@diousk/pi-subagents-fast 0.20.0

package/dist/agent-runner.js

@@ -0,0 +1,1034 @@
+/**
+ * agent-runner.ts — Core execution engine: creates sessions, runs agents, collects results.
+ */
+import { readFileSync } from "node:fs";
+import { homedir } from "node:os";
+import { basename, dirname, isAbsolute, join, resolve } from "node:path";
+import { createAgentSession, DefaultResourceLoader, getAgentDir, SessionManager, SettingsManager, } from "@earendil-works/pi-coding-agent";
+import { BUILTIN_TOOL_NAMES, getAgentConfig, getConfig, getMemoryToolNames, getReadOnlyMemoryToolNames, getToolNamesForType } from "./agent-types.js";
+import { runInChildSessionContext } from "./child-context.js";
+import { buildParentContext, extractText } from "./context.js";
+import { DEFAULT_AGENTS } from "./default-agents.js";
+import { detectEnv } from "./env.js";
+import { buildMemoryBlock, buildReadOnlyMemoryBlock } from "./memory.js";
+import { createNestedSubagentTools, getMaxSubagentDepth } from "./nested-tools.js";
+import { buildAgentPrompt } from "./prompts.js";
+import { preloadSkills } from "./skill-loader.js";
+import { createStructuredCapture, createStructuredOutputTool, structuredRetryPrompt } from "./structured-output.js";
+/**
+ * Tool names registered by THIS extension. Single source of truth so the
+ * registration sites (index.ts) and the subagent exclusion list below can't
+ * drift apart. These are our own tools, not pi built-ins, so they can't be
+ * derived from pi — but they only need defining once.
+ */
+export const SUBAGENT_TOOL_NAMES = {
+    AGENT: "Agent",
+    WORKFLOW: "SubagentWorkflow",
+    GET_RESULT: "get_subagent_result",
+    STEER: "steer_subagent",
+};
+/** Names of tools registered by this extension that subagents must NOT inherit. */
+const EXCLUDED_TOOL_NAMES = Object.values(SUBAGENT_TOOL_NAMES);
+/** APIs whose request payloads support OpenAI service tiers. */
+const SERVICE_TIER_APIS = new Set(["openai-codex-responses", "openai-responses"]);
+/** Whether an API accepts the OpenAI `service_tier` request field. */
+export function isServiceTierApi(api) {
+    return api !== undefined && SERVICE_TIER_APIS.has(api);
+}
+function isObjectPayload(payload) {
+    return typeof payload === "object" && payload !== null && !Array.isArray(payload);
+}
+/**
+ * Add a custom agent's service tier to compatible provider requests.
+ *
+ * The existing payload hook belongs to pi's extension lifecycle, so it must be
+ * called first and its replacement preserved. An undefined replacement means
+ * "keep the original payload" in pi-ai's callback contract.
+ */
+export function installServiceTierPayload(session, serviceTier) {
+    if (serviceTier === undefined)
+        return;
+    const priorOnPayload = session.agent.onPayload;
+    session.agent.onPayload = async (payload, requestModel) => {
+        const replacement = priorOnPayload
+            ? await priorOnPayload(payload, requestModel)
+            : undefined;
+        const effectivePayload = replacement === undefined ? payload : replacement;
+        if (!isServiceTierApi(requestModel.api) || !isObjectPayload(effectivePayload)) {
+            return effectivePayload;
+        }
+        return { ...effectivePayload, service_tier: serviceTier };
+    };
+}
+/**
+ * Canonical name of an extension for `extensions: [...]` allowlist matching.
+ * Lowercased — extension names match case-insensitively so `extensions: [Mcp]`
+ * resolves the same as `[mcp]`. Tool names within `ext:foo/bar` are not affected.
+ * Directory extensions (`foo/index.ts`) resolve to the parent directory name;
+ * single-file extensions to the basename minus `.ts`/`.js`.
+ */
+export function extensionCanonicalName(extPath) {
+    const base = basename(extPath);
+    const name = base === "index.ts" || base === "index.js"
+        ? basename(dirname(extPath))
+        : base.replace(/\.(ts|js)$/, "");
+    return name.toLowerCase();
+}
+/**
+ * The unscoped, lowercased npm short name of the pi package that DECLARES
+ * `extPath` as an extension entry — or undefined if the entry doesn't belong to
+ * such a package.
+ *
+ * Climbs from the entry's directory looking for the package that owns it, and
+ * stays strictly within that package's tree by stopping at two structural
+ * boundaries — no hardcoded depth:
+ *   - the FIRST `package.json` found (the package root); the entry's own
+ *     manifest always sits at the root, above the entry, below any node_modules.
+ *   - a `node_modules` directory: a package never spans one (it's where OTHER
+ *     packages live), so reaching it means we've climbed out of the package —
+ *     stop before reading a consumer's or parent package's manifest.
+ * The name is then taken only when that root's `pi.extensions` manifest actually
+ * lists this entry. That "declares this entry" check is deliberate: our own test
+ * fixtures live under this repo, whose root manifest declares `./src/index.ts`
+ * as `@tintinweb/pi-subagents`, so a looser rule would misattribute every
+ * co-located file to `pi-subagents`.
+ */
+function extensionPackageName(extPath) {
+    const entry = resolve(extPath);
+    let dir = dirname(extPath);
+    for (;;) {
+        // Climbing into node_modules means we've left the owning package's tree.
+        if (basename(dir) === "node_modules")
+            return undefined;
+        let pkg;
+        try {
+            pkg = JSON.parse(readFileSync(join(dir, "package.json"), "utf-8"));
+        }
+        catch {
+            const parent = dirname(dir);
+            if (parent === dir)
+                return undefined; // walked to the filesystem root
+            dir = parent;
+            continue;
+        }
+        // First package.json wins — it's the package root; decide here.
+        const entries = pkg.pi?.extensions;
+        if (typeof pkg.name === "string" &&
+            Array.isArray(entries) &&
+            entries.some((e) => typeof e === "string" && resolve(dir, e) === entry)) {
+            const short = pkg.name.startsWith("@") ? pkg.name.slice(pkg.name.indexOf("/") + 1) : pkg.name;
+            return short.toLowerCase();
+        }
+        return undefined;
+    }
+}
+/**
+ * All names an extension answers to for allowlist matching (lowercased): its
+ * path-derived {@link extensionCanonicalName} plus, when a pi package manifest
+ * declares this entry, that package's unscoped short name (`@scope/foo` → `foo`).
