fix: align compaction cuts and native usage accounting

src/server/chat-handler.ts

@@ -333,6 +333,7 @@ export async function handleChatInput(
 			outputTokens: usage.output,
 			cacheReadTokens: usage.cacheRead,
 			cacheWriteTokens: usage.cacheWrite,
+			cacheWrite1hTokens: usage.cacheWrite1h ?? 0,
 			// Optional in pi's Usage — persisted as 0, never dropped.
 			reasoningTokens: usage.reasoning ?? 0,
 		});

src/server/compaction.ts

@@ -175,7 +175,7 @@ async function compact(deps: CompactionDeps, chatId: string): Promise<void> {
 	if (tokensBefore <= deps.thresholdTokens) return; // settled beneath while we waited
 
 	const cut = compactionCutIndex(messages);
-	if (cut >= messages.length) return; // nothing to drop (defensive)
+	if (cut <= 0 || cut >= messages.length) return; // no history can be dropped
 
 	// The previous compaction's kept boundary (chained compactions
 	// summarize the old summary + what rode after it).
@@ -240,6 +240,7 @@ async function compact(deps: CompactionDeps, chatId: string): Promise<void> {
 					outputTokens: usage.output,
 					cacheReadTokens: usage.cacheRead,
 					cacheWriteTokens: usage.cacheWrite,
+					cacheWrite1hTokens: usage.cacheWrite1h ?? 0,
 					reasoningTokens: usage.reasoning ?? 0,
 				},
 			],

src/server/graph.ts

@@ -11,6 +11,7 @@ import {
 } from "../core/graph-deletion.ts";
 import { extractionStatuses } from "../agent/subagents/extraction.ts";
 import { stores } from "./stores.ts";
+import { callCost } from "./pricing.ts";
 
 /**
  * Server-side graph entry for the home view (GET /api/graph). A PURE READ:
@@ -41,7 +42,7 @@ export function graphCosts(): GraphCostSummary {
 		.readSources()
 		.filter((s) => s.extractions?.length)
 		.map((s) => ({ filePath: s.filePath, extractions: s.extractions! }));
-	return graphCostSummary({ chats, sources });
+	return graphCostSummary({ chats, sources }, callCost);
 }
 
 /** The graph's version stamp (session-deletion wave 6): the max mtime

src/server/index.ts

@@ -43,3 +43,4 @@ export * from "./compaction.ts";
 export { runPiTurn, piTurnResponse, type PiTurnHandle, type PiTurnSpec } from "./pi/ui-stream.ts";
 export * from "./observability.ts";
 export * from "./model-catalog.ts";
+export { sessionCost } from "./pricing.ts";

src/server/observability.ts

@@ -24,6 +24,7 @@ import { chatStore } from "../store/chats.ts";
 import { graphRevision } from "./graph.ts";
 import { catalogContextLimits } from "./model-catalog.ts";
 import { listJobs } from "./jobs.ts";
+import { jobCost } from "./pricing.ts";
 import type { JobRecord } from "../core/contracts.ts";
 
 export { abortExtraction, extractionStatuses, type ExtractionPhase, type RunningExtraction } from "../agent/subagents/extraction.ts";
@@ -36,6 +37,7 @@ export { chatStore } from "../store/chats.ts";
  *  model is missing from the catalog (the blob then renders no context
  *  line — never a fake window). */
 export interface ActivityJob extends JobRecord {
+	costUsd: number | null;
 	title: string;
 	contextLimit: number | null;
 }
@@ -56,9 +58,8 @@ export interface ActivityExtraction {
 
 /** The dock's whole world in one payload: every running process, every view.
  *  Jobs are RUNNING-ONLY — finished jobs belong to their chat's transcript;
- *  the blob disappears when the process ends. Cost math is NOT done here:
- *  raw usage rides on the records and the client folds it with the pure
- *  helpers in core/prices.ts (transport stays thin). */
+ *  the blob disappears when the process ends. Costs are joined here through
+ *  the server-side Pi pricing adapter; clients never reconstruct rates. */
 /** The dock's whole world + the graph's version stamp (wave 6 session
  *  deletion): the payload carries graphRevision (max mtime across the
  *  graph data files — see server/graph.ts) so any view can detect "the
@@ -78,6 +79,7 @@ export function activityFeed(): {
 			if (job.status === "running") {
 				jobs.push({
 					...job,
+					costUsd: jobCost(job),
 					title: chat.title ?? chat.id,
 					contextLimit: job.usage ? (limits.get(job.usage.model) ?? null) : null,
 				});

src/server/pricing.ts

@@ -0,0 +1,50 @@
+/** Read-time API-dollar estimates. Pi owns both rates and arithmetic.
+ * Catalog resolution is key-independent; no provider dispatch/auth is needed.
+ * Never mutate saved native messages: calculateCost mutates its usage.cost.
+ */
+import { calculateCost, type Usage } from "@earendil-works/pi-ai";
+import { piModels } from "../agent/models.ts";
+import type { CallUsage, ChatRecord, JobRecord, TurnUsage } from "../core/contracts.ts";
+import { isChatGPTPlan, sessionCost as foldSessionCost, turnCost as foldTurnCost, jobCost as foldJobCost } from "../core/prices.ts";
+
+function catalogModel(id: string) {
+	const colon = id.indexOf(":");
+	if (colon < 0 || isChatGPTPlan(id)) return undefined;
+	return piModels().getModel(id.slice(0, colon), id.slice(colon + 1));
+}
+
+export function callCost(call: Partial<CallUsage>, modelId: string): number | null {
+	const model = catalogModel(modelId);
+	if (!model || call.inputTokens === undefined || call.outputTokens === undefined) return null;
+	const input = call.inputTokens;
+	const output = call.outputTokens;
+	const cacheRead = call.cacheReadTokens ?? 0;
+	const cacheWrite = call.cacheWriteTokens ?? 0;
+	const cacheWrite1h = call.cacheWrite1hTokens ?? 0;
+	if ([input, output, cacheRead, cacheWrite, cacheWrite1h].some((v) => !Number.isFinite(v) || v < 0) || cacheWrite1h > cacheWrite) return null;
+	// Old records discarded this split. Only price them if it cannot affect
+	// the result; do not invent short-retention precision from a missing field.
+	if (cacheWrite > 0 && call.cacheWrite1hTokens === undefined &&
+		[model.cost, ...(model.cost.tiers ?? [])].some((rate) => rate.cacheWrite !== 2 * rate.input)) return null;
+	const usage: Usage = {
+		input, output, cacheRead, cacheWrite, cacheWrite1h,
+		totalTokens: input + output + cacheRead + cacheWrite,
+		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+	};
+	const total = calculateCost(model, usage).total;
+	return Number.isFinite(total) && total >= 0 ? total : null;
+}
+
+export function turnCost(turn: TurnUsage): number | null {
+	return foldTurnCost(turn, callCost);
+}
+export function sessionCost(record: Pick<ChatRecord, "turns">) {
+	return foldSessionCost(record, callCost);
+}
+export function jobCost(job: Pick<JobRecord, "usage">): number | null {
+	const usage = job.usage;
+	if (!usage || usage.steps < 1) return null;
+	const hasCalls = usage.calls?.length === usage.steps;
+	if (!hasCalls && usage.steps > 1 && catalogModel(usage.model)?.cost.tiers?.length) return null;
+	return foldJobCost(job, callCost);
+}