAM,aAAA,CAAc,GAAA,EAAK,OAAO,CAAA;AACtC,EAAA,MAAM,QAAQ,aAAA,CAAc,GAAA,CAAI,GAAG,CAAA,IAAK,QAAQ,OAAA,EAAQ;AACxD,EAAA,MAAM,MAAA,GAAS,KAAA,CAAM,IAAA,CAAK,EAAE,CAAA;AAC5B,EAAA,aAAA,CAAc,GAAA;AAAA,IACZ,GAAA;AAAA,IACA,MAAA,CAAO,IAAA;AAAA,MACL,MAAM,MAAA;AAAA,MACN,MAAM;AAAA;AACR,GACF;AACA,EAAA,OAAO,MAAA;AACT","file":"chunk-7Y23B7XX.js","sourcesContent":["import type { SessionStore } from \"../../types/session-store.js\";\nimport {\n  type PersistTurnInput,\n  persistTurn,\n  readSessionMessages,\n  type TranscriptLocation,\n} from \"./agent-session-store.js\";\n\n/**\n * Per-agent conversation history kept across runs (and across `Agent.resume()`\n * within the same process). Lets the fixture responder recall prior facts when\n * the user asks a follow-up question.\n *\n * SE40 — disk persistence IS the native Claude-shaped `.jsonl` transcript. The\n * in-memory cache holds the narrowed `SessionMessage[]` (user/assistant text) so\n * a send can read `priorMessages` synchronously; the whole rich turn (user +\n * assistant + tool blocks) lands on disk once per send via {@link persistTurnToTranscript}\n * (fed by `run.conversation()`). Hydration on resume reconstructs the session by\n * walking the transcript DAG (`readSessionMessages`).\n *\n * @internal\n */\n\n// `SessionMessage` lives in `./session-types.ts` (leaf types file). Re-exported\n// for back-compat with downstream importers that pulled it from here.\nexport type { SessionMessage } from \"./types.js\";\n\nimport type { SessionMessage } from \"./types.js\";\n\n// M75 — the cache lives in a leaf; see session-cache.ts for the reason (cycle broken by extraction).\nexport {\n  hydratedKeys,\n  invalidateSessionCache,\n  sessions,\n  transcriptKey,\n} from \"./session-cache.js\";\n\nimport { diag } from \"../diagnostics.js\";\nimport { globalSingleton } from \"../global-singleton.js\";\nimport { hydratedKeys, sessions, transcriptKey } from \"./session-cache.js\";\n\nconst pendingWrites = new Map<string, Promise<void>>();\nconst recordCounts = new Map<string, number>();\n\n/**\n * Append a session message to the in-memory cache only. Disk persistence for the\n * whole turn happens once per send via {@link persistTurnToTranscript}; the cache\n * feeds `priorMessages` / `onBeforeSend.previousMessageCount` synchronously.\n *\n * @internal\n */\nexport function appendSessionMessage(agentId: string, message: SessionMessage): void {\n  const existing = sessions.get(agentId) ?? [];\n  existing.push(message);\n  // `delete` + `set` reinserts at the END: a JS `Map` preserves insertion order, so the first key\n  // is always the least recently touched. That is the entire LRU, with no new structure (parsimony\n  // rungs 2/5 — the ordering we need is already a language guarantee).\n  sessions.delete(agentId);\n  sessions.set(agentId, existing);\n  enforceCeiling();\n}\n\n/**\n * Ceiling on sessions kept in memory.\n *\n * The runtime only reads the **active** session; the rest is pure cache, rebuildable from the\n * on-disk transcript. 32 is deliberately generous — the primary removal path is the explicit\n * `discardSession()` at the end of the agent's life, and this ceiling is a safety net against a\n * long-lived process running hundreds of sessions (plan risk #2: a tight ceiling could evict a\n * session still referenced by an in-flight async path).\n */\nexport const MAX_CACHED_SESSIONS = 32;\n\nfunction enforceCeiling(): void {\n  while (sessions.size > MAX_CACHED_SESSIONS) {\n    const oldest = sessions.keys().next().value;\n    if (oldest === undefined) return;\n    sessions.delete(oldest);\n    forgetBookkeeping(oldest);\n  }\n}\n\n/**\n * Erases the bookkeeping for an `agentId` across EVERY `cwd` it appears in.