hob-bot 1.0.4 → 1.2.0

package/vendor/claude-bridge/src/index.ts

@@ -0,0 +1,2503 @@
+import { createAssistantMessageEventStream, type AssistantMessage, type AssistantMessageEventStream, type Context, type ImageContent, type Model, type SimpleStreamOptions, type TextContent, type Tool, type UserMessage } from "@earendil-works/pi-ai";
+import { getModels } from "@earendil-works/pi-ai/compat";
+import { buildSessionContext, keyHint, type ExtensionAPI, type ExtensionContext, type ExtensionUIContext } from "@earendil-works/pi-coding-agent";
+import { query, type EffortLevel, type SDKMessage, type SettingSource } from "@anthropic-ai/claude-agent-sdk";
+import type { Base64ImageSource, ContentBlockParam } from "@anthropic-ai/sdk/resources";
+import { Text } from "@earendil-works/pi-tui";
+import { appendFileSync, mkdirSync } from "fs";
+import { dirname, join } from "path";
+import { PROVIDER_ID, messageContentToText } from "./convert.js";
+import { debugLogPath, diagLogPath } from "./log-paths.js";
+import { applyLongContext, buildModels, claudeCodeModelId, type LongContextSettings, resolveModel as _resolveModel } from "./models.js";
+import { MCP_SERVER_NAME, MCP_TOOL_PREFIX, renderSkillsBlock } from "./skills.js";
+import { extractAllToolResults as _extractAllToolResults, type McpResult } from "./extract-tool-results.js";
+import { QueryContext, ctx } from "./query-state.js";
+import { makePromptStream, userMessage, type PromptStream } from "./prompt-stream.js";
+import { claudeCodeSettings, loadConfig, markAskClaudeDeprecationNoticeShown, markStartupNoticeShown, type Config } from "./config.js";
+import {
+	collectPromptSkills,
+	projectPromptCapture,
+	sharedPromptCaptures,
+	type PromptCapture,
+} from "./prompt-capture.js";
+import { createToolServer } from "./mcp-server.js";
+import { buildActionSummary, type ToolCallState } from "./askclaude-ui.js";
+import { askClaudeCallTags, askClaudeToolDescription, buildAskClaudeParams, resolveAskClaudeDefaults, resolveAskClaudeMode, type AskClaudeMode } from "./askclaude-schema.js";
+import { nonSystemMessages, toBridgeContext } from "./transcript.js";
+import { forgetSessionCwd, rememberSessionCwd, rememberSummarySession, sessionCwdFor, summarySessionFor } from "./session-cwd.js";
+import { updateUsage, type SdkUsage } from "./usage.js";
+import { rateLimitResponse } from "./rate-limit.js";
+
+// --- Debug logging ---
+// CLAUDE_BRIDGE_DEBUG=1 enables debug logging to the bridge log in pi's agent
+// dir (log-paths.ts), not a fixed ~/.pi/agent.
+
+const DEBUG = process.env.CLAUDE_BRIDGE_DEBUG === "1";
+
+// Applied to every Claude Code subprocess the bridge spawns — provider, AskClaude
+// and the compact summary. One place, so a guard is added once rather than three
+// times, and so a missing one is visible.
+//
+// - ENABLE_CLAUDEAI_MCP_SERVERS=0: keep the user's claude.ai-connected MCP servers
+//   out of a pi session, which serves its own tools.
+// - DISABLE_AUTO_COMPACT=1: pi owns compaction; CC compacting its own copy would
+//   diverge from pi's history, which is the source of truth for every rebuild.
+const CC_CHILD_ENV = {
+	ENABLE_CLAUDEAI_MCP_SERVERS: "0",
+	DISABLE_AUTO_COMPACT: "1",
+} as const;
+
+// Pi owns context files on the provider path, so Claude Code must not load its
+// own on top: otherwise project CLAUDE.md/AGENTS.md files arrive twice, and
+// ~/.claude/CLAUDE.md — a user persona written for a harness that is not the
+// one running — arrives at all, stamped "These instructions OVERRIDE any default
+// behavior" and outranking Pi's own AGENTS.md.
+//
+// Excludes rather than settingSources: the source gate that suppresses CLAUDE.md
+// is the same one that reads settings.json, where Bedrock/Vertex users keep
+// `env` and `apiKeyHelper`. Patterns are matched with picomatch against absolute
+// paths; the filename globs cover user, ancestor, project and .claude/ copies,
+// while rules need their own. Managed/policy memory is not excludable by design.
+const CLAUDE_MD_EXCLUDES = ["**/CLAUDE.md", "**/AGENTS.md", "**/.claude/rules/**"];
+
+// Unique per module evaluation — confirms whether subagents share module state
+const moduleInstanceId = Math.random().toString(36).slice(2, 8);
+
+function debug(...args: unknown[]) {
+	if (!DEBUG) return;
+	const logPath = debugLogPath();
+	try { mkdirSync(dirname(logPath), { recursive: true }); } catch { /* ignore */ }
+	const ts = new Date().toISOString();
+	const fmt = (a: unknown): string => {
+		if (typeof a === "string") return a;
+		if (a instanceof Error) return `${a.name}: ${a.message}${a.stack ? "\n" + a.stack : ""}`;
+		return JSON.stringify(a);
+	};
+	const msg = args.map(fmt).join(" ");
+	appendFileSync(logPath, `[${ts}] [${moduleInstanceId}] ${msg}\n`);
+}
+
+// Per-query CLI debug capture. When CLAUDE_BRIDGE_DEBUG=1, ask the Claude Code
+// CLI subprocess to write its own debug log to a file we choose, and also
+// forward its stderr into our debug stream. Drops straight into the real SDK's
+// Options — see @anthropic-ai/claude-agent-sdk sdk.d.ts:1245 (debug, debugFile,
+// stderr). Without this, CC's internal view of the world is invisible to us
+// and "No conversation found" / empty-error reports are unactionable.
+let nextCliDebugSeq = 1;
+function makeCliDebugOptions(tag: string): { debug?: boolean; debugFile?: string; stderr?: (data: string) => void } {
+	if (!DEBUG) return {};
+	const seq = nextCliDebugSeq++;
+	const ts = new Date().toISOString().replace(/[:.]/g, "-");
+	const logDir = join(dirname(debugLogPath()), "cc-cli-logs");
+	try { mkdirSync(logDir, { recursive: true }); } catch { /* ignore */ }
+	const debugFile = join(logDir, `${ts}-${tag}-${seq}.log`);
+	debug(`cli-debug: ${tag} #${seq} → ${debugFile}`);
+	return {
+		debug: true,
+		debugFile,
+		stderr: (data: string) => {
+			for (const line of data.split(/\r?\n/)) {
+				if (line) debug(`[cli-stderr ${tag}#${seq}] ${line}`);
+			}
+		},
+	};
+}
+
+/** Unconditional diagnostic dump — for "should never happen" paths. Creates pi's agent
+ *  dir itself: callers run inside streamSimple, where a missing dir must not throw. */
+function diagDump(label: string, data: Record<string, unknown>) {
+	const ts = new Date().toISOString();
+	const entry = { ts, moduleInstanceId, label, ...data };
+	const logPath = diagLogPath();
+	mkdirSync(dirname(logPath), { recursive: true });
+	appendFileSync(logPath, JSON.stringify(entry) + "\n");
+	debug(`DIAG: ${label} (see ${logPath})`);
+}
+
+// --- Constants ---
+
+// Marks which bridge module instance owns the registered provider's stream fn.
+// Full registration policy (first vs later instances, shared vs own registry):
+// see the "--- Provider ---" block in activate() below.
+
+// Claude Code's own builtin tools, for the AskClaude path where CC really runs
+// them. The provider path never sees these — it starts CC with `tools: []`.
+const SDK_TO_PI_TOOL_NAME: Record<string, string> = {
+	read: "read", write: "write", edit: "edit", bash: "bash",
+};
+
+// MODELS is buildModels(getModels("anthropic")) — projection kept in models.js.
+const MODELS = buildModels(getModels("anthropic"));
+let providerSettings: NonNullable<Config["provider"]> = {};
+let longContextSettings: LongContextSettings = { plan: "pro", longContextExtraUsage: false };
+
+function resolveModel(input: string) {
+	return _resolveModel(MODELS, input);
+}
+
+// --- Error handling ---
+
+function errorMessage(err: unknown): string {
+	if (err instanceof Error) return err.message;
+	if (err && typeof err === "object") {
+		const obj = err as Record<string, unknown>;
+		if (typeof obj.message === "string") return obj.message;
+		if (typeof obj.error === "string") return obj.error;
+		try { return JSON.stringify(err); } catch {}
+	}
+	return String(err);
+}
+
+// AskClaude mode presets — controls which CC tools are blocked per mode.
+// Only block tools that can't work (no pi TUI for user interaction).
+// Other CC tools (Agent, SendMessage, RemoteTrigger, Tasks, etc.) are intentionally not blocked.
+const ASKCLAUDE_ALWAYS_BLOCKED = [
+	"AskUserQuestion", "EnterPlanMode", "ExitPlanMode",
+	"ToolSearch", // probes for blocked tools, wastes tokens
+	"ScheduleWakeup", // no harness to fire wakeup from inside a delegated subagent
+];
+const MODE_DISALLOWED_TOOLS: Record<AskClaudeMode, string[]> = {
+	full: ASKCLAUDE_ALWAYS_BLOCKED,
+	read: [
+		...ASKCLAUDE_ALWAYS_BLOCKED,
+		"Write", "Edit", "Bash", "NotebookEdit",
+		"EnterWorktree", "ExitWorktree", "CronCreate", "CronDelete", "TeamCreate", "TeamDelete",
+	],
+	none: [
+		...ASKCLAUDE_ALWAYS_BLOCKED,
+		"Read", "Write", "Edit", "Glob", "Grep", "Bash", "Agent",
+		"NotebookEdit", "EnterWorktree", "ExitWorktree",
+		"CronCreate", "CronDelete", "TeamCreate", "TeamDelete",
+		"WebFetch", "WebSearch",
+	],
+};
+
+// --- Session persistence ---
+
+interface SessionState {
+	sessionId: string;
+	cursor: number;
+	cwd: string;
+	// The pi session this CC conversation serves, from the provider call's
+	// options.sessionId. Attribution for history rewrites: a subagent's
+	// session_compact must not force a rebuild of a conversation that belongs
+	// to a different pi session. Null until a provider call recorded it.
+	piSessionId?: string;
+	// Context rewrites and interrupted queries require a fresh SDK replay.
+	needsRebuild?: boolean;
+	forceRotate?: boolean;
+}
+
+/** The session key a pi session's provider calls address. Unattributed calls
+ *  (no options.sessionId — AskClaude's direct sync, or a host that omits it)
+ *  share the "(none)" bucket: they cannot be told apart, so they share the
+ *  pre-existing single-slot semantics. */
+function sessionKey(piSessionId: string | null | undefined): string {
+	return piSessionId ?? "(none)";
+}
+
+/** Mirror of the CC conversation one pi session's turns are running on. One
+ *  entry per pi session: a bridge process serves several sessions at once
+ *  (pi-subagents children run their own AgentSessions), and a single shared
+ *  slot forced them to fight over it — a length-matching foreign sync could
+ *  REUSE or rebuild another session's CC file, and the completion capture was
+ *  last-writer-wins (a foreground child in the parent's first turn permanently
+ *  reassigned the parent's conversation). Keyed lookup removes the fight: each
+ *  session's reads, writes and teardown marks touch only its own mirror. */
+const sharedSessions = new Map<string, SessionState>();
+
+/** The mirror for `piSessionId`, or null when this session has none yet. */
+function sessionStateFor(piSessionId: string | null | undefined): SessionState | null {
+	return sharedSessions.get(sessionKey(piSessionId)) ?? null;
+}
+
+/** Replace (or plant) the mirror for `piSessionId`. */
+function setSessionStateFor(piSessionId: string | null | undefined, state: SessionState | null): void {
+	if (state === null) sharedSessions.delete(sessionKey(piSessionId));
+	else sharedSessions.set(sessionKey(piSessionId), state);
+}
+
+// pi replaced one of its sessions' history (compact, tree) rather than appending
+// to it. Read on the tool-result path — the one provider call that never reaches
+// syncSharedSession — so a query parked at a tool boundary is discarded rather
+// than resumed (issue #101). Keyed by pi session id, not process-global: a bridge
+// process serves several pi sessions at once (subagents run their own
+// AgentSessions and can compact mid-run while the parent is parked), and
+// marking across that boundary kills healthy queries (or, worse, rebuilds the
+// parent around a compaction that never touched it).
+//
+// Holds real pi session ids only, never the "(none)" key: no production caller
+// marks with null (the rewrite events attribute via ctx.sessionManager), and
+// every reader guards on a non-null piSessionId — so a "(none)" entry could
+// never be matched or consumed, only leaked.
+const historyRewrittenBySession = new Set<string>();
+
+/** Handlers that arm rewrite staleness, one per module instance. Worktree-
+ *  spawned subagents can load this module fresh (pi's loader cache is keyed on
+ *  cwd, and a worktree cwd clears it), while the serving instance — whose
+ *  streamSimple the pi sessions actually call — is whoever registered first.
+ *  A fresh instance must forward its session's rewrites to the serving one.
+ *  Symbol.for: one registry per process, like SHARED_CAPTURES_KEY below. */
+const MARK_REBUILD_HOOKS_KEY = Symbol.for("claude-bridge:markRebuildHooks");
+type MarkRebuildHook = (piSession: string | null, event: string) => void;
+const markRebuildHooks: Set<MarkRebuildHook> =
+	((globalThis as Record<symbol, unknown>)[MARK_REBUILD_HOOKS_KEY] as Set<MarkRebuildHook> | undefined) ?? new Set();
+(globalThis as Record<symbol, unknown>)[MARK_REBUILD_HOOKS_KEY] = markRebuildHooks;
+
+/** pi mutated its messages array out from under us: force the next
+ *  syncSharedSession down REBUILD, and arm the discard above. `piSession`
+ *  is never null from an event handler (each pi session has its own runner).
+ *  The "(none)" key still covers the mirror for a hypothetical direct caller
+ *  with no session id, so such a rewrite forces the REBUILD side; the discard
+ *  set holds real ids only — see the comment on historyRewrittenBySession. */
+function markRebuildForSession(piSession: string | null, event: string): void {
+	// The rewriting session's own mirror: the rewrite changed the history it was
+	// built from, so its next sync must REBUILD rather than REUSE. Every other
+	// session's mirror stays untouched — its conversation was never rewritten.
+	const key = sessionKey(piSession);
+	const state = sharedSessions.get(key);
+	if (!state) {
+		debug(`${event}: history rewritten, no session to mark yet`);
+	} else {
+		sharedSessions.set(key, { ...state, needsRebuild: true });
+		debug(`${event}: marking needsRebuild on session ${state.sessionId.slice(0, 8)}`);
+	}
+	// Arming parked contexts cannot wait for delivery: the entry checks
+	// `resultCtx.historyStale`, and a rewrite usually lands *while* the query is
+	// parked (compaction runs inside pi's turn loop, not between provider calls).
+	if (piSession) historyRewrittenBySession.add(piSession);
+	armStaleContexts();
+}
+
+/** Copy each armed session's mark onto the parked queries built from it. */
+function armStaleContexts(): void {
+	for (const c of activeQueryContexts) {
+		if (c.piSessionId && historyRewrittenBySession.has(c.piSessionId)) c.historyStale = true;
+	}
+}
+
+// This instance enlists. Session ids reaching any hook equal options.sessionId
+// on the serving instance's provider calls, so forwarding is safe: only the
+// owning instance's contexts and SessionState match the key.
+markRebuildHooks.add(markRebuildForSession);
+
+/** Event handlers call this: it fans the rewrite out to every module instance
+ *  in the process, of which exactly one is serving provider traffic for any
+ *  given pi session. */
+function sponsorMarkRebuildForSession(piSession: string | null, event: string): void {
+	for (const hook of markRebuildHooks) hook(piSession, event);
+}
+
+// Pi doesn't pass tool results directly — it appends them to the context and calls
+// the provider again. Thin wrapper over extract-tool-results.js that adds per-turn
+// debug logging at the extraction boundary.
