@bastani/atomic 0.8.30-alpha.1 → 0.8.30-alpha.3

package/CHANGELOG.md

@@ -13,6 +13,8 @@
 
 ### Fixed
 
+- Fixed Anthropic/GitHub Copilot extended-thinking replay by repairing provider payloads so same-model `thinking` and `redacted_thinking` blocks are restored byte-for-byte after provider conversion and extension `before_provider_request` hooks, and by making compaction treat retained thinking-bearing assistant messages as all-or-nothing so sibling tool calls/text cannot be partially removed. This covers signed empty thinking blocks, sanitized thinking text, and compacted sessions that previously triggered 400 `thinking blocks ... cannot be modified` failures.
+- Fixed post-compaction context usage accounting so the footer no longer trusts provider `totalTokens` when normalized usage components are available and avoids double-counting Anthropic-compatible cache buckets that mirror `input` tokens, preventing compacted sessions from displaying roughly doubled context-window percentages such as ~117% when the active prompt is closer to ~58%.
 - Fixed package commands to drain stdout/stderr before the forced post-command exit so piped `atomic list`, help, progress, and self-update fallback output is not truncated under Node while still terminating leaked extension handles.
 - Fixed custom `models.json` providers whose `apiKey` references an unset explicit environment variable so their models are omitted from `/model`, `--list-models`, available-model RPC responses, and automatic fallback candidates until the environment variable is configured.
 - Fixed bash/child-process output draining so late stdout/stderr arriving after process exit continues draining while active, quiet inherited pipes release promptly, endlessly noisy detached descendants are bounded by a longer active-drain cap, and the built-in `bash` tool ignores output after its result accumulator has been finalized.
@@ -26,6 +28,8 @@
 - Fixed bundled workflow and subagent structured-output gates to recover from missing or invalid `structured_output` final answers by issuing up to three corrective retries that echo the actual contract or schema-validation error before failing.
 - Fixed bundled workflow failed-stage metadata so error-stage transcripts remain discoverable and follow-up messaging resumes from the failed conversation instead of resetting to an empty session.
 - Fixed context compaction so older assistant `thinking` and `redacted_thinking` blocks can be removed like other stale blocks, while `thinking` or `redacted_thinking` blocks in the latest assistant message remain rejected by validation and paired tool-result restoration still preserves active context integrity ([#1386](https://github.com/bastani-inc/atomic/issues/1386)).
+- Fixed bundled workflow graph rendering/runtime state for limited-concurrency `ctx.parallel` fan-outs so queued branches now keep sibling parentage after earlier branch failures.
+- Fixed context compaction to universally protect every content block in the latest retained assistant message when that message contains `thinking` or `redacted_thinking`, so `context_delete` and `context_grep_delete` cannot remove visible sibling blocks or make an older partially-filtered thinking-bearing assistant become the latest retained assistant ([#1405](https://github.com/bastani-inc/atomic/issues/1405)).
 
 ## [0.8.29] - 2026-06-15
 

package/dist/builtin/cursor/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/cursor",
-  "version": "0.8.30-alpha.1",
+  "version": "0.8.30-alpha.3",
   "private": true,
   "description": "Experimental first-party Atomic extension for Cursor OAuth, model discovery, and streaming provider registration.",
   "contributors": [
@@ -40,7 +40,7 @@
     }
   },
   "dependencies": {
-    "@bastani/atomic-natives": "0.8.30-alpha.1",
+    "@bastani/atomic-natives": "0.8.30-alpha.3",
     "@bufbuild/protobuf": "^2.0.0"
   }
 }

package/dist/builtin/intercom/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/intercom",
-  "version": "0.8.30-alpha.1",
+  "version": "0.8.30-alpha.3",
   "private": true,
   "description": "Atomic extension providing a private coordination channel between parent and child agent sessions. Fork of: https://github.com/nicobailon/pi-intercom",
   "contributors": [

package/dist/builtin/mcp/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/mcp",
-  "version": "0.8.30-alpha.1",
+  "version": "0.8.30-alpha.3",
   "private": true,
   "description": "Atomic extension that adapts MCP (Model Context Protocol) servers into the coding agent. Fork of: https://github.com/nicobailon/pi-mcp-adapter",
   "contributors": [

package/dist/builtin/subagents/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/subagents",
-  "version": "0.8.30-alpha.1",
+  "version": "0.8.30-alpha.3",
   "private": true,
   "description": "Atomic extension for delegating tasks to subagents with chains, parallel execution, and TUI clarification. Fork of: https://github.com/nicobailon/pi-subagents",
   "contributors": [

package/dist/builtin/web-access/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/web-access",
-  "version": "0.8.30-alpha.1",
+  "version": "0.8.30-alpha.3",
   "private": true,
   "description": "Atomic extension for web search, URL fetching, GitHub repo cloning, PDF/video extraction. Fork of: https://github.com/nicobailon/pi-web-access",
   "contributors": [

package/dist/builtin/workflows/CHANGELOG.md

@@ -16,6 +16,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 - Fixed schema-backed workflow stages to send up to three corrective follow-up prompts when a turn finishes without the required `structured_output` call or with an invalid `structured_output` call, echoing the concrete contract/validation error before failing the stage.
 - Fixed failed workflow stages to retain and persist SDK `sessionId`/`sessionFile` metadata, so post-error transcript inspection and follow-up messaging resume from the failed conversation instead of silently creating a fresh empty session.
 - Fixed schema-backed workflow stages with `noTools: "all"` to keep the restrictive allowlist while still exposing the required `structured_output` final-answer tool.
+- Fixed `ctx.parallel` graph inference so queued branches launched under a limited `concurrency` setting keep the same parent frontier as their sibling branches, even when an earlier sibling fails with `failFast: false`, instead of appearing as downstream children of failed siblings.
 
 ## [0.8.29] - 2026-06-15
 

package/dist/builtin/workflows/README.md

@@ -68,7 +68,7 @@ export default defineWorkflow("summarize-pr")
 
