@oh-my-pi/pi-agent-core 18.15.0 → 18.16.0

package/CHANGELOG.md

@@ -2,6 +2,21 @@
 
 ## [Unreleased]
 
+## [18.16.0] - 2026-11-02
+
+### Changed
+
+- Deprecated `createFileOps`, `computeFileLists`, `formatFileOperations`, `upsertFileOperations`, `isUrlSchemePath` and `FileOperations` in `@oh-my-pi/pi-agent-core/compaction`; they now re-export `@oh-my-pi/snapcompact`, which callers should import directly
+
+### Fixed
+
+- Context estimates for a model without a known tokenizer family (Gemini, Grok and the like), and token counts made without a model (compress metrics, the legacy `estimateTokens` export), count text in o200k_base instead of guessing 4 bytes per token, which undercounted what those providers bill by 11–48% on code-heavy sessions and let compaction fire late after a switch to them ([#17607](https://github.com/can1357/oh-my-pi/pull/17607) by [@will-bogusz](https://github.com/will-bogusz))
+- Context estimates no longer charge a reader another model's thinking signatures or redacted thinking, which the request strips: a Claude session switched to Grok or Gemini, including Gemini through Factory, counted them as text ([#17607](https://github.com/can1357/oh-my-pi/pull/17607) by [@will-bogusz](https://github.com/will-bogusz))
+
+### Removed
+
+- Removed `PI_TOKENIZER_ACCURATE`: every model is now counted with a native tokenizer ([#17607](https://github.com/can1357/oh-my-pi/pull/17607) by [@will-bogusz](https://github.com/will-bogusz))
+
 ## [18.15.0] - 2026-11-01
 
 ### Fixed

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

@@ -5,12 +5,12 @@
  * a summary of the branch being left so context isn't lost.
  */
 import type { Api, ApiKey, AssistantMessage, Context, Model, SimpleStreamOptions } from "@oh-my-pi/pi-ai";
+import * as snapcompact from "@oh-my-pi/snapcompact";
 import { type AgentTelemetry } from "../telemetry.js";
 import { Tokenizer } from "../tokenizer.js";
 import type { AgentMessage } from "../types.js";
 import type { ReadonlySessionManager, SessionEntry } from "./entries.js";
 import { type ConvertToLlm } from "./messages.js";
-import { type FileOperations } from "./utils.js";
 export interface BranchSummaryResult {
     summary?: string;
     readFiles?: string[];
@@ -23,12 +23,11 @@ export interface BranchSummaryDetails {
     readFiles: string[];
     modifiedFiles: string[];
 }
-export type { FileOperations } from "./utils.js";
 export interface BranchPreparation {
     /** Messages extracted for summarization, in chronological order */
     messages: AgentMessage[];
     /** File operations extracted from tool calls */
-    fileOps: FileOperations;
+    fileOps: snapcompact.FileOperations;
     /** Total estimated tokens in messages */
     totalTokens: number;
 }

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

@@ -5,13 +5,13 @@
  * and after compaction the session is reloaded.
  */
 import { type Api, type ApiKey, type AssistantMessage, type CodexCompactionContext, type Context, type FetchImpl, type MessageAttribution, type Model, type OneshotRetryOptions, type ProviderSessionState, type SimpleStreamOptions, type Tool, type Usage } from "@oh-my-pi/pi-ai";
+import * as snapcompact from "@oh-my-pi/snapcompact";
 import { type AgentTelemetry } from "../telemetry.js";
 import { ThinkingLevel } from "../thinking.js";
 import { Tokenizer } from "../tokenizer.js";
 import type { AgentMessage } from "../types.js";
 import type { SessionEntry } from "./entries.js";
 import { type ConvertToLlm } from "./messages.js";
-import { type FileOperations } from "./utils.js";
 /** Details stored in CompactionEntry.details for file tracking */
 export interface CompactionDetails {
     readFiles: string[];
@@ -368,7 +368,7 @@ export interface CompactionPreparation {
     /** Preserved opaque compaction payload from the previous compaction, if any. */
     previousPreserveData?: Record<string, unknown>;
     /** File operations extracted from messagesToSummarize */
-    fileOps: FileOperations;
+    fileOps: snapcompact.FileOperations;
     /** Compaction settions from settings.jsonl	*/
     settings: CompactionSettings;
 }