+ * #143: an extension installed via `pi.extensions: ["./src/index.ts"]` would
+ * otherwise only ever match as `src` (the source directory), never by its
+ * package name. The path-derived name is preserved, so it keeps matching too.
+ */
+export function extensionCanonicalNames(extPath) {
+    const canonical = extensionCanonicalName(extPath);
+    const pkg = extensionPackageName(extPath);
+    return pkg && pkg !== canonical ? [canonical, pkg] : [canonical];
+}
+/**
+ * Classify `extensions: string[]` frontmatter entries for the loader-level filter.
+ *
+ * An entry is a PATH iff it contains a path separator or starts with `~`; otherwise
+ * it is a NAME. `"*"` sets the wildcard flag (keep all default-discovered extensions).
+ *
+ * Path entries are resolved (`~` expanded, made absolute against `cwd`) into `paths`
+ * — and their canonical name is also added to `names`. The loader override matches
+ * everything by canonical name, so path-loaded extensions are matched via their name
+ * rather than their post-staging `Extension.path`.
+ */
+export function parseExtensionsSpec(entries, cwd) {
+    const names = new Set();
+    const paths = [];
+    let wildcard = false;
+    for (const entry of entries) {
+        if (!entry)
+            continue;
+        if (entry === "*") {
+            wildcard = true;
+            continue;
+        }
+        const isPathEntry = entry.includes("/") || entry.includes("\\") || entry.startsWith("~");
+        if (!isPathEntry) {
+            names.add(entry.toLowerCase());
+            continue;
+        }
+        let p = entry;
+        if (p === "~" || p.startsWith("~/") || p.startsWith("~\\")) {
+            p = homedir() + p.slice(1);
+        }
+        const abs = isAbsolute(p) ? p : resolve(cwd, p);
+        paths.push(abs);
+        names.add(extensionCanonicalName(abs));
+    }
+    return { names, paths, wildcard };
+}
+/**
+ * Parse raw `ext:` selector strings (from the `tools:` CSV) into the set of
+ * extension names to keep loaded and a per-extension tool-narrowing map.
+ *
+ * `ext:foo` → `extNames` has `foo`, no narrowing entry (all of foo's tools).
+ * `ext:foo/bar` → `extNames` has `foo`, `narrowing.foo` has `bar` (only `bar`).
+ * A name lands in `narrowing` only when a `/tool` form is seen, so a bare
+ * `ext:foo` alongside `ext:foo/bar` leaves narrowing in effect (narrowing wins).
+ * The split is on the first `/`; extension canonical names never contain `/`.
+ */
+export function parseExtSelectors(entries) {
+    const extNames = new Set();
+    const narrowing = new Map();
+    for (const raw of entries) {
+        if (!raw)
+            continue;
+        const body = raw.slice("ext:".length);
+        const slash = body.indexOf("/");
+        // Extension name matches case-insensitively (matches the loader-side canonical
+        // name). Tool names are case-preserved — they're matched against pi-mono's
+        // registered identifiers, which are case-sensitive.
+        const name = (slash === -1 ? body : body.slice(0, slash)).trim().toLowerCase();
+        if (!name)
+            continue;
+        extNames.add(name);
+        if (slash === -1)
+            continue;
+        const tool = body.slice(slash + 1).trim();
+        if (!tool)
+            continue;
+        let set = narrowing.get(name);
+        if (!set) {
+            set = new Set();
+            narrowing.set(name, set);
+        }
+        set.add(tool);
+    }
+    return { extNames, narrowing };
+}
+/**
+ * Keep a subagent's tool scope correct as extensions register tools over time.
+ *
+ * Extensions may call `registerTool` long after load — pi-mcp from `session_start`,
+ * context-mode from `before_agent_start` — so scope has to be re-derived rather than
+ * snapshotted. `registerTool` writes into the very `extension.tools` maps this reads,
+ * so `inScope()` sees late arrivals on the next call.
+ *
+ * Two enforcement points, because neither covers the whole picture:
+ *
+ *   - `turn_end` re-narrows the ACTIVE set. pi emits `turn_end` immediately before
+ *     `prepareNextTurn` re-snapshots `agent.state.tools`, and session listeners run
+ *     synchronously, so the narrow lands in time for turns 2..N.
+ *   - `beforeToolCall` blocks out-of-scope calls. Turn 1 cannot be narrowed at all:
+ *     `before_agent_start` fires INSIDE `prompt()` and may widen the tool set, but
+ *     `createContextSnapshot()` freezes that turn's tools immediately after — there
+ *     is no hook in between. A call-time check is the only correct guard there.
+ *
+ * Both are installed on the session and deliberately NOT unsubscribed: they must
+ * outlive the `runAgent` call so resumed/steered turns stay scoped. pi's `dispose()`
+ * clears `_eventListeners`, so they die with the session rather than leaking.
+ *
+ * Only meaningful when extensions are loaded — under `noExtensions`/`isolated` the
+ * static `allowedToolNames` allowlist already gates the registry itself.
+ */
+export function installExtensionToolScope(session, ctx) {
+    const { loader, toolNames, disallowedSet, extNames, narrowing, readmitToolNames } = ctx;
+    // The names allowed right now. Mirrors the `ext:` opt-in flip: when any `ext:`
+    // selector is present, extension tools become an explicit allowlist — a loaded
+    // extension not named by a selector contributes nothing (its handlers still ran),
+    // and `ext:foo/bar` narrows `foo` to just `bar`.
+    const inScope = () => {
+        const keep = new Set(toolNames.filter((t) => !disallowedSet?.has(t)));
+        const optInActive = extNames.size > 0;
+        for (const extension of loader.getExtensions().extensions) {
+            const canons = extensionCanonicalNames(extension.path);
+            if (optInActive && !canons.some((c) => extNames.has(c)))
+                continue;
+            // First alias that carries a narrowing set — a user won't narrow one
+            // extension under two different names, so first-match is correct.
