@tangle-network/agent-runtime 0.71.1 → 0.73.0

package/dist/worktree-DaxOvw-C.d.ts

@@ -0,0 +1,702 @@
+import { ChildProcess } from 'node:child_process';
+import { DefaultVerdict } from '@tangle-network/agent-eval';
+import { AgentProfile } from '@tangle-network/agent-interface';
+import { BackendType } from '@tangle-network/sandbox';
+import { L as LoopTokenUsage, b as RuntimeHooks } from './types-JufmXF2a.js';
+
+/**
+ * @experimental
+ *
+ * Subprocess wrappers for the local coding-harness CLIs installed in the
+ * sandbox image (claude-code, codex, opencode). Used by the in-process
+ * delegation executor (`createInProcessExecutor`) so a `delegate_code` call
+ * spawns a real harness on a real git worktree instead of provisioning a
+ * sibling sandbox.
+ *
+ * All harness invocations:
+ *   - run with `cwd` set to the worktree
+ *   - inherit env from the parent (the MCP server inside the sandbox has
+ *     the harness's auth already)
+ *   - capture stdout/stderr
+ *   - support cancellation via AbortSignal
+ *   - enforce a wall-clock timeout
+ */
+
+/** Local coding harness available inside the sandbox. */
+type LocalHarness = 'claude' | 'codex' | 'opencode';
+/** @experimental */
+interface RunLocalHarnessOptions {
+    harness: LocalHarness;
+    /** Working directory for the subprocess (typically a worktree path). */
+    cwd: string;
+    /** Prompt forwarded as the harness CLI's task argument. */
+    taskPrompt: string;
+    /**
+     * Pre-built command + args (e.g. from `harnessInvocation` so the full authored
+     * `AgentProfile` — systemPrompt + model — reaches the harness). When set it OVERRIDES the
+     * default prompt-only `buildArgs(taskPrompt)` path; `command` defaults to the harness's
+     * default binary when only `args` is supplied. When absent the legacy prompt-only shape
+     * is used unchanged.
+     */
+    invocation?: {
+        command?: string;
+        args: ReadonlyArray<string>;
+    };
+    /** Wall-clock kill deadline (ms). Default 5 min. Subprocess SIGTERMed on expiry. */
+    timeoutMs?: number;
+    /** Caller cancellation. SIGTERM is sent on abort. */
+    signal?: AbortSignal;
+    /** Override env (defaults to inheriting from the parent). */
+    env?: NodeJS.ProcessEnv;
+    /**
+     * Test seam — inject a custom spawner so unit tests can mock the
+     * subprocess without touching the OS. Defaults to node's `child_process.spawn`.
+     */
+    spawn?: (command: string, args: ReadonlyArray<string>, opts: {
+        cwd: string;
+        env: NodeJS.ProcessEnv;
+        stdio: 'pipe';
+    }) => ChildProcess;
+}
+/** @experimental */
+interface LocalHarnessResult {
+    /** OS exit code. `null` when killed before exit. */
+    exitCode: number | null;
+    /** Concatenated stdout. */
+    stdout: string;
+    /** Concatenated stderr. */
+    stderr: string;
+    /** Set when the process exited via signal (timeout / abort). */
+    killedBySignal: NodeJS.Signals | null;
+    /** Wall-clock duration ms (spawn → exit). */
+    durationMs: number;
+    /** Set when timeoutMs elapsed before exit. */
+    timedOut: boolean;
+}
+/**
+ * Spawn a local coding harness CLI as a subprocess + collect its output.
+ *
+ * NOT responsible for parsing the harness's output or extracting a diff —
+ * the in-process executor's `streamPrompt` orchestrates `git diff` against
+ * the worktree after this resolves. This function is intentionally narrow:
+ * spawn, wait, capture, return.
+ *
+ * Fails loud — throws when:
+ *   - `cwd` doesn't exist (subprocess emits ENOENT; surfaced as Error)
+ *   - the harness binary is not on PATH (ENOENT)
+ *
+ * Does NOT throw when:
+ *   - the subprocess exits non-zero (`result.exitCode` carries the code)
+ *   - the subprocess is aborted / timed out (`result.killedBySignal` /
+ *     `result.timedOut` carries the reason)
+ *
+ * @experimental
+ */
+declare function runLocalHarness(options: RunLocalHarnessOptions): Promise<LocalHarnessResult>;
+
+/**
+ * @experimental
+ *
+ * Recursive execution atom — the FROZEN type surface (the keystone contract).