 ### Example 2 — Parallel fan-out with `ctx.parallel`
 
-Use `ctx.parallel` for independent specialist work. The aggregator receives the specialist outputs through typed task results instead of manual stage/session plumbing.
+Use `ctx.parallel` for independent specialist work. The aggregator receives the specialist outputs through typed task results instead of manual stage/session plumbing. The runtime snapshots the parent graph frontier when the fan-out starts, so every branch shares the same parents even when limited `concurrency` queues later branches or an earlier sibling fails with `failFast: false`.
 
 ```typescript
 import { defineWorkflow, Type } from "@bastani/workflows";

package/dist/builtin/workflows/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/workflows",
-  "version": "0.8.30-alpha.1",
+  "version": "0.8.30-alpha.3",
   "private": true,
   "description": "Atomic extension for multi-stage workflow authoring and execution.",
   "contributors": [

package/dist/builtin/workflows/src/runs/foreground/executor.ts

@@ -2575,6 +2575,7 @@ interface ParallelFailFastScope {
   failed: boolean;
   firstFailure?: unknown;
   readonly activeStages: Map<string, ParallelFailFastStage>;
+  readonly parentIds?: readonly string[];
 }
 
 // ---------------------------------------------------------------------------
@@ -4041,13 +4042,18 @@ export async function run<TInputs extends WorkflowInputValues>(
 
       // b. tracker.onSpawn → provisional parentIds
       const provisionalParentIds = tracker.onSpawn(stageId, name);
+      const scopedParentIds = opts.continuation === undefined ? stageFailFastScope?.parentIds : undefined;
+      const initialParentIds = scopedParentIds === undefined ? provisionalParentIds : [...scopedParentIds];
+      if (scopedParentIds !== undefined && !sameStringSet(scopedParentIds, provisionalParentIds)) {
+        tracker.replaceParents(stageId, scopedParentIds);
+      }
 
       // c. Create StageSnapshot as "pending"
       const replayKey = `stage:${name}`;
       const replayDecision = replayIndex.decide({
         displayName: name,
         replayKey,
-        parentIds: provisionalParentIds,
+        parentIds: initialParentIds,
         stageId,
         kind: "stage",
       });
@@ -4599,7 +4605,7 @@ export async function run<TInputs extends WorkflowInputValues>(
           throw err;
         }
 
-        if (opts.continuation === undefined && stageSnapshot.startedAt === undefined) {
+        if (opts.continuation === undefined && stageSnapshot.startedAt === undefined && stageFailFastScope?.parentIds === undefined) {
           const actualParentIds = tracker.currentParents();
           if (!sameStringSet(actualParentIds, stageSnapshot.parentIds)) {
             tracker.replaceParents(stageId, actualParentIds);
@@ -4956,22 +4962,25 @@ export async function run<TInputs extends WorkflowInputValues>(
     async parallel(steps: readonly WorkflowTaskStep[], options: WorkflowParallelOptions = {}): Promise<WorkflowTaskResult[]> {
       throwIfWorkflowExitSelected();
       const fallback = parallelFallbackTask(steps, options);
-      const failFastScope: ParallelFailFastScope | undefined = options.failFast === false
-        ? undefined
-        : { failed: false, activeStages: new Map<string, ParallelFailFastStage>() };
+      const failFastEnabled = options.failFast !== false;
+      const parallelScope: ParallelFailFastScope = {
+        failed: false,
+        activeStages: new Map<string, ParallelFailFastStage>(),
+        parentIds: Object.freeze(tracker.currentParents()),
+      };
       return mapParallelSteps(steps, options.concurrency, options.failFast, async (step) => {
         throwIfWorkflowExitSelected();
         const prompt = replaceTaskPlaceholder(step.prompt ?? step.task ?? fallback, options.task ?? fallback);
         return await (ctx.task as typeof ctx.task & ((taskName: string, taskOptions: WorkflowTaskOptions, scope?: ParallelFailFastScope) => Promise<WorkflowTaskResult>))(
           step.name,
           taskWithSharedDefaults(taskOptionsFromStep(step, prompt, taskPrevious(step)), options),
-          failFastScope,
+          parallelScope,
         );
       }, (error) => {
-        if (failFastScope === undefined) return;
-        failFastScope.failed = true;
-        failFastScope.firstFailure = error;
-        for (const stage of failFastScope.activeStages.values()) {
+        if (!failFastEnabled) return;
+        parallelScope.failed = true;
+        parallelScope.firstFailure = error;
+        for (const stage of parallelScope.activeStages.values()) {
           stage.skip();
         }
       }, {

package/dist/core/anthropic-thinking-guard.d.ts

@@ -0,0 +1,15 @@
+import type { Api, Message, Model } from "@earendil-works/pi-ai";
+/**
+ * Restore same-model Anthropic replay thinking blocks after provider payload
+ * construction and extension payload hooks.
+ *
+ * Anthropic requires thinking/redacted_thinking blocks from replayed assistant
+ * messages to remain byte-for-byte identical to the original response. The
+ * upstream pi-ai Anthropic converter currently still sanitizes thinking text,
+ * drops signed empty thinking, and does not understand raw redacted_thinking
+ * blocks. This guard repairs the already-built Anthropic payload from the
+ * pre-provider LLM messages while leaving non-Anthropic and cross-model payloads
+ * unchanged.
+ */
+export declare function restoreAnthropicReplayThinkingBlocks(payload: unknown, sourceMessages: readonly Message[], model: Model<Api>): unknown;
+//# sourceMappingURL=anthropic-thinking-guard.d.ts.map

package/dist/core/anthropic-thinking-guard.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"anthropic-thinking-guard.d.ts","sourceRoot":"","sources":["../../src/core/anthropic-thinking-guard.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,GAAG,EAAoB,OAAO,EAAE,KAAK,EAAE,MAAM,uBAAuB,CAAC;AAuLnF;;;;;;;;;;;GAWG;AACH,wBAAgB,oCAAoC,CAAC,OAAO,EAAE,OAAO,EAAE,cAAc,EAAE,SAAS,OAAO,EAAE,EAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,GAAG,OAAO,CAyCrI","sourcesContent":["import type { Api, AssistantMessage, Message, Model } from \"@earendil-works/pi-ai\";\n\ntype JsonObject = Record<string, unknown>;\n\ntype AnthropicContentBlock = JsonObject & { type: string };\n\ntype ReplayThinkingBlock =\n\t| { type: \"thinking\"; thinking: string; signature: string }\n\t| { type: \"redacted_thinking\"; data: string };\n\nfunction isObject(value: unknown): value is JsonObject {\n\treturn value !== null && typeof value === \"object\" && !Array.isArray(value);\n}\n\nfunction isAnthropicContentBlock(value: unknown): value is AnthropicContentBlock {\n\treturn isObject(value) && typeof value.type === \"string\";\n}\n\nfunction isThinkingLikeAnthropicBlock(value: unknown): boolean {\n\tif (!isAnthropicContentBlock(value)) return false;\n\treturn value.type === \"thinking\" || value.type === \"redacted_thinking\";\n}\n\nfunction isSameModelAssistant(message: AssistantMessage, model: Model<Api>): boolean {\n\treturn message.provider === model.provider && message.api === model.api && message.model === model.id;\n}\n\nfunction readReplayThinkingBlock(block: unknown): ReplayThinkingBlock | undefined {\n\tif (!isObject(block) || typeof block.type !== \"string\") return undefined;\n\n\tif (block.type === \"redacted_thinking\") {\n\t\treturn typeof block.data === \"string\" ? { type: \"redacted_thinking\", data: block.data } : undefined;\n\t}\n\n\tif (block.type !== \"thinking\") return undefined;\n\n\tif (block.redacted === true) {\n\t\treturn typeof block.thinkingSignature === \"string\"\n\t\t\t? { type: \"redacted_thinking\", data: block.thinkingSignature }\n\t\t\t: undefined;\n\t}\n\n\treturn typeof block.thinking === \"string\" && typeof block.thinkingSignature === \"string\"\n\t\t? { type: \"thinking\", thinking: block.thinking, signature: block.thinkingSignature }\n\t\t: undefined;\n}\n\nfunction hasReplayThinkingBlock(message: AssistantMessage): boolean {\n\treturn message.content.some((block) => readReplayThinkingBlock(block) !== undefined);\n}\n\nfunction legacyProviderWouldEmitAssistantBlock(block: unknown, allowEmptySignature: boolean): boolean {\n\tif (!isObject(block) || typeof block.type !== \"string\") return false;\n\n\tif (block.type === \"text\") {\n\t\treturn typeof block.text === \"string\" && block.text.trim().length > 0;\n\t}\n\n\tif (block.type === \"toolCall\") {\n\t\treturn true;\n\t}\n\n\tif (block.type !== \"thinking\") {\n\t\treturn false;\n\t}\n\n\tif (block.redacted === true) {\n\t\treturn true;\n\t}\n\n\tif (typeof block.thinking !== \"string\") return false;\n\tif (block.thinking.trim().length === 0) return false;\n\n\tconst signature = typeof block.thinkingSignature === \"string\" ? block.thinkingSignature : undefined;\n\tif (signature !== undefined && signature.trim().length > 0) return true;\n\treturn allowEmptySignature;\n}\n\nfunction legacyProviderWouldEmitAssistant(message: AssistantMessage, model: Model<Api>, allowEmptySignature: boolean): boolean {\n\tif (message.stopReason === \"error\" || message.stopReason === \"aborted\") return false;\n\n\t// This mirrors the relevant same-model branch in @earendil-works/pi-ai's\n\t// transformMessages() + Anthropic convertMessages() pipeline closely enough\n\t// to keep assistant ordinal mapping aligned with the provider payload.\n\tif (isSameModelAssistant(message, model)) {\n\t\treturn message.content.some((block) => legacyProviderWouldEmitAssistantBlock(block, allowEmptySignature));\n\t}\n\n\treturn message.content.some((block) => {\n\t\tif (!isObject(block) || typeof block.type !== \"string\") return false;\n\t\tif (block.type === \"text\") return typeof block.text === \"string\" && block.text.trim().length > 0;\n\t\tif (block.type === \"toolCall\") return true;\n\t\tif (block.type !== \"thinking\") return false;\n\t\tif (block.redacted === true) return false;\n\t\treturn typeof block.thinking === \"string\" && block.thinking.trim().length > 0;\n\t});\n}\n\nfunction fallbackTextBlock(block: JsonObject): AnthropicContentBlock | undefined {\n\treturn typeof block.text === \"string\" && block.text.trim().length > 0\n\t\t? { type: \"text\", text: block.text }\n\t\t: undefined;\n}\n\nfunction fallbackToolUseBlock(block: JsonObject): AnthropicContentBlock | undefined {\n\tif (typeof block.id !== \"string\" || typeof block.name !== \"string\") return undefined;\n\treturn { type: \"tool_use\", id: block.id, name: block.name, input: block.arguments ?? {} };\n}\n\nfunction takeProviderBlock(\n\tproviderBlocks: readonly AnthropicContentBlock[],\n\tproviderIndex: number,\n\texpectedType: string,\n): { block: AnthropicContentBlock | undefined; nextProviderIndex: number } {\n\tconst providerBlock = providerBlocks[providerIndex];\n\tif (providerBlock?.type === expectedType) {\n\t\treturn { block: providerBlock, nextProviderIndex: providerIndex + 1 };\n\t}\n\treturn { block: undefined, nextProviderIndex: providerIndex };\n}\n\nfunction repairAssistantContent(\n\toriginalMessage: AssistantMessage,\n\tproviderBlocks: readonly AnthropicContentBlock[],\n): AnthropicContentBlock[] {\n\tconst repaired: AnthropicContentBlock[] = [];\n\tlet providerIndex = 0;\n\n\tfor (const originalBlock of originalMessage.content) {\n\t\tconst replayBlock = readReplayThinkingBlock(originalBlock);\n\t\tif (replayBlock) {\n\t\t\tif (isThinkingLikeAnthropicBlock(providerBlocks[providerIndex])) {\n\t\t\t\tproviderIndex += 1;\n\t\t\t}\n\t\t\trepaired.push({ ...replayBlock });\n\t\t\tcontinue;\n\t\t}\n\n\t\tif (!isObject(originalBlock) || typeof originalBlock.type !