\n *\n * The three maps are keyed by `transcriptKey(cwd, agentId)`; `sessions` is the only one keyed by the\n * raw `agentId`. The ceiling's first version deleted from all three using `sessions`' key — that is,\n * deleted nothing — and left `hydratedKeys` **orphaned**. Since `hydrateSession` returns early when\n * the marker is present, an evicted session came back **empty** instead of rehydrating from disk:\n * silent amnesia, and a regression new to M95, because before it nothing evicted.\n *\n * The ceiling only knows the `agentId`, not the `cwd`, so it scans by suffix — which is the format\n * `transcriptKey` produces. A scan rather than an index because these maps hold tens of entries, not\n * thousands: a reverse index here would be new structure for a problem that does not exist.\n */\nfunction forgetBookkeeping(agentId: string): void {\n  const suffix = transcriptKey(\"\", agentId).slice(0 - agentId.length - 2);\n  for (const k of [...hydratedKeys]) if (k.endsWith(suffix)) hydratedKeys.delete(k);\n  for (const k of [...pendingWrites.keys()]) if (k.endsWith(suffix)) pendingWrites.delete(k);\n  for (const k of [...recordCounts.keys()]) if (k.endsWith(suffix)) recordCounts.delete(k);\n}\n\n/**\n * Erases the agent's module bookkeeping and returns how many entries were removed.\n *\n * M95 — `invalidateSessionCache` used to clear **two** of the four maps (`sessions`, `hydratedKeys`);\n * `pendingWrites` and `recordCounts` were never touched by id, so they grew for the life of the\n * process. Neither is large per entry — the leak is in count, not volume — but a cache with no owner\n * for removal is a cache that only grows.\n *\n * Returns the count so the caller can prove the removal; a second discard returns 0, which is what\n * makes the idempotency test possible without exposing the maps.\n */\nexport function discardSession(cwd: string, agentId: string): number {\n  const key = transcriptKey(cwd, agentId);\n  let removed = 0;\n  // `sessions` is NOT erased here — and the distinction is measured, not aesthetic. It holds the\n  // readable conversation, and there is a legitimate reader AFTER dispose: the golden\n  // `two-concurrent-sends-serialize` calls `getSessionMessages(agentId)` after `agent.dispose()`.\n  // Erasing it here returned an empty list and broke two goldens. What bounds it is the LRU ceiling\n  // above; the three below are pure bookkeeping, with no post-dispose reader.\n  //\n  // The key is `transcriptKey(cwd, agentId)` in ALL THREE — not the raw `agentId`. The first version\n  // erased two of them by `agentId` and therefore **never erased anything**; the test did not catch\n  // it because it only asserted that the SECOND call returns 0, which is true either way.\n  if (hydratedKeys.delete(key)) removed++;\n  if (pendingWrites.delete(key)) removed++;\n  if (recordCounts.delete(key)) removed++;\n  return removed;\n}\n\nexport function getSessionMessages(agentId: string): SessionMessage[] {\n  return sessions.get(agentId) ?? [];\n}\n\n/**\n * Persist a full conversation turn (user + assistant + tool blocks) to the native\n * transcript. Chained per-(agent, transcript) so on-disk order matches send order,\n * and fire-and-forget so `send()` is not blocked by disk I/O. Every\n * M50 — when the caller supplies `turn.autoCompact`, size-driven auto-compaction (usage real vs\n * the model's context window) runs in this same chain, surfaced via the optional `onCompact` observer.