@linxiraos/pi-ai 1.1.14 → 1.1.15

package/CHANGELOG.md

@@ -2,79 +2,7 @@
 
 ## [Unreleased]
 
-## [1.1.14] - 2026-09-12
+## [1.1.15] - 2026-09-16
 
-- 随 1.1.14 版本线发布:bazel 构建面(crates/*/BUILD.bazel)版本号纳入一致性检查,CI 原生构建与桌面冒烟守卫修复。
+- 上游 v18.1.17–v18.1.21 同步:provider 修复与补全回退链强化(保序、预算、同模型区分)随同步带过。
 
-## [1.1.13] - 2026-09-10
-
-- 上游 v18.1.16 同步:AuthStorage 合并封锁契约——凭据级 `blockedUntilMs` 与 `providerTimed` 时间线合并,先到的更长封锁在短提示到来时保持有效;GitHub Copilot OAuth 拆分公共 GitHub / GHE 双 client-id;Codex WebSocket 传输 abort 携带 cause 链。
-
-## [1.1.12] - 2026-09-10
-
-- 品牌与合并工具链维护版本;无本包用户可见变更。
-
-## [1.1.11] - 2026-09-08
-
-### Added
-
-- Muse Code subscription sign-in, credential refresh, inference, and quota reporting in `/usage`, with durable rate-limit backoff so quota refresh recovers instead of repeatedly retrying.
-
-## [1.1.10-omp18.1.12] - 2026-09-06
-
-### Added
-
-- Added Muse Code subscription sign-in, credential refresh, inference, and quota reporting in `/usage`, with durable rate-limit backoff so quota refresh recovers instead of repeatedly retrying.
-
-## [1.1.10-omp18.1.11] - 2026-09-05
-
-### Fixed
-
-- Fixed OpenCode Go usage polls (`GET /zen/go/v1/usage`) missing `x-opencode-session` and omp's `User-Agent`: background polls now attribute with the stable install id so the requests OpenCode flags as `Bun fetch` carry the required session header.
-- GitHub Copilot sign-in now requests only basic profile access, restoring login for Enterprise organizations that reject repository, gist, and Codespaces permissions ([#10656](https://github.com/can1357/oh-my-pi/issues/10656)).
-
-## [1.1.10] - 2026-09-07
-
-- GitHub Copilot sign-in now requests only basic profile access, restoring login for Enterprise organizations that reject repository, gist, and Codespaces permissions.
-- Transient gateway stream failures are now retried instead of surfacing as session errors.
-
-## [1.1.9] - 2026-09-05
-
-- Z.ai OAuth key name sends zeta (merge restored the upstream oh-my-pi literal in tests); xAI/OpenAI-compatible requests send the zeta User-Agent again.
-
-## [1.1.6] - 2026-08-30
-
-- 同步上游 OMP v18.0.9（`cc14e04f075d`）。
-
-## [1.1.5] - 2026-08-26
-
-- 同步上游 OMP v18.0.5 / v18.0.6：新增 Yolo-Auto / OpenRouter 浏览器登录与 DeepInfra 支持，空补全重试重构（withReplaySafeStreamRetry）。
-
-## [1.1.2] - 2026-08-25
-
-### Fixed
-
-- Republished as 1.1.2 to reset the `latest` tag after the broken 1.1.0 (no functional change over 1.1.1).
-
-## [1.1.1] - 2026-08-25
-
-### Fixed
-
-- Published tarballs now carry real dependency versions instead of Bun's `catalog:` protocol (1.1.0 installs failed with "Unsupported URL Type catalog:").
-
-## [1.1.0] - 2026-08-25
-
-### Changed
-
-- 同步上游 OMP v18.0.3 / v18.0.4（内部运行时与构建改进，无独立用户可见变更）。
-
-## [1.0.1] - 2026-08-14
-
-- Reset the version to 1.0.0 and republished under the `@linxiraos/*` scope, breaking from the `@linxiraos` version lineage.
-
-## [1.0.0] - 2026-08-13
-
-### Changed
-
-- Reset the version to 1.0.0 and republished under the `@linxiraos/*` scope, breaking from the `@linxiraos` version lineage.
-- Fixed Gemini thought summaries occasionally leaking a raw `` ```thinking `` / `` ``````thinking `` fence delimiter into the reasoning block, so it no longer shows up as fence spam in the thinking display or persisted transcripts ([#8719](https://github.com/can1357/oh-my-pi/issues/8719)).