== \"string\") continue;\n\n\t\tif (originalBlock.type === \"text\") {\n\t\t\tconst providerText = takeProviderBlock(providerBlocks, providerIndex, \"text\");\n\t\t\tif (providerText.block) {\n\t\t\t\trepaired.push(providerText.block);\n\t\t\t\tproviderIndex = providerText.nextProviderIndex;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tconst fallback = fallbackTextBlock(originalBlock);\n\t\t\tif (fallback) repaired.push(fallback);\n\t\t\tcontinue;\n\t\t}\n\n\t\tif (originalBlock.type === \"toolCall\") {\n\t\t\tconst providerToolUse = takeProviderBlock(providerBlocks, providerIndex, \"tool_use\");\n\t\t\tif (providerToolUse.block) {\n\t\t\t\trepaired.push(providerToolUse.block);\n\t\t\t\tproviderIndex = providerToolUse.nextProviderIndex;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tconst fallback = fallbackToolUseBlock(originalBlock);\n\t\t\tif (fallback) repaired.push(fallback);\n\t\t}\n\t}\n\n\tfor (; providerIndex < providerBlocks.length; providerIndex += 1) {\n\t\trepaired.push(providerBlocks[providerIndex]);\n\t}\n\n\treturn repaired;\n}\n\nfunction getAnthropicMessages(payload: unknown): unknown[] | undefined {\n\tif (!isObject(payload)) return undefined;\n\treturn Array.isArray(payload.messages) ? payload.messages : undefined;\n}\n\nfunction getAnthropicAssistantContent(message: unknown): AnthropicContentBlock[] | undefined {\n\tif (!isObject(message) || message.role !== \"assistant\") return undefined;\n\tif (!Array.isArray(message.content)) return undefined;\n\tconst content = message.content.filter(isAnthropicContentBlock);\n\treturn content.length === message.content.length ? content : undefined;\n}\n\n/**\n * Restore same-model Anthropic replay thinking blocks after provider payload\n * construction and extension payload hooks.\n *\n * Anthropic requires thinking/redacted_thinking blocks from replayed assistant\n * messages to remain byte-for-byte identical to the original response. The\n * upstream pi-ai Anthropic converter currently still sanitizes thinking text,\n * drops signed empty thinking, and does not understand raw redacted_thinking\n * blocks. This guard repairs the already-built Anthropic payload from the\n * pre-provider LLM messages while leaving non-Anthropic and cross-model payloads\n * unchanged.\n */\nexport function restoreAnthropicReplayThinkingBlocks(payload: unknown, sourceMessages: readonly Message[], model: Model<Api>): unknown {\n\tif (model.api !== \"anthropic-messages\") return payload;\n\n\tconst payloadMessages = getAnthropicMessages(payload);\n\tif (!payloadMessages) return payload;\n\n\tconst assistantPayloads = payloadMessages\n\t\t.map((message, index) => ({ message, index, content: getAnthropicAssistantContent(message) }))\n\t\t.filter(\n\t\t\t(entry): entry is { message: JsonObject; index: number; content: AnthropicContentBlock[] } =>\n\t\t\t\tentry.content !== undefined && isObject(entry.message),\n\t\t);\n\n\tif (assistantPayloads.length === 0) return payload;\n\n\tconst allowEmptySignature = isObject(model.compat) && model.compat.allowEmptySignature === true;\n\tlet assistantPayloadOrdinal = 0;\n\tlet nextPayloadMessages: unknown[] | undefined;\n\n\tfor (const sourceMessage of sourceMessages) {\n\t\tif (sourceMessage.role !== \"assistant\") continue;\n\t\tif (!legacyProviderWouldEmitAssistant(sourceMessage, model, allowEmptySignature)) continue;\n\n\t\tconst payloadAssistant = assistantPayloads[assistantPayloadOrdinal];\n\t\tassistantPayloadOrdinal += 1;\n\t\tif (!payloadAssistant) break;\n\n\t\tif (!isSameModelAssistant(sourceMessage, model) || !hasReplayThinkingBlock(sourceMessage)) continue;\n\n\t\tconst repairedContent = repairAssistantContent(sourceMessage, payloadAssistant.content);\n\t\tif (JSON.stringify(repairedContent) === JSON.stringify(payloadAssistant.content)) continue;\n\n\t\tnextPayloadMessages ??= [...payloadMessages];\n\t\tnextPayloadMessages[payloadAssistant.index] = {\n\t\t\t...payloadAssistant.message,\n\t\t\tcontent: repairedContent,\n\t\t};\n\t}\n\n\tif (!nextPayloadMessages || !isObject(payload)) return payload;\n\treturn { ...payload, messages: nextPayloadMessages };\n}\n"]}

package/dist/core/anthropic-thinking-guard.js

@@ -0,0 +1,205 @@
+function isObject(value) {
+    return value !== null && typeof value === "object" && !Array.isArray(value);
+}
+function isAnthropicContentBlock(value) {
+    return isObject(value) && typeof value.type === "string";
+}
+function isThinkingLikeAnthropicBlock(value) {
+    if (!isAnthropicContentBlock(value))
+        return false;
+    return value.type === "thinking" || value.type === "redacted_thinking";
+}
+function isSameModelAssistant(message, model) {
+    return message.provider === model.provider && message.api === model.api && message.model === model.id;
+}
+function readReplayThinkingBlock(block) {
+    if (!isObject(block) || typeof block.type !== "string")
+        return undefined;
+    if (block.type === "redacted_thinking") {
+        return typeof block.data === "string" ? { type: "redacted_thinking", data: block.data } : undefined;
+    }
+    if (block.type !== "thinking")
+        return undefined;
+    if (block.redacted === true) {
+        return typeof block.thinkingSignature === "string"
+            ? { type: "redacted_thinking", data: block.thinkingSignature }
+            : undefined;
+    }
+    return typeof block.thinking === "string" && typeof block.thinkingSignature === "string"
+        ? { type: "thinking", thinking: block.thinking, signature: block.thinkingSignature }
+        : undefined;
+}
+function hasReplayThinkingBlock(message) {
+    return message.content.some((block) => readReplayThinkingBlock(block) !== undefined);
+}
+function legacyProviderWouldEmitAssistantBlock(block, allowEmptySignature) {
+    if (!isObject(block) || typeof block.type !== "string")
+        return false;
+    if (block.type === "text") {
+        return typeof block.text === "string" && block.text.trim().length > 0;
+    }
+    if (block.type === "toolCall") {
+        return true;
+    }
+    if (block.type !== "thinking") {
+        return false;
+    }
+    if (block.redacted === true) {
+        return true;
+    }
+    if (typeof block.thinking !== "string")
+        return false;
+    if (block.thinking.trim().length === 0)
+        return false;
+    const signature = typeof block.thinkingSignature === "string" ? block.thinkingSignature : undefined;
+    if (signature !== undefined && signature.trim().length > 0)
+        return true;
+    return allowEmptySignature;
+}
+function legacyProviderWouldEmitAssistant(message, model, allowEmptySignature) {
+    if (message.stopReason === "error" || message.stopReason === "aborted")
+        return false;
+    // This mirrors the relevant same-model branch in @earendil-works/pi-ai's
+    // transformMessages() + Anthropic convertMessages() pipeline closely enough
+    // to keep assistant ordinal mapping aligned with the provider payload.
+    if (isSameModelAssistant(message, model)) {
+        return message.content.some((block) => legacyProviderWouldEmitAssistantBlock(block, allowEmptySignature));
+    }
+    return message.content.some((block) => {
+        if (!isObject(block) || typeof block.type !== "string")
+            return false;
+        if (block.type === "text")
+            return typeof block.text === "string" && block.text.trim().length > 0;
+        if (block.type === "toolCall")
+            return true;
+        if (block.type !== "thinking")
+            return false;
+        if (block.redacted === true)
+            return false;
+        return typeof block.thinking === "string" && block.thinking.trim().length > 0;
+    });
+}
+function fallbackTextBlock(block) {
+    return typeof block.text === "string" && block.text.trim().length > 0
+        ? { type: "text", text: block.text }
+        : undefined;
+}
+function fallbackToolUseBlock(block) {
+    if (typeof block.id !== "string" || typeof block.name !== "string")
+        return undefined;
+    return { type: "tool_use", id: block.id, name: block.name, input: block.arguments ?? {} };
+}
+function takeProviderBlock(providerBlocks, providerIndex, expectedType) {
+    const providerBlock = providerBlocks[providerIndex];
+    if (providerBlock?.type === expectedType) {
+        return { block: providerBlock, nextProviderIndex: providerIndex + 1 };
+    }
+    return { block: undefined, nextProviderIndex: providerIndex };
+}
+function repairAssistantContent(originalMessage, providerBlocks) {
+    const repaired = [];
+    let providerIndex = 0;
+    for (const originalBlock of originalMessage.content) {
+        const replayBlock = readReplayThinkingBlock(originalBlock);
+        if (replayBlock) {
+            if (isThinkingLikeAnthropicBlock(providerBlocks[providerIndex])) {
+                providerIndex += 1;
+            }
+            repaired.push({ ...replayBlock });
+            continue;
+        }
+        if (!isObject(originalBlock) || typeof originalBlock.type !== "string")
+            continue;
+        if (originalBlock.type === "text") {
+            const providerText = takeProviderBlock(providerBlocks, providerIndex, "text");
+            if (providerText.block) {
+                repaired.push(providerText.block);
+                providerIndex = providerText.nextProviderIndex;
+                continue;
+            }
+            const fallback = fallbackTextBlock(originalBlock);
+            if (fallback)
+                repaired.push(fallback);
+            continue;
+        }
+        if (originalBlock.type === "toolCall") {
+            const providerToolUse = takeProviderBlock(providerBlocks, providerIndex, "tool_use");
+            if (providerToolUse.block) {
+                repaired.push(providerToolUse.block);
+                providerIndex = providerToolUse.nextProviderIndex;
+                continue;
+            }
+            const fallback = fallbackToolUseBlock(originalBlock);
+            if (fallback)
+                repaired.push(fallback);
+        }
+    }
+    for (; providerIndex < providerBlocks.length; providerIndex += 1) {
+        repaired.push(providerBlocks[providerIndex]);
+    }
+    return repaired;
+}
+function getAnthropicMessages(payload) {
+    if (!isObject(payload))
+        return undefined;
+    return Array.isArray(payload.messages) ? payload.messages : undefined;
+}
+function getAnthropicAssistantContent(message) {
+    if (!isObject(message) || message.role !== "assistant")
+        return undefined;
+    if (!Array.isArray(message.content))
+        return undefined;
+    const content = message.content.filter(isAnthropicContentBlock);
+    return content.length === message.content.length ? content : undefined;
+}
+/**
+ * Restore same-model Anthropic replay thinking blocks after provider payload
+ * construction and extension payload hooks.
+ *
+ * Anthropic requires thinking/redacted_thinking blocks from replayed assistant
+ * messages to remain byte-for-byte identical to the original response. The
+ * upstream pi-ai Anthropic converter currently still sanitizes thinking text,
+ * drops signed empty thinking, and does not understand raw redacted_thinking
+ * blocks. This guard repairs the already-built Anthropic payload from the
+ * pre-provider LLM messages while leaving non-Anthropic and cross-model payloads
+ * unchanged.
+ */
+export function restoreAnthropicReplayThinkingBlocks(payload, sourceMessages, model) {
+    if (model.api !== "anthropic-messages")
+        return payload;
+    const payloadMessages = getAnthropicMessages(payload);
+    if (!payloadMessages)
+        return payload;