\n *\n * @internal\n */\nexport function persistTurnToTranscript(\n  store: SessionStore,\n  loc: TranscriptLocation,\n  sessionId: string,\n  turn: PersistTurnInput,\n  onCompact?: () => void,\n): void {\n  const key = transcriptKey(loc.cwd, loc.agentId);\n  // PRE-EXISTING debt, exposed when M75 fixed the Biome config that used to abort before\n  // sweeping these files (a nested root under refactor/). It is not new code and was not touched\n  // by M75; refactoring SDK internals without review would trade a visible problem for a diff\n  // risky. Tracked in usetheodev/theokit-sdk#151.\n  // biome-ignore lint/complexity/noExcessiveCognitiveComplexity: see the reason just above\n  const chained = (pendingWrites.get(key) ?? Promise.resolve()).then(async () => {\n    try {\n      await persistTurn(store, loc, sessionId, turn);\n      const count = (recordCounts.get(key) ?? 0) + 1;\n      recordCounts.set(key, count);\n      // M50 — the 50-turn no-summary boundary stub is GONE (it silently amnesia'd the session).\n      // Auto-compaction is now size-driven with a real summary: see `maybeAutoCompact` below,\n      // invoked in this same write chain by the post-run lifecycle when usage is known.\n      if (turn.autoCompact !== undefined) {\n        const { autoCompactIfNeeded } = await import(\"./compact-session.js\");\n        const fired = await autoCompactIfNeeded({\n          store,\n          loc,\n          sessionId,\n          usageTotal: turn.autoCompact.usageTotal,\n          contextWindow: turn.autoCompact.contextWindow,\n          turnCount: count,\n          summarize: turn.autoCompact.summarize,\n        });\n        if (fired) onCompact?.();\n      }\n    } catch (cause) {\n      const msg = cause instanceof Error ? cause.message : String(cause);\n      diag(`[theokit-sdk] session transcript write failed (${loc.agentId}): ${msg}\\n`);\n    }\n  });\n  pendingWrites.set(\n    key,\n    chained.then(\n      () => undefined,\n      () => undefined,\n    ),\n  );\n}\n\n/**\n * Load the persisted transcript into the in-memory cache. Idempotent per\n * (baseDir, cwd, agentId). Call once per agent lifecycle before the first read.\n *\n * @internal\n */\nexport async function hydrateSession(\n  agentId: string,\n  loc: { store: SessionStore; cwd: string },\n): Promise<void> {\n  const key = transcriptKey(loc.cwd, agentId);\n  if (hydratedKeys.has(key)) return;\n  hydratedKeys.add(key);\n\n  const persisted = await readSessionMessages(loc.store, agentId);\n  if (persisted.length === 0) return;\n  // M51 review F4 — the DISK is the source of truth at hydration time: after an invalidation\n  // (compact/inject), an in-flight turn may have repopulated the cache with a SINGLE message before\n  // this hydrate ran; the old \"skip when non-empty\" guard then pinned the parent to a 1-message\n  // context (history + injected pair lost until restart). The persist chain serializes writes, so\n  // the disk already contains that in-flight turn — replacing is always correct.\n  sessions.set(agentId, persisted);\n}\n\n/**\n * Wait for all pending transcript writes to settle. Used by tests and by the\n * agent dispose path so on-disk state matches in-memory before the caller proceeds.\n *\n * @internal\n */\nexport async function flushSessionWrites(): Promise<void> {\n  while (pendingWrites.size > 0) {\n    const all = Array.from(pendingWrites.values());\n    pendingWrites.clear();\n    await Promise.all(all);\n  }\n}\n\n/**\n * Drop the in-memory message cache for `agentId` and NOTHING else — the hydration marker in\n * `hydratedKeys` deliberately stays.