+            const narrowed = canons.map((c) => narrowing.get(c)).find(Boolean);
+            for (const name of extension.tools.keys()) {
+                if (narrowed && !narrowed.has(name))
+                    continue;
+                if (disallowedSet?.has(name))
+                    continue;
+                keep.add(name);
+            }
+        }
+        for (const name of EXCLUDED_TOOL_NAMES)
+            keep.delete(name);
+        // Injected tools are legitimately active for this agent — re-admit them so
+        // the renarrow keeps them in the active set and beforeToolCall doesn't
+        // block them. Already vetted against `disallowed_tools` by the caller,
+        // which is the only place that knows which kind may be taken back.
+        for (const name of readmitToolNames)
+            keep.add(name);
+        return keep;
+    };
+    const renarrow = () => {
+        const allowed = inScope();
+        const next = session.getAllTools().map((t) => t.name).filter((n) => allowed.has(n));
+        const current = session.getActiveToolNames();
+        // setActiveToolsByName unconditionally rebuilds the system prompt, so skip
+        // the no-op that steady-state turns would otherwise pay for every turn.
+        if (next.length !== current.length || next.some((n, i) => n !== current[i])) {
+            session.setActiveToolsByName(next);
+        }
+    };
+    // Activate what registered during session_start (eager MCP servers); pi would
+    // otherwise leave only its four default built-ins active at turn 1.
+    renarrow();
+    session.subscribe((event) => {
+        if (event.type === "turn_end")
+            renarrow();
+    });
+    const priorBeforeToolCall = session.agent.beforeToolCall;
+    session.agent.beforeToolCall = async (context, signal) => {
+        if (!inScope().has(context.toolCall.name)) {
+            return {
+                block: true,
+                reason: `Tool "${context.toolCall.name}" is not available to this subagent.`,
+            };
+        }
+        return priorBeforeToolCall?.(context, signal);
+    };
+}
+/** Default max turns. undefined = unlimited (no turn limit). */
+let defaultMaxTurns;
+/** Normalize max turns. undefined or 0 = unlimited, otherwise minimum 1. */
+export function normalizeMaxTurns(n) {
+    if (n == null || n === 0)
+        return undefined;
+    return Math.max(1, n);
+}
+/** Get the default max turns value. undefined = unlimited. */
+export function getDefaultMaxTurns() { return defaultMaxTurns; }
+/** Set the default max turns value. undefined or 0 = unlimited, otherwise minimum 1. */
+export function setDefaultMaxTurns(n) { defaultMaxTurns = normalizeMaxTurns(n); }
+/**
+ * The turn limit a run of `type` will actually enforce: an explicit value if the
+ * caller supplied one, else the agent's own `max_turns`, else the project
+ * default. `undefined` = unlimited.
+ *
+ * Exported because the widget's turn counter (`↻3≤20`) has to predict this
+ * before the run starts, and a second copy of the expression would drift from
+ * the one below that enforces it.
+ */
+export function resolveEffectiveMaxTurns(type, explicit) {
+    return normalizeMaxTurns(explicit ?? getAgentConfig(type)?.maxTurns ?? defaultMaxTurns);
+}
+/**
+ * Project default for `persist_session`, from the `rememberAgents` setting.
+ * On by default: a persisted session is what lets `@handle` reopen an agent's
+ * conversation after its record has been evicted, which is the whole point of
+ * addressing an agent by a name that outlives one run. Per-agent frontmatter
+ * still overrides it in both directions.
+ */
+let rememberAgents = true;
+/** Whether subagent sessions are persisted by default. */
+export function getRememberAgents() { return rememberAgents; }
+/** Set whether subagent sessions are persisted by default. */
+export function setRememberAgents(b) { rememberAgents = b; }
+/** Additional turns allowed after the soft limit steer message. */
+let graceTurns = 5;
+/** Get the grace turns value. */
+export function getGraceTurns() { return graceTurns; }
+/** Set the grace turns value (minimum 1). */
+export function setGraceTurns(n) { graceTurns = Math.max(1, n); }
+/**
+ * Try to find the right model for an agent type.
+ * Priority: explicit option > config.model > parent model.
+ */
+export function resolveDefaultModel(parentModel, registry, configModel) {
+    if (configModel) {
+        const slashIdx = configModel.indexOf("/");
+        if (slashIdx !== -1) {
+            const provider = configModel.slice(0, slashIdx);
+            const modelId = configModel.slice(slashIdx + 1);
+            // Build a set of available model keys for fast lookup
+            const available = registry.getAvailable?.();
+            const availableKeys = available
+                ? new Set(available.map((m) => `${m.provider}/${m.id}`))
+                : undefined;
+            const isAvailable = (p, id) => !availableKeys || availableKeys.has(`${p}/${id}`);
+            const found = registry.find(provider, modelId);
+            if (found && isAvailable(provider, modelId))
+                return found;
+        }
+    }
+    return parentModel;
+}
+/**
+ * Subscribe to a session and collect the last assistant message text.
+ * Returns an object with a `getText()` getter and an `unsubscribe` function.
+ */
+function collectResponseText(session) {
+    let text = "";
+    const unsubscribe = session.subscribe((event) => {
+        // message_start also fires for user and toolResult messages — resetting on
+        // those would wipe assistant text already collected. Reset only when a new
+        // ASSISTANT message begins, so getText() is the last assistant message's text.
+        if (event.type === "message_start" && event.message.role === "assistant") {
+            text = "";
+        }
+        if (event.type === "message_update" && event.assistantMessageEvent.type === "text_delta") {
+            text += event.assistantMessageEvent.delta;
+        }
+    });
+    return { getText: () => text, unsubscribe };
+}
+/**
+ * Get the last non-empty assistant text produced during THIS invocation.
+ * `startIndex` is the message count captured before the prompt, so the walk-back
+ * never crosses into a previous turn: on a resume whose new turn failed empty,
+ * this returns "" instead of the prior turn's answer (#144). Defaults to 0 (a
+ * fresh spawn, where the whole history belongs to this run).
