@linxiraos/pi-agent-core 1.2.1 → 1.3.1

package/src/tokenizer.ts

@@ -1,11 +1,23 @@
-import type { Model } from "@linxiraos/pi-ai";
+import type { AssistantMessage, Model } from "@linxiraos/pi-ai";
+import { renderDemotedThinking } from "@linxiraos/pi-ai/dialect/demotion";
+import { ollamaChatDropsOwnThinking } from "@linxiraos/pi-ai/providers/ollama";
+import { isColdReplayableResponsesTurn } from "@linxiraos/pi-ai/providers/openai-shared";
+import {
+	isSameModelTurn,
+	shouldDropTruncatedThinkingOnlyAssistant,
+} from "@linxiraos/pi-ai/providers/transform-messages";
+import {
+	getOpenAIResponsesHistoryItems,
+	sanitizeOpenAIResponsesAssistantHistoryItemsForReplay,
+} from "@linxiraos/pi-ai/utils";
+import { resolveThinkingReplayTokens, type ThinkingReplayTokens } from "@linxiraos/pi-catalog/compat/thinking-replay";
 import type { ModelTokenizer } from "@linxiraos/pi-catalog/types";
 import * as natives from "@linxiraos/pi-natives";
 import { materializeString, stringifyJson } from "@linxiraos/pi-utils";
 import { LRUCache } from "@linxiraos/pi-utils/lru";
 import * as snapcompact from "@linxiraos/pi-snapcompact";
 import { isEstimateCacheable, messageEstimateVersion } from "./compaction/message-cache";
-import { base64ImageSize, estimateImageContentTokens } from "./image-tokens";
+import { base64ImageSize, estimateImageTokens, type ImageDetail, type ImageSize } from "./image-tokens";
 import type { AgentMessage } from "./types";
 
 const testEnv = Bun.env.NODE_ENV === "test";
@@ -20,6 +32,7 @@ const NATIVE_ENCODING: Record<ModelTokenizer, natives.Encoding> = {
 	"deepseek-v3": natives.Encoding.DeepSeekV3,
 	"kimi-k2": natives.Encoding.KimiK2,
 	glm5: natives.Encoding.Glm5,
+	o200k: natives.Encoding.O200kBase,
 };
 
 /** Maps the catalog-resolved tokenizer family to its native implementation. */
@@ -27,6 +40,41 @@ export function tokenizerEncodingForModel(model: Pick<Model, "tokenizer"> | null
 	return model?.tokenizer ? NATIVE_ENCODING[model.tokenizer] : null;
 }
 