package/dist/types/error/flags.d.ts

@@ -28,6 +28,12 @@ export declare const Flag: {
 };
 export type Flag = (typeof Flag)[keyof typeof Flag];
 export declare const STREAM_READ_ERROR_PATTERN: RegExp;
+/** Python h2/httpx diagnostics forwarded through provider or proxy error events. */
+export declare const PYTHON_HTTP2_STREAM_RESET_PATTERN: RegExp;
+/** Python h11/httpx EOF while reading an HTTP/1.1 chunked response body. */
+export declare const PYTHON_HTTP_INCOMPLETE_CHUNK_PATTERN: RegExp;
+/** reqwest body-frame failures forwarded by the Codex HTTP proxy. */
+export declare const CODEX_HTTP_BODY_READ_ERROR_PATTERN: RegExp;
 export declare const TRANSIENT_TRANSPORT_PATTERN: RegExp;
 /**
  * Local llama.cpp / Ollama deterministic tool-call argument JSON parse failure.

package/dist/types/providers/anthropic-wire.d.ts

@@ -135,7 +135,21 @@ export type FallbackBlockParam = {
         model: string;
     };
 };
-export type ContentBlockParam = TextBlockParam | ImageBlockParam | ToolUseBlockParam | ToolResultBlockParam | ServerToolUseBlockParam | WebSearchToolResultBlockParam | ToolSearchToolResultBlockParam | ToolAdditionBlockParam | ToolRemovalBlockParam | ThinkingBlockParam | RedactedThinkingBlockParam | FallbackBlockParam;
+/** Beta enabling server-side compaction (`compact_20260112` edit, `compaction` blocks). */
+export declare const COMPACTION_BETA = "compact-2026-01-12";
+/**
+ * Server-side compaction summary (compact-2026-01-12). Returned at the start
+ * of the assistant response that crossed the trigger; on replay the API drops
+ * every block that precedes it, so it may open the messages array. The
+ * `encrypted_content` is opaque provider state, round-tripped verbatim.
+ */
+export type CompactionBlockParam = {
+    type: "compaction";
+    content: string;
+    encrypted_content?: string | null;
+    cache_control?: CacheControlEphemeral | null;
+};
+export type ContentBlockParam = TextBlockParam | ImageBlockParam | ToolUseBlockParam | ToolResultBlockParam | ServerToolUseBlockParam | WebSearchToolResultBlockParam | ToolSearchToolResultBlockParam | ToolAdditionBlockParam | ToolRemovalBlockParam | ThinkingBlockParam | RedactedThinkingBlockParam | FallbackBlockParam | CompactionBlockParam;
 /**
  * A single conversation turn.
  *
@@ -229,12 +243,24 @@ export type FallbackParam = {
     output_config?: OutputConfig;
     speed?: "fast";
 };
+/** Server-side compaction edit (compact-2026-01-12). */
+export type CompactionEdit = {
+    type: "compact_20260112";
+    /** `input_tokens` is the only trigger; `value` must be at least 50,000. */
+    trigger?: {
+        type: "input_tokens";
+        value: number;
+    };
+    pause_after_compaction?: boolean;
+    /** Replaces the API's default summarization prompt entirely. */
+    instructions?: string;
+};
 /** Claude Code context-management beta payload. */
 export type ContextManagement = {
     edits: Array<{
         type: "clear_thinking_20251015";
         keep: "all";
-    }>;
+    } | CompactionEdit>;
 };
 export type MessageCreateParams = {
     model: string;
@@ -267,7 +293,7 @@ export type MessageCreateParams = {
 export type MessageCreateParamsStreaming = MessageCreateParams & {
     stream: true;
 };
-export type StopReason = "end_turn" | "max_tokens" | "stop_sequence" | "tool_use" | "pause_turn" | "refusal" | "sensitive" | "model_context_window_exceeded";
+export type StopReason = "end_turn" | "max_tokens" | "stop_sequence" | "tool_use" | "pause_turn" | "refusal" | "sensitive" | "model_context_window_exceeded" | "compaction";
 export type CacheCreation = {
     ephemeral_5m_input_tokens?: number | null;
     ephemeral_1h_input_tokens?: number | null;
@@ -278,12 +304,15 @@ export type ServerToolUsage = {
 };
 /**
  * Per-attempt token accounting inside a multi-run turn
- * (server-side-fallback-2026-06-01). Populated whenever a fallback chain
- * ran, including sticky-served turns with no `fallback` content block.
- * A `fallback_message` entry is the definitive "served by fallback" signal.
+ * (server-side-fallback-2026-06-01, compact-2026-01-12). Populated whenever
+ * a fallback chain ran, including sticky-served turns with no `fallback`
+ * content block, and whenever the compaction beta is active. A
+ * `fallback_message` entry is the definitive "served by fallback" signal; a
+ * `compaction` entry is the summarization sampling the top-level usage
+ * excludes.
  */
 export type UsageIteration = {
-    type?: "message" | "fallback_message" | string;
+    type?: "message" | "fallback_message" | "compaction" | string;
     model?: string | null;
     input_tokens?: number | null;
     output_tokens?: number | null;
@@ -343,6 +372,10 @@ export type ResponseContentBlock = {
     to: {
         model: string;
     };
+} | {
+    type: "compaction";
+    content?: string | null;
+    encrypted_content?: string | null;
 };
 export type ContentBlockDelta = {
     type: "text_delta";
@@ -356,6 +389,10 @@ export type ContentBlockDelta = {
 } | {
     type: "signature_delta";
     signature: string;
+} | {
+    type: "compaction_delta";
+    content?: string | null;
+    encrypted_content?: string | null;
 };
 export type StopDetails = {
     type: string;