+function extractAllToolResults(context: Context): McpResult[] {
+	const { results, stopIdx } = _extractAllToolResults(context.messages as unknown as Array<{ role: string; [key: string]: unknown }>);
+	debug(`extractAllToolResults: ${results.length} results from ${context.messages.length} msgs, stopped at index ${stopIdx}`);
+	debug(`extractAllToolResults: all msg roles:`, context.messages.map((m, i) => `[${i}]${m.role}`).join(" "));
+	for (let r = 0; r < results.length; r++) {
+		debug(`extractAllToolResults: result[${r}] id=${results[r].toolCallId}${results[r].isError ? " ERROR" : ""} preview:`, JSON.stringify(results[r].content).slice(0, 150));
+	}
+	return results;
+}
+
+/** Index of the first message of the current user turn — the trailing run of
+ *  user messages that has not been written into the Claude Code session yet.
+ *  Equals messages.length when the last message is not a user message.
+ *
+ *  Single source of truth for the history/prompt split: everything before this
+ *  index is replayed as session history, everything from it onward becomes the
+ *  prompt. Deriving both halves from one index is what keeps a message from
+ *  landing in both — an extension appending a display-only user message after
+ *  the real one (see issue #34) makes the turn longer than one message. */
+function turnStart(messages: Context["messages"]): number {
+	let i = messages.length;
+	while (i > 0 && messages[i - 1].role === "user") i--;
+	return i;
+}
+
+/** Extract the current user turn as a prompt string. Returns null if the last message is not a user message. */
+function extractUserPrompt(messages: Context["messages"]): string | null {
+	const turn = messages.slice(turnStart(messages)) as UserMessage[];
+	if (turn.length === 0) return null;
+	// Drop empties before joining so an all-empty turn still yields "" and trips
+	// the caller's empty-prompt guard rather than sending bare newlines.
+	return turn
+		.map((m) => (typeof m.content === "string" ? m.content : messageContentToText(m.content)))
+		.filter((text) => text)
+		.join("\n");
+}
+
+/** Extract the current user turn as ContentBlockParam[] (preserving images).
+ *  Returns null if no images — caller should fall back to string prompt. */
+function extractUserPromptBlocks(messages: Context["messages"]): ContentBlockParam[] | null {
+	const turn = messages.slice(turnStart(messages)) as UserMessage[];
+	if (turn.length === 0) return null;
+
+	let hasImage = false;
+	const blocks: ContentBlockParam[] = [];
+	for (const message of turn) {
+		const content: (TextContent | ImageContent)[] = typeof message.content === "string"
+			? [{ type: "text", text: message.content }]
+			: message.content;
+		// Off-type content violates UserMessage's contract, so fail rather than
+		// degrade — but name the shape, since the cause is almost always another
+		// extension appending a malformed message, not this file.
+		if (!Array.isArray(content)) {
+			throw new Error(
+				`extractUserPromptBlocks: user message content must be a string or block array, got ${typeof content} — likely a malformed message from another extension`,
+			);
+		}
+		for (const block of content) {
+			if (block.type === "text" && block.text) {
+				blocks.push({ type: "text", text: block.text });
+			} else if (block.type === "image") {
+				// Guard before logging: data-less image blocks do occur, and reading
+				// .length off the missing field in the debug template would throw
+				// before this check ever runs (template args evaluate unconditionally).
+				if (!block.data || !block.mimeType) {
+					debug(`image block missing data or mimeType, skipping: keys=${Object.keys(block).join(",")}`);
+					continue;
+				}
+				debug(`image block: mimeType=${block.mimeType}, data length=${block.data.length}`);
+				hasImage = true;
+				blocks.push({
+					type: "image",
+					source: {
+						type: "base64",
+						media_type: block.mimeType as Base64ImageSource["media_type"],
+						data: block.data,
+					},
+				});
+			}
+		}
+	}
+	debug(`extractUserPromptBlocks: ${turn.length} msgs in turn, ${blocks.length} blocks, types=${blocks.map((b) => b.type).join(",")}`);
+	return hasImage ? blocks : null;
+}
+
+// A session cannot be resumed after a restart or context rewrite. Replay only
+// Hob's own history as SDK input; Claude Code alone manages its session files.
+// Historical images are never re-sent: the whole history travels inside ONE
+// SDK user message, so an inline image reads as an attachment of the current
+// turn (the model answered a text-only message by describing "the three images
+// attached here" — issue #71), and every replay re-uploads all of them. Each
+// becomes a text placeholder instead; the current turn's own images still
+// travel via extractUserPromptBlocks after the 「以下是当前消息：」 marker.
+function replayBlocks(messages: Context["messages"]): ContentBlockParam[] {
+	const blocks: ContentBlockParam[] = [{ type: "text", text: "以下是此前的对话与工具结果，按顺序作为历史参考；历史里的图片不重发，只以占位文字标注；最后一段是当前消息。" }];
+	for (const message of messages) {
+		const m = message as unknown as Record<string, unknown>;
+		const text = JSON.stringify({ role: m.role, content: m.content, toolCallId: m.toolCallId, toolName: m.toolName, isError: m.isError }, (_key, value) => {
+			if (value && typeof value === "object" && value.type === "image" && typeof value.data === "string" && typeof value.mimeType === "string") {
+				return "[图片：这条历史消息当时带了一张图片，不在当前消息里]";
+			}
+			return value;
+		});
+		blocks.push({ type: "text", text: `历史消息：${text}` });
+	}
+	blocks.push({ type: "text", text: "以下是当前消息：" });
+	return blocks;
+}
+
+function newAssistantOutput(model: Model<any>, text: string, stopReason: AssistantMessage["stopReason"], errorMessage?: string): AssistantMessage {
+	return {
+		role: "assistant",
+		content: text ? [{ type: "text", text }] : [],
+		api: model.api,
+		provider: model.provider,
+		model: model.id,
+		usage: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, totalTokens: 0,
+			cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } },
+		stopReason,
+		...(errorMessage ? { errorMessage } : {}),
+		timestamp: Date.now(),
+	};
+}
+
+function extractIsolatedSummaryPrompt(messages: Context["messages"]): string {
+	if (messages.length !== 1 || messages[0].role !== "user") {
+		throw new Error(
+			`isolatedStreamFn: expected exactly 1 user message, got ${messages.length} ` +
+			`(${messages.map((m) => m.role).join(",")})`,
+		);
+	}
+	const promptText = extractUserPrompt(messages);
+	if (!promptText) throw new Error("isolatedStreamFn: summarization prompt is empty");
+	return promptText;
+}
+
+/** Failure text for an SDK result, or undefined when it succeeded. CC reports API failures
+ *  (429 capacity, overload, prompt-too-long) with `is_error` on an otherwise success-shaped
+ *  result; the dedicated error subtypes carry `errors` instead. */
+function resultErrorText(message: SDKMessage): string | undefined {
+	const result = message as SDKMessage & { subtype?: string; is_error?: boolean; result?: string; errors?: unknown; error?: unknown };
+	if (result.subtype === "success") return result.is_error ? result.result || "Claude Code reported an error" : undefined;
+	if (Array.isArray(result.errors) && result.errors.length) return result.errors.map(String).join("\n");
+	if (typeof result.error === "string") return result.error;
+	return `Claude Code failed: ${result.subtype ?? "unknown result"}`;
+}
+
+/** Name a failure as a rate limit when a rejection preceded it.
+ *
+ *  pi has no typed rate-limit error — `stopReason` is only ever `"error"` and the sole carrier
+ *  is `errorMessage` — so everything that reacts to a rate limit pattern-matches that string:
+ *  pi-subagents gates `fallbackModels` on its own pattern list, and key-rotating extensions use
+ *  their own. Claude Code words a subscription limit as "You're out of extra usage · resets
+ *  6:30pm", which matches none of them, so an exhausted quota reads as a fatal error and the
+ *  fallback chain never runs (issue #58).
+ *
+ *  Leading with "Claude quota exceeded — rate limit" rather than appending keeps both phrases
+ *  in any truncated render, and avoids the `<tool> failed (exit N):` shape that pi-subagents
+ *  treats as a tool failure and refuses to retry. "quota exceeded" is on pi-ai's non-retryable
+ *  list, which wins over its retryable "rate limit" pattern, so pi hands the rejected window
+ *  straight to fallback instead of retrying a limit that resets in hours (issue #183). */
+function describeRateLimitFailure(rejection: { rateLimitType?: string; resetsAt?: number }, failure: string): string {
+	const kind = rejection.rateLimitType ? ` (${rejection.rateLimitType})` : "";
+	const resets = rejection.resetsAt ? ` — resets ${new Date(rejection.resetsAt * 1000).toLocaleTimeString()}` : ""; // resetsAt: Unix seconds (unit undocumented in the SDK; observed)
+	return `Claude quota exceeded — rate limit${kind}${resets}: ${failure}`;
+}
+
+/** Reports to the call that owns the open stream, before its next event, as pi's own
+ *  providers do. A failing consumer must not end the turn it is observing. */
+async function reportHeldRateLimit(c: QueryContext, model: Model<any>): Promise<void> {
+	const info = c.heldRateLimit;
+	if (!info || !c.currentPiStream || !c.currentOnResponse) return;
+	c.heldRateLimit = undefined;
+	try {
+		await c.currentOnResponse(rateLimitResponse(info), model);
+	} catch (error) {
+		debug("rate limit: onResponse failed", error);
+	}
+}
+
+function isolatedStreamFn(model: Model<any>, context: Context, options?: SimpleStreamOptions): AssistantMessageEventStream {
+	const stream = createAssistantMessageEventStream();
+	void runIsolatedSummary(model, context, options, stream);
+	return stream;
+}
+
+async function runIsolatedSummary(
+	model: Model<any>,
+	context: Context,
+	options: SimpleStreamOptions | undefined,
+	stream: AssistantMessageEventStream,
+): Promise<void> {
+	// pi delivers compaction/branch-summary requests as a transcript: the summarization
+	// prompt folded into a leading system message ahead of the lone user message
+	// (issue #106). toBridgeContext restores the prompt/tools fields the extraction
+	// assertion below assumes; the summarization prompt still reaches CC as its systemPrompt.
+	context = toBridgeContext(context);
+	let sdkQuery: ReturnType<typeof query> | undefined;
+	let wasAborted = false;
+	const onAbort = () => {
+		wasAborted = true;
+		void sdkQuery?.interrupt().catch(() => {});
+		try { sdkQuery?.close(); } catch {}
+	};
+
+	try {
+		// One-off summarizer calls (compaction, branch summary, turn prefix, bug report —
+		// anything routed through pi's completeSummarization) are marked cacheRetention:
+		// "none". Any of them may appear in a future pi release without a bridge change,
+		// so route on the marker, not on which summarizer is calling. Non-summarizer calls
+		// must still match the [system,user] compaction shape exactly.
+		const isOneOffSummary = options?.cacheRetention === "none";
+		const promptText = isOneOffSummary
+			? extractUserPrompt(context.messages)
+			: extractIsolatedSummaryPrompt(context.messages);
+		if (!promptText) throw new Error("runIsolatedSummary: one-off summary without a user prompt (last message is not user?)");
+		const cwd = sessionCwdFor(summarySessionFor(options?.signal) ?? options?.sessionId);
+		const compactProviderSettings = loadConfig(cwd).provider;
+		const claudeExecutable = compactProviderSettings?.pathToClaudeCodeExecutable;
+		const cliModel = claudeCodeModelId(model, longContextSettings);
+		// Honour the caller's reasoning level, same mapping as the provider path.
+		const effort = effortForReasoning(model, options?.reasoning);
+		debug(`compact summary: spawn model=${cliModel} registeredModel=${model.id} effort=${effort ?? "default"} promptLen=${promptText.length}`);
+
+		sdkQuery = queryImpl({
+			prompt: promptText,
+			options: {
+				cwd,
+				env: { ...process.env, ...CC_CHILD_ENV },
+				settings: { autoMemoryEnabled: false },
+				tools: [],
+				strictMcpConfig: true,
+				extraArgs: { "safe-mode": null },
+				skills: [],
+				persistSession: false,
+				systemPrompt: context.systemPrompt,
+				model: cliModel,
+				maxTurns: 1,
+				...(effort ? { effort } : {}),
+				...(claudeExecutable ? { pathToClaudeCodeExecutable: claudeExecutable } : {}),
+				...makeCliDebugOptions("compact-summary"),
+			},
+		});
+
+		if (options?.signal) {
+			if (options.signal.aborted) onAbort();
+			else options.signal.addEventListener("abort", onAbort, { once: true });
+		}
+
+		let assistantText = "";
+		let finalText = "";
+		let errorText: string | undefined;
+		let firstEventLogged = false;
+
+		for await (const message of sdkQuery) {
+			if (!firstEventLogged) {
+				debug(`compact summary: first event type=${message.type}`);
+				firstEventLogged = true;
+			}
+			if (wasAborted) break;
+
+			if (message.type === "assistant") {
+				for (const block of (message as any).message?.content ?? []) {
+					if (block.type === "text" && typeof block.text === "string") assistantText += block.text;
+				}
+			} else if (message.type === "result") {
+				logServedContextWindow("compact summary", message, model);
+				errorText = resultErrorText(message);
+				if (!errorText && message.subtype === "success") finalText = message.result || assistantText;
+			}
+		}
+
+		if (wasAborted) {
+			const output = newAssistantOutput(model, "", "aborted", "Operation aborted");
+			debug("compact summary: aborted");
+			stream.push({ type: "error", reason: "aborted", error: output });
+			stream.end();
+			return;
+		}
+
+		const text = finalText || assistantText;
+		if (errorText || !text.trim()) {
+			const msg = errorText ?? "Claude Code summary returned empty text";
+			debug(`compact summary: error ${msg}`);
+			stream.push({ type: "error", reason: "error", error: newAssistantOutput(model, "", "error", msg) });
+			stream.end();
+			return;
+		}
+
+		debug(`compact summary: done textLen=${text.length}`);
+		stream.push({ type: "done", reason: "stop", message: newAssistantOutput(model, text, "stop") });
+		stream.end();
+	} catch (err) {
+		const msg = errorMessage(err);
+		debug("runIsolatedSummary threw; pushing terminal error", err);
+		stream.push({ type: "error", reason: "error", error: newAssistantOutput(model, "", "error", msg) });
+		stream.end();
+	} finally {
+		options?.signal?.removeEventListener("abort", onAbort);
+		try { sdkQuery?.close(); } catch {}
+	}
+}
+
+interface SyncResult {
+	sessionId: string | null;
+	preserveSharedSession?: boolean;
+	replay?: Context["messages"];
+}
+
+/**
+ * Ensure the shared session has all messages up to (but not including) the last user message.
+ * Returns session ID to resume from, or null if no resume needed.
+ */
+// Resume an unchanged Claude session or replay Hob history through the SDK.
+	function syncSharedSession(
+	messages: Context["messages"],
+	cwd: string,
+	customToolNameToSdk?: Map<string, string>,
+	modelId?: string,
+	piSessionId?: string | null,
+): SyncResult {
+	// System messages are pi's transcript representation of prompt and tool state, not
+	// conversation history — they are never imported into a CC session, so exclude them from
+	// the history space (priorMessages, cursor, missed) everywhere below (issue #106).
+	// The mirror this sync coordinates belongs to the syncing pi session alone:
+	// every read and write below addresses sessionStateFor(piSessionId), so a
+	// foreign session's shape-matching context can never REUSE or rebuild another
+	// session's CC file.
+	const sharedSession = sessionStateFor(piSessionId);
+	const history = nonSystemMessages(messages);
+	const priorMessages = history.slice(0, turnStart(history)); // everything before the current user turn
+
+	// REUSE path
+	//
+	// Guard on priorMessages.length >= cursor: a shorter incoming context cannot
+	// be a continuation of the cached session. This is the general invariant for
+	// pi-side history rewrites such as /compact and session_tree: without it,
+	// missed = [].slice(cursor) can falsely hit REUSE and resume an unrelated
+	// longer CC session. See issue #25.
