@linxiraos/pi-agent-core 1.1.19 → 1.1.22

package/CHANGELOG.md

@@ -2,9 +2,248 @@
 
 ## [Unreleased]
 
-## [1.1.19] - 2026-09-22
+## [1.1.22] - 2026-09-30
 
-- 版本线推进;本版无独立用户可见变化。
+- 随 1.1.22 版本线:CLI bin 命名(zeta-c)相关的用量文案与文档对齐。
+
+## [14.9.5] - 2026-05-12
+
+### Added
+
+- Added an `isError?: boolean` field on `AgentToolResult` so tools can flag a non-throwing failure (e.g. an aggregator that catches per-entry errors). `coerceToolResult` preserves the flag and the agent loop surfaces it as a tool error on the wire.
+
+## [14.9.3] - 2026-05-10
+
+### Added
+
+- Added `onHarmonyLeak` option on `Agent`/loop config to receive GPT-5 Harmony leak audit callbacks
+- Added harmony-leak detection and audit exports to the package index for programmatic leak detection and recovery hooks
+
+### Changed
+
+- Changed OpenAI Codex model runs to detect GPT-5 Harmony protocol leakage during streaming and automatically retry or recover tool calls instead of sending contaminated arguments downstream
+
+### Security
+
+- Hardened tool-call handling against leaked `to=functions.*` protocol tails by truncating or retrying before execution
+- Hardened failure handling so repeated GPT-5 Harmony leak mitigation is retried only up to two times before escalating to an explicit error
+
+## [14.9.0] - 2026-05-10
+
+### Added
+
+- Added `Agent#metadata` field forwarded to every API request; callers can set arbitrary provider metadata (e.g. `metadata.user_id`) once and have it applied to all subsequent stream calls without modifying per-call options
+- Added `Agent#setMetadataResolver(fn)` for installing a function that resolves request metadata at call time. The `metadata` getter dispatches through the resolver on every read (including the snapshot taken per `prompt()`), so callers reflect mutable external state (e.g. live OAuth account UUID after a token refresh) without manual re-syncs. Plain `agent.metadata = …` continues to set a static value and clears any installed resolver.
+- Added an `onSseEvent` agent option and loop config forwarding path for raw provider SSE diagnostics.
+
+## [14.7.6] - 2026-05-07
+
+### Added
+
+- Added `hideThinkingSummary` option/getter/setter on `Agent` and `AgentLoopConfig`. Forwarded to the underlying stream call so providers can omit reasoning/thinking summaries on demand.
+
+## [14.7.2] - 2026-05-06
+
+### Added
+
+- Added `loadMode` option to `AgentTool` to mark built-in tools as `essential` for initial loading or `discoverable` for search activation
+- Added optional `summary` field to `AgentTool` definitions for one-line text used in tool discovery indexes
+
+## [14.7.0] - 2026-05-04
+
+### Breaking Changes
+
+- Changed `Agent` API types so `systemPrompt` is now a list of prompt strings, requiring callers to pass and update system prompts via string arrays
+
+### Changed
+
+- Removed automatic project-context injection into each model call from loop logic
+
+### Removed
+
+- Removed the `projectPrompt` field from agent state/context and the `setProjectPrompt` mutator
+
+## [14.6.2] - 2026-05-03
+
+### Fixed
+
+- Fixed unhandled promise rejection when `getApiKey` or any other async error occurs during `streamAssistantResponse`: agent loop IIFEs now catch and route errors through `EventStream.fail()`, which terminates the `for await` loop and lets `Agent#runLoop`'s catch block create a proper error assistant message instead of crashing
+
+## [14.6.0] - 2026-05-02
+
+### Fixed
+
+- Fixed request cancellation before provider events by emitting an aborted assistant message and ending the stream with `stopReason: "aborted"`
+
+## [14.5.10] - 2026-04-30
+
+### Added
+
+- Added an `onResponse` stream option for observing provider response metadata after response headers arrive.
+
+## [14.2.0] - 2026-04-23
+
+### Changed
+
+- Changed tool dispatch to match model-returned tool calls by either internal tool name or custom wire name, enabling custom OpenAI tool names such as `apply_patch`.
+
+## [14.0.1] - 2026-04-08
+
+### Added
+
+- Added `onAssistantMessageEvent` callback option to inspect assistant streaming events before they are emitted, enabling abort decisions before buffered events continue flowing
+- Added `setAssistantMessageEventInterceptor()` method to dynamically set or update the assistant message event interceptor
+
+## [13.13.0] - 2026-03-18
+
+### Added
+
+- Added `startup.checkUpdate` setting, set to `true` by default, can be disabled to skip the update check on agent initialization
+
+## [13.12.7] - 2026-03-16
+
+### Added
+
+- Added overload for `prompt()` method accepting a string input with optional options parameter
+
+### Fixed
+
