fix: align compaction cuts and native usage accounting

src/core/compaction.ts

@@ -1,7 +1,8 @@
 /**
  * COMPACTION (owner round 14, 2026-10-05): the context-fitting mechanic,
- * mirroring pi's design (docs aside, this is OUR implementation — pi's
- * compaction lives in the coding-agent package we don't depend on).
+ * with a bounded local port of Pi coding-agent 1.4.0 cutoff selection.
+ * No coding-agent dependency. See docs/compaction-accounting.md for the
+ * behavior boundary, native golden fixtures, and upstream MIT notice.
  *
  * THE CONTRACT: full history stays persisted in the record — extraction
  * and the user's transcript keep reading everything. A compaction marker
@@ -38,7 +39,7 @@ import { COMPACTED_HISTORY_HEADER } from "./contracts/chat.ts";
 import type { LhMessage, LhMarkerMessage } from "./native-messages.ts";
 import { newMessageId } from "./native-messages.ts";
 import type { AgentMessage } from "@earendil-works/pi-agent-core";
-import type { AssistantMessage, ToolResultMessage } from "@earendil-works/pi-ai";
+import type { AssistantMessage, ToolResultMessage, SystemMessage } from "@earendil-works/pi-ai";
 
 // ─── constants (the tunables, owner-approved) ────────────────────────────────
 
@@ -77,9 +78,8 @@ export function calculateContextTokens(usage: {
 	);
 }
 
-/** Estimate a message's context tokens from its WIRE-relevant text only
- * (pi's estimator): text + thinking + toolCall name/args for assistants,
- * content text for user/toolResult/custom. NEVER JSON.stringify of the
+/** Pi 1.4.0's heuristic: text + thinking + toolCall name/args for
+ * assistants, content text plus a fixed image allowance for other roles. NEVER JSON.stringify of the
  * whole message — record scaffolding (ids, usage, responseId, app-side
  * details, escaped quotes/backslashes) inflates ~3x (measured on the
  * 2026-10-05 relational-databases incident: 84k estimated vs ~30k real). */
@@ -100,23 +100,32 @@ export function estimateMessageTokens(message: LhMessage | AgentMessage): number
 		}
 	} else {
 		chars += textCharsOfContent(content);
+		if (message.role === "system") {
+			const system = message as SystemMessage;
+			for (const section of Object.values(system.sections ?? {})) chars += section?.length ?? 0;
+			if (system.toolsAdded) chars += JSON.stringify(system.toolsAdded).length;
+		}
 	}
 	return Math.ceil(chars / 4);
 }
 
+// Pi coding-agent 1.4.0 estimateTokens: fixed image allowance, NOT a
+// measurement of encrypted reasoning or actual provider vision tokens.
+const ESTIMATED_IMAGE_CHARS = 4_800;
 function textCharsOfContent(content: unknown): number {
 	if (typeof content === "string") return content.length;
 	if (!Array.isArray(content)) return 0;
 	let chars = 0;
 	for (const block of content as Array<string | { type?: string; text?: string }>) {
 		if (typeof block === "string") chars += block.length;
-		else if (block?.type === "text" || "text" in block) chars += block.text?.length ?? 0;
+		else if (block?.type === "image") chars += ESTIMATED_IMAGE_CHARS;
+		else if (block?.type === "text") chars += block.text?.length ?? 0;
 	}
 	return chars;
 }
 
 /** The newest assistant message carrying VALID provider-reported usage
- * (pi's rule: skip aborted messages and all-zero usage — they carry no
+ * (pi's rule: skip aborted/error messages and all-zero usage — they carry no
  * truth about the wire). The scan STOPS at the newest compaction
  * marker: usage captured before a compaction described the pre-cut
  * context — the next request carries the summary + kept tail instead,
@@ -131,7 +140,7 @@ export function lastAssistantUsageIndex(
 		if ((m as { role?: string }).role === "custom" && (m as LhMarkerMessage).customType === "compaction") return null;
 		if ((m as { role?: string }).role !== "assistant") continue;
 		const a = m as AssistantMessage & { stopReason?: string };
-		if (a.stopReason === "aborted") continue;
+		if (a.stopReason === "aborted" || a.stopReason === "error") continue;
 		if (calculateContextTokens((a.usage ?? {}) as Parameters<typeof calculateContextTokens>[0]) === 0) continue;
 		return i;
 	}
@@ -180,51 +189,53 @@ export function estimateTokens(messages: readonly (LhMessage | AgentMessage)[]):
 
