Merge remote-tracking branch 'origin/main' into add-claude-code-support

# Conflicts:
#	package-lock.json
#	src/agent/models.ts

src/agent/chatgpt-bridge.ts

@@ -0,0 +1,128 @@
+import type { Model } from "@earendil-works/pi-ai";
+import type { ChatGPTStatus } from "../core/contracts/chatgpt.ts";
+import type { AdmissionScope, NoDispatchKind } from "./chatgpt-request-cycle.ts";
+export type { ChatGPTStatus } from "../core/contracts/chatgpt.ts";
+export interface Stamp { epoch: number; authorityRevision: number; revision: number }
+interface StatusValue extends Stamp { status: ChatGPTStatus }
+interface ModelsValue extends StatusValue { accountId: string | null; models: { id: string; name: string }[] }
+interface LeaseValue extends Stamp { accessToken: string; accountId: string }
+interface Pending { op: string; resolve: (v: unknown) => void; reject: (e: Error) => void; timer: ReturnType<typeof setTimeout> }
+interface BridgeState {
+ status: ChatGPTStatus; stamp: Stamp; models: Model<"openai-responses">[]; catalogAccount: string | null;
+ pending: Map<string, Pending>; controllers: Map<AbortController, boolean>; listeners: Set<() => void>; counter: number; catalogRequest: number;
+}
+const key = Symbol.for("lh.chatgpt.bridge");
+const unavailable = (): ChatGPTStatus => ({ available: false, pending: false, allowBackgroundUse: true, activeId: null, accounts: [], error: null });
+const validStamp = (s: Stamp) => s && [s.epoch, s.authorityRevision, s.revision].every((v) => Number.isSafeInteger(v) && v >= 0);
+/** Next dev also uses parent/child IPC. IPC alone is not our desktop
+ * transport: require the same launch markers set by Electron startServer.
+ * These identify the transport, not a substitute for admission/lease checks. */
+function desktopIPCConnected(): boolean {
+ return process.env.LH_DESKTOP_SERVER === "1" && (process.env.LH_DESKTOP_API_SECRET?.length ?? 0) >= 64 && typeof process.send === "function" && process.connected === true;
+}
+function bridge(): BridgeState {
+ const global = globalThis as typeof globalThis & { [key]?: BridgeState };
+ if (!global[key]) {
+  const state: BridgeState = { status: unavailable(), stamp: { epoch: 0, authorityRevision: 0, revision: 0 }, models: [], catalogAccount: null, pending: new Map(), controllers: new Map(), listeners: new Set(), counter: 0, catalogRequest: 0 };
+  global[key] = state;
+  process.on("message", (raw: unknown) => {
+   const m = raw as { channel?: string; kind?: string; id?: string; op?: string; ok?: boolean; value?: unknown; stamp: Stamp; status: ChatGPTStatus; invalidate?: string; error?: string };
+   if (!desktopIPCConnected() || m?.channel !== "lh:chatgpt") return;
+   if (m.kind === "state" && validStamp(m.stamp) && m.stamp.revision >= state.stamp.revision) {
+    acceptStatus(state, { ...m.stamp, status: m.status });
+    if (m.invalidate === "all") { state.catalogRequest++; clearCatalog(state); }
+    if (m.invalidate === "all" || m.invalidate === "background") {
+     for (const [controller, background] of state.controllers) if (m.invalidate === "all" || background) controller.abort(m.invalidate === "all" ? "all" : "background");
+    }
+    for (const listener of state.listeners) listener();
+   }
+   if (m.id) {
+    const call = state.pending.get(m.id);
+    if (!call || call.op !== m.op) return;
+    clearTimeout(call.timer); state.pending.delete(m.id);
+    if (m.ok === true) call.resolve(m.value);
+    else call.reject(new Error(m.error ?? "ChatGPT connection failed."));
+   }
+  });
+  process.on("disconnect", () => {
+   state.status = unavailable(); clearCatalog(state); state.catalogRequest++;
+   for (const controller of state.controllers.keys()) controller.abort("all");
+   for (const call of state.pending.values()) { clearTimeout(call.timer); call.reject(new Error("The desktop connection closed.")); }
+   state.pending.clear();
+   for (const listener of state.listeners) listener();
+  });
+ }
+ return global[key];
+}
+function clearCatalog(state: BridgeState) { state.models = []; state.catalogAccount = null; }
+function acceptStatus(state: BridgeState, value: StatusValue) {
+ if (!validStamp(value) || value.revision < state.stamp.revision) return;
+ if (value.epoch !== state.stamp.epoch || value.authorityRevision !== state.stamp.authorityRevision || value.status.activeId !== state.status.activeId) clearCatalog(state);
+ state.status = value.status; state.stamp = { epoch: value.epoch, authorityRevision: value.authorityRevision, revision: value.revision };
+}
+/** Runtime-only inherited IPC. No HTTP credential route or renderer method. */
+export function chatGPTRequest<T>(op: "status" | "models" | "lease", fields: Record<string, unknown> = {}): Promise<T> {
+ const state = bridge();
+ if (!desktopIPCConnected()) return Promise.reject(new Error("ChatGPT sign-in is available only in the desktop app with a private server connection."));
+ return new Promise<T>((resolve, reject) => {
+  const id = `${process.pid}:${++state.counter}`;
+  const timer = setTimeout(() => { state.pending.delete(id); reject(new Error("The ChatGPT desktop connection timed out.")); }, 45000);
+  state.pending.set(id, { op, resolve: (v) => resolve(v as T), reject, timer });
+  process.send!({ ...fields, channel: "lh:chatgpt", id, op }, (error) => {
+   if (error) { clearTimeout(timer); state.pending.delete(id); reject(new Error("The ChatGPT desktop connection closed.")); }
+  });
+ });
+}
+export function chatGPTStatus(): ChatGPTStatus { return bridge().status; }
+export function chatGPTConnected(): boolean { const s = chatGPTStatus(); return desktopIPCConnected() && s.available && s.accounts.some((a) => a.id === s.activeId && a.connected && a.planEnabled === true); }
+export function chatGPTBackgroundAllowed(): boolean { const s = chatGPTStatus(); return chatGPTConnected() && s.allowBackgroundUse === true; }
+export function onChatGPTChange(listener: () => void): () => void { bridge().listeners.add(listener); return () => { bridge().listeners.delete(listener); }; }
+export function chatGPTModels(): Model<"openai-responses">[] { return bridge().models; }
+export function chatGPTModel(id: string): Model<"openai-responses"> | undefined { return chatGPTModels().find((m) => m.id === id); }
+export function chatGPTScope(selectedModelId: string): AdmissionScope {
+ const state = bridge();
+ if (!chatGPTConnected() || !state.catalogAccount || state.catalogAccount !== state.status.activeId || !chatGPTModel(selectedModelId)) throw new Error("ChatGPT account model catalog is unavailable");
+ return { accountId: state.catalogAccount, epoch: state.stamp.epoch, authorityRevision: state.stamp.authorityRevision, selectedModelId };
+}
+export function chatGPTScopeCurrent(scope: AdmissionScope): boolean {
+ const s = bridge();
+ return desktopIPCConnected() && chatGPTConnected() && scope.accountId === s.catalogAccount && scope.accountId === s.status.activeId && scope.epoch === s.stamp.epoch && scope.authorityRevision === s.stamp.authorityRevision && !!chatGPTModel(scope.selectedModelId);
+}
+export async function chatGPTLease(scope: AdmissionScope, background: boolean): Promise<LeaseValue> {
+ if (!chatGPTScopeCurrent(scope)) throw new Error("ChatGPT catalog authority changed");
+ const value = await chatGPTRequest<LeaseValue>("lease", { background, expectedAccountId: scope.accountId, expectedAuthorityRevision: scope.authorityRevision, selectedModelId: scope.selectedModelId });
+ if (!validStamp(value) || !chatGPTScopeCurrent(scope) || value.epoch !== scope.epoch || value.authorityRevision !== scope.authorityRevision || value.accountId !== scope.accountId || typeof value.accessToken !== "string" || !value.accessToken || value.accessToken.startsWith("sk-")) throw new Error("ChatGPT lease is stale");
+ return value;
+}
+/** Safe status confirmation can establish consent-only denial, never parse error text. */
+export async function chatGPTDenial(scope: AdmissionScope | undefined, controller: AbortController, callerAborted: boolean, background: boolean): Promise<NoDispatchKind> {
+ if (callerAborted || controller.signal.reason === "all") return "all";
+ if (!scope || !chatGPTScopeCurrent(scope)) return "other";
+ if (controller.signal.reason === "background") return "background";
+ if (!background) return "other";
+ try { const value = await chatGPTRequest<StatusValue>("status"); acceptStatus(bridge(), value); }
+ catch { return "other"; }
+ if (controller.signal.reason === "all" || !chatGPTScopeCurrent(scope)) return "all";
+ return !chatGPTBackgroundAllowed() ? "background" : "other";
+}
+/** Preserve account order; unknown capabilities/windows are never fabricated. */
+export async function refreshChatGPTModels(supported: Model<"openai-responses">[]): Promise<void> {
+ const state = bridge();
+ if (!desktopIPCConnected()) { clearCatalog(state); return; }
+ const sequence = ++state.catalogRequest;
+ try {
+  const value = await chatGPTRequest<ModelsValue>("models");
+  if (sequence !== state.catalogRequest || !validStamp(value)) return;
+  acceptStatus(state, value);
+  if (value.epoch !== state.stamp.epoch || value.authorityRevision !== state.stamp.authorityRevision || value.accountId !== state.status.activeId || !chatGPTConnected()) return;
+  state.catalogAccount = value.accountId;
+  state.models = value.models.flatMap((entry) => {
+   const known = supported.find((m) => m.id === entry.id && m.api === "openai-responses");
+   return known ? [{ ...known, provider: "chatgpt", name: entry.name, headers: undefined, samplingParams: undefined, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 } }] : [];
+  });
+ } catch { if (sequence === state.catalogRequest) clearCatalog(state); }
+}
+export function chatGPTAbortController(background = true): { controller: AbortController; release: () => void } {
+ const state = bridge(); const controller = new AbortController(); state.controllers.set(controller, background);
+ return { controller, release: () => { state.controllers.delete(controller); } };
+}