+- Fixed stale forced toolChoice being passed to provider after tools are refreshed mid-turn
+
+## [13.9.16] - 2026-03-10
+
+### Added
+
+- Added `onPayload` option to `AgentOptions` to inspect or replace provider payloads before they are sent
+
+## [13.9.3] - 2026-03-07
+
+### Added
+
+- Exported `ThinkingLevel` selector constants and types for configuring agent reasoning behavior
+- Added `inherit` thinking level option to defer reasoning configuration to higher-level selectors
+- Added `serviceTier` option to configure service tier for agent requests
+
+### Changed
+
+- Changed `thinkingLevel` from required string to optional `Effort` type, allowing undefined state
+- Updated `setThinkingLevel()` method to accept `Effort | undefined` instead of `ThinkingLevel` string
+
+## [13.4.0] - 2026-03-01
+
+### Added
+
+- Added `getToolChoice` option to dynamically override tool choice per LLM call
+
+## [13.3.8] - 2026-02-28
+
+### Changed
+
+- Changed intent field name from `agent__intent` to `_i` in tool schemas
+
+### Fixed
+
+- Fixed synthetic tool result text formatting so aborted/error tool results no longer emit `Tool execution was aborted.: Request was aborted` style punctuation.
+
+## [13.3.7] - 2026-02-27
+
+### Added
+
+- Added `lenientArgValidation` option to tools to allow graceful handling of argument validation errors by passing raw arguments to execute() instead of returning an error to the LLM
+
+## [13.3.1] - 2026-02-26
+
+### Added
+
+- Added `topP`, `topK`, `minP`, `presencePenalty`, and `repetitionPenalty` options to `AgentOptions` for fine-grained sampling control
+- Added getter and setter properties for sampling parameters on the `Agent` class to allow runtime configuration
+
+## [13.1.0] - 2026-02-23
+
+### Changed
+
+- Removed per-tool `agent__intent` field description from injected schema to reduce token usage; intent format is now documented once in the system prompt instead of repeated in every tool definition
+
+## [12.19.0] - 2026-02-22
+
+### Changed
+
+- Updated tool result messages to include error details when tool execution fails
+
+## [12.14.0] - 2026-02-19
+
+### Added
+
+- Added `intentTracing` option to enable intent goal extraction from tool calls, allowing models to specify high-level goals via a required `_intent` field that is automatically injected into tool schemas and stripped from arguments before execution
+
+## [12.11.0] - 2026-02-19
+
+### Added
+
+- Exported `AgentBusyError` exception class for handling concurrent agent operations
+
+### Changed
+
+- Agent now throws `AgentBusyError` instead of generic `Error` when attempting concurrent operations
+
+## [12.8.0] - 2026-02-16
+
+### Added
+
+- Added `transformToolCallArguments` option to `AgentOptions` and `AgentLoopConfig` for transforming tool call arguments before execution (e.g. secret deobfuscation)
+
+## [12.2.0] - 2026-02-13
+
+### Added
+
+- Added `providerSessionState` option to share provider state map for session-scoped transport and session caches
+- Added `preferWebsockets` option to hint that websocket transport should be preferred when supported by the provider implementation
+
+## [11.10.0] - 2026-02-10
+
+### Added
+
+- Added `temperature` option to `AgentOptions` to control LLM sampling temperature
+- Added `temperature` getter and setter to `Agent` class for runtime configuration
+
+## [11.6.0] - 2026-02-07
+
+### Added
+
+- Added `hasQueuedMessages()` method to check for pending steering/follow-up messages
+- Resume queued steering and follow-up messages from `continue()` after auto-compaction
+
+### Changed
+
+- Extracted `dequeueSteeringMessages()` and `dequeueFollowUpMessages()` from inline config callbacks
+- Added `skipInitialSteeringPoll` option to `_runLoop()` for correct queue resume ordering
+
+## [11.3.0] - 2026-02-06
+
+### Added
+
+- Added `maxRetryDelayMs` option to AgentOptions to cap server-requested retry delays, allowing higher-level retry logic to handle long waits with user visibility
+
+### Changed
+
+- Updated ThinkingLevel documentation to include support for gpt-5.3 and gpt-5.3-codex models with 'xhigh' thinking level
+
+## [11.2.0] - 2026-02-05
+
+### Fixed
+
+- Fixed handling of aborted requests to properly throw abort errors when stream terminates without a terminal event
+
+## [10.5.0] - 2026-02-04
+
+### Added
+
+- Added `concurrency` option to `AgentTool` to control tool scheduling: "shared" (default, runs in parallel) or "exclusive" (runs alone)
+- Implemented parallel execution of shared tools within a single agent turn for improved performance
+
+### Changed
+
+- Refactored tool execution to support concurrent scheduling with proper interrupt handling and steering message checks
 