+type TokenizerModel = Pick<Model, "tokenizer"> &
+	snapcompact.ShapeTarget & { compat?: Model["compat"]; thinking?: Model["thinking"] };
+
+/** The model an assistant turn's native reasoning (signed thinking, Responses history) is replayed to. */
+interface ReasoningReplayReader {
+	api: string;
+	provider: string;
+	id: string;
+}
+
+function reasoningReplayReader(model: TokenizerModel | null | undefined): ReasoningReplayReader | undefined {
+	if (!model?.api || !model.provider || !model.id) return undefined;
+	// Filtering targets strip native reasoning items before the request is encoded.
+	const compat = model.compat;
+	if (compat && "filterReasoningHistory" in compat && compat.filterReasoningHistory) return undefined;
+	return { api: model.api, provider: model.provider, id: model.id };
+}
+
+/** Identity of the reader whose replayed reasoning a {@link Tokenizer} counts; equal keys count it alike. */
+export function reasoningReplayKey(model: TokenizerModel | null | undefined): string {
+	const reader = reasoningReplayReader(model);
+	return reader ? `${reader.api}\0${reader.provider}\0${reader.id}` : "";
+}
+
+/** Request state a {@link Tokenizer} counts for, beyond its model. */
+export interface TokenizerOptions {
+	/**
+	 * Whether the request replays native Responses history (pi-ai's
+	 * `openAIResponsesReplaysNativeHistory`). A cold session rebuilds a turn
+	 * holding encrypted reasoning from its text and tool calls, without the
+	 * reasoning or its signature. Default: true.
+	 */
+	nativeHistoryReplay?: boolean;
+}
+
 /**
  * `strict` always tries an exact native count (the catalog-resolved tokenizer
  * when known, o200k_base otherwise), falling back to the byte upper bound when
@@ -43,11 +91,15 @@ export interface MessageCountOptions {
 	/**
 	 * Drop opaque provider reasoning payloads (`thinkingSignature`,
 	 * `redactedThinking`, native server-tool blocks) from the estimate. Those
-	 * are billed by the provider on replay, so the default counts them — but
-	 * their *local* byte size can diverge wildly from what the provider
-	 * charges, so the compaction floor (which only needs the reliably-countable,
-	 * on-wire-compressible content) excludes them to avoid false triggers on
-	 * thinking-heavy turns.
+	 * are billed by the provider on replay, so the default counts them (a reader
+	 * billing replayed thinking at `output` size: the reasoning, at its turn's
+	 * output size) — but their *local* byte size can diverge wildly from what
+	 * the provider charges, so the compaction floor (which only needs the
+	 * reliably-countable, on-wire-compressible content) excludes them to avoid
+	 * false triggers on thinking-heavy turns.
+	 * Native Responses reasoning that actually replays is not an opaque byte
+	 * guess: it is charged at the turn's provider-reported reasoning tokens in
+	 * both arms.
 	 */
 	excludeEncryptedReasoning?: boolean;
 }
@@ -149,15 +201,25 @@ interface MessageEstimate {
  * Model-aware local token counter. Immutable: the catalog-resolved encoding
  * is fixed at construction, so a cached count can never straddle two
  * encodings. An `Agent` owns one for its active model (swapping the instance
- * when the model's encoding or snapcompact frame pricing changes); one-shot
- * flows construct their own for the model that will be billed. Known
- * tokenizer families use exact native counts; unknown models keep the fast
- * byte estimate (or o200k when `PI_TOKENIZER_ACCURATE=1`).
+ * when the model's encoding, image or thinking-replay pricing, replay
+ * identity or native-history replay state changes); one-shot flows construct
+ * their own for the model that will be billed. Known tokenizer families use
+ * exact native counts; unknown models keep the fast byte estimate (or o200k
+ * when `PI_TOKENIZER_ACCURATE=1`).
  */
 export class Tokenizer {
 	readonly #encoding: natives.Encoding | null;
 	readonly #frameBilling: snapcompact.FrameBilling;
 	readonly frameBillingKey: string;
+	readonly #reasoningReader: ReasoningReplayReader | undefined;
+	readonly #dropsOllamaThinking: boolean;
+	/** Claude's own API and prefix-bound hosts keep another model's signature from their deployment. */
+	readonly #readsDeploymentSignatures: boolean;
+	readonly reasoningReplayKey: string;
+	/** How the reader bills replayed signed thinking (catalog `thinking-replay-tokens`). */
+	readonly thinkingReplayTokens: ThinkingReplayTokens;
+	/** Whether the counted request replays native Responses history; see {@link TokenizerOptions}. */
+	readonly nativeHistoryReplay: boolean;
 
 	/** Exact counts only; byte fallbacks remain mode-dependent and uncached. */
 	readonly #nativeCounts = new LRUCache<string, number>({
@@ -176,10 +238,27 @@ export class Tokenizer {
 	 */
 	#estimates = new WeakMap<AgentMessage, MessageEstimate>();
 