src/agent/chatgpt-request-cycle.ts

@@ -0,0 +1,122 @@
+import type { AgentEvent, AgentMessage, PrepareRequestContext } from "@earendil-works/pi-agent-core";
+import type { AssistantMessage } from "@earendil-works/pi-ai";
+import { nativeInputCause, needsContinuation, type LhMessage } from "../core/native-messages.ts";
+
+export interface AdmissionScope { accountId: string; epoch: number; authorityRevision: number; selectedModelId: string }
+export type NoDispatchKind = "background" | "all" | "other";
+export interface ChatGPTDispatchLifecycle {
+ captured(scope: AdmissionScope): void;
+ beforeDispatch(scope: AdmissionScope): void;
+ noDispatch(kind: NoDispatchKind): void;
+}
+interface CycleOptions {
+ initialMessages: AgentMessage[];
+ staged?: AgentMessage[];
+ foregroundOwed: boolean;
+ purpose: { background: boolean };
+ scopeCurrent: (scope: AdmissionScope) => boolean;
+ commitStaged?: () => void;
+ restoreAutomatic: (automaticIds: ReadonlySet<string>, retainedIds: ReadonlySet<string>) => void;
+}
+const idOf = (m: AgentMessage) => (m as { id?: string }).id;
+/** One logical turn, not an SDK loop. No credentials, persisted provenance or wire protocol. */
+export class ChatGPTRequestCycle implements ChatGPTDispatchLifecycle {
+ private release?: (event: AgentEvent) => void;
+ private held: AgentEvent[] = [];
+ private snapshot: AgentMessage[] = [];
+ private seen: Set<string>;
+ private automatic = new Set<string>();
+ private committed = false;
+ private denied?: NoDispatchKind;
+ private scope?: AdmissionScope;
+ private originalScope?: AdmissionScope;
+ private foregroundOwed: boolean;
+ private inheritedBackground: boolean;
+ private resumed = false;
+ private userAborted = false;
+ private stagedCommitted = false;
+ private started = false;
+ private anyDispatch = false;
+ private firstResolved = false;
+ private resolveAdmission!: (admitted: boolean) => void;
+ readonly admission = new Promise<boolean>((resolve) => { this.resolveAdmission = resolve; });
+ constructor(private readonly options: CycleOptions) {
+  this.seen = new Set(options.initialMessages.map(idOf).filter((id): id is string => !!id));
+  for (const message of options.staged ?? []) { const id = idOf(message); if (id) this.seen.delete(id); }
+  this.foregroundOwed = options.foregroundOwed;
+  this.inheritedBackground = options.purpose.background;
+ }
+ get dispatched(): boolean { return this.anyDispatch; }
+ get deniedBeforeDispatch(): boolean { return this.denied !== undefined; }
+ abort(): void { this.userAborted = true; }
+ bindRelease(release: (event: AgentEvent) => void): () => void { this.release = release; return () => { this.release = undefined; }; }
+ prepare(request: PrepareRequestContext): void {
+  const messages = request.context.messages as AgentMessage[];
+  this.snapshot = [...messages]; this.automatic = new Set();
+  let learner = false;
+  for (const m of messages) {
+   const id = idOf(m); if (!id || this.seen.has(id)) continue;
+   const cause = nativeInputCause(m, messages.slice(0, messages.indexOf(m)));
+   if (cause === "automatic") this.automatic.add(id);
+   if (cause === "learner") learner = true;
+  }
+  if (this.automatic.size > 0) this.options.purpose.background = true;
+  const retained = messages.filter((m) => !this.automatic.has(idOf(m) ?? ""));
+  this.foregroundOwed ||= learner || (!this.inheritedBackground && needsContinuation(retained as LhMessage[]));
+ }
+ hold(event: AgentEvent): boolean {
+  if (event.type === "agent_start") { if (this.started) return true; this.started = true; return false; }
+  if (event.type === "agent_end") return true; // settle only after native activeRun is cleared
+  if (event.type === "turn_start") {
+   this.committed = false; this.denied = undefined; this.scope = undefined;
+   // A retained learner batch from the denied cycle already owns its turn_start.
+   if (!this.held.some((e) => e.type === "turn_start")) this.held.push(event);
+   return true;
+  }
+  if (!this.committed) { this.held.push(event); return true; }
+  return false;
+ }
+ captured(scope: AdmissionScope): void {
+  if (this.originalScope && this.resumed && JSON.stringify(scope) !== JSON.stringify(this.originalScope)) throw new Error("ChatGPT recovery scope changed");
+  this.scope = { ...scope };
+  this.originalScope ??= { ...scope };
+ }
+ beforeDispatch(scope: AdmissionScope): void {
+  if (this.userAborted || this.denied || this.committed || !this.scope || JSON.stringify(scope) !== JSON.stringify(this.scope) || !this.options.scopeCurrent(scope)) throw new Error("ChatGPT admission cancelled");
+  if (!this.stagedCommitted && (this.options.staged?.length ?? 0) > 0) {
+   this.options.commitStaged?.(); // synchronous durable prefix, before visibility and fetch
+   this.stagedCommitted = true;
+   const stagedEvents: AgentEvent[] = this.options.staged!.flatMap((message) => [{ type: "message_start" as const, message }, { type: "message_end" as const, message }]);
+   const boundary = this.held.findIndex((event) => event.type === "turn_start");
+   this.held.splice(boundary < 0 ? 0 : boundary + 1, 0, ...stagedEvents);
+  }
+  this.committed = true; this.anyDispatch = true; this.foregroundOwed = false;
+  this.inheritedBackground = this.options.purpose.background;
+  for (const m of this.snapshot) { const id = idOf(m); if (id) this.seen.add(id); }
+  const events = this.held; this.held = [];
+  for (const event of events) this.release?.(event);
+  this.resolveFirst(true);
+ }
+ noDispatch(kind: NoDispatchKind): void { if (!this.committed) this.denied = this.userAborted ? "all" : kind; }
+ acceptsUsage(_message: AssistantMessage): boolean { return this.committed; }
+ private resolveFirst(value: boolean) { if (!this.firstResolved) { this.firstResolved = true; this.resolveAdmission(value); } }
+ finishNativeCycle(): { action: "settle" | "resumeForeground"; messages: AgentMessage[] } {
+  if (!this.denied) { this.resolveFirst(this.anyDispatch); return { action: "settle", messages: [] }; }
+  const messages = this.snapshot.filter((m) => !this.automatic.has(idOf(m) ?? ""));
+  const retainedIds = new Set(messages.map(idOf).filter((id): id is string => !!id));
+  this.options.restoreAutomatic(this.automatic, retainedIds);
+  this.resolveFirst(false); // queue deferral is installed before initial fire cargo returns
+  this.held = this.held.filter((e) => (e.type === "turn_start") || ((e.type === "message_start" || e.type === "message_end") && e.message.role !== "assistant" && !this.automatic.has(idOf(e.message) ?? "")));
+  const retry = this.denied === "background" && !this.userAborted && !this.resumed && this.foregroundOwed && !!this.scope && this.options.scopeCurrent(this.scope);
+  if (retry) {
+   this.resumed = true; this.options.purpose.background = false;
+   this.seen = retainedIds;
+   this.denied = undefined;
+   // Staged automatic input must not be resurrected on foreground recovery.
+   this.options.staged = undefined;
+   return { action: "resumeForeground", messages };
+  }
+  this.held = [];
+  return { action: "settle", messages };
+ }
+}