+    const assistantPayloads = payloadMessages
+        .map((message, index) => ({ message, index, content: getAnthropicAssistantContent(message) }))
+        .filter((entry) => entry.content !== undefined && isObject(entry.message));
+    if (assistantPayloads.length === 0)
+        return payload;
+    const allowEmptySignature = isObject(model.compat) && model.compat.allowEmptySignature === true;
+    let assistantPayloadOrdinal = 0;
+    let nextPayloadMessages;
+    for (const sourceMessage of sourceMessages) {
+        if (sourceMessage.role !== "assistant")
+            continue;
+        if (!legacyProviderWouldEmitAssistant(sourceMessage, model, allowEmptySignature))
+            continue;
+        const payloadAssistant = assistantPayloads[assistantPayloadOrdinal];
+        assistantPayloadOrdinal += 1;
+        if (!payloadAssistant)
+            break;
+        if (!isSameModelAssistant(sourceMessage, model) || !hasReplayThinkingBlock(sourceMessage))
+            continue;
+        const repairedContent = repairAssistantContent(sourceMessage, payloadAssistant.content);
+        if (JSON.stringify(repairedContent) === JSON.stringify(payloadAssistant.content))
+            continue;
+        nextPayloadMessages ??= [...payloadMessages];
+        nextPayloadMessages[payloadAssistant.index] = {
+            ...payloadAssistant.message,
+            content: repairedContent,
+        };
+    }
+    if (!nextPayloadMessages || !isObject(payload))
+        return payload;
+    return { ...payload, messages: nextPayloadMessages };
+}
+//# sourceMappingURL=anthropic-thinking-guard.js.map

package/dist/core/anthropic-thinking-guard.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"anthropic-thinking-guard.js","sourceRoot":"","sources":["../../src/core/anthropic-thinking-guard.ts"],"names":[],"mappings":"AAUA,SAAS,QAAQ,CAAC,KAAc;IAC/B,OAAO,KAAK,KAAK,IAAI,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;AAC7E,CAAC;AAED,SAAS,uBAAuB,CAAC,KAAc;IAC9C,OAAO,QAAQ,CAAC,KAAK,CAAC,IAAI,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ,CAAC;AAC1D,CAAC;AAED,SAAS,4BAA4B,CAAC,KAAc;IACnD,IAAI,CAAC,uBAAuB,CAAC,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC;IAClD,OAAO,KAAK,CAAC,IAAI,KAAK,UAAU,IAAI,KAAK,CAAC,IAAI,KAAK,mBAAmB,CAAC;AACxE,CAAC;AAED,SAAS,oBAAoB,CAAC,OAAyB,EAAE,KAAiB;IACzE,OAAO,OAAO,CAAC,QAAQ,KAAK,KAAK,CAAC,QAAQ,IAAI,OAAO,CAAC,GAAG,KAAK,KAAK,CAAC,GAAG,IAAI,OAAO,CAAC,KAAK,KAAK,KAAK,CAAC,EAAE,CAAC;AACvG,CAAC;AAED,SAAS,uBAAuB,CAAC,KAAc;IAC9C,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ;QAAE,OAAO,SAAS,CAAC;IAEzE,IAAI,KAAK,CAAC,IAAI,KAAK,mBAAmB,EAAE,CAAC;QACxC,OAAO,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;IACrG,CAAC;IAED,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU;QAAE,OAAO,SAAS,CAAC;IAEhD,IAAI,KAAK,CAAC,QAAQ,KAAK,IAAI,EAAE,CAAC;QAC7B,OAAO,OAAO,KAAK,CAAC,iBAAiB,KAAK,QAAQ;YACjD,CAAC,CAAC,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,KAAK,CAAC,iBAAiB,EAAE;YAC9D,CAAC,CAAC,SAAS,CAAC;IACd,CAAC;IAED,OAAO,OAAO,KAAK,CAAC,QAAQ,KAAK,QAAQ,IAAI,OAAO,KAAK,CAAC,iBAAiB,KAAK,QAAQ;QACvF,CAAC,CAAC,EAAE,IAAI,EAAE,UAAU,EAAE,QAAQ,EAAE,KAAK,CAAC,QAAQ,EAAE,SAAS,EAAE,KAAK,CAAC,iBAAiB,EAAE;QACpF,CAAC,CAAC,SAAS,CAAC;AACd,CAAC;AAED,SAAS,sBAAsB,CAAC,OAAyB;IACxD,OAAO,OAAO,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,uBAAuB,CAAC,KAAK,CAAC,KAAK,SAAS,CAAC,CAAC;AACtF,CAAC;AAED,SAAS,qCAAqC,CAAC,KAAc,EAAE,mBAA4B;IAC1F,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ;QAAE,OAAO,KAAK,CAAC;IAErE,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;QAC3B,OAAO,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ,IAAI,KAAK,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC,CAAC;IACvE,CAAC;IAED,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;QAC/B,OAAO,IAAI,CAAC;IACb,CAAC;IAED,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;QAC/B,OAAO,KAAK,CAAC;IACd,CAAC;IAED,IAAI,KAAK,CAAC,QAAQ,KAAK,IAAI,EAAE,CAAC;QAC7B,OAAO,IAAI,CAAC;IACb,CAAC;IAED,IAAI,OAAO,KAAK,CAAC,QAAQ,KAAK,QAAQ;QAAE,OAAO,KAAK,CAAC;IACrD,IAAI,KAAK,CAAC,QAAQ,CAAC,IAAI,EAAE,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC;IAErD,MAAM,SAAS,GAAG,OAAO,KAAK,CAAC,iBAAiB,KAAK,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,iBAAiB,CAAC,CAAC,CAAC,SAAS,CAAC;IACpG,IAAI,SAAS,KAAK,SAAS,IAAI,SAAS,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,IAAI,CAAC;IACxE,OAAO,mBAAmB,CAAC;AAC5B,CAAC;AAED,SAAS,gCAAgC,CAAC,OAAyB,EAAE,KAAiB,EAAE,mBAA4B;IACnH,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,OAAO,CAAC,UAAU,KAAK,SAAS;QAAE,OAAO,KAAK,CAAC;IAErF,yEAAyE;IACzE,4EAA4E;IAC5E,uEAAuE;IACvE,IAAI,oBAAoB,CAAC,OAAO,EAAE,KAAK,CAAC,EAAE,CAAC;QAC1C,OAAO,OAAO,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,qCAAqC,CAAC,KAAK,EAAE,mBAAmB,CAAC,CAAC,CAAC;IAC3G,CAAC;IAED,OAAO,OAAO,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,EAAE;QACrC,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QACrE,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM;YAAE,OAAO,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ,IAAI,KAAK,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC,CAAC;QACjG,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU;YAAE,OAAO,IAAI,CAAC;QAC3C,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU;YAAE,OAAO,KAAK,CAAC;QAC5C,IAAI,KAAK,CAAC,QAAQ,KAAK,IAAI;YAAE,OAAO,KAAK,CAAC;QAC1C,OAAO,OAAO,KAAK,CAAC,QAAQ,KAAK,QAAQ,IAAI,KAAK,CAAC,QAAQ,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC,CAAC;IAC/E,CAAC,CAAC,CAAC;AACJ,CAAC;AAED,SAAS,iBAAiB,CAAC,KAAiB;IAC3C,OAAO,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ,IAAI,KAAK,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC;QACpE,CAAC,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE;QACpC,CAAC,CAAC,SAAS,CAAC;AACd,CAAC;AAED,SAAS,oBAAoB,CAAC,KAAiB;IAC9C,IAAI,OAAO,KAAK,CAAC,EAAE,KAAK,QAAQ,IAAI,OAAO,KAAK,CAAC,IAAI,KAAK,QAAQ;QAAE,OAAO,SAAS,CAAC;IACrF,OAAO,EAAE,IAAI,EAAE,UAAU,EAAE,EAAE,EAAE,KAAK,CAAC,EAAE,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE,KAAK,EAAE,KAAK,CAAC,SAAS,IAAI,EAAE,EAAE,CAAC;AAC3F,CAAC;AAED,SAAS,iBAAiB,CACzB,cAAgD,EAChD,aAAqB,EACrB,YAAoB;IAEpB,MAAM,aAAa,GAAG,cAAc,CAAC,aAAa,CAAC,CAAC;IACpD,IAAI,aAAa,EAAE,IAAI,KAAK,YAAY,EAAE,CAAC;QAC1C,OAAO,EAAE,KAAK,EAAE,aAAa,EAAE,iBAAiB,EAAE,aAAa,GAAG,CAAC,EAAE,CAAC;IACvE,CAAC;IACD,OAAO,EAAE,KAAK,EAAE,SAAS,EAAE,iBAAiB,EAAE,aAAa,EAAE,CAAC;AAC/D,CAAC;AAED,SAAS,sBAAsB,CAC9B,eAAiC,EACjC,cAAgD;IAEhD,MAAM,QAAQ,GAA4B,EAAE,CAAC;IAC7C,IAAI,aAAa,GAAG,CAAC,CAAC;IAEtB,KAAK,MAAM,aAAa,IAAI,eAAe,CAAC,OAAO,EAAE,CAAC;QACrD,MAAM,WAAW,GAAG,uBAAuB,CAAC,aAAa,CAAC,CAAC;QAC3D,IAAI,WAAW,EAAE,CAAC;YACjB,IAAI,4BAA4B,CAAC,cAAc,CAAC,aAAa,CAAC,CAAC,EAAE,CAAC;gBACjE,aAAa,IAAI,CAAC,CAAC;YACpB,CAAC;YACD,QAAQ,CAAC,IAAI,CAAC,EAAE,GAAG,WAAW,EAAE,CAAC,CAAC;YAClC,SAAS;QACV,CAAC;QAED,IAAI,CAAC,QAAQ,CAAC,aAAa,CAAC,IAAI,OAAO,aAAa,CAAC,IAAI,KAAK,QAAQ;YAAE,SAAS;QAEjF,IAAI,aAAa,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YACnC,MAAM,YAAY,GAAG,iBAAiB,CAAC,cAAc,EAAE,aAAa,EAAE,MAAM,CAAC,CAAC;YAC9E,IAAI,YAAY,CAAC,KAAK,EAAE,CAAC;gBACxB,QAAQ,CAAC,IAAI,CAAC,YAAY,CAAC,KAAK,CAAC,CAAC;gBAClC,aAAa,GAAG,YAAY,CAAC,iBAAiB,CAAC;gBAC/C,SAAS;YACV,CAAC;YACD,MAAM,QAAQ,GAAG,iBAAiB,CAAC,aAAa,CAAC,CAAC;YAClD,IAAI,QAAQ;gBAAE,QAAQ,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;YACtC,SAAS;QACV,CAAC;QAED,IAAI,aAAa,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;YACvC,MAAM,eAAe,GAAG,iBAAiB,CAAC,cAAc,EAAE,aAAa,EAAE,UAAU,CAAC,CAAC;YACrF,IAAI,eAAe,CAAC,KAAK,EAAE,CAAC;gBAC3B,QAAQ,CAAC,IAAI,CAAC,eAAe,CAAC,KAAK,CAAC,CAAC;gBACrC,aAAa,GAAG,eAAe,CAAC,iBAAiB,CAAC;gBAClD,SAAS;YACV,CAAC;YACD,MAAM,QAAQ,GAAG,oBAAoB,CAAC,aAAa,CAAC,CAAC;YACrD,IAAI,QAAQ;gBAAE,QAAQ,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;QACvC,CAAC;IACF,CAAC;IAED,OAAO,aAAa,GAAG,cAAc,CAAC,MAAM,EAAE,aAAa,IAAI,CAAC,EAAE,CAAC;QAClE,QAAQ,CAAC,IAAI,CAAC,cAAc,CAAC,aAAa,CAAC,CAAC,CAAC;IAC9C,CAAC;IAED,OAAO,QAAQ,CAAC;AACjB,CAAC;AAED,SAAS,oBAAoB,CAAC,OAAgB;IAC7C,IAAI,CAAC,QAAQ,CAAC,OAAO,CAAC;QAAE,OAAO,SAAS,CAAC;IACzC,OAAO,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,QAAQ,CAAC,CAAC,CAAC,SAAS,CAAC;AACvE,CAAC;AAED,SAAS,4BAA4B,CAAC,OAAgB;IACrD,IAAI,CAAC,QAAQ,CAAC,OAAO,CAAC,IAAI,OAAO,CAAC,IAAI,KAAK,WAAW;QAAE,OAAO,SAAS,CAAC;IACzE,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC,OAAO,CAAC;QAAE,OAAO,SAAS,CAAC;IACtD,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC,uBAAuB,CAAC,CAAC;IAChE,OAAO,OAAO,CAAC,MAAM,KAAK,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,SAAS,CAAC;AACxE,CAAC;AAED;;;;;;;;;;;GAWG;AACH,MAAM,UAAU,oCAAoC,CAAC,OAAgB,EAAE,cAAkC,EAAE,KAAiB;IAC3H,IAAI,KAAK,CAAC,GAAG,KAAK,oBAAoB;QAAE,OAAO,OAAO,CAAC;IAEvD,MAAM,eAAe,GAAG,oBAAoB,CAAC,OAAO,CAAC,CAAC;IACtD,IAAI,CAAC,eAAe;QAAE,OAAO,OAAO,CAAC;IAErC,MAAM,iBAAiB,GAAG,eAAe;SACvC,GAAG,CAAC,CAAC,OAAO,EAAE,KAAK,EAAE,EAAE,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,EAAE,OAAO,EAAE,4BAA4B,CAAC,OAAO,CAAC,EAAE,CAAC,CAAC;SAC7F,MAAM,CACN,CAAC,KAAK,EAAqF,EAAE,CAC5F,KAAK,CAAC,OAAO,KAAK,SAAS,IAAI,QAAQ,CAAC,KAAK,CAAC,OAAO,CAAC,CACvD,CAAC;IAEH,IAAI,iBAAiB,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,OAAO,CAAC;IAEnD,MAAM,mBAAmB,GAAG,QAAQ,CAAC,KAAK,CAAC,MAAM,CAAC,IAAI,KAAK,CAAC,MAAM,CAAC,mBAAmB,KAAK,IAAI,CAAC;IAChG,IAAI,uBAAuB,GAAG,CAAC,CAAC;IAChC,IAAI,mBAA0C,CAAC;IAE/C,KAAK,MAAM,aAAa,IAAI,cAAc,EAAE,CAAC;QAC5C,IAAI,aAAa,CAAC,IAAI,KAAK,WAAW;YAAE,SAAS;QACjD,IAAI,CAAC,gCAAgC,CAAC,aAAa,EAAE,KAAK,EAAE,mBAAmB,CAAC;YAAE,SAAS;QAE3F,MAAM,gBAAgB,GAAG,iBAAiB,CAAC,uBAAuB,CAAC,CAAC;QACpE,uBAAuB,IAAI,CAAC,CAAC;QAC7B,IAAI,CAAC,gBAAgB;YAAE,MAAM;QAE7B,IAAI,CAAC,oBAAoB,CAAC,aAAa,EAAE,KAAK,CAAC,IAAI,CAAC,sBAAsB,CAAC,aAAa,CAAC;YAAE,SAAS;QAEpG,MAAM,eAAe,GAAG,sBAAsB,CAAC,aAAa,EAAE,gBAAgB,CAAC,OAAO,CAAC,CAAC;QACxF,IAAI,IAAI,CAAC,SAAS,CAAC,eAAe,CAAC,KAAK,IAAI,CAAC,SAAS,CAAC,gBAAgB,CAAC,OAAO,CAAC;YAAE,SAAS;QAE3F,mBAAmB,KAAK,CAAC,GAAG,eAAe,CAAC,CAAC;QAC7C,mBAAmB,CAAC,gBAAgB,CAAC,KAAK,CAAC,GAAG;YAC7C,GAAG,gBAAgB,CAAC,OAAO;YAC3B,OAAO,EAAE,eAAe;SACxB,CAAC;IACH,CAAC;IAED,IAAI,CAAC,mBAAmB,IAAI,CAAC,QAAQ,CAAC,OAAO,CAAC;QAAE,OAAO,OAAO,CAAC;IAC/D,OAAO,EAAE,GAAG,OAAO,EAAE,QAAQ,EAAE,mBAAmB,EAAE,CAAC;AACtD,CAAC","sourcesContent":["import type { Api, AssistantMessage, Message, Model } from \"@earendil-works/pi-ai\";\n\ntype JsonObject = Record<string, unknown>;\n\ntype AnthropicContentBlock = JsonObject & { type: string };\n\ntype ReplayThinkingBlock =\n\t| { type: \"thinking\"; thinking: string; signature: string }\n\t| { type: \"redacted_thinking\"; data: string };\n\nfunction isObject(value: unknown): value is JsonObject {\n\treturn value !== null && typeof value === \"object\" && !Array.isArray(value);\n}\n\nfunction isAnthropicContentBlock(value: unknown): value is AnthropicContentBlock {\n\treturn isObject(value) && typeof value.type === \"string\";\n}\n\nfunction isThinkingLikeAnthropicBlock(value: unknown): boolean {\n\tif (!isAnthropicContentBlock(value)) return false;\n\treturn value.type === \"thinking\" || value.type === \"redacted_thinking\";\n}\n\nfunction isSameModelAssistant(message: AssistantMessage, model: Model<Api>): boolean {\n\treturn message.provider === model.provider && message.api === model.api && message.model === model.id;\n}\n\nfunction readReplayThinkingBlock(block: unknown): ReplayThinkingBlock | undefined {\n\tif (!isObject(block) || typeof block.type !== \"string\") return undefined;\n\n\tif (block.type === \"redacted_thinking\") {\n\t\treturn typeof block.data === \"string\" ? { type: \"redacted_thinking\", data: block.data } : undefined;\n\t}\n\n\tif (block.type !