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

@@ -165,6 +165,14 @@ export type AnthropicClientOptionsArgs = {
     fetch?: FetchImpl;
     maxRetryDelayMs?: number;
     sessionId?: string;
+    /** Working-identity cache key for this credential+host; undefined off the Copilot path. */
+    copilotCacheKey?: string;
+    /**
+     * Build-time cache provenance for the wrapper: the cached value the
+     * outgoing headers were built from, or `null` when the cache was empty at
+     * build. `undefined` rereads the cache at dispatch.
+     */
+    copilotCacheSnapshot?: string | null;
 };
 export type AnthropicClientOptionsResult = {
     isOAuthToken: boolean;
@@ -204,6 +212,32 @@ export type AnthropicUsageLike = {
  * zero-valued objects clear prior extras from earlier stream usage snapshots.
  */
 export declare function applyAnthropicUsageExtras(usage: Usage, source: AnthropicUsageLike): void;
+/**
+ * Whether this model's requests reach the official Anthropic API, resolved the
+ * way the transport resolves it — including the Foundry and
+ * `ANTHROPIC_BASE_URL` reroutes that leave `compat.officialEndpoint` stale.
+ */
+export declare function resolvesToOfficialAnthropicEndpoint(model: Model<"anthropic-messages">): boolean;
+/**
+ * Whether server-side compaction (`compact-2026-01-12`) may be spoken for
+ * this model to the endpoint a request actually reaches: a model line the
+ * beta supports (`compat.supportsServerCompaction`, rule-owned in the
+ * catalog), on the official API for the first-party provider or on any
+ * endpoint that opted in through `remoteCompaction.enabled`, and never on one
+ * whose deployment contract excludes context management. The same predicate
+ * gates emitting the edit, attaching the beta, and replaying a persisted
+ * block, so a route or model change can never leave a session sending a block
+ * its endpoint rejects.
+ */
+export declare function supportsAnthropicCompaction(model: Model<"anthropic-messages">, effectiveBaseUrl?: string): boolean;
+/**
+ * {@link supportsAnthropicCompaction} for a request on a caller-owned client:
+ * the endpoint is whatever the client targets (an `AnthropicVertex` client
+ * carries an Anthropic model to Vertex), never the model's own routing. SDK
+ * clients expose it as `baseURL`; a client that exposes no endpoint only
+ * compacts through an explicit `remoteCompaction.enabled` opt-in.
+ */
+export declare function supportsAnthropicCompactionOnClient(model: Model<"anthropic-messages">, client: AnthropicMessagesClientLike): boolean;
 /** Detects the preserved-thinking error caused by rewriting a signed block's conversation prefix. */
 export declare function isThinkingPrefixBindingError(message: string): boolean;
 export declare function isInvalidThinkingSignatureError(message: string): boolean;
@@ -250,9 +284,17 @@ export type AnthropicMessageParam = MessageParam;
  * `fallback` content block from a prior turn be replayed on the wire;
  * otherwise the block is dropped to avoid a 400 on non-fallback requests
  * that don't send the beta.
+ *
+ * `opts.replayCompaction` — replay a user-role compaction summary that
+ * carries an {@link AnthropicCompactionPayload} from this provider as a
+ * native `compaction` block instead of its text. The API drops every block
+ * before the compaction block, so the assistant turn carrying it may open
+ * the conversation; the request must send the compaction beta (the stream
+ * entry point adds it whenever such a payload is present).
  */
 export declare function convertAnthropicMessages(messages: Message[], model: Model<"anthropic-messages">, isOAuthToken: boolean, opts?: {
     serverSideFallbackEnabled?: boolean;
+    replayCompaction?: boolean;
     dropAllThinking?: boolean;
     droppedThinkingBlocks?: ReadonlySet<string>;
 }): AnthropicMessageParam[];

