@code-yeongyu/senpi 2026.8.9 → 2026.8.11

package/docs/extensions.md

@@ -1437,6 +1437,82 @@ pi.registerTool({
 });
 ```
 
+### pi.registerFilesystemPolicy(policy)
+
+Register a filesystem access policy for Senpi's built-in `read`, `write`,
+`edit`, `ls`, `find`, and `grep` tools. This is a **factory-time only** API: call
+it directly from the extension factory so the checker is compiled before the
+built-in tools are constructed.
+
+The policy receives a canonical absolute target:
+
+```typescript
+interface FilesystemPolicyRequest {
+  operation: "read" | "enumerate" | "write";
+  canonicalPath: string;
+  toolName: string;
+}
+```
+
+Operation mapping:
+
+| Operation | Built-in tools |
+|---|---|
+| `read` | `read` |
+| `write` | `write`, `edit` |
+| `enumerate` | `ls`, `find`, `grep` |
+
+Existing targets are resolved through filesystem symlinks. A missing write
+target is resolved through its nearest existing real parent, then its missing
+suffix is reattached. The policy runs after that resolution and immediately
+before the tool's target I/O. A denial is returned as a normal tool error with
+the policy's reason.
+
+```typescript
+import { isAbsolute, relative, resolve, sep } from "node:path";
+import type { ExtensionAPI } from "@code-yeongyu/senpi";
+
+function isContained(root: string, candidate: string): boolean {
+  const pathFromRoot = relative(root, candidate);
+  return pathFromRoot === "" || (
+    pathFromRoot !== ".." &&
+    !pathFromRoot.startsWith(`..${sep}`) &&
+    !isAbsolute(pathFromRoot)
+  );
+}
+
+export default function (pi: ExtensionAPI) {
+  const writableRoot = resolve(process.cwd());
+
+  pi.registerFilesystemPolicy({
+    check({ operation, canonicalPath }) {
+      if (operation !== "write" || isContained(writableRoot, canonicalPath)) {
+        return { allow: true };
+      }
+      return {
+        allow: false,
+        reason: `Writes are limited to ${writableRoot}`,
+      };
+    },
+  });
+}
+```
+
+Policies run in extension load and registration order. Every policy must allow a
+request; the first denial wins. A thrown policy error also stops the tool before
+its target I/O.
+
+An optional `deniedRoots: readonly string[]` field exposes metadata for future
+inherited process sandbox support. Use canonical absolute roots. The current
+runtime does not interpret this metadata and does not apply filesystem policies
+to `bash` or descendants; command-string rewriting is not a filesystem security
+boundary.
+
+This enforcement is separate from `tool_call` permission and approval hooks.
+An unrestricted approval mode, or a hook that returns `{ block: false }`, cannot
+bypass the built-in executor check. Custom extension tools are not covered
+automatically and must call their own policy if they access the filesystem.
+
 ### pi.registerMcpServer(name, config)
 
 Register an MCP server that the agent can use. This is a **factory-time only**

package/docs/providers.md

@@ -88,8 +88,18 @@ If your Claude Pro/Max subscription usage through `claude-sdk-oauth` feels unexp
 1. **Upgrade to v2026.8.3 or later.** Resume-first session continuity (#634-637) landed on 2026-08-03. On older builds, every turn after a divergence (compaction, abort, model switch, restart, failover) re-sends the entire conversation, which is the dominant token-burn mechanism.
 2. **Check which lane you are on.** The `ambient` lane (default) inherits the environment. `oauth-slots` and `config-dir` are managed lanes set via `SENPI_CLAUDE_SDK_OAUTH_TOKEN_INJECTION`. The `config-dir` lane keeps each account's credentials in its own `CLAUDE_CONFIG_DIR`; no official SDK API moves a transcript across roots, so account failover on that lane always flattens (re-sends the full history) — this is a declared residual, not a bug.
 3. **Read the continuity observations.** Tail the session log and filter for `flatten` — each `flatten` line means the lane re-sent the whole conversation and lost prompt-cache hits. A healthy conversation shows one `bootstrap` followed by `delta` lines. Common flatten reasons: `transcript_missing`, `registry_miss`, `resume_initialization_failed`, `cross_root_unsupported` (config-dir only).