 ## [9.2.2] - 2026-01-31
 
@@ -136,6 +375,10 @@
 
 Initial release under @oh-my-pi scope. See previous releases at [badlogic/pi-mono](https://github.com/badlogic/pi-mono).
 
+## [1.1.19] - 2026-09-22
+
+- 版本线推进;本版无独立用户可见变化。
+
 ## [1.1.18] - 2026-09-22
 
 - 包元数据:author/maintainer 更新为 Linxira-OS,LICENSE 追加 Linxira-OS 版权行(发行面变更)。
@@ -167,14 +410,14 @@ Initial release under @oh-my-pi scope. See previous releases at [badlogic/pi-mon
 ### Breaking Changes
 
 - **Queue API replaced with steer/followUp**: The `queueMessage()` method has been split into two methods with different delivery semantics ([#403](https://github.com/badlogic/pi-mono/issues/403)):
-  - `steer(msg)`: Interrupts the agent mid-run. Delivered after current tool execution, skips remaining tools.
-  - `followUp(msg)`: Waits until the agent finishes. Delivered only when there are no more tool calls or steering messages.
+   - `steer(msg)`: Interrupts the agent mid-run. Delivered after current tool execution, skips remaining tools.
+   - `followUp(msg)`: Waits until the agent finishes. Delivered only when there are no more tool calls or steering messages.
 - **Queue mode renamed**: `queueMode` option renamed to `steeringMode`. Added new `followUpMode` option. Both control whether messages are delivered one-at-a-time or all at once.
 - **AgentLoopConfig callbacks renamed**: `getQueuedMessages` split into `getSteeringMessages` and `getFollowUpMessages`.
 - **Agent methods renamed**:
-  - `queueMessage()` → `steer()` and `followUp()`
-  - `clearMessageQueue()` → `clearSteeringQueue()`, `clearFollowUpQueue()`, `clearAllQueues()`
-  - `setQueueMode()`/`getQueueMode()` → `setSteeringMode()`/`getSteeringMode()` and `setFollowUpMode()`/`getFollowUpMode()`
+   - `queueMessage()` → `steer()` and `followUp()`
+   - `clearMessageQueue()` → `clearSteeringQueue()`, `clearFollowUpQueue()`, `clearAllQueues()`
+   - `setQueueMode()`/`getQueueMode()` → `setSteeringMode()`/`getSteeringMode()` and `setFollowUpMode()`/`getFollowUpMode()`
 
 ### Fixed
 
@@ -186,9 +429,9 @@ Initial release under @oh-my-pi scope. See previous releases at [badlogic/pi-mon
 
