Harden session persistence, worker recovery, and daemon refreshes

packages/ai/.changes/service-tier-default-omitted.md

@@ -0,0 +1 @@
+- Fixed GitHub Copilot requests to omit unsupported service tiers while preserving explicit tiers for other providers, and corrected Anthropic cache-write pricing when streaming usage changes.

packages/ai/src/providers/anthropic.ts

@@ -6,7 +6,11 @@ import type {
 	MessageParam,
 	RawMessageStreamEvent,
 } from "@anthropic-ai/sdk/resources/messages.js";
-import { getAnthropicCacheWriteCost, hasStandardAnthropicCachePricing } from "../cache-pricing.js";
+import {
+	type AnthropicCacheCreationUsage,
+	getAnthropicCacheWriteCost,
+	hasStandardAnthropicCachePricing,
+} from "../cache-pricing.js";
 import { getEnvApiKey } from "../env-api-keys.js";
 import { calculateCost, clampThinkingLevel } from "../models.js";
 import type {
@@ -695,6 +699,16 @@ export const streamAnthropic: StreamFunction<"anthropic-messages", AnthropicOpti
 					if (event.usage.cache_creation_input_tokens != null) {
 						output.usage.cacheWrite = event.usage.cache_creation_input_tokens;
 					}
+					// The SDK's MessageDeltaUsage type omits cache_creation, but the wire carries it.
+					const deltaCacheCreation = (event.usage as { cache_creation?: AnthropicCacheCreationUsage | null })
+						.cache_creation;
+					if (cacheControl && usesAnthropicCachePricing && deltaCacheCreation) {
+						cacheWriteCost = getAnthropicCacheWriteCost(
+							model.cost.input,
+							cacheControl.ttl === "1h" ? "1h" : "5m",
+							deltaCacheCreation,
+						);
+					}
 					output.usage.totalTokens =
 						output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
 					calculateCost(

packages/ai/src/providers/openai-responses.ts

@@ -237,7 +237,9 @@ function buildParams(model: Model<"openai-responses">, context: Context, options
 		params.temperature = options?.temperature;
 	}
 
-	if (options?.serviceTier !== undefined) {
+	// GitHub Copilot rejects the service_tier FIELD itself (400) for every value.
+	// Elsewhere it is always sent: absence means "auto" (project tier), not "default".
+	if (options?.serviceTier !== undefined && model.provider !== "github-copilot") {
 		params.service_tier = options.serviceTier;
 	}
 

packages/ai/test/anthropic-sse-parsing.test.ts

@@ -78,11 +78,13 @@ function createFakeAnthropicClient(response: Response): Anthropic {
 	} as unknown as Anthropic;
 }
 
-function createCacheUsageEvents(cacheCreation: {
-	ephemeral_5m_input_tokens: number;
-	ephemeral_1h_input_tokens: number;
-}): Array<{ event: string; data: string }> {
-	const cacheWriteTokens = cacheCreation.ephemeral_5m_input_tokens + cacheCreation.ephemeral_1h_input_tokens;
+type CacheCreation = { ephemeral_5m_input_tokens: number; ephemeral_1h_input_tokens: number };
+
+function createCacheUsageEvents(
+	cacheCreation: CacheCreation,
+	deltaCacheCreation?: CacheCreation,
+): Array<{ event: string; data: string }> {
+	const tokens = (c: CacheCreation) => c.ephemeral_5m_input_tokens + c.ephemeral_1h_input_tokens;
 	return [
 		{
 			event: "message_start",
@@ -94,7 +96,7 @@ function createCacheUsageEvents(cacheCreation: {
 						input_tokens: 12,
 						output_tokens: 0,
 						cache_read_input_tokens: 0,
-						cache_creation_input_tokens: cacheWriteTokens,
+						cache_creation_input_tokens: tokens(cacheCreation),
 						cache_creation: cacheCreation,
 					},
 				},
@@ -109,7 +111,8 @@ function createCacheUsageEvents(cacheCreation: {
 					input_tokens: 12,
 					output_tokens: 5,
 					cache_read_input_tokens: 0,
-					cache_creation_input_tokens: cacheWriteTokens,
+					cache_creation_input_tokens: tokens(deltaCacheCreation ?? cacheCreation),
+					...(deltaCacheCreation ? { cache_creation: deltaCacheCreation } : {}),
 				},
 			}),
 		},
@@ -153,6 +156,24 @@ describe("Anthropic raw SSE parsing", () => {
 		expect(result.usage.cost.cacheWrite).toBeCloseTo(testCase.expectedCacheWriteCost);
 	});
 