+ *
+ * One self-similar `Agent` atom runs inside a budget-conserving reactive `Scope`,
+ * orchestrated by a `Supervisor` over an event-sourced `SpawnJournal`. A leaf is an
+ * `Agent` that never calls `scope.spawn`; a driver is an `Agent` that spawns and runs
+ * a policy over its children's streaming results.
+ *
+ * Two invariants the surface exists to make enforceable:
+ *  - Budget is an atomically-reserved CONSERVED pool, so `Σk(treatment) ≡ Σk(blind)` by
+ *    construction (reserve-on-spawn, refund-unspent-on-settle, fail-closed admission).
+ *  - The journal records a content-addressed `outRef` per child result, so replay
+ *    rehydrates the exact `Settled` the driver branched on (the replay invariant below).
+ *
+ * The leaf RUNTIME is one OPEN `Executor` interface, not a closed `inline|sandbox|cli`
+ * union the call site switches on. The built-ins (router/inline, sandbox, cli) are the
+ * initial IMPLEMENTATIONS; any user agent is first-class the moment it implements the
+ * interface. The interface IS the extension point — no per-vendor adapters live here.
+ *
+ * Layering: substrate types (`DefaultVerdict`) come from `@tangle-network/agent-eval`;
+ * runtime-shaped types (everything else) live here. Pure types/interfaces only — this
+ * module typechecks standalone and is imported by every keystone impl.
+ */
+
+/**
+ * One self-similar atom. A leaf is an `Agent` that never calls `scope.spawn`; a driver
+ * is an `Agent` whose `act` spawns children and reacts to them via `scope.next()`. An
+ * analyst is an `Agent` whose task is "read these traces → findings" — `where` it runs
+ * is its executor, not a separate type.
+ *
+ * `act` MUST be replay-safe: it may read `verdict`, `spent`, and `out` (rehydrated by
+ * `outRef`) off each `Settled`; it MUST NOT read `Date.now`, `Math.random`, or any
+ * unordered collection. `scope.next()` delivers strictly in recorded `seq` order.
+ */
+interface Agent<Task, Out> {
+    readonly name: string;
+    act(task: Task, scope: Scope<Out>): Promise<Out>;
+}
+/**
+ * The leaf runtime — ONE open interface, not a closed union. `execute` returns a
+ * `Promise<ExecutorResult>` for one-shot executors OR an `AsyncIterable<UsageEvent>` for
+ * streaming ones; a streaming executor reports incremental normalized usage as it runs
+ * (the budget pool reconciles against it) and exposes its terminal artifact via
+ * `resultArtifact()`. Both shapes normalize usage to `UsageEvent` so the conserved pool
+ * meters every runtime identically.
+ *
+ * Built-in implementations (in `runtime.ts`, NOT variants here): router/inline (a direct
+ * Router/HTTP inference call, no box), sandbox (COMPOSES `runLoop` as a leaf, forwarding
+ * PR #150's optional `lineage` passthrough — does NOT reinvent checkpoint/fork), cli
+ * (Halo/RLM subprocess; `budgetExempt`, excluded from equal-k by construction). A user's
+ * own agent (mastra/agno/raw HTTP/anything) is first-class by implementing this interface.
+ */
+interface Executor<Out> {
+    /** Stable runtime tag for traces + the equal-k exemption check. */
+    readonly runtime: Runtime;
+    /**
+     * When true, this executor's spend is NOT metered against the conserved pool and its
+     * iterations are excluded from the equal-k assertion (a `cli` subprocess without
+     * token accounting). Fail-loud everywhere else: a metered executor MUST report usage.
+     */
+    readonly budgetExempt?: boolean;
+    /**
+     * One-shot → resolves a `ExecutorResult`; streaming → yields incremental `UsageEvent`s and
+     * the terminal artifact is read from `resultArtifact()` after the stream drains.
+     * `signal` is the spawn-scoped abort (chains the acquire lifecycle for sandbox).