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

@@ -3,13 +3,12 @@
  */
 import type { Message } from "@oh-my-pi/pi-ai";
 import { type Dialect } from "@oh-my-pi/pi-ai/dialect";
+import * as snapcompact from "@oh-my-pi/snapcompact";
 import type { AgentMessage } from "../types.js";
-export interface FileOperations {
-    read: Set<string>;
-    written: Set<string>;
-    edited: Set<string>;
-}
-export declare function createFileOps(): FileOperations;
+/** @deprecated Import these from `@oh-my-pi/snapcompact`. */
+export { computeFileLists, createFileOps, type FileOperations, isUrlSchemePath, upsertFileOperations, } from "@oh-my-pi/snapcompact";
+/** @deprecated Use `upsertFileOperations("", readFiles, modifiedFiles, readSet)` from `@oh-my-pi/snapcompact`. */
+export declare function formatFileOperations(readFiles: string[], modifiedFiles: string[], readSet?: ReadonlySet<string>): string;
 /** Error text for a summarization call that hit its output limit (`stopReason: "length"`): partial text is never used. */
 export declare function incompleteOutputError(prefix: string): string;
 /**
@@ -33,25 +32,10 @@ export declare function stripReadSelector(path: string): string;
  * than its plain lines: verbatim `raw` bytes or the merge `conflicts` view.
  */
 export declare function isAlternateFormReadSelector(sel: string): boolean;
-/**
- * Whether `path` references a `scheme://` URL (internal URI or web URL) rather
- * than a filesystem path that belongs in the compaction `<files>` summary.
- */
-export declare function isUrlSchemePath(path: string): boolean;
 /**
  * Extract file operations from tool calls in an assistant message.
  */
-export declare function extractFileOpsFromMessage(message: AgentMessage, fileOps: FileOperations): void;
-/**
- * Compute final file lists from file operations.
- * Returns readFiles (files only read, not modified) and modifiedFiles.
- */
-export declare function computeFileLists(fileOps: FileOperations): {
-    readFiles: string[];
-    modifiedFiles: string[];
-};
-export declare function formatFileOperations(readFiles: string[], modifiedFiles: string[], readSet?: ReadonlySet<string>): string;
-export declare function upsertFileOperations(summary: string, readFiles: string[], modifiedFiles: string[], readSet?: ReadonlySet<string>): string;
+export declare function extractFileOpsFromMessage(message: AgentMessage, fileOps: snapcompact.FileOperations): void;
 /**
  * Truncate tool results to the same representation used in summarization prompts.
  */