-	constructor(model?: (Pick<Model, "tokenizer"> & snapcompact.ShapeTarget) | null) {
+	/**
+	 * Counts what `streamSimple` sends `model`. Its catalog compat decides
+	 * whether Responses reasoning history is filtered: `streamOpenAIResponses`'
+	 * per-request `filterReasoningHistory` override is a debugging hook
+	 * `streamSimple` never forwards, so a caller using it passes `model` with
+	 * that compat set.
+	 */
+	constructor(model?: TokenizerModel | null, options?: TokenizerOptions) {
 		this.#encoding = tokenizerEncodingForModel(model);
 		this.#frameBilling = snapcompact.frameBilling(model ?? undefined);
 		this.frameBillingKey = snapcompact.frameBillingKey(this.#frameBilling);
+		this.#reasoningReader = reasoningReplayReader(model);
+		this.#dropsOllamaThinking = ollamaChatDropsOwnThinking(model);
+		const compat = model?.compat;
+		this.#readsDeploymentSignatures =
+			model?.api === "anthropic-messages" &&
+			((compat !== undefined && "officialEndpoint" in compat && compat.officialEndpoint) ||
+				model.thinking?.prefixBinding === true);
+		this.reasoningReplayKey = reasoningReplayKey(model);
+		this.thinkingReplayTokens = resolveThinkingReplayTokens(model ?? {});
+		this.nativeHistoryReplay = options?.nativeHistoryReplay ?? true;
 	}
 
 	get encoding(): natives.Encoding | null {
@@ -218,12 +297,12 @@ export class Tokenizer {
 	 * Settled historical messages are counted once and reused until an owner
 	 * (prune/shake/strip-images) calls `invalidateMessageCache`; streaming
 	 * assistants bypass the memo entirely (see the message-cache settle-gate
-	 * invariant). Image blocks charge {@link estimateImageContentTokens}, the
-	 * same dimension-based estimate the native remote-compaction probe applies.
+	 * invariant). Image blocks charge {@link imageTokens}, the same estimate
+	 * the native remote-compaction probe applies.
 	 */
 	countMessage(message: AgentMessage, options?: MessageCountOptions): number {
 		const floored = options?.excludeEncryptedReasoning === true;
-		if (!isEstimateCacheable(message)) return this.#measureMessage(message, floored);
+		if (!isEstimateCacheable(message)) return this.#measureMessage(message, floored, false);
 		const version = messageEstimateVersion(message);
 		let entry = this.#estimates.get(message);
 		if (entry === undefined || entry.version !== version) {
@@ -232,12 +311,25 @@ export class Tokenizer {
 		}
 		const cached = floored ? entry.floored : entry.default;
 		if (cached !== undefined) return cached;
-		const result = this.#measureMessage(message, floored);
+		const result = this.#measureMessage(message, floored, false);
 		if (floored) entry.floored = result;
 		else entry.default = result;
 		return result;
 	}
 
+	/**
+	 * {@link countMessage} for the transcript's latest surviving assistant turn
+	 * (`isSurvivingAssistant`), which an Anthropic Messages replay keeps
+	 * byte-for-byte (`transformMessages`): a turn whose tool calls ended without
+	 * a tool-use stop keeps its signatures then, so its reasoning is charged. A
+	 * count of one message cannot see its position; a caller walking back from
+	 * the end of the transcript counts that turn here. Not memoized: it differs
+	 * from {@link countMessage} only for that one turn.
+	 */
+	countLatestTurn(message: AgentMessage, options?: MessageCountOptions): number {
+		return this.#measureMessage(message, options?.excludeEncryptedReasoning === true, true);
+	}
+
 	/** Sum of {@link countMessage} over `messages`. */
 	countMessages(messages: readonly AgentMessage[], options?: MessageCountOptions): number {
 		let total = 0;
@@ -245,7 +337,51 @@ export class Tokenizer {
 		return total;
 	}
 