src/agent/chatgpt-stream.ts

@@ -0,0 +1,118 @@
+import type { Api, AssistantMessage, Model, SimpleStreamOptions, StreamFunction, TranscriptContext } from "@earendil-works/pi-ai";
+import { streamSimple } from "@earendil-works/pi-ai/api/openai-responses";
+import { AssistantMessageEventStream } from "@earendil-works/pi-ai/utils/event-stream";
+import { transformMessages } from "@earendil-works/pi-ai/api/transform-messages";
+import { chatGPTAbortController, chatGPTScope, chatGPTScopeCurrent, chatGPTLease, chatGPTDenial } from "./chatgpt-bridge.ts";
+import type { AdmissionScope, ChatGPTDispatchLifecycle } from "./chatgpt-request-cycle.ts";
+
+const USAGE = "https://chatgpt.com/settings/usage";
+export function chatGPTStreamError(message: string): string {
+	if (/subscription_sharing_usage_limit_exceeded|quota|429/.test(message)) return `ChatGPT plan usage limit reached. Manage usage at ${USAGE}. No API-key fallback was used.`;
+	if (/scope|authorization_context|permission/.test(message)) return "ChatGPT plan permission is missing. Reconnect in Providers and authorize plan use. No API-key fallback was used.";
+	if (/expired|invalid_grant|401|invalid_user/.test(message)) return "Your ChatGPT session expired. Reconnect in Providers. No API-key fallback was used.";
+	if (/background|Background/.test(message)) return "ChatGPT background use is off in this app. Enable Allow background use for all ChatGPT accounts in Providers to run this work. No API-key fallback was used.";
+	if (/eligible|403/.test(message)) return "This ChatGPT account or workspace does not permit this request. Check ChatGPT settings; reconnecting may not change eligibility.";
+	if (/cancel|abort|closed|disconnect/i.test(message)) return "ChatGPT request cancelled or connection closed. No API-key fallback was used.";
+	if (/catalog|model/.test(message)) return "The selected model is unavailable for this ChatGPT account. Refresh Providers and explicitly choose an available model. No API-key fallback was used.";
+	if (/incomplete|terminal|interrupted|stop reason/.test(message)) return "ChatGPT returned an incomplete response. Retry when ready; no API-key fallback was used.";
+	return "ChatGPT request failed. Check your connection and ChatGPT usage, then retry. No API-key fallback was used.";
+}
+
+/** Restrict the body at the final payload seam, even if caller samplingParams add fields. */
+export function chatGPTPayload(raw: unknown, selectedModelId: string): Record<string, unknown> {
+	const payload = raw as Record<string, unknown>;
+	const out: Record<string, unknown> = { model: selectedModelId, store: false, stream: true };
+	for (const key of ["instructions", "reasoning", "include", "text", "tool_choice", "parallel_tool_calls", "prompt_cache_key"]) {
+		if (payload[key] !== undefined) out[key] = payload[key];
+	}
+	if (!Array.isArray(payload.input)) throw new Error("ChatGPT requires stateless input history");
+	out.input = payload.input.map((item) => item?.role === "system" ? { ...item, role: "developer" } : item);
+	if (Array.isArray(payload.tools) && payload.tools.length) {
+		if (payload.tools.some((tool) => tool.type !== "function" && tool.type !== "custom")) throw new Error("Unsupported ChatGPT tool capability");
+		out.tools = [{ type: "namespace", name: "learnharness", description: "LearnHarness local learning tools", tools: payload.tools }];
+	}
+	return out;
+}
+
+/** Only ChatGPT history is translated. Mixed API-key history keeps its actual identity. */
+export function chatGPTHistory(context: TranscriptContext, logicalModel: Model<Api>): TranscriptContext {
+	// Apply Pi's normal foreign-model signature treatment BEFORE the alias
+	// becomes openai, or real API-key turns could look like subscription turns.
+	const messages = transformMessages(context.messages, logicalModel);
+	return { ...context, messages: messages.map((m) => m.role === "assistant" && m.provider === "chatgpt" ? { ...m, provider: "openai" } : m) };
+}
+function empty(model: Model<Api>): AssistantMessage {
+	return { role: "assistant", api: model.api, provider: "chatgpt", model: model.id, content: [], stopReason: "error", timestamp: Date.now(), usage: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, totalTokens: 0, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } } };
+}
+export function chatGPTStream(background: () => boolean, lifecycle?: ChatGPTDispatchLifecycle): StreamFunction<Api, SimpleStreamOptions> {
+	return (logicalModel, context, options) => {
+		const output = new AssistantMessageEventStream();
+		const purpose = background();
+		const { controller, release } = chatGPTAbortController(purpose);
+		const signal = options?.signal ? AbortSignal.any([options.signal, controller.signal]) : controller.signal;
+		void (async () => {
+			let partial = empty(logicalModel);
+			let completed = false;
+			let failure = "";
+			let scope: AdmissionScope | undefined;
+			let committed = false;
+			let terminal = false;
+			const check = () => { if (signal.aborted || !scope || !chatGPTScopeCurrent(scope)) throw new Error("cancelled or catalog authority changed"); };
+			try {
+				if (logicalModel.provider !== "chatgpt") throw new Error("ChatGPT account catalog has no selected model");
+				if (Object.keys(options?.headers ?? {}).some((key) => key.toLowerCase() === "authorization")) throw new Error("Authorization override is forbidden");
+				scope = chatGPTScope(logicalModel.id);
+				lifecycle?.captured(scope);
+				const credentials = await chatGPTLease(scope, purpose);
+				check();
+				// NEVER dispatch through Models under openai: that would expose ambient API-key fallback.
+				const transport: Model<"openai-responses"> = { ...logicalModel, api: "openai-responses", provider: "openai", baseUrl: "https://api.openai.com/v1", headers: undefined, samplingParams: undefined, compat: { ...logicalModel.compat, supportsDeveloperRole: true } };
+				const delegate = options?.fetch ?? globalThis.fetch;
+				const source = streamSimple(transport, chatGPTHistory(context, logicalModel), {
+					...options,
+					apiKey: credentials.accessToken,
+					signal,
+					maxRetries: 0,
+					headers: { "OpenAI-Beta": null, ...options?.headers },
+					fetch: (input, init) => {
+						check();
+						if (committed) throw new Error("ChatGPT does not retry admitted requests");
+						lifecycle?.beforeDispatch(scope!);
+						committed = true;
+						return delegate(input, init);
+					},
+					onPayload: async (raw, model) => {
+						const replacement = await options?.onPayload?.(raw, model);
+						const payload = chatGPTPayload(replacement === undefined ? raw : replacement, logicalModel.id);
+						check();
+						return payload;
+					},
+					onProviderStreamEvent: async (raw, model) => {
+						const event = raw as { type?: string; code?: string; error?: { code?: string }; response?: { error?: { code?: string }; status?: string } };
+						if (event.type === "response.completed" && event.response?.status === "completed" && !event.response?.error) completed = true;
+						if (["response.incomplete", "response.failed", "error"].includes(event.type ?? "") || event.response?.error) failure = event.code || event.error?.code || event.response?.error?.code || "incomplete terminal response";
+						await options?.onProviderStreamEvent?.(raw, model);
+					},
+				});
+				for await (const event of source) {
+					const message = "partial" in event ? event.partial : event.type === "done" ? event.message : event.error;
+					// Copy the message, not its tool/reasoning blocks: IDs/signatures remain byte-for-byte intact.
+					partial = { ...message, provider: "chatgpt" };
+					if (event.type === "done") {
+						if (!completed || failure) throw new Error(failure || "missing terminal response.completed");
+						terminal = true;
+						output.push({ ...event, message: partial });
+					} else if (event.type === "error") {
+						throw new Error(failure || event.error.errorMessage || "ChatGPT stream error");
+					} else output.push({ ...event, partial });
+				}
+			} catch (err) {
+				if (terminal) return;
+				if (!committed) lifecycle?.noDispatch(await chatGPTDenial(scope, controller, options?.signal?.aborted === true, purpose));
+				partial = { ...partial, stopReason: signal.aborted ? "aborted" : "error", errorMessage: chatGPTStreamError(err instanceof Error ? err.message : "") };
+				output.push({ type: "error", reason: partial.stopReason as "error" | "aborted", error: partial });
+			} finally { release(); output.end(); }
+		})();
+		return output;
+	};
+}

src/agent/models.ts

@@ -12,6 +12,9 @@ import type { ThinkingLevel } from "@earendil-works/pi-agent-core";
 import { type ModelRole } from "../../config.ts";
 import { DEFAULT_THINKING_ORDER } from "../core/settings.ts";
 import packageMetadata from "../../package.json" with { type: "json" };
+import { chatGPTModel } from "./chatgpt-bridge.ts";
+import { chatGPTStream } from "./chatgpt-stream.ts";
+import type { ChatGPTDispatchLifecycle } from "./chatgpt-request-cycle.ts";
 import { createClaudeCodeProvider } from "./claude-code/provider.ts";
 import { streamClaudeCode } from "./claude-code/stream.ts";
 
@@ -96,9 +99,9 @@ function defaultThinkingLevel(model: Model<Api>): ThinkingLevel {
  *  key is resolved by pi-ai itself from the environment per request (the
  *  keys seam reflects file-saved keys into env), so the engine never
  *  hardcodes OpenRouter's. */
-export function piEngineForSpec(spec: ResolvedModelSpec): PiEngine {
+export function piEngineForSpec(spec: ResolvedModelSpec, background: boolean | (() => boolean) = true, lifecycle?: ChatGPTDispatchLifecycle): PiEngine {
 	const models: Models = piModels();
-	const model = models.getModel(spec.provider, spec.id);
+	const model = spec.provider === "chatgpt" ? chatGPTModel(spec.id) : models.getModel(spec.provider, spec.id);
 	if (!model) {
 		throw new Error(`${spec.provider} catalog has no model ${spec.id}`);
 	}
@@ -114,11 +117,10 @@ export function piEngineForSpec(spec: ResolvedModelSpec): PiEngine {
 		// bound streamSimple: the agent loop forwards state.thinkingLevel as
 		// `reasoning` (undefined = off on pi-ai's scale); the clamp above keeps
 		// the pin honest even when config or settings drift off the map.
-		streamFn: (requestModel, context, options) =>
-			models.streamSimple(requestModel, context, {
-				...options,
-				reasoning: level === "off" ? undefined : level,
-			}),
+		streamFn: (requestModel, context, options) => {
+			const stream = spec.provider === "chatgpt" ? chatGPTStream(typeof background === "function" ? background : () => background, lifecycle) : models.streamSimple.bind(models);
+			return stream(requestModel, context, { ...options, reasoning: level === "off" ? undefined : level });
+		},
 	};
 }
 

src/agent/subagents/extraction.ts

@@ -777,6 +777,7 @@ export function callUsage(usage: Usage): CallUsage {
 		outputTokens: usage.output,
 		cacheReadTokens: usage.cacheRead,
 		cacheWriteTokens: usage.cacheWrite,
+		cacheWrite1hTokens: usage.cacheWrite1h ?? 0,
 		// Optional in pi's Usage (providers that never emit reasoning
 		// tokens leave it unset) — persisted as 0, never dropped.
 		reasoningTokens: usage.reasoning ?? 0,

src/agent/subagents/headless.ts

@@ -137,6 +137,9 @@ export function browserServiceSpawnEnv(
 	const out: Record<string, string | undefined> = { ...env };
 	delete out.ELECTRON_RUN_AS_NODE;
 	delete out.LH_DEV_URL;
+	// The artifact renderer is not an authorized app API client.
+	delete out.LH_DESKTOP_API_SECRET;
+	delete out.LH_DESKTOP_SERVER;
 	return out;
 }
 

src/agent/subagents/researcher.ts

@@ -49,6 +49,8 @@ Tool outputs may be error strings (e.g. search unconfigured, a page that failed
 Adapt: try another URL or another search angle; if research is genuinely impossible, say so
 in the brief instead of inventing facts.
 