package/dist/types/tokenizer.d.ts

@@ -10,6 +10,7 @@ type TokenizerModel = Pick<Model, "tokenizer"> & snapcompact.ShapeTarget & {
     compat?: Model["compat"];
     thinking?: Model["thinking"];
     reasoning?: boolean;
+    requestModelId?: string;
 };
 /** Identity of the reader whose replayed reasoning a {@link Tokenizer} counts; equal keys count it alike. */
 export declare function reasoningReplayKey(model: TokenizerModel | null | undefined): string;
@@ -24,13 +25,12 @@ export interface TokenizerOptions {
     nativeHistoryReplay?: boolean;
 }
 /**
- * `strict` always tries an exact native count (the catalog-resolved tokenizer
- * when known, o200k_base otherwise), falling back to the byte upper bound when
- * the loaded addon does not recognize that encoding. `approximate` and
- * `upperbound` prefer the same exact count for known tokenizer families or
- * when `PI_TOKENIZER_ACCURATE=1` is set; otherwise they use a cheap heuristic:
- * `approximate` a bytes/4 guess, `upperbound` the raw byte length (never
- * undercounts).
+ * Every mode counts with the native tokenizer: the catalog-resolved family
+ * when known, o200k_base otherwise. They differ only in the fallback when the
+ * loaded addon does not recognize that encoding: `strict` and `upperbound`
+ * the raw byte length (never undercounts), `approximate` a bytes/4 guess.
+ * Tests (`NODE_ENV=test`) skip the native count outside `strict` and use that
+ * fallback.
  */
 export type TokenCountMode = "strict" | "approximate" | "upperbound";
 /** Options for {@link Tokenizer.countMessage} / {@link Tokenizer.countMessages}. */
@@ -69,8 +69,8 @@ export interface TokenBudgetCheck {
  * when the model's encoding, image or thinking-replay pricing, replay
  * identity or native-history replay state changes); one-shot flows construct
  * their own for the model that will be billed. Known tokenizer families use
- * exact native counts; unknown models keep the fast byte estimate (or o200k
- * when `PI_TOKENIZER_ACCURATE=1`).
+ * exact native counts; other models are counted in o200k_base, which bills
+ * unknown readers closer than a bytes/4 guess.
  */
 export declare class Tokenizer {
     #private;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-agent-core",
-	"version": "18.15.0",
+	"version": "18.16.0",
 	"description": "General-purpose agent with transport abstraction, state management, and attachment support",
 	"homepage": "https://omp.sh",
 	"author": {
@@ -38,16 +38,16 @@
 		"fmt": "oxfmt --no-error-on-unmatched-pattern 'src/**/*.{ts,tsx}' '{test,bench,examples,scripts}/**/*.ts' '*.ts'"
 	},
 	"dependencies": {
-		"@oh-my-pi/pi-ai": "18.15.0",
-		"@oh-my-pi/pi-catalog": "18.15.0",
-		"@oh-my-pi/pi-natives": "18.15.0",
-		"@oh-my-pi/pi-utils": "18.15.0",
-		"@oh-my-pi/pi-wire": "18.15.0",
-		"@oh-my-pi/snapcompact": "18.15.0",
+		"@oh-my-pi/pi-ai": "18.16.0",
+		"@oh-my-pi/pi-catalog": "18.16.0",
+		"@oh-my-pi/pi-natives": "18.16.0",
+		"@oh-my-pi/pi-utils": "18.16.0",
+		"@oh-my-pi/pi-wire": "18.16.0",
+		"@oh-my-pi/snapcompact": "18.16.0",
 		"@opentelemetry/api": "^1.9.1"
 	},
 	"devDependencies": {
-		"@oh-my-pi/omptype": "18.15.0",
+		"@oh-my-pi/omptype": "18.16.0",
 		"@opentelemetry/context-async-hooks": "^2.9.0",
 		"@opentelemetry/sdk-trace-base": "^2.9.0",
 		"@types/bun": "^1.3.14"

package/src/agent-loop.ts

@@ -53,7 +53,7 @@ import {
 	signalListLabel,
 } from "@oh-my-pi/pi-ai/utils/harmony-leak";
 import { isDsmlLeakRecoveryTarget, removeDsmlToolMarkupLeak } from "@oh-my-pi/pi-ai/utils/dsml-leak";
-import { cloneJsonTree, logger, sanitizeText, structuredCloneJSON } from "@oh-my-pi/pi-utils";
+import { cloneJsonTree, logger, sanitizeText, structuredCloneJSON, toError } from "@oh-my-pi/pi-utils";
 import { INTENT_FIELD } from "@oh-my-pi/pi-wire";
 import { LiveSteeringChannel } from "./live-steering";
 import { agentPauseGate } from "./pause";
@@ -1188,7 +1188,7 @@ function resolveAsides(entries: AsideMessage[] | undefined): AgentMessage[] {
 			if (message) out.push(message);
 		}
 	} catch (error) {
-		discardAsides(out, error instanceof Error ? error : new Error(String(error)));
+		discardAsides(out, toError(error));
 		throw error;
 	}
 	return out;

package/src/compaction/branch-summarization.ts

@@ -24,16 +24,12 @@ import {
 import branchSummaryPrompt from "./prompts/branch-summary.md" with { type: "text" };
 import branchSummaryPreamble from "./prompts/branch-summary-preamble.md" with { type: "text" };
 import {
-	computeFileLists,
-	createFileOps,
 	extractFileOpsFromMessage,
-	type FileOperations,
 	incompleteOutputError,
 	SUMMARIZATION_SYSTEM_PROMPT,
 	serializeConversationForSummary,
 	stripReadSelector,
 	truncateToolResultForSummary,
-	upsertFileOperations,
 } from "./utils";
 
 // ============================================================================
@@ -54,13 +50,11 @@ export interface BranchSummaryDetails {
 	modifiedFiles: string[];
 }
 
-export type { FileOperations } from "./utils";
-
 export interface BranchPreparation {
 	/** Messages extracted for summarization, in chronological order */
 	messages: AgentMessage[];
 	/** File operations extracted from tool calls */
-	fileOps: FileOperations;
+	fileOps: snapcompact.FileOperations;
 	/** Total estimated tokens in messages */
 	totalTokens: number;
 }
@@ -252,7 +246,7 @@ export function prepareBranchEntries(
 	tokenBudget: number = 0,
 ): BranchPreparation {
 	const messages: AgentMessage[] = [];
-	const fileOps = createFileOps();
+	const fileOps = snapcompact.createFileOps();
 	let totalTokens = 0;
 