+     */
+    execute(task: unknown, signal: AbortSignal): Promise<ExecutorResult<Out>> | AsyncIterable<UsageEvent>;
+    /**
+     * Optional inbox: receive an out-of-band message from the driver mid-run (the `send`/`steer_agent`
+     * verb). A streaming executor drains pending messages between turns and folds them into the next
+     * step (a steer / interrupt / resume). A one-shot executor that can't be steered mid-flight omits
+     * this; `Scope.send` then returns `false` for it. Never throws — a malformed message is the
+     * executor's to ignore.
+     */
+    deliver?(msg: unknown): void;
+    /**
+     * Tear the executor's resources down. `grace` mirrors the OTP shutdown spec
+     * (`'brutalKill'` = immediate, a number = ms grace, `'infinity'` = await clean exit).
+     */
+    teardown(grace: number | 'brutalKill' | 'infinity'): Promise<{
+        destroyed: boolean;
+    }>;
+    /**
+     * The replay source (B1): the content-addressed `outRef` + the materialized output the
+     * driver branched on, its verdict, and the conserved spend. Read once, after settle.
+     */
+    resultArtifact(): {
+        outRef: string;
+        out: Out;
+        verdict?: DefaultVerdict;
+        spent: Spend;
+    };
+    /**
+     * A driver-executor's OWN-inference subtree total (rolled up from its nested tree's `metered`
+     * events) — the parent scope journals it as a `metered` event for this node on settle, on BOTH
+     * the done AND the down/crash paths, so a crashed sub-driver's partial inference still re-homes
+     * (the pool already debited it via `observe`; the journal must match). NOT reconciled, so it never
+     * trips the reservation clamp. Read on settle, valid after `execute` resolves OR throws. Leaf
+     * executors omit it (returns `undefined`).
+     */
+    metered?(): Spend | undefined;
+}
+/** Terminal artifact of a one-shot `Executor.execute`. */
+interface ExecutorResult<Out> {
+    outRef: string;
+    out: Out;
+    verdict?: DefaultVerdict;
+    spent: Spend;
+}
+/**
+ * Normalized usage event — the single channel every executor reports through, so the
+ * conserved pool meters all runtimes identically. `tokens` carries `LoopTokenUsage`'s
+ * `{ input, output }`; `usd` is a SEPARATE channel (never folded into tokens).
+ */
+type UsageEvent = {
+    kind: 'tokens';
+    input: number;
+    output: number;
+} | {
+    kind: 'cost';
+    usd: number;
+} | {
+    kind: 'iteration';
+};
+/** The runtime tag of a `Executor` impl. Open by intent — `string` so a BYO executor
+ *  names its own runtime; the built-ins use these literals. */
+type Runtime = 'router' | 'inline' | 'sandbox' | 'cli' | (string & {});
+/**
+ * `AgentProfile` does NOT carry a `harness`/backend field — `harness` lives on the
+ * sandbox SDK's `BackendConfig`, not the portable profile. So an agent is mapped to its
+ * executor through this MINIMAL wrapper, never by fabricating a field onto `AgentProfile`.
+ *
+ * Resolution (in `runtime.ts`):
+ *  - `executor` present        → BYO: use it verbatim (a user's own `Executor`).
+ *  - `harness === null`        → router/inline: a direct Router call, no box.
+ *  - `harness` is a `BackendType` → sandbox: compose `runLoop` against `profile` on that backend.
+ * Fail loud on an unresolvable spec (no executor and an unknown harness).
+ */
+interface AgentSpec {
+    readonly profile: AgentProfile;
+    /** `null` selects router/inline; a `BackendType` selects the sandboxed harness. */
+    readonly harness: BackendType | null;
+    /** Bring-your-own executor: when set, overrides harness-based resolution entirely. */
+    readonly executor?: Executor<unknown>;
+}
+/**
+ * Builds a fresh `Executor` for one spawn from the resolved spec. Per-spawn (not
+ * shared) so each child owns its own box/abort/teardown lifecycle. A BYO factory lets a
+ * user supply construction args without pre-instantiating.
+ */
+type ExecutorFactory<Out> = (spec: AgentSpec, ctx: ExecutorContext) => Executor<Out>;
+/** Construction context handed to a `ExecutorFactory` — the seams a built-in needs
+ *  (sandbox client for the sandbox executor, router config for router/inline) without
+ *  the factory reaching into module globals. */
+interface ExecutorContext {
+    readonly signal: AbortSignal;
+    /** Opaque seams the registry threads through; a built-in narrows what it needs. */
+    readonly seams: Readonly<Record<string, unknown>>;
+}
+/**
+ * The OPEN resolver: maps an `AgentSpec` to a `ExecutorFactory`. The default
+ * registry resolves the three built-ins AND accepts a BYO `executor`/factory; callers
+ * register more runtimes by name. NOT a closed switch — registration is the extension
+ * point, mirroring the open `Executor` interface.