-4. **Prompt-cache retention.** On the direct Anthropic API path (`anthropic` provider with an OAuth token), cache retention defaults to `long` with a 1-hour TTL on the native endpoint. Set `PI_CACHE_RETENTION=long` to ensure 1h caching. The SDK OAuth lane manages its own caching internally and senpi cannot add `cache_control` breakpoints to it.
-5. **Directive-block deduplication.** As of v2026.8.4, the flatten serialization collapses repeated `<ultrawork-mode>` directive blocks to a single copy, preventing the issue-#494 scenario where duplicated ~17KB directive blocks consumed up to 73% of the re-sent prompt. The continuity observation reports how many were collapsed and the payload size.
+4. **Prompt-cache retention.** Effective cache TTL depends on which lane you are on:
+
+   | Lane | Effective TTL | Who controls it | How to override |
+   | --- | --- | --- | --- |
+   | Claude SDK OAuth (subscription, `claude-sdk-oauth`) | 5 minutes | The Claude SDK owns `cache_control`; senpi cannot add breakpoints. senpi reports 300s for this lane so cache-aware budgets (tool waits, goal timing) size themselves correctly. | Not overridable |
+   | Direct Anthropic API (`api.anthropic.com`, API key or OAuth token) | 1 hour | senpi defaults retention to `long` on this lane. Note: Anthropic's own API default is 5 minutes; the 1h TTL is senpi's choice, and 1h cache writes cost 2x base input vs 1.25x for 5m ([Anthropic prompt caching](https://docs.claude.com/en/docs/build-with-claude/prompt-caching)). | `PI_CACHE_RETENTION`, `cacheRetention` |
+   | Anthropic-compatible providers (kimi-coding, fireworks, gateways) | 5 minutes | The 1h TTL is gated on the native `api.anthropic.com` base URL, so these lanes stay short. | `cacheRetention` |
+
+   Override precedence: `cacheRetention` in `models.json` / the model catalog wins over everything. `PI_CACHE_RETENTION=long` selects long; any other set value forces short; unset falls back to the lane default above.
+5. **Goal-monitor timing.** The goal monitor's continuation backstop is derived from the model's cache-safe wait (TTL minus `promptCache.safetyBufferSeconds`, default 30), capped by `promptCache.goalBackstopMaxSeconds` (default 3570), instead of a fixed 4 minutes. On 5m lanes that means wakes every ~4m30s; on the 1h direct-API lane, up to 59m30s. Cache-warm notices show which warm iteration you are on.
+6. **Wake sources that hold the goal backstop.** Anything that can wake a parked session publishes a `wake_source_state` event (`{source, activeCount}`): terminal monitors (`terminal-monitors`), background bash sessions including auto-detached and killed ones (`terminal-background-sessions`), detached `eval` cells (`senpi-codemode`), and omo-senpi background task children plus owned team members (`senpi-task`). The goal extension sums every source, so a goal waits inside the prompt-cache TTL while ANY of them is on duty instead of continuing immediately. The legacy `terminal_monitor_state` event is still emitted for external consumers and is folded onto the same `terminal-monitors` count.
+7. **Directive-block deduplication.** As of v2026.8.4, the flatten serialization collapses repeated `<ultrawork-mode>` directive blocks to a single copy, preventing the issue-#494 scenario where duplicated ~17KB directive blocks consumed up to 73% of the re-sent prompt. The continuity observation reports how many were collapsed and the payload size.
 
 ### GitHub Copilot
 
@@ -108,6 +118,16 @@ If your Claude Pro/Max subscription usage through `claude-sdk-oauth` feels unexp
 - On remote/headless machines (e.g. over SSH) the browser cannot reach the loopback callback; paste the final redirect URL (or the authorization code) into the login prompt instead
 - `OPENROUTER_API_KEY` remains available through **Use an API key**
 
+#### OpenRouter prompt caching and sticky routing
+
+senpi pins every OpenRouter request to the same upstream from the first call by sending the session id both as the `x-session-id` header and as the request-body `session_id` field. OpenRouter's precedence is body > header > `prompt_cache_key`, with a 256-character limit on the value; specifying a manual `provider.order` disables sticky routing entirely ([OpenRouter prompt caching](https://openrouter.ai/docs/guides/best-practices/prompt-caching)).
+
+Upstreams that require explicit cache breakpoints get `cache_control` blocks through OpenRouter: model ids starting with `anthropic/`, `qwen/`, or `google/` (catalog ids carrying one leading `~` are matched after stripping it). Other upstreams (OpenAI, DeepSeek, Moonshot, Groq, Z.AI, Grok) cache automatically and need no markers.