 	// First pass: collect file ops from ALL entries (even if they don't fit in token budget)
@@ -400,8 +394,8 @@ export async function generateBranchSummary(
 	summary = BRANCH_SUMMARY_PREAMBLE + summary;
 
 	// Compute file lists and append to summary
-	const { readFiles, modifiedFiles } = computeFileLists(fileOps);
-	summary = upsertFileOperations(summary, readFiles, modifiedFiles, fileOps.read);
+	const { readFiles, modifiedFiles } = snapcompact.computeFileLists(fileOps);
+	summary = snapcompact.upsertFileOperations(summary, readFiles, modifiedFiles, fileOps.read);
 
 	return {
 		summary: summary || "No summary generated",

package/src/compaction/compaction-v2-streaming.ts

@@ -32,7 +32,7 @@ import {
 	OPENAI_HEADER_VALUES,
 	OPENAI_HEADERS,
 } from "@oh-my-pi/pi-catalog/wire/codex";
-import { $env, isUnexpectedSocketCloseMessage, logger, ptree, stringifyJson } from "@oh-my-pi/pi-utils";
+import { $env, isUnexpectedSocketCloseMessage, logger, ptree, stringifyJson, toError } from "@oh-my-pi/pi-utils";
 import { appendAzureApiVersion, resolveAzureOpenAiBaseUrl } from "./azure-openai-endpoint";
 import { prepareBedrockCompactionRequest } from "./bedrock";
 
@@ -320,7 +320,7 @@ export async function requestCompactionV2Streaming(
 				codexAccessPrograms: options?.codexAccessPrograms,
 			});
 		} catch (err) {
-			const error = err instanceof Error ? err : new Error(String(err));
+			const error = toError(err);
 			if (signal?.aborted) throw error;
 
 			if (isRetryableCompactionError(error) && attempt < V2_COMPACTION_MAX_RETRIES) {

package/src/compaction/compaction.ts

@@ -86,16 +86,12 @@ import handoffDocumentPrompt from "./prompts/handoff-document.md" with { type: "
 import snapcompactArchiveContextPrompt from "./prompts/snapcompact-archive-context.md" with { type: "text" };
 
 import {
-	computeFileLists,
-	createFileOps,
 	escapeSummaryBoundaryTags,
 	extractFileOpsFromMessage,
-	type FileOperations,
 	incompleteOutputError,
 	SUMMARIZATION_SYSTEM_PROMPT,
 	serializeConversationForSummary,
 	stripReadSelector,
-	upsertFileOperations,
 } from "./utils";
 
 // ============================================================================
@@ -115,8 +111,8 @@ function extractFileOperations(
 	messages: AgentMessage[],
 	entries: SessionEntry[],
 	prevCompactionIndex: number,
-): FileOperations {
-	const fileOps = createFileOps();
+): snapcompact.FileOperations {
+	const fileOps = snapcompact.createFileOps();
 
 	// Collect from previous compaction's details (if pi-generated)
 	if (prevCompactionIndex >= 0) {
@@ -1384,7 +1380,7 @@ export interface CompactionPreparation {
 	/** Preserved opaque compaction payload from the previous compaction, if any. */
 	previousPreserveData?: Record<string, unknown>;
 	/** File operations extracted from messagesToSummarize */
-	fileOps: FileOperations;
+	fileOps: snapcompact.FileOperations;
 	/** Compaction settions from settings.jsonl	*/
 	settings: CompactionSettings;
 }
@@ -2343,8 +2339,8 @@ export async function compact(
 			});
 