+ */
+interface ExecutorRegistry {
+    /** Register a factory for a named runtime. Throws on a duplicate name (fail loud). */
+    register<Out>(runtime: Runtime, factory: ExecutorFactory<Out>): void;
+    /**
+     * Resolve a spec to a factory. Precedence: a BYO `spec.executor` → a trivial factory
+     * returning it; else `harness === null` → the `'router'` factory; else a registered
+     * factory for the harness-derived runtime. Returns a typed outcome — the caller
+     * inspects `succeeded` before `value` (no silent fallback).
+     */
+    resolve<Out>(spec: AgentSpec): {
+        succeeded: true;
+        value: ExecutorFactory<Out>;
+    } | {
+        succeeded: false;
+        error: string;
+    };
+}
+/** A budget envelope on a spawn or the root. All ceilings; the pool reserves against them. */
+interface Budget {
+    readonly maxIterations: number;
+    readonly maxTokens: number;
+    readonly maxUsd?: number;
+    readonly deadlineMs?: number;
+}
+/** Conserved spend, reconciled from the normalized `UsageEvent` stream. Tokens and usd
+ *  are separate channels (never folded). */
+interface Spend {
+    iterations: number;
+    tokens: LoopTokenUsage;
+    usd: number;
+    ms: number;
+}
+/** OTP child-spec restart class. */
+type Restart = 'temporary' | 'transient' | 'permanent';
+/** `'acquiring'` is first-class (M1): a node spends real time + reaps an orphan box
+ *  during sandbox acquire BEFORE it is `running`, so abort must be defined over it. */
+type NodeStatus = 'pending' | 'acquiring' | 'running' | 'done' | 'failed' | 'cancelled';
+/** Deterministic node id — `${parent}:s${seq}` from the cursor order, never wall-clock. */
+type NodeId = string;
+interface SpawnOpts {
+    readonly budget: Budget;
+    readonly label: string;
+    readonly restart?: Restart;
+    /** Teardown grace handed to the executor when this node is reaped. */
+    readonly shutdown?: number | 'brutalKill' | 'infinity';
+}
+/**
+ * A live child handle. `abort()` is defined over the ACQUIRE lifecycle: it chains into
+ * the `acquireSandbox` signal and reaps a find-by-name orphan box, so a node aborted
+ * mid-acquire never leaks (M1).
+ */
+interface Handle<Out> {
+    readonly id: NodeId;
+    readonly label: string;
+    readonly status: NodeStatus;
+    abort(reason?: string): void;
+    /** Phantom: binds the handle to the child's output type so `spawn<C>` returns a
+     *  `Handle<C>` distinct from a `Handle<other>`. Type-only — never present at runtime. */
+    readonly __out?: Out;
+}
+/**
+ * A settled child, delivered by `scope.next()`. `seq` is the monotonic cursor order
+ * `next()` yielded this settlement (B2) — NOT wall-clock — and replay delivers strictly
+ * in `seq` order. `outRef` rehydrates `out` from the `ResultBlobStore` on replay.
+ */
+type Settled<Out> = {
+    kind: 'done';
+    handle: Handle<Out>;
+    out: Out;
+    outRef: string;
+    verdict?: DefaultVerdict;
+    spent: Spend;
+    seq: number;
+} | {
+    kind: 'down';
+    handle: Handle<Out>;
+    reason: string;
+    /** True = infrastructure failure (excluded from merge `n` / equal-k), not a bad result. */
+    infra: boolean;
+    restartCount: number;
+    seq: number;
+};
+/**
+ * The budget-conserving reactive scope an `Agent.act` runs inside. `spawn` reserves
+ * budget atomically from the shared pool and fails closed when the pool cannot cover it.
+ * `next()` waits for one settlement from this scope's live set; `view` reads live state,
+ * not the replay log.