+	it("reprices cache writes from a message_delta usage breakdown", async () => {
+		const model = getModel("anthropic", "claude-haiku-4-5");
+		const response = createSseResponse(
+			createCacheUsageEvents(
+				{ ephemeral_5m_input_tokens: 1000, ephemeral_1h_input_tokens: 0 },
+				{ ephemeral_5m_input_tokens: 0, ephemeral_1h_input_tokens: 2000 },
+			),
+		);
+		const result = await streamAnthropic(
+			model,
+			{ messages: [{ role: "user", content: "Say hello.", timestamp: Date.now() }] },
+			{ client: createFakeAnthropicClient(response), cacheRetention: "long" },
+		).result();
+
+		expect(result.usage.cacheWrite).toBe(2000);
+		// 2000 one-hour tokens at 2x input cost, not the stale 1.25x rate from message_start.
+		expect(result.usage.cost.cacheWrite).toBeCloseTo(0.004, 6);
+	});
 	it("preserves configured cache write pricing for non-Anthropic models", async () => {
 		const model = getModel("minimax", "MiniMax-M2.7-highspeed");
 		const response = createSseResponse(

packages/ai/test/openai-codex-stream.test.ts

@@ -562,6 +562,8 @@ describe("openai-codex streaming", () => {
 	});
 
 	it.each([
+		// "default" must stay on the wire: absence means "auto" (the project tier).
+		["gpt-5.1-codex", "default", 1],
 		["gpt-5.1-codex", "flex", 0.5],
 		["gpt-5.1-codex", "priority", 2],
 		["gpt-5.4", "priority", 2],

packages/ai/test/openai-responses-copilot-provider.test.ts

@@ -212,6 +212,41 @@ describe("openai-responses provider defaults", () => {
 		expect(captured).toEqual({ sessionId: null, clientRequestId: null });
 	});
 
+	it.each([
+		["github-copilot" as const, "auto" as const, false],
+		["github-copilot" as const, "default" as const, false],
+		["openai" as const, "default" as const, true],
+	])("scopes service_tier serialization to the provider (%s, %s)", async (provider, serviceTier, expected) => {
+		const base = getModel("openai", "gpt-5.4");
+		const model = { ...base, provider };
+		const sse = `data: ${JSON.stringify({
+			type: "response.completed",
+			response: {
+				status: "completed",
+				usage: { input_tokens: 1, output_tokens: 1, total_tokens: 2, input_tokens_details: { cached_tokens: 0 } },
+			},
+		})}\n\n`;
+		let wireBody: Record<string, unknown> | undefined;
+		vi.spyOn(globalThis, "fetch").mockImplementation(async (_input, init) => {
+			wireBody = JSON.parse(String(init?.body)) as Record<string, unknown>;
+			return new Response(sse, { status: 200, headers: { "content-type": "text/event-stream" } });
+		});
+
+		const result = await streamOpenAIResponses(
+			model,
+			{ systemPrompt: "sys", messages: [{ role: "user", content: "hi", timestamp: Date.now() }] },
+			{ apiKey: "test-key", serviceTier },
+		).result();
+
+		expect(result.stopReason).toBe("stop");
+		// Copilot rejects the FIELD for every value; elsewhere absence means "auto"
+		// (the project tier), so an explicit "default" must stay on the wire.
+		expect(wireBody && "service_tier" in wireBody).toBe(expected);
+		if (expected) {
+			expect((wireBody as Record<string, unknown>).service_tier).toBe(serviceTier);
+		}
+	});
+
 	it.each([
 		["gpt-5.4", "priority", 2],
 		["gpt-5.5", "priority", 2.5],

packages/coding-agent/.changes/dead-kernel-memo.md

@@ -0,0 +1 @@
+- A Python kernel that dies after a successful startup is restarted on the next use instead of every call being handed the dead kernel forever, and skill-MCP tools advertise their real input schemas again under mcp>=2 (the SDK renamed the field to input_schema).

packages/coding-agent/.changes/res-1260-semantic-edges-on-event-log.md

@@ -0,0 +1 @@
+- Moved the semantic-edge ledger's append and replay IO onto the shared event-log substrate. One behavior unified across both ledgers: an unterminated final line is an uncommitted append — skipped on read and truncated before the next append, never newline-completed.