 // ─── the cut walk ────────────────────────────────────────────────────────────
 
-/**
- * Find the compaction cut index: walk BACKWARDS from the end accumulating
- * estimates until ~keepRecentTokens are kept; the cut lands there.
- *
- * HARD INVARIANT — a tool result is never separated from its call: if the
- * walk stops on a toolResult, the KEPT region extends to the assistant
- * message carrying its call (keeping the result means keeping the call).
- * This is the provider protocol's shape law: dropping a call while
- * keeping its result strands an orphan the provider rejects. A PENDING
- * question/problem call (no result yet — the pose→answer window) sits at
- * the transcript end and is always kept: the walk never crosses it.
- *
- * Returns 0 when everything must be summarized (the whole history is the
- * span — valid), or messages.length when there is nothing meaningful to
- * cut (kept budget covers everything — caller checks before compacting).
- */
-export function compactionCutIndex(messages: readonly LhMessage[], keepRecentTokens = KEEP_RECENT_TOKENS): number {
-	let kept = 0;
-	for (let i = messages.length - 1; i >= 0; i--) {
-		if (kept >= keepRecentTokens) {
-			// The cut candidate: extend past orphaned tool results to the
-			// assistant message carrying their calls.
-			let cut = i + 1;
-			while (cut < messages.length && (messages[cut] as { role?: string }).role === "toolResult") {
-				const callIndex = assistantIndexOfCall(messages, (messages[cut] as ToolResultMessage).toolCallId, cut);
-				if (callIndex === -1) break; // no call found (defensive): stop extending
-				cut = callIndex;
-			}
-			return cut;
-		}
-		kept += estimateMessageTokens(messages[i]!);
-	}
-	return 0; // the whole history fits inside the kept budget — summarize all
+/** Model-visible entry adapter for the local Pi 1.4.0 cutoff port.
+ * Compaction entries are boundaries, not ordinary cut candidates; UI-only
+ * markers have no context, just like native metadata entries. */
+function cutMessage(message: LhMessage): LhMessage | null {
+	if (message.role !== "custom") return message;
+	if (["compaction", "plan-block", "problem-dismissed"].includes(message.customType)) return null;
+	return message;
 }
 
-/** The newest assistant message at or before `before` whose content
- * carries the tool call `toolCallId` — the call a result answers. */
-function assistantIndexOfCall(messages: readonly LhMessage[], toolCallId: string, before: number): number {
-	for (let i = before - 1; i >= 0; i--) {
-		const m = messages[i]!;
-		if ((m as { role?: string }).role !== "assistant") continue;
-		const assistant = m as AssistantMessage;
-		if (assistant.content.some((b) => b.type === "toolCall" && b.id === toolCallId)) return i;
+/** Local port of pi-coding-agent 1.4.0 findCutPoint (MIT).
+ * Walk backwards to >= budget, then choose the nearest valid cut AT OR
+ * AFTER that entry. If oversized trailing results exhaust the budget,
+ * retain their preceding assistant call, not the entire session.
+ * No encrypted-signature sizing and no exact retained-token guarantee.
+ * Returns the first kept record index; a zero/no-progress cut drops nothing.
+ * Keep this and estimateMessageTokens aligned deliberately when upgrading Pi.
+ */
+export function compactionCutIndex(messages: readonly LhMessage[], keepRecentTokens = KEEP_RECENT_TOKENS): number {
+	const previous = activeCompactionOf(messages);
+	const previousKept = previous ? keptIndexFor(messages, previous) : -1;
+	const start = Math.max(0, previousKept);
+	const points: number[] = [];
+	for (let i = start; i < messages.length; i++) {
+		const message = cutMessage(messages[i]!);
+		if (message && (message.role === "user" || message.role === "assistant" || message.role === "custom")) points.push(i);
 	}
-	return -1;
+	if (!points.length) return start;
+	let cut = points[0]!;
+	let accumulated = 0;
+	for (let i = messages.length - 1; i >= start; i--) {
+		const message = cutMessage(messages[i]!);
+		const tokens = message ? estimateMessageTokens(message) : 0;
+		if (tokens === 0) continue;
+		accumulated += tokens;
+		if (accumulated >= keepRecentTokens) {
+			cut = points.find((candidate) => candidate >= i) ?? points.at(-1)!;
+			break;
+		}
+	}
+	// Native entries preserve adjacent non-context metadata, but never walk
+	// backwards across a previous compaction entry.
+	while (cut > start) {
+		const before = messages[cut - 1]!;
+		if ((before.role === "custom" && before.customType === "compaction") || cutMessage(before)) break;
+		cut--;
+	}
+	return cut;
 }
 