 	// Compute file lists and append to summary
-	const { readFiles, modifiedFiles } = computeFileLists(fileOps);
-	summary = upsertFileOperations(summary, readFiles, modifiedFiles, fileOps.read);
+	const { readFiles, modifiedFiles } = snapcompact.computeFileLists(fileOps);
+	summary = snapcompact.upsertFileOperations(summary, readFiles, modifiedFiles, fileOps.read);
 	if (nativeSummary !== undefined) {
 		// The replayed block (summary and any `tool_changes`) stays
 		// byte-identical to the API's output so it matches its signature. The
@@ -2352,7 +2348,7 @@ export async function compact(
 		// summary message with the block and skips its text, so they would
 		// otherwise be invisible to this provider. Every other provider keeps
 		// reading the entry text.
-		const filesText = upsertFileOperations("", readFiles, modifiedFiles, fileOps.read) || undefined;
+		const filesText = snapcompact.upsertFileOperations("", readFiles, modifiedFiles, fileOps.read) || undefined;
 		preserveData = withAnthropicCompactionPreserveData(preserveData, {
 			provider: model.provider,
 			content: nativeSummary,

package/src/compaction/utils.ts

@@ -5,27 +5,27 @@
 import type { Message, ToolCall } from "@oh-my-pi/pi-ai";
 import { type Dialect, getDialectDefinition } from "@oh-my-pi/pi-ai/dialect";
 import { escapeHarmonyControlTokens } from "@oh-my-pi/pi-ai/utils/harmony-leak";
-import { formatGroupedPaths, prompt, stringifyJson } from "@oh-my-pi/pi-utils";
+import { prompt, stringifyJson } from "@oh-my-pi/pi-utils";
+import * as snapcompact from "@oh-my-pi/snapcompact";
 import type { AgentMessage } from "../types";
-import fileOperationsTemplate from "./prompts/file-operations.md" with { type: "text" };
 import summarizationSystemPrompt from "./prompts/summarization-system.md" with { type: "text" };
 
-// ============================================================================
-// File Operation Tracking
-// ============================================================================
+/** @deprecated Import these from `@oh-my-pi/snapcompact`. */
+export {
+	computeFileLists,
+	createFileOps,
+	type FileOperations,
+	isUrlSchemePath,
+	upsertFileOperations,
+} from "@oh-my-pi/snapcompact";
 
-export interface FileOperations {
-	read: Set<string>;
-	written: Set<string>;
-	edited: Set<string>;
-}
-
-export function createFileOps(): FileOperations {
-	return {
-		read: new Set(),
-		written: new Set(),
-		edited: new Set(),
-	};
+/** @deprecated Use `upsertFileOperations("", readFiles, modifiedFiles, readSet)` from `@oh-my-pi/snapcompact`. */
+export function formatFileOperations(
+	readFiles: string[],
+	modifiedFiles: string[],
+	readSet?: ReadonlySet<string>,
+): string {
+	return snapcompact.upsertFileOperations("", readFiles, modifiedFiles, readSet);
 }
 
 /** Error text for a summarization call that hit its output limit (`stopReason: "length"`): partial text is never used. */
@@ -91,27 +91,10 @@ export function isAlternateFormReadSelector(sel: string): boolean {
 	return sel.split(":").some(part => READ_ALTERNATE_FORM_RE.test(part));
 }
 
-/**
- * A real filesystem path never contains a `scheme://` URL. Tool-call paths that
- * do — `conflict://1`, `artifact://3`, `local://ctx.md`, `history://…`,
- * `issue://12`, `https://…`, and the tolerated `file.ts:conflict://1` prefix
- * form — are session-scoped or remote resources, not files the post-compaction
- * agent can re-ground on. Keep them out of the `<files>` summary.
- */
-const URL_SCHEME_RE = /[a-z][a-z0-9+.-]*:\/\//i;
-
-/**
- * Whether `path` references a `scheme://` URL (internal URI or web URL) rather
- * than a filesystem path that belongs in the compaction `<files>` summary.
- */
-export function isUrlSchemePath(path: string): boolean {
-	return URL_SCHEME_RE.test(path);
-}
-
 /**
  * Extract file operations from tool calls in an assistant message.
  */
-export function extractFileOpsFromMessage(message: AgentMessage, fileOps: FileOperations): void {
+export function extractFileOpsFromMessage(message: AgentMessage, fileOps: snapcompact.FileOperations): void {
 	if (message.role !== "assistant") return;
 	if (!("content" in message) || !Array.isArray(message.content)) return;
 
@@ -128,7 +111,7 @@ export function extractFileOpsFromMessage(message: AgentMessage, fileOps: FileOp
 
 		// Internal URIs (conflict://, artifact://, local://, history://, …) and
 		// web URLs are not re-groundable files — keep them out of `<files>`.
-		if (isUrlSchemePath(path)) continue;
+		if (snapcompact.isUrlSchemePath(path)) continue;
 
 		switch (block.name) {
 			case "read":
@@ -144,68 +127,6 @@ export function extractFileOpsFromMessage(message: AgentMessage, fileOps: FileOp
 	}
 }
 