 - **Transport abstraction removed**: `ProviderTransport`, `AppTransport`, and `AgentTransport` interface have been removed. Use the `streamFn` option directly for custom streaming implementations.
 - **Agent options renamed**:
-  - `transport` → removed (use `streamFn` instead)
-  - `messageTransformer` → `convertToLlm`
-  - `preprocessor` → `transformContext`
+   - `transport` → removed (use `streamFn` instead)
+   - `messageTransformer` → `convertToLlm`
+   - `preprocessor` → `transformContext`
 - **`AppMessage` renamed to `AgentMessage`**: All references to `AppMessage` have been renamed to `AgentMessage` for consistency.
 - **`CustomMessages` renamed to `CustomAgentMessages`**: The declaration merging interface has been renamed.
 - **`UserMessageWithAttachments` and `Attachment` types removed**: Attachment handling is now the responsibility of the `convertToLlm` function.

package/README.md

@@ -154,8 +154,9 @@ const agent = new Agent({
   // Dynamic model-scoped API key resolution (for expiring OAuth tokens)
   getApiKey: async (model) => tokenForModel(model),
 
-  // Tool execution context (late-bound UI/session access)
-  getToolContext: () => ({ /* app-defined */ }),
+  // Tool execution context (late-bound UI/session access). Surface the loop's
+  // passive-context sink so tools can call ctx.addAdditionalContext(...).
+  getToolContext: toolCall => ({ addAdditionalContext: toolCall?.addAdditionalContext /* app-defined */ }),
 });
 ```
 
@@ -461,7 +462,7 @@ const runCoverage = aggregateAgentRunCoverage(coverages);
 
 ### Tool status reporting
 
-`execute_tool` spans carry `pi.gen_ai.tool.status` ∈
+`execute_tool` spans carry `omp.gen_ai.tool.status` ∈
 `"ok" | "error" | "skipped" | "blocked" | "timeout" | "aborted"`.
 `beforeToolCall` blocks throw a distinguishable `ToolCallBlockedError`
 internally; the catch path reports `status: "blocked"` instead of conflating

package/THIRD-PARTY-NOTICES.txt

@@ -78,8 +78,8 @@ below.
 uutils coreutils (https://github.com/uutils/coreutils), MIT
 --------------------------------------------------------------------------------
 Covers: base32, base64, basename, cat, cksum (shared checksum machinery),
-b2sum, md5sum, sha1sum, sha224sum, sha256sum, sha384sum, sha512sum, comm, cut,
-date, dirname, head, hostname, ln, ls, mkdir, mktemp, mv, nproc, paste,
+b2sum, md5sum, sha1sum, sha224sum, sha256sum, sha384sum, sha512sum, comm, cp,
+cut, date, dirname, head, hostname, ln, ls, mkdir, mktemp, mv, nproc, paste,
 printenv, readlink, realpath, rm, seq, sort, stat, tac, tail, tee, touch, tr,
 truncate, uname, uniq, wc, whoami, yes.
 
@@ -293,7 +293,7 @@ MIT License
 Copyright (c) 2026 Sander Land
   (measured tokenizer vocabulary data, reconstruction model, and reference
   implementation: https://github.com/sanderland/ctok)
-Copyright (c) 2026 Can Bölük and the Oh My Pi contributors
+Copyright (c) 2026 Can Bölük and the omp contributors
 Copyright (c) 2026 Stencil Labs, Inc.
   (Rust implementation and the compact binary vocabulary encoding in
   crates/pi-natives/src/utok/claude)
@@ -335,43 +335,6 @@ This license allows the work and adaptations of it to be shared and used
 commercially, as long as it is attributed to Poppy Works. The font is bundled
 here (crates/pi-natives/src/fonts/Silver.ttf) as a CJK/Unicode bitmap fallback.
 
--------------------------------------------------------------------------------
-packages/utils/src/vendor/mermaid-ascii/NOTICE
-
-This directory contains an in-house Mermaid-diagram-to-ASCII renderer adapted
-from beautiful-mermaid (https://github.com/lukilabs/beautiful-mermaid), used
-under the MIT License.
-
-  Copyright (c) 2026 Craft Docs
-
-Only the ASCII rendering pipeline is ported (flowchart/state, sequence, class,
-ER, and xychart diagrams); the SVG renderer and its `elkjs` graph-layout
-dependency, the browser entry point, and the SVG theme/style modules were
-dropped. Terminal display width is reimplemented on `Bun.stringWidth`, and
-inline label formatting (HTML tags, markdown emphasis) is reduced to plain text
-for ASCII output. Layout and edge-routing logic is preserved faithfully so
