@linxiraos/pi-agent-core 1.3.0 → 1.3.2

package/CHANGELOG.md

@@ -1,42 +1,8 @@
 ## [Unreleased]
 
-## [1.3.0] - 2026-10-18
-
-### Breaking Changes
+## [1.3.2] - 2026-10-19
 
-- Removed `estimateImageContentTokens()` from `@linxiraos/pi-agent-core/image-tokens`; use `Tokenizer.imageTokens(size, detail)`, which prices with the reading model's catalog image rule ([#15446](https://github.com/can1357/oh-my-pi/pull/15446) by [@will-bogusz](https://github.com/will-bogusz))
-- `applyShakeRegion()` and `applyShakeRegions()` take the replacement's token count (`replacementTokens`); a block rewrite of an assistant turn adds the tokens it removes to `elidedOutputTokens`, so a replayed Claude turn's count stays lowered after a reload ([#15446](https://github.com/can1357/oh-my-pi/pull/15446) by [@will-bogusz](https://github.com/will-bogusz))
-- Removed `ApprovalDetailsOptions`; `formatApprovalDetails` takes only `args` again, since the compact approval prompt no longer shows tool details ([#15642](https://github.com/can1357/oh-my-pi/pull/15642) by [@H4vC](https://github.com/H4vC))
-
-### Added
-
-- Added `AgentLoopConfig.onDetachedToolUpdate`: a tool's `onUpdate` frames that arrive after its run's event stream ended (a returned call whose background job keeps reporting) reach it instead of being dropped; `Agent` forwards them to its subscribers as `tool_execution_update` events
-- `AgentTool.formatApprovalDetails` receives `ApprovalDetailsOptions`; `compact: true` asks for untruncated bodies so a compact approval prompt can count their lines ([#15550](https://github.com/can1357/oh-my-pi/pull/15550) by [@H4vC](https://github.com/H4vC))
-- `Tokenizer.countLatestTurn`: counts a message as the transcript's latest assistant turn, which an Anthropic Messages replay keeps byte-for-byte ([#15791](https://github.com/can1357/oh-my-pi/pull/15791) by [@will-bogusz](https://github.com/will-bogusz))
-- `Agent.prioritizeSteering`: moves given pending steering messages ahead of older ones so one-at-a-time delivery takes them first ([#15642](https://github.com/can1357/oh-my-pi/pull/15642))
-- `AgentTool.delegationGuidance`: instructions a same-name extension replacement inherits in its description when it delegates to the native tool ([#15477](https://github.com/can1357/oh-my-pi/pull/15477) by [@iliaal](https://github.com/iliaal))
-- `CompactionPreparation.lastSummarizedEntryId` and `CompactionResult.providerReplayThroughEntryId`: a remote compaction that could only fit the history before the kept tail reports the last entry its native history covers; later entries replay normally ([#15712](https://github.com/can1357/oh-my-pi/pull/15712) by [@H4vC](https://github.com/H4vC))
-- `agent_start` carries `resumedToolCallIds` when a continued run first executes the calls of an unpaired assistant tail
-- `prepareQueuedMessages` receives which queue (`steering` or `followUp`) a batch came from ([#16026](https://github.com/can1357/oh-my-pi/pull/16026))
-
-### Changed
-
-- Before-model-call hooks now receive the loop context as a second argument, and `buildProviderContext` receives the summarizer model ([#15436](https://github.com/can1357/oh-my-pi/pull/15436) by [@will-bogusz](https://github.com/will-bogusz))
-
-### Fixed
-
-- Fixed Claude context counts reading Opus 5.5 requests about 55% over their billed input: a replayed thinking turn now counts at the output size Claude billed for it, not its signature's bytes ([#15446](https://github.com/can1357/oh-my-pi/pull/15446) by [@will-bogusz](https://github.com/will-bogusz))
-- Fixed screenshots and other non-frame images being counted at the OpenAI rate for every model instead of the reading model's own; Claude billed them about 16% higher ([#15446](https://github.com/can1357/oh-my-pi/pull/15446) by [@will-bogusz](https://github.com/will-bogusz))
-- Fixed OpenAI Responses and Codex context estimates leaving out replayed reasoning: each turn whose encrypted reasoning actually replays now counts its reported reasoning tokens (its ciphertext when the backend reports none), and none while a cold session rebuilds it out of the request, so compaction and snapcompact sizing match billing; branch summaries no longer budget for encrypted reasoning they never send ([#15447](https://github.com/can1357/oh-my-pi/pull/15447) by [@will-bogusz](https://github.com/will-bogusz))
-- Fixed a Claude compaction or branch summarizer receiving assistant reasoning from a snapcompact archive rendered under another model, which Claude refuses (`reasoning_extraction`) ([#15436](https://github.com/can1357/oh-my-pi/pull/15436) by [@will-bogusz](https://github.com/will-bogusz))
-- Fixed a Cursor tool call re-sent by the server persisting two results under one id; the re-run's result replaces the first ([#15749](https://github.com/can1357/oh-my-pi/issues/15749))
-- Fixed compaction keeping more recent history than `keepRecentTokens` when the latest Claude turn on the Anthropic Messages API ended its tool calls on `stop` or `length`: the replay keeps that turn's signed reasoning, but the cut point counted only its visible blocks ([#15791](https://github.com/can1357/oh-my-pi/pull/15791) by [@will-bogusz](https://github.com/will-bogusz))
-- Fixed native (OpenAI/Codex) remote compaction refusing or truncating sessions that fit the window: its preflight counted the JSON-escaped request, inflating quote-, backslash- and newline-heavy tool output such as disassembly by ~10–20%; it now counts text as the model reads it ([#15712](https://github.com/can1357/oh-my-pi/pull/15712) by [@H4vC](https://github.com/H4vC))
-- Fixed OpenAI V1/V2 remote compaction failing outright when the history, kept tail included, exceeds the context window (a large final response or tool output after the last check); it retries on the history before the kept tail ([#15712](https://github.com/can1357/oh-my-pi/pull/15712) by [@H4vC](https://github.com/H4vC))
-- Fixed fitted output caps overflowing the context window on strict hosts (llama.cpp-family) after turns with many parallel tool results; chat-template framing is now charged per locally counted message ([#15699](https://github.com/can1357/oh-my-pi/issues/15699))
-- Fixed leaked DeepSeek DSML markup reaching saved history from reasoning blocks and from turns that also made a valid tool call ([#15485](https://github.com/can1357/oh-my-pi/issues/15485))
-- Cancelling an `ask` after a background job finishes now reports the user cancellation instead of telling the model to retry a skipped tool ([#15985](https://github.com/can1357/oh-my-pi/issues/15985))
-- Fixed Cursor turns failing with `PROVIDER_ERROR` when a follow-up (a finished background job, a late diagnostic, a reminder) arrived while the previous turn was still streaming ([#15821](https://github.com/can1357/oh-my-pi/pull/15821) by [@robfeldmann](https://github.com/robfeldmann))
+- Version line moves to 1.3.2 with the rest of the suite; no user-visible change in this package.
 