+If web_search reports that search is unavailable because credentials are missing, do not retry it with different queries. Use web_fetch when you have source URLs, and continue with any other available sources. If the missing search capability materially limits your work, state the limitation and the missing-key setup needed in your brief. Never invent search results or sources.
+
 Your FINAL assistant message is your entire deliverable — it must stand alone, using this format:
 
 ## Summary
@@ -70,8 +72,7 @@ export interface ResearcherRunOptions {
 	/** The search seam (search port wave, 2026-09-20): resolves the LIVE
 	 *  provider per call, null = honestly unavailable (the web_search tool
 	 *  reports that in its output — no crash, no silent fallback). Absent
-	 *  seam reads as unavailable: the server wiring always injects one; the
-	 *  dispatch-level drop (jobTools) keeps doomed jobs from even starting. */
+	 *  seam reads as unavailable; web_fetch remains usable without it. */
 	search?: () => SearchProvider | null;
 	/** The researcher's PDF dir (PDF wave, 2026-10-07): when present, the
 	 *  toolset gains web_fetch_pdf + pdf_info/pdf_extract/pdf_search scoped

src/agent/subagents/runner.ts

@@ -82,9 +82,8 @@ export interface JobRunnerDeps {
 	/** The search seam (search port wave, 2026-09-20): resolves the LIVE
 	 *  provider fresh per job — server/search.ts composition injected by
 	 *  the wiring; null = honestly unavailable (the researcher's web_search
-	 *  reports it). Absent seam (tests) reads the same. The dispatch-level
-	 *  drop (jobTools' searchAvailable) uses the same signal — jobs already
-	 *  in flight when a key disappears still fail honestly, never crash. */
+	 *  reports it as recoverable text). Absent seam (tests) reads the same;
+	 *  research dispatch and web_fetch do not require search credentials. */
 	searchProvider?: () => SearchProvider | null;
 	/** Key-presence gate (provider-key interlude, 2026-09-25): given the
 	 *  resolved engine's provider + the kind's slot tier, return the

src/agent/teacher-agent.ts

@@ -18,7 +18,8 @@ import { practiceTools, type PracticeToolContext } from "./tools/practice.ts";
 import { healToolRecord } from "./tools/heal.ts";
 import { quizInputObject, quizInputProblems } from "../core/question-core.ts";
 import { QUESTION_TOOLS, type EnabledTools } from "../core/pending.ts";
-import { toLlmMessages } from "../core/native-messages.ts";
+import { toLlmMessages, nativeInputCause } from "../core/native-messages.ts";
+import type { ChatGPTRequestCycle } from "./chatgpt-request-cycle.ts";
 
 /**
  * The two teaching-side agents (docs/probe-teach-split.md): the PROBER and
@@ -97,6 +98,8 @@ export type StepUsageCollector = (usage: Usage) => void;
 export interface AgentBuildOptions {
 	/** Model + stream function — tests inject a scripted faux engine. */
 	engine?: PiEngine;
+	callPurpose?: { background: boolean };
+	requestCycle?: ChatGPTRequestCycle;
 	/** The initial transcript, as pi messages (the record IS agent state). */
 	messages?: AgentMessage[];
 	/** Money channel: one call per model request. */
@@ -178,6 +181,7 @@ function baseAgentOptions(kind: AgentKind, opts: AgentBuildOptions, tools: Agent
 	// LH_PROVIDER_EVENTS=1 and the caller handed a chatId; undefined is
 	// the normal no-observer path (tests, env-off).
 	const providerEvents = opts.chatId ? providerEventObserver(opts.chatId, kind) : undefined;
+	const seenInputs = new Set((opts.messages ?? []).map((m) => (m as { id?: string }).id).filter(Boolean));
 	const agent = new Agent({
 		identity: APP_IDENTITY,
 		sessionId: opts.chatId,
@@ -211,13 +215,31 @@ function baseAgentOptions(kind: AgentKind, opts: AgentBuildOptions, tools: Agent
 		// an abort's dangling question call gets its INTERRUPTED result. It
 		// must never throw (pi's convertToLlm contract).
 		convertToLlm: toLlmMessages,
+		prepareRequest: (request) => {
+			if (opts.callPurpose) {
+				let learner = false, automatic = false;
+				const messages = request.context.messages;
+				for (const message of messages) {
+					const id = (message as { id?: string }).id;
+					const cause = nativeInputCause(message, messages.slice(0, messages.indexOf(message)));
+					if (cause === "neutral") continue;
+					if (id && seenInputs.has(id)) continue;
+					if (id) seenInputs.add(id);
+					if (cause === "learner" && id) learner = true;
+					else automatic = true; // missing/unrecognized identity never grants foreground
+				}
+				if (automatic) opts.callPurpose.background = true;
+				else if (learner) opts.callPurpose.background = false;
+			}
+			opts.requestCycle?.prepare(request);
+		},
 	});
 	if (opts.onUsage) {
 		agent.subscribe((event: AgentEvent) => {
 			if (event.type === "message_end" && event.message.role === "assistant" && "usage" in event.message) {
 				// A cancelled call is not a billed step (partial/estimated
 				// usage on an aborted message is not money spent).
-				if (event.message.stopReason === "aborted") return;
+				if (event.message.stopReason === "aborted" || (opts.requestCycle && !opts.requestCycle.acceptsUsage(event.message))) return;
 				opts.onUsage!(event.message.usage);
 			}
 		});

src/agent/tools.ts

@@ -268,8 +268,7 @@ export interface ProberToolContext extends ProbeDoneContext {
 	jobStore: JobStore;
 	/** Fire the background runner for a freshly created job. */
 	spawn: (job: JobRecord, input: ResearchInput | DemoInput | VisualInput | PdfInput) => void;
-	/** Search availability — the research dispatch tool's honest-disable
-	 *  signal (same gate as the teacher's). Absent = available. */
+	/** Retained search-status seam; research dispatch is available keylessly. */
 	searchAvailable?: () => boolean;
 	/** The prober's graph-map reader (wave 7): present = the two map tools
 	 *  (list_sessions / list_propositions) are registered. The composition
@@ -331,13 +330,12 @@ export function createProberTools(
 				// research + pdf are ALWAYS available in every phase — the
 				// prober can verify facts or read learner PDFs while assessing.
 				// demo/visual stay teach-phase-only (they need the lesson's plan
-				// context). Same gating as the teacher: toggle-filtered here,
-				// research subject to the searchAvailable gate (honestly absent
-				// when no engine is live).
+				// context). Same user-toggle filtering as the teacher; research
+				// remains available without a search-provider key.
 				...(jobs.research ? { research: jobs.research } : {}),
 				pdf: jobs.pdf!,
 				// The map reader (wave 7): present only when the settings gate
-				// opened — same registration-time drop as the search gate.
+				// opened.
 				...(ctx.graphStore
 					? proberGraphTools({ chatStore: ctx.chatStore, graphStore: ctx.graphStore })
 					: {}),
@@ -400,9 +398,7 @@ export interface TeacherToolContext extends PlanToolContext {
 	 *  rides along — the spawn fires mid-stream, before the tool part is
 	 *  persisted to the transcript (a transcript scan would race it). */
 	spawn: (job: JobRecord, input: ResearchInput | DemoInput | VisualInput | PdfInput) => void;
-	/** Search availability (search port wave, 2026-09-20): the research
-	 *  dispatch tool's honest-disable signal — no live engine = the tool
-	 *  drops from this registry (JobToolContext.searchAvailable). */
+	/** Retained search-status seam; research dispatch is available keylessly. */
 	searchAvailable?: () => boolean;
 }
 
@@ -462,10 +458,6 @@ export function createTeacherTools(
 			find: findTool,
 			grep: grepTool,
 			...planTools(ctx),
-			// Fix wave (2026-09-20): searchAvailable was RECEIVED here but
-			// DROPPED when building the inner JobToolContext — the drop-gate
-			// always took its ?? true default. Threaded now: the live signal
-			// reaches jobTools' research gate.
 			...jobTools({ chatId: ctx.chatId, jobStore: ctx.jobStore, spawn: ctx.spawn, searchAvailable: ctx.searchAvailable }),
 		},
 		withoutMechanics(enabled),

src/agent/tools/jobs.ts

@@ -23,11 +23,9 @@ export interface JobToolContext {
 	 *  input rides along because the spawn happens MID-STREAM, before
 	 *  onFinish persists the tool part — a transcript scan would race it. */
 	spawn: (job: JobRecord, input: ResearchInput | DemoInput | VisualInput | PdfInput) => void;
-	/** Search availability (search port wave, 2026-09-20): the honest
-	 *  disable signal for the research dispatch tool — no live engine
-	 *  means NO research tool in the registry (owner decision: registry-
-	 *  level drop, no doomed dispatch, no silent fallback). Absent =
-	 *  available (tests + any seam without search wiring). */
+	/** Search-provider availability, retained for existing composition seams.
+	 *  Research dispatch does not depend on it: web_fetch is keyless and
+	 *  web_search reports a missing provider as a recoverable tool result. */
 	searchAvailable?: () => boolean;
 }
 