+	if (sharedSession && !sharedSession.needsRebuild && priorMessages.length >= sharedSession.cursor) {
+		const missed = priorMessages.slice(sharedSession.cursor);
+		const trailingAssistantOnly =
+			missed.length === 1 && (missed[0] as { role?: string }).role === "assistant";
+		if (missed.length === 0 || trailingAssistantOnly) {
+		if (trailingAssistantOnly) {
+			setSessionStateFor(piSessionId, { ...sharedSession, cursor: priorMessages.length, cwd });
+			debug(`Case 3: advanced cursor past trailing assistant, resuming session ${sharedSession.sessionId.slice(0, 8)}, cursor=${priorMessages.length}`);
+		} else {
+			debug(`Case 3: resuming session ${sharedSession.sessionId.slice(0, 8)}, cursor=${sharedSession.cursor}`);
+		}
+		debug(`syncResult: path=reuse sessionId=${sharedSession.sessionId} cursor=${sharedSession?.cursor}`);
+		return { sessionId: sharedSession.sessionId };
+		}
+	}
+	// This is what keeps a caller with a pruned or short context from resuming
+	// — then overwriting — the bucket's session: shorter-than-cursor means the
+	// incoming history cannot be a continuation, so start clean and preserve.
+	// Historically this also caught reentrant subagents (a subagent's priors are
+	// shorter than the parent's cursor); with per-session mirrors it now catches
+	// the pruned-context shapes on a session's own bucket, and the non-isolated
+	// AskClaude path on the "(none)" bucket. The captured ephemeral session is
+	// deleted once its query completes (see preserveSharedSession in the
+	// completion handler).
+	//
+	// It is NOT, despite an earlier comment here, the isolated compact-summary
+	// path: runIsolatedSummary never calls syncSharedSession at all.
+	//
+	// Only reachable when needsRebuild is false — user-facing history rewrites
+	// (/compact, session_tree, /new, fork) always set needsRebuild or clear
+	// sharedSession before the next syncSharedSession call.
+	if (sharedSession && !sharedSession.needsRebuild && priorMessages.length < sharedSession.cursor) {
+		debug(`Case 1 synthetic: clean start for shorter context, preserving shared session ${sharedSession.sessionId.slice(0, 8)}, cursor=${sharedSession.cursor}`);
+		debug(`syncResult: path=clean-start preserve-shared sessionId=${sharedSession.sessionId} cursor=${sharedSession.cursor}`);
+		return { sessionId: null, preserveSharedSession: true, replay: priorMessages };
+	}
+
+	// REBUILD path
+	if (priorMessages.length === 0) {
+		setSessionStateFor(piSessionId, null);
+		debug(`Case 1: clean start, ${history.length} total messages`);
+		debug(`syncResult: path=clean-start`);
+		return { sessionId: null };
+	}
+	// Replay Hob's own history through the SDK rather than editing Claude Code files.
+	setSessionStateFor(piSessionId, null);
+	debug(`syncResult: path=replay priors=${priorMessages.length}`);
+	return { sessionId: null, replay: priorMessages };
+}
+
+// The SDK's query(), or a test double (see setQuery).
+let queryImpl: typeof query = query;
+
+// @internal
+export const __test = {
+	setQuery(fn: typeof query | null) {
+		queryImpl = fn ?? query;
+	},
+	resetSharedSession(piSessionId?: string | null) {
+		// No id: full reset (the pre-map semantics — tests start from a blank slate).
+		if (piSessionId === undefined) sharedSessions.clear();
+		else setSessionStateFor(piSessionId, null);
+		historyRewrittenBySession.clear();
+	},
+	markRebuildForSession,
+	getHistoryRewritten: () => historyRewrittenBySession.size > 0,
+	historyRewrittenBySession,
+	armStaleContexts,
+	discardRewrittenQuery,
+	contextForToolResults,
+	isQueryAbandoned: (q: object) => abandonedQueries.has(q),
+	get activeQueryContexts() {
+		return activeQueryContexts;
+	},
+	setSharedSession(piSessionId: string | null, state: SessionState | null) {
+		setSessionStateFor(piSessionId, state);
+	},
+	getSharedSession(piSessionId: string | null = null) {
+		return sessionStateFor(piSessionId);
+	},
+	setPiUI(ui: ExtensionUIContext | null) {
+		piUI = ui;
+	},
+	toBridgeContext,
+	isolatedStreamFn,
+	effortForReasoning,
+	syncSharedSession,
+	extractUserPromptBlocks,
+	replayBlocks,
+	consumeQuery,
+	finalizeCurrentStream,
+	resultErrorText,
+	deliverToolResults,
+	drainForAbort,
+	CC_CHILD_ENV,
+	buildMcpServers,
+	get promptCaptures() {
+		return promptCaptures;
+	},
+};
+
+// --- Provider helpers: tool name mapping ---
+
+// AskClaude path: CC runs its own tools, so builtin names are real.
+function mapToolName(name: string): string {
+	const normalized = name.toLowerCase();
+	const builtin = SDK_TO_PI_TOOL_NAME[normalized];
+	if (builtin) return builtin;
+	if (normalized.startsWith(MCP_TOOL_PREFIX)) return name.slice(MCP_TOOL_PREFIX.length);
+	return name;
+}
+
+// Provider path: the query runs with `tools: []`, so the only tools CC can
+// legitimately call are the pi tools we serve over MCP. Any other name is the
+// model hallucinating a builtin (`bash`, `Bash`, `Edit`, an MCP server we don't
+// serve). CC answers those itself with "No such tool available" and retries
+// inside the same query, never dispatching them to our MCP server — so a tool
+// call under such a name must not reach pi. Forwarding one ran a tool CC never
+// dispatched (real side effects) and, because the retry carries a fresh
+// tool_use id, left the handler for the retry with no result to release it:
+// pi's result arrived keyed to the dead id, and both sides deadlocked.
+function piToolNameFor(name: string, customToolNameToPi: Map<string, string>): string | undefined {
+	return customToolNameToPi.get(name) ?? customToolNameToPi.get(name.toLowerCase());
+}
+
+// Renames for Claude Code SDK param names that differ from pi's native names.
+// Keys not listed here pass through unchanged, so new pi params work automatically.
+const SDK_KEY_RENAMES: Record<string, Record<string, string>> = {
+	read:  { file_path: "path" },
+	write: { file_path: "path" },
+	edit:  { file_path: "path", old_string: "oldText", new_string: "newText", old_text: "oldText", new_text: "newText" },
+};
+
+// Maps SDK tool args to pi tool args via key renaming + pass-through.
+// Pi's own prepareArguments hooks handle any structural transforms (e.g. edit oldText/newText → edits[]).
+function mapToolArgs(
+	toolName: string, args: Record<string, unknown> | undefined,
+): Record<string, unknown> {
+	const input = args ?? {};
+	const renames = SDK_KEY_RENAMES[toolName.toLowerCase()];
+	const result: Record<string, unknown> = {};
+	for (const [key, value] of Object.entries(input)) {
+		const piKey = renames?.[key] ?? key;
+		if (!(piKey in result)) result[piKey] = value; // first alias wins
+	}
+	// Pi bash has no default timeout; add a safety default
+	if (toolName.toLowerCase() === "bash" && result.timeout == null) {
+		result.timeout = 120;
+	}
+	return result;
+}
+
+// --- Query state ---
+// QueryContext lives in query-state.js so tests can import it without
+// activating the extension.
+
+// Global (not query state):
+let piUI: ExtensionUIContext | null = null;
+let piMode: ExtensionContext["mode"] | null = null;
+const activeQueryContexts = new Set<QueryContext>();
+
+// Defaults that silently cost the user something (no Opus 1M on Max) are
+// announced once. Deferred to the first bridge query rather
+// than session_start: the notice persists a flag to the global config, and
+// firing it on startup would write that file for every pi session that merely
+// has this extension installed. One message, because consecutive info notifies
+// overwrite each other in the TUI.
+let pendingNotices: string[] = [];
+
+function showStartupNoticeOnce(): void {
+	// `hasUI` is true in RPC mode too — it means dialogs are possible, not that a
+	// human is watching. Only a terminal user can act on this.
+	if (pendingNotices.length === 0 || piMode !== "tui") return;
+	const notices = pendingNotices;
+	pendingNotices = [];
+	const path = markStartupNoticeShown();
+	// pi wraps the whole notify string in the theme's dim foreground; the inner reset
+	// drops back to the terminal default rather than dim, which is fine here.
+	const title = `\x1b[33mWelcome to pi-claude-bridge\x1b[39m — settings live in ${path}`;
+	const bullets = [...notices, "This message only appears once. See README.md for more."].map((n) => `• ${n}`);
+	piUI?.notify([title, ...bullets, "─".repeat(64)].join("\n"), "info");
+}
+
+// Captures of what pi assembled per agent; see src/prompt-capture.ts for why this
+// is keyed rather than held in a single slot. One process-wide instance, shared
+// across every extension module instance: isolated subagents re-evaluate this
+// module, and the pinned stream they all route through resolves against it.
+const promptCaptures = sharedPromptCaptures((diagnostic) => {
+	const first = diagnostic.matches[0];
+	debug(
+		`prompt-capture: no match for ${diagnostic.systemPrompt.length}-char system prompt. `
+		+ (first
+			? `closest known (${first.key.length}-char) shares its first ${first.firstDivergent} chars and diverges at offset ${first.firstDivergent}: `
+			  + JSON.stringify(diagnostic.systemPrompt.slice(first.firstDivergent - 40, first.firstDivergent + 60))
+			: "no known captures to compare against."
+		) + ` known keys=${diagnostic.matches.length}`,
+	);
+});
+
+/** Whatever a settled session left behind, named in one greppable line.
+ *
+ *  Every one of these should be empty once the last turn ends, and each is a leak
+ *  that costs something real: a retained context routes a later orphaned tool result
+ *  into the delivery path and returns a stream nobody ends; a pending tool call is an
+ *  MCP handler Claude Code is still waiting on; a live prompt stream is an unresolved
+ *  ack. The activeQueryContexts leak was present on every single happy-path run and
+ *  no test noticed, because nothing asserted that anything ends clean — so assert it
+ *  where the real sessions are, and let diag/audit-warnings.mjs scan for it. */
+function reportLeaks(label: string): void {
+	const pendingCalls = [...activeQueryContexts].reduce((n, c) => n + c.pendingToolCalls.size, 0);
+	const liveStreams = [...activeQueryContexts].filter((c) => c.promptStream !== null).length;
+	if (activeQueryContexts.size === 0 && pendingCalls === 0 && liveStreams === 0) return;
+	debug(
+		`WARNING: ${label} left state behind — contexts=${activeQueryContexts.size} `
+		+ `pendingToolCalls=${pendingCalls} promptStreams=${liveStreams}`,
+	);
+}
+
+function contextForToolResults(results: McpResult[]): QueryContext | undefined {
+	for (const result of results) {
+		const id = result.toolCallId;
+		if (!id) continue;
+		for (const queryCtx of activeQueryContexts) {
+			if (queryCtx.pendingToolCalls.has(id) || queryCtx.pendingResults.has(id) || queryCtx.turnToolCallIds.includes(id)) {
+				return queryCtx;
+			}
+		}
+	}
+	return undefined;
+}
+
+function resolveMcpTools(context: Context, excludeToolName?: string): {
+	mcpTools: Tool[];
+	customToolNameToSdk: Map<string, string>;
+	customToolNameToPi: Map<string, string>;
+} {
+	const mcpTools: Tool[] = [];
+	const customToolNameToSdk = new Map<string, string>();
+	const customToolNameToPi = new Map<string, string>();
+
+	if (!context.tools) return { mcpTools, customToolNameToSdk, customToolNameToPi };
+
+	for (const tool of context.tools) {
+		if (tool.name === excludeToolName) continue;
+		const sdkName = `${MCP_TOOL_PREFIX}${tool.name}`;
+		mcpTools.push(tool);
+		customToolNameToSdk.set(tool.name, sdkName);
+		customToolNameToSdk.set(tool.name.toLowerCase(), sdkName);
+		customToolNameToPi.set(sdkName, tool.name);
+		customToolNameToPi.set(sdkName.toLowerCase(), tool.name);
+	}
+
+	return { mcpTools, customToolNameToSdk, customToolNameToPi };
+}
+
+// Creates an MCP server that bridges pi tools to the SDK. Each tool handler
+// blocks on a Promise until pi delivers the tool result via streamSimple.
+// Handlers receive their toolCallId from Claude's tools/call _meta, so results
+// are matched by ID end to end.
+//
+// The handler and pi's result can arrive in either order, hence the two maps:
+// a result that lands first waits in `pendingResults` for the handler to claim
+// it, and a handler that runs first parks its resolver in `pendingToolCalls`.
+// Handlers close over the captured `queryCtx`, ensuring they operate on the
+// correct query's state while multiple queries run concurrently.
+function buildMcpServers(tools: Tool[], queryCtx: QueryContext): Record<string, ReturnType<typeof createToolServer>> | undefined {
+	if (!tools.length) return undefined;
+	const mcpTools = tools.map((tool) => ({
+		name: tool.name,
+		description: tool.description,
+		inputSchema: tool.parameters,
+		handler: async (toolCallId: string) => {
+			if (queryCtx.pendingResults.has(toolCallId)) {
+				const result = queryCtx.pendingResults.get(toolCallId)!;
+				queryCtx.pendingResults.delete(toolCallId);
+				debug(`mcp handler: ${tool.name} [${toolCallId}] → resolved from queue (${queryCtx.pendingResults.size} remaining)`);
+				return result;
+			}
+			debug(`mcp handler: ${tool.name} [${toolCallId}] → waiting`);
+			return new Promise<McpResult>((resolve) => {
+				queryCtx.pendingToolCalls.set(toolCallId, { toolName: tool.name, resolve });
+			});
+		},
+	}));
+	return { [MCP_SERVER_NAME]: createToolServer(MCP_SERVER_NAME, mcpTools) };
+}
+
+// --- Usage helpers ---
+
+// The counter mapping lives in usage.ts; the debug line is this side's job, so every
+// call site logs the same way rather than three times over.
+function recordUsage(output: AssistantMessage, usage: SdkUsage, model: Model<any>): void {
+	const { cachePct, reasoning } = updateUsage(output, usage, model);
+	const reasoningText = reasoning == null ? "" : ` reasoning=${reasoning}`;
+	debug(`usage: in=${output.usage.input} out=${output.usage.output} cacheRead=${output.usage.cacheRead} cacheWrite=${output.usage.cacheWrite} total=${output.usage.totalTokens}${reasoningText} cachePct=${cachePct}% model=${model.id}`);
+}
+
+// Log the *served* context window reported by an SDK result message
+// (modelUsage[id].contextWindow), which can differ from the window pi
+// registered (model.contextWindow) when the runtime entitlement doesn't
+// match the docs — e.g. bare Opus served 200K on Pro, or [1m] not honored.
+// The result message's modelUsage is otherwise discarded; this makes the
+// gap observable. See issue #18.
+function logServedContextWindow(label: string, message: SDKMessage, model: Model<any>): void {
+	const modelUsage = (message as any).modelUsage as Record<string, { contextWindow?: number; maxOutputTokens?: number }> | undefined;
+	if (!modelUsage) return;
+	for (const [k, v] of Object.entries(modelUsage)) {
+		debug(`${label}: served contextWindow=${v.contextWindow ?? "?"} maxOutputTokens=${v.maxOutputTokens ?? "?"} servedModel=${k} registered=${model.contextWindow}`);
+	}
+}
+
+// --- Effort level mapping ---
+// Pi reasoning levels → CC SDK effort levels
+
+const REASONING_TO_EFFORT: Record<string, EffortLevel> = {
+	minimal: "low", low: "low", medium: "medium", high: "high", xhigh: "max", max: "max",
+};
+
+const VALID_EFFORTS = new Set<string>(["low", "medium", "high", "xhigh", "max"]);
+
+// Prefer the model's own thinkingLevelMap (per-model overrides — e.g. a map can
+// route xhigh→xhigh where the generic table maps xhigh→max). pi-ai's catalog
+// ships a map for most Claude models; the table above covers models without
+// one, and the levels a map leaves unnamed. A null entry means the level is
+// unsupported on that model: no effort argument is sent, so Claude Code's own
+// default applies rather than the generic table's value. Map values are
+// provider-generic strings, so a map value is trusted only when it names a
+// level CC accepts.