+
+#### Moonshot / Kimi prompt caching
+
+senpi sends `prompt_cache_key` (set to the session id) on Moonshot requests. Kimi documents the field as required for the Kimi Code Plan and recommended for any multi-turn agent ([Kimi context caching](https://platform.kimi.ai/docs/guide/use-context-caching-feature-of-kimi-api)). Kimi reports cache hits as a flat `usage.cached_tokens` field, which senpi parses as cache-read tokens.
+
 ### Radius
 
 Radius is a dynamic `pi-messages` gateway. `/login radius` stores OAuth tokens in `auth.json`; the gateway catalog is refreshed independently and cached in `models-store.json`. Custom Radius gateways can be declared in `models.json` with `"oauth": "radius"` and a gateway `baseUrl`.
@@ -296,7 +316,7 @@ Also supports ECS task roles (`AWS_CONTAINER_CREDENTIALS_*`) and IRSA (`AWS_WEB_
 pi --provider amazon-bedrock --model us.anthropic.claude-sonnet-4-20250514-v1:0
 ```
 
-Prompt caching is enabled automatically for Claude models whose ID contains a recognizable model name (base models and system-defined inference profiles). For application inference profiles (whose ARNs don't contain the model name), set `AWS_BEDROCK_FORCE_CACHE=1` to enable cache points:
+Prompt caching is enabled automatically for Claude models whose ID contains a recognizable model name (base models and system-defined inference profiles). With `PI_CACHE_RETENTION=long`, the 1-hour cache TTL is only requested for Claude Opus 4.5, Sonnet 4.5, and Haiku 4.5, the models AWS documents as supporting it ([Bedrock prompt caching](https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html)); every other cacheable Bedrock Claude model uses 5 minutes on both the wire and the TTL estimate. For application inference profiles (whose ARNs don't contain the model name), set `AWS_BEDROCK_FORCE_CACHE=1` to enable cache points:
 
 ```bash
 export AWS_BEDROCK_FORCE_CACHE=1

package/node_modules/@code-yeongyu/senpi-codemode/CHANGELOG.md

@@ -12,6 +12,44 @@
 
 ### Removed
 
+## [2026.8.11] - 2026-08-10
+
+### Breaking Changes
+
+### Added
+
+### Changed
+
+### Fixed
+
+### Removed
+
+## [2026.8.10] - 2026-08-10
+
+### Breaking Changes
+
+### Added
+
+### Changed
+
+### Fixed
+
+### Removed
+
+## [2026.8.9-2] - 2026-08-09
+
+### Breaking Changes
+
+### Added
+
+### Changed
+
+- Detached eval cells now emit the shared `wake_source_state` event under source `senpi-codemode` when they detach, complete, stop, or are disposed. The optional host event passthrough remains guarded, synchronous cells emit no lifecycle transition, and per-cell snapshot metadata is preserved.
+
+### Fixed
+
+### Removed
+
 ## [2026.8.9] - 2026-08-09
 
 ### Breaking Changes

package/node_modules/@code-yeongyu/senpi-codemode/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@code-yeongyu/senpi-codemode",
-	"version": "2026.8.9",
+	"version": "2026.8.11",
 	"description": "Source-only senpi extension package for codemode evaluation tools",
 	"type": "module",
 	"main": "./src/index.ts",
@@ -30,14 +30,14 @@
 	},
 	"dependencies": {
 		"@babel/parser": "8.0.4",
-		"@earendil-works/pi-ai": "npm:@code-yeongyu/senpi-ai@2026.8.9",
+		"@earendil-works/pi-ai": "npm:@code-yeongyu/senpi-ai@2026.8.11",
 		"typebox": "1.3.8"
 	},
 	"peerDependencies": {
-		"@code-yeongyu/senpi": "2026.8.9"
+		"@code-yeongyu/senpi": "2026.8.11"
 	},
 	"devDependencies": {
-		"@code-yeongyu/senpi": "2026.8.9"
+		"@code-yeongyu/senpi": "2026.8.11"
 	},
 	"keywords": [
 		"senpi",

package/node_modules/@code-yeongyu/senpi-codemode/src/extension/{resumption-channel.ts → wake-source-state.ts}

@@ -1,29 +1,29 @@
 /**
- * Resumption-channel liveness contract shared with the goal builtin.