\n *\n * That asymmetry is the feature, not an oversight, and it is what separates this from\n * `invalidateSessionCache`, which drops both. The single caller is the personality switch's\n * `reset: true` path (`internal/personality/switch.ts`): leaving the marker set makes\n * `hydrateSession` return early, so the next send starts from an EMPTY context instead of\n * replaying the pre-reset transcript off disk. Clearing the marker here would re-hydrate the\n * conversation the reset was asked to discard.\n *\n * The same shape read as a bug one function up: `forgetBookkeeping` documents an evicted session\n * coming back empty as \"silent amnesia\". The difference is intent — eviction wants the transcript\n * back, a reset does not. Anyone reaching for this to invalidate a cache wants\n * `invalidateSessionCache(cwd, agentId)` instead.\n */\nexport function clearSession(agentId: string): void {\n  sessions.delete(agentId);\n}\n\n/**\n * M50 review F5 — run `fn` serialized on the SAME per-(cwd,agentId) write chain the per-turn\n * persistence uses, so a manual `Agent.compact` can never interleave with an in-flight turn's\n * writes (boundary landing mid-turn would orphan the turn from the replay).\n */\nexport function enqueueSessionWrite<T>(\n  cwd: string,\n  agentId: string,\n  fn: () => Promise<T>,\n): Promise<T> {\n  const key = transcriptKey(cwd, agentId);\n  const prior = pendingWrites.get(key) ?? Promise.resolve();\n  const result = prior.then(fn);\n  pendingWrites.set(\n    key,\n    result.then(\n      () => undefined,\n      () => undefined,\n    ),\n  );\n  return result;\n}\n\n/** Test-only: drop every cached session and hydration marker. @internal */\nexport function clearAllSessions(): void {\n  sessions.clear();\n  hydratedKeys.clear();\n  recordCounts.clear();\n  // M50 review F11 — the auto-compact attempt marks live on globalThis; tests reset them here.\n  // Through the helper, with the SAME key compact-session.ts uses: two files hand-rolling one slot is\n  // how two copies of a mechanism drift apart.\n  globalSingleton(\n    \"theokit-sdk.session.auto-compact-attempts\",\n    () => new Map<string, number>(),\n  ).clear();\n}\n"]}

package/dist/{chunk-QY3CIMUV.cjs → chunk-BZKF7TXX.cjs}

@@ -328,7 +328,10 @@ async function consumeAnthropicResponse(response) {
     toolCalls: extractAnthropicToolCalls(data),
     stopReason: mapAnthropicStopReason(data.stop_reason ?? null),
     inputTokens: data.usage?.input_tokens,
-    outputTokens: data.usage?.output_tokens
+    outputTokens: data.usage?.output_tokens,
+    // B-412 — the counters the Bedrock and Vertex Anthropic clients used to drop.
+    cacheReadTokens: data.usage?.cache_read_input_tokens,
+    cacheWriteTokens: data.usage?.cache_creation_input_tokens
   };
 }
 async function parseHttpErrorBody(response) {
@@ -541,13 +544,9 @@ var AnthropicClient = class {
     }
     const accumulator = new AnthropicStreamAccumulator();
     for await (const record of parseSseStream(response.body, signal)) {
-      if (record.event === "ping" || record.event === "message_start") continue;
-      let parsed;
-      try {
-        parsed = JSON.parse(record.data);
-      } catch {
-        continue;
-      }
+      if (record.event === "ping") continue;
+      const parsed = parseAnthropicFrame(record);
+      if (parsed === void 0) continue;
       const events = accumulator.consume(parsed);
       for (const event of events) yield event;
     }
@@ -559,6 +558,30 @@ var AnthropicClient = class {
     return accumulator.finish();
   }
 };
+var CONTENT_FRAMES = /* @__PURE__ */ new Set([
+  "content_block_start",
+  "content_block_delta",
+  "message_delta"
+]);
+function parseAnthropicFrame(record) {
+  try {
+    return JSON.parse(record.data);
+  } catch (parseErr) {
+    const event = record.event;
+    const size = `${record.data.length} chars`;