packages/coding-agent/.changes/res-1267-atomic-persistence.md

@@ -0,0 +1 @@
+- Made every durable JSON/JSONL state write crash-safe through one shared atomic-write owner (temp file + rename, Windows rename retry): auth.json is no longer written in place (an interrupted write can no longer log you out everywhere), the auth migration writes its destination before destroying its sources, racing first-time settings writers no longer silently discard each other, and the kernel bootstrap lock can no longer be stolen mid-reclaim. Session files now repair crash damage (torn tails, zero-filled records) at open instead of silently losing the next message, and a session lease whose owner file is momentarily unreadable is no longer treated as stale and destroyed.

packages/coding-agent/.changes/res-1272-coalesced-bookkeeping-appends.md

@@ -0,0 +1 @@
+- Fixed unbounded session-journal growth from derived bookkeeping: child usage attribution now flushes one entry per child turn instead of one per model request, and idle status sweeps no longer persist fabricated fallback verdicts, duplicate statuses, or retry failed summary generations (including paid model calls) every 25 seconds on unchanged content.

packages/coding-agent/.changes/res-1272-incremental-session-scans.md

@@ -0,0 +1 @@
+- Fixed session-list refreshes re-reading entire session files on every change: metadata scans now resume from the last scanned byte offset, stop at the file size seen at scan start, and concurrent readers of the same session share one scan.

packages/coding-agent/.changes/res-1272-memoized-passive-topology.md

@@ -0,0 +1 @@
+- Fixed daemon request latency on large agent trees: the passive-subagent topology is derived once and memoized, with every consumer (session list, snapshots, cron recovery, agent messaging, passivation) reading the cached walk until the spawn ledger, residency, or a child session file changes.

packages/coding-agent/.changes/session-io-defects.md

@@ -0,0 +1 @@
+- Seven small correctness fixes: compaction keeps only the final turn when the budget is crossed inside trailing tool results (instead of silently keeping everything); a retry whose scheduled continue cannot run ends the retry instead of leaving the session stuck retrying; saved subagent sessions with a lost parent edge still display as subagents; tail truncation rescues an oversized final line even when output ends with a newline; a failed output-spill stream degrades to the in-memory tail instead of crashing the process; piped stdin and a CLI instruction are joined with a blank line instead of glued together; and frontmatter parses behind a UTF-8 BOM.

packages/coding-agent/.changes/session-reader-safety.md

@@ -0,0 +1,2 @@
+- Prevented session export and daemon-client startup from repairing or rewriting transcripts owned by another process.
+- Enforced the retained session-scan usage cache limit for oversized transcripts.

packages/coding-agent/.changes/webp-exif-scan-termination.md

@@ -0,0 +1 @@
+- The WebP EXIF chunk scan reads chunk sizes as unsigned, so a crafted or corrupt image can no longer hang the process in an infinite scan loop.

packages/coding-agent/.changes/worker-snapshot-cursor.md

@@ -0,0 +1 @@
+- Fixed chunked session-snapshot transfers so the transfer id names the exact materialized snapshot cut, and a mismatched or restarted transfer now fails only that transfer (clients resync) instead of bouncing the whole worker channel.

packages/coding-agent/.changes/worker-state-one-liveness-probe.md

@@ -0,0 +1 @@
+- Fixed zombie processes being treated as live owners by the daemon supervisor ownership registry, session leases, supervisor launch locks, `daemon ps` process stops, and update-restart liveness checks; all process liveness probes now share the zombie-aware helper.

packages/coding-agent/.changes/worker-state-recovery.md

@@ -0,0 +1 @@
+- Fixed daemon sessions bricking behind a terminal failed worker state: attach, create, and retry now re-run recovery for a failed worker whose process is verified alive, and a known-but-still-recovering session answers with a structured retryable error instead of "Unknown active session".

packages/coding-agent/.changes/zai-default-model.md

@@ -0,0 +1 @@
+- The zai provider default model now points at glm-5.3; the previous default was removed from the catalog and silently fell back to a template model.

packages/coding-agent/src/cli/daemon-ps.ts

@@ -19,7 +19,12 @@ import {
 import { defaultDaemonSocketDir, defaultDaemonSocketPath, normalizeSocketPath } from "../modes/daemon/daemon-socket.js";
 import { acquireDaemonShutdownAdmission } from "../modes/daemon/daemon-supervisor-ownership.js";
 import type { DaemonWorkerDescriptor } from "../modes/daemon/daemon-worker-protocol.js";
-import { signalProcessGroupOrProcess } from "../utils/child-process.js";
+import {
+	isProcessAlive,
+	processGroupHasLiveMember,
+	processIdExists,
+	signalProcessGroupIfHeld,
+} from "../utils/child-process.js";
 import { formatDaemonListTable } from "./daemon-ps-format.js";
 import { promptYesNo } from "./daemon-stop-confirm.js";
 