+ */
+interface Scope<Out> {
+    /**
+     * Spawn a child. Reserves `opts.budget` from the conserved pool atomically; refunds the
+     * unspent remainder on settle. Returns a typed outcome — fail-closed on an exhausted
+     * pool or an exceeded depth ceiling (the caller inspects `ok` before `handle`).
+     */
+    spawn<C extends Out>(agent: Agent<unknown, C>, task: unknown, opts: SpawnOpts): {
+        ok: true;
+        handle: Handle<C>;
+    } | {
+        ok: false;
+        reason: 'budget-exhausted' | 'depth-exceeded';
+    };
+    /** ray.wait n=1 over this scope's in-memory live set; resolves as each child settles;
+     *  `null` when the live set is empty. */
+    next(): Promise<Settled<Out> | null>;
+    /**
+     * Steer a RUNNING child out-of-band — deliver a message to its executor's inbox (the driver's
+     * `send` verb: next-instruction, interrupt, or resume). Returns `true` if the message was
+     * delivered to a live child whose executor accepts delivery, `false` otherwise (unknown id,
+     * already settled, or an executor with no inbox). The executor drains its inbox between turns;
+     * a leaf that does not implement `deliver` simply cannot be steered mid-flight. In-process this
+     * is a direct call; the sandbox/Agent-Bus transports surface the SAME verb as an MCP tool.
+     */
+    send(nodeId: NodeId, msg: unknown): boolean;
+    /** This scope's abort signal — aborted when the run is cancelled, a breaker trips, the pool
+     *  is exhausted, or a parent scope cascades. A long-running driver `act` over this scope reads
+     *  it to break promptly (the conserved pool + driver-stop are the other bounds). A nested
+     *  scope carries its own signal, chained off its driver child's abort. */
+    readonly signal: AbortSignal;
+    /**
+     * Meter the driver's OWN compute against the conserved pool — its inference turns, which are
+     * real tokens/usd but not a spawned child (no reserve/reconcile). A direct `free → committed`
+     * debit, so equal-k counts the driver's tokens AND the in-loop budget guard (`budget.tokensLeft`)
+     * halts a driver that thinks the pool dry. `detail` rides an `agent.turn` trace event for live
+     * observability (turn index, tool calls, cumulative spend). It also journals a `metered` event —
+     * the durable twin of the pool debit (as `settled` is the twin of `reconcile`) — so every
+     * journal-based cost reader (`spentFromJournal`, `trajectoryReport`) sums driver inference
+     * automatically. A leaf never calls this; a driver meters each chat turn and awaits it (the
+     * metered event is cost-critical, so it lands before the join-barrier roll-up).
+     */
+    meter(spend: Spend, detail?: Record<string, unknown>): Promise<void>;
+    /** The live tree — reads the in-memory nursery, not the journal. */
+    readonly view: TreeView;
+    /** Conserved-pool readouts (post-reservation). */
+    readonly budget: Readonly<{
+        tokensLeft: number;
+        usdLeft: number;
+        usdCapped: boolean;
+        deadlineMs: number;
+        reservedTokens: number;
+    }>;
+}
+interface NodeSnapshot {
+    readonly id: NodeId;
+    readonly parent?: NodeId;
+    readonly label: string;
+    readonly status: NodeStatus;
+    readonly runtime: Runtime;
+    readonly budget: Budget;
+    /** Conserved spend so far for this node. */
+    readonly spent: Spend;
+    /** `outRef` once the node is `done` (the replay/result pointer). */
+    readonly outRef?: string;
+}
+/** The live tree — what `scope.view` / `RootHandle.view()` materialize for a viewer. */
+interface TreeView {
+    readonly root: NodeId;
+    readonly nodes: ReadonlyArray<NodeSnapshot>;
+    /** Count of nodes in `running` or `acquiring` — the "what's in flow?" answer. */
+    readonly inFlight: number;
+}
+/** Journaled spawn-tree events (B1/B2). `seq` is the cursor order; `at` is an ISO
+ *  timestamp for human inspection only (NOT a replay input). */
+type SpawnEvent = {
+    kind: 'spawned';
+    id: NodeId;
+    parent?: NodeId;
+    label: string;
+    budget: Budget;
+    runtime: Runtime;
+    seq: number;
+    at: string;
+} | {
+    kind: 'settled';
+    id: NodeId;
+    status: 'done' | 'down';
+    /** Content-addressed result pointer; rehydrates `out` from `ResultBlobStore`. */
+    outRef?: string;
+    verdict?: DefaultVerdict;
+    spent: Spend;
+    infra?: boolean;
+    seq: number;
+    at: string;
+} | {
+    kind: 'cancelled';
+    id: NodeId;
+    reason: string;
+    seq: number;
+    at: string;
+} | {
+    /** A driver's OWN inference spend, journaled separately from spawned-child work — the journal
+     *  TWIN of `BudgetPool.observe`, exactly as `settled` is the twin of `reconcile`. So every
+     *  journal-based cost reader sums it automatically — the journal is the single cost ledger.