+    if (CONTENT_FRAMES.has(event)) {
+      throw new chunkUNFBHV46_cjs.NetworkError(
+        `Anthropic SSE ${event} frame (${size}) is not valid JSON; the turn would lose content`,
+        { code: "stream_malformed_frame", cause: parseErr }
+      );
+    }
+    chunk6LHQPOMI_cjs.diag(
+      `[theokit-sdk] skipped an Anthropic SSE ${event} frame (${size}) that is not valid JSON
+`
+    );
+    return void 0;
+  }
+}
 var AnthropicStreamAccumulator = class {
   text = "";
   stopReason = "end_turn";
@@ -593,6 +616,11 @@ var AnthropicStreamAccumulator = class {
     if (event.type === "message_delta") {
       this.handleMessageDelta(event);
     }
+    if (event.type === "message_start") {
+      const usage = event.message?.usage;
+      if (typeof usage?.input_tokens === "number") this.inputTokens = usage.input_tokens;
+      this.readCacheCounters(usage);
+    }
     return [];
   }
   handleToolStart(start) {
@@ -645,14 +673,26 @@ var AnthropicStreamAccumulator = class {
       this.sawStopReason = true;
     }
     this.stopReason = mapAnthropicStopReason(md.delta.stop_reason);
-    if (md.usage?.input_tokens !== void 0) this.inputTokens = md.usage.input_tokens;
-    if (md.usage?.output_tokens !== void 0) this.outputTokens = md.usage.output_tokens;
-    if (md.usage?.cache_creation_input_tokens !== void 0 && md.usage.cache_creation_input_tokens > 0) {
-      this.cacheWriteTokens = md.usage.cache_creation_input_tokens;
-    }
-    if (md.usage?.cache_read_input_tokens !== void 0 && md.usage.cache_read_input_tokens > 0) {
-      this.cacheReadTokens = md.usage.cache_read_input_tokens;
+    const input = md.usage?.input_tokens;
+    if (typeof input === "number" && (input > 0 || this.inputTokens === void 0)) {
+      this.inputTokens = input;
     }
+    if (md.usage?.output_tokens !== void 0) this.outputTokens = md.usage.output_tokens;
+    this.readCacheCounters(md.usage);
+  }
+  /**
+   * T3.8 / B-412 / B-475 — the cache counters, from `message_start` and then `message_delta`. A
+   * value above 0 replaces the earlier one. A zero, a null or an absent field leaves it in place,
+   * so a `message_delta` that zero-fills its usage does not erase what `message_start` reported;
+   * a counter no event reports above 0 stays unreported (usage-accumulator strips zero entries).
+   * The `typeof` check is there because the wire is JSON: a null arrives although the type
+   * declares none.
+   */
+  readCacheCounters(usage) {
+    const write = usage?.cache_creation_input_tokens;
+    if (typeof write === "number" && write > 0) this.cacheWriteTokens = write;
+    const read = usage?.cache_read_input_tokens;
+    if (typeof read === "number" && read > 0) this.cacheReadTokens = read;
   }
   /** M2 #61 — whether the terminal `message_delta` (stop_reason) was seen. */
   get finishReasonSeen() {
@@ -748,7 +788,7 @@ async function resolveBedrockToken(region) {
     return cachedToken.value;
   }
   try {
-    const r = module$1.createRequire((typeof document === 'undefined' ? require('u' + 'rl').pathToFileURL(__filename).href : (_documentCurrentScript && _documentCurrentScript.tagName.toUpperCase() === 'SCRIPT' && _documentCurrentScript.src || new URL('chunk-QY3CIMUV.cjs', document.baseURI).href)));
+    const r = module$1.createRequire((typeof document === 'undefined' ? require('u' + 'rl').pathToFileURL(__filename).href : (_documentCurrentScript && _documentCurrentScript.tagName.toUpperCase() === 'SCRIPT' && _documentCurrentScript.src || new URL('chunk-BZKF7TXX.cjs', document.baseURI).href)));
     const gen = r("@aws/bedrock-token-generator");
     if (typeof gen.getToken !== "function") return void 0;