-	#measureMessage(message: AgentMessage, excludeEncryptedReasoning: boolean): number {
+	/**
+	 * Input tokens for one image (not a snapcompact frame) under the reader's
+	 * catalog image rule, or the generic estimate when it has none.
+	 */
+	imageTokens(size: ImageSize | null | undefined, detail?: ImageDetail): number {
+		return estimateImageTokens(size, detail, this.#frameBilling.rule);
+	}
+
+	/**
+	 * Whether `message`'s signed reasoning reaches this reader natively: from
+	 * the same model, host and wire, or from another Claude model on the same
+	 * Anthropic deployment, which keeps the signature (`transformMessages`).
+	 * Claude bills that reasoning when the reader can read it and drops the
+	 * block otherwise, so such a turn counts high rather than low.
+	 */
+	#replaysReasoningNatively(message: AssistantMessage): boolean {
+		const reader = this.#reasoningReader;
+		return (
+			reader !== undefined &&
+			message.api === reader.api &&
+			message.provider === reader.provider &&
+			(message.model === reader.id || this.#readsDeploymentSignatures)
+		);
+	}
+
+	/**
+	 * The replayed native Responses reasoning {@link countMessage} charges
+	 * across `messages`: the part of a count that changes with the reading
+	 * model. A replay whose backend reported no reasoning tokens counts at the
+	 * signature price `countMessage` gives it.
+	 */
+	replayedReasoningTokens(messages: readonly AgentMessage[]): number {
+		let total = 0;
+		for (const message of messages) {
+			if (message.role !== "assistant") continue;
+			const replayed = this.#replayedReasoningTokens(message);
+			total +=
+				replayed === null
+					? this.countMessage(message) - this.countMessage(message, { excludeEncryptedReasoning: true })
+					: (replayed ?? 0);
+		}
+		return total;
+	}
+
+	#measureMessage(message: AgentMessage, excludeEncryptedReasoning: boolean, latestTurn: boolean): number {
 		const fragments: string[] = [];
 		let extra = 0;
 		// Declaration-merged app roles (the coding-agent's bashExecution) are
@@ -273,35 +409,78 @@ export class Tokenizer {
 						if (block.type === "text" && block.text) {
 							fragments.push(block.text);
 						} else if (block.type === "image") {
-							extra += estimateImageContentTokens(block);
+							extra += this.imageTokens(base64ImageSize(block.data), block.detail);
 						}
 					}
 				}
 				break;
 			}
 			case "assistant": {
-				for (const block of message.content) {
+				// Replayed native Responses reasoning is billed as input at the turn's
+				// reported reasoning tokens, not at its ~3x larger ciphertext size, so it
+				// is a token price and counts in both arms. Its thinking signature never
+				// reaches the wire then (the payload items replay instead); a turn whose
+				// backend reported no reasoning tokens keeps the signature as its price.
+				const replayedReasoning = this.#replayedReasoningTokens(message);
+				const ollamaReader = this.#reasoningReader?.api === "ollama-chat" ? this.#reasoningReader : undefined;
+				const ownOllamaTurn = ollamaReader !== undefined && isSameModelTurn(message, ollamaReader);
+				if (typeof replayedReasoning === "number") extra += replayedReasoning;
+				const signaturesPriced = this.thinkingReplayTokens === "signature";
+				// The replay's signature rule (`transformMessages`): an aborted or
+				// errored turn's final block can hold a partial signature, and a turn
+				// whose tool calls ended without a tool-use stop is abandoned, so all
+				// its signatures are stripped. Every other block completed. On the
+				// Anthropic Messages API the latest assistant turn keeps an abandoned
+				// turn's signatures byte-for-byte (Bedrock's Converse replay strips
+				// them too); a per-message count cannot see position, so it follows
+				// the rule for older turns unless the caller says it is the latest
+				// (`countLatestTurn`).
+				const content = message.content;
+				const invalidStop = message.stopReason === "aborted" || message.stopReason === "error";
+				const keepsLatestSignatures = latestTurn && message.api === "anthropic-messages";
+				const abandonedToolUse =
+					!keepsLatestSignatures &&
+					!invalidStop &&
+					message.stopReason !== "toolUse" &&
+					content.some(block => block.type === "toolCall");
+				const untrustedIndex = invalidStop ? content.length - 1 : -1;
+				let signedReasoning = false;
+				for (let index = 0; index < content.length; index++) {
+					const block = content[index];
 					if (block.type === "text") {
 						fragments.push(block.text);
 					} else if (block.type === "thinking") {
+						// `transformMessages` drops another model's empty thinking and demotes the
+						// rest to visible text; only the reader's own thinking stays native.
+						if (ollamaReader && !ownOllamaTurn) {
+							if (block.thinking.trim()) fragments.push(renderDemotedThinking(ollamaReader.id, block.thinking));
+							continue;
+						}
+						if (this.#dropsOllamaThinking && ownOllamaTurn) continue;
 						fragments.push(block.thinking);
 						// Providers charge for the opaque signature/reasoning payload that
 						// rides alongside the thinking text (OpenAI Responses encrypted
-						// reasoning items, Anthropic signed thinking blocks, etc.). Without
-						// counting it, this estimator can read ~half of the provider-reported
-						// usage on thinking-heavy turns — see #2275 for the resulting
-						// compaction-trigger / post-check metric divergence. The compaction
-						// floor excludes it (its local byte size diverges from provider billing).
-						if (block.thinkingSignature && !excludeEncryptedReasoning) {
+						// reasoning items, Gemini thought signatures). A reader billing at
+						// `output` size (Claude) charges the reasoning, priced below, not
+						// these bytes. The compaction floor excludes it (its local byte size
+						// diverges from provider billing).
+						if (
+							block.thinkingSignature &&
+							!excludeEncryptedReasoning &&
+							signaturesPriced &&
+							typeof replayedReasoning !== "number"
+						) {
 							fragments.push(block.thinkingSignature);
 						}
+						if (block.thinkingSignature && !abandonedToolUse && index !== untrustedIndex) signedReasoning = true;
 					} else if (block.type === "toolCall") {
 						fragments.push(block.name);
 						fragments.push(stringifyJson(block.arguments) ?? "null");
 					} else if (block.type === "redactedThinking") {
 						// Encrypted reasoning blob the provider still bills for on replay;
 						// excluded from the compaction floor for the same reason as above.
-						if (!excludeEncryptedReasoning) fragments.push(block.data);
+						signedReasoning = true;
+						if (!excludeEncryptedReasoning && signaturesPriced) fragments.push(block.data);
 					} else if (block.type === "anthropicServerTool") {
 						// Native Anthropic server-tool call/result replayed verbatim on the
 						// wire (server_tool_use input and opaque result content). The provider
@@ -311,6 +490,19 @@ export class Tokenizer {
 						if (!excludeEncryptedReasoning) fragments.push(stringifyJson(block.block) ?? "null");
 					}
 				}
+				// The visible text undercounts a native replay: the provider rebuilds
+				// the full reasoning from its signature, billed at the turn's output
+				// size less the visible text shake has since elided from it. The
+				// floor keeps the visible count.
+				if (
+					signedReasoning &&
+					!excludeEncryptedReasoning &&
+					!signaturesPriced &&
+					this.#replaysReasoningNatively(message)
+				) {
+					const visible = fragments.length === 0 ? extra : extra + this.countTokens(fragments);
+					return Math.max(visible, message.usage.output - (message.elidedOutputTokens ?? 0));
+				}
 				break;
 			}
 			case "custom":
@@ -323,7 +515,7 @@ export class Tokenizer {
 						if (block.type === "text" && block.text) {
 							fragments.push(block.text);
 						} else if (block.type === "image") {
-							extra += estimateImageContentTokens(block);
+							extra += this.imageTokens(base64ImageSize(block.data), block.detail);
 						}
 					}
 				}
@@ -354,6 +546,42 @@ export class Tokenizer {
 		return extra + this.countTokens(fragments);
 	}
 