@@ -127,25 +125,20 @@ const PDF_DESCRIPTION =
  * context (chatId + jobStore + the spawn seam), so concurrent chats never
  * share state. chat-handler wires it each turn; tests wire their own.
  *
- * The research tool drops from the registry when ctx.searchAvailable reports
- * no live engine (2026-09-20) — the teacher never sees it, so a keyless
- * server can't dispatch doomed research jobs. Jobs ALREADY in flight when
- * a key disappears still run to their honest failure (runner backstop).
+ * Research remains available without a search-provider key; the role's
+ * existing user-toggle filtering still applies outside this registry.
  */
 export function jobTools(ctx: JobToolContext): Record<string, AgentTool<any>> {
-	const searchOn = ctx.searchAvailable?.() ?? true;
 	return {
-		...(searchOn && {
-			research: {
-				name: "research",
-				label: "research",
-				description: RESEARCH_DESCRIPTION,
-				parameters: researchInputSchema,
-				execute: serverExecute("research", researchInputSchema, (input, { toolCallId }) =>
-					dispatch(ctx, "research", toolCallId, RESEARCH_ACK, input),
-				),
-			},
-		}),
+		research: {
+			name: "research",
+			label: "research",
+			description: RESEARCH_DESCRIPTION,
+			parameters: researchInputSchema,
+			execute: serverExecute("research", researchInputSchema, (input, { toolCallId }) =>
+				dispatch(ctx, "research", toolCallId, RESEARCH_ACK, input),
+			),
+		},
 		demo: {
 			name: "demo",
 			label: "demo",

src/agent/tools/practice.ts

@@ -40,8 +40,8 @@ export interface PracticeToolContext {
 	graphNodes?: () => ReadonlyArray<Pick<import("../../core/contracts/graph.ts").GraphNode, "id" | "name">>;
 	/** THE RARELY-NEEDED SET (owner round 10, 2026-10-05): the map readers
 	 *  behind the SAME settings gate as the prober's (graphToolsEnabled at
-	 *  the composition root); the research + pdf jobs with the honest
-	 *  searchAvailable drop. Each absent (tests) = absent from the registry. */
+	 *  the composition root); the research + pdf jobs (both keyless).
+	 *  Each absent context (tests) = absent from the registry. */
 	graph?: GraphToolContext;
 	jobs?: JobToolContext;
 }
@@ -225,8 +225,8 @@ export function practiceTools(ctx: PracticeToolContext): Record<string, AgentToo
 	const gate = createPracticeGate(ctx);
 	// THE RARELY-NEEDED SET (owner round 10): read/ls/find/grep over the
 	// learner's files (always — the same read-only set the teacher and
-	// prober carry), the research + pdf jobs (research honestly absent
-	// when no search engine is live; demo/visual stay teach-only), and
+	// prober carry), the research + pdf jobs (available without search
+	// keys; demo/visual stay teach-only), and
 	// the prober's map readers when the settings gate opened. None are
 	// toggle-filtered: problem/conclude are the session; these are
 	// exceptions the doctrine says to reach for rarely.

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/chatgpt.ts

@@ -0,0 +1,22 @@
+/** Client-safe status only. Credentials and OAuth transaction data never enter this contract. */
+export interface ChatGPTStatus {
+	available: boolean;
+	pending: boolean;
+	/** App-wide ChatGPT background permission, independent of account selection. */
+	allowBackgroundUse: boolean;
+	activeId: string | null;
+	accounts: { id: string; email: string; registration: string; connected: boolean; planEnabled?: boolean }[];
+	error: string | null;
+	catalogError?: string | null;
+}
+export interface ChatGPTControlBridge {
+	status(): Promise<ChatGPTStatus>;
+	login(id?: string): Promise<ChatGPTStatus>;
+	cancel(): Promise<ChatGPTStatus>;
+	select(id: string): Promise<ChatGPTStatus>;
+	/** Explicitly forget every local registration/credential for this verified identity. */
+	disconnect(id: string): Promise<ChatGPTStatus>;
+	logout(id: string): Promise<ChatGPTStatus>;
+	consent(enabled: boolean): Promise<ChatGPTStatus>;
+	onChange(callback: () => void): () => void;
+}

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.
  *
@@ -18,13 +18,15 @@
  */
 
 import type { CallUsage, EmissionStats, ExtractionRunUsage } from "./contracts.ts";
-import { callCost, sumSteps } from "./prices.ts";
+import { sumSteps, isChatGPTPlan, type CallPricer } from "./prices.ts";
 
 /** One pipeline stage's cumulative cost across every run. */
 export interface StageCost {
+	/** ChatGPT plan stage-runs, excluded from API dollars and unpriced counts. */
+	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;
@@ -34,6 +36,7 @@ export interface StageCost {
 
 /** One run in the recent-runs list (extraction or naming). */
 export interface RecentRun {
+	subscriptionRuns?: number;
 	startedAt: number;
 	/** ISO wall-clock. */
 	at: string;
@@ -51,9 +54,10 @@ export interface RecentRun {
 
 /** The cost pill's whole payload: stage totals + recent runs. */
 export interface GraphCostSummary {
+	subscriptionRuns?: number;
 	/** Inferred USD across all priced stages of all runs. */
 	totalDollars: number;
-	/** Runs where EVERY stage was unpriced (totalDollars excludes them). */
+	/** Runs with no priced API stage and at least one unknown stage. */
 	unpricedRuns: number;
 	/** The billable pipeline stages. (Wave 7, owner ruling 2026-09-23:
 	 *  the embedding stage is gone with the provider embedder; the
@@ -88,10 +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 {
+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;
 	}
@@ -101,6 +106,9 @@ function stageDollars(model: string, steps: readonly CallUsage[]): number | null
 /** Merge one step-set's tokens into an accumulating CallUsage. */
 function mergeTokens(into: CallUsage, steps: readonly CallUsage[]): void {
 	const summed = sumSteps(steps);
+	// Fold this before updating total writes, so an older write-bearing
+	// aggregate with a missing split remains unknown even after fresh data.
+	into.cacheWrite1hTokens = sumSteps([into, summed]).cacheWrite1hTokens;
 	into.inputTokens = (into.inputTokens ?? 0) + (summed.inputTokens ?? 0);
 	into.outputTokens = (into.outputTokens ?? 0) + (summed.outputTokens ?? 0);
 	into.cacheReadTokens = (into.cacheReadTokens ?? 0) + (summed.cacheReadTokens ?? 0);
@@ -110,7 +118,7 @@ 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[] = [];
@@ -120,14 +128,16 @@ 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);
-		if (extDollars === null) extraction.unpricedRuns++;
+		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);
-		if (audDollars === null) audit.unpricedRuns++;
+		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;
 
 		// Run-level: null only when EVERY stage is unpriced.
@@ -135,8 +145,10 @@ export function graphCostSummary(input: CostSummaryInput): GraphCostSummary {
 		const runDollars = allStages.every((d) => d === null)
 			? null
 			: allStages.reduce<number>((a, d) => a + (d ?? 0), 0);
+		const subscriptionRuns = [stages.extractor, stages.auditor].filter((stage) => isChatGPTPlan(stage.model)).length;
 		const allSteps = [...stages.extractor.steps, ...stages.auditor.steps];
 		runs.push({
+			...(subscriptionRuns ? { subscriptionRuns } : {}),
 			startedAt: run.startedAt,
 			at: run.at,
 			durationMs: run.durationMs,
@@ -159,8 +171,10 @@ export function graphCostSummary(input: CostSummaryInput): GraphCostSummary {
 	const capped = runs.slice(0, input.recentCap ?? 10);
 
 	// A run is "unpriced" when it billed nothing countable at all.
-	const unpricedRuns = runs.filter((r) => r.dollars === null).length;
+	const unpricedRuns = runs.filter((r) => r.dollars === null && (r.subscriptionRuns ?? 0) < 2).length;
+	const subscriptionRuns = (extraction.subscriptionRuns ?? 0) + (audit.subscriptionRuns ?? 0);
 	return {
+		...(subscriptionRuns ? { subscriptionRuns } : {}),
 		totalDollars: extraction.dollars + audit.dollars,
 		unpricedRuns,
 		stages: { extraction, audit },

src/core/index.ts

@@ -48,6 +48,8 @@ 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";
 // settings.ts + settings-resolution.ts (settings-storage wave, 2026-09-19):

src/core/native-messages.ts

@@ -674,6 +674,57 @@ export function endsInDismissal(messages: readonly LhMessage[]): boolean {
 	return last !== undefined && isDismissedResult(last);
 }
 
+export type RequestCause = "foreground" | "background";
+export type ReconstructedCause = RequestCause | "ambiguous";
+export type NativeInputCause = "learner" | "automatic" | "neutral" | "unrecognizable";
+/** Native causes only: flattened user roles, IDs' prefixes and prose confer no permission. */
+export function nativeInputCause(message: AgentMessage, history: readonly AgentMessage[]): NativeInputCause {
+	if (message.role === "user") return "learner";
+	if (message.role === "system" || message.role === "assistant") return "neutral";
+	if (message.role === "custom") {
+		if (["probe-handoff", "subagent-steer", "subagent-abort"].includes(message.customType)) return "automatic";
+		if (["practice-thinking", "practice-next"].includes(message.customType)) return "learner";
+		return ["graph-context", "plan-block", "problem-dismissed", "compaction"].includes(message.customType) ? "neutral" : "unrecognizable";
+	}
+	if (message.role === "toolResult") {
+		if (!(QUESTION_TOOLS as readonly string[]).includes(message.toolName) || message.isError || message.details == null || (message.details as { dismissed?: boolean }).dismissed) return "neutral";
+		const details = message.details as Partial<QuestionAnswerData>;
+		if (!["quiz", "ask", "explain"].includes(details.kind ?? "") || typeof details.questionText !== "string" || typeof details.dontKnow !== "boolean" || !Array.isArray(details.answers) || typeof details.answeredAt !== "string") return "unrecognizable";
+		const calls = history.flatMap((m) => m.role === "assistant" ? m.content.filter((b) => b.type === "toolCall" && b.id === message.toolCallId) : []);
+		return calls.length === 1 && calls[0]?.type === "toolCall" && calls[0].name === message.toolName ? "learner" : "unrecognizable";
+	}
+	return "unrecognizable";
+}
+/** Full persisted native record, including pre-compaction causes. Never guessed foreground. */
+export function reconstructRequestCause(messages: readonly LhMessage[]): ReconstructedCause {
+	let inherited: ReconstructedCause = "ambiguous";
+	let triggers: NativeInputCause[] = [];
+	const seen = new Map<string, string>();
+	const history: AgentMessage[] = [];
+	const resolve = (): ReconstructedCause => {
+		if (triggers.includes("unrecognizable")) return "ambiguous";
+		const lastLearner = triggers.lastIndexOf("learner"), lastAutomatic = triggers.lastIndexOf("automatic");
+		if (lastLearner < 0 && lastAutomatic < 0) return inherited;
+		if (lastAutomatic < 0) return "foreground";
+		if (lastLearner < lastAutomatic) return "background";
+		return "ambiguous";
+	};
+	for (const message of messages) {
+		const cause = nativeInputCause(message, history);
+		const id = (message as { id?: string }).id;
+		if (id) {
+			const value = JSON.stringify(message);
+			if (seen.has(id)) { if (seen.get(id) !== value) triggers.push("unrecognizable"); continue; }
+			seen.set(id, value);
+		} else if (cause !== "neutral") triggers.push("unrecognizable");
+		if (message.role === "assistant" && !(message.provider === "synthetic" && message.api === "faux" && message.model === "synthetic-conclude" && message.content.some((b) => b.type === "toolCall" && b.name === "conclude" && b.arguments?.outcome === "dismissed"))) {
+			inherited = resolve(); triggers = [];
+		} else if (cause !== "neutral") triggers.push(cause);
+		history.push(message);
+	}
+	return resolve();
+}
+
 export function needsContinuation(messages: readonly LhMessage[]): boolean {
 	const last = messages[messages.length - 1];
 	if (!last || (last as { role?: string }).role !== "toolResult") return false;