+     *  It carries spend only and is NOT a settlement: replay + `materializeTreeView` skip it for
+     *  structure, and its `seq` lives outside the cursor-uniqueness namespace. A
+     *  driver re-homes its nested subtree's metered total up to its parent (like settled spend),
+     *  so summing any sub-tree root yields that sub-tree's true driver-inference cost. */
+    kind: 'metered';
+    id: NodeId;
+    spend: Spend;
+    seq: number;
+    at: string;
+};
+/**
+ * The spawn-tree event source (mirrors `ConversationJournal`'s begin/append/load shape).
+ * `loadTree` replays the full ordered event list for resume/replay; `appendEvent` is
+ * called only AFTER the event is observed-committed (never speculative).
+ */
+interface SpawnJournal {
+    loadTree(root: NodeId): Promise<SpawnEvent[] | undefined>;
+    beginTree(root: NodeId, at: string): Promise<void>;
+    appendEvent(root: NodeId, ev: SpawnEvent): Promise<void>;
+}
+/** Content-addressed result blobs (the `outRef` → artifact map) backing the replay
+ *  invariant. Split from the journal so the journal stays small (decisions) and the
+ *  payloads (evidence) live where a viewer/replayer rehydrates them. */
+interface ResultBlobStore {
+    put(outRef: string, artifact: unknown): Promise<void>;
+    get(outRef: string): Promise<unknown | undefined>;
+}
+/**
+ * Owns the conserved pool, the spawn log, the abort cascade, the OTP intensity breaker,
+ * and the root handle. `run` executes the root `Agent` to completion; `attach` wires a
+ * live `RootHandle` (the Q2 substrate the chat/pi-viz client later consumes).
+ */
+interface Supervisor<Task, Out> {
+    run(root: Agent<Task, Out>, task: Task, opts: SupervisorOpts): Promise<SupervisedResult<Out>>;
+    attach(h: RootHandle<Out>): void;
+}
+interface SupervisorOpts {
+    /** The root conserved-pool ceiling (tokens + usd + iterations + deadline). */
+    readonly budget: Budget;
+    /** Trace-correlation root + the journal/blob root key. */
+    readonly runId: NodeId;
+    /** Event source — defaults to the in-memory journal in the impl; pass JSONL/FS for durability. */
+    readonly journal: SpawnJournal;
+    /** Result payload store backing `outRef` rehydration. */
+    readonly blobs: ResultBlobStore;
+    /** Executor resolution — the open registry mapping `AgentSpec` → `Executor`. */
+    readonly executors: ExecutorRegistry;
+    /** Runtime recursion-depth ceiling (paired with the conserved pool per R3). */
+    readonly maxDepth?: number;
+    /**
+     * OTP intensity breaker: more than `maxRestarts` child restarts within `withinMs`
+     * trips the supervisor to `no-winner` rather than restarting forever.
+     */
+    readonly maxRestarts?: number;
+    readonly withinMs?: number;
+    readonly now?: () => number;
+    readonly signal?: AbortSignal;
+    /** Lifecycle stream sink, threaded into the root `Scope` so every `spawn`/settle emits on the
+     *  same `agent.spawn`/`agent.child` stream `runLoop` feeds — one observable recursive tree. */
+    readonly hooks?: RuntimeHooks;
+}
+/** Typed terminal result (M2) — a no-winner is NEVER coerced to a best-effort output. */
+type SupervisedResult<Out> = {
+    kind: 'winner';
+    out: Out;
+    outRef: string;
+    verdict?: DefaultVerdict;
+    tree: TreeView;
+    spentTotal: Spend;
+    /** Where `spentTotal` went: `driverInference` = the drivers' own chat turns (metered via
+     *  `Scope.meter`); `childWork` = every spawned child's reconciled spend (the journal sum).