+	/**
+	 * Input tokens this reader bills for replaying `message`'s native Responses
+	 * reasoning; `undefined` when the reader re-encodes the turn from its
+	 * blocks instead, `null` when it replays but the backend reported no
+	 * reasoning tokens to price it by. The payload goes back on the wire only
+	 * to the api, model and provider that produced it (`convertMessages` in
+	 * pi-ai's Responses providers), and then only what survives replay: a
+	 * truncated thinking-only turn is dropped, a hidden-empty or reasoning-only
+	 * payload sanitizes to nothing, and a cold session rebuilds the turn without
+	 * its reasoning, so those send no reasoning and no signature (0).
+	 */
+	#replayedReasoningTokens(message: AssistantMessage): number | null | undefined {
+		const reader = this.#reasoningReader;
+		if (!reader || message.api !== reader.api || message.model !== reader.id) return undefined;
+		const items = getOpenAIResponsesHistoryItems(message.providerPayload, reader.provider, message.provider);
+		if (!items) return undefined;
+		if (!items.some(item => item.type === "reasoning") || shouldDropTruncatedThinkingOnlyAssistant(message)) return 0;
+		const replay = sanitizeOpenAIResponsesAssistantHistoryItemsForReplay(items);
+		if (!replay?.some(item => item.type === "reasoning")) return 0;
+		// Cold, only plaintext reasoning bound to no server state replays natively.
+		if (!this.nativeHistoryReplay && !isColdReplayableResponsesTurn(replay)) return 0;
+		return message.usage?.reasoningTokens ?? null;
+	}
+
+	/** Whether a Tokenizer built for `model` with `options` would count every message as this one does. */
+	countsAlike(model: TokenizerModel | null | undefined, options?: TokenizerOptions): boolean {
+		return (
+			tokenizerEncodingForModel(model) === this.#encoding &&
+			snapcompact.frameBillingKey(snapcompact.frameBilling(model ?? undefined)) === this.frameBillingKey &&
+			reasoningReplayKey(model) === this.reasoningReplayKey &&
+			resolveThinkingReplayTokens(model ?? {}) === this.thinkingReplayTokens &&
+			ollamaChatDropsOwnThinking(model) === this.#dropsOllamaThinking &&
+			(options?.nativeHistoryReplay ?? true) === this.nativeHistoryReplay
+		);
+	}
+
 	/** One snapcompact frame at the active model's price for its pixel size; unreadable frames cost the ceiling. */
 	#frameTokens(data: string): number {
 		const size = base64ImageSize(data);