src/server/chat-handler.ts

@@ -7,6 +7,10 @@ import {
 	type TeachingAgent,
 } from "../agent/teacher-agent.ts";
 import { piEngineForSpec, type PiEngine } from "../agent/models.ts";
+import { chatGPTBackgroundAllowed, chatGPTScopeCurrent } from "../agent/chatgpt-bridge.ts";
+import { chatGPTStreamError } from "../agent/chatgpt-stream.ts";
+import { ChatGPTRequestCycle, type ChatGPTDispatchLifecycle } from "../agent/chatgpt-request-cycle.ts";
+import { reconstructRequestCause } from "../core/native-messages.ts";
 import { createJobRunner } from "../agent/subagents/runner.ts";
 import { createExtractionPipeline } from "../agent/subagents/extraction.ts";
 import type { ModelRole } from "../../config.ts";
@@ -15,6 +19,7 @@ import {
 	noteSteerVisible,
 	reconcileUserSteers,
 	registerLiveAgent,
+	restoreUncommittedAutomatic,
 	setJobRespawn,
 	setTurnRunner,
 	settleSteers,
@@ -41,7 +46,7 @@ import { persistTurn, triggerExtraction } from "./turn-service.ts";
 import { hasProbeDone, danglingQuestionOf } from "../core/native-messages.ts";
 import { composeInstructions } from "../agent/teacher-agent.ts";
 import { ensureCompacted, kickCompaction, effectiveThresholdTokens } from "./compaction.ts";
-import { catalogContextLimits } from "./model-catalog.ts";
+import { catalogContextLimits, refreshConnectionCatalog } from "./model-catalog.ts";
 import { emit } from "../core/events.ts";
 import { settledTurnResponse } from "./turn-runs.ts";
 import { asAgentMessages } from "../core/native-messages.ts";
@@ -201,6 +206,12 @@ export async function handleChatInput(
 	// persist (nothing durable changes); wake-pipeline turns ride the same
 	// path. The affected tier is strong — the conversation chain's own tier
 	// (a chat.model override still resolves through it).
+	if (conversation.provider === "chatgpt") {
+		await refreshConnectionCatalog();
+		if (input.type === "steers" && !chatGPTBackgroundAllowed()) {
+			return Response.json({ error: "ChatGPT background use is off in this app. Automatic follow-up remains deferred in its queue." }, { status: 403 });
+		}
+	}
 	if (!providerKeyPresent(conversation.provider)) {
 		const envelope = providerKeyMissingEnvelope(conversation.provider, ["strong"]);
 		console.info(`[chat ${chatId}] turn refused — ${JSON.stringify(envelope)}`);
@@ -216,12 +227,22 @@ export async function handleChatInput(
 	// placement mirrors the BYOK/key gates — BEFORE appends and
 	// beginTurn, no liveness effects — and the record re-reads after
 	// because compaction appends the marker.
-	const conversationEngine = piEngineForSpec(conversation);
+	const callPurpose = { background: input.type === "steers" || (input.type === "continue" && reconstructRequestCause(record.messages) !== "foreground") };
+	let requestCycle: ChatGPTRequestCycle | undefined;
+	const lifecycle: ChatGPTDispatchLifecycle = {
+		captured: (scope) => requestCycle?.captured(scope),
+		beforeDispatch: (scope) => requestCycle?.beforeDispatch(scope),
+		noDispatch: (kind) => requestCycle?.noDispatch(kind),
+	};
+	let conversationEngine: PiEngine;
+	try { conversationEngine = piEngineForSpec(conversation, () => callPurpose.background, lifecycle); }
+	catch { return Response.json({ error: conversation.provider === "chatgpt" ? chatGPTStreamError("model catalog") : "The selected model is unavailable." }, { status: 503 }); }
 	const modelIdForCompaction = `${conversationEngine.model.provider}:${conversationEngine.model.id}`;
 	await ensureCompacted(
 		{
 			chatStore: stores.chatStore,
-			engine: conversationEngine,
+			engine: piEngineForSpec(conversation, true),
+			canStart: () => conversation.provider !== "chatgpt" || chatGPTBackgroundAllowed(),
 			modelId: modelIdForCompaction,
 			kind: record.kind === "practice" ? "practice" : hasProbeDone(record.messages) ? "teach" : "probe",
 			thresholdTokens: effectiveThresholdTokens(
@@ -287,12 +308,14 @@ export async function handleChatInput(
 	// ride this same write-ahead - the pump de-queued them at fire, so this
 	// is the ONLY copy ever appended.)
 	const merged = [...record.messages, ...appends];
+	const stagedAutomatic = conversation.provider === "chatgpt" && input.type === "steers";
+	let persistBaseCount = stagedAutomatic ? record.messages.length : merged.length;
 	try {
-		stores.chatStore.saveMessages(chatId, merged);
+		if (!stagedAutomatic) stores.chatStore.saveMessages(chatId, merged);
 		// Dock fix (owner ruling 2026-09-26): persisted markers are visible
 		// in the record — flip their jobs now (the wake's steers ride this
 		// write-ahead; the delivering turn may run for minutes).
-		for (const appended of appends) {
+		for (const appended of stagedAutomatic ? [] : appends) {
 			if ((appended as { role?: string }).role === "custom") noteSteerVisible(appended as SteerMessage);
 		}
 	} catch (err) {
@@ -310,6 +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,
 		});
@@ -339,8 +363,19 @@ export async function handleChatInput(
 	// enabled). The definitive toggleable set is TOGGLEABLE_TOOLS — the
 	// sub-agent JOB tools (research/demo/visual/pdf; adjudicated with the
 	// user: quiz/ask/fs/plan/probe_done are always-on).
+	if (conversation.provider === "chatgpt") requestCycle = new ChatGPTRequestCycle({
+		initialMessages: asAgentMessages(merged),
+		staged: stagedAutomatic ? asAgentMessages(appends) : undefined,
+		foregroundOwed: !callPurpose.background,
+		purpose: callPurpose,
+		scopeCurrent: chatGPTScopeCurrent,
+		commitStaged: () => { stores.chatStore.saveMessages(chatId, merged); persistBaseCount = merged.length; },
+		restoreAutomatic: (automaticIds, retainedIds) => restoreUncommittedAutomatic(chatId, constructed, automaticIds, retainedIds),
+	});
 	const agentOpts: AgentBuildOptions = {
 		engine,
+		callPurpose,
+		requestCycle,
 		messages: asAgentMessages(merged),
 		onUsage: collectUsage,
 		toolToggles: effective.agent.tools,
@@ -439,6 +474,7 @@ export async function handleChatInput(
 		providerEventNote(chatId, "turn", `run start (base: ${merged.length} messages)`);
 		const handle = runPiTurn({
 			agent: constructed,
+			requestCycle,
 			// The coalescing window seam (COMPACTION LAW, 2026-10-09): tests
 			// tighten it via LH_UPDATE_FLUSH_MS; production rides the 100ms
 			// default in ui-stream.
@@ -454,13 +490,14 @@ export async function handleChatInput(
 			// route through here: they ride the error-result path and reach the
 			// transcript directly (the retry context the model saw).
 			onError: (error) => {
+				if (conversation.provider === "chatgpt") return chatGPTStreamError(String(error));
 				console.error(`[chat ${chatId}] turn error:`, error);
 				return "An error occurred.";
 			},
 			onFinish: (turnMessages, isAborted) => {
 				// Wave B: the record write is the agent's own final state —
 				// the new messages are the slice past the merged base.
-				const stamped = persistTurn(stores.chatStore, chatId, merged.length, turnMessages, isAborted);
+				const stamped = persistTurn(stores.chatStore, chatId, persistBaseCount, turnMessages, isAborted);
 				// usage log append (idempotent per turn id) — the stats bar
 				// refolds this on the client after the turn settles. The
 				// provenance message is the turn's FIRST new assistant message.
@@ -479,7 +516,7 @@ export async function handleChatInput(
 				// PRACTICE (owner, 2026-10-04): practice sessions write nothing —
 				// no extraction, no graph, no coverage (v1: the session itself
 				// is the deliverable).
-				if (record.kind !== "practice" && extractionEnabled(effective.graph)) {
+				if ((!requestCycle || requestCycle.dispatched) && record.kind !== "practice" && extractionEnabled(effective.graph)) {
 					// Fresh per-turn read of the graph section's chunk budget
 					// (extraction-UX redesign, 2026-09-24): the trigger compares
 					// against it and the pipeline's batches share it.
@@ -520,18 +557,20 @@ export async function handleChatInput(
 				// beneath-threshold record is a silent no-op. Skipped under
 				// tests (the extraction precedent — scripted chats are
 				// fixtures; the service itself is pinned directly).
-				if (process.env.NODE_ENV !== "test")
+				if (process.env.NODE_ENV !== "test" && (!requestCycle || requestCycle.dispatched))
 					kickCompaction(() => {
 						const fresh = effectiveSettings();
 						const spec = resolveConversationSpec(stores.chatStore.getChat(chatId)?.settings, fresh.models.strong);
-						if (!spec) return null;
-						const kickEngine = piEngineForSpec(spec);
+						if (!spec || (spec.provider === "chatgpt" && !chatGPTBackgroundAllowed())) return null;
+						let kickEngine: PiEngine;
+						try { kickEngine = piEngineForSpec(spec, true); } catch { return null; }
 						const kickModelId = `${kickEngine.model.provider}:${kickEngine.model.id}`;
 						const freshRecord = stores.chatStore.getChat(chatId);
 						if (!freshRecord) return null;
 						return {
 							chatStore: stores.chatStore,
 							engine: kickEngine,
+							canStart: () => spec.provider !== "chatgpt" || chatGPTBackgroundAllowed(),
 							modelId: kickModelId,
 							kind: freshRecord.kind === "practice" ? "practice" : hasProbeDone(freshRecord.messages) ? "teach" : "probe",
 							thresholdTokens: effectiveThresholdTokens(
@@ -558,6 +597,7 @@ export async function handleChatInput(
 		// must cost zero permanently-held slots beyond the one live view. The
 		// turn itself is a server-side process (ruling 2026-09-25) — it never
 		// needed this connection, only a view did.
+		if (stagedAutomatic && requestCycle && !(await requestCycle.admission)) return Response.json({ error: "Automatic follow-up remains deferred in its queue." }, { status: 403 });
 		return Response.json({ ok: true });
 	} catch (err) {
 		// The stream never started, so no onFinish will fire — clear the registry