+ */
+function getLastAssistantText(session, startIndex = 0) {
+    for (let i = session.messages.length - 1; i >= startIndex; i--) {
+        const msg = session.messages[i];
+        if (msg.role !== "assistant")
+            continue;
+        const text = extractText(msg.content).trim();
+        if (text)
+            return text;
+    }
+    return "";
+}
+/**
+ * Error message of THIS invocation's final assistant message, when that turn
+ * failed. Two failure shapes, both keyed off how the final turn STOPPED:
+ *   - stopReason "error": a provider failure pi resolved instead of rejecting
+ *     (any text; partial output is surfaced separately).
+ *   - stopReason "length" with NO text: a silent max-token death — the run hit
+ *     the output-token ceiling before writing anything, which would otherwise
+ *     land as a "completed" run with an empty result (the #144 symptom).
+ * Everything else completes: a clean "stop"/"toolUse" final, and — crucially — a
+ * "length" stop that DID produce text (a legitimate truncated-but-useful answer).
+ * "aborted" is handled by the manager's abort flag / "stopped" guard, not here.
+ * Bounded by `startIndex` (like the text fallback) so a resume that produced no
+ * assistant message of its own never inherits a PRIOR turn's stop reason.
+ */
+function finalTurnError(session, startIndex = 0) {
+    for (let i = session.messages.length - 1; i >= startIndex; i--) {
+        const msg = session.messages[i];
+        if (msg.role !== "assistant")
+            continue;
+        if (msg.stopReason === "error") {
+            return msg.errorMessage?.trim() || "provider error with no output";
+        }
+        if (msg.stopReason === "length" && !extractText(msg.content).trim()) {
+            return "run hit the output token limit before producing any text";
+        }
+        return undefined;
+    }
+    return undefined;
+}
+/**
+ * Wire an AbortSignal to abort a session.
+ * Returns a cleanup function to remove the listener.
+ */
+function forwardAbortSignal(session, signal) {
+    if (!signal)
+        return () => { };
+    const onAbort = () => session.abort();
+    signal.addEventListener("abort", onAbort, { once: true });
+    return () => signal.removeEventListener("abort", onAbort);
+}
+function resolveConfiguredSessionDir(sessionDir, cwd) {
+    if (!sessionDir)
+        return undefined;
+    if (sessionDir === "~" || sessionDir.startsWith("~/"))
+        return resolve(homedir(), sessionDir.slice(2));
+    if (isAbsolute(sessionDir))
+        return sessionDir;
+    return resolve(cwd, sessionDir);
+}
+export async function runAgent(ctx, type, prompt, options) {
+    const config = getConfig(type);
+    const agentConfig = getAgentConfig(type);
+    // Resolve working directory: worktree override > parent cwd
+    const effectiveCwd = options.cwd ?? ctx.cwd;
+    // Filesystem work happens in effectiveCwd; config discovery in configCwd.
+    // They differ only for SpawnOptions.cwd spawns (config stays with the parent).
+    const configCwd = options.configCwd ?? effectiveCwd;
+    const env = await detectEnv(options.pi, effectiveCwd);
+    // Get parent system prompt for append-mode agents
+    const parentSystemPrompt = ctx.getSystemPrompt();
+    // Build prompt extras (memory, skill preloading)
+    const extras = {};
+    if (options.worktreeBase)
+        extras.worktreeBase = options.worktreeBase;
+    if (options.workflow && !options.structuredOutput)
+        extras.workflowChild = true;
+    // Resolve extensions/skills: isolated overrides to false
+    const extensions = options.isolated ? false : config.extensions;
+    // Nulling excludes under isolated also suppresses the orphaned-exclude warning —
+    // isolation is an intentional override, not a misconfiguration.
+    const excludeExtensions = options.isolated ? undefined : config.excludeExtensions;
+    const skills = options.isolated ? false : config.skills;
+    // Skill preloading: when skills is string[], preload their content into prompt
+    if (Array.isArray(skills)) {
+        const loaded = preloadSkills(skills, configCwd);
+        if (loaded.length > 0) {
+            extras.skillBlocks = loaded;
+        }
+    }
+    let toolNames = getToolNamesForType(type);
+    // Persistent memory: detect write capability and branch accordingly.
+    // Account for disallowedTools — a tool in the base set but on the denylist is not truly available.
+    if (agentConfig?.memory) {
+        const existingNames = new Set(toolNames);
+        const denied = agentConfig.disallowedTools ? new Set(agentConfig.disallowedTools) : undefined;
+        const effectivelyHas = (name) => existingNames.has(name) && !denied?.has(name);
+        const hasWriteTools = effectivelyHas("write") || effectivelyHas("edit");
+        if (hasWriteTools) {
+            // Read-write memory: add any missing memory tool names (read/write/edit)
+            const extraNames = getMemoryToolNames(existingNames);
+            if (extraNames.length > 0)
+                toolNames = [...toolNames, ...extraNames];
+            extras.memoryBlock = buildMemoryBlock(agentConfig.name, agentConfig.memory, configCwd);
+        }
+        else {
+            // Read-only memory: only add read tool name, use read-only prompt
+            const extraNames = getReadOnlyMemoryToolNames(existingNames);
+            if (extraNames.length > 0)
+                toolNames = [...toolNames, ...extraNames];
+            extras.memoryBlock = buildReadOnlyMemoryBlock(agentConfig.name, agentConfig.memory, configCwd);
+        }
+    }
+    // Build system prompt from agent config
+    let systemPrompt;
+    if (agentConfig) {
+        systemPrompt = buildAgentPrompt(agentConfig, effectiveCwd, env, parentSystemPrompt, extras);
+    }
+    else {
+        // Unknown type fallback: spread the canonical general-purpose config (defensive —
+        // unreachable in practice since index.ts resolves unknown types before calling runAgent).
+        const fallback = DEFAULT_AGENTS.get("general-purpose");
+        if (!fallback)
+            throw new Error(`No fallback config available for unknown type "${type}"`);
+        systemPrompt = buildAgentPrompt({ ...fallback, name: type }, effectiveCwd, env, parentSystemPrompt, extras);
+    }
+    // When skills is string[], we've already preloaded them into the prompt.
+    // Still pass noSkills: true since we don't need the skill loader to load them again.