package/dist/types/providers/github-copilot-headers.d.ts

@@ -1,4 +1,4 @@
-import type { Message } from "../types.js";
+import type { FetchImpl, Message } from "../types.js";
 /**
  * Infer whether the current request to Copilot is user-initiated or agent-initiated.
  * Accepts `unknown[]` because providers may pass pre-converted message shapes.
@@ -11,6 +11,70 @@ export type CopilotDynamicHeaders = {
     premiumRequests: CopilotPremiumRequests;
 };
 export declare function resolveGitHubCopilotBaseUrl(baseUrl: string | undefined, apiKey: string | undefined): string | undefined;
+/**
+ * Opt-in `Copilot-Integration-Id` override for chat and model-policy requests.
+ * Reads `COPILOT_INTEGRATION_ID`; unset/invalid keeps the chat-surface default
+ * (`COPILOT_CHAT_INTEGRATION_ID`). Model discovery keeps the CLI identity: it
+ * unlocks enterprise/experimental models and listing models is not
+ * policy-gated the way chat completions are (#11372).
+ */
+export declare function resolveCopilotIntegrationIdOverride(env?: Record<string, string | undefined>): string | undefined;
+/**
+ * Effective identity before the chat-surface default: explicit value, then
+ * request headers, then `COPILOT_INTEGRATION_ID`. Pure given its inputs, so
+ * tests inject literals instead of mutating process state.
+ */
+export declare function resolveCopilotRequestIdentity(headers?: Record<string, string>, explicit?: unknown, env?: Record<string, string | undefined>): string | undefined;
+/**
+ * Stable cache key for a raw Copilot API key envelope on one effective host.
+ * Hashes the bearer with `Bun.hash` (repo-approved hashing API; same
+ * credential-scoped pattern as the GitLab Duo and Codex account keys) so token
+ * bytes never sit in the map as keys; enterprise/business routing inputs and
+ * the normalized effective base URL participate so the same token on two hosts
+ * does not share an entry.
+ */
+export declare function getCopilotIntegrationCacheKey(apiKeyRaw: string | undefined, baseUrl?: string): string | undefined;
+/** Cached working identity for a cache key, if one was learned. */
+export declare function getCachedCopilotIntegrationId(cacheKey: string | undefined): string | undefined;
+/**
+ * Remember the identity that cleared the identity gate for a credential.
+ * Every store refreshes recency, so hot credentials survive eviction.
+ */
+export declare function rememberCopilotWorkingIntegrationId(cacheKey: string | undefined, integrationId: unknown): void;
+/** Clear one cached identity, or the whole cache when no key is given. */
+export declare function clearCopilotIntegrationCache(cacheKey?: string): void;
+/**
+ * Reissue Copilot client-identity denials once with the other surface.
+ *
+ * Chat is the default surface (`COPILOT_CHAT_INTEGRATION_ID`) because Business
+ * organizations that gate premium models per client surface commonly allow
+ * chat while blocking CLI/agentic clients (issue #11372). Other Business and
+ * Enterprise orgs do the opposite and reject the chat identity — as an HTTP 403
+ * or, on `api.business.githubcopilot.com`, an HTTP 400 `model_not_supported`
+ * (issue #11669). Both denials retry once as the CLI. The retry fires only for
+ * requests carrying the chat default and only when the caller resolved no
+ * explicit identity — an explicit choice is never second-guessed. The denied
+ * body is drained before reissuing, and the retry carries the CLI identity so
+ * the guard passes it through: at most two requests, never a loop.
+ *
+ * When `cacheKey` is set, a 2xx retry remembers its identity via
+ * `rememberCopilotWorkingIntegrationId`, so later streams for the same
+ * credential start at the working shape. A cached CLI start that is itself
+ * denied (stale after an org-policy flip) retries once as chat and relearns.
+ * Only a 2xx retry proves its identity — 401s deny every identity equally and
+ * 408/429/5xx are transport-retryable (the transport resends the *original*
+ * headers), so those must never be recorded as working. Any non-2xx retry
+ * clears the entry instead: the next stream rediscovers rather than pinning a
+ * shape that just failed.
+ */
+export declare function wrapFetchForCopilotFallback(base: FetchImpl | undefined, enabled: boolean, integrationId?: unknown, cacheKey?: string, 
+/**
+ * Build-time cache provenance: the exact cached value the outgoing headers
+ * were built from. `undefined` rereads the cache at dispatch (direct
+ * callers); `null` pins "cache was empty at build" so a sibling learning
+ * mid-flight cannot change this request's retry decision.
+ */
+cacheSnapshot?: string | null): FetchImpl;
 export declare function inferCopilotInitiator(messages: unknown[]): CopilotInitiator;
 /** Check whether any message in the conversation contains image content. */
 export declare function hasCopilotVisionInput(messages: Message[]): boolean;
@@ -37,4 +101,10 @@ export declare function buildCopilotDynamicHeaders(params: {
     headers?: Record<string, string>;
     initiatorOverride?: CopilotInitiator;
     planTier?: string;
+    /** Enterprise login domain; Enterprise keeps the CLI identity that its private endpoint accepts. */
+    enterpriseUrl?: string;
+    /** Raw explicit identity; validated here, chat default when absent/invalid. */
+    integrationId?: unknown;
+    /** Learned working identity for this credential; explicit still wins, then this, then the defaults. */
+    cachedIntegrationId?: unknown;
 }): CopilotDynamicHeaders;