 ## [9.2.2] - 2026-01-31
 
@@ -168,6 +134,54 @@
 
 Initial release under @oh-my-pi scope. See previous releases at [badlogic/pi-mono](https://github.com/badlogic/pi-mono).
 
+## [1.3.1] - 2026-10-19
+
+### Changed
+
+- The failed-tool-call correction sent to DeepSeek models now also covers chat-template tool-call tokens written out as text, not only malformed DSML ([#16216](https://github.com/can1357/oh-my-pi/pull/16216) by [@H4vC](https://github.com/H4vC))
+
+### Fixed
+
+- Fixed Ollama chat history estimates counting stripped native thinking and missing foreign thinking replayed as text, which skewed compaction after model switches ([#16164](https://github.com/can1357/oh-my-pi/issues/16164))
+
+## [1.3.0] - 2026-10-18
+
+### Fixed
+
+- Fixed Cursor turns failing with `PROVIDER_ERROR` when a follow-up (a finished background job, a late diagnostic, a reminder) arrived while the previous turn was still streaming ([#15821](https://github.com/can1357/oh-my-pi/pull/15821) by [@robfeldmann](https://github.com/robfeldmann))
+- Cancelling an `ask` after a background job finishes now reports the user cancellation instead of telling the model to retry a skipped tool ([#15985](https://github.com/can1357/oh-my-pi/issues/15985))
+- Fixed a Cursor tool call re-sent by the server persisting two results under one id; the re-run's result replaces the first ([#15749](https://github.com/can1357/oh-my-pi/issues/15749))
+- Fixed compaction keeping more recent history than `keepRecentTokens` when the latest Claude turn on the Anthropic Messages API ended its tool calls on `stop` or `length`: the replay keeps that turn's signed reasoning, but the cut point counted only its visible blocks ([#15791](https://github.com/can1357/oh-my-pi/pull/15791) by [@will-bogusz](https://github.com/will-bogusz))
+- Fixed Claude context counts reading Opus 5.5 requests about 55% over their billed input: a replayed thinking turn now counts at the output size Claude billed for it, not its signature's bytes ([#15446](https://github.com/can1357/oh-my-pi/pull/15446) by [@will-bogusz](https://github.com/will-bogusz))
+- Fixed screenshots and other non-frame images being counted at the OpenAI rate for every model instead of the reading model's own; Claude billed them about 16% higher ([#15446](https://github.com/can1357/oh-my-pi/pull/15446) by [@will-bogusz](https://github.com/will-bogusz))
+- Fixed OpenAI Responses and Codex context estimates leaving out replayed reasoning: each turn whose encrypted reasoning actually replays now counts its reported reasoning tokens (its ciphertext when the backend reports none), and none while a cold session rebuilds it out of the request, so compaction and snapcompact sizing match billing; branch summaries no longer budget for encrypted reasoning they never send ([#15447](https://github.com/can1357/oh-my-pi/pull/15447) by [@will-bogusz](https://github.com/will-bogusz))
+- Fixed a Claude compaction or branch summarizer receiving assistant reasoning from a snapcompact archive rendered under another model, which Claude refuses (`reasoning_extraction`) ([#15436](https://github.com/can1357/oh-my-pi/pull/15436) by [@will-bogusz](https://github.com/will-bogusz))
+- Fixed native (OpenAI/Codex) remote compaction refusing or truncating sessions that fit the window: its preflight counted the JSON-escaped request, inflating quote-, backslash- and newline-heavy tool output such as disassembly by ~10–20%; it now counts text as the model reads it ([#15712](https://github.com/can1357/oh-my-pi/pull/15712) by [@H4vC](https://github.com/H4vC))