+ * Wake-source liveness contract shared with the goal builtin.
  *
- * The goal builtin delays its hidden continuation while any live resumption
- * channel can still wake the session. Each emitter publishes a full per-source
+ * The goal builtin delays its hidden continuation while any live wake source
+ * can still wake the session. Each emitter publishes a full per-source
  * snapshot on every liveness transition and once on `session_start`.
  *
  * The event name is duplicated here on purpose: senpi-codemode is a separate
  * package and must not import across the packages/coding-agent boundary, so
  * the sentinel test pins this literal to the exact cross-package contract.
  */
-export const RESUMPTION_CHANNEL_STATE_EVENT = "resumption_channel_state";
+export const WAKE_SOURCE_STATE_EVENT = "wake_source_state";
 
 /** Source key identifying detached eval cells in the per-source snapshot. */
-export const EVAL_DETACHED_CHANNEL_SOURCE = "eval-detached";
+export const SENPI_CODEMODE_WAKE_SOURCE = "senpi-codemode";
 
-/** One live channel as broadcast on the resumption channel state event. */
-export interface ResumptionChannelEntry {
+/** One detached cell as broadcast on the wake-source state event. */
+export interface WakeSourceStateItem {
 	readonly id: string;
 	readonly description: string;
 	/** Epoch milliseconds when the channel registered; lets consumers render their own elapsed labels. */
 	readonly startedAtMs: number;
 }
 
-export interface ResumptionChannelState {
+export interface WakeSourceState {
 	readonly source: string;
 	readonly activeCount: number;
-	readonly channels?: readonly ResumptionChannelEntry[];
+	readonly items?: readonly WakeSourceStateItem[];
 }

package/node_modules/@code-yeongyu/senpi-codemode/src/index.ts

@@ -7,7 +7,6 @@ import { defaultCodemodeSettings } from "./config/settings.ts";
 import { EvalNotifier } from "./extension/eval-notifier.ts";
 import { EVAL_CELLS_STATUS_KEY } from "./extension/eval-status.ts";
 import { EvalStatusTicker } from "./extension/eval-status-ticker.ts";
-import { RESUMPTION_CHANNEL_STATE_EVENT, type ResumptionChannelState } from "./extension/resumption-channel.ts";
 import {
 	createExecuteTool,
 	createRuntime,
@@ -16,6 +15,7 @@ import {
 } from "./extension/runtime-factory.ts";
 import type { CodemodeSessionManager, CreateCodemodeSessionManagerOptions } from "./extension/session-manager.ts";
 import { SessionManagerProxy } from "./extension/session-manager-proxy.ts";
+import { WAKE_SOURCE_STATE_EVENT, type WakeSourceState } from "./extension/wake-source-state.ts";
 import { EvalDetachedCellManager, type EvalDetachedCellStatusEntry } from "./tool/detached-cell-manager.ts";
 import { createEvalTool } from "./tool/eval-tool.ts";
 import { renderEvalCall, renderEvalResult } from "./tool/render.ts";
@@ -39,7 +39,7 @@ export interface CodemodeExtensionAPI {
 	getActiveTools(): string[];
 	getAllTools(): readonly EvalSchemaToolInfo[];
 	sendUserMessage(content: string, options?: { deliverAs?: "steer" | "followUp" }): void;
-	/** Optional host event bus; a host without one turns resumption-channel emission into a harmless no-op. */
+	/** Optional host event bus; a host without one turns wake-source emission into a harmless no-op. */
 	events?: { emit(name: string, data: unknown): void };
 }
 
@@ -82,8 +82,8 @@ export default function senpiCodemode(pi: CodemodeExtensionAPI, options: SenpiCo
 	const showDetachedCells = (entries: readonly EvalDetachedCellStatusEntry[]): void => {
 		statusTicker.sync(entries);
 	};
-	const emitChannelState = (state: ResumptionChannelState): void => {
-		pi.events?.emit(RESUMPTION_CHANNEL_STATE_EVENT, state);
+	const emitWakeSourceState = (state: WakeSourceState): void => {
+		pi.events?.emit(WAKE_SOURCE_STATE_EVENT, state);
 	};
 	const registerEvalForRuntime = (
 		runtime: SessionRuntime,
@@ -132,7 +132,7 @@ export default function senpiCodemode(pi: CodemodeExtensionAPI, options: SenpiCo
 			cellManager: new EvalDetachedCellManager({
 				notifier,
 				onStatusChange: showDetachedCells,
-				onChannelState: emitChannelState,
+				onWakeSourceState: emitWakeSourceState,
 				...(options.now === undefined ? {} : { now: options.now }),
 			}),
 			executionTracker: manager,
@@ -163,12 +163,12 @@ export default function senpiCodemode(pi: CodemodeExtensionAPI, options: SenpiCo
 			artifactsDir: runtime.artifactsDir,
 			notifier,
 			onStatusChange: showDetachedCells,
-			onChannelState: emitChannelState,
+			onWakeSourceState: emitWakeSourceState,
 			...(options.now === undefined ? {} : { now: options.now }),
 		});
 		activeCells = cellManager;
 		// The goal builtin clears its per-session counts at session_start; re-publish our snapshot.