-/**
- * Compute final file lists from file operations.
- * Returns readFiles (files only read, not modified) and modifiedFiles.
- */
-export function computeFileLists(fileOps: FileOperations): { readFiles: string[]; modifiedFiles: string[] } {
-	// Drop any `scheme://` URLs (e.g. legacy `conflict://`/`artifact://` entries
-	// rehydrated straight into `fileOps` from a pre-fix compaction summary) — only
-	// real files belong in `<files>`. New tool-call scans are already filtered.
-	const modified = new Set([...fileOps.edited, ...fileOps.written].filter(f => !isUrlSchemePath(f)));
-	const readOnly = [...fileOps.read].filter(f => !isUrlSchemePath(f) && !modified.has(f)).sort();
-	const modifiedFiles = [...modified].sort();
-	return { readFiles: readOnly, modifiedFiles };
-}
-
-/**
- * Format file operations as one `<files>` tag: a grouped, prefix-folded
- * directory tree (find-tool shape — `# dir/` headers, bare basenames) with a
- * ` (Read)` / ` (Write)` / ` (RW)` marker per file instead of separate
- * read/modified lists. `readSet` is the cumulative read set (`fileOps.read`),
- * used to tell modified files that were also read (RW) from blind writes.
- */
-const FILE_OPERATION_SUMMARY_LIMIT = 20;
-
-function stripFileOperationTags(summary: string): string {
-	// Legacy <read-files>/<modified-files> tags are still stripped so summaries
-	// written before the combined <files> tag self-heal on the next compaction.
-	return summary
-		.replace(/<files>[\s\S]*?<\/files>\s*/g, "")
-		.replace(/<read-files>[\s\S]*?<\/read-files>\s*/g, "")
-		.replace(/<modified-files>[\s\S]*?<\/modified-files>\s*/g, "")
-		.trimEnd();
-}
-export function formatFileOperations(
-	readFiles: string[],
-	modifiedFiles: string[],
-	readSet?: ReadonlySet<string>,
-): string {
-	if (readFiles.length === 0 && modifiedFiles.length === 0) return "";
-	const mode = new Map<string, "Read" | "Write" | "RW">();
-	for (const file of readFiles) mode.set(file, "Read");
-	for (const file of modifiedFiles) mode.set(file, readSet?.has(file) ? "RW" : "Write");
-	const all = [...mode.keys()].sort();
-	let files = formatGroupedPaths(all.slice(0, FILE_OPERATION_SUMMARY_LIMIT), path => ` (${mode.get(path)})`);
-	if (all.length > FILE_OPERATION_SUMMARY_LIMIT) {
-		files += `\n[…${all.length - FILE_OPERATION_SUMMARY_LIMIT} files elided…]`;
-	}
-	return prompt.render(fileOperationsTemplate, { files });
-}
-
-export function upsertFileOperations(
-	summary: string,
-	readFiles: string[],
-	modifiedFiles: string[],
-	readSet?: ReadonlySet<string>,
-): string {
-	const baseSummary = stripFileOperationTags(summary);
-	const fileOperations = formatFileOperations(readFiles, modifiedFiles, readSet);
-	if (!fileOperations) return baseSummary;
-	if (!baseSummary) return fileOperations;
-	return `${baseSummary}\n\n${fileOperations}`;
-}
-
 // ============================================================================
 // Message Serialization
 // ============================================================================

package/src/tokenizer.ts

@@ -1,9 +1,11 @@
 import type { AssistantMessage, Model } from "@oh-my-pi/pi-ai";
 import { renderDemotedThinking } from "@oh-my-pi/pi-ai/dialect/demotion";
+import { isFactoryDroidGeminiWire, keepsFactoryGeminiReasoning } from "@oh-my-pi/pi-ai/providers/factory-droid/gemini";
 import { ollamaChatDropsOwnThinking } from "@oh-my-pi/pi-ai/providers/ollama";
 import { completionsDropsOwnThinking, isLiteLLMThinkingSignature } from "@oh-my-pi/pi-ai/providers/openai-completions";
 import { isColdReplayableResponsesTurn } from "@oh-my-pi/pi-ai/providers/openai-shared";
 import {
+	dropsForeignReasoningPayload,
 	isSameModelTurn,
 	shouldDropTruncatedThinkingOnlyAssistant,
 } from "@oh-my-pi/pi-ai/providers/transform-messages";
@@ -22,7 +24,6 @@ import { base64ImageSize, estimateImageTokens, type ImageDetail, type ImageSize
 import type { AgentMessage } from "./types";
 