+    const noSkills = skills === false || Array.isArray(skills);
+    const agentDir = getAgentDir();
+    // Extension loading:
+    // - true  → all default-discovered extensions
+    // - false → none (noExtensions)
+    // - string[] → loader-level allowlist. Bare names keep the matching
+    //   default-discovered extension; path entries load that extension fresh;
+    //   "*" keeps all default-discovered extensions. Excluded extensions never
+    //   bind handlers or register tools (their factory still runs once).
+    //
+    // Suppress AGENTS.md/CLAUDE.md and APPEND_SYSTEM.md — upstream's
+    // buildSystemPrompt() re-appends both AFTER systemPromptOverride, which
+    // would defeat prompt_mode: replace and isolated: true. Parent context, if
+    // wanted, reaches the subagent via prompt_mode: append (parentSystemPrompt
+    // is embedded in systemPromptOverride) or inherit_context (conversation).
+    // `ext:` selectors from the `tools:` CSV narrow which extension tools surface to
+    // the LLM. They do NOT control loading — `extensions:` is the sole authority for
+    // which extensions load. `ext:foo` against an extension that `extensions:` excluded
+    // is an orphan and warns after reload. `isolated` means no extension tools at all.
+    const { extNames, narrowing } = parseExtSelectors(options.isolated ? [] : (agentConfig?.extSelectors ?? []));
+    const noExtensions = extensions === false;
+    const extensionsSpec = Array.isArray(extensions)
+        ? parseExtensionsSpec(extensions, configCwd)
+        : undefined;
+    const keepNames = extensionsSpec?.names ?? new Set();
+    // `exclude_extensions:` is a denylist applied AFTER the include set — exclude wins.
+    // Plain canonical names only (case-insensitive). Note: excluded extensions'
+    // factories still run once during reload() (see comment above) — exclusion
+    // suppresses handler binding and tool registration; it is not a sandbox.
+    const excludeNames = new Set((excludeExtensions ?? []).map((n) => n.toLowerCase()));
+    const hasExcludes = excludeNames.size > 0;
+    // The override filters loaded extensions down to `keepNames` minus `excludeNames`.
+    // It's only needed when we're neither loading everything without excludes
+    // (`extensions: true` or a `"*"` wildcard) nor nothing (`noExtensions`).
+    const loadAll = extensions === true || extensionsSpec?.wildcard === true;
+    const additionalExtensionPaths = extensionsSpec?.paths.length ? extensionsSpec.paths : undefined;
+    // Pre-filter discovered set, captured by the override — the exclude-typo warning
+    // must compare against this, not the surviving set (absence from survivors is
+    // an exclude *succeeding*).
+    let discoveredNames;
+    const extensionsOverride = noExtensions || (loadAll && !hasExcludes)
+        ? undefined
+        : (base) => {
+            discoveredNames = new Set(base.extensions.flatMap((e) => extensionCanonicalNames(e.path)));
+            return {
+                ...base,
+                extensions: base.extensions.filter((e) => {
+                    const canons = extensionCanonicalNames(e.path);
+                    if (canons.some((n) => excludeNames.has(n)))
+                        return false; // exclude wins
+                    return loadAll || canons.some((n) => keepNames.has(n));
+                }),
+            };
+        };
+    const loader = new DefaultResourceLoader({
+        cwd: configCwd,
+        agentDir,
+        noExtensions,
+        additionalExtensionPaths,
+        extensionsOverride,
+        noSkills,
+        noPromptTemplates: true,
+        noThemes: true,
+        noContextFiles: true,
+        systemPromptOverride: () => systemPrompt,
+        appendSystemPromptOverride: () => [],
+    });
+    await runInChildSessionContext(() => loader.reload());
+    // Plain entries in `tools:` are expected to be built-in names (extension tools
+    // go through `ext:`), so an unknown name there is unambiguously a typo. Previously
+    // this produced a silently broken agent (#75) — pi-mono accepted the bogus name
+    // into the allowlist, then dropped it at registration with no signal back.
+    if (agentConfig?.builtinToolNames?.length) {
+        const knownBuiltins = new Set(BUILTIN_TOOL_NAMES);
+        for (const name of agentConfig.builtinToolNames) {
+            if (!knownBuiltins.has(name)) {
+                options.onToolActivity?.({
+                    type: "end",
+                    toolName: `tools-error:tool "${name}" requested by agent "${type}" is not a known built-in`,
+                });
+            }
+        }
+    }
+    // A subagent spawns mid-task, so a bad `extensions:`/`ext:` entry warns rather
+    // than aborts. Two distinct misconfigurations to catch:
+    //   - `extensions: [foo]` but no extension named foo was discovered (typo or
+    //     path that failed to load — path entries fold their canonical name into
+    //     `keepNames`, so this covers them too).
+    //   - `tools: ext:foo` but foo isn't in the loaded set (because `extensions:`
+    //     didn't include it). Since v0.9, `ext:` no longer pulls extensions in;
+    //     loading is `extensions:`-authoritative.
+    // An exclude_extensions: alongside extensions: false is contradictory — nothing
+    // loads, so there is nothing to exclude.
+    if (hasExcludes && noExtensions) {
+        options.onToolActivity?.({
+            type: "end",
+            toolName: `extension-error:exclude_extensions has no effect for agent "${type}" — extensions: false loads nothing`,
+        });
+    }
+    // Exclude typo check: compares against the PRE-filter discovered set (an excluded
+    // name absent from the surviving set is the exclude working as intended). Also
+    // flags path-like and "*" entries — excludes are plain names only.