-		cellManager.publishChannelState();
+		cellManager.publishWakeSourceState();
 		activeRuntime = runtime;
 		activeModelId = ctx.model?.id;
 		registerEvalForRuntime(runtime, activeModelId, cellManager);

package/node_modules/@code-yeongyu/senpi-codemode/src/tool/detached-cell-manager.ts

@@ -1,5 +1,5 @@
 import type { AgentToolResult } from "@code-yeongyu/senpi";
-import { EVAL_DETACHED_CHANNEL_SOURCE, type ResumptionChannelState } from "../extension/resumption-channel.ts";
+import { SENPI_CODEMODE_WAKE_SOURCE, type WakeSourceState } from "../extension/wake-source-state.ts";
 import { detachedNotificationSpillPath } from "./detached-cell-notification.ts";
 import { currentDetachedResult, detachedErrorResult, snapshotDetachedCell } from "./detached-cell-snapshot.ts";
 import {
@@ -60,14 +60,14 @@ export interface EvalDetachedCellManagerOptions {
 	readonly notifier?: EvalDetachedCellNotifier;
 	readonly onStatusChange?: (entries: readonly EvalDetachedCellStatusEntry[]) => void;
 	/** Receives a full per-source liveness snapshot on every detached-cell transition; used by the goal builtin. */
-	readonly onChannelState?: (state: ResumptionChannelState) => void;
+	readonly onWakeSourceState?: (state: WakeSourceState) => void;
 	readonly now?: () => number;
 }
 
 export class EvalDetachedCellManager {
 	readonly #artifactsDir: string | undefined;
 	readonly #onStatusChange: ((entries: readonly EvalDetachedCellStatusEntry[]) => void) | undefined;
-	readonly #onChannelState: ((state: ResumptionChannelState) => void) | undefined;
+	readonly #onWakeSourceState: ((state: WakeSourceState) => void) | undefined;
 	readonly #cells = new Map<string, ManagedCell>();
 	readonly #detachedByLanguage = new Map<EvalLanguage, ManagedCell>();
 	readonly #notificationQueue: DetachedNotificationQueue;
@@ -76,7 +76,7 @@ export class EvalDetachedCellManager {
 	constructor(options: EvalDetachedCellManagerOptions = {}) {
 		this.#artifactsDir = options.artifactsDir;
 		this.#onStatusChange = options.onStatusChange;
-		this.#onChannelState = options.onChannelState;
+		this.#onWakeSourceState = options.onWakeSourceState;
 		this.#notificationQueue = new DetachedNotificationQueue(options.notifier, options.artifactsDir);
 		this.#now = options.now ?? Date.now;
 	}
@@ -160,6 +160,7 @@ export class EvalDetachedCellManager {
 		await Promise.allSettled(
 			detached.map(async (cell) => await this.stop(cell.cellId, "Session ended; detached eval cell cancelled")),
 		);
+		if (detached.length === 0) this.#emitWakeSourceState([]);
 		await this.#notificationQueue.flush();
 	}
 
@@ -168,8 +169,8 @@ export class EvalDetachedCellManager {
 	}
 
 	/** Re-publish the current snapshot; consumers reset their per-source counts at session_start. */
-	publishChannelState(): void {
-		this.#emitChannelState([...this.#detachedByLanguage.values()]);
+	publishWakeSourceState(): void {
+		this.#emitWakeSourceState([...this.#detachedByLanguage.values()]);
 	}
 
 	#settle(