+     *  `driverInference + childWork === spentTotal`. Present whenever any driver metered. */
+    spentBreakdown?: {
+        driverInference: Spend;
+        childWork: Spend;
+    };
+} | {
+    kind: 'no-winner';
+    reason: 'all-children-down' | 'budget-exhausted' | 'aborted';
+    tree: TreeView;
+    downCount: number;
+    /** The conserved spend incurred before the run failed — real cost is paid even when no
+     *  worker delivers, so the caller always learns what the delegation actually spent. Summed
+     *  off the same journal the `winner` path reads. */
+    spentTotal: Spend;
+};
+/** Live root handle — the substrate a chat/pi-viz client attaches to (Q2). `signal`
+ *  delivers an out-of-band message to the running root; `view()` materializes the tree. */
+interface RootHandle<Out> {
+    view(): TreeView;
+    signal(msg: RootSignal): void;
+    abort(reason?: string): void;
+    /** Phantom: binds the handle to the supervised run's output type. Type-only — never
+     *  present at runtime; lets `attach(h: RootHandle<Out>)` stay output-typed. */
+    readonly __out?: Out;
+}
+/** Out-of-band message to a running root. Open by intent — a client extends it. */
+type RootSignal = {
+    kind: 'pause';
+} | {
+    kind: 'resume';
+} | {
+    kind: 'cancel';
+    reason?: string;
+} | {
+    kind: 'ask';
+    question: string;
+};
+/**
+ * The progressive-widening gate (MCTS-PW). Decides whether a settled child is
+ * `promising` enough to spawn another under the remaining pool. DEFAULTS TO FLAT
+ * (`shouldWiden` always false) so a gate run never widens and the selector≠judge
+ * firewall conflict (R2) stays dormant. When widening IS enabled, `promising` MUST be
+ * derived from TRACE findings (`analyses`), never raw `verdict` — or the gate carries
+ * an explicit, argued `judgeExempt: true` (the documented escape hatch, off by default).
+ */
+interface WidenGate<Out> {
+    /** Default impl returns false for every settlement (flat — never widens). */
+    shouldWiden(settled: Settled<Out>, budget: Scope<Out>['budget']): boolean;
+    /** When true, widening may read `verdict` directly (collides with the steer firewall —
+     *  must be explicitly argued per cell, never defaulted on). */
+    readonly judgeExempt?: boolean;
+}
+
+/**
+ * @experimental
+ *
+ * The completion-oracle: **settled ⟺ DELIVERED.**
+ *
+ * Foreman's one hard lesson (0/18 self-improvement deliverables) — "done" must mean a check
+ * PASSED, not the agent's say-so. `gateOnDeliverable` wraps an `Executor` so its settlement
+ * is `valid` ONLY when the deliverable check passes. The child still RUNS and settles (its
+ * spend is conserved into the pool either way), but a child that ran WITHOUT delivering
+ * settles `valid:false` — so a keep-best driver never counts it as done, and a gate never
+ * inflates with self-judged wins.
+ *
+ * Dual-purpose by construction:
+ *  - product: the agent fleet only advances on real, checked deliverables.
+ *  - proof: the gate's `valid` is the honest settle — equal-k comparisons can't be gamed by an
+ *    arm that "ran" without producing the artifact.
+ *
+ * The check is a DEPLOYABLE oracle — a test command, a state verifier, the commit0 judge —
+ * read off the child's output, never the model judging itself. A throwing check is
+ * fail-closed (not delivered), never a crash.
+ */
+
+/**
+ * The deployable completion oracle passed to {@link gateOnDeliverable}: a `check` that
+ * decides DELIVERED (settles `valid` ⟺ it resolves true) plus an optional `describe` of
+ * what the spawn was supposed to produce. The check reads the child's output — never the
+ * model judging itself.
+ */
+interface DeliverableSpec<Out = unknown> {
+    /** The deployable check that decides DELIVERED. `settled.valid ⟺ this resolves true`. */
+    check: (out: Out) => boolean | Promise<boolean>;
+    /** What the spawn was supposed to produce — surfaced in traces/reports. */
+    describe?: string;
+}
+/**
+ * Wrap an `Executor` so its settlement `valid` reflects the deliverable check, not the
+ * inner verdict. Handles both `execute` shapes (one-shot `Promise<ExecutorResult>` and
+ * streaming `AsyncIterable<UsageEvent>` + `resultArtifact()`); the check runs once the inner
+ * executor has produced its output. The inner `score` is preserved; only `valid` is gated.