+function effortForReasoning(model: Model<any>, reasoning: string | undefined): EffortLevel | undefined {
+	if (!reasoning) return undefined;
+	const mapped = model.thinkingLevelMap?.[reasoning];
+	if (mapped === undefined) return REASONING_TO_EFFORT[reasoning];
+	return VALID_EFFORTS.has(mapped as EffortLevel) ? mapped as EffortLevel : undefined;
+}
+
+// --- Provider helpers: misc ---
+
+function mapStopReason(reason: string | undefined): "stop" | "length" | "toolUse" {
+	switch (reason) {
+		case "tool_use": return "toolUse";
+		case "max_tokens": return "length";
+		case "end_turn": default: return "stop";
+	}
+}
+
+function parsePartialJson(input: string, fallback: Record<string, unknown>): Record<string, unknown> {
+	if (!input) return fallback;
+	try { return JSON.parse(input); } catch { return fallback; }
+}
+
+
+// --- Provider: streaming function ---
+//
+// Push-based streaming with MCP tool bridge:
+// 1. streamSimple starts a query() and kicks off consumeQuery() in background
+// 2. consumeQuery() iterates the SDK generator, pushing events to currentPiStream
+// 3. On tool_use: ends the current pi stream, nulls it out. The MCP handler
+//    blocks the generator naturally — no events arrive until resolved.
+// 4. Pi executes the tool, calls streamSimple again. We swap in the new stream,
+//    resolve the MCP handler, and the generator unblocks — events flow to new stream.
+//
+// Note: resetTurnState clears turnSawStreamEvent while the generator may still
+// have queued messages from the previous turn. This is safe because step 3 nulls
+// currentPiStream, so any leftover messages hit the `!ctx().currentPiStream` guard
+// in consumeQuery and are skipped before resetTurnState runs.
+
+const completedStreams = new WeakSet<object>();
+
+function markStreamComplete(stream: AssistantMessageEventStream | null): void {
+	if (stream) completedStreams.add(stream as object);
+}
+
+function claimCurrentPiStream(stream: AssistantMessageEventStream, label: string, c: QueryContext, options: SimpleStreamOptions | undefined): void {
+	if (c.currentPiStream && !completedStreams.has(c.currentPiStream as object)) {
+		debug(`WARNING: currentPiStream overwritten before terminal event (${label}); activeQuery=${Boolean(c.activeQuery)} pendingHandlers=${c.pendingToolCalls.size}`);
+	}
+	c.currentPiStream = stream;
+	c.currentOnResponse = options?.onResponse;
+}
+
+function ensureTurnStarted(c: QueryContext): void {
+	if (!c.turnStarted && c.currentPiStream && c.turnOutput) {
+		c.currentPiStream!.push({ type: "start", partial: c.turnOutput });
+		c.turnStarted = true;
+	}
+}
+
+function finalizeCurrentStream(c: QueryContext, stopReason?: string): void {
+	if (!c.currentPiStream || !c.turnOutput) return;
+	debug(`provider: finalizeCurrentStream called, stopReason=${stopReason}, turnOutput=${JSON.stringify({stopReason: c.turnOutput!.stopReason, error: c.turnOutput!.errorMessage})}`);
+	if (!c.turnStarted) ensureTurnStarted(c);
+	const stream = c.currentPiStream;
+	if (c.turnOutput.stopReason === "error") {
+		stream!.push({ type: "error", reason: "error", error: c.turnOutput });
+	} else {
+		const reason = stopReason === "length" ? "length" : "stop";
+		stream!.push({ type: "done", reason, message: c.turnOutput });
+	}
+	markStreamComplete(stream);
+	stream!.end();
+	c.currentPiStream = null;
+}
+
+/** Maps Anthropic stream events to pi stream events (text, thinking, toolcall).
+ *  On message_stop with tool_use: ends currentPiStream so pi can execute the tool. */
+function processStreamEvent(
+	message: SDKMessage,
+	customToolNameToPi: Map<string, string>,
+	model: Model<any>,
+	c: QueryContext,
+): void {
+	if (!c.currentPiStream || !c.turnOutput) return;
+	c.turnSawStreamEvent = true;
+	const event = (message as SDKMessage & { event: any }).event;
+
+	if (event?.type === "message_start") {
+		// Still open from an earlier message_start: Claude Code gave up on that
+		// stream and is retrying it. Its blocks were never completed.
+		if (c.turnStreamOpen) dropAbandonedStreamBlocks(c, `restreamed as ${event.message?.id}`);
+		c.turnToolCallIds = [];
+		c.turnStreamMessageId = event.message?.id;
+		c.turnStreamOpen = true;
+		c.turnStreamBlockStart = c.turnBlocks.length;
+		if (event.message?.usage) recordUsage(c.turnOutput, event.message.usage, model);
+		return;
+	}
+
+	if (event?.type === "content_block_start") {
+		ensureTurnStarted(c);
+		if (event.content_block?.type === "text") {
+			c.turnBlocks.push({ type: "text", text: "", index: event.index });
+			c.currentPiStream!.push({ type: "text_start", contentIndex: c.turnBlocks.length - 1, partial: c.turnOutput });
+		} else if (event.content_block?.type === "thinking") {
+			c.turnBlocks.push({ type: "thinking", thinking: "", thinkingSignature: "", index: event.index });
+			c.currentPiStream!.push({ type: "thinking_start", contentIndex: c.turnBlocks.length - 1, partial: c.turnOutput });
+		} else if (event.content_block?.type === "tool_use") {
+			const piName = piToolNameFor(event.content_block.name, customToolNameToPi);
+			if (!piName) {
+				debug(`processStreamEvent: skipping tool_use for unserved tool ${event.content_block.name} [${event.content_block.id}] — CC rejects it and retries`);
+				return;
+			}
+			c.turnToolCallIds.push(event.content_block.id);
+			c.turnBlocks.push({
+				type: "toolCall", id: event.content_block.id,
+				name: piName,
+				arguments: (event.content_block.input as Record<string, unknown>) ?? {},
+				partialJson: "", index: event.index,
+			});
+			c.currentPiStream!.push({ type: "toolcall_start", contentIndex: c.turnBlocks.length - 1, partial: c.turnOutput });
+		} else {
+			debug("processStreamEvent: unhandled content_block_start type", event.content_block?.type);
+		}
+		return;
+	}
+
+	if (event?.type === "content_block_delta") {
+		const index = c.turnBlocks.findIndex((b: any) => b.index === event.index);
+		const block = c.turnBlocks[index];
+		if (!block) return;
+		if (event.delta?.type === "text_delta" && block.type === "text") {
+			block.text += event.delta.text;
+			c.currentPiStream!.push({ type: "text_delta", contentIndex: index, delta: event.delta.text, partial: c.turnOutput });
+		} else if (event.delta?.type === "thinking_delta" && block.type === "thinking") {
+			block.thinking += event.delta.thinking;
+			c.currentPiStream!.push({ type: "thinking_delta", contentIndex: index, delta: event.delta.thinking, partial: c.turnOutput });
+		} else if (event.delta?.type === "input_json_delta" && block.type === "toolCall") {
+			block.partialJson += event.delta.partial_json;
+			block.arguments = parsePartialJson(block.partialJson, block.arguments);
+			c.currentPiStream!.push({ type: "toolcall_delta", contentIndex: index, delta: event.delta.partial_json, partial: c.turnOutput });
+		} else if (event.delta?.type === "signature_delta" && block.type === "thinking") {
+			block.thinkingSignature = (block.thinkingSignature ?? "") + event.delta.signature;
+		} else {
+			debug("processStreamEvent: unhandled content_block_delta type", event.delta?.type);
+		}
+		return;
+	}
+
+	if (event?.type === "content_block_stop") {
+		const index = c.turnBlocks.findIndex((b: any) => b.index === event.index);
+		const block = c.turnBlocks[index];
+		if (!block) return;
+		delete block.index;
+		if (block.type === "text") {
+			c.currentPiStream!.push({ type: "text_end", contentIndex: index, content: block.text, partial: c.turnOutput });
+		} else if (block.type === "thinking") {
+			c.currentPiStream!.push({ type: "thinking_end", contentIndex: index, content: block.thinking, partial: c.turnOutput });
+		} else if (block.type === "toolCall") {
+			c.turnSawToolCall = true;
+			block.arguments = mapToolArgs(
+				block.name, parsePartialJson(block.partialJson, block.arguments),
+			);
+			delete block.partialJson;
+			c.currentPiStream!.push({ type: "toolcall_end", contentIndex: index, toolCall: block, partial: c.turnOutput });
+		}
+		return;
+	}
+
+	if (event?.type === "message_delta") {
+		c.turnOutput.stopReason = mapStopReason(event.delta?.stop_reason);
+		if (event.usage) recordUsage(c.turnOutput, event.usage, model);
+		return;
+	}
+
+	if (event?.type === "message_stop") {
+		dropUnfinishedStreamBlocks(c, `message_stop for ${c.turnStreamMessageId}`);
+		c.turnStreamOpen = false;
+	}
+
+	if (event?.type === "message_stop" && c.turnSawToolCall) {
+		// Tool call complete — end this pi stream. The SDK will still yield an
+		// assistant message for this turn, but currentPiStream=null causes
+		// consumeQuery to skip it. The MCP handler blocks the generator until
+		// pi delivers the tool result via the next streamSimple call.
+		c.turnOutput.stopReason = "toolUse";
+		const stream = c.currentPiStream;
+		stream!.push({ type: "done", reason: "toolUse", message: c.turnOutput });
+		markStreamComplete(stream);
+		stream!.end();
+		c.currentPiStream = null;
+
+		// Cursor is updated by the next streamSimple call (tool result delivery path)
+		// which sets cursor = context.messages.length with the post-tool-result context.
+		return;
+	}
+
+	if (event?.type !== "message_stop" && event?.type !== "ping") {
+		debug("processStreamEvent: unhandled event type", event?.type);
+	}
+}
+
+/** Remove the blocks of the message that just stopped which never got their own
+ *  content_block_stop, because the stream was cut inside them. A tool call left
+ *  that way carries whatever JSON survived the cut — `{}` when no delta landed at
+ *  all — under an id Claude Code drops from its own conversation before resuming
+ *  the turn, so it will never dispatch that id and nothing could answer a result
+ *  keyed to it; pi would reject the call, and its result would sit unmatched while
+ *  the resumed call waited in a handler. A thinking block left that way has no
+ *  signature.
+ *
+ *  Per block rather than per message: Claude Code keeps the blocks that did
+ *  complete, tool calls included, and one it dispatched while pi was never told is
+ *  the same deadlock the other way round. Nothing here ends the pi turn — with no
+ *  finished tool call left the resume is more of this same turn, and a turn ended
+ *  here would leave the real call with no stream to arrive on. */
+function dropUnfinishedStreamBlocks(c: QueryContext, why: string): void {
+	const dropped: string[] = [];
+	for (let i = c.turnBlocks.length - 1; i >= c.turnStreamBlockStart; i--) {
+		// `index` is deleted at content_block_stop, so it marks a block that never closed.
+		if (c.turnBlocks[i].index === undefined) continue;
+		dropped.push(`${c.turnBlocks[i].type}${c.turnBlocks[i].id ? `[${c.turnBlocks[i].id}]` : ""}`);
+		c.turnBlocks.splice(i, 1);
+	}
+	if (!dropped.length) return;
+	debug(`dropUnfinishedStreamBlocks: ${why}; dropped ${dropped.reverse().join(",")}`);
+	c.turnToolCallIds = c.turnBlocks.filter((b: any) => b.type === "toolCall").map((b: any) => b.id);
+	c.turnSawToolCall = c.turnBlocks.some((b: any) => b.type === "toolCall");
+}
+
+/** Remove the blocks a stream Claude Code abandoned mid-message. Their own
+ *  content_block_stop never came, so a thinking block has no signature and a tool
+ *  call is one CC will never dispatch; left in, pi would run the tool and the turn
+ *  would wait on a handler that never comes, or the next request would replay a
+ *  broken block.
+ *  The fallback then restarts those indices. pi's normal provider path tolerates that;
+ *  pi-agent-core's experimental harness frame encoder keys blocks by contentIndex and
+ *  rejects a repeated start, so it would need a change there to drive this provider. */
+function dropAbandonedStreamBlocks(c: QueryContext, why: string): void {
+	const dropped = c.turnBlocks.splice(c.turnStreamBlockStart);
+	debug(`dropAbandonedStreamBlocks: ${why}; dropped ${dropped.length} blocks from ${c.turnStreamMessageId} types=${dropped.map((b: any) => b.type).join(",")}`);
+	c.turnToolCallIds = [];
+	c.turnSawToolCall = c.turnBlocks.some((b: any) => b.type === "toolCall");
+	c.turnStreamOpen = false;
+}
+
+// The SDK always yields `assistant` messages (completed content blocks) after streaming.
+// When stream_events already delivered the content, this is a no-op. But after
+// resetTurnState (e.g. tool result delivery), if the next turn's assistant message
+// arrives before any stream_events, this is the primary content path. Must maintain
+// the same stream lifecycle as processStreamEvent — including ending the stream on
+// tool_use to prevent deadlock with the MCP handler.
+//
+// It is also the content path when a stream stalls: Claude Code drops it and asks
+// again without streaming ("Error streaming, falling back to non-streaming mode"),
+// and the answer arrives as one assistant message, under a new message id, with no
+// stream_events of its own. turnSawStreamEvent is already set by the dead stream,
+// so gating on it alone dropped that message: its tool calls never reached pi, CC
+// sat in the MCP handler waiting for their results, and the turn hung on "Working"
+// until the user aborted it.
+function processAssistantMessage(message: SDKMessage, model: Model<any>, customToolNameToPi: Map<string, string>, c: QueryContext): void {
+	const assistantMsg = (message as any).message;
+	if (!assistantMsg?.content) return;
+	// Claude Code reports API failures and its own quota/context notices as a
+	// `<synthetic>` assistant message: a report, not model output. Streaming its text
+	// pins the turn's output before the failure it describes, and a consumer that only
+	// fails over before output commits (pi-model-fallback-alias) then cannot reach the
+	// next provider. Keep the wording on the failed turn — the error event carries it —
+	// but emit no events, so the turn still reads as a call that produced no output.
+	// Issue #162.
+	if (assistantMsg.model === "<synthetic>") {
+		// The report can follow a stalled stream whose non-streaming retry also failed;
+		// drop the abandoned partial blocks (unsigned thinking, a tool call CC will never
+		// dispatch) the way the fallback path below would.
+		if (c.turnSawStreamEvent && c.turnStreamOpen) dropAbandonedStreamBlocks(c, "synthetic failure report");
+		debug(`processAssistantMessage: <synthetic> message, keeping ${assistantMsg.content.length} block(s) off the stream`);
+		for (const block of assistantMsg.content) {
+			if (block.type === "text" && block.text) c.turnBlocks.push({ type: "text", text: block.text });
+			else debug("processAssistantMessage: unhandled <synthetic> block type", block.type);
+		}
+		if (assistantMsg.usage && c.turnOutput) recordUsage(c.turnOutput, assistantMsg.usage, model);
+		return;
+	}
+	if (c.turnSawStreamEvent) {
+		// Same id was already delivered; a new id is CC's non-streaming fallback.
+		// Drop the stalled stream's partial blocks if it never stopped. Deliberately
+		// deliver even if it stopped, at the risk of duplication if CC renumbers it.