== \"thinking\") return undefined;\n\n\tif (block.redacted === true) {\n\t\treturn typeof block.thinkingSignature === \"string\"\n\t\t\t? { type: \"redacted_thinking\", data: block.thinkingSignature }\n\t\t\t: undefined;\n\t}\n\n\treturn typeof block.thinking === \"string\" && typeof block.thinkingSignature === \"string\"\n\t\t? { type: \"thinking\", thinking: block.thinking, signature: block.thinkingSignature }\n\t\t: undefined;\n}\n\nfunction hasReplayThinkingBlock(message: AssistantMessage): boolean {\n\treturn message.content.some((block) => readReplayThinkingBlock(block) !== undefined);\n}\n\nfunction legacyProviderWouldEmitAssistantBlock(block: unknown, allowEmptySignature: boolean): boolean {\n\tif (!isObject(block) || typeof block.type !== \"string\") return false;\n\n\tif (block.type === \"text\") {\n\t\treturn typeof block.text === \"string\" && block.text.trim().length > 0;\n\t}\n\n\tif (block.type === \"toolCall\") {\n\t\treturn true;\n\t}\n\n\tif (block.type !== \"thinking\") {\n\t\treturn false;\n\t}\n\n\tif (block.redacted === true) {\n\t\treturn true;\n\t}\n\n\tif (typeof block.thinking !== \"string\") return false;\n\tif (block.thinking.trim().length === 0) return false;\n\n\tconst signature = typeof block.thinkingSignature === \"string\" ? block.thinkingSignature : undefined;\n\tif (signature !== undefined && signature.trim().length > 0) return true;\n\treturn allowEmptySignature;\n}\n\nfunction legacyProviderWouldEmitAssistant(message: AssistantMessage, model: Model<Api>, allowEmptySignature: boolean): boolean {\n\tif (message.stopReason === \"error\" || message.stopReason === \"aborted\") return false;\n\n\t// This mirrors the relevant same-model branch in @earendil-works/pi-ai's\n\t// transformMessages() + Anthropic convertMessages() pipeline closely enough\n\t// to keep assistant ordinal mapping aligned with the provider payload.\n\tif (isSameModelAssistant(message, model)) {\n\t\treturn message.content.some((block) => legacyProviderWouldEmitAssistantBlock(block, allowEmptySignature));\n\t}\n\n\treturn message.content.some((block) => {\n\t\tif (!isObject(block) || typeof block.type !== \"string\") return false;\n\t\tif (block.type === \"text\") return typeof block.text === \"string\" && block.text.trim().length > 0;\n\t\tif (block.type === \"toolCall\") return true;\n\t\tif (block.type !== \"thinking\") return false;\n\t\tif (block.redacted === true) return false;\n\t\treturn typeof block.thinking === \"string\" && block.thinking.trim().length > 0;\n\t});\n}\n\nfunction fallbackTextBlock(block: JsonObject): AnthropicContentBlock | undefined {\n\treturn typeof block.text === \"string\" && block.text.trim().length > 0\n\t\t? { type: \"text\", text: block.text }\n\t\t: undefined;\n}\n\nfunction fallbackToolUseBlock(block: JsonObject): AnthropicContentBlock | undefined {\n\tif (typeof block.id !== \"string\" || typeof block.name !== \"string\") return undefined;\n\treturn { type: \"tool_use\", id: block.id, name: block.name, input: block.arguments ?? {} };\n}\n\nfunction takeProviderBlock(\n\tproviderBlocks: readonly AnthropicContentBlock[],\n\tproviderIndex: number,\n\texpectedType: string,\n): { block: AnthropicContentBlock | undefined; nextProviderIndex: number } {\n\tconst providerBlock = providerBlocks[providerIndex];\n\tif (providerBlock?.type === expectedType) {\n\t\treturn { block: providerBlock, nextProviderIndex: providerIndex + 1 };\n\t}\n\treturn { block: undefined, nextProviderIndex: providerIndex };\n}\n\nfunction repairAssistantContent(\n\toriginalMessage: AssistantMessage,\n\tproviderBlocks: readonly AnthropicContentBlock[],\n): AnthropicContentBlock[] {\n\tconst repaired: AnthropicContentBlock[] = [];\n\tlet providerIndex = 0;\n\n\tfor (const originalBlock of originalMessage.content) {\n\t\tconst replayBlock = readReplayThinkingBlock(originalBlock);\n\t\tif (replayBlock) {\n\t\t\tif (isThinkingLikeAnthropicBlock(providerBlocks[providerIndex])) {\n\t\t\t\tproviderIndex += 1;\n\t\t\t}\n\t\t\trepaired.push({ ...replayBlock });\n\t\t\tcontinue;\n\t\t}\n\n\t\tif (!isObject(originalBlock) || typeof originalBlock.type !== \"string\") continue;\n\n\t\tif (originalBlock.type === \"text\") {\n\t\t\tconst providerText = takeProviderBlock(providerBlocks, providerIndex, \"text\");\n\t\t\tif (providerText.block) {\n\t\t\t\trepaired.push(providerText.block);\n\t\t\t\tproviderIndex = providerText.nextProviderIndex;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tconst fallback = fallbackTextBlock(originalBlock);\n\t\t\tif (fallback) repaired.push(fallback);\n\t\t\tcontinue;\n\t\t}\n\n\t\tif (originalBlock.type === \"toolCall\") {\n\t\t\tconst providerToolUse = takeProviderBlock(providerBlocks, providerIndex, \"tool_use\");\n\t\t\tif (providerToolUse.block) {\n\t\t\t\trepaired.push(providerToolUse.block);\n\t\t\t\tproviderIndex = providerToolUse.nextProviderIndex;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tconst fallback = fallbackToolUseBlock(originalBlock);\n\t\t\tif (fallback) repaired.push(fallback);\n\t\t}\n\t}\n\n\tfor (; providerIndex < providerBlocks.length; providerIndex += 1) {\n\t\trepaired.push(providerBlocks[providerIndex]);\n\t}\n\n\treturn repaired;\n}\n\nfunction getAnthropicMessages(payload: unknown): unknown[] | undefined {\n\tif (!isObject(payload)) return undefined;\n\treturn Array.isArray(payload.messages) ? payload.messages : undefined;\n}\n\nfunction getAnthropicAssistantContent(message: unknown): AnthropicContentBlock[] | undefined {\n\tif (!isObject(message) || message.role !== \"assistant\") return undefined;\n\tif (!Array.isArray(message.content)) return undefined;\n\tconst content = message.content.filter(isAnthropicContentBlock);\n\treturn content.length === message.content.length ? content : undefined;\n}\n\n/**\n * Restore same-model Anthropic replay thinking blocks after provider payload\n * construction and extension payload hooks.\n *\n * Anthropic requires thinking/redacted_thinking blocks from replayed assistant\n * messages to remain byte-for-byte identical to the original response. The\n * upstream pi-ai Anthropic converter currently still sanitizes thinking text,\n * drops signed empty thinking, and does not understand raw redacted_thinking\n * blocks. This guard repairs the already-built Anthropic payload from the\n * pre-provider LLM messages while leaving non-Anthropic and cross-model payloads\n * unchanged.\n */\nexport function restoreAnthropicReplayThinkingBlocks(payload: unknown, sourceMessages: readonly Message[], model: Model<Api>): unknown {\n\tif (model.api !== \"anthropic-messages\") return payload;\n\n\tconst payloadMessages = getAnthropicMessages(payload);\n\tif (!payloadMessages) return payload;\n\n\tconst assistantPayloads = payloadMessages\n\t\t.map((message, index) => ({ message, index, content: getAnthropicAssistantContent(message) }))\n\t\t.filter(\n\t\t\t(entry): entry is { message: JsonObject; index: number; content: AnthropicContentBlock[] } =>\n\t\t\t\tentry.content !== undefined && isObject(entry.message),\n\t\t);\n\n\tif (assistantPayloads.length === 0) return payload;\n\n\tconst allowEmptySignature = isObject(model.compat) && model.compat.allowEmptySignature === true;\n\tlet assistantPayloadOrdinal = 0;\n\tlet nextPayloadMessages: unknown[] | undefined;\n\n\tfor (const sourceMessage of sourceMessages) {\n\t\tif (sourceMessage.role !== \"assistant\") continue;\n\t\tif (!legacyProviderWouldEmitAssistant(sourceMessage, model, allowEmptySignature)) continue;\n\n\t\tconst payloadAssistant = assistantPayloads[assistantPayloadOrdinal];\n\t\tassistantPayloadOrdinal += 1;\n\t\tif (!payloadAssistant) break;\n\n\t\tif (!isSameModelAssistant(sourceMessage, model) || !hasReplayThinkingBlock(sourceMessage)) continue;\n\n\t\tconst repairedContent = repairAssistantContent(sourceMessage, payloadAssistant.content);\n\t\tif (JSON.stringify(repairedContent) === JSON.stringify(payloadAssistant.content)) continue;\n\n\t\tnextPayloadMessages ??= [...payloadMessages];\n\t\tnextPayloadMessages[payloadAssistant.index] = {\n\t\t\t...payloadAssistant.message,\n\t\t\tcontent: repairedContent,\n\t\t};\n\t}\n\n\tif (!nextPayloadMessages || !isObject(payload)) return payload;\n\treturn { ...payload, messages: nextPayloadMessages };\n}\n"]}

package/dist/core/compaction/compaction.d.ts

@@ -16,8 +16,14 @@ export interface CompactionSettings {
 }
 export declare const DEFAULT_COMPACTION_SETTINGS: CompactionSettings;
 /**
- * Calculate total context tokens from usage.
- * Uses the native totalTokens field when available, falls back to computing from components.
+ * Calculate active context-window tokens from provider usage.
+ *
+ * Prefer normalized component fields over `totalTokens`: some providers expose
+ * `totalTokens` as a billing/cumulative total, while the footer needs an active
+ * context estimate. Anthropic-compatible endpoints can also mirror cached input
+ * in both `input` and `cacheRead`/`cacheWrite`; when cache buckets are nearly the
+ * same size as `input`, treat `input` as the full prompt instead of counting the
+ * same cached prompt twice.
  */
 export declare function calculateContextTokens(usage: Usage): number;
 /**

package/dist/core/compaction/compaction.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"compaction.d.ts","sourceRoot":"","sources":["../../../src/core/compaction/compaction.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,+BAA+B,CAAC;AAClE,OAAO,KAAK,EAAoB,KAAK,EAAE,MAAM,uBAAuB,CAAC;AACrE,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,uBAAuB,CAAC;AAE1D,MAAM,WAAW,kBAAkB;IAClC,OAAO,EAAE,OAAO,CAAC;IACjB,aAAa,EAAE,MAAM,CAAC;IACtB,gFAAgF;IAChF,iBAAiB,EAAE,MAAM,CAAC;IAC1B,+EAA+E;IAC/E,eAAe,EAAE,MAAM,CAAC;IACxB,+FAA+F;IAC/F,KAAK,CAAC,EAAE,MAAM,CAAC;CACf;AAED,eAAO,MAAM,2BAA2B,EAAE,kBAKzC,CAAC;AAEF;;;GAGG;AACH,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE3D;AAgBD;;GAEG;AACH,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,YAAY,EAAE,GAAG,KAAK,GAAG,SAAS,CAShF;AAED,MAAM,WAAW,oBAAoB;IACpC,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,cAAc,EAAE,MAAM,CAAC;IACvB,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAUD;;;GAGG;AACH,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,YAAY,EAAE,GAAG,oBAAoB,CA4BpF;AAED;;GAEG;AACH,wBAAgB,aAAa,CAAC,aAAa,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM,EAAE,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAGjH;AAoBD;;;GAGG;AACH,wBAAgB,cAAc,CAAC,OAAO,EAAE,YAAY,GAAG,MAAM,CAuC5D","sourcesContent":["/**\n * Neutral context-usage metrics for deciding when a session needs compaction.\n */\n\nimport type { AgentMessage } from \"@earendil-works/pi-agent-core\";\nimport type { AssistantMessage, Usage } from \"@earendil-works/pi-ai\";\nimport type { SessionEntry } from \"../session-manager.ts\";\n\nexport interface CompactionSettings {\n\tenabled: boolean;\n\treserveTokens: number;\n\t/** Fraction of compactable context to keep. 0.3 is aggressive, 0.7 is light. */\n\tcompression_ratio: number;\n\t/** Number of recent context-eligible messages to preserve in standard mode. */\n\tpreserve_recent: number;\n\t/** Focus query for relevance-based pruning; auto-detected when omitted in settings/options. */\n\tquery?: string;\n}\n\nexport const DEFAULT_COMPACTION_SETTINGS: CompactionSettings = {\n\tenabled: true,\n\treserveTokens: 16384,\n\tcompression_ratio: 0.5,\n\tpreserve_recent: 2,\n};\n\n/**\n * Calculate total context tokens from usage.\n * Uses the native totalTokens field when available, falls back to computing from components.\n */\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\n}\n\n/**\n * Get usage from an assistant message if available.\n * Skips aborted and error messages as they don't have valid usage data.\n */\nfunction getAssistantUsage(msg: AgentMessage): Usage | undefined {\n\tif (msg.role === \"assistant\" && \"usage\" in msg) {\n\t\tconst assistantMsg = msg as AssistantMessage;\n\t\tif (assistantMsg.stopReason !