 const testEnv = Bun.env.NODE_ENV === "test";
-const accurate = process.env.PI_TOKENIZER_ACCURATE === "1" && !testEnv;
 
 const NATIVE_ENCODING: Record<ModelTokenizer, natives.Encoding> = {
 	"claude-v3": natives.Encoding.ClaudeV3,
@@ -33,6 +34,7 @@ const NATIVE_ENCODING: Record<ModelTokenizer, natives.Encoding> = {
 	"deepseek-v3": natives.Encoding.DeepSeekV3,
 	"kimi-k2": natives.Encoding.KimiK2,
 	glm5: natives.Encoding.Glm5,
+	cl100k: natives.Encoding.Cl100kBase,
 	o200k: natives.Encoding.O200kBase,
 };
 
@@ -42,7 +44,12 @@ export function tokenizerEncodingForModel(model: Pick<Model, "tokenizer"> | null
 }
 
 type TokenizerModel = Pick<Model, "tokenizer"> &
-	snapcompact.ShapeTarget & { compat?: Model["compat"]; thinking?: Model["thinking"]; reasoning?: boolean };
+	snapcompact.ShapeTarget & {
+		compat?: Model["compat"];
+		thinking?: Model["thinking"];
+		reasoning?: boolean;
+		requestModelId?: string;
+	};
 
 /** The model an assistant turn's native reasoning (signed thinking, Responses history) is replayed to. */
 interface ReasoningReplayReader {
@@ -77,13 +84,12 @@ export interface TokenizerOptions {
 }
 
 /**
- * `strict` always tries an exact native count (the catalog-resolved tokenizer
- * when known, o200k_base otherwise), falling back to the byte upper bound when
- * the loaded addon does not recognize that encoding. `approximate` and
- * `upperbound` prefer the same exact count for known tokenizer families or
- * when `PI_TOKENIZER_ACCURATE=1` is set; otherwise they use a cheap heuristic:
- * `approximate` a bytes/4 guess, `upperbound` the raw byte length (never
- * undercounts).
+ * Every mode counts with the native tokenizer: the catalog-resolved family
+ * when known, o200k_base otherwise. They differ only in the fallback when the
+ * loaded addon does not recognize that encoding: `strict` and `upperbound`
+ * the raw byte length (never undercounts), `approximate` a bytes/4 guess.
+ * Tests (`NODE_ENV=test`) skip the native count outside `strict` and use that
+ * fallback.
  */
 export type TokenCountMode = "strict" | "approximate" | "upperbound";
 
@@ -205,8 +211,8 @@ interface MessageEstimate {
  * when the model's encoding, image or thinking-replay pricing, replay
  * identity or native-history replay state changes); one-shot flows construct
  * their own for the model that will be billed. Known tokenizer families use
- * exact native counts; unknown models keep the fast byte estimate (or o200k
- * when `PI_TOKENIZER_ACCURATE=1`).
+ * exact native counts; other models are counted in o200k_base, which bills
+ * unknown readers closer than a bytes/4 guess.
  */
 export class Tokenizer {
 	readonly #encoding: natives.Encoding | null;
@@ -217,6 +223,8 @@ export class Tokenizer {
 	readonly #dropsCompletionsThinking: boolean;
 	/** Claude's own API and prefix-bound hosts keep another model's signature from their deployment. */
 	readonly #readsDeploymentSignatures: boolean;
+	/** Factory's Gemini generate wire, whose replay pass keeps only Google-origin reasoning. */
+	readonly #factoryGeminiWire: boolean;
 	readonly reasoningReplayKey: string;
 	/** How the reader bills replayed signed thinking (catalog `thinking-replay-tokens`). */
 	readonly thinkingReplayTokens: ThinkingReplayTokens;
@@ -259,6 +267,10 @@ export class Tokenizer {
 			model?.api === "anthropic-messages" &&
 			((compat !== undefined && "officialEndpoint" in compat && compat.officialEndpoint) ||
 				model.thinking?.prefixBinding === true);
+		this.#factoryGeminiWire =
+			model?.api === "factory-droid-agent" &&
+			model.id !== undefined &&
+			isFactoryDroidGeminiWire({ id: model.id, requestModelId: model.requestModelId });
 		this.reasoningReplayKey = reasoningReplayKey(model);
 		this.thinkingReplayTokens = resolveThinkingReplayTokens(model ?? {});
 		this.nativeHistoryReplay = options?.nativeHistoryReplay ?? true;
@@ -269,11 +281,8 @@ export class Tokenizer {
 	}
 