 // ─── the marker ──────────────────────────────────────────────────────────────

src/core/context-usage.ts

@@ -0,0 +1,27 @@
+import type { AssistantMessage } from "@earendil-works/pi-ai";
+import type { LhMessage } from "./native-messages.ts";
+
+/** Latest completed conversation measurement, never a request-size estimate.
+ * Ordered compaction markers invalidate every earlier reading. Live callers
+ * must exclude assistant slots that have not received message_end.
+ * Reasoning is already inside output; totalTokens is deliberately not used.
+ */
+export function measuredContextTokens(
+	messages: readonly LhMessage[],
+	incompleteIds: ReadonlySet<string> = new Set(),
+): number | null {
+	for (let i = messages.length - 1; i >= 0; i--) {
+		const message = messages[i]!;
+		if (message.role === "custom" && message.customType === "compaction") return null;
+		if (message.role !== "assistant" || incompleteIds.has(message.id)) continue;
+		const assistant = message as AssistantMessage;
+		if (!["stop", "length", "toolUse"].includes(assistant.stopReason)) continue;
+		const usage = assistant.usage;
+		if (!usage) continue;
+		const buckets = [usage.input, usage.output, usage.cacheRead, usage.cacheWrite];
+		if (buckets.some((value) => !Number.isFinite(value) || value < 0)) continue;
+		const tokens = buckets.reduce((sum, value) => sum + value, 0);
+		if (tokens > 0) return tokens;
+	}
+	return null;
+}

src/core/contracts/chat.ts

@@ -87,8 +87,8 @@ export interface ChatSettings {
  *  billed (CallUsage per agent step). Steps, not a sum: cost and context
  *  folds happen at read time (core/prices.ts). */
 export interface ExtractionStageUsage {
-	/** The model that ran this stage, provider-qualified ("provider:id" —
-	 *  model slots wave, 2026-09-20; matches PRICES' uniform keys). */
+	/** The model that ran this stage, provider-qualified ("provider:id")
+	 * for key-independent Pi catalog lookup at read time. */
 	model: string;
 	steps: CallUsage[];
 }
@@ -118,7 +118,7 @@ export interface ExtractionRunUsage {
  * flattened to the fields we persist (undefined = provider didn't report;
  * renderers show absent, never zero). */
 export interface CallUsage {
-	/** Total input (prompt) tokens, cached portion included. */
+	/** Pi-normalized fresh input tokens; cache buckets are disjoint. */
 	inputTokens: number | undefined;
 	outputTokens: number | undefined;
 	/** Cached input tokens READ (billed at the cache-read rate, if any). */
@@ -126,6 +126,9 @@ export interface CallUsage {
 	/** Input tokens written to the cache this call (OpenRouter bills writes
 	 * for some providers; undefined = not reported/not billed). */
 	cacheWriteTokens: number | undefined;
+	/** Subset of cacheWriteTokens written with one-hour retention. Missing
+	 * on older logs; new native captures explicitly record zero if absent. */
+	cacheWrite1hTokens?: number;
 	/** The reasoning portion of the output tokens. */
 	reasoningTokens: number | undefined;
 }
@@ -144,8 +147,8 @@ export interface TurnUsage {
 	 *  entry to the transcript (the block this turn "belongs" to). */
 	messageId: string;
 	at: string;
-	/** The model that ran the turn, provider-qualified ("provider:id" —
-	 *  model slots wave, 2026-09-20; matches PRICES' uniform keys). */
+	/** The model that ran the turn, provider-qualified ("provider:id")
+	 * for key-independent Pi catalog lookup at read time. */
 	model: string;
 	/** One entry per agent step (model call) in this turn. */
 	steps: CallUsage[];