== \"aborted\" && assistantMsg.stopReason !== \"error\" && assistantMsg.usage) {\n\t\t\treturn assistantMsg.usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n/**\n * Find the last non-aborted assistant message usage from session entries.\n */\nexport function getLastAssistantUsage(entries: SessionEntry[]): Usage | undefined {\n\tfor (let i = entries.length - 1; i >= 0; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type === \"message\") {\n\t\t\tconst usage = getAssistantUsage(entry.message);\n\t\t\tif (usage) return usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\nexport interface ContextUsageEstimate {\n\ttokens: number;\n\tusageTokens: number;\n\ttrailingTokens: number;\n\tlastUsageIndex: number | null;\n}\n\nfunction getLastAssistantUsageInfo(messages: AgentMessage[]): { usage: Usage; index: number } | undefined {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst usage = getAssistantUsage(messages[i]);\n\t\tif (usage) return { usage, index: i };\n\t}\n\treturn undefined;\n}\n\n/**\n * Estimate context tokens from messages, using the last assistant usage when available.\n * If there are messages after the last usage, estimate their tokens with estimateTokens.\n */\nexport function estimateContextTokens(messages: AgentMessage[]): ContextUsageEstimate {\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\n\tif (!usageInfo) {\n\t\tlet estimated = 0;\n\t\tfor (const message of messages) {\n\t\t\testimated += estimateTokens(message);\n\t\t}\n\t\treturn {\n\t\t\ttokens: estimated,\n\t\t\tusageTokens: 0,\n\t\t\ttrailingTokens: estimated,\n\t\t\tlastUsageIndex: null,\n\t\t};\n\t}\n\n\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\tlet trailingTokens = 0;\n\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\ttrailingTokens += estimateTokens(messages[i]);\n\t}\n\n\treturn {\n\t\ttokens: usageTokens + trailingTokens,\n\t\tusageTokens,\n\t\ttrailingTokens,\n\t\tlastUsageIndex: usageInfo.index,\n\t};\n}\n\n/**\n * Check if compaction should trigger based on context usage.\n */\nexport function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean {\n\tif (!settings.enabled) return false;\n\treturn contextTokens > contextWindow - settings.reserveTokens;\n}\n\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nfunction estimateTextAndImageContentChars(content: string | Array<{ type: string; text?: string }>): number {\n\tif (typeof content === \"string\") {\n\t\treturn content.length;\n\t}\n\n\tlet chars = 0;\n\tfor (const block of content) {\n\t\tif (block.type === \"text\" && block.text) {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"image\") {\n\t\t\tchars += ESTIMATED_IMAGE_CHARS;\n\t\t}\n\t}\n\treturn chars;\n}\n\n/**\n * Estimate token count for a message using chars/4 heuristic.\n * This is conservative (overestimates tokens).\n */\nexport function estimateTokens(message: AgentMessage): number {\n\tlet chars = 0;\n\n\tswitch (message.role) {\n\t\tcase \"user\": {\n\t\t\tchars = estimateTextAndImageContentChars(\n\t\t\t\t(message as { content: string | Array<{ type: string; text?: string }> }).content,\n\t\t\t);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"assistant\": {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\tfor (const block of assistant.content) {\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tchars += block.text.length;\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\tchars += block.thinking.length;\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tchars += block.name.length + JSON.stringify(block.arguments).length;\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"custom\":\n\t\tcase \"toolResult\": {\n\t\t\tchars = estimateTextAndImageContentChars(message.content);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"bashExecution\": {\n\t\t\tchars = message.command.length + message.output.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"branchSummary\": {\n\t\t\tchars = message.summary.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t}\n\n\treturn 0;\n}\n"]}
+{"version":3,"file":"compaction.d.ts","sourceRoot":"","sources":["../../../src/core/compaction/compaction.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,+BAA+B,CAAC;AAClE,OAAO,KAAK,EAAoB,KAAK,EAAE,MAAM,uBAAuB,CAAC;AACrE,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,uBAAuB,CAAC;AAE1D,MAAM,WAAW,kBAAkB;IAClC,OAAO,EAAE,OAAO,CAAC;IACjB,aAAa,EAAE,MAAM,CAAC;IACtB,gFAAgF;IAChF,iBAAiB,EAAE,MAAM,CAAC;IAC1B,+EAA+E;IAC/E,eAAe,EAAE,MAAM,CAAC;IACxB,+FAA+F;IAC/F,KAAK,CAAC,EAAE,MAAM,CAAC;CACf;AAED,eAAO,MAAM,2BAA2B,EAAE,kBAKzC,CAAC;AAEF;;;;;;;;;GASG;AACH,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAY3D;AAgBD;;GAEG;AACH,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,YAAY,EAAE,GAAG,KAAK,GAAG,SAAS,CAShF;AAED,MAAM,WAAW,oBAAoB;IACpC,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,cAAc,EAAE,MAAM,CAAC;IACvB,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAUD;;;GAGG;AACH,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,YAAY,EAAE,GAAG,oBAAoB,CA4BpF;AAED;;GAEG;AACH,wBAAgB,aAAa,CAAC,aAAa,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM,EAAE,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAGjH;AAoBD;;;GAGG;AACH,wBAAgB,cAAc,CAAC,OAAO,EAAE,YAAY,GAAG,MAAM,CAuC5D","sourcesContent":["/**\n * Neutral context-usage metrics for deciding when a session needs compaction.\n */\n\nimport type { AgentMessage } from \"@earendil-works/pi-agent-core\";\nimport type { AssistantMessage, Usage } from \"@earendil-works/pi-ai\";\nimport type { SessionEntry } from \"../session-manager.ts\";\n\nexport interface CompactionSettings {\n\tenabled: boolean;\n\treserveTokens: number;\n\t/** Fraction of compactable context to keep. 0.3 is aggressive, 0.7 is light. */\n\tcompression_ratio: number;\n\t/** Number of recent context-eligible messages to preserve in standard mode. */\n\tpreserve_recent: number;\n\t/** Focus query for relevance-based pruning; auto-detected when omitted in settings/options. */\n\tquery?: string;\n}\n\nexport const DEFAULT_COMPACTION_SETTINGS: CompactionSettings = {\n\tenabled: true,\n\treserveTokens: 16384,\n\tcompression_ratio: 0.5,\n\tpreserve_recent: 2,\n};\n\n/**\n * Calculate active context-window tokens from provider usage.\n *\n * Prefer normalized component fields over `totalTokens`: some providers expose\n * `totalTokens` as a billing/cumulative total, while the footer needs an active\n * context estimate. Anthropic-compatible endpoints can also mirror cached input\n * in both `input` and `cacheRead`/`cacheWrite`; when cache buckets are nearly the\n * same size as `input`, treat `input` as the full prompt instead of counting the\n * same cached prompt twice.\n */\nexport function calculateContextTokens(usage: Usage): number {\n\tconst input = Math.max(0, usage.input || 0);\n\tconst output = Math.max(0, usage.output || 0);\n\tconst cacheRead = Math.max(0, usage.cacheRead || 0);\n\tconst cacheWrite = Math.max(0, usage.cacheWrite || 0);\n\tconst cacheTokens = cacheRead + cacheWrite;\n\tconst hasComponents = input > 0 || output > 0 || cacheTokens > 0;\n\tif (!hasComponents) return Math.max(0, usage.totalTokens || 0);\n\n\tconst cacheMirrorsInput = input > 0 && cacheTokens > 0 && cacheTokens >= input * 0.9 && cacheTokens <= input * 1.1;\n\tconst promptTokens = cacheMirrorsInput ? input : input + cacheTokens;\n\treturn promptTokens + output;\n}\n\n/**\n * Get usage from an assistant message if available.\n * Skips aborted and error messages as they don't have valid usage data.\n */\nfunction getAssistantUsage(msg: AgentMessage): Usage | undefined {\n\tif (msg.role === \"assistant\" && \"usage\" in msg) {\n\t\tconst assistantMsg = msg as AssistantMessage;\n\t\tif (assistantMsg.stopReason !== \"aborted\" && assistantMsg.stopReason !== \"error\" && assistantMsg.usage) {\n\t\t\treturn assistantMsg.usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n/**\n * Find the last non-aborted assistant message usage from session entries.\n */\nexport function getLastAssistantUsage(entries: SessionEntry[]): Usage | undefined {\n\tfor (let i = entries.length - 1; i >= 0; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type === \"message\") {\n\t\t\tconst usage = getAssistantUsage(entry.message);\n\t\t\tif (usage) return usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\nexport interface ContextUsageEstimate {\n\ttokens: number;\n\tusageTokens: number;\n\ttrailingTokens: number;\n\tlastUsageIndex: number | null;\n}\n\nfunction getLastAssistantUsageInfo(messages: AgentMessage[]): { usage: Usage; index: number } | undefined {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst usage = getAssistantUsage(messages[i]);\n\t\tif (usage) return { usage, index: i };\n\t}\n\treturn undefined;\n}\n\n/**\n * Estimate context tokens from messages, using the last assistant usage when available.\n * If there are messages after the last usage, estimate their tokens with estimateTokens.\n */\nexport function estimateContextTokens(messages: AgentMessage[]): ContextUsageEstimate {\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\n\tif (!usageInfo) {\n\t\tlet estimated = 0;\n\t\tfor (const message of messages) {\n\t\t\testimated += estimateTokens(message);\n\t\t}\n\t\treturn {\n\t\t\ttokens: estimated,\n\t\t\tusageTokens: 0,\n\t\t\ttrailingTokens: estimated,\n\t\t\tlastUsageIndex: null,\n\t\t};\n\t}\n\n\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\tlet trailingTokens = 0;\n\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\ttrailingTokens += estimateTokens(messages[i]);\n\t}\n\n\treturn {\n\t\ttokens: usageTokens + trailingTokens,\n\t\tusageTokens,\n\t\ttrailingTokens,\n\t\tlastUsageIndex: usageInfo.index,\n\t};\n}\n\n/**\n * Check if compaction should trigger based on context usage.\n */\nexport function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean {\n\tif (!settings.enabled) return false;\n\treturn contextTokens > contextWindow - settings.reserveTokens;\n}\n\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nfunction estimateTextAndImageContentChars(content: string | Array<{ type: string; text?: string }>): number {\n\tif (typeof content === \"string\") {\n\t\treturn content.length;\n\t}\n\n\tlet chars = 0;\n\tfor (const block of content) {\n\t\tif (block.type === \"text\" && block.text) {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"image\") {\n\t\t\tchars += ESTIMATED_IMAGE_CHARS;\n\t\t}\n\t}\n\treturn chars;\n}\n\n/**\n * Estimate token count for a message using chars/4 heuristic.\n * This is conservative (overestimates tokens).\n */\nexport function estimateTokens(message: AgentMessage): number {\n\tlet chars = 0;\n\n\tswitch (message.role) {\n\t\tcase \"user\": {\n\t\t\tchars = estimateTextAndImageContentChars(\n\t\t\t\t(message as { content: string | Array<{ type: string; text?: string }> }).content,\n\t\t\t);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"assistant\": {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\tfor (const block of assistant.content) {\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tchars += block.text.length;\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\tchars += block.thinking.length;\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tchars += block.name.length + JSON.stringify(block.arguments).length;\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"custom\":\n\t\tcase \"toolResult\": {\n\t\t\tchars = estimateTextAndImageContentChars(message.content);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"bashExecution\": {\n\t\t\tchars = message.command.length + message.output.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"branchSummary\": {\n\t\t\tchars = message.summary.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t}\n\n\treturn 0;\n}\n"]}