@@ -207,14 +208,14 @@ export class EvalDetachedCellManager {
 				...(cell.input.summary === undefined ? {} : { summary: cell.input.summary }),
 			})),
 		);
-		this.#emitChannelState(liveCells);
+		this.#emitWakeSourceState(liveCells);
 	}
 
-	#emitChannelState(liveCells: readonly ManagedCell[]): void {
-		this.#onChannelState?.({
-			source: EVAL_DETACHED_CHANNEL_SOURCE,
+	#emitWakeSourceState(liveCells: readonly ManagedCell[]): void {
+		this.#onWakeSourceState?.({
+			source: SENPI_CODEMODE_WAKE_SOURCE,
 			activeCount: liveCells.length,
-			channels: liveCells.map((cell) => ({
+			items: liveCells.map((cell) => ({
 				id: cell.cellId,
 				description:
 					cell.input.summary === undefined || cell.input.summary.length === 0 ? cell.cellId : cell.input.summary,

package/node_modules/@earendil-works/pi-agent-core/package.json

@@ -1,7 +1,7 @@
 {
 	"name": "@earendil-works/pi-agent-core",
 	"private": true,
-	"version": "2026.8.9",
+	"version": "2026.8.11",
 	"description": "General-purpose agent with transport abstraction, state management, and attachment support",
 	"type": "module",
 	"main": "./dist/index.js",
@@ -30,7 +30,7 @@
 		"prepublishOnly": "npm run build"
 	},
 	"dependencies": {
-		"@earendil-works/pi-ai": "npm:@code-yeongyu/senpi-ai@2026.8.9",
+		"@earendil-works/pi-ai": "npm:@code-yeongyu/senpi-ai@2026.8.11",
 		"diff": "9.0.0",
 		"ignore": "7.0.6",
 		"typebox": "1.3.8",

package/node_modules/@earendil-works/pi-ai/dist/api/anthropic-messages.d.ts

@@ -1,5 +1,6 @@
 import Anthropic from "@anthropic-ai/sdk";
-import type { SimpleStreamOptions, StreamFunction, StreamOptions } from "../types.ts";
+import type { MessageCreateParamsNonStreaming } from "@anthropic-ai/sdk/resources/messages.js";
+import type { Context, Model, SimpleStreamOptions, StreamFunction, StreamOptions } from "../types.ts";
 export { getAnthropicCompat } from "../utils/prompt-cache-ttl.ts";
 export type AnthropicEffort = "low" | "medium" | "high" | "xhigh" | "max";
 export type AnthropicThinkingDisplay = "summarized" | "omitted";
@@ -69,4 +70,5 @@ export interface AnthropicOptions extends StreamOptions {
 }
 export declare const stream: StreamFunction<"anthropic-messages", AnthropicOptions>;
 export declare const streamSimple: StreamFunction<"anthropic-messages", SimpleStreamOptions>;
+export declare function buildAnthropicWarmPromptCacheParams(model: Model<"anthropic-messages">, context: Context, options?: AnthropicOptions): MessageCreateParamsNonStreaming;
 //# sourceMappingURL=anthropic-messages.d.ts.map

package/node_modules/@earendil-works/pi-ai/dist/api/anthropic-messages.