package/dist/types/providers/openai-completions.d.ts

@@ -42,5 +42,5 @@ export interface OpenAICompletionsOptions extends StreamOptions {
  * assistant output commits the attempt.
  */
 export declare const streamOpenAICompletions: StreamFunction<"openai-completions">;
-export declare function parseChunkUsage(rawUsage: object, model: Model<"openai-completions">, premiumRequests: number | undefined): AssistantMessage["usage"];
+export declare function parseChunkUsage(rawUsage: object, model: Model<"openai-completions">, premiumRequests: number | undefined, timestamp?: number): AssistantMessage["usage"];
 export declare function convertMessages(model: Model<"openai-completions">, context: Context, compat: ResolvedOpenAICompat): ChatCompletionMessageParam[];

package/dist/types/providers/openai-shared.d.ts

@@ -65,6 +65,14 @@ export interface OpenAIRequestSetup {
     headers: Record<string, string>;
     query: Record<string, string> | undefined;
     requestHeaders: Record<string, string>;
+    /** Working-identity cache key for this credential+host; undefined off the Copilot path. */
+    copilotCacheKey: string | undefined;
+    /**
+     * Build-time cache provenance for the wrapper: the cached value the
+     * outgoing headers were built from, or `null` when the cache was empty at
+     * build. `undefined` off the Copilot path (wrapper rereads at dispatch).
+     */
+    copilotCacheSnapshot: string | null | undefined;
 }
 export declare function resolveOpenAIRequestSetup(model: OpenAIRequestSetupModel, options: OpenAIRequestSetupOptions): OpenAIRequestSetup;
 export declare function applyOpenAIServiceTier(params: {