+    if (hasExcludes && discoveredNames) {
+        for (const name of excludeNames) {
+            if (!discoveredNames.has(name)) {
+                options.onToolActivity?.({
+                    type: "end",
+                    toolName: `extension-error:exclude_extensions: "${name}" for agent "${type}" did not match any discovered extension`,
+                });
+            }
+        }
+    }
+    if (keepNames.size > 0 || extNames.size > 0) {
+        const survivingNames = new Set(loader.getExtensions().extensions.flatMap((e) => extensionCanonicalNames(e.path)));
+        for (const name of keepNames) {
+            if (!survivingNames.has(name)) {
+                options.onToolActivity?.({
+                    type: "end",
+                    toolName: excludeNames.has(name)
+                        ? `extension-error:extension "${name}" is in both extensions: and exclude_extensions: for agent "${type}" — exclude wins`
+                        : `extension-error:extension "${name}" requested by agent "${type}" was not loaded`,
+                });
+            }
+        }
+        for (const name of extNames) {
+            if (!survivingNames.has(name)) {
+                options.onToolActivity?.({
+                    type: "end",
+                    toolName: `extension-error:ext:${name} referenced by agent "${type}" but extension "${name}" is not loaded (check extensions:/exclude_extensions:)`,
+                });
+            }
+        }
+    }
+    // Resolve model: explicit option > config.model > parent model
+    const model = options.model ?? resolveDefaultModel(ctx.model, ctx.modelRegistry, agentConfig?.model);
+    // Resolve thinking level: explicit option > agent config > undefined (inherit)
+    const thinkingLevel = options.thinkingLevel ?? agentConfig?.thinking;
+    const disallowedSet = agentConfig?.disallowedTools
+        ? new Set(agentConfig.disallowedTools)
+        : undefined;
+    // Nested delegation tools (opt-in, ownership-scoped). Empty unless the agent
+    // set `allowed_subagents` and a nestedRuntime was provided — and never when
+    // isolated. Their names collide with EXCLUDED_TOOL_NAMES by design, so the
+    // scoping below re-admits them explicitly (registry deny + active-set narrow).
+    const effectiveMaxDepth = options.nestedRuntime?.maxSubagentDepth ?? getMaxSubagentDepth();
+    // At (or past) the cap this agent can never spawn, so it can never own a child
+    // to fetch from or steer either — inject nothing rather than three tools whose
+    // every call is an error. This is also what makes `maxSubagentDepth` 0/1 mean
+    // "nesting off" instead of "nesting always fails".
+    const nestedRuntime = options.nestedRuntime && options.nestedRuntime.depth < effectiveMaxDepth
+        ? options.nestedRuntime
+        : undefined;
+    const nestedTools = agentConfig?.allowedSubagents && nestedRuntime && !options.isolated
+        ? createNestedSubagentTools({
+            manager: nestedRuntime.manager,
+            pi: options.pi,
+            parentAgentId: nestedRuntime.parentAgentId,
+            depth: nestedRuntime.depth,
+            maxSubagentDepth: effectiveMaxDepth,
+            allowedSubagents: agentConfig.allowedSubagents,
+            configCwd,
+        })
+        : [];
+    const nestedToolNames = new Set(nestedTools.map(tool => tool.name));
+    // The `agent({ schema })` contract: this child reports its answer by calling
+    // StructuredOutput, and `structuredJson` below is what the caller reads. The
+    // schema was already compiled by whoever asked for it, so a bad one failed
+    // before any of this ran.
+    const structuredCapture = options.structuredOutput ? createStructuredCapture() : undefined;
+    const structuredTools = options.structuredOutput && structuredCapture
+        ? [createStructuredOutputTool(options.structuredOutput, structuredCapture)]
+        : [];
+    const structuredToolNames = new Set(structuredTools.map(tool => tool.name));
+    // Re-admitted together at every gate below. Kept as one set so a new injected
+    // tool cannot be added to some of the three gates and forgotten at the rest.
+    //
+    // `disallowed_tools` is applied HERE rather than at the gates, because the two
+    // kinds answer to it differently: a nested delegation tool is an opt-in the
+    // agent's own frontmatter can take back, while StructuredOutput exists only
+    // because this call asked for a schema — removing it would make the request
+    // unsatisfiable by construction rather than merely restricted.
+    const readmitToolNames = new Set([
+        ...[...nestedToolNames].filter(name => !disallowedSet?.has(name)),
+        ...structuredToolNames,
+    ]);
+    // ─── Tool scoping ───────────────────────────────────────────────────────
+    //
+    // Some extensions register their tools ASYNCHRONOUSLY, long after the
+    // `loader.reload()` above: pi-mcp calls registerTool from `session_start`
+    // (once its MCP servers connect), context-mode from `before_agent_start`.
+    // That is deliberate on their part — eagerly spawning an MCP bridge during
+    // extension discovery orphans child processes on pi's non-agent code paths
+    // (--help, config, trust probing).
+    //
+    // So the tool set cannot be snapshotted here. pi's `allowedToolNames` gates
+    // tool *registration* (`_refreshToolRegistry`'s `isAllowedTool`), not merely
+    // the active set, and is frozen at construction — a name absent from the
+    // snapshot is dropped forever, even once the tool actually registers (#125).
+    //
+    // Whenever extensions are in play we therefore:
+    //   - leave `allowedToolNames` unset, so pi's live gate admits tools whenever
+    //     they register;
+    //   - express the name-stable, permanent part of the scope (our own
+    //     orchestration tools, built-ins the agent didn't ask for, and
+    //     `disallowedTools`) as `excludeTools`, which pi re-applies on every
+    //     registry refresh;
+    //   - enforce `ext:` narrowing on the ACTIVE set via the live `inScope()`
+    //     predicate installed after bind — the active set is what the LLM sees,
+    //     so a registry tool that is never activated is invisible and uncallable.
+    //
+    // `noExtensions`/`isolated` keeps the historical static allowlist: nothing
+    // async can appear there, and a hard registry gate is the correct boundary.
+    const builtinToolNameSet = new Set(toolNames);
+    let sessionTools;
+    let sessionExcludeTools;
+    if (noExtensions) {
+        // Strict allowlist: built-ins the agent asked for, plus any opt-in nested
+        // tools (whose names would otherwise be dropped as EXCLUDED_TOOL_NAMES).
+        sessionTools = [
+            ...toolNames.filter((t) => !EXCLUDED_TOOL_NAMES.includes(t) && !disallowedSet?.has(t)),
+            ...[...nestedToolNames].filter((t) => !disallowedSet?.has(t)),
+            // Not filtered through `disallowedSet`, unlike the nested tools above:
+            // the caller asked for a schema, and removing the only tool that can
+            // satisfy it would make the request unsatisfiable by construction rather
+            // than merely restricted.
+            ...structuredToolNames,
+        ];
+    }
+    else {
+        // Deny the orchestration tools EXCEPT the nested ones this agent opted into —
+        // those are injected as customTools and must survive the registry gate.
+        const denyTools = new Set(EXCLUDED_TOOL_NAMES.filter((t) => !nestedToolNames.has(t)));
+        // Keep only the built-ins the agent asked for — deny the rest.