-ASCII output matches the upstream package.
-
-MIT License
-
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
-
 -------------------------------------------------------------------------------
 packages/coding-agent/src/markit/NOTICE
 

package/dist/types/agent-loop.d.ts

@@ -5,7 +5,7 @@
 import { type AssistantMessage, type Context, EventStream, type ToolResultMessage } from "@linxiraos/pi-ai";
 import { type Dialect } from "@linxiraos/pi-ai/dialect";
 import { type AgentRunCoverage, type AgentRunSummary } from "./run-collector.js";
-import type { AgentContext, AgentEvent, AgentLoopConfig, AgentMessage, StreamFn } from "./types.js";
+import type { AgentContext, AgentEvent, AgentLoopConfig, AgentMessage, SteeringQueueState, StreamFn } from "./types.js";
 /** Stop-details marker for a provider error after assistant content/tool args already streamed. */
 export declare const STREAM_INTERRUPTED_AFTER_CONTENT_STOP_DETAIL = "stream_interrupted_after_content";
 /**
@@ -32,6 +32,12 @@ export declare function createToolScopedAbortReason(message: string, toolCallMes
  * boundary; this reason stops after persisting the completed tool batch.
  */
 export declare const TERMINAL_TOOL_RESULT_ABORT_REASON: unique symbol;
+/**
+ * Abort reason carried by an interruptible tool's signal when queued steering,
+ * a peer IRC, or a background completion cut it short. Lets a wait tell the
+ * designed wake path apart from an external/user abort of the run.
+ */
+export declare const TOOL_INTERRUPT_ABORT_REASON: unique symbol;
 export declare function resolveOwnedDialectFromEnv(value: string | undefined): Dialect | undefined;
 /**
  * Start an agent loop with a new prompt message.
@@ -100,6 +106,13 @@ export interface NormalizeToolsOptions {
     pruneDescriptions?: boolean;
 }
 export declare function normalizeTools(tools: AgentContext["tools"], options: NormalizeToolsOptions): Context["tools"];
+/**
+ * Classify the first `count` steering messages for a tool-batch interrupt:
+ * any user-authored message wins, then agent-attributed user messages, else
+ * system steering (advisor cards, hidden directives). Shared by the agent's
+ * queue peek and the loop's live-taken steering.
+ */
+export declare function steeringQueueState(messages: readonly AgentMessage[], count?: number): SteeringQueueState;
 /** Resolve the human-readable reason an abort carried. A caller that aborts via
  *  `AbortController.abort(reason)` with a string or a non-`AbortError` `Error`
  *  (e.g. the coding agent's user-interrupt label) gets that text surfaced on the

package/dist/types/agent.d.ts

@@ -159,6 +159,11 @@ export interface AgentOptions {
     pruneToolDescriptions?: boolean;
     /** Owned tool-calling dialect. Undefined keeps provider-native tool calling. */
     dialect?: Dialect;
+    /**
+     * Per-request owned-dialect resolver, consulted with the model being requested.
+     * Authoritative when set (like {@link serviceTierResolver}): replaces {@link dialect}.
+     */
+    dialectResolver?: (model: Model) => Dialect | undefined;
     /**
      * When owned tool calling is active and the model fabricates a tool result
      * mid-turn: `true` (default) aborts the provider request immediately; `false`
@@ -216,6 +221,8 @@ export interface AgentOptions {
      * tool-call arguments). See {@link AgentLoopConfig.transformAssistantMessage}.
      */
     transformAssistantMessage?: AgentLoopConfig["transformAssistantMessage"];
+    /** See {@link AgentLoopConfig.transformAssistantMessagePreservesToolCalls}. */
+    transformAssistantMessagePreservesToolCalls?: boolean;
     /**
      * Opt-in OpenTelemetry instrumentation. Passing `{}` enables the loop's
      * GenAI-semantic-convention spans using the global tracer provider. See
@@ -252,10 +259,17 @@ export declare class Agent {
      * UI emission, and tool dispatch. Reassign at any time to swap the implementation.
      */
     transformAssistantMessage?: AgentLoopConfig["transformAssistantMessage"];
+    /** Declares {@link transformAssistantMessage} never rewrites streamed tool calls; reassign alongside it. */
+    transformAssistantMessagePreservesToolCalls?: boolean;