src/server/compaction.ts

@@ -89,6 +89,8 @@ async function subagentSummarizer(prompt: string, engine: PiEngine): Promise<{ s
 /** Everything the service needs from its host (chat-handler wires the
  *  production versions; tests inject). */
 export interface CompactionDeps {
+	/** Eligibility only for NEW runs; structural waits still apply. */
+	canStart?: () => boolean;
 	chatStore: ChatStore;
 	/** The chat's conversation engine — also the summarizer's. */
 	engine: PiEngine;
@@ -173,7 +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).
@@ -238,6 +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,
 				},
 			],
@@ -267,6 +270,7 @@ export async function ensureCompacted(deps: CompactionDeps, chatId: string): Pro
 	// injected summarizer.
 	if (process.env.NODE_ENV === "test" && !deps.summarize) return;
 	await awaitCompaction(chatId);
+	if (deps.canStart?.() === false) return;
 	const record = deps.chatStore.getChat(chatId);
 	if (!record) return;
 	if (projectedContextTokens(record.messages, deps.systemPromptTokens) <= deps.thresholdTokens) return;
@@ -291,7 +295,7 @@ export async function ensureCompacted(deps: CompactionDeps, chatId: string): Pro
 export function kickCompaction(buildDeps: () => CompactionDeps | null, chatId: string): void {
 	if (state().inFlight.has(chatId)) return;
 	const deps = buildDeps();
-	if (!deps) return;
+	if (!deps || deps.canStart?.() === false) return;
 	const record = deps.chatStore.getChat(chatId);
 	if (!record) return;
 	if (projectedContextTokens(record.messages, deps.systemPromptTokens) <= deps.thresholdTokens) return;

src/server/desktop-access.ts

@@ -0,0 +1,17 @@
+import { timingSafeEqual } from "node:crypto";
+
+/** Per-launch desktop HTTP capability. This is not an OS-user malware
+ * boundary and is not an OAuth credential. Browser-only mode is unchanged. */
+export function desktopAPIAuthorized(
+	headers: Headers,
+	env: Record<string, string | undefined> = process.env,
+): boolean {
+	const expected = env.LH_DESKTOP_API_SECRET;
+	if (env.LH_DESKTOP_SERVER !== "1" && !expected) return true;
+	if (!expected || expected.length < 64) return false;
+	const supplied = headers.get("x-learnharness-desktop");
+	if (!supplied) return false;
+	const actual = Buffer.from(supplied);
+	const wanted = Buffer.from(expected);
+	return actual.length === wanted.length && timingSafeEqual(actual, wanted);
+}

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/jobs.ts

@@ -66,7 +66,9 @@ import {
 import type { TurnInput } from "../core/input-events.ts";
 import { danglingQuestionOf, endsInDismissal, type LhMessage } from "../core/native-messages.ts";
 import { fireJobAbort, isJobLive } from "../agent/subagents/aborts.ts";
-import { stores } from "./stores.ts";
+import { stores, effectiveSettings } from "./stores.ts";
+import { resolveConversationSpec } from "../core/settings-resolution.ts";
+import { chatGPTBackgroundAllowed, onChatGPTChange } from "../agent/chatgpt-bridge.ts";
 import { isTurnRunning } from "./turn-liveness.ts";
 import type { Agent } from "@earendil-works/pi-agent-core";
 
@@ -111,6 +113,8 @@ interface WakeState {
 	pumping: Set<string>;
 	/** Throttle bookkeeping for the stale-job sweep. */
 	lastSweepAt: number;
+	stopConnectionWatch?: () => void;
+	automaticDeferred: WeakSet<Agent>;
 }
 
 interface LhWakeGlobal {
@@ -119,7 +123,7 @@ interface LhWakeGlobal {
 
 function state(): WakeState {
 	const g = globalThis as typeof globalThis & LhWakeGlobal;
-	return (g.__lhWakePipeline ??= { queues: new Map(), liveAgents: new Map(), userSteers: new Map(), pumping: new Set(), lastSweepAt: 0 });
+	return (g.__lhWakePipeline ??= { queues: new Map(), liveAgents: new Map(), userSteers: new Map(), pumping: new Set(), lastSweepAt: 0, automaticDeferred: new WeakSet() });
 }
 
 /** chat-handler wires this at module init: its adapter around handleChatRequest. */
@@ -210,9 +214,18 @@ export function deliverJobResult(steer: SteerMessage): Promise<void> {
  *  beats the loop's synchronous start-poll drain and the marker rides the
  *  run's FIRST LLM call (the native equivalent of the old continuation
  *  coalescing). */
+export function automaticFollowUpAllowed(chatId: string, agent?: Agent): boolean {
+	if (agent && state().automaticDeferred.has(agent)) return false;
+	const provider = agent?.state?.model?.provider ?? resolveConversationSpec(stores.chatStore.getChat(chatId)?.settings, effectiveSettings().models.strong)?.provider;
+	return provider !== "chatgpt" || chatGPTBackgroundAllowed();
+}
+
 function handOffEligible(chatId: string, agent: Agent): void {
 	const queue = state().queues.get(chatId);
-	if (queue && queue.length > 0) {
+	// Keep automatic results in their owning registry. Never mix them into a
+	// foreground batch without express account consent. Learner steers below
+	// remain deliverable, including the first poll of an answer run.
+	if (automaticFollowUpAllowed(chatId, agent) && queue && queue.length > 0) {
 		for (const entry of queue) {
 			if (entry.handedTo !== agent && entry.steer.customType !== "probe-handoff") {
 				agent.steer(entry.steer);
@@ -290,9 +303,25 @@ export function reconcileUserSteers(chatId: string, persisted: readonly LhMessag
 /** Clear the live agent at settlement — MUST run before settleSteers so the
  *  reconciliation's re-pump can never hand a marker to the dying agent. */
 export function unregisterLiveAgent(chatId: string): void {
+	const agent = state().liveAgents.get(chatId);
+	if (agent) state().automaticDeferred.delete(agent);
 	state().liveAgents.delete(chatId);
 }
 
+/** Restore exact registry objects after native drain but before dispatch.
+ * The vehicle is disposable; result ownership/promises are not. */
+export function restoreUncommittedAutomatic(chatId: string, agent: Agent, automaticIds: ReadonlySet<string>, retainedIds: ReadonlySet<string>): void {
+	const s = state();
+	s.automaticDeferred.add(agent);
+	for (const entry of s.queues.get(chatId) ?? []) {
+		if (entry.handedTo === agent && automaticIds.has(entry.steer.id)) entry.handedTo = undefined;
+	}
+	agent.clearSteeringQueue();
+	for (const entry of s.userSteers.get(chatId) ?? []) {
+		if (!retainedIds.has(entry.message.id)) { agent.steer(entry.message); entry.handedTo = agent; }
+	}
+}
+
 /** TURN SETTLEMENT RECONCILIATION (Wave C): the handler calls this AFTER
  * unregisterLiveAgent, with the run's persisted messages. Two jobs:
  *
@@ -419,6 +448,7 @@ async function pump(chatId: string): Promise<void> {
 	const s = state();
 	if (s.pumping.has(chatId)) return;
 	s.pumping.add(chatId);
+	let parked = false;
 	try {
 		for (let attempt = 0; ; ) {
 			const queue = s.queues.get(chatId);
@@ -452,6 +482,7 @@ async function pump(chatId: string): Promise<void> {
 				return;
 			}
 			s.queues.set(chatId, stillPending);
+			if (!automaticFollowUpAllowed(chatId, s.liveAgents.get(chatId))) { parked = true; return; }
 
 			// VEHICLE 1 — THE LIVE RUN: pi's steering queue. The agent of the
 			// streaming turn takes every eligible marker; pi injects at the
@@ -525,7 +556,8 @@ async function pump(chatId: string): Promise<void> {
 			// hands them to that turn once the text-claim lands — until then,
 			// settlement re-pumps and lands them right after it). The pause is
 			// derived from the tail, so any user message un-holds by construction.
-			if (endsInDismissal(fresh.messages)) {
+			if (endsInDismissal(fresh.messages) || !automaticFollowUpAllowed(chatId)) {
+				parked = true;
 				return;
 			}
 
@@ -544,6 +576,15 @@ async function pump(chatId: string): Promise<void> {
 				console.error(`[steer ${steers.map((x) => x.id).join(",")}] wake turn crashed:`, err);
 				return -1;
 			});
+			if (status === 403) {
+				// Consent may have changed after the last gate. Preserve completed
+				// job/result ownership and wait for an explicit connection change.
+				const back = s.queues.get(chatId) ?? [];
+				back.push(...fire);
+				s.queues.set(chatId, back);
+				parked = true;
+				return;
+			}
 			if (status === 409) {
 				const back = s.queues.get(chatId) ?? [];
 				back.push(...fire);
@@ -573,10 +614,30 @@ async function pump(chatId: string): Promise<void> {
 		// a live run are NOT stragglers (settleSteers owns their resurrection)
 		// — restarting for them would just park in waitForIdle forever.
 		const q = state().queues.get(chatId) ?? [];
-		if (q.some((e) => e.handedTo === undefined)) void pump(chatId);
+		if (!parked && automaticFollowUpAllowed(chatId, state().liveAgents.get(chatId)) && q.some((e) => e.handedTo === undefined)) void pump(chatId);
 	}
 }
 