+- Fixed OpenAI V1/V2 remote compaction failing outright when the history, kept tail included, exceeds the context window (a large final response or tool output after the last check); it retries on the history before the kept tail ([#15712](https://github.com/can1357/oh-my-pi/pull/15712) by [@H4vC](https://github.com/H4vC))
+- Fixed fitted output caps overflowing the context window on strict hosts (llama.cpp-family) after turns with many parallel tool results; chat-template framing is now charged per locally counted message ([#15699](https://github.com/can1357/oh-my-pi/issues/15699))
+- Fixed leaked DeepSeek DSML markup reaching saved history from reasoning blocks and from turns that also made a valid tool call ([#15485](https://github.com/can1357/oh-my-pi/issues/15485))
+
+### Added
+
+- `prepareQueuedMessages` receives which queue (`steering` or `followUp`) a batch came from ([#16026](https://github.com/can1357/oh-my-pi/pull/16026))
+- `agent_start` carries `resumedToolCallIds` when a continued run first executes the calls of an unpaired assistant tail
+- `Tokenizer.countLatestTurn`: counts a message as the transcript's latest assistant turn, which an Anthropic Messages replay keeps byte-for-byte ([#15791](https://github.com/can1357/oh-my-pi/pull/15791) by [@will-bogusz](https://github.com/will-bogusz))
+- `Agent.prioritizeSteering`: moves given pending steering messages ahead of older ones so one-at-a-time delivery takes them first ([#15642](https://github.com/can1357/oh-my-pi/pull/15642))
+- Added `AgentLoopConfig.onDetachedToolUpdate`: a tool's `onUpdate` frames that arrive after its run's event stream ended (a returned call whose background job keeps reporting) reach it instead of being dropped; `Agent` forwards them to its subscribers as `tool_execution_update` events
+- `AgentTool.formatApprovalDetails` receives `ApprovalDetailsOptions`; `compact: true` asks for untruncated bodies so a compact approval prompt can count their lines ([#15550](https://github.com/can1357/oh-my-pi/pull/15550) by [@H4vC](https://github.com/H4vC))
+- `AgentTool.delegationGuidance`: instructions a same-name extension replacement inherits in its description when it delegates to the native tool ([#15477](https://github.com/can1357/oh-my-pi/pull/15477) by [@iliaal](https://github.com/iliaal))
+- `CompactionPreparation.lastSummarizedEntryId` and `CompactionResult.providerReplayThroughEntryId`: a remote compaction that could only fit the history before the kept tail reports the last entry its native history covers; later entries replay normally ([#15712](https://github.com/can1357/oh-my-pi/pull/15712) by [@H4vC](https://github.com/H4vC))
+
+### Breaking Changes
+
+- Removed `estimateImageContentTokens()` from `@linxiraos/pi-agent-core/image-tokens`; use `Tokenizer.imageTokens(size, detail)`, which prices with the reading model's catalog image rule ([#15446](https://github.com/can1357/oh-my-pi/pull/15446) by [@will-bogusz](https://github.com/will-bogusz))
+- `applyShakeRegion()` and `applyShakeRegions()` take the replacement's token count (`replacementTokens`); a block rewrite of an assistant turn adds the tokens it removes to `elidedOutputTokens`, so a replayed Claude turn's count stays lowered after a reload ([#15446](https://github.com/can1357/oh-my-pi/pull/15446) by [@will-bogusz](https://github.com/will-bogusz))
+- Removed `ApprovalDetailsOptions`; `formatApprovalDetails` takes only `args` again, since the compact approval prompt no longer shows tool details ([#15642](https://github.com/can1357/oh-my-pi/pull/15642) by [@H4vC](https://github.com/H4vC))
+
+### Changed
+
+- Before-model-call hooks now receive the loop context as a second argument, and `buildProviderContext` receives the summarizer model ([#15436](https://github.com/can1357/oh-my-pi/pull/15436) by [@will-bogusz](https://github.com/will-bogusz))
+
 ## [1.2.0] - 2026-10-12
 