src/core/contracts/jobs.ts

@@ -43,7 +43,7 @@ export interface JobRecord {
  * money-channel doctrine; cost renders "—", never $0). */
 export interface JobUsage {
 	/** The model that ran the sub-agent, provider-qualified ("provider:id" —
-	 *  model slots wave, 2026-09-20; matches PRICES' uniform keys). */
+	 *  model slots wave, 2026-09-20; matches Pi catalog lookup keys). */
 	model: string;
 	/** Running totals summed over the run's model calls so far. */
 	inputTokens: number | undefined;
@@ -51,6 +51,10 @@ export interface JobUsage {
 	cacheReadTokens: number | undefined;
 	cacheWriteTokens: number | undefined;
 	reasoningTokens: number | undefined;
+	cacheWrite1hTokens?: number;
+	/** Individual calls for current-catalog per-call tier pricing. Absent on
+	 * older jobs; never backfill them from cumulative totals. */
+	calls?: CallUsage[];
 	/** How many model calls the run has made (each bills separately). */
 	steps: number;
 	/** The MOST RECENT model call's accounting (2026-09 dock contract): the
@@ -67,7 +71,7 @@ export interface JobUsage {
  * object with these fields folds, including the real thing and test fakes). */
 export interface UsageStepLike {
 	inputTokens?: number | undefined;
-	inputTokenDetails?: { cacheReadTokens?: number; cacheWriteTokens?: number } | undefined;
+	inputTokenDetails?: { cacheReadTokens?: number; cacheWriteTokens?: number; cacheWrite1hTokens?: number } | undefined;
 	outputTokens?: number | undefined;
 	outputTokenDetails?: { reasoningTokens?: number } | undefined;
 }
@@ -80,12 +84,13 @@ export function piUsageToStep(usage: {
 	output: number;
 	cacheRead: number;
 	cacheWrite: number;
+	cacheWrite1h?: number;
 	reasoning?: number;
 }): UsageStepLike {
 	return {
 		inputTokens: usage.input,
 		outputTokens: usage.output,
-		inputTokenDetails: { cacheReadTokens: usage.cacheRead, cacheWriteTokens: usage.cacheWrite },
+		inputTokenDetails: { cacheReadTokens: usage.cacheRead, cacheWriteTokens: usage.cacheWrite, cacheWrite1hTokens: usage.cacheWrite1h ?? 0 },
 		outputTokenDetails: usage.reasoning !== undefined ? { reasoningTokens: usage.reasoning } : undefined,
 	};
 }
@@ -101,6 +106,14 @@ export function accumulateJobUsage(
 ): JobUsage {
 	const add = (a: number | undefined, b: number | undefined): number | undefined =>
 		a === undefined && b === undefined ? undefined : (a ?? 0) + (b ?? 0);
+	const call: CallUsage = {
+		inputTokens: step.inputTokens,
+		outputTokens: step.outputTokens,
+		cacheReadTokens: step.inputTokenDetails?.cacheReadTokens,
+		cacheWriteTokens: step.inputTokenDetails?.cacheWriteTokens,
+		cacheWrite1hTokens: step.inputTokenDetails?.cacheWrite1hTokens,
+		reasoningTokens: step.outputTokenDetails?.reasoningTokens,
+	};
 	return {
 		model,
 		inputTokens: add(total?.inputTokens, step.inputTokens),
@@ -108,14 +121,12 @@ export function accumulateJobUsage(
 		cacheReadTokens: add(total?.cacheReadTokens, step.inputTokenDetails?.cacheReadTokens),
 		cacheWriteTokens: add(total?.cacheWriteTokens, step.inputTokenDetails?.cacheWriteTokens),
 		reasoningTokens: add(total?.reasoningTokens, step.outputTokenDetails?.reasoningTokens),
+		cacheWrite1hTokens: ((total?.cacheWriteTokens ?? 0) > 0 && total?.cacheWrite1hTokens === undefined) ||
+			((call.cacheWriteTokens ?? 0) > 0 && call.cacheWrite1hTokens === undefined)
+			? undefined : add(total?.cacheWrite1hTokens, call.cacheWrite1hTokens),
 		steps: (total?.steps ?? 0) + 1,
-		lastStep: {
-			inputTokens: step.inputTokens,
-			outputTokens: step.outputTokens,
-			cacheReadTokens: step.inputTokenDetails?.cacheReadTokens,
-			cacheWriteTokens: step.inputTokenDetails?.cacheWriteTokens,
-			reasoningTokens: step.outputTokenDetails?.reasoningTokens,
-		},
+		...(!total || total.calls ? { calls: [...(total?.calls ?? []), call] } : {}),
+		lastStep: call,
 	};
 }
 

src/core/contracts/models.ts

@@ -10,7 +10,7 @@ import type { BackgroundModelRole, ModelSlotName } from "../settings.ts";
 