+        for (const name of BUILTIN_TOOL_NAMES) {
+            if (!builtinToolNameSet.has(name))
+                denyTools.add(name);
+        }
+        if (disallowedSet) {
+            // disallowed_tools wins even over an opt-in nested tool of the same name.
+            // Not over StructuredOutput, though — see the allowlist branch above.
+            for (const name of disallowedSet) {
+                if (!structuredToolNames.has(name))
+                    denyTools.add(name);
+            }
+        }
+        sessionExcludeTools = [...denyTools];
+    }
+    const settingsManager = SettingsManager.create(configCwd, agentDir);
+    const configuredSessionDir = resolveConfiguredSessionDir(agentConfig?.sessionDir, effectiveCwd);
+    const defaultSessionDir = process.env.PI_CODING_AGENT_SESSION_DIR ?? settingsManager.getSessionDir?.();
+    // Frontmatter wins when it says anything; otherwise the project default,
+    // which `rememberAgents` supplies for top-level agents only. Same precedence
+    // as `outputTranscript`.
+    const persistSession = agentConfig?.persistSession ?? (options.nested ? false : rememberAgents);
+    const sessionManager = options.resumeSessionFile
+        // Reopening an existing conversation: the file already carries its own
+        // header (cwd, parent) and history, so none of the create-time options
+        // apply. `sessionDir` still matters for a later /new or /branch off it.
+        ? SessionManager.open(options.resumeSessionFile, configuredSessionDir ?? defaultSessionDir)
+        : persistSession
+            ? SessionManager.create(effectiveCwd, configuredSessionDir ?? defaultSessionDir, {
+                // Optional metadata — it only nests the subagent under its spawner in
+                // `/resume`. Until `rememberAgents` this ran solely for the rare
+                // `persist_session: true` agent; now it runs for every spawn, so a
+                // context without a session manager (a bare programmatic ctx) must
+                // still persist rather than take the whole spawn down.
+                parentSession: ctx.sessionManager?.getSessionFile?.(),
+            })
+            : SessionManager.inMemory(effectiveCwd);
+    // Pi 0.80.8 replaced createAgentSession's modelRegistry option with
+    // modelRuntime, but ExtensionContext still exposes only the registry facade.
+    // Pass both so the full supported Pi range retains the parent's providers.
+    const parentModelRuntime = ctx.modelRegistry.runtime;
+    const sessionOpts = {
+        cwd: effectiveCwd,
+        agentDir,
+        sessionManager,
+        settingsManager,
+        modelRegistry: ctx.modelRegistry,
+        // `as never` is what keeps this assignable across the supported Pi range:
+        // pre-0.80.8 the field exists only via the `modelRuntime?: unknown` shim
+        // above, while newer Pi types it as `ModelRuntime` — a shape an opaque
+        // `unknown` read off the private facade field can never satisfy.
+        ...(parentModelRuntime !== undefined && { modelRuntime: parentModelRuntime }),
+        model,
+        tools: sessionTools,
+        customTools: [...nestedTools, ...structuredTools],
+        resourceLoader: loader,
+    };
+    if (sessionExcludeTools) {
+        sessionOpts.excludeTools = sessionExcludeTools;
+    }
+    if (thinkingLevel) {
+        sessionOpts.thinkingLevel = thinkingLevel;
+    }
+    const { session } = await runInChildSessionContext(() => createAgentSession(sessionOpts));
+    const baseSessionName = agentConfig?.name ?? type;
+    session.setSessionName(options.agentId ? `${baseSessionName}#${options.agentId.slice(0, 8)}` : baseSessionName);
+    // Bind extensions so that session_start fires and extensions can initialize
+    // (e.g. loading credentials, setting up state). Tool gating already happened
+    // at session construction via the `tools:` allowlist above — no separate
+    // post-bind filter is needed. All ExtensionBindings fields are optional.
+    await session.bindExtensions({
+        onError: (err) => {
+            options.onToolActivity?.({
+                type: "end",
+                toolName: `extension-error:${err.extensionPath}`,
+            });
+        },
+    });
+    installServiceTierPayload(session, agentConfig?.serviceTier);
+    // With `allowedToolNames` unset, the registry is scoped by `excludeTools` but
+    // the ACTIVE set still needs managing: pi activates only its four default
+    // built-ins at turn 1, and `ext:` narrowing has no registry-level expression
+    // (we can't deny the name of a tool that hasn't registered yet). Both are
+    // handled below by re-deriving scope from the loader's live extension maps —
+    // `registerTool` writes into those same maps, so late arrivals are judged too.
+    if (!noExtensions) {
+        installExtensionToolScope(session, {
+            loader,
+            toolNames,
+            disallowedSet,
+            extNames,
+            narrowing,
+            readmitToolNames,
+        });
+    }
+    options.onSessionCreated?.(session);
+    // Track turns for graceful max_turns enforcement
+    let turnCount = 0;
+    const maxTurns = resolveEffectiveMaxTurns(type, options.maxTurns);
+    let softLimitReached = false;
+    let aborted = false;
+    let currentMessageText = "";
+    const unsubTurns = session.subscribe((event) => {
+        if (event.type === "turn_end") {
+            turnCount++;
+            options.onTurnEnd?.(turnCount);
+            if (maxTurns != null) {
+                if (!softLimitReached && turnCount >= maxTurns) {
+                    softLimitReached = true;
+                    session.steer("You have reached your turn limit. Wrap up immediately — provide your final answer now.");
+                }
+                else if (softLimitReached && turnCount >= maxTurns + graceTurns) {
+                    aborted = true;
+                    session.abort();
+                }
+            }
+        }
+        if (event.type === "message_start") {
+            currentMessageText = "";
+        }
+        if (event.type === "message_update" && event.assistantMessageEvent.type === "text_delta") {
+            currentMessageText += event.assistantMessageEvent.delta;
+            options.onTextDelta?.(event.assistantMessageEvent.delta, currentMessageText);
+        }
+        if (event.type === "tool_execution_start") {
+            options.onToolActivity?.({ type: "start", toolName: event.toolName });
+        }
+        if (event.type === "tool_execution_end") {
+            options.onToolActivity?.({ type: "end", toolName: event.toolName });
+        }
+        if (event.type === "message_end" && event.message.role === "assistant") {
+            const u = event.message.usage;
+            if (u)
+                options.onAssistantUsage?.({
+                    input: u.input ?? 0,
+                    output: u.output ?? 0,
+                    cacheWrite: u.cacheWrite ?? 0,
+                    cacheRead: u.cacheRead ?? 0,
+                    cost: u.cost?.total ?? 0,
+                });
+        }
+        if (event.type === "compaction_end" && !event.aborted && event.result) {
+            options.onCompaction?.({ reason: event.reason, tokensBefore: event.result.tokensBefore });
+        }
+    });
+    const collector = collectResponseText(session);
+    const cleanupAbort = forwardAbortSignal(session, options.signal);
+    // Build the effective prompt: optionally prepend parent context
+    let effectivePrompt = prompt;
+    if (options.inheritContext) {
+        const parentContext = buildParentContext(ctx);
+        if (parentContext) {
+            effectivePrompt = parentContext + prompt;
+        }
+    }
+    // Boundary for the history fallback: only assistant text produced from here
+    // on counts as this run's output (a fresh session, so usually 0).