 - 版本线推进至 1.2.0；本版无独立用户可见变化。

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@linxiraos/pi-agent-core",
-	"version": "1.3.0",
+	"version": "1.3.2",
 	"description": "General-purpose agent with transport abstraction, state management, and attachment support",
 	"homepage": "https://linxira-os.github.io/zeta/",
 	"author": "Linxira-OS",
@@ -35,16 +35,16 @@
 		"fmt": "oxfmt --no-error-on-unmatched-pattern 'src/**/*.{ts,tsx}' '{test,bench,examples,scripts}/**/*.ts' '*.ts'"
 	},
 	"dependencies": {
-		"@linxiraos/pi-ai": "1.3.0",
-		"@linxiraos/pi-catalog": "1.3.0",
-		"@linxiraos/pi-natives": "1.3.0",
-		"@linxiraos/pi-utils": "1.3.0",
-		"@linxiraos/pi-wire": "1.3.0",
-		"@linxiraos/pi-snapcompact": "1.3.0",
+		"@linxiraos/pi-ai": "1.3.2",
+		"@linxiraos/pi-catalog": "1.3.2",
+		"@linxiraos/pi-natives": "1.3.2",
+		"@linxiraos/pi-utils": "1.3.2",
+		"@linxiraos/pi-wire": "1.3.2",
+		"@linxiraos/pi-snapcompact": "1.3.2",
 		"@opentelemetry/api": "^1.9.1"
 	},
 	"devDependencies": {
-		"@linxiraos/pi-omptype": "1.3.0",
+		"@linxiraos/pi-omptype": "1.3.2",
 		"@opentelemetry/context-async-hooks": "^2.9.0",
 		"@opentelemetry/sdk-trace-base": "^2.9.0",
 		"@types/bun": "1.4.3"

package/src/prompts/dsml-tool-call-leak.md

@@ -1,4 +1,4 @@
-Tool call failed: your previous response tried to invoke a tool, but the DSML markup was malformed, so it arrived as plain text, no tool was executed, and the broken markup was removed from your response. Usually the opening `<｜DSML｜tool_calls>` and/or `<｜DSML｜invoke name="…">` tags are missing, or the tool name was written as a plain line.
+Tool call failed: your previous response tried to invoke a tool, but the tool-call markup was malformed, so it arrived as plain text, no tool was executed, and the broken markup was removed from your response. Usually the opening `<｜DSML｜tool_calls>` and/or `<｜DSML｜invoke name="…">` tags are missing, the tool name was written as a plain line, or the chat template's tool-call special tokens were typed out as text.
 
 A tool call must be one complete envelope:
 

package/src/tokenizer.ts

@@ -1,6 +1,11 @@
 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 { shouldDropTruncatedThinkingOnlyAssistant } from "@linxiraos/pi-ai/providers/transform-messages";
+import {
+	isSameModelTurn,
+	shouldDropTruncatedThinkingOnlyAssistant,
+} from "@linxiraos/pi-ai/providers/transform-messages";
 import {
 	getOpenAIResponsesHistoryItems,
 	sanitizeOpenAIResponsesAssistantHistoryItemsForReplay,
@@ -207,6 +212,7 @@ export class Tokenizer {
 	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;
@@ -244,6 +250,7 @@ export class Tokenizer {
 		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" &&
@@ -415,6 +422,8 @@ export class Tokenizer {
 				// 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
@@ -441,6 +450,13 @@ export class Tokenizer {
 					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
@@ -561,6 +577,7 @@ export class Tokenizer {
 			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
 		);
 	}