@@ -1061,34 +1066,47 @@ async function stopTrackedProcess(
 	expectedStartId: string | undefined,
 	assertAdmission: () => Promise<void>,
 ): Promise<boolean> {
-	if (!isProcessAlive(pid)) {
+	if (trackedProcessStopped(pid)) {
 		return true;
 	}
-	if (!expectedStartId || getProcessStartId(pid) !== expectedStartId) {
+	if (!expectedStartId || !trackedLeaderIdentityCurrent(pid, expectedStartId)) {
 		return false;
 	}
 	await assertAdmission();
-	if (getProcessStartId(pid) !== expectedStartId) {
+	if (!trackedLeaderIdentityCurrent(pid, expectedStartId)) {
 		return false;
 	}
-	signalProcessGroupOrProcess(pid, "SIGTERM");
+	signalProcessGroupIfHeld(pid, "SIGTERM");
 	let deadline = Date.now() + 500;
-	while (isProcessAlive(pid) && Date.now() < deadline) {
+	while (!trackedProcessStopped(pid) && Date.now() < deadline) {
 		await delay(25);
 	}
-	if (!isProcessAlive(pid)) {
+	if (trackedProcessStopped(pid)) {
 		return true;
 	}
 	await assertAdmission();
-	if (getProcessStartId(pid) !== expectedStartId) {
+	if (!trackedLeaderIdentityCurrent(pid, expectedStartId)) {
 		return false;
 	}
-	signalProcessGroupOrProcess(pid, "SIGKILL");
+	signalProcessGroupIfHeld(pid, "SIGKILL");
 	deadline = Date.now() + 1000;
-	while (isProcessAlive(pid) && Date.now() < deadline) {
+	while (!trackedProcessStopped(pid) && Date.now() < deadline) {
 		await delay(25);
 	}
-	return !isProcessAlive(pid);
+	return trackedProcessStopped(pid);
+}
+
+/** A GROUP stop completes when the leader is gone AND no live member remains; unreaped zombies do not block it. */
+function trackedProcessStopped(pid: number): boolean {
+	return !isProcessAlive(pid) && !processGroupHasLiveMember(pid);
+}
+
+/** Identity gates guard pid reuse, so they apply only while the leader exists; a pgid cannot be reused while members hold it. */
+function trackedLeaderIdentityCurrent(pid: number, expectedStartId: string): boolean {
+	if (!processIdExists(pid)) {
+		return true;
+	}
+	return getProcessStartId(pid) === expectedStartId;
 }
 
 export async function runReap(json: boolean, force: boolean): Promise<void> {
@@ -1220,15 +1238,6 @@ async function forceKillDaemon(pid: number): Promise<void> {
 	}
 }
 
-function isProcessAlive(pid: number): boolean {
-	try {
-		process.kill(pid, 0);
-		return true;
-	} catch (error) {
-		return (error as NodeJS.ErrnoException).code === "EPERM";
-	}
-}
-
 function delay(ms: number): Promise<void> {
 	return new Promise((resolve) => setTimeout(resolve, ms));
 }

packages/coding-agent/src/cli/daemon-update-restart.ts

@@ -14,6 +14,7 @@ import {
 	DAEMON_WORKER_SUPERVISOR_SOCKET_ENV,
 	DAEMON_WORKER_TOKEN_ENV,
 } from "../modes/daemon/daemon-worker-protocol.js";
+import { isProcessAlive } from "../utils/child-process.js";
 import { createCliSubprocessLaunchSpec } from "./subprocess-launch.js";
 
 export const DAEMON_UPDATE_RESTART_COORDINATOR_FLAG = "--internal-update-restart-coordinator";
@@ -354,15 +355,6 @@ async function withCoordinatorRegistryGuard<T>(registryDir: string, action: () =
 	}
 }
 
-function isProcessAlive(pid: number): boolean {
-	try {
-		process.kill(pid, 0);
-	} catch (error) {
-		return (error as NodeJS.ErrnoException).code !== "ESRCH";
-	}
-	return true;
-}
-
 function matchesProcessStartId(identity: DaemonUpdateRestartProcessIdentity): boolean {
 	if (!identity.processStartId) {
 		return true;

packages/coding-agent/src/cli/initial-message.ts

@@ -37,7 +37,7 @@ export function buildInitialMessage({
 	}
 
 	return {
-		initialMessage: parts.length > 0 ? parts.join("") : undefined,
+		initialMessage: parts.length > 0 ? parts.join("\n\n") : undefined,
 		initialImages: fileImages && fileImages.length > 0 ? fileImages : undefined,
 	};
 }