 /** One picker row: everything the UI shows about a model, projected from
  *  pi-ai's catalog. Costs are USD per 1M tokens (the same catalog, the same
- *  units core/prices.ts's generated table carries). */
+ *  units as Pi's server-side pricing calculator). */
 export interface ModelCatalogEntry {
 	id: string;
 	name: string;

src/core/cost-summary.ts

@@ -9,7 +9,7 @@
  * the same ruling as wave 7's stages.embedding.)
  *
  * MONEY CHANNEL LAW: dollars are INFERRED here at read time via
- * core/prices.ts — never stored anywhere. An unpriced model contributes
+ * an injected server-side Pi pricer — never stored anywhere. An unpriced model contributes
  * ZERO dollars but is counted in `unpricedRuns` so the UI can render
  * "— unpriced" honestly, never a silent zero. Tokens always sum.
  *
@@ -18,7 +18,7 @@
  */
 
 import type { CallUsage, EmissionStats, ExtractionRunUsage } from "./contracts.ts";
-import { callCost, sumSteps, isChatGPTPlan } from "./prices.ts";
+import { sumSteps, isChatGPTPlan, type CallPricer } from "./prices.ts";
 
 /** One pipeline stage's cumulative cost across every run. */
 export interface StageCost {
@@ -26,7 +26,7 @@ export interface StageCost {
 	subscriptionRuns?: number;
 	/** Inferred USD summed over PRICED runs (partial when unpricedRuns > 0). */
 	dollars: number;
-	/** Stage-runs whose model is missing from PRICES (render "—", never zero). */
+	/** Stage-runs whose model is missing from Pi's catalog (render "—", never zero). */
 	unpricedRuns: number;
 	/** Runs that billed this stage. */
 	runs: number;
@@ -92,11 +92,11 @@ const emptyStage = (): StageCost => ({ dollars: 0, unpricedRuns: 0, runs: 0, tok
 
 /** One stage's dollars over its steps; null when the model is unpriced
  *  (all-or-nothing: a stage runs one model). */
-function stageDollars(model: string, steps: readonly CallUsage[]): number | null {
-	if (isChatGPTPlan(model)) return null;
+function stageDollars(model: string, steps: readonly CallUsage[], price: CallPricer): number | null {
+	if (isChatGPTPlan(model) || steps.length === 0) return null;
 	let total = 0;
 	for (const step of steps) {
-		const c = callCost(step, model);
+		const c = price(step, model);
 		if (c === null) return null;
 		total += c;
 	}
@@ -115,7 +115,7 @@ function mergeTokens(into: CallUsage, steps: readonly CallUsage[]): void {
 
 /** Fold every usage channel into the pill's payload. Pure: same input,
  *  same output — no clock, no fs, no store reads. */
-export function graphCostSummary(input: CostSummaryInput): GraphCostSummary {
+export function graphCostSummary(input: CostSummaryInput, price: CallPricer): GraphCostSummary {
 	const extraction = emptyStage();
 	const audit = emptyStage();
 	const runs: RecentRun[] = [];
@@ -125,14 +125,14 @@ export function graphCostSummary(input: CostSummaryInput): GraphCostSummary {
 		// Per-stage: tokens always count; dollars only when priced.
 		extraction.runs++;
 		mergeTokens(extraction.tokens, stages.extractor.steps);
-		const extDollars = stageDollars(stages.extractor.model, stages.extractor.steps);
+		const extDollars = stageDollars(stages.extractor.model, stages.extractor.steps, price);
 		if (isChatGPTPlan(stages.extractor.model)) extraction.subscriptionRuns = (extraction.subscriptionRuns ?? 0) + 1;
 		else if (extDollars === null) extraction.unpricedRuns++;
 		else extraction.dollars += extDollars;
 
 		audit.runs++;
 		mergeTokens(audit.tokens, stages.auditor.steps);
-		const audDollars = stageDollars(stages.auditor.model, stages.auditor.steps);
+		const audDollars = stageDollars(stages.auditor.model, stages.auditor.steps, price);
 		if (isChatGPTPlan(stages.auditor.model)) audit.subscriptionRuns = (audit.subscriptionRuns ?? 0) + 1;
 		else if (audDollars === null) audit.unpricedRuns++;
 		else audit.dollars += audDollars;

src/core/index.ts

@@ -48,6 +48,7 @@ export * from "./input-events.ts";
 export * from "./view.ts";
 export * from "./tool-parts.ts";
 export * from "./prices.ts";
+export * from "./context-usage.ts";
 export type { ChatGPTStatus, ChatGPTControlBridge } from "./contracts/chatgpt.ts";
 export * from "./cost-summary.ts";
 export * from "./fuzzy.ts";