package/dist/types/providers/vision-guard.d.ts

@@ -1,4 +1,4 @@
-import type { ImageContent, Model, TextContent } from "../types.js";
+import type { Api, ImageContent, Model, TextContent } from "../types.js";
 export declare const NON_VISION_IMAGE_PLACEHOLDER = "[image omitted: model does not support vision]";
 export declare function partitionVisionContent(content: ReadonlyArray<TextContent | ImageContent>, supportsImages: boolean): {
     textBlocks: TextContent[];
@@ -12,4 +12,18 @@ export declare function joinTextWithImagePlaceholder(text: string, omittedImages
  * misconfigured provider descriptors or user model entries (e.g. text-only
  * DashScope Qwen SKUs, DeepSeek models) whose endpoints reject `image_url`.
  */
-export declare function isOpenAICompletionsVisionSupported(model: Model<"openai-completions">): boolean;
+export declare function isOpenAICompletionsVisionSupported(model: Model<"openai-completions" | "openrouter">): boolean;
+/**
+ * Whether the transport that will carry `model` sends image content on the wire.
+ *
+ * The `pi-native` transport forwards the original context (images included) to
+ * the gateway, which resolves its own model server-side, so the Chat
+ * Completions guard below never runs client-side and the declared input
+ * applies. Otherwise the OpenAI Chat Completions path applies the text-only
+ * guard, as does the OpenRouter chat fallback (`PI_OPENROUTER_RESPONSES=0`,
+ * which dispatches `openrouter` models through `streamOpenAICompletions`);
+ * every other API ships the modalities the model declares. Callers that report
+ * or gate on the wire (for example the `omp models` table) read this
+ * predicate; declared capability reads `model.input`.
+ */
+export declare function sendsImageInputOnWire(model: Model<Api>): boolean;

package/dist/types/registry/oauth/github-copilot.d.ts

@@ -8,6 +8,7 @@ type GitHubCopilotLoginOptions = {
         allowEmpty?: boolean;
     }) => Promise<string>;
     onProgress?: (message: string) => void;
+    copilotIntegrationId?: unknown;
     signal?: AbortSignal;
     pollIntervalFloorMs?: number;
     pollIntervalScaleMs?: number;

package/dist/types/registry/oauth/muse-code.d.ts

@@ -6,8 +6,8 @@ declare const museCodeKeyResponseSchema: import("@linxiraos/pi-omptype").FluentT
     is_subs_active?: boolean | undefined;
     require_payment?: boolean | undefined;
     require_payment_action_url?: string | undefined;
-    subs_tier_id?: string | undefined;
-    subs_tier_name?: string | undefined;
+    subs_tier_id?: string | null | undefined;
+    subs_tier_name?: string | null | undefined;
     subs_usage?: {
         weekly?: {
             resets_at?: string | number | undefined;
@@ -28,8 +28,8 @@ declare const museCodeKeyResponseSchema: import("@linxiraos/pi-omptype").FluentT
     is_subs_active?: boolean | undefined;
     require_payment?: boolean | undefined;
     require_payment_action_url?: string | undefined;
-    subs_tier_id?: string | undefined;
-    subs_tier_name?: string | undefined;
+    subs_tier_id?: string | null | undefined;
+    subs_tier_name?: string | null | undefined;
     subs_usage?: {
         weekly?: {
             resets_at?: string | number | undefined;