     /**
      * Hook that peeks whether interrupting IRC asides are queued for the next boundary.
      */
     hasIrcInterrupts?: AgentLoopConfig["hasIrcInterrupts"];
+    /**
+     * Hook that peeks whether background completions (jobs, supervised processes)
+     * are queued for the next boundary.
+     */
+    hasBackgroundCompletions?: AgentLoopConfig["hasBackgroundCompletions"];
     constructor(opts?: AgentOptions);
     /**
      * Get the current session ID used for provider caching.
@@ -355,6 +369,15 @@ export declare class Agent {
     set serviceTierResolver(value: ((model: Model) => ServiceTier | undefined) | undefined);
     get hideThinkingSummary(): boolean | undefined;
     set hideThinkingSummary(value: boolean | undefined);
+    /** Strip tool descriptions from provider-bound specs; read per request. */
+    get pruneToolDescriptions(): boolean;
+    set pruneToolDescriptions(value: boolean);
+    /** Inject/strip the intent field on tool calls; applies from the next prompt run. */
+    get intentTracing(): boolean;
+    set intentTracing(value: boolean);
+    /** Abort the provider request on a fabricated tool result; applies from the next prompt run. */
+    get abortOnFabricatedToolResult(): boolean | undefined;
+    set abortOnFabricatedToolResult(value: boolean | undefined);
     /**
      * Get the current max retry delay in milliseconds.
      */
@@ -392,6 +415,14 @@ export declare class Agent {
     addBeforeQueuedMessageDequeueHook(hook: (signal?: AbortSignal) => Promise<void> | void): () => void;
     /** Register an independently removable hook that runs immediately before each model call. */
     addBeforeModelCallHook(hook: (signal?: AbortSignal) => Promise<void> | void): () => void;
+    /**
+     * Take back live-steered messages ahead of an abort (Esc restores them to the editor):
+     * the aborted run then neither records nor requeues them.
+     */
+    withdrawLiveSteering(): AgentMessage[];
+    /** Steering live steering took for the streaming response; the transcript records it once
+     *  that response (or its tool batch) ends, which is when the model switches to it. */
+    peekLiveSteeredMessages(): AgentMessage[];
     setProviderResponseInterceptor(fn: SimpleStreamOptions["onResponse"] | undefined): void;
     setRawSseEventInterceptor(fn: SimpleStreamOptions["onSseEvent"] | undefined): void;
     setAssistantMessageEventInterceptor(fn: ((message: AssistantMessage, event: AssistantMessageEvent) => void) | undefined): void;
@@ -457,6 +488,11 @@ export declare class Agent {
      *  single source of truth. Includes exclusively claimed originals while
      *  preparation is pending, so editor restoration can cancel their delivery. */
     peekSteeringQueue(): readonly AgentMessage[];
+    /** Dequeued messages not yet in the transcript, e.g. steering a provider
+     *  took into its in-flight response via live steering. Aborting the run
+     *  requeues them, so the session's empty-submit interrupt counts them as
+     *  pending input even though {@link peekSteeringQueue} no longer does. */
+    peekUndeliveredQueuedMessages(): AgentMessage[];
     /** Non-consuming view of the pending follow-up queue. See
      *  {@link peekSteeringQueue}. */
     peekFollowUpQueue(): readonly AgentMessage[];

package/dist/types/compaction/anthropic.d.ts

@@ -1,33 +1,21 @@
 /**
- * Anthropic server-side compaction (`compact-2026-01-12` beta).
+ * Anthropic on-demand compaction (`compact-2026-09-04` beta).
  *
- * The compaction request is the live turn's own request shape — same system
- * prompt, tools, and message history — plus the `compact_20260112` edit with
- * `pause_after_compaction`. The API summarizes the prompt from the already
- * cached prefix and stops; the summary arrives as a `compaction` block that
- * the provider surfaces as an `anthropicCompaction` payload. The summary is
- * plain text, so it doubles as the compaction entry's readable summary for
- * every other provider, while the Anthropic provider replays it as a native
- * block (the API drops everything that precedes it). The retained tail after