     const providers = r("@aws-sdk/credential-providers");
@@ -1231,6 +1271,117 @@ function toOllamaTools(tools) {
   }));
 }
 
+// src/internal/budget/normalize-usage.ts
+function int(v) {
+  if (typeof v === "number") return Number.isFinite(v) ? Math.max(0, Math.trunc(v)) : 0;
+  if (typeof v === "string") {
+    const n = Number.parseInt(v, 10);
+    return Number.isFinite(n) ? Math.max(0, n) : 0;
+  }
+  return 0;
+}
+function buildTotal(buckets) {
+  return buckets.inputTokens + buckets.outputTokens + buckets.cacheReadTokens + buckets.cacheWriteTokens;
+}
+function omitUndefined(usage) {
+  return {
+    inputTokens: usage.inputTokens,
+    outputTokens: usage.outputTokens,
+    ...usage.cacheReadTokens > 0 ? { cacheReadTokens: usage.cacheReadTokens } : {},
+    ...usage.cacheWriteTokens > 0 ? { cacheWriteTokens: usage.cacheWriteTokens } : {},
+    ...usage.reasoningTokens > 0 ? { reasoningTokens: usage.reasoningTokens } : {},
+    totalTokens: usage.totalTokens
+  };
+}
+function inferApiMode(provider) {
+  const p = provider.toLowerCase();
+  if (p === "anthropic" || p === "claude" || p === "bedrock_anthropic") {
+    return "anthropic_messages";
+  }
+  if (p === "openai-codex" || p === "codex") return "openai_responses";
+  return "openai_chat_completions";
+}
+function normalizeUsage(rawUsage, opts) {
+  if (rawUsage === null || rawUsage === void 0) {
+    return { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
+  }
+  if (typeof rawUsage !== "object") {
+    return { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
+  }
+  const raw = rawUsage;
+  const mode = opts.apiMode ?? inferApiMode(opts.provider);
+  if (mode === "anthropic_messages") return normalizeAnthropic(raw);
+  if (mode === "openai_responses") return normalizeOpenAIResponses(raw);
+  return normalizeOpenAIChat(raw);
+}
+function normalizeAnthropic(raw) {
+  const inputTokens = int(raw.input_tokens);
+  const outputTokens = int(raw.output_tokens);
+  const cacheReadTokens = int(raw.cache_read_input_tokens);
+  const cacheWriteTokens = int(raw.cache_creation_input_tokens);
+  const usage = {
+    inputTokens,
+    outputTokens,
+    cacheReadTokens,
+    cacheWriteTokens,
+    reasoningTokens: 0,
+    totalTokens: buildTotal({ inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens })
+  };
+  return omitUndefined(usage);
+}
+function normalizeOpenAIResponses(raw) {
+  const inputTotal = int(raw.input_tokens);
+  const outputTokens = int(raw.output_tokens);
+  const inputDetails = raw.input_tokens_details ?? {};
+  const outputDetails = raw.output_tokens_details ?? {};
+  const cacheReadTokens = int(inputDetails.cached_tokens);
+  const cacheWriteTokens = int(inputDetails.cache_creation_tokens);
+  const reasoningTokens = int(outputDetails.reasoning_tokens);
+  const inputTokens = Math.max(0, inputTotal - cacheReadTokens - cacheWriteTokens);
+  const usage = {
+    inputTokens,
+    outputTokens,
+    cacheReadTokens,
+    cacheWriteTokens,
+    reasoningTokens,
+    totalTokens: buildTotal({ inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens })
+  };
+  return omitUndefined(usage);
+}
+function sent(v) {
+  return v !== void 0 && v !== null;
+}
+function splitChatCompletionsPrompt(raw) {
+  const details = raw.prompt_tokens_details ?? {};
+  const bucket = (nested, topLevel) => sent(nested) || sent(topLevel) ? int(nested) || int(topLevel) : void 0;
+  const cacheReadTokens = bucket(details.cached_tokens, raw.cache_read_input_tokens);
+  const cacheWriteTokens = bucket(details.cache_write_tokens, raw.cache_creation_input_tokens);
+  const cached = (cacheReadTokens ?? 0) + (cacheWriteTokens ?? 0);