+// Connection changes are the only wake for consent-deferred automatic work.
+// No retry timer, inference, or API-key fallback is started while consent is off.
+state().stopConnectionWatch?.();
+state().stopConnectionWatch = onChatGPTChange(() => {
+	const s = state();
+	for (const [chatId, agent] of s.liveAgents) {
+		if (agent.state?.model?.provider !== "chatgpt" || automaticFollowUpAllowed(chatId, agent)) continue;
+		const consumed = new Set(agent.state.messages.map((m) => (m as { id?: string }).id));
+		const pending = (s.queues.get(chatId) ?? []).filter((entry) => entry.handedTo === agent && !consumed.has(entry.steer.id));
+		if (pending.length === 0) continue;
+		agent.clearSteeringQueue();
+		for (const entry of pending) entry.handedTo = undefined;
+		for (const entry of s.userSteers.get(chatId) ?? []) if (!consumed.has(entry.message.id)) entry.handedTo = undefined;
+		handOffEligible(chatId, agent); // only unconsumed learner input; never replay consumed steers
+	}
+	for (const chatId of s.queues.keys()) {
+		if (automaticFollowUpAllowed(chatId, s.liveAgents.get(chatId))) void pump(chatId);
+	}
+});
+
 /** Flip the job to its final status now that the lesson has resumed. */
 function finalizeJob(jobId: string): void {
 	const job = stores.jobStore.getJob(jobId);

src/server/keys.ts

@@ -52,6 +52,7 @@ import {
 	validateKeyShape,
 } from "../core/auth-keys.ts";
 import { authStore } from "../store/auth.ts";
+import { chatGPTConnected } from "../agent/chatgpt-bridge.ts";
 
 // The once-per-process boot record — globalThis anchor (HMR re-imports the
 // module; the snapshot must not re-run and misread fill-gapped vars as real
@@ -200,6 +201,7 @@ export function keyValue(id: string): string | undefined {
  * (gate by what would run). Providers absent from AUTH_PROVIDER_KEYS take
  * no key at all: nothing to gate, present by definition. */
 export function providerKeyPresent(provider: string): boolean {
+	if (provider === "chatgpt") return chatGPTConnected();
 	const info = AUTH_PROVIDER_KEYS.find((p) => p.id === provider);
 	return info ? keyValue(info.envVar) !== undefined : true;
 }

src/server/model-catalog.ts

@@ -20,6 +20,15 @@
 import { clampThinkingLevel, getSupportedThinkingLevels, type Api, type Model } from "@earendil-works/pi-ai";
 import type { ThinkingLevel } from "@earendil-works/pi-agent-core";
 import { piModels } from "../agent/models.ts";
+import { chatGPTConnected, chatGPTModel, chatGPTModels, refreshChatGPTModels } from "../agent/chatgpt-bridge.ts";
+
+export async function refreshConnectionCatalog(): Promise<void> {
+	await refreshChatGPTModels(piModels().getModels("openai").filter((m): m is Model<"openai-responses"> => m.api === "openai-responses"));
+}
+function connectionModel(provider: string, id: string): Model<Api> | undefined {
+	return provider === "chatgpt" ? chatGPTModel(id) : piModels().getModel(provider, id);
+}
+function connectionProviderIds(): string[] { return [...piModels().getProviders().map((p) => p.id), "chatgpt"]; }
 import { settingsStore } from "../store/settings.ts";
 import { effectiveSettings } from "./stores.ts";
 import { keyValue } from "./keys.ts";
@@ -126,7 +135,7 @@ export function catalogProviders(): CatalogProvider[] {
 		id: p.id,
 		envVar: p.envVar,
 		keyPresent: keyValue(p.envVar) !== undefined,
-	})).sort((a, b) => a.id.localeCompare(b.id));
+	})).concat([{ id: "chatgpt", envVar: "", keyPresent: chatGPTConnected() }]).sort((a, b) => a.id.localeCompare(b.id));
 }
 
 /** The providers whose key is present — the picker federates exactly
@@ -173,6 +182,7 @@ export function catalogContextLimits(): Map<string, number> {
 			index.set(`${provider.id}:${model.id}`, model.contextWindow);
 		}
 	}
+	for (const model of chatGPTModels()) index.set(`chatgpt:${model.id}`, model.contextWindow);
 	return index;
 }
 
@@ -187,7 +197,8 @@ export function modelCatalog(): ModelCatalogResponse {
 		.filter((p) => keyed.has(p.id))
 		.flatMap((p) => piModels().getModels(p.id))
 		.map(toModelDto)
-		.sort((a, b) => (a.provider === b.provider ? a.id.localeCompare(b.id) : a.provider.localeCompare(b.provider)));
+		.sort((a, b) => (a.provider === b.provider ? a.id.localeCompare(b.id) : a.provider.localeCompare(b.provider)))
+		.concat(chatGPTModels().map(toModelDto));
 	return {
 		models,
 		defaults: catalogDefaults(),
@@ -209,13 +220,13 @@ export type SettingsValidation =
  *    the model the picker writes against (settings.model, else config's
  *    prober/teacher pin; one controller, so the two specs share it). */
 export function validateChatSettings(settings: ChatSettings): SettingsValidation {
-	const providers = piModels().getProviders().map((p) => p.id);
+	const providers = connectionProviderIds();
 	if (settings.model) {
 		const { provider, id } = settings.model;
 		if (!providers.includes(provider)) {
 			return { ok: false, error: `Unknown provider "${provider}" — configured: ${providers.join(", ")}` };
 		}
-		if (!piModels().getModel(provider, id)) {
+		if (!connectionModel(provider, id)) {
 			return { ok: false, error: `No model "${id}" in the ${provider} catalog` };
 		}
 	}
@@ -239,13 +250,13 @@ export function validateModelsSettings(settings: ModelsSettings): SettingsValida
 }
 
 function validateSlotSpec(name: ModelSlotName, slot: { model?: { provider: string; id: string }; thinkingLevel?: ThinkingLevel }): SettingsValidation {
-	const providers = piModels().getProviders().map((p) => p.id);
+	const providers = connectionProviderIds();
 	if (slot.model) {
 		const { provider, id } = slot.model;
 		if (!providers.includes(provider)) {
 			return { ok: false, error: `Unknown provider "${provider}" — configured: ${providers.join(", ")}` };
 		}
-		if (!piModels().getModel(provider, id)) {
+		if (!connectionModel(provider, id)) {
 			return { ok: false, error: `No model "${id}" in the ${provider} catalog` };
 		}
 	}
@@ -267,7 +278,7 @@ function validateSlotSpec(name: ModelSlotName, slot: { model?: { provider: strin
  *  or the level already supported — the existence gate rejects unknown
  *  models separately, so this only ever sees real ones in practice). */
 function clampLevelFor(provider: string, id: string, level: ThinkingLevel): ThinkingLevel {
-	const model = piModels().getModel(provider, id);
+	const model = connectionModel(provider, id);
 	return model ? clampThinkingLevel(model, level) : level;
 }
 

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/pi/ui-stream.ts

@@ -2,6 +2,7 @@ import type { Agent, AgentEvent, AgentMessage } from "@earendil-works/pi-agent-c
 import type { SteerMessage } from "../../core/contracts.ts";
 import { noteSteerVisible } from "../jobs.ts";
 import { ToolCallIdUniquifier } from "./call-ids.ts";
+import type { ChatGPTRequestCycle } from "../../agent/chatgpt-request-cycle.ts";
 
 /**
  * THE NATIVE WIRE (Wave D, owner-approved 2026-09-26): one agent turn, streamed
@@ -71,6 +72,7 @@ export type TurnWireEvent =
 	| { type: "error"; errorText: string };
 
 export interface PiTurnSpec {
+	requestCycle?: ChatGPTRequestCycle;
 	/** The per-turn agent (stateless engine — built from the record). */
 	agent: Agent;
 	/** The handler's error-masking policy for LOOP-level failures (the
@@ -117,7 +119,7 @@ export interface PiTurnHandle {
 }
 
 export function runPiTurn(spec: PiTurnSpec): PiTurnHandle {
-	const { agent, onError, onFinish, usedToolCallIds, updateFlushMs = 100 } = spec;
+	const { agent, onError, onFinish, usedToolCallIds, updateFlushMs = 100, requestCycle } = spec;
 	const uniquifier = usedToolCallIds ? new ToolCallIdUniquifier(usedToolCallIds) : undefined;
 
 	let settled = false;
@@ -248,11 +250,24 @@ export function runPiTurn(spec: PiTurnSpec): PiTurnHandle {
 			settleWith();
 		}
 	};
-	agent.subscribe(handle);
+	requestCycle?.bindRelease(handle);
+	agent.subscribe((event) => {
+		if (!requestCycle?.hold(event)) handle(event);
+	});
 
 	void (async () => {
 		try {
-			await agent.continue();
+			do {
+				await agent.continue();
+				if (!requestCycle) break;
+				const denied = requestCycle.deniedBeforeDispatch;
+				const outcome = requestCycle.finishNativeCycle();
+				if (denied) agent.state.messages = outcome.messages;
+				if (outcome.action !== "resumeForeground") {
+					handle({ type: "agent_end", messages: [...agent.state.messages] });
+					break;
+				}
+			} while (!sawAbort);
 		} catch (err) {
 			// A loop-level failure (not a provider error — those ride the
 			// stream as stopReason "error"): surface, close, persist. If the
@@ -318,6 +333,7 @@ export function runPiTurn(spec: PiTurnSpec): PiTurnHandle {
 		abort() {
 			if (settled) return;
 			sawAbort = true;
+			requestCycle?.abort();
 			agent.abort();
 		},
 	};

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);
+}