+		const id = assistantMsg.id;
+		if (!id || !c.turnStreamMessageId || id === c.turnStreamMessageId) return;
+		if (c.turnStreamOpen) dropAbandonedStreamBlocks(c, `non-streaming fallback ${id}`);
+	}
+	c.turnToolCallIds = [];
+	debug(`processAssistantMessage fallback: ${assistantMsg.content.length} blocks, types=${assistantMsg.content.map((b: any) => b.type).join(",")}`);
+	for (const block of assistantMsg.content) {
+		if (block.type === "text" && block.text) {
+			ensureTurnStarted(c);
+			c.turnBlocks.push({ type: "text", text: block.text });
+			const idx = c.turnBlocks.length - 1;
+			c.currentPiStream?.push({ type: "text_start", contentIndex: idx, partial: c.turnOutput });
+			c.currentPiStream?.push({ type: "text_delta", contentIndex: idx, delta: block.text, partial: c.turnOutput });
+			c.currentPiStream?.push({ type: "text_end", contentIndex: idx, content: block.text, partial: c.turnOutput });
+		} else if (block.type === "thinking") {
+			ensureTurnStarted(c);
+			c.turnBlocks.push({ type: "thinking", thinking: block.thinking ?? "", thinkingSignature: block.signature ?? "" });
+			const idx = c.turnBlocks.length - 1;
+			c.currentPiStream?.push({ type: "thinking_start", contentIndex: idx, partial: c.turnOutput });
+			if (block.thinking) c.currentPiStream?.push({ type: "thinking_delta", contentIndex: idx, delta: block.thinking, partial: c.turnOutput });
+			c.currentPiStream?.push({ type: "thinking_end", contentIndex: idx, content: block.thinking ?? "", partial: c.turnOutput });
+		} else if (block.type === "tool_use") {
+			const piName = piToolNameFor(block.name, customToolNameToPi);
+			if (!piName) {
+				debug(`processAssistantMessage: skipping tool_use for unserved tool ${block.name} [${block.id}] — CC rejects it and retries`);
+				continue;
+			}
+			ensureTurnStarted(c);
+			c.turnSawToolCall = true;
+			c.turnToolCallIds.push(block.id);
+			c.turnBlocks.push({
+				type: "toolCall", id: block.id,
+				name: piName,
+				arguments: mapToolArgs(piName, block.input),
+			});
+			const idx = c.turnBlocks.length - 1;
+			const toolBlock = c.turnBlocks[idx];
+			c.currentPiStream?.push({ type: "toolcall_start", contentIndex: idx, partial: c.turnOutput });
+			c.currentPiStream?.push({ type: "toolcall_end", contentIndex: idx, toolCall: toolBlock as any, partial: c.turnOutput });
+		} else {
+			debug("processAssistantMessage: unhandled block type", block.type);
+		}
+	}
+	if (assistantMsg.usage && c.turnOutput) recordUsage(c.turnOutput, assistantMsg.usage, model);
+
+	// End the stream on tool_use, same as processStreamEvent's message_stop handler.
+	if (c.turnSawToolCall && c.currentPiStream && c.turnOutput) {
+		c.turnOutput.stopReason = "toolUse";
+		const stream = c.currentPiStream;
+		stream.push({ type: "done", reason: "toolUse", message: c.turnOutput });
+		markStreamComplete(stream);
+		stream.end();
+		c.currentPiStream = null;
+	}
+}
+
+/** Background consumer: iterates the SDK generator, pushing events to currentPiStream.
+ *  Runs until the query ends. Per turn, the SDK yields stream_events (deltas), then
+ *  an assistant message (completed blocks). On tool_use, the stream is ended by
+ *  whichever path handles it first (processStreamEvent or processAssistantMessage),
+ *  and the MCP handler blocks the generator until pi delivers the tool result. */
+async function consumeQuery(
+	sdkQuery: ReturnType<typeof query>,
+	customToolNameToPi: Map<string, string>,
+	model: Model<any>,
+	wasAborted: () => boolean,
+	queryCtx: QueryContext,
+): Promise<{ capturedSessionId?: string }> {
+	let capturedSessionId: string | undefined;
+
+	for await (const message of sdkQuery) {
+		// CLAUDE_BRIDGE_RECORD_STREAM=<path> appends every SDK message consumeQuery
+		// sees, one JSON object per line. Used by tests/lib/record-sdk-streams.mjs to
+		// capture replay fixtures, so unit tests assert against message shapes Claude
+		// Code really emitted rather than ones we imagined.
+		const recordStreamPath = process.env.CLAUDE_BRIDGE_RECORD_STREAM;
+		if (recordStreamPath) appendFileSync(recordStreamPath, `${JSON.stringify(message)}\n`);
+		if (wasAborted()) break;
+		await reportHeldRateLimit(queryCtx, model);
+		// Everything below the currentPiStream guard is content, which there is
+		// nowhere to put once a turn has ended on a tool call. These three are not
+		// content and must not share that gate:
+		//
+		// - stdin: nothing else closes the CLI's stdin now that the prompt is a
+		//   streamed generator (isSingleUserTurn=false), so missing this hangs the query.
+		// - the failure a `result` carries: it is the only record that the turn
+		//   failed at all. Behind the guard, a 429 arriving at a tool boundary set
+		//   no stopReason, no errorMessage, and logged nothing — the turn simply
+		//   ended empty.
+		// - rate-limit events: notifications to the user, which are most likely to
+		//   fire during exactly the long tool-using turns the guard was skipping.
+		let resultError: string | undefined;
+		if (message.type === "result") {
+			queryCtx.promptStream?.end();
+			logServedContextWindow("result", message, model);
+			resultError = resultErrorText(message);
+			if (resultError !== undefined) {
+				// Consume the rejection alongside the failure it caused, so a later
+				// unrelated failure on this query doesn't inherit the label.
+				if (queryCtx.rateLimitRejection) {
+					resultError = describeRateLimitFailure(queryCtx.rateLimitRejection, resultError);
+					queryCtx.rateLimitRejection = null;
+				}
+				debug(`consumeQuery: error result, subtype=${message.subtype}, error=${resultError}`);
+				if (queryCtx.turnOutput) {
+					queryCtx.turnOutput.stopReason = "error";
+					queryCtx.turnOutput.errorMessage = resultError;
+				}
+			}
+		}
+		if (message.type === "rate_limit_event") {
+			const info = (message as any).rate_limit_info;
+			debug("consumeQuery: rate_limit_event", JSON.stringify(info).slice(0, 300));
+			if (info) queryCtx.heldRateLimit = info;
+			await reportHeldRateLimit(queryCtx, model);
+			if (info?.status === "rejected") {
+				// Held so the failure Claude Code sends next can be named as a rate limit.
+				queryCtx.rateLimitRejection = info;
+				// The "rate limit reached" notice below supersedes warnings; re-arm so the next
+				// window's warnings fire even if it opens straight into allowed_warning.
+				queryCtx.lastRateLimitWarnStep = null;
+				queryCtx.lastRateLimitWarnThreshold = undefined;
+				// resetsAt is Unix seconds, not milliseconds.
+				const resetsAt = info.resetsAt ? new Date(info.resetsAt * 1000).toLocaleTimeString() : "unknown";
+				piUI?.notify(`Claude subscription rate limit reached (${info.rateLimitType ?? "unknown"}) — resets at ${resetsAt}`, "warning");
+			} else if (info?.status === "allowed") {
+				// Back under the threshold (window reset) — re-arm the warning dedupe.
+				queryCtx.lastRateLimitWarnStep = null;
+				queryCtx.lastRateLimitWarnThreshold = undefined;
+			} else if (info?.status === "allowed_warning") {
+				// utilization is a fraction (0..1); allowed_warning fires once it crosses surpassedThreshold.
+				const percent = Math.round((info.utilization ?? 0) * 100);
+				// The SDK emits one event per request, so only re-notify when the level
+				// rises past a new 5% step or the threshold changes.
+				const step = Math.floor(percent / 5);
+				const rose = queryCtx.lastRateLimitWarnStep === null || step > queryCtx.lastRateLimitWarnStep;
+				if (rose || info.surpassedThreshold !== queryCtx.lastRateLimitWarnThreshold) {
+					queryCtx.lastRateLimitWarnStep = step;
+					queryCtx.lastRateLimitWarnThreshold = info.surpassedThreshold;
+					piUI?.notify(`Claude rate limit warning: ${percent}% used (${info.rateLimitType ?? ""})`, "warning");
+				}
+			}
+			continue;
+		}
+		if (!queryCtx.currentPiStream || !queryCtx.turnOutput) continue;
+
+		switch (message.type) {
+			case "stream_event":
+				processStreamEvent(message, customToolNameToPi, model, queryCtx);
+				break;
+			case "assistant":
+				processAssistantMessage(message, model, customToolNameToPi, queryCtx);
+				break;
+			case "result": {
+					// The failure itself was recorded above the guard, along with the served
+					// context window. What is left here is the success path: push the result
+					// text when no assistant message already delivered it.
+					if (resultError === undefined && !queryCtx.turnSawStreamEvent && message.subtype === "success") {
+					ensureTurnStarted(queryCtx);
+					const text = message.result || "";
+					queryCtx.turnBlocks.push({ type: "text", text });
+					const idx = queryCtx.turnBlocks.length - 1;
+					queryCtx.currentPiStream?.push({ type: "text_start", contentIndex: idx, partial: queryCtx.turnOutput });
+					queryCtx.currentPiStream?.push({ type: "text_delta", contentIndex: idx, delta: text, partial: queryCtx.turnOutput });
+					queryCtx.currentPiStream?.push({ type: "text_end", contentIndex: idx, content: text, partial: queryCtx.turnOutput });
+				}
+				break;
+			}
+			case "system":
+				if ((message as any).subtype === "init" && (message as any).session_id) {
+					capturedSessionId = (message as any).session_id;
+				}
+				break;
+			case "user":
+				// SDK echo of the user prompt — no stream events to emit. Note it
+				// carries only prompts and tool results: a steer CC drained at a
+				// tool boundary is recorded in its session transcript as a
+				// `queued_command` attachment and never reaches this stream, which
+				// is why the mid-turn steering tripwire has to live in the
+				// integration test.
+				break;
+			default:
+				debug("consumeQuery: unhandled SDK message type", message.type);
+				break;
+		}
+	}
+
+	// DEBUG: trace when consumeQuery exits
+	debug(`consumeQuery: for-await loop exited, wasAborted=${wasAborted()}, capturedSessionId=${capturedSessionId?.slice(0, 8) ?? "none"}`);
+
+	return { capturedSessionId };
+}
+
+/** The trailing user turn as content blocks, or null if there isn't one.
+ *  Blocks rather than text so image steers keep their images. */
+function steerBlocks(messages: Context["messages"]): ContentBlockParam[] | null {
+	const blocks = extractUserPromptBlocks(messages);
+	if (blocks) return blocks;
+	const text = extractUserPrompt(messages);
+	return text ? [{ type: "text", text }] : null;
+}
+
+/** A steer that never made it into CC's session. The cursor has already counted
+ *  it, so count-based sync would skip it forever — rebuild instead, which
+ *  re-imports the message from pi's context. */
+function steerMissedSession(c: QueryContext, text: string): void {
+	c.missedSteer = true;
+	const state = sessionStateFor(c.piSessionId);
+	if (state) setSessionStateFor(c.piSessionId, { ...state, needsRebuild: true });
+	debug(`provider: steer never reached CC, marked query for rebuild: ${text.slice(0, 60)}`);
+}
+
+/** Releases this turn's tool results to their MCP handlers, after first pushing
+ *  any steer to CC.
+ *
+ *  The ordering is mandatory, not an optimization. The steer and the MCP tool
+ *  result travel back to CC over the same stdin FIFO. Awaiting the push ack
+ *  (which resolves only once the SDK's write to stdin completed) before
+ *  resolving any handler guarantees CC enqueues the steer *before* it reads the
+ *  tool result, so its post-tool-call drain sees it and acts on it this turn.
+ *  Resolve first and the steer misses the drain, silently degrading to
+ *  follow-up semantics.
+ *
+ *  Both the post-tool-call drain and the FIFO ordering are CC CLI internals,
+ *  not SDK contract — tests/int-tool-message.mjs is the tripwire if they move. */
+async function deliverToolResults(
+	c: QueryContext,
+	results: McpResult[],
+	steer: ContentBlockParam[] | null,
+	contextLength: number,
+): Promise<void> {
+	if (steer) {
+		const text = steer.map((b) => (b.type === "text" ? b.text : "[image]")).join("\n");
+		if (!c.promptStream) {
+			debug(`WARNING: steer with no prompt stream, dropping: ${text.slice(0, 60)}`);
+			steerMissedSession(c, text);
+		} else {
+			try {
+				await c.promptStream.push(userMessage(steer, "next"));
+				debug(`provider: steer written to CC stdin before tool result: ${text.slice(0, 60)}`);
+			} catch (error) {
+				// The query is ending — pushing further input would wedge tool-result
+				// delivery, so the steer doesn't reach this query. It is still in
+				// pi's context, and the caller has already advanced the session
+				// cursor past it, so force a rebuild or CC would never see it.
+				debug(`provider: steer push rejected, delivering tool result anyway:`, error);
+				steerMissedSession(c, text);
+			}
+		}
+	}
+
+	debug(`provider: tool results, ${results.length} results, ${c.pendingToolCalls.size} waiting handlers, ctx.msgs=${contextLength}`);
+	for (const result of results) {
+		const id = result.toolCallId;
+		if (id && c.pendingToolCalls.has(id)) {
+			const pending = c.pendingToolCalls.get(id)!;
+			c.pendingToolCalls.delete(id);
+			debug(`provider: resolving ${pending.toolName} [${id}]${result.isError ? " (error)" : ""}`, JSON.stringify(result.content).slice(0, 200));
+			pending.resolve(result);
+		} else if (id) {
+			c.pendingResults.set(id, result);
+			debug(`provider: queued result [${id}] (${c.pendingResults.size} pending)`);
+		} else {
+			debug(`WARNING: tool result without toolCallId, cannot match`);
+		}
+		if (c.pendingToolCalls.size > 0 && c.pendingResults.size > 0) {
+			debug(`BUG: both maps non-empty! handlers=${c.pendingToolCalls.size} results=${c.pendingResults.size}`);
+		}
+	}
+	if (c.pendingToolCalls.size > 0) {
+		debug(`WARNING: ${c.pendingToolCalls.size} MCP handlers still waiting after delivering ${results.length} results`);
+		piUI?.notify(`Claude bridge: ${c.pendingToolCalls.size} tool handler(s) still waiting — provider may be stuck`, "warning");
+	}
+}
+
+/** Abort teardown for one query: settle everything that would otherwise be left
+ *  awaiting a subprocess we are about to kill. The pump abandons iteration on
+ *  abort, so an in-flight prompt-stream push would hang forever and take
+ *  tool-result delivery with it. */
+function drainForAbort(c: QueryContext, promptStream: PromptStream): void {
+	promptStream.fail(new Error("Operation aborted"));
+	c.releasePendingToolCalls("Operation aborted");
+}
+
+/** Queries pi's history moved out from under. Their completion must not touch
+ *  `sharedSession` or the pi stream: the query that took over the turn has
+ *  already rebuilt both from the new history, and this one's session id names the
+ *  conversation pi just discarded. */
+const abandonedQueries = new WeakSet<object>();
+
+/** The prompt a continuation query is opened with: the pi turn goes on, but its
+ *  last message is a tool result rather than a prompt, and query() cannot resume
+ *  a session without one. */
+const CONTINUE_AFTER_REWRITE_PROMPT =
+	"[Your context was compacted. What precedes this is a summary plus the most recent messages, "
+	+ "ending with the tool result you were waiting for. Continue the task from there.]";
+
+/** The prompt a resumed tool result is opened with: no query of this process
+ *  was parked on it (a restart, a crash, a stall), so the turn goes on from
+ *  history that ends at the result. */
+const CONTINUE_AFTER_RESUME_PROMPT =
+	"[Your session was interrupted and resumed. What precedes this ends with the tool result "
+	+ "you were waiting for. Continue the task from there.]";
+
+/** Drop a Claude Code query parked at a tool boundary whose conversation pi has
+ *  since rewritten (/compact, tree navigation).
+ *
+ *  Delivering the turn's tool result into that query hands Claude Code the
+ *  context pi just shrank: one pi turn is one CC query, and the query keeps its
+ *  own context inside the CLI whatever pi does to its transcript. It answers off
+ *  the pre-compaction conversation, reports the pre-compaction usage back, and pi
+ *  crosses the same threshold at the next boundary — measured as one compaction
+ *  per tool call with usage never dropping (issue #101). `needsRebuild` does not
+ *  prevent it: only syncSharedSession reads that flag, and tool-result delivery
+ *  is the one call that never syncs.
+ *
+ *  The caller then takes the fresh-query path, where REBUILD imports pi's
+ *  rewritten history — this tool result included, since it is already in that
+ *  history — so the turn continues instead of ending here. Nothing is lost by
+ *  killing the subprocess: pi owns the only copy of the conversation that counts. */
+function discardRewrittenQuery(c: QueryContext): void {
+	const discarded = c.activeQuery as { interrupt?: () => Promise<unknown>; close?: () => void } | null;
+	if (discarded) abandonedQueries.add(discarded);
+	c.activeQuery = null;
+	// Leaving the routing set is what stops this result coming straight back here:
+	// contextForToolResults only matches ids against contexts still in it.