 	countTokens(text: string | string[], mode: TokenCountMode = "approximate"): number {
-		if (mode === "strict") return countTokensNat(text, this.#encoding, mode, this.#nativeCounts).tokens;
-		if (!testEnv && this.#encoding !== null)
-			return countTokensNat(text, this.#encoding, mode, this.#nativeCounts).tokens;
-		if (accurate) return countTokensNat(text, undefined, mode, this.#nativeCounts).tokens;
-		return sumFragments(text, mode === "upperbound" ? byteLength : byteEstimate);
+		if (testEnv && mode !== "strict") return sumFragments(text, mode === "upperbound" ? byteLength : byteEstimate);
+		return countTokensNat(text, this.#encoding, mode, this.#nativeCounts).tokens;
 	}
 
 	/**
@@ -365,6 +374,21 @@ export class Tokenizer {
 		);
 	}
 
+	/**
+	 * Whether `message`'s opaque reasoning payload (thinking signatures, redacted
+	 * thinking) may reach this reader: everything except what `transformMessages`
+	 * certainly strips from another model's turn, or what Factory's Gemini wire
+	 * drops from a turn not of Google origin. Without a replay reader (no
+	 * model, or a `filterReasoningHistory` target) every payload is charged,
+	 * which overcounts a filtering reader's foreign turns.
+	 */
+	#receivesReasoningPayload(message: AssistantMessage): boolean {
+		const reader = this.#reasoningReader;
+		if (reader === undefined) return true;
+		if (this.#factoryGeminiWire) return keepsFactoryGeminiReasoning(message);
+		return !dropsForeignReasoningPayload(message, reader);
+	}
+
 	/**
 	 * The replayed native Responses reasoning {@link countMessage} charges
 	 * across `messages`: the part of a count that changes with the reading
@@ -468,15 +492,17 @@ export class Tokenizer {
 						fragments.push(block.thinking);
 						// Providers charge for the opaque signature/reasoning payload that
 						// rides alongside the thinking text (OpenAI Responses encrypted
-						// reasoning items, Gemini thought signatures). A reader billing at
-						// `output` size (Claude) charges the reasoning, priced below, not
-						// these bytes. The compaction floor excludes it (its local byte size
-						// diverges from provider billing).
+						// reasoning items, Gemini thought signatures) wherever it survives
+						// replay; `transformMessages` strips another model's on most wires.
+						// A reader billing at `output` size (Claude) charges the reasoning,
+						// priced below, not these bytes. The compaction floor excludes it
+						// (its local byte size diverges from provider billing).
 						if (
 							block.thinkingSignature &&
 							!excludeEncryptedReasoning &&
 							signaturesPriced &&
-							typeof replayedReasoning !== "number"
+							typeof replayedReasoning !== "number" &&
+							this.#receivesReasoningPayload(message)
 						) {
 							fragments.push(block.thinkingSignature);
 						}
@@ -485,10 +511,13 @@ export class Tokenizer {
 						fragments.push(block.name);
 						fragments.push(stringifyJson(block.arguments) ?? "null");
 					} else if (block.type === "redactedThinking") {
-						// Encrypted reasoning blob the provider still bills for on replay;
-						// excluded from the compaction floor for the same reason as above.
+						// Encrypted reasoning blob the provider still bills for wherever it
+						// survives replay, as above; excluded from the compaction floor for
+						// the same reason.
 						signedReasoning = true;
-						if (!excludeEncryptedReasoning && signaturesPriced) fragments.push(block.data);
+						if (!excludeEncryptedReasoning && signaturesPriced && this.#receivesReasoningPayload(message)) {
+							fragments.push(block.data);
+						}
 					} else if (block.type === "anthropicServerTool") {
 						// Native Anthropic server-tool call/result replayed verbatim on the
 						// wire (server_tool_use input and opaque result content). The provider

package/src/compaction/prompts/file-operations.md

@@ -1,5 +0,0 @@
-{{#if files}}
-{{#xml "files"}}
-{{files}}
-{{/xml}}
-{{/if}}