package/dist/types/types.d.ts

@@ -195,6 +195,18 @@ export interface CodexCompactionMetadata {
 export interface CodexCompactionRequestContext extends CodexCompactionMetadata {
     operationId: string;
 }
+/** Anthropic `compact_20260112` context-management edit (`compact-2026-01-12` beta). */
+export interface AnthropicCompactionRequest {
+    /**
+     * Prompt input-token count at which the API compacts. The API enforces a
+     * 50,000-token floor and defaults to 150,000 when omitted.
+     */
+    triggerInputTokens?: number;
+    /** Stop after the compaction block instead of continuing the response. */
+    pauseAfterCompaction?: boolean;
+    /** Custom summarization prompt; replaces the API default entirely when set. */
+    instructions?: string;
+}
 /** OpenAI's GPT-5.6+ explicit prompt-cache controls. */
 export interface OpenAIPromptCacheOptions {
     /** `explicit` disables OpenAI's automatic latest-message breakpoint. */
@@ -243,6 +255,15 @@ export interface StreamOptions {
     anthropicPrefixMismatchBehavior?: "drop_block" | "error";
     /** @internal Marks a replay-only Anthropic request that must use non-streaming `max_tokens: 0`. */
     anthropicCacheRefreshRequest?: boolean;
+    /**
+     * Anthropic server-side compaction (`compact-2026-01-12` beta). Sends the
+     * `compact_20260112` context-management edit so the API summarizes the
+     * prompt in-band once its input reaches the trigger; the resulting summary
+     * arrives as an {@link AnthropicCompactionPayload} on the assistant message.
+     * Ignored by every other provider and by Anthropic-compatible endpoints
+     * without context-management support.
+     */
+    anthropicCompaction?: AnthropicCompactionRequest;
     /**
      * Additional headers to include in provider requests.
      * These are merged on top of model-defined headers.
@@ -703,7 +724,32 @@ export interface AnthropicMessagePayload {
         name: string;
     }>;
 }
-export type ProviderPayload = OpenAIResponsesHistoryPayload | AnthropicMessagePayload;
+/**
+ * Anthropic server-side compaction summary (`compact-2026-01-12` beta).
+ *
+ * Produced by the Anthropic provider on the assistant message of a request
+ * that streamed a `compaction` content block, and attached to the user-role
+ * compaction summary message that replaces the compacted history so the
+ * provider can replay the block verbatim: the API drops every block that
+ * precedes it. `content` is the plain-text summary, so every other provider
+ * reads the message text and ignores the payload.
+ */
+export interface AnthropicCompactionPayload {
+    type: "anthropicCompaction";
+    /** Provider that produced the summary; only that provider replays it natively. */
+    provider: string;
+    content: string;
+    /** Opaque provider state the API attached to the block; replayed verbatim when present. */
+    encryptedContent?: string;
+    /**
+     * Harness-appended file metadata (`<files>` section) kept out of the
+     * byte-identical block. Replayed as a user message after the native block:
+     * the converter replaces the summary message with the block and skips its
+     * text, so without this the metadata would be invisible to this provider.
+     */
+    filesText?: string;
+}
+export type ProviderPayload = OpenAIResponsesHistoryPayload | AnthropicMessagePayload | AnthropicCompactionPayload;
 /** Provider-reported rewrite applied to request content before inference. */
 export interface ProviderInputTransformation {
     type: string;

package/dist/types/usage/charm-hyper.d.ts

@@ -0,0 +1,2 @@
+import type { UsageProvider } from "../usage.js";
+export declare const charmHyperUsageProvider: UsageProvider;

package/dist/types/utils/block-symbols.d.ts

@@ -72,3 +72,44 @@ export type DemotedThinkingCarrier = object & {
 };
 /** True for text blocks synthesized by cross-model thinking demotion. */
 export declare function isDemotedThinking(block: DemotedThinkingCarrier | null | undefined): boolean;
+/**
+ * Marks an Anthropic wire message that was serialized from a source
+ * `role: "user"` message.
+ *
+ * The wire role alone cannot answer this. `convertAnthropicMessages` emits
+ * `role: "user"` for several things that are not a conversational turn:
+ * `developer` messages on models without mid-conversation `system` support,
+ * `tool_result` runs, and the synthetic `Continue.` pads inserted between
+ * adjacent assistants. Prompt-cache decimation counts conversational turns,
+ * so it reads this marker instead of guessing from wire content.
+ *
+ * Symbol-keyed so the marker never persists across the JSONL round-trip and
+ * never reaches the wire.
+ */
+export declare const kConversationalUser: unique symbol;
+/** Carries the conversational-user marker without exposing a string-keyed property. */
+export type ConversationalUserCarrier = object & {
+    [kConversationalUser]?: boolean;
+};
+/** True for wire messages serialized from a source `role: "user"` message. */
+export declare function isConversationalUser(message: ConversationalUserCarrier | null | undefined): boolean;
+/**
+ * Marks a `role: "user"` message that `transformMessages` synthesized rather
+ * than one the user sent.
+ *
+ * The stale-tool-result note is deliberately emitted as `user` so no provider
+ * elevates untrusted tool output to instruction priority, which leaves it
+ * indistinguishable from a real turn by role alone. Prompt-cache decimation
+ * must not count it, or an orphan result appearing or disappearing in a
+ * compacted history shifts every later checkpoint.
+ *
+ * Symbol-keyed so the marker never persists across the JSONL round-trip and
+ * never reaches the wire.
+ */
+export declare const kSyntheticUser: unique symbol;
+/** Carries the synthetic-user marker without exposing a string-keyed property. */
+export type SyntheticUserCarrier = object & {
+    [kSyntheticUser]?: boolean;
+};
+/** True for `user` messages synthesized by message transformation. */
+export declare function isSyntheticUser(message: SyntheticUserCarrier | null | undefined): boolean;