+	activeQueryContexts.delete(c);
+	c.turnToolCallIds = [];
+	c.promptStream?.fail(new Error("conversation rewritten"));
+	c.promptStream = null;
+	// Settle the parked handlers before killing the CLI, for drainForAbort's
+	// reason: one left awaiting a dead subprocess never settles.
+	c.releasePendingToolCalls("Context was compacted; this query was discarded.");
+	void discarded?.interrupt?.().catch(() => {});
+	try { discarded?.close?.(); } catch {}
+	// The CLI we just killed may still flush a record into the session JSONL, and
+	// the rebuild is the next thing that happens — so rotate rather than race it,
+	// exactly as after an abort. Only this session's mirror: the discarding query
+	// proves its own conversation is the one being rebuilt around.
+	const state = sessionStateFor(c.piSessionId);
+	if (state) setSessionStateFor(c.piSessionId, { ...state, needsRebuild: true, forceRotate: true });
+	if (c.piSessionId) historyRewrittenBySession.delete(c.piSessionId);
+	debug("provider: history rewritten under a parked query — discarded it, rebuilding from current history");
+}
+
+/** Provider entry point. Pi calls this for each new prompt and each tool result.
+ *  Two cases: tool result delivery (active query) or fresh query. */
+function streamClaudeAgentSdk(model: Model<any>, context: Context, options?: SimpleStreamOptions): AssistantMessageEventStream {
+	showStartupNoticeOnce();
+	// pi hands providers a transcript (prompt/tools folded into system messages) — fold it
+	// back out to the prompt/tools fields every cursor write, syncSharedSession call and
+	// prompt-capture lookup below assumes (issue #106).
+	context = toBridgeContext(context);
+
+	// Pi's compaction, branch-summary and bug-report calls use cacheRetention:"none"
+	// in completeSummarization. Their prompts bypass the capture boundaries, so route
+	// them to an isolated persistSession:false CC process without session sync.
+	if (options?.cacheRetention === "none") {
+		debug(`provider: one-off summarizer call (cacheRetention none) routed to isolated summary, msgs=${context.messages.length}`);
+		return isolatedStreamFn(model, context, options);
+	}
+
+	const stream = createAssistantMessageEventStream();
+
+	// DEBUG: trace followUp message triggering
+	const lastMsgRole = context.messages[context.messages.length - 1]?.role;
+	debug(`provider: streamClaudeAgentSdk called, activeQuery=${!!ctx().activeQuery}, lastMsgRole=${lastMsgRole}, isReentrant=${ctx().activeQuery !== null}`);
+
+	let activeQuery = ctx().activeQuery !== null;
+	const allResults = activeQueryContexts.size > 0 ? extractAllToolResults(context) : [];
+	let resultCtx = allResults.length > 0 ? contextForToolResults(allResults) : undefined;
+
+	// pi rewrote its history while this query sat parked at a tool boundary, so the
+	// query answers about a conversation that no longer exists. Discard it and let
+	// this tool result carry the turn into a fresh query over the rewritten history.
+	// The staleness mark is per pi session: a subagent's compaction (its own
+	// AgentSession, sharing this process) must not discard the parent's parked
+	// query, and vice versa.
+	const rewrittenUnderQuery = Boolean(resultCtx?.historyStale);
+	if (resultCtx && rewrittenUnderQuery) {
+		discardRewrittenQuery(resultCtx);
+		resultCtx = undefined;
+		// Recomputed, not cleared: a reentrant subagent may still hold a query of its own.
+		activeQuery = ctx().activeQuery !== null;
+	}
+
+	const isReentrantUserQuery = activeQuery && lastMsgRole === "user" && allResults.length === 0;
+	if (isReentrantUserQuery) {
+		debug(`provider: active query user-only call treated as reentrant fresh query, waitingHandlers=${ctx().pendingToolCalls.size}, ctx.msgs=${context.messages.length}`);
+	}
+
+	// --- Tool result delivery ---
+	// Pi appends tool results to context and calls back. Extract this turn's results
+	// (everything after the last assistant message) and match against waiting MCP
+	// handlers. Results that arrive before their handler get queued in pendingResults.
+	if (resultCtx) {
+		claimCurrentPiStream(stream, "tool-result", resultCtx, options);
+		resultCtx.resetTurnState(model);
+		// A rewrite that armed the mark after this query parked gets copied here,
+		// though markRebuildForSession usually reaches it directly.
+		if (!resultCtx.historyStale && resultCtx.piSessionId && historyRewrittenBySession.has(resultCtx.piSessionId)) {
+			resultCtx.historyStale = true;
+		}
+		// User messages (steer/followUp) pi injected into context during the
+		// active query: a steer sent while a tool was executing, drained by pi at
+		// the turn boundary and appended alongside the tool result.
+		const steer = lastMsgRole === "user" ? steerBlocks(context.messages) : null;
+		// Delivery is async because the steer must reach CC's stdin *before* the
+		// tool result does — see deliverToolResults. Detached so the provider
+		// still returns its stream synchronously.
+		void deliverToolResults(resultCtx, allResults, steer, context.messages.length);
+		// The shared cursor tracks the top-level conversation. A reentrant subagent
+		// delivering its own results would drag it to that subagent's message count
+		// — observed pulling a parent from 5 back to 3, which cost the parent's next
+		// turn a full rebuild and a flushed prompt cache.
+		const state = sessionStateFor(resultCtx.piSessionId);
+		if (state) state.cursor = context.messages.length;
+		resultCtx.latestCursor = Math.max(resultCtx.latestCursor, context.messages.length);
+		return stream;
+	}
+
+	// --- Orphaned tool result (e.g. user aborted a tool call) ---
+	// The query is gone but pi still delivered the result. Nothing to do — just
+	// emit end_turn so pi waits for the next real user message. The discard
+	// branch above already siphoned off the stale-query case, which goes on to a
+	// rebuild instead — that one has somewhere to deliver the result to.
+	// Only when this call's own signal is aborted: after an abort during a tool
+	// call, pi calls back with the run's (already aborted) signal. A result that
+	// arrives otherwise -- after a restart (a durable host resuming a run), a
+	// crash, a stall, or an overflow/quota retry -- has no parked query either, and
+	// falls through to the fresh-query path, which resumes the CC session and
+	// delivers it.
+	const lastMsg = context.messages[context.messages.length - 1];
+	const unparkedToolResult = lastMsg?.role === "toolResult" && !rewrittenUnderQuery;
+	if (unparkedToolResult && options?.signal?.aborted) {
+		debug(`provider: orphaned tool result after abort, emitting end_turn`);
+		// With no query in flight anywhere, the top-level session this result
+		// belongs to is the one whose turn just ended: its cursor advances to
+		// count the result (options.sessionId is that session — pi emits the
+		// result event through the same session's streamSimple call).
+		const orphanState = sessionStateFor(options?.sessionId ?? null);
+		if (orphanState && activeQueryContexts.size === 0) orphanState.cursor = context.messages.length;
+		// No query owns this result, so there is no context to reset: resetTurnState
+		// on the top-level ctx() would replace a live parent's turnOutput mid-stream,
+		// stranding the blocks it had already emitted. A throwaway context just
+		// supplies the empty message this turn ends with.
+		const c = new QueryContext();
+		c.resetTurnState(model);
+		queueMicrotask(() => {
+			stream.push({ type: "done", reason: "stop", message: c.turnOutput });
+			markStreamComplete(stream);
+			stream.end();
+		});
+		return stream;
+	}
+
+	// --- Fresh query ---
+
+	// 1. Determine reentrancy. Reentrant queries get their own QueryContext so
+	//    background subagents can run concurrently with the parent query.
+	const isReentrant = activeQuery;
+	const queryCtx = isReentrant ? new QueryContext() : ctx();
+	debug(`provider: fresh query setup, isReentrant=${isReentrant}, activeContexts=${activeQueryContexts.size}`);
+
+	// Resolved first: an unaccountable system prompt fails this query before anything
+	// is claimed or reset, leaving no half-built query behind — in particular no stream
+	// claimed on the shared context that nobody will ever end.
+	const { mcpTools, customToolNameToSdk, customToolNameToPi } = resolveMcpTools(context, askClaudeToolName);
+	// Build from what Pi loaded for this run, so `--no-context-files` and
+	// `--no-skills` reach Claude Code by leaving nothing to forward. A sub-agent's
+	// custom override embeds its parent's assembled Pi prompt; recursive projection
+	// replaces that exact inherited prompt with its already-safe portable parts.
+	// Derive the key from the transcript replay (toBridgeContext), NOT from the
+	// recorded keys: under a forced prompt the transcript head is projected via
+	// transformContext after turn_start, so ctx.getSystemPrompt() is not the head.
+	let promptCapture: PromptCapture | undefined;
+	let systemPromptAppend: string | undefined;
+	try {
+		promptCapture = promptCaptures.resolveOrDerive(context.systemPrompt);
+		systemPromptAppend = promptCapture
+			? projectPromptCapture(promptCapture, {
+				skillReadTool: mcpTools.some((tool) => tool.name === "read") ? "mcp" : "none",
+			})
+			: undefined;
+	} catch (err) {
+		// resolveOrDerive and projectPromptCapture throw to stop a turn that would lose
+		// its instructions or leak pi's harness text. Report it on the stream, as pi-ai's
+		// provider contract expects, so any caller — not only pi's agent loop, which
+		// catches a throw — sees a failed turn rather than a synchronous exception.
+		const output = newAssistantOutput(model, "", "error", errorMessage(err));
+		queueMicrotask(() => {
+			stream.push({ type: "error", reason: "error", error: output });
+			markStreamComplete(stream);
+			stream.end();
+		});
+		diagDump("prompt_capture_unresolved", {
+			promptChars: context.systemPrompt?.length ?? 0,
+			knownKeys: promptCaptures.size,
+			reentrantUserQuery: isReentrantUserQuery,
+			error: errorMessage(err),
+		});
+		return stream;
+	}
+
+	// 2. Fresh child context — constructor already gave us clean Maps and empty
+	//    arrays. For a reused top-level context, clear explicitly.
+	claimCurrentPiStream(stream, "fresh-query", queryCtx, options);
+	// A held event belongs to the query that received it, not to this request.
+	queryCtx.heldRateLimit = undefined;
+	queryCtx.pendingToolCalls.clear();
+	queryCtx.pendingResults.clear();
+	// Stale ids would let a late result from the previous query route here via
+	// contextForToolResults — which now means pushing its steer into this
+	// query's stdin, not just mismatching a map.
+	queryCtx.turnToolCallIds = [];
+	queryCtx.resetTurnState(model);
+	queryCtx.latestCursor = 0;
+	// The served pi session, for rewrite attribution on delivery (issue #101
+	// follow-up) and on SessionState. A fresh instance of this module inside a
+	// worktree-spawned subagent has its own contexts; each records its own.
+	queryCtx.piSessionId = options?.sessionId ?? null;
+	// A discarded query's replacement reuses this context; without the reset its
+	// first tool result would sit on armed staleness again (the mark is consumed
+	// from the set, not from here) and re-discard a healthy query.
+	queryCtx.historyStale = false;
+	queryCtx.missedSteer = false;
+
+	const cwd = sessionCwdFor(options?.sessionId);
+	// cliModel is the actual id sent to Claude Code (may carry [1m]); model.id is the
+	// pi-registered id. Log cliModel so debug lines reflect what CC actually received.
+	const cliModel = claudeCodeModelId(model, longContextSettings);
+	// Which pi session this query serves — the attribution key for history
+	// rewrites (session_compact / session_tree) and for SessionState above.
+	const piSessionId = options?.sessionId ?? null;
+	// Read before the consume below: with no query parked (an overflow retried
+	// after compaction), the session's mark is the only sign of the rewrite.
+	const historyRewritten = rewrittenUnderQuery || (piSessionId !== null && historyRewrittenBySession.has(piSessionId));
+	const syncResult = syncSharedSession(context.messages, cwd, customToolNameToSdk, cliModel, piSessionId);
+	// This query starts from the history pi has now: consume this session's
+	// armed rewrite — a sibling pi session's stays armed for its own queries.
+	if (piSessionId) historyRewrittenBySession.delete(piSessionId);
+	const { sessionId: resumeSessionId } = syncResult;
+	const promptBlocks = extractUserPromptBlocks(context.messages);
+	let promptText = extractUserPrompt(context.messages) ?? "";
+
+	// A turn continuing past a rewrite ends at its tool result, not at a
+	// prompt, so say what happened rather than falling into the empty-prompt
+	// recovery below — that one is for a shape we do not expect, and this is one
+	// we do. The rebuilt session already ends with the tool result, placed after
+	// the tool call it answers.
+	if (historyRewritten && !promptText && !promptBlocks) {
+		promptText = CONTINUE_AFTER_REWRITE_PROMPT;
+		debug(`provider: continuing the turn after a rewritten history, ${context.messages.length} msgs rebuilt`);
+	}
+	if (unparkedToolResult && !promptText && !promptBlocks) {
+		promptText = CONTINUE_AFTER_RESUME_PROMPT;
+		debug(`provider: resuming the turn at a tool result with no parked query, ${context.messages.length} msgs`);
+	}
+
+	// Guard: empty prompt means the last context message isn't a user message.
+	// This should never happen with per-query state — dump diagnostics if it does.
+	if (!promptText && !promptBlocks) {
+		diagDump("empty_prompt", {
+			contextLength: context.messages.length,
+			lastMsgRole: lastMsg?.role,
+			isReentrant,
+			activeQueryContexts: activeQueryContexts.size,
+			activeQueryExists: queryCtx.activeQuery !== null,
+			sharedSession: sessionStateFor(piSessionId) ? { sessionId: sessionStateFor(piSessionId)!.sessionId.slice(0, 8), cursor: sessionStateFor(piSessionId)!.cursor } : (sessionStateFor(null) ? { sessionId: sessionStateFor(null)!.sessionId.slice(0, 8), cursor: sessionStateFor(null)!.cursor } : null),
+			messageRoles: context.messages.map((m, i) => `[${i}]${m.role}`).join(" "),
+		});
+		// Recover: use a continuation prompt so the SDK doesn't send an empty text block
+		promptText = "[continue]";
+	}
+
+	// Always stream the prompt rather than passing a string: a parked input
+	// generator is what lets us write steers to CC's stdin mid-turn. The cost is
+	// that `isSingleUserTurn` is false, so the SDK no longer closes stdin on the
+	// first result — consumeQuery ends the stream explicitly instead, or the
+	// query would never terminate.
+	const promptStream = makePromptStream();
+	const currentBlocks: ContentBlockParam[] = promptBlocks ?? [{ type: "text", text: promptText }];
+	void promptStream.push(userMessage(syncResult.replay?.length ? [...replayBlocks(syncResult.replay), ...currentBlocks] : currentBlocks))
+		.catch((error) => debug(`provider: initial prompt push rejected:`, error));
+	queryCtx.promptStream = promptStream;
+	const mcpServers = buildMcpServers(mcpTools, queryCtx);
+
+	// MCP auto-loading suppression: CC reads MCP servers from ~/.claude.json (top-level
+	// + per-project) and .mcp.json. Since pi executes tools (not CC), those are pure
+	// token overhead. --strict-mcp-config tells the binary to use ONLY mcpServers passed
+	// programmatically and ignore filesystem MCP entries — applied unconditionally because
+	// settingSources is left at CC's default, which loads all sources.
+	const strictMcpConfigEnabled = providerSettings.strictMcpConfig !== false;
+	const claudeExecutable = providerSettings.pathToClaudeCodeExecutable;
+
+	const effort = effortForReasoning(model, options?.reasoning);
+
+	const extraArgs: Record<string, string | null> = { model: cliModel, "safe-mode": null };
+	if (strictMcpConfigEnabled) extraArgs["strict-mcp-config"] = null;
+	// Opus 4.7 defaults thinking.display to "omitted" (empty thinking text in stream).
+	// Force summarized so thinking_delta events arrive. See anthropics/claude-agent-sdk-python#830.
+	if (effort) extraArgs["thinking-display"] = "summarized";
+
+	// Suppress claude.ai cloud MCP servers (Figma/Canva/etc. auto-discovered via OAuth
+	// when the user is logged into Anthropic). These are a separate code path from
+	// filesystem MCP and are NOT blocked by --strict-mcp-config or settingSources=undefined.