+ */
+declare function gateOnDeliverable<Out>(inner: Executor<Out>, deliverable: DeliverableSpec<Out>): Executor<Out>;
+
+/**
+ * @experimental
+ *
+ * Git worktree helpers for the in-process delegation executor. Each
+ * delegation runs in its own worktree so multiple parallel harness
+ * subprocesses (claude / codex / opencode in a 3-way fanout) don't clobber
+ * each other's edits on the shared workspace.
+ *
+ * Worktrees live under `<repoRoot>/.agent-worktrees/<runId>/`. After the
+ * harness exits + the diff is captured, the worktree is removed.
+ *
+ * All operations spawn `git` via `child_process.spawn` synchronously
+ * (via a `runGit` helper). Stays narrow on purpose: no commits, no rebases.
+ * Diff capture stages all changes (`git add -A`) into the ephemeral worktree's
+ * index so created (untracked) files appear in the `--cached` diff.
+ */
+/** @experimental */
+interface WorktreeHandle {
+    /** Absolute path to the worktree directory. */
+    path: string;
+    /** SHA the worktree was created at. */
+    baseSha: string;
+    /** Branch name created for this worktree (typically `delegate/<runId>`). */
+    branch: string;
+}
+/** @experimental */
+interface CreateWorktreeOptions {
+    /** Absolute path to the main git checkout. */
+    repoRoot: string;
+    /** Unique id for the worktree path + branch. Use the delegation run id. */
+    runId: string;
+    /** Parent directory the worktree lives under. Defaults to `.agent-worktrees`. */
+    variantsDir?: string;
+    /** Override the base ref (default `HEAD`). */
+    baseRef?: string;
+    /** Test seam — inject a custom git runner. */
+    runGit?: GitRunner;
+}
+/** @experimental */
+interface DiffOptions {
+    /** Worktree to diff. */
+    worktree: WorktreeHandle;
+    /** What to compare against. Default `worktree.baseSha`. */
+    baseRef?: string;
+    /** Test seam. */
+    runGit?: GitRunner;
+}
+/** @experimental */
+interface DiffResult {
+    patch: string;
+    stats: {
+        filesChanged: number;
+        insertions: number;
+        deletions: number;
+    };
+}
+/** @experimental */
+interface RemoveWorktreeOptions {
+    worktree: WorktreeHandle;
+    repoRoot: string;
+    /** Force removal even if dirty (default true; the loser of a fanout has uncommitted changes). */
+    force?: boolean;
+    /** Test seam. */
+    runGit?: GitRunner;
+}
+/** Pluggable git runner (sync) — replaceable in tests. */
+type GitRunner = (args: ReadonlyArray<string>, opts: {
+    cwd: string;
+}) => {
+    stdout: string;
+    stderr: string;
+    exitCode: number;
+};
+/** @experimental */
+declare function createWorktree(options: CreateWorktreeOptions): Promise<WorktreeHandle>;
+/** @experimental */
+declare function captureWorktreeDiff(options: DiffOptions): Promise<DiffResult>;
+/** @experimental */
+declare function removeWorktree(options: RemoveWorktreeOptions): Promise<void>;
+
+export { type AgentSpec as A, type Budget as B, type CreateWorktreeOptions as C, type DiffOptions as D, type ExecutorRegistry as E, type GitRunner as G, type LocalHarness as L, type NodeId as N, type RemoveWorktreeOptions as R, type SpawnJournal as S, type TreeView as T, type UsageEvent as U, type WorktreeHandle as W, type DiffResult as a, type LocalHarnessResult as b, type RunLocalHarnessOptions as c, captureWorktreeDiff as d, createWorktree as e, removeWorktree as f, type Agent as g, type ResultBlobStore as h, type RootHandle as i, type SupervisedResult as j, type Settled as k, type Spend as l, type Scope as m, type Executor as n, type DeliverableSpec as o, type ExecutorFactory as p, type SpawnEvent as q, runLocalHarness as r, type Supervisor as s, type ExecutorContext as t, type ExecutorResult as u, type SupervisorOpts as v, type WidenGate as w, gateOnDeliverable as x };