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@linxiraos/pi-ai",
-	"version": "1.1.14",
+	"version": "1.1.15",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"keywords": [
 		"ai",
@@ -124,11 +124,11 @@
 		"fmt": "oxfmt --no-error-on-unmatched-pattern 'src/**/*.{ts,tsx}' '{test,bench,examples,scripts}/**/*.ts' '*.ts'"
 	},
 	"dependencies": {
-		"@linxiraos/pi-omptype": "1.1.14",
-		"@linxiraos/pi-catalog": "1.1.14",
-		"@linxiraos/pi-natives": "1.1.14",
-		"@linxiraos/pi-utils": "1.1.14",
-		"@linxiraos/pi-wire": "1.1.14"
+		"@linxiraos/pi-omptype": "1.1.15",
+		"@linxiraos/pi-catalog": "1.1.15",
+		"@linxiraos/pi-natives": "1.1.15",
+		"@linxiraos/pi-utils": "1.1.15",
+		"@linxiraos/pi-wire": "1.1.15"
 	},
 	"devDependencies": {
 		"@types/bun": "^1.3.14"

package/src/auth-storage.ts

@@ -51,6 +51,7 @@ import type {
 } from "./usage";
 import { resolveUsedFraction } from "./usage";
 import { alibabaTokenPlanRankingStrategy, alibabaTokenPlanUsageProvider } from "./usage/alibaba-token-plan";
+import { charmHyperUsageProvider } from "./usage/charm-hyper";
 import { claudeRankingStrategy, claudeUsageProvider } from "./usage/claude";
 import { clinePassUsageProvider } from "./usage/cline-pass";
 import { cursorUsageProvider } from "./usage/cursor";
@@ -677,6 +678,7 @@ const DEFAULT_USAGE_PROVIDERS: UsageProvider[] = [
 	syntheticUsageProvider,
 	xaiOauthUsageProvider,
 	devinUsageProvider,
+	charmHyperUsageProvider,
 ];
 
 const DEFAULT_USAGE_PROVIDER_MAP = new Map<Provider, UsageProvider>(

package/src/error/finalize.ts

@@ -45,7 +45,8 @@ export interface FinalizeResult {
  */
 export async function finalize(error: unknown, opts: FinalizeOptions = {}): Promise<FinalizeResult> {
 	const aborted = opts.abortTracker ? opts.abortTracker.wasCallerAbort() : opts.signal?.aborted === true;
-	const currentStatus = status(error) ?? opts.capturedErrorResponse?.status;
+	const errorStatus = status(error);
+	const currentStatus = errorStatus ?? opts.capturedErrorResponse?.status;
 
 	let message: string;
 	try {
@@ -55,11 +56,17 @@ export async function finalize(error: unknown, opts: FinalizeOptions = {}): Prom
 		message = error instanceof Error ? error.message : String(error);
 	}
 
+	// A captured status is transport context for the original error. Put it at
+	// the root of the classification cause chain so terminal 4xx policy governs
+	// nested diagnostics before they can contribute transient flags.
+	const classificationError =
+		errorStatus === undefined && currentStatus !== undefined ? { status: currentStatus, cause: error } : error;
+
 	const id = classifyMessage({
 		api: opts.api,
 		provider: opts.provider,
 		model: opts.model,
-		errorId: classify(error, opts.api),
+		errorId: classify(classificationError, opts.api),
 		errorMessage: message,
 		errorStatus: currentStatus,
 	});