package/dist/core/compaction/compaction.js

@@ -8,11 +8,27 @@ export const DEFAULT_COMPACTION_SETTINGS = {
     preserve_recent: 2,
 };
 /**
- * Calculate total context tokens from usage.
- * Uses the native totalTokens field when available, falls back to computing from components.
+ * Calculate active context-window tokens from provider usage.
+ *
+ * Prefer normalized component fields over `totalTokens`: some providers expose
+ * `totalTokens` as a billing/cumulative total, while the footer needs an active
+ * context estimate. Anthropic-compatible endpoints can also mirror cached input
+ * in both `input` and `cacheRead`/`cacheWrite`; when cache buckets are nearly the
+ * same size as `input`, treat `input` as the full prompt instead of counting the
+ * same cached prompt twice.
  */
 export function calculateContextTokens(usage) {
-    return usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;
+    const input = Math.max(0, usage.input || 0);
+    const output = Math.max(0, usage.output || 0);
+    const cacheRead = Math.max(0, usage.cacheRead || 0);
+    const cacheWrite = Math.max(0, usage.cacheWrite || 0);
+    const cacheTokens = cacheRead + cacheWrite;
+    const hasComponents = input > 0 || output > 0 || cacheTokens > 0;
+    if (!hasComponents)
+        return Math.max(0, usage.totalTokens || 0);
+    const cacheMirrorsInput = input > 0 && cacheTokens > 0 && cacheTokens >= input * 0.9 && cacheTokens <= input * 1.1;
+    const promptTokens = cacheMirrorsInput ? input : input + cacheTokens;
+    return promptTokens + output;
 }
 /**
  * Get usage from an assistant message if available.

package/dist/core/compaction/compaction.js.map

@@ -1 +1 @@
-{"version":3,"file":"compaction.js","sourceRoot":"","sources":["../../../src/core/compaction/compaction.ts"],"names":[],"mappings":"AAAA;;GAEG;AAiBH,MAAM,CAAC,MAAM,2BAA2B,GAAuB;IAC9D,OAAO,EAAE,IAAI;IACb,aAAa,EAAE,KAAK;IACpB,iBAAiB,EAAE,GAAG;IACtB,eAAe,EAAE,CAAC;CAClB,CAAC;AAEF;;;GAGG;AACH,MAAM,UAAU,sBAAsB,CAAC,KAAY;IAClD,OAAO,KAAK,CAAC,WAAW,IAAI,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC,MAAM,GAAG,KAAK,CAAC,SAAS,GAAG,KAAK,CAAC,UAAU,CAAC;AAC7F,CAAC;AAED;;;GAGG;AACH,SAAS,iBAAiB,CAAC,GAAiB;IAC3C,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,IAAI,OAAO,IAAI,GAAG,EAAE,CAAC;QAChD,MAAM,YAAY,GAAG,GAAuB,CAAC;QAC7C,IAAI,YAAY,CAAC,UAAU,KAAK,SAAS,IAAI,YAAY,CAAC,UAAU,KAAK,OAAO,IAAI,YAAY,CAAC,KAAK,EAAE,CAAC;YACxG,OAAO,YAAY,CAAC,KAAK,CAAC;QAC3B,CAAC;IACF,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,qBAAqB,CAAC,OAAuB;IAC5D,KAAK,IAAI,CAAC,GAAG,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC9C,MAAM,KAAK,GAAG,OAAO,CAAC,CAAC,CAAC,CAAC;QACzB,IAAI,KAAK,CAAC,IAAI,KAAK,SAAS,EAAE,CAAC;YAC9B,MAAM,KAAK,GAAG,iBAAiB,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;YAC/C,IAAI,KAAK;gBAAE,OAAO,KAAK,CAAC;QACzB,CAAC;IACF,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AASD,SAAS,yBAAyB,CAAC,QAAwB;IAC1D,KAAK,IAAI,CAAC,GAAG,QAAQ,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC/C,MAAM,KAAK,GAAG,iBAAiB,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;QAC7C,IAAI,KAAK;YAAE,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC;IACvC,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,qBAAqB,CAAC,QAAwB;IAC7D,MAAM,SAAS,GAAG,yBAAyB,CAAC,QAAQ,CAAC,CAAC;IAEtD,IAAI,CAAC,SAAS,EAAE,CAAC;QAChB,IAAI,SAAS,GAAG,CAAC,CAAC;QAClB,KAAK,MAAM,OAAO,IAAI,QAAQ,EAAE,CAAC;YAChC,SAAS,IAAI,cAAc,CAAC,OAAO,CAAC,CAAC;QACtC,CAAC;QACD,OAAO;YACN,MAAM,EAAE,SAAS;YACjB,WAAW,EAAE,CAAC;YACd,cAAc,EAAE,SAAS;YACzB,cAAc,EAAE,IAAI;SACpB,CAAC;IACH,CAAC;IAED,MAAM,WAAW,GAAG,sBAAsB,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC;IAC5D,IAAI,cAAc,GAAG,CAAC,CAAC;IACvB,KAAK,IAAI,CAAC,GAAG,SAAS,CAAC,KAAK,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAC5D,cAAc,IAAI,cAAc,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;IAC/C,CAAC;IAED,OAAO;QACN,MAAM,EAAE,WAAW,GAAG,cAAc;QACpC,WAAW;QACX,cAAc;QACd,cAAc,EAAE,SAAS,CAAC,KAAK;KAC/B,CAAC;AACH,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,aAAa,CAAC,aAAqB,EAAE,aAAqB,EAAE,QAA4B;IACvG,IAAI,CAAC,QAAQ,CAAC,OAAO;QAAE,OAAO,KAAK,CAAC;IACpC,OAAO,aAAa,GAAG,aAAa,GAAG,QAAQ,CAAC,aAAa,CAAC;AAC/D,CAAC;AAED,MAAM,qBAAqB,GAAG,IAAI,CAAC;AAEnC,SAAS,gCAAgC,CAAC,OAAwD;IACjG,IAAI,OAAO,OAAO,KAAK,QAAQ,EAAE,CAAC;QACjC,OAAO,OAAO,CAAC,MAAM,CAAC;IACvB,CAAC;IAED,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC7B,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,IAAI,KAAK,CAAC,IAAI,EAAE,CAAC;YACzC,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC;QAC5B,CAAC;aAAM,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;YACnC,KAAK,IAAI,qBAAqB,CAAC;QAChC,CAAC;IACF,CAAC;IACD,OAAO,KAAK,CAAC;AACd,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,cAAc,CAAC,OAAqB;IACnD,IAAI,KAAK,GAAG,CAAC,CAAC;IAEd,QAAQ,OAAO,CAAC,IAAI,EAAE,CAAC;QACtB,KAAK,MAAM,EAAE,CAAC;YACb,KAAK,GAAG,gCAAgC,CACtC,OAAwE,CAAC,OAAO,CACjF,CAAC;YACF,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;QAC7B,CAAC;QACD,KAAK,WAAW,EAAE,CAAC;YAClB,MAAM,SAAS,GAAG,OAA2B,CAAC;YAC9C,KAAK,MAAM,KAAK,IAAI,SAAS,CAAC,OAAO,EAAE,CAAC;gBACvC,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;oBAC3B,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC;gBAC5B,CAAC;qBAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;oBACtC,KAAK,IAAI,KAAK,CAAC,QAAQ,CAAC,MAAM,CAAC;gBAChC,CAAC;qBAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;oBACtC,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,MAAM,CAAC;gBACrE,CAAC;YACF,CAAC;YACD,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;QAC7B,CAAC;QACD,KAAK,QAAQ,CAAC;QACd,KAAK,YAAY,EAAE,CAAC;YACnB,KAAK,GAAG,gCAAgC,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;YAC1D,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;QAC7B,CAAC;QACD,KAAK,eAAe,EAAE,CAAC;YACtB,KAAK,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,MAAM,CAAC;YACvD,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;QAC7B,CAAC;QACD,KAAK,eAAe,EAAE,CAAC;YACtB,KAAK,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC;YAC/B,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;QAC7B,CAAC;IACF,CAAC;IAED,OAAO,CAAC,CAAC;AACV,CAAC","sourcesContent":["/**\n * Neutral context-usage metrics for deciding when a session needs compaction.\n */\n\nimport type { AgentMessage } from \"@earendil-works/pi-agent-core\";\nimport type { AssistantMessage, Usage } from \"@earendil-works/pi-ai\";\nimport type { SessionEntry } from \"../session-manager.ts\";\n\nexport interface CompactionSettings {\n\tenabled: boolean;\n\treserveTokens: number;\n\t/** Fraction of compactable context to keep. 0.3 is aggressive, 0.7 is light. */\n\tcompression_ratio: number;\n\t/** Number of recent context-eligible messages to preserve in standard mode. */\n\tpreserve_recent: number;\n\t/** Focus query for relevance-based pruning; auto-detected when omitted in settings/options. */\n\tquery?: string;\n}\n\nexport const DEFAULT_COMPACTION_SETTINGS: CompactionSettings = {\n\tenabled: true,\n\treserveTokens: 16384,\n\tcompression_ratio: 0.5,\n\tpreserve_recent: 2,\n};\n\n/**\n * Calculate total context tokens from usage.\n * Uses the native totalTokens field when available, falls back to computing from components.\n */\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\n}\n\n/**\n * Get usage from an assistant message if available.\n * Skips aborted and error messages as they don't have valid usage data.\n */\nfunction getAssistantUsage(msg: AgentMessage): Usage | undefined {\n\tif (msg.role === \"assistant\" && \"usage\" in msg) {\n\t\tconst assistantMsg = msg as AssistantMessage;\n\t\tif (assistantMsg.stopReason !== \"aborted\" && assistantMsg.stopReason !== \"error\" && assistantMsg.usage) {\n\t\t\treturn assistantMsg.usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n/**\n * Find the last non-aborted assistant message usage from session entries.\n */\nexport function getLastAssistantUsage(entries: SessionEntry[]): Usage | undefined {\n\tfor (let i = entries.length - 1; i >= 0; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type === \"message\") {\n\t\t\tconst usage = getAssistantUsage(entry.message);\n\t\t\tif (usage) return usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\nexport interface ContextUsageEstimate {\n\ttokens: number;\n\tusageTokens: number;\n\ttrailingTokens: number;\n\tlastUsageIndex: number | null;\n}\n\nfunction getLastAssistantUsageInfo(messages: AgentMessage[]): { usage: Usage; index: number } | undefined {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst usage = getAssistantUsage(messages[i]);\n\t\tif (usage) return { usage, index: i };\n\t}\n\treturn undefined;\n}\n\n/**\n * Estimate context tokens from messages, using the last assistant usage when available.\n * If there are messages after the last usage, estimate their tokens with estimateTokens.\n */\nexport function estimateContextTokens(messages: AgentMessage[]): ContextUsageEstimate {\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\n\tif (!usageInfo) {\n\t\tlet estimated = 0;\n\t\tfor (const message of messages) {\n\t\t\testimated += estimateTokens(message);\n\t\t}\n\t\treturn {\n\t\t\ttokens: estimated,\n\t\t\tusageTokens: 0,\n\t\t\ttrailingTokens: estimated,\n\t\t\tlastUsageIndex: null,\n\t\t};\n\t}\n\n\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\tlet trailingTokens = 0;\n\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\ttrailingTokens += estimateTokens(messages[i]);\n\t}\n\n\treturn {\n\t\ttokens: usageTokens + trailingTokens,\n\t\tusageTokens,\n\t\ttrailingTokens,\n\t\tlastUsageIndex: usageInfo.index,\n\t};\n}\n\n/**\n * Check if compaction should trigger based on context usage.\n */\nexport function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean {\n\tif (!settings.enabled) return false;\n\treturn contextTokens > contextWindow - settings.reserveTokens;\n}\n\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nfunction estimateTextAndImageContentChars(content: string | Array<{ type: string; text?: string }>): number {\n\tif (typeof content === \"string\") {\n\t\treturn content.length;\n\t}\n\n\tlet chars = 0;\n\tfor (const block of content) {\n\t\tif (block.type === \"text\" && block.text) {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"image\") {\n\t\t\tchars += ESTIMATED_IMAGE_CHARS;\n\t\t}\n\t}\n\treturn chars;\n}\n\n/**\n * Estimate token count for a message using chars/4 heuristic.\n * This is conservative (overestimates tokens).\n */\nexport function estimateTokens(message: AgentMessage): number {\n\tlet chars = 0;\n\n\tswitch (message.role) {\n\t\tcase \"user\": {\n\t\t\tchars = estimateTextAndImageContentChars(\n\t\t\t\t(message as { content: string | Array<{ type: string; text?: string }> }).content,\n\t\t\t);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"assistant\": {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\tfor (const block of assistant.content) {\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tchars += block.text.length;\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\tchars += block.thinking.length;\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tchars += block.name.length + JSON.stringify(block.arguments).length;\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"custom\":\n\t\tcase \"toolResult\": {\n\t\t\tchars = estimateTextAndImageContentChars(message.content);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"bashExecution\": {\n\t\t\tchars = message.command.length + message.output.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"branchSummary\": {\n\t\t\tchars = message.summary.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t}\n\n\treturn 0;\n}\n"]}