- * the cut point is replayed from session entries exactly like a local summary.
+ * The request sends only the prefix to summarize, with the live conversation's
+ * system prompt, tools and thinking settings. The returned signed block
+ * replaces that prefix; the retained tail is replayed from session entries.
  */
 import type { AnthropicCompactionPayload, ApiKey, Effort, Message, Model, SimpleStreamOptions, Tool, Usage } from "@linxiraos/pi-ai";
 import { type InstrumentedChatSpanOptions } from "../telemetry.js";
+import type { AgentMessage } from "../types.js";
 export declare const ANTHROPIC_COMPACTION_PRESERVE_KEY = "anthropicCompaction";
-/** The API rejects a `compact_20260112` trigger below this many input tokens. */
-export declare const ANTHROPIC_COMPACTION_MIN_TRIGGER_TOKENS = 50000;
-/**
- * Smallest context the native lane accepts. The trigger sits at the API
- * floor, so a prompt that lands below it is answered instead of compacted;
- * the margin over the floor absorbs the difference between the last reported
- * context size and the compaction request's own input.
- */
-export declare const ANTHROPIC_COMPACTION_MIN_CONTEXT_TOKENS = 55000;
 /** Summary persisted under {@link ANTHROPIC_COMPACTION_PRESERVE_KEY}. */
 export interface AnthropicCompactionPreserveData {
     provider: string;
     content: string;
-    /** Opaque provider state the API attached to the block; replayed verbatim. */
+    /** Signature attached to an on-demand block; replayed verbatim. */
+    signature?: string;
+    /** Legacy threshold block state; replay-only. */
     encryptedContent?: string;
     /** Harness file metadata (`<files>` section) replayed after the native block. */
     filesText?: string;
@@ -51,33 +39,17 @@ export declare function withAnthropicCompactionPreserveData(preserveData: Record
 /** Replay payload for a compaction summary the active model produced natively. */
 export declare function getAnthropicCompactionPayload(preserveData: Record<string, unknown> | undefined): AnthropicCompactionPayload | undefined;
 /**
- * The retained tail as the model will see it, for the summarization
- * instructions: how many of the conversation's final wire messages stay in
- * context verbatim, and the role of the first. The compaction request carries
- * the whole conversation so the prompt cache the live turn wrote is read, but
- * the summary must cover only the history before that tail — the local
- * summarizer never sees the tail, and the rebuilt context replays it after the
- * summary. Counting mirrors the provider's message conversion (consecutive
- * tool results collapse into one user message; developer messages are user
- * messages). Structured for the prompt template, which renders the
- * singular/plural wording; the description quotes no content: quoting the
- * tail would hand the summarizer the very facts it must leave to the tail.
+ * Move the existing keep-tail boundary forward until the summary/tail boundary
+ * alternates wire roles and no tool call is separated from its result. If no
+ * boundary is safe, the request summarizes the entire snapshot (empty tail).
  */
-export interface RetainedTailScope {
-    count: number;
-    role: "assistant" | "user";
-}
-export declare function describeRetainedTail(messages: readonly Message[]): RetainedTailScope | undefined;
-/**
- * Summarization prompt sent as the edit's `instructions`, which replace the
- * API default entirely. The template lays out the retained-tail boundary
- * first, so the summary covers only the history the rebuilt context drops,
- * then the caller's extra context, the same structure prompt as the local
- * summarizer, the caller's focus, and the tool-abstention clause the API
- * recommends when tools are defined (a summarization pass that calls a tool
- * yields no summary).
- */
-export declare function buildAnthropicCompactionInstructions(basePrompt: string, customInstructions: string | undefined, extraContext: string | undefined, retainedTail: RetainedTailScope | undefined): string;
+export declare function findAnthropicCompactionCut(messages: readonly (AgentMessage | {