+    const startLen = session.messages.length;
+    let structuredRetried = false;
+    try {
+        await session.prompt(effectivePrompt);
+        // One more prompt when a schema was asked for and nothing usable came back
+        // — the model answered in prose, or only ever called the tool invalidly.
+        // Inside this `try`, so the turn tracking, the text collector and above all
+        // the abort forwarding are still live: torn down first, a retry would be
+        // unkillable.
+        if (structuredCapture !== undefined && structuredCapture.json === undefined
+            && !aborted && options.signal?.aborted !== true) {
+            structuredRetried = true;
+            await session.prompt(structuredRetryPrompt(structuredCapture));
+        }
+    }
+    finally {
+        unsubTurns();
+        collector.unsubscribe();
+        cleanupAbort();
+    }
+    const responseText = collector.getText().trim() || getLastAssistantText(session, startLen);
+    // A child asked for structured output that never gave any has failed, however
+    // articulate its prose was. Reported through `failure` so it travels the same
+    // path as a provider error rather than arriving as a successful empty answer.
+    const structuredFailure = structuredCapture !== undefined && structuredCapture.json === undefined
+        ? structuredCapture.lastError !== undefined
+            ? `The agent's StructuredOutput call did not match the required schema: ${structuredCapture.lastError}`
+            : "The agent did not report its answer through StructuredOutput."
+        : undefined;
+    return {
+        responseText,
+        session,
+        aborted,
+        steered: softLimitReached,
+        failure: finalTurnError(session, startLen) ?? structuredFailure,
+        ...(structuredCapture?.json !== undefined ? { structuredJson: structuredCapture.json } : {}),
+        ...(structuredRetried ? { structuredRetried } : {}),
+    };
+}
+/**
+ * Send a new prompt to an existing session (resume).
+ */
+export async function resumeAgent(session, prompt, options = {}) {
+    // Boundary for the history fallback: the session already holds prior turns,
+    // so only assistant text produced by THIS resume prompt counts as its output
+    // — a failed resume must not surface the previous turn's answer (#144).
+    const startLen = session.messages.length;
+    const collector = collectResponseText(session);
+    const cleanupAbort = forwardAbortSignal(session, options.signal);
+    const unsubEvents = (options.onToolActivity || options.onAssistantUsage || options.onCompaction)
+        ? session.subscribe((event) => {
+            if (event.type === "tool_execution_start")
+                options.onToolActivity?.({ type: "start", toolName: event.toolName });
+            if (event.type === "tool_execution_end")
+                options.onToolActivity?.({ type: "end", toolName: event.toolName });
+            if (event.type === "message_end" && event.message.role === "assistant") {
+                const u = event.message.usage;
+                if (u)
+                    options.onAssistantUsage?.({
+                        input: u.input ?? 0,
+                        output: u.output ?? 0,
+                        cacheWrite: u.cacheWrite ?? 0,
+                        cacheRead: u.cacheRead ?? 0,
+                        cost: u.cost?.total ?? 0,
+                    });
+            }
+            if (event.type === "compaction_end" && !event.aborted && event.result) {
+                options.onCompaction?.({ reason: event.reason, tokensBefore: event.result.tokensBefore });
+            }
+        })
+        : () => { };
+    try {
+        await session.prompt(prompt);
+    }
+    finally {
+        collector.unsubscribe();
+        unsubEvents();
+        cleanupAbort();
+    }
+    return {
+        text: collector.getText().trim() || getLastAssistantText(session, startLen),
+        failure: finalTurnError(session, startLen),
+    };
+}
+/**
+ * Send a steering message to a running subagent.
+ * The message will interrupt the agent after its current tool execution.
+ */
+export async function steerAgent(session, message) {
+    await session.steer(message);
+}
+/**
+ * Get the subagent's conversation messages as formatted text.
+ */
+export function getAgentConversation(session) {
+    const parts = [];
+    for (const msg of session.messages) {
+        if (msg.role === "user") {
+            const text = typeof msg.content === "string"
+                ? msg.content
+                : extractText(msg.content);
+            if (text.trim())
+                parts.push(`[User]: ${text.trim()}`);
+        }
+        else if (msg.role === "assistant") {
+            const textParts = [];
+            const toolCalls = [];
+            for (const c of msg.content) {
+                if (c.type === "text" && c.text)
+                    textParts.push(c.text);
+                else if (c.type === "toolCall")
+                    toolCalls.push(`  Tool: ${c.name ?? c.toolName ?? "unknown"}`);
+            }
+            if (textParts.length > 0)
+                parts.push(`[Assistant]: ${textParts.join("\n")}`);
+            if (toolCalls.length > 0)
+                parts.push(`[Tool Calls]:\n${toolCalls.join("\n")}`);
+        }
+        else if (msg.role === "toolResult") {
+            const text = extractText(msg.content);
+            const truncated = text.length > 200 ? text.slice(0, 200) + "..." : text;
+            parts.push(`[Tool Result (${msg.toolName})]: ${truncated}`);
+        }
+    }
+    return parts.join("\n\n");
+}