+  return {
+    ...sent(raw.prompt_tokens) ? { inputTokens: Math.max(0, int(raw.prompt_tokens) - cached) } : {},
+    ...cacheReadTokens !== void 0 ? { cacheReadTokens } : {},
+    ...cacheWriteTokens !== void 0 ? { cacheWriteTokens } : {}
+  };
+}
+function normalizeOpenAIChat(raw) {
+  const outputTokens = int(raw.completion_tokens);
+  const completionDetails = raw.completion_tokens_details ?? {};
+  const split = splitChatCompletionsPrompt(raw);
+  const inputTokens = split.inputTokens ?? 0;
+  const cacheReadTokens = split.cacheReadTokens ?? 0;
+  const cacheWriteTokens = split.cacheWriteTokens ?? 0;
+  const reasoningTokens = int(completionDetails.reasoning_tokens);
+  const usage = {
+    inputTokens,
+    outputTokens,
+    cacheReadTokens,
+    cacheWriteTokens,
+    reasoningTokens,
+    totalTokens: buildTotal({ inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens })
+  };
+  return omitUndefined(usage);
+}
+
 // src/internal/base-url.ts
 function getConfiguredBaseUrl() {
   const value = process.env.THEOKIT_API_BASE_URL;
@@ -1507,6 +1658,16 @@ function assistantMessage(message) {
   return result;
 }
 
+// src/internal/llm/openrouter-system.ts
+function encodeOpenRouterSystemContent(system) {
+  if (!Array.isArray(system) || !system.some((block) => block.cacheable === true)) {
+    return void 0;
+  }
+  return system.map(
+    (block) => block.cacheable === true ? { type: "text", text: block.text, cache_control: { type: "ephemeral" } } : { type: "text", text: block.text }
+  );
+}
+
 // src/internal/llm/openai.ts
 function deriveChatPath(baseUrl) {
   try {
@@ -1649,11 +1810,8 @@ var OpenAIStreamAccumulator = class {
   allowedToolNames;
   text = "";
   stopReason = "end_turn";
-  inputTokens;
-  outputTokens;
-  cacheReadTokens;
-  cacheWriteTokens;
-  reasoningTokens;
+  /** The stream's `usage` object, a later chunk replacing an earlier one per top-level field. */
+  usage = {};
   toolCalls = /* @__PURE__ */ new Map();
   /** R7: present only when recovery is enabled AND the request declares tools — holds suspected
    *  leaked-dialect content back from the `text_delta` stream. `undefined` ⇒ stream immediately. */
@@ -1687,18 +1845,7 @@ var OpenAIStreamAccumulator = class {
     return { type: "reasoning_delta", text: reasoning };
   }
   applyUsage(usage) {
-    if (usage?.prompt_tokens !== void 0) this.inputTokens = usage.prompt_tokens;
-    if (usage?.completion_tokens !== void 0) this.outputTokens = usage.completion_tokens;
-    const cachedDetail = usage?.prompt_tokens_details?.cached_tokens;
-    if (cachedDetail !== void 0) this.cacheReadTokens = cachedDetail;
-    const reasoningDetail = usage?.completion_tokens_details?.reasoning_tokens;
-    if (reasoningDetail !== void 0) this.reasoningTokens = reasoningDetail;
-    if (usage?.cache_read_input_tokens !== void 0) {
-      this.cacheReadTokens = usage.cache_read_input_tokens;
-    }
-    if (usage?.cache_creation_input_tokens !== void 0) {
-      this.cacheWriteTokens = usage.cache_creation_input_tokens;
-    }
+    if (usage !== void 0) this.usage = { ...this.usage, ...usage };
   }
   applyContentDelta(content) {
     if (typeof content !== "string" || content.length === 0) return void 0;
@@ -1769,11 +1916,10 @@ var OpenAIStreamAccumulator = class {
       stopReason,
       text,
       toolCalls,
-      inputTokens: this.inputTokens,
-      outputTokens: this.outputTokens,
-      cacheReadTokens: this.cacheReadTokens,
-      cacheWriteTokens: this.cacheWriteTokens,
-      reasoningTokens: this.reasoningTokens
+      // D376: the prompt split (cache buckets out of inputTokens) has one owner, normalize-usage.ts.