package/src/types.ts

@@ -38,10 +38,14 @@ export interface QueuedMessagePreparation {
 	commit(): readonly AgentMessage[] | undefined;
 }
 
+/** Which agent queue a batch came from: steering amends the running task, follow-ups start the next one. */
+export type QueuedMessageQueue = "steering" | "followUp";
+
 /** Prepare an exclusively claimed batch. Undefined delivers unchanged; the signal also aborts when the claim is cancelled. */
 export type PrepareQueuedMessages = (
 	messages: readonly AgentMessage[],
 	signal: AbortSignal,
+	queue: QueuedMessageQueue,
 ) => QueuedMessagePreparation | undefined | Promise<QueuedMessagePreparation | undefined>;
 
 /** Called once an aside has been inserted into the agent's live context. */
@@ -338,6 +342,13 @@ export interface AgentLoopConfig extends SimpleStreamOptions {
 	 * tools finish first.
 	 */
 	hasQueuedAsides?: () => boolean | Promise<boolean>;
+	/**
+	 * Receives a tool's `onUpdate` frames that arrive after this run's event
+	 * stream ended: a call that returned while its background job keeps
+	 * reporting (`details.async`), whose run yielded before the job settled.
+	 * Without it those frames are dropped.
+	 */
+	onDetachedToolUpdate?: (event: Extract<AgentEvent, { type: "tool_execution_update" }>) => void;
 
 	/**
 	 * Returns follow-up messages to process after the agent would otherwise stop.
@@ -1135,6 +1146,11 @@ export interface AgentTool<
 	loadMode?: ToolLoadMode;
 	/** Short one-line summary used for tool discovery indexes. */
 	summary?: string;
+	/**
+	 * Instructions a same-name extension replacement inherits in its description, because results it
+	 * obtains by delegating to this native tool through `ctx.invokeTool` follow this tool's contract.
+	 */
+	readonly delegationGuidance?: string;
 	/**
 	 * On-demand documentation topics (`topic → markdown`), readable as
 	 * `xd://<tool>/<topic>`. Lets a tool keep large sub-surfaces out of its
@@ -1240,7 +1256,14 @@ export interface AgentContext {
  */
 export type AgentEvent =
 	// Agent lifecycle
-	| { type: "agent_start" }
+	| {
+			type: "agent_start";
+			/**
+			 * Calls of the unpaired assistant tail a continued run executes before its
+			 * first model request. They run without a new assistant `message_end`.
+			 */
+			resumedToolCallIds?: string[];
+	  }
 	| {
 			type: "agent_end";
 			messages: AgentMessage[];