+	// The native CC binary gates them on env var ENABLE_CLAUDEAI_MCP_SERVERS: setting it
+	// to "0"/"false"/"no"/"off" makes the loader return early before any cloud fetch.
+	// DISABLE_AUTO_COMPACT=1: pi owns context-management and propagates its own
+	// /compact via session_compact (see handler in default export). Letting CC
+	// also autocompact would double-flush the prompt cache and races pi's
+	// threshold with CC's, including CC's anti-thrashing guard (issue #8).
+	// Manual /compact in CC still works (we never invoke it).
+	debug("provider: system prompt append:", systemPromptAppend ?? "");
+	const childEnv = { ...process.env, ...CC_CHILD_ENV };
+	const queryOptions: NonNullable<Parameters<typeof query>[0]["options"]> = {
+		cwd,
+		env: childEnv,
+		tools: [],
+		permissionMode: "bypassPermissions",
+		includePartialMessages: true,
+		// includeGitInstructions:false drops the gitStatus block from the preset.
+		// That block is the trailing suffix of the cached system block, and a
+		// git-state transition (new file, staging, commit) rewrites it — busting
+		// the prompt cache for the whole conversation from there on (see
+		// diag/probe-git-cache.mjs). The bridge re-invokes CC per turn, so this
+		// hit on every transition. Cost here is nil: the setting also strips
+		// CC's git-workflow guidance from its Bash tool prompt, but the provider
+		// path runs CC with `tools: []`, so those definitions never ship.
+		// AskClaude keeps CC's native tools and its guidance — unaffected.
+		settings: {
+			...claudeCodeSettings(providerSettings),
+			claudeMdExcludes: CLAUDE_MD_EXCLUDES,
+			includeGitInstructions: false,
+		},
+		systemPrompt: {
+			type: "preset", preset: "claude_code",
+			append: systemPromptAppend ? systemPromptAppend : undefined,
+		},
+		extraArgs,
+		...(effort ? { effort } : {}),
+		...(mcpServers ? { mcpServers } : {}),
+		...(resumeSessionId ? { resume: resumeSessionId } : {}),
+		...(claudeExecutable ? { pathToClaudeCodeExecutable: claudeExecutable } : {}),
+		...makeCliDebugOptions("provider"),
+	};
+
+	debug("provider: fresh query",
+		`model=${cliModel} msgs=${context.messages.length} tools=${mcpTools.length}`,
+		`resume=${resumeSessionId?.slice(0, 8) ?? "none"} effort=${effort ?? "default"}`,
+		`ctxFiles=${promptCapture?.contextFiles.length ?? 0} strictMcp=${strictMcpConfigEnabled}`,
+		`prompt=${promptText.slice(0, 60)}${promptBlocks ? " [+images]" : ""}`);
+
+	// 3. Start SDK query and claim it for this context
+	let wasAborted = false;
+	const sdkQuery = queryImpl({ prompt: promptStream.stream, options: queryOptions });
+	queryCtx.activeQuery = sdkQuery;
+	activeQueryContexts.add(queryCtx);
+
+	// 4. Capture context for abort handling
+	const abortCtx = queryCtx;
+
+	const requestAbort = () => {
+		// interrupt() asks the CLI to stop gracefully; close() kills it immediately.
+		// Both are needed — interrupt alone lets the current API call finish.
+		void sdkQuery.interrupt().catch(() => {});
+		try { sdkQuery.close(); } catch {}
+	};
+	const onAbort = () => {
+		wasAborted = true;
+		drainForAbort(abortCtx, promptStream);
+		requestAbort();
+	};
+	if (options?.signal) {
+		if (options.signal.aborted) onAbort();
+		else options.signal.addEventListener("abort", onAbort, { once: true });
+	}
+
+	// Background consumer — runs until query ends
+	consumeQuery(sdkQuery, customToolNameToPi, model, () => wasAborted, queryCtx)
+		.then(async ({ capturedSessionId }) => {
+			debug(`provider: consumeQuery completed, stopReason=${queryCtx.turnOutput?.stopReason}, error=${queryCtx.turnOutput?.errorMessage}, aborted=${wasAborted}`);
+
+			// Discarded out from under: the query continuing the turn owns the context,
+			// the session and the stream. Capturing this one's session id here would put
+			// Claude Code back on the conversation it was discarded for.
+			if (abandonedQueries.has(sdkQuery)) {
+				debug("provider: discarded query completed, leaving session and stream to its replacement");
+				return;
+			}
+
+			// --- Abort detection in normal completion path ---
+			if (wasAborted || options?.signal?.aborted) {
+				// The killed subprocess may flush a late record into this session's
+				// JSONL — its own mirror's next sync must rebuild and rotate.
+				const state = sessionStateFor(queryCtx.piSessionId);
+				if (state) setSessionStateFor(queryCtx.piSessionId, { ...state, needsRebuild: true, forceRotate: true });
+				debug(`provider: abort detected, marked sharedSession needsRebuild + forceRotate`);
+				if (queryCtx.turnOutput) {
+					queryCtx.turnOutput.stopReason = "aborted";
+					queryCtx.turnOutput.errorMessage = "Operation aborted";
+				}
+				const stream = queryCtx.currentPiStream;
+				stream?.push({ type: "error", reason: "aborted", error: queryCtx.turnOutput! });
+				markStreamComplete(stream);
+				stream?.end();
+				queryCtx.currentPiStream = null;
+				queryCtx.heldRateLimit = undefined;
+				return;
+			}
+
+			// --- Capture session ID ---
+			// This query's own mirror — a reentrant subagent completing does not
+			// reassign the parent's conversation to the child's CC file.
+			if (syncResult.preserveSharedSession) {
+				debug(`provider: query done, ignoring captured session ${capturedSessionId?.slice(0, 8) ?? "none"} to preserve shared session`);
+			} else {
+				const state = sessionStateFor(queryCtx.piSessionId);
+				const sessionId = capturedSessionId ?? state?.sessionId;
+				if (sessionId) {
+					const cursor = Math.max(context.messages.length, queryCtx.latestCursor, state?.cursor ?? 0);
+					debug(`provider: query done, session=${sessionId.slice(0, 8)}, cursor=${cursor}`);
+					// A missed steer may precede the first mirror or arrive while this
+					// query is still able to complete. Preserve both rebuild signals.
+					setSessionStateFor(queryCtx.piSessionId, { ...state, sessionId, cursor, cwd, piSessionId: queryCtx.piSessionId ?? undefined, needsRebuild: queryCtx.missedSteer || state?.needsRebuild });
+				}
+			}
+
+			if (!isReentrant && queryCtx.activeQuery === sdkQuery) {
+				debug("provider: clearing activeQuery before final stream completion");
+				queryCtx.activeQuery = null;
+			}
+			finalizeCurrentStream(queryCtx, queryCtx.turnOutput?.stopReason);
+		})
+		.catch((error) => {
+			debug(`provider: query error, model=${cliModel}, aborted=${Boolean(options?.signal?.aborted)}, error=`, error);
+			if (abandonedQueries.has(sdkQuery)) {
+				debug("provider: discarded query ended in error, leaving session and stream to its replacement");
+				return;
+			}
+			if ((wasAborted || options?.signal?.aborted)) {
+				const state = sessionStateFor(queryCtx.piSessionId);
+				if (state) setSessionStateFor(queryCtx.piSessionId, { ...state, needsRebuild: true, forceRotate: true });
+				queryCtx.heldRateLimit = undefined;
+			} else {
+				// Drop this session's mirror: its conversation is in an unknown
+				// state after the error. Other sessions' mirrors stay — one
+				// session's failure says nothing about another's conversation.
+				setSessionStateFor(queryCtx.piSessionId, null);
+			}
+			promptStream.fail(error instanceof Error ? error : new Error(String(error)));
+			if (queryCtx.turnOutput) {
+				queryCtx.turnOutput.stopReason = options?.signal?.aborted ? "aborted" : "error";
+				// The SDK drops its copy of the result text if any message follows the error
+				// result, so prefer the cause consumeQuery recorded off the result itself.
+				queryCtx.turnOutput.errorMessage ??= error instanceof Error ? error.message : String(error);
+			}
+			if (!isReentrant && queryCtx.activeQuery === sdkQuery) {
+				queryCtx.releasePendingToolCalls("Query ended");
+				debug("provider: clearing activeQuery before error stream completion");
+				queryCtx.activeQuery = null;
+			}
+			const stream = queryCtx.currentPiStream;
+			stream?.push({ type: "error", reason: (queryCtx.turnOutput?.stopReason ?? "error") as "aborted" | "error", error: queryCtx.turnOutput! });
+			markStreamComplete(stream);
+			stream?.end();
+			queryCtx.currentPiStream = null;
+		})
+		.finally(() => {
+			if (options?.signal) options.signal.removeEventListener("abort", onAbort);
+			// Settle any ack still parked in the generator — the CLI is gone, so
+			// nothing will resume it. Clear the handle only if a later query
+			// hasn't already claimed the shared context.
+			promptStream.fail(new Error("query ended"));
+			if (queryCtx.promptStream === promptStream) queryCtx.promptStream = null;
+			// A later query claiming this context sets activeQuery to its own handle;
+			// null means the .then/.catch above cleared ours and nothing replaced it.
+			// Testing only for `=== sdkQuery` would never fire on the non-reentrant
+			// path, leaving the top-level context in the routing set forever — where a
+			// later orphaned tool result matches its stale turnToolCallIds and takes
+			// the delivery branch, returning a stream nothing ends.
+			if (queryCtx.activeQuery === sdkQuery || queryCtx.activeQuery === null) {
+				queryCtx.releasePendingToolCalls("Query ended");
+				queryCtx.activeQuery = null;
+				activeQueryContexts.delete(queryCtx);
+			}
+			sdkQuery.close();
+		});
+
+	return stream;
+}
+
+// --- AskClaude: prompt and wait ---
+
+async function promptAndWait(
+	prompt: string,
+	mode: "full" | "read" | "none",
+	toolCalls: Map<string, ToolCallState>,
+	signal?: AbortSignal,
+	options?: {
+		systemPrompt?: string;
+		appendSkills?: boolean;
+		onStreamUpdate?: (responseText: string) => void;
+		model?: string;
+		thinking?: string;
+		isolated?: boolean;
+		context?: Context["messages"];
+		/** pi session the calling tool ran in — AskClaude's conversation continues
+	 *  the session that called it, so its sync and capture key that mirror. */
+		piSessionId?: string | null;
+	},
+): Promise<{ responseText: string; stopReason: string }> {
+	const cwd = sessionCwdFor(options?.piSessionId);
+	const requestedModel = options?.model ?? "opus";
+	const model = resolveModel(requestedModel);
+	const modelId = model?.id ?? requestedModel;
+	const cliModel = model ? claudeCodeModelId(model, longContextSettings) : modelId;
+
+	// Session resume for shared mode — reuse provider's session if it exists,
+	// otherwise create one from pi's context.
+	// Note: doesn't update the mirror's cursor after completion, so the next
+	// provider call will see missed messages and trigger a Case 4 rebuild.
+	// AskClaude has a random-origin context handed to it, but the session it
+	// belongs to is the one whose tool ran — the extension API's execute ctx
+	// carries it — passed here as piSessionId and used for every map access.
+	const askClaudeSessionId = options?.piSessionId ?? null;
+	let resumeSessionId: string | null = null;
+	if (!options?.isolated && options?.context?.length) {
+		const askClaudeState = sessionStateFor(askClaudeSessionId);
+		if (askClaudeState) {
+			// Provider already has a session — just resume from it
+			// Any missed messages from other providers were already handled by the provider's Case 4
+			resumeSessionId = askClaudeState.sessionId;
+		} else {
+			// No provider session yet — create one from pi's context
+			const contextWithPrompt = [...options.context, { role: "user" as const, content: prompt, timestamp: Date.now() }];
+			const sync = syncSharedSession(contextWithPrompt as Context["messages"], cwd, undefined, cliModel, askClaudeSessionId);
+			resumeSessionId = sync.sessionId;
+		}
+	}
+
+	// Mode → disallowed tools
+	const disallowedTools = MODE_DISALLOWED_TOOLS[mode];
+
+	// AskClaude uses Claude Code's native Read tool rather than Pi's MCP bridge.
+	// Same resolver as the provider path: a prompt neither recorded nor derivable
+	// throws here too, rather than silently sending Claude Code no skills.
+	//
+	// Resolved only when the answer would be used. The throw is justified by what a
+	// miss would cost, so where it costs nothing — skills switched off, or no reader
+	// to open a skill file with — an unrelated miss must not fail the call.
+	const skillReadTool = disallowedTools.includes("Read") ? "none" : "native";
+	const skillCapture = options?.appendSkills !== false && skillReadTool !== "none"
+		? promptCaptures.resolveOrDerive(options?.systemPrompt)
+		: undefined;
+	const skillsBlock = skillCapture
+		? renderSkillsBlock(collectPromptSkills(skillCapture), skillReadTool)
+		: undefined;
+
+	// Effort
+	const effort = options?.thinking && options.thinking !== "off"
+		? REASONING_TO_EFFORT[options.thinking] : undefined;
+
+	const claudeExecutable = providerSettings.pathToClaudeCodeExecutable;
+
+	const extraArgs: Record<string, string | null> = {
+		"strict-mcp-config": null,
+		model: cliModel,
+	};
+	if (effort) extraArgs["thinking-display"] = "summarized";
+
+	debug("askClaude:",
+		`mode=${mode} model=${modelId} cliModel=${cliModel} effort=${effort ?? "default"}`,
+		`isolated=${options?.isolated ?? false} resume=${resumeSessionId?.slice(0, 8) ?? "none"}`,
+		`skills=${Boolean(skillsBlock)} promptLen=${prompt.length}`);
+
+	// skills: [] suppresses Claude Code's own skill listing, a system-reminder naming every
+	// skill under the ~/.claude estate. The provider path gets this for free — `tools: []`
+	// removes the Skill tool and the listing with it — but AskClaude runs on CC's native
+	// tools, so it has to be asked for. Pi-side skills still arrive via skillsBlock below,
+	// which is meant to be the only channel.
+	const sdkQuery = query({
+		prompt,
+		options: {
+			cwd,
+			env: { ...process.env, ...CC_CHILD_ENV },
+			permissionMode: "bypassPermissions",
+			settings: { ...claudeCodeSettings(providerSettings), claudeMdExcludes: CLAUDE_MD_EXCLUDES },
+			skills: [],
+			...(disallowedTools.length ? { disallowedTools } : {}),
+			...(effort ? { effort } : {}),
+			// Preset unconditionally: omitting it leaves the child on the SDK's bare default,
+			// without the tool and permission guidance the bridge relies on everywhere else.
+			// Whether pi has skills to append is unrelated to whether the child needs that.