+      ...splitChatCompletionsPrompt(this.usage),
+      outputTokens: this.usage.completion_tokens,
+      reasoningTokens: this.usage.completion_tokens_details?.reasoning_tokens
     });
   }
 };
@@ -1794,10 +1940,10 @@ function applyToolsRequest(body, request) {
 }
 function buildOpenAIBody(request, providerName) {
   const messages = [];
+  const systemParts = providerName === "openrouter" ? encodeOpenRouterSystemContent(request.system) : void 0;
   const systemText = openAISystemText(request.system);
-  if (systemText.length > 0) {
-    messages.push({ role: "system", content: systemText });
-  }
+  if (systemParts !== void 0) messages.push({ role: "system", content: systemParts });
+  else if (systemText.length > 0) messages.push({ role: "system", content: systemText });
   for (const message of request.messages) {
     for (const out of toOpenAIMessages(message)) messages.push(out);
   }
@@ -2516,7 +2662,7 @@ async function resolveVertexAccessToken() {
 async function getAuthClientOrThrow() {
   let mod;
   try {
-    const r = module$1.createRequire((typeof document === 'undefined' ? require('u' + 'rl').pathToFileURL(__filename).href : (_documentCurrentScript && _documentCurrentScript.tagName.toUpperCase() === 'SCRIPT' && _documentCurrentScript.src || new URL('chunk-QY3CIMUV.cjs', document.baseURI).href)));
+    const r = module$1.createRequire((typeof document === 'undefined' ? require('u' + 'rl').pathToFileURL(__filename).href : (_documentCurrentScript && _documentCurrentScript.tagName.toUpperCase() === 'SCRIPT' && _documentCurrentScript.src || new URL('chunk-BZKF7TXX.cjs', document.baseURI).href)));
     mod = r("google-auth-library");
   } catch (err) {
     const isModuleMissing = err instanceof Error && (err.message.includes("Cannot find module") || err.code === "MODULE_NOT_FOUND");
@@ -3305,11 +3451,13 @@ exports.compactSessionTranscript = compactSessionTranscript;
 exports.discoverProviderPlugins = discoverProviderPlugins;
 exports.getConfiguredBaseUrl = getConfiguredBaseUrl;
 exports.httpRequest = httpRequest;
+exports.inferApiMode = inferApiMode;
 exports.isCompactSummary = isCompactSummary;
 exports.isFixtureApiKey = isFixtureApiKey;
 exports.mapOllamaHttpError = mapOllamaHttpError;
 exports.mapOllamaTransportError = mapOllamaTransportError;
 exports.mergeExplicitApiKey = mergeExplicitApiKey;
+exports.normalizeUsage = normalizeUsage;
 exports.parseSseStream = parseSseStream;
 exports.presentProviderCredentialEnvVars = presentProviderCredentialEnvVars;
 exports.readErrorResponseBody = readErrorResponseBody;
@@ -3323,5 +3471,5 @@ exports.shouldUseFixtureMode = shouldUseFixtureMode;
 exports.shouldUseRealLocalRuntime = shouldUseRealLocalRuntime;
 exports.tryFirstEvent = tryFirstEvent;
 exports.validateApiKeyShape = validateApiKeyShape;
-//# sourceMappingURL=chunk-QY3CIMUV.cjs.map
-//# sourceMappingURL=chunk-QY3CIMUV.cjs.map
+//# sourceMappingURL=chunk-BZKF7TXX.cjs.map
+//# sourceMappingURL=chunk-BZKF7TXX.cjs.map