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"anthropic-messages.d.ts","sourceRoot":"","sources":["../../src/api/anthropic-messages.ts"],"names":[],"mappings":"AAAA,OAAO,SAAS,MAAM,mBAAmB,CAAC;AAU1C,OAAO,KAAK,EAWX,mBAAmB,EAEnB,cAAc,EACd,aAAa,EAOb,MAAM,aAAa,CAAC;AAkCrB,OAAO,EAAE,kBAAkB,EAAE,MAAM,8BAA8B,CAAC;AA6JlE,MAAM,MAAM,eAAe,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,GAAG,KAAK,CAAC;AAE1E,MAAM,MAAM,wBAAwB,GAAG,YAAY,GAAG,SAAS,CAAC;AAoEhE,MAAM,WAAW,gBAAiB,SAAQ,aAAa;IACtD;;;;;;OAMG;IACH,eAAe,CAAC,EAAE,OAAO,CAAC;IAC1B;;;;OAIG;IACH,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B;;;;;;;;;;;OAWG;IACH,MAAM,CAAC,EAAE,eAAe,CAAC;IACzB;;;;;;;;;;;OAWG;IACH,eAAe,CAAC,EAAE,wBAAwB,CAAC;IAC3C;;;;;OAKG;IACH,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B;;;;OAIG;IACH,UAAU,CAAC,EAAE,MAAM,GAAG,KAAK,GAAG,MAAM,GAAG;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,CAAC;IACtE;;;;OAIG;IACH,MAAM,CAAC,EAAE,SAAS,CAAC;CACnB;AA4zBD,eAAO,MAAM,MAAM,EAAE,cAAc,CAAC,oBAAoB,EAAE,gBAAgB,CA6ZzE,CAAC;AA0FF,eAAO,MAAM,YAAY,EAAE,cAAc,CAAC,oBAAoB,EAAE,mBAAmB,CA2ClF,CAAC"}
+{"version":3,"file":"anthropic-messages.d.ts","sourceRoot":"","sources":["../../src/api/anthropic-messages.ts"],"names":[],"mappings":"AAAA,OAAO,SAAS,MAAM,mBAAmB,CAAC;AAC1C,OAAO,KAAK,EAGX,+BAA+B,EAK/B,MAAM,yCAAyC,CAAC;AAEjD,OAAO,KAAK,EAKX,OAAO,EAGP,KAAK,EAGL,mBAAmB,EAEnB,cAAc,EACd,aAAa,EAOb,MAAM,aAAa,CAAC;AAkCrB,OAAO,EAAE,kBAAkB,EAAE,MAAM,8BAA8B,CAAC;AA6JlE,MAAM,MAAM,eAAe,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,GAAG,KAAK,CAAC;AAE1E,MAAM,MAAM,wBAAwB,GAAG,YAAY,GAAG,SAAS,CAAC;AAoEhE,MAAM,WAAW,gBAAiB,SAAQ,aAAa;IACtD;;;;;;OAMG;IACH,eAAe,CAAC,EAAE,OAAO,CAAC;IAC1B;;;;OAIG;IACH,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B;;;;;;;;;;;OAWG;IACH,MAAM,CAAC,EAAE,eAAe,CAAC;IACzB;;;;;;;;;;;OAWG;IACH,eAAe,CAAC,EAAE,wBAAwB,CAAC;IAC3C;;;;;OAKG;IACH,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B;;;;OAIG;IACH,UAAU,CAAC,EAAE,MAAM,GAAG,KAAK,GAAG,MAAM,GAAG;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,CAAC;IACtE;;;;OAIG;IACH,MAAM,CAAC,EAAE,SAAS,CAAC;CACnB;AA4zBD,eAAO,MAAM,MAAM,EAAE,cAAc,CAAC,oBAAoB,EAAE,gBAAgB,CA6ZzE,CAAC;AA0FF,eAAO,MAAM,YAAY,EAAE,cAAc,CAAC,oBAAoB,EAAE,mBAAmB,CA2ClF,CAAC;AAsIF,wBAAgB,mCAAmC,CAClD,KAAK,EAAE,KAAK,CAAC,oBAAoB,CAAC,EAClC,OAAO,EAAE,OAAO,EAChB,OAAO,CAAC,EAAE,gBAAgB,GACxB,+BAA+B,CAkBjC"}

package/node_modules/@earendil-works/pi-ai/dist/api/anthropic-messages.js

@@ -1438,6 +1438,19 @@ function createClient(model, apiKey, interleavedThinking, useFineGrainedToolStre
     });
     return { client, isOAuthToken: false };
 }
+export function buildAnthropicWarmPromptCacheParams(model, context, options) {
+    const params = buildParams(model, context, false, {
+        ...options,
+        maxTokens: 0,
+        thinkingEnabled: undefined,
+        toolChoice: undefined,
+    });
+    const { stream: _stream, thinking: _thinking, output_config: _outputConfig, tool_choice: _toolChoice, ...nonStreaming } = params;
+    return sanitizeAnthropicToolPairs({
+        ...nonStreaming,
+        max_tokens: 0,
+    });
+}
 function buildParams(model, context, isOAuthToken, options, unsignedThinkingReplay = getAnthropicCompat(model).unsignedThinkingReplay) {
     const compat = getAnthropicCompat(model);
     const { cacheControl } = getCacheControl(model, options?.cacheRetention ?? model.cacheRetention, options?.env);