+    role: "system";
+    content: string;
+    timestamp: number;
+})[], initialCut: number): number;
+/** Instructions replace the API default; the request contains only summarized messages. */
+export declare function buildAnthropicCompactionInstructions(basePrompt: string, customInstructions: string | undefined, extraContext: string | undefined): string;
 export interface AnthropicNativeCompactionRequest {
     systemPrompt: string[];
     messages: Message[];
@@ -88,21 +60,19 @@ export interface AnthropicNativeCompactionRequest {
 }
 export interface AnthropicNativeCompactionResponse {
     content: string;
-    encryptedContent?: string;
+    signature: string;
     usage: Usage;
     model: string;
 }
 export interface AnthropicNativeCompactionOptions extends Pick<SimpleStreamOptions, "initiatorOverride" | "metadata" | "fetch" | "sessionId" | "promptCacheKey" | "providerSessionState" | "maxInFlightRequests">, Pick<InstrumentedChatSpanOptions, "completeImpl" | "telemetry" | "retry"> {
 }
 /**
- * Run one compaction request and return the summary the API wrote, with the
- * opaque `encrypted_content` the API attached for the replay. `completeSimple`
+ * Run one compaction request and return the summary and signature the API wrote. `completeSimple`
  * resolves terminal failures as messages, so their classification is restored
  * here: an aborted response is an `AbortError` (a cancellation, never a native
  * failure) and an error response keeps its HTTP status, so auth and timeout
  * handling downstream classify it the same way as the OpenAI lanes. A response
- * without a summary is a native failure — the API answers the prompt instead
- * when its input never reached the trigger, and returns an empty block when
- * the model called a tool during summarization.
+ * without a summary is a native failure, including tool use, refusals and
+ * output limits; the configured method order can then choose a fallback.
  */
 export declare function requestAnthropicNativeCompaction(model: Model<"anthropic-messages">, apiKey: ApiKey, request: AnthropicNativeCompactionRequest, signal: AbortSignal | undefined, options: AnthropicNativeCompactionOptions): Promise<AnthropicNativeCompactionResponse>;

package/dist/types/compaction/branch-summarization.d.ts

@@ -55,7 +55,7 @@ export interface GenerateBranchSummaryOptions {
     convertToLlm?: ConvertToLlm;
     /**
      * Optional telemetry handle. When provided, the branch summary LLM call is
-     * wrapped in an OTEL chat span tagged with `pi.gen_ai.oneshot.kind = "branch_summary"`.
+     * wrapped in an OTEL chat span tagged with `omp.gen_ai.oneshot.kind = "branch_summary"`.
      */
     telemetry?: AgentTelemetry;
     /**

package/dist/types/compaction/compaction.d.ts

@@ -180,7 +180,7 @@ export interface SummaryOptions {
     /**
      * Optional telemetry handle. When provided, every LLM call emitted during
      * compaction is wrapped in an OTEL chat span tagged with
-     * `pi.gen_ai.oneshot.kind` (`compaction_summary`, `compaction_short_summary`,
+     * `omp.gen_ai.oneshot.kind` (`compaction_summary`, `compaction_short_summary`,
      * or `compaction_turn_prefix`). `undefined` keeps the call paths zero-cost.
      */
     telemetry?: AgentTelemetry;
@@ -243,7 +243,7 @@ export interface HandoffOptions {
     metadata?: Record<string, unknown>;
     /**
      * Optional telemetry handle. When provided, the handoff LLM call is
-     * wrapped in an OTEL chat span tagged with `pi.gen_ai.oneshot.kind = "handoff"`.
+     * wrapped in an OTEL chat span tagged with `omp.gen_ai.oneshot.kind = "handoff"`.
      */
     telemetry?: AgentTelemetry;
     /**
@@ -297,6 +297,8 @@ export interface CompactionPreparation {
     turnPrefixMessages: AgentMessage[];
     /** Messages kept in full after compaction (recent history) */
     recentMessages: AgentMessage[];
+    /** Entry IDs parallel to recentMessages, for an Anthropic-safe keep-tail boundary. */
+    recentEntryIds?: string[];
     /** Whether this is a split turn (cut point in middle of turn) */
     isSplitTurn: boolean;
     tokensBefore: number;