+			systemPrompt: { type: "preset", preset: "claude_code", append: skillsBlock },
+			settingSources: ["user", "project"] as SettingSource[],
+			extraArgs,
+			...(resumeSessionId ? { resume: resumeSessionId } : {}),
+			...(options?.isolated ? { persistSession: false } : {}),
+			...(claudeExecutable ? { pathToClaudeCodeExecutable: claudeExecutable } : {}),
+			...makeCliDebugOptions("askclaude"),
+		},
+	});
+
+	// Abort handling
+	let wasAborted = false;
+	const onAbort = () => {
+		wasAborted = true;
+		sdkQuery.interrupt().catch(() => { try { sdkQuery.close(); } catch {} });
+	};
+	if (signal?.aborted) { onAbort(); throw new Error("Aborted"); }
+	signal?.addEventListener("abort", onAbort, { once: true });
+
+	let responseText = "";
+	let sdkMessageCount = 0;
+	let textDeltaCount = 0;
+	let resultSubtype: string | undefined;
+
+	try {
+		for await (const message of sdkQuery) {
+			if (wasAborted) break;
+			sdkMessageCount++;
+
+			switch (message.type) {
+				case "stream_event": {
+					const event = (message as SDKMessage & { event: any }).event;
+					// Text deltas → accumulate and stream
+					if (event?.type === "content_block_delta" && event.delta?.type === "text_delta") {
+						responseText += event.delta.text;
+						textDeltaCount++;
+						options?.onStreamUpdate?.(responseText);
+					}
+					// Tool call start → track for action summary progress
+					if (event?.type === "content_block_start" && event.content_block?.type === "tool_use") {
+						debug(`askClaude: tool_use start: ${event.content_block.name}`);
+						toolCalls.set(event.content_block.id, {
+							name: mapToolName(event.content_block.name),
+							status: "running",
+						});
+					}
+					break;
+				}
+				case "assistant": {
+					// Update tool calls with full input for action summary
+					for (const block of (message as any).message?.content ?? []) {
+						if (block.type === "tool_use") {
+							toolCalls.set(block.id, {
+								name: mapToolName(block.name),
+								status: "complete",
+								rawInput: block.input,
+							});
+						}
+					}
+					break;
+				}
+				case "result": {
+					resultSubtype = message.subtype;
+					const r = message as any;
+					if (r.usage) {
+						debug(`askClaude: result usage: in=${r.usage.input_tokens} out=${r.usage.output_tokens} cacheRead=${r.usage.cache_read_input_tokens ?? 0} cacheWrite=${r.usage.cache_creation_input_tokens ?? 0} turns=${r.num_turns ?? "?"}`);
+					}
+					// Claude Code reports an API failure with `is_error` on a result whose
+					// subtype is still "success", so without this the error text was returned
+					// as Claude's answer and pi's model read a 429 as content. Throwing hands
+					// it to the tool's own catch, which renders it as an error result.
+					const failure = wasAborted ? undefined : resultErrorText(message);
+					if (failure) throw new Error(failure);
+					if (!responseText && message.subtype === "success" && message.result) {
+						responseText = message.result;
+					}
+					break;
+				}
+			}
+		}
+
+		const stopReason = wasAborted ? "cancelled" : "stop";
+		debug(`askClaude: done`,
+			`stopReason=${stopReason} resultSubtype=${resultSubtype ?? "none"}`,
+			`sdkMessages=${sdkMessageCount} textDeltas=${textDeltaCount} responseLen=${responseText.length}`,
+			`toolCalls=${toolCalls.size}`);
+		return { responseText, stopReason };
+	} finally {
+		signal?.removeEventListener("abort", onAbort);
+		sdkQuery.close();
+	}
+}
+
+// --- Extension registration ---
+
+const PREVIEW_MAX_CHARS = 1000;
+const PREVIEW_MAX_LINES = 6;
+
+let askClaudeToolName = "AskClaude";
+
+export default function (pi: ExtensionAPI) {
+	const config = loadConfig(process.cwd());
+	let askClaudeDeprecationNoticeShown = Boolean(config.askClaudeDeprecationShown);
+	debug("loadConfig:", JSON.stringify(config));
+	providerSettings = config.provider ?? {};
+	// We need these settings to know if we're eligible for 1M context on certain models
+	// Validate at the boundary: a non-array here would throw inside every
+	// claudeCodeModelId call and brick the extension at activation.
+	const forceTwoHundredK = Array.isArray(providerSettings.forceTwoHundredK)
+		? providerSettings.forceTwoHundredK.filter((id): id is string => typeof id === "string")
+		: undefined;
+	longContextSettings = {
+		plan: providerSettings.plan ?? "pro",
+		longContextExtraUsage: providerSettings.longContextExtraUsage ?? false,
+		forceTwoHundredK,
+	};
+	const registeredModels = applyLongContext(MODELS, longContextSettings);
+	if (registeredModels.length === 0) {
+		console.error("claude-bridge: no models available from pi-ai's anthropic catalog — update @earendil-works/pi-ai (requires >=0.86.1)");
+	}
+
+	if (!config.startupNoticeShown && config.provider?.plan === undefined) {
+		pendingNotices.push('Are you using a Max plan? You need to set provider.plan to "max" to unlock 1M context in Opus.');
+	}
+
+	// Hob hosts concurrent independent sessions in one process.
+	const clearSession = (id: string) => {
+		sharedSessions.delete(sessionKey(id));
+		historyRewrittenBySession.delete(id);
+	};
+	pi.on("session_start", (event, ctx) => {
+		piUI = ctx.ui;
+		piMode = ctx.mode;
+		rememberSessionCwd(ctx.sessionManager.getSessionId(), ctx.cwd);
+		if (event.reason === "new" || event.reason === "resume" || event.reason === "fork") {
+			clearSession(ctx.sessionManager.getSessionId());
+		}
+	});
+	pi.on("session_before_compact", (event, ctx) => {
+		rememberSummarySession(event.signal, ctx.sessionManager.getSessionId());
+	});
+	pi.on("session_before_tree", (event, ctx) => {
+		rememberSummarySession(event.signal, ctx.sessionManager.getSessionId());
+	});
+	// `--system-prompt` replaces pi's default rather than adding to it, but Claude
+	// Code's preset carries its own tool and permission guidance that the bridge
+	// still depends on, so both flags are forwarded as an append.
+	//
+	// The options (custom/append/contextFiles/skills/sections) are pi config, stable across a
+	// turn; only the auto-generated tool list in the rendered prompt varies. Stash them
+	// at before_agent_start so the agent_start recording below can reuse them.
+	type RecordOptions = Parameters<typeof recordSystemPrompt>[2];
+	let lastSystemPromptOptions: RecordOptions | undefined;
+	function recordSystemPrompt(source: string, systemPrompt: string | undefined, options: {
+		customPrompt?: string;
+		appendSystemPrompt?: string;
+		contextFiles?: { path: string; content: string }[];
+		skills?: Parameters<typeof promptCaptures.record>[1]["skills"];
+		sections?: Record<string, string>;
+		selectedTools?: string[];
+	} | undefined) {
+		if (!systemPrompt) return;
+		const hasRead = !options?.selectedTools || options.selectedTools.includes("read");
+		promptCaptures.record(systemPrompt, {
+			custom: options?.customPrompt,
+			append: options?.appendSystemPrompt,
+			contextFiles: options?.contextFiles ?? [],
+			skills: hasRead ? options?.skills ?? [] : [],
+			sections: options?.sections,
+		}, source);
+	}
+	pi.on("before_agent_start", (event) => {
+		lastSystemPromptOptions = event.systemPromptOptions;
+		recordSystemPrompt("before_agent_start", event.systemPrompt, event.systemPromptOptions);
+	});
+	// The prompt the provider actually queries with is the fully-widened one: MCP tool
+	// descriptions merge into the system prompt only after their servers connect, which
+	// is after before_agent_start. ctx.getSystemPrompt() returns that widened prompt by
+	// agent_start (verified: before_agent_start=10,988 chars vs agent_start/query=23,479).
+	// A subagent embeds the widened parent prompt verbatim (pi-subagents reads
+	// ctx.getSystemPrompt() at dispatch), so unless the widened prompt is a capture key
+	// too, the child's turn resolves against nothing, falls to a verbatim side request,
+	// and ships pi's harness — tripping the server's third-party plan-eligibility check
+	// ("out of extra usage"). Recording it here, before the query, restores the match.
+	//
+	// agent_start also captures a handler-returned forceSystemPrompt, which
+	// buildSystemPrompt renders verbatim.
+	pi.on("agent_start", (_event, ctx) => {
+		recordSystemPrompt("agent_start", ctx.getSystemPrompt(), lastSystemPromptOptions);
+	});
+
+	// Mid-run re-renders: turn_start fires before every turn (first turn included)
+	// after the turn's prompt is final: prepareNextTurnWithContext has
+	// re-rendered the options (pi's section-based prompt) and any mid-run
+	// setActiveToolsByName rebuild has already landed. Re-keying at each boundary
+	// the prompt can change at keeps exact-match alive mid-run. The stashed options can
+	// lag a mid-run tool-loadout change, which skews the hasRead skills filter until the
+	// next before_agent_start — accepted: a stale skills list beats failing the turn.
+	pi.on("turn_start", (_event, ctx) => {
+		recordSystemPrompt("turn_start", ctx.getSystemPrompt(), lastSystemPromptOptions);
+	});
+	pi.on("session_shutdown", (_event, ctx) => {
+		forgetSessionCwd(ctx.sessionManager.getSessionId());
+		reportLeaks("session_shutdown");
+		clearSession(ctx.sessionManager.getSessionId());
+	});
+
+	// Pi and summary extensions own compaction and branch summaries; native calls use the
+	// isolated provider path. Observe the completed rewrite below, not its owner.
+
+	// pi /compact and session-tree navigation (rewind / fork-at-point /
+	// branch switch) both mutate pi's messages array out from under the
+	// bridge. syncSharedSession's REUSE check would otherwise see
+	// slice(cursor) === [] (or skip entries) and keep --resume'ing a CC
+	// session that no longer matches pi's history. /compact in particular
+	// triggers CC's autocompact-thrashing guard (issue #8). Force the next
+	// call down the REBUILD path so CC sees the current history — and, when a
+	// query is parked at a tool boundary while this fires, discard that query
+	// instead of resuming it (markRebuild, discardRewrittenQuery).
+	//
+	// Attributed to the compacting session (ctx.sessionManager belongs to the
+	// session whose runner fired this), so a subagent compacting while its
+	// parent sits parked on the Agent tool result discards nothing — the parent's
+	// query is live and its conversation untouched. Registered by every instance
+	// of this module; instances sponsoring stale marks forward them to the
+	// serving instance via sponsorMarkRebuildForSession.
+	pi.on("session_compact", (event, ctx) =>
+		sponsorMarkRebuildForSession(ctx.sessionManager.getSessionId(), `session_compact:${event.reason}:willRetry=${event.willRetry}`));
+	pi.on("session_tree", (_event, ctx) => sponsorMarkRebuildForSession(ctx.sessionManager.getSessionId(), "session_tree"));
+
+	// Hob uses the same module instance for catalog registration and all sessions.
+
+	const providerConfig = {
+		baseUrl: "claude-bridge",
+		apiKey: "not-used",
+		api: "claude-bridge",
+		models: registeredModels,
+		// Cast: the Provider interface passes a TranscriptContext; the bridge takes plain
+		// Context models (toBridgeContext normalizes at the stream entry points).
+		streamSimple: streamClaudeAgentSdk as any,
+	};
+	pi.registerProvider(PROVIDER_ID, providerConfig);
+
+	// --- AskClaude tool ---
+
+	const askConf = config.askClaude;
+	const askDefaults = resolveAskClaudeDefaults(askConf);
+	askClaudeToolName = askConf?.name ?? "AskClaude";
+
+	if (askConf?.enabled) {
+		const askClaudeParams = buildAskClaudeParams(askDefaults);
+		pi.registerTool<typeof askClaudeParams>({
+			name: askConf?.name ?? "AskClaude",
+			label: askConf?.label ?? "Ask Claude Code",
+			description: askClaudeToolDescription(askDefaults, askConf?.description),
+			parameters: askClaudeParams,
+			renderCall(args, theme) {
+				let text = theme.fg("mdLink", theme.bold("AskClaude "));
+				const tags = askClaudeCallTags(args, askDefaults);
+				if (tags.length) text += `${theme.fg("accent", `[${tags.join(", ")}]`)} `;
+				const truncated = args.prompt.length > PREVIEW_MAX_CHARS ? args.prompt.substring(0, PREVIEW_MAX_CHARS) : args.prompt;
+				const lines = truncated.split("\n").slice(0, PREVIEW_MAX_LINES);
+				text += theme.fg("muted", `"${lines.join("\n")}"`);
+				if (args.prompt.length > PREVIEW_MAX_CHARS || args.prompt.split("\n").length > PREVIEW_MAX_LINES) text += theme.fg("dim", " …");
+				return new Text(text, 0, 0);
+			},
+			renderResult(result, { expanded, isPartial }, theme) {
+				if (isPartial) {
+					const status = result.content[0]?.type === "text" ? result.content[0].text : "working...";
+					return new Text(theme.fg("mdLink", "◉ Claude Code ") + theme.fg("muted", status), 0, 0);
+				}
+
+				const details = result.details as { prompt?: string; executionTime?: number; actions?: string; error?: boolean } | undefined;
+				const body = result.content[0]?.type === "text" ? result.content[0].text : "";
+
+				let text = details?.error
+					? theme.fg("error", "✗ Claude Code error")
+					: theme.fg("mdLink", "✓ Claude Code");
+
+				if (details?.executionTime) text += ` ${theme.fg("dim", `${(details.executionTime / 1000).toFixed(1)}s`)}`;
+				if (details?.actions) text += ` ${theme.fg("muted", details.actions)}`;
+
+				if (expanded) {
+					if (details?.prompt) text += `\n${theme.fg("dim", `Prompt: ${details.prompt}`)}`;
+					if (details?.prompt && body) text += `\n${theme.fg("dim", "─".repeat(40))}`;
+					if (body) text += `\n${theme.fg("toolOutput", body)}`;
+				} else {
+					const truncated = body.length > PREVIEW_MAX_CHARS ? body.substring(0, PREVIEW_MAX_CHARS) : body;
+					const lines = truncated.split("\n").slice(0, PREVIEW_MAX_LINES);
+					if (lines.length) text += `\n${theme.fg("toolOutput", lines.join("\n"))}`;
+					if (body.length > PREVIEW_MAX_CHARS || body.split("\n").length > PREVIEW_MAX_LINES) text += `\n${theme.fg("dim", `… (${keyHint("app.tools.expand", "to expand")})`)}`;
+
+				}
+
+				return new Text(text, 0, 0);
+			},
+			async execute(_id, params, signal, onUpdate, ctx) {
+				if (!askClaudeDeprecationNoticeShown && piMode === "tui") {
+					ctx.ui.notify("AskClaude is deprecated and may be removed in a future release. The Claude Code provider remains supported.", "warning");
+					markAskClaudeDeprecationNoticeShown();
+					askClaudeDeprecationNoticeShown = true;
+				}
+
+				// Guard: circular delegation
+				if (ctx.model?.baseUrl === "claude-bridge") {
+					debug("askClaude: blocked circular delegation (active provider is claude-bridge)");
+					return {
+						content: [{ type: "text" as const, text: "Error: AskClaude cannot be used when the active provider is claude-bridge — you're already running through Claude Code." }],
+						details: { error: true },
+					};
+				}
+
+				const mode = resolveAskClaudeMode(params.mode, askDefaults);
+				const isolated = params.isolated ?? askDefaults.isolated;
+				const toolCalls = new Map<string, ToolCallState>();
+				const start = Date.now();
+
+				const progressInterval = setInterval(() => {
+					const elapsed = ((Date.now() - start) / 1000).toFixed(0);
+					const summary = buildActionSummary(toolCalls);
+					const status = summary ? `${elapsed}s — ${summary}` : `${elapsed}s — working...`;
+					onUpdate?.({
+						content: [{ type: "text", text: status }],
+						details: { prompt: params.prompt, executionTime: Date.now() - start },
+					});
+				}, 1000);
+
+				try {
+					const result = await promptAndWait(params.prompt, mode, toolCalls, signal, {
+						systemPrompt: ctx.getSystemPrompt(),
+						appendSkills: askConf?.appendSkills,
+						model: params.model,
+						thinking: params.thinking,
+						isolated,
+						context: isolated ? undefined : buildSessionContext(ctx.sessionManager.getBranch()).messages as Context["messages"],
+						piSessionId: ctx.sessionManager.getSessionId(),
+					});
+					clearInterval(progressInterval);
+					onUpdate?.({ content: [{ type: "text", text: "" }], details: {} });
+					const executionTime = Date.now() - start;
+					const actions = buildActionSummary(toolCalls);
+
+					const text = actions
+						? `${result.responseText}\n\n[Claude Code actions: ${actions}]`
+						: result.responseText;
+					return {
+						content: [{ type: "text" as const, text }],
+						details: { prompt: params.prompt, executionTime, actions },
+					};
+				} catch (err) {
+					clearInterval(progressInterval);
+					debug(`askClaude error: mode=${mode}, model=${params.model ?? "default"}, isolated=${isolated}, elapsed=${((Date.now() - start) / 1000).toFixed(1)}s, error=`, err);
+					const msg = errorMessage(err);
+					return {
+						content: [{ type: "text" as const, text: `Error: ${msg}` }],
+						details: { prompt: params.prompt, executionTime: Date.now() - start, error: true },
+					};
+				}
+			},
+		});
+	}
+}