+{"version":3,"file":"compaction.js","sourceRoot":"","sources":["../../../src/core/compaction/compaction.ts"],"names":[],"mappings":"AAAA;;GAEG;AAiBH,MAAM,CAAC,MAAM,2BAA2B,GAAuB;IAC9D,OAAO,EAAE,IAAI;IACb,aAAa,EAAE,KAAK;IACpB,iBAAiB,EAAE,GAAG;IACtB,eAAe,EAAE,CAAC;CAClB,CAAC;AAEF;;;;;;;;;GASG;AACH,MAAM,UAAU,sBAAsB,CAAC,KAAY;IAClD,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,KAAK,IAAI,CAAC,CAAC,CAAC;IAC5C,MAAM,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,MAAM,IAAI,CAAC,CAAC,CAAC;IAC9C,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,SAAS,IAAI,CAAC,CAAC,CAAC;IACpD,MAAM,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,UAAU,IAAI,CAAC,CAAC,CAAC;IACtD,MAAM,WAAW,GAAG,SAAS,GAAG,UAAU,CAAC;IAC3C,MAAM,aAAa,GAAG,KAAK,GAAG,CAAC,IAAI,MAAM,GAAG,CAAC,IAAI,WAAW,GAAG,CAAC,CAAC;IACjE,IAAI,CAAC,aAAa;QAAE,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,WAAW,IAAI,CAAC,CAAC,CAAC;IAE/D,MAAM,iBAAiB,GAAG,KAAK,GAAG,CAAC,IAAI,WAAW,GAAG,CAAC,IAAI,WAAW,IAAI,KAAK,GAAG,GAAG,IAAI,WAAW,IAAI,KAAK,GAAG,GAAG,CAAC;IACnH,MAAM,YAAY,GAAG,iBAAiB,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,KAAK,GAAG,WAAW,CAAC;IACrE,OAAO,YAAY,GAAG,MAAM,CAAC;AAC9B,CAAC;AAED;;;GAGG;AACH,SAAS,iBAAiB,CAAC,GAAiB;IAC3C,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,IAAI,OAAO,IAAI,GAAG,EAAE,CAAC;QAChD,MAAM,YAAY,GAAG,GAAuB,CAAC;QAC7C,IAAI,YAAY,CAAC,UAAU,KAAK,SAAS,IAAI,YAAY,CAAC,UAAU,KAAK,OAAO,IAAI,YAAY,CAAC,KAAK,EAAE,CAAC;YACxG,OAAO,YAAY,CAAC,KAAK,CAAC;QAC3B,CAAC;IACF,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,qBAAqB,CAAC,OAAuB;IAC5D,KAAK,IAAI,CAAC,GAAG,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC9C,MAAM,KAAK,GAAG,OAAO,CAAC,CAAC,CAAC,CAAC;QACzB,IAAI,KAAK,CAAC,IAAI,KAAK,SAAS,EAAE,CAAC;YAC9B,MAAM,KAAK,GAAG,iBAAiB,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;YAC/C,IAAI,KAAK;gBAAE,OAAO,KAAK,CAAC;QACzB,CAAC;IACF,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AASD,SAAS,yBAAyB,CAAC,QAAwB;IAC1D,KAAK,IAAI,CAAC,GAAG,QAAQ,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC/C,MAAM,KAAK,GAAG,iBAAiB,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;QAC7C,IAAI,KAAK;YAAE,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC;IACvC,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,qBAAqB,CAAC,QAAwB;IAC7D,MAAM,SAAS,GAAG,yBAAyB,CAAC,QAAQ,CAAC,CAAC;IAEtD,IAAI,CAAC,SAAS,EAAE,CAAC;QAChB,IAAI,SAAS,GAAG,CAAC,CAAC;QAClB,KAAK,MAAM,OAAO,IAAI,QAAQ,EAAE,CAAC;YAChC,SAAS,IAAI,cAAc,CAAC,OAAO,CAAC,CAAC;QACtC,CAAC;QACD,OAAO;YACN,MAAM,EAAE,SAAS;YACjB,WAAW,EAAE,CAAC;YACd,cAAc,EAAE,SAAS;YACzB,cAAc,EAAE,IAAI;SACpB,CAAC;IACH,CAAC;IAED,MAAM,WAAW,GAAG,sBAAsB,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC;IAC5D,IAAI,cAAc,GAAG,CAAC,CAAC;IACvB,KAAK,IAAI,CAAC,GAAG,SAAS,CAAC,KAAK,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAC5D,cAAc,IAAI,cAAc,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;IAC/C,CAAC;IAED,OAAO;QACN,MAAM,EAAE,WAAW,GAAG,cAAc;QACpC,WAAW;QACX,cAAc;QACd,cAAc,EAAE,SAAS,CAAC,KAAK;KAC/B,CAAC;AACH,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,aAAa,CAAC,aAAqB,EAAE,aAAqB,EAAE,QAA4B;IACvG,IAAI,CAAC,QAAQ,CAAC,OAAO;QAAE,OAAO,KAAK,CAAC;IACpC,OAAO,aAAa,GAAG,aAAa,GAAG,QAAQ,CAAC,aAAa,CAAC;AAC/D,CAAC;AAED,MAAM,qBAAqB,GAAG,IAAI,CAAC;AAEnC,SAAS,gCAAgC,CAAC,OAAwD;IACjG,IAAI,OAAO,OAAO,KAAK,QAAQ,EAAE,CAAC;QACjC,OAAO,OAAO,CAAC,MAAM,CAAC;IACvB,CAAC;IAED,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC7B,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,IAAI,KAAK,CAAC,IAAI,EAAE,CAAC;YACzC,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC;QAC5B,CAAC;aAAM,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;YACnC,KAAK,IAAI,qBAAqB,CAAC;QAChC,CAAC;IACF,CAAC;IACD,OAAO,KAAK,CAAC;AACd,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,cAAc,CAAC,OAAqB;IACnD,IAAI,KAAK,GAAG,CAAC,CAAC;IAEd,QAAQ,OAAO,CAAC,IAAI,EAAE,CAAC;QACtB,KAAK,MAAM,EAAE,CAAC;YACb,KAAK,GAAG,gCAAgC,CACtC,OAAwE,CAAC,OAAO,CACjF,CAAC;YACF,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;QAC7B,CAAC;QACD,KAAK,WAAW,EAAE,CAAC;YAClB,MAAM,SAAS,GAAG,OAA2B,CAAC;YAC9C,KAAK,MAAM,KAAK,IAAI,SAAS,CAAC,OAAO,EAAE,CAAC;gBACvC,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;oBAC3B,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC;gBAC5B,CAAC;qBAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;oBACtC,KAAK,IAAI,KAAK,CAAC,QAAQ,CAAC,MAAM,CAAC;gBAChC,CAAC;qBAAM,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;oBACtC,KAAK,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,MAAM,CAAC;gBACrE,CAAC;YACF,CAAC;YACD,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;QAC7B,CAAC;QACD,KAAK,QAAQ,CAAC;QACd,KAAK,YAAY,EAAE,CAAC;YACnB,KAAK,GAAG,gCAAgC,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;YAC1D,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;QAC7B,CAAC;QACD,KAAK,eAAe,EAAE,CAAC;YACtB,KAAK,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,MAAM,CAAC;YACvD,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;QAC7B,CAAC;QACD,KAAK,eAAe,EAAE,CAAC;YACtB,KAAK,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC;YAC/B,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;QAC7B,CAAC;IACF,CAAC;IAED,OAAO,CAAC,CAAC;AACV,CAAC","sourcesContent":["/**\n * Neutral context-usage metrics for deciding when a session needs compaction.\n */\n\nimport type { AgentMessage } from \"@earendil-works/pi-agent-core\";\nimport type { AssistantMessage, Usage } from \"@earendil-works/pi-ai\";\nimport type { SessionEntry } from \"../session-manager.ts\";\n\nexport interface CompactionSettings {\n\tenabled: boolean;\n\treserveTokens: number;\n\t/** Fraction of compactable context to keep. 0.3 is aggressive, 0.7 is light. */\n\tcompression_ratio: number;\n\t/** Number of recent context-eligible messages to preserve in standard mode. */\n\tpreserve_recent: number;\n\t/** Focus query for relevance-based pruning; auto-detected when omitted in settings/options. */\n\tquery?: string;\n}\n\nexport const DEFAULT_COMPACTION_SETTINGS: CompactionSettings = {\n\tenabled: true,\n\treserveTokens: 16384,\n\tcompression_ratio: 0.5,\n\tpreserve_recent: 2,\n};\n\n/**\n * Calculate active context-window tokens from provider usage.\n *\n * Prefer normalized component fields over `totalTokens`: some providers expose\n * `totalTokens` as a billing/cumulative total, while the footer needs an active\n * context estimate. Anthropic-compatible endpoints can also mirror cached input\n * in both `input` and `cacheRead`/`cacheWrite`; when cache buckets are nearly the\n * same size as `input`, treat `input` as the full prompt instead of counting the\n * same cached prompt twice.\n */\nexport function calculateContextTokens(usage: Usage): number {\n\tconst input = Math.max(0, usage.input || 0);\n\tconst output = Math.max(0, usage.output || 0);\n\tconst cacheRead = Math.max(0, usage.cacheRead || 0);\n\tconst cacheWrite = Math.max(0, usage.cacheWrite || 0);\n\tconst cacheTokens = cacheRead + cacheWrite;\n\tconst hasComponents = input > 0 || output > 0 || cacheTokens > 0;\n\tif (!hasComponents) return Math.max(0, usage.totalTokens || 0);\n\n\tconst cacheMirrorsInput = input > 0 && cacheTokens > 0 && cacheTokens >= input * 0.9 && cacheTokens <= input * 1.1;\n\tconst promptTokens = cacheMirrorsInput ? input : input + cacheTokens;\n\treturn promptTokens + output;\n}\n\n/**\n * Get usage from an assistant message if available.\n * Skips aborted and error messages as they don't have valid usage data.\n */\nfunction getAssistantUsage(msg: AgentMessage): Usage | undefined {\n\tif (msg.role === \"assistant\" && \"usage\" in msg) {\n\t\tconst assistantMsg = msg as AssistantMessage;\n\t\tif (assistantMsg.stopReason !== \"aborted\" && assistantMsg.stopReason !== \"error\" && assistantMsg.usage) {\n\t\t\treturn assistantMsg.usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n/**\n * Find the last non-aborted assistant message usage from session entries.\n */\nexport function getLastAssistantUsage(entries: SessionEntry[]): Usage | undefined {\n\tfor (let i = entries.length - 1; i >= 0; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type === \"message\") {\n\t\t\tconst usage = getAssistantUsage(entry.message);\n\t\t\tif (usage) return usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\nexport interface ContextUsageEstimate {\n\ttokens: number;\n\tusageTokens: number;\n\ttrailingTokens: number;\n\tlastUsageIndex: number | null;\n}\n\nfunction getLastAssistantUsageInfo(messages: AgentMessage[]): { usage: Usage; index: number } | undefined {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst usage = getAssistantUsage(messages[i]);\n\t\tif (usage) return { usage, index: i };\n\t}\n\treturn undefined;\n}\n\n/**\n * Estimate context tokens from messages, using the last assistant usage when available.\n * If there are messages after the last usage, estimate their tokens with estimateTokens.\n */\nexport function estimateContextTokens(messages: AgentMessage[]): ContextUsageEstimate {\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\n\tif (!usageInfo) {\n\t\tlet estimated = 0;\n\t\tfor (const message of messages) {\n\t\t\testimated += estimateTokens(message);\n\t\t}\n\t\treturn {\n\t\t\ttokens: estimated,\n\t\t\tusageTokens: 0,\n\t\t\ttrailingTokens: estimated,\n\t\t\tlastUsageIndex: null,\n\t\t};\n\t}\n\n\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\tlet trailingTokens = 0;\n\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\ttrailingTokens += estimateTokens(messages[i]);\n\t}\n\n\treturn {\n\t\ttokens: usageTokens + trailingTokens,\n\t\tusageTokens,\n\t\ttrailingTokens,\n\t\tlastUsageIndex: usageInfo.index,\n\t};\n}\n\n/**\n * Check if compaction should trigger based on context usage.\n */\nexport function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean {\n\tif (!settings.enabled) return false;\n\treturn contextTokens > contextWindow - settings.reserveTokens;\n}\n\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nfunction estimateTextAndImageContentChars(content: string | Array<{ type: string; text?: string }>): number {\n\tif (typeof content === \"string\") {\n\t\treturn content.length;\n\t}\n\n\tlet chars = 0;\n\tfor (const block of content) {\n\t\tif (block.type === \"text\" && block.text) {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"image\") {\n\t\t\tchars += ESTIMATED_IMAGE_CHARS;\n\t\t}\n\t}\n\treturn chars;\n}\n\n/**\n * Estimate token count for a message using chars/4 heuristic.\n * This is conservative (overestimates tokens).\n */\nexport function estimateTokens(message: AgentMessage): number {\n\tlet chars = 0;\n\n\tswitch (message.role) {\n\t\tcase \"user\": {\n\t\t\tchars = estimateTextAndImageContentChars(\n\t\t\t\t(message as { content: string | Array<{ type: string; text?: string }> }).content,\n\t\t\t);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"assistant\": {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\tfor (const block of assistant.content) {\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tchars += block.text.length;\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\tchars += block.thinking.length;\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tchars += block.name.length + JSON.stringify(block.arguments).length;\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"custom\":\n\t\tcase \"toolResult\": {\n\t\t\tchars = estimateTextAndImageContentChars(message.content);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"bashExecution\": {\n\t\t\tchars = message.command.length + message.output.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"branchSummary\": {\n\t\t\tchars = message.summary.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t}\n\n\treturn 0;\n}\n"]}