@sema-agent/core/dist/orchestration/goal.d.ts

Stateless, task-oriented AI agent core

/**
 * design/96 §C (S2) — **Goal 模式**:LLM 自报完成 + core 校验续跑(CC goal 对标:目标 + 完成判定 + 续跑)。
 * `runGoal` 是 `runRepairLoop`/`verifyCompleted` 的 SIBLING —— thin composition over `runner.runTask`,零
 * Runner core 改动。详 `design/96-references/GOAL-IMPLEMENTER-SPEC.md`(v3+r2 异源证伪式对照收敛)。
 *
 * 🔴 **完成判定双闸(reward-hack 红线)**:① LLM 自报完成 = 调内置 `declare_done` 工具(或 `submit_output`,有
 * outputSchema 时)= **机器信号**,非泛化 `completed`;② core `doneCheck`(机械/oracle 校验)。**AND**:仅二者
 * 都满足才停 —— LLM 自报但 doneCheck 否决 → 续跑(把 reason 围栏喂回),这是防"模型说我完成了就算数"的核心。
 *
 * 🔴 **G1 脱钩(不可绕)**:`status:"achieved"` = 双闸过 → **停迭代 + SURFACE**,`accepted` 恒 false。runGoal
 * **无任何 commit/accept 副作用**(纯 loop,同 runRepairLoop SAFE-tier 永不 `fixed`)。要 accept,caller 自走
 * design/77 Gate-1 out-of-process oracle 隔离边界,不在 goal helper 开后门。
 */
import type { RunnerSelfSeat } from "../core/runner/contracts.js";
import type { CheckpointToken } from "../core/checkpoint-store.js";
import type { TaskResult, TaskSpec } from "../core/types.js";
/** Read-only snapshot a `doneCheck` sees after one iteration. */
export interface GoalTurnState {
    /** This iteration's full {@link TaskResult}. */
    readonly result: TaskResult;
    /** 1-based iteration index. */
    readonly iteration: number;
    /** The session threaded through the whole goal (so a doneCheck can read the session/working tree). */
    readonly sessionId: string;
    /** Cumulative tokens (own + nested) across iterations — observe-only (the budget gate is in the engine). */
    readonly cumulativeTokens: number;
}
export interface GoalVerdict {
    /** core verification passed? (AND-ed with the LLM self-report — the double gate). */
    readonly done: boolean;
    /** When not done: the specific feedback fed back to the LLM (fenced as untrusted data). */
    readonly reason: string;
}
/** The caller's HONEST declaration of how `doneCheck` verifies — surfaced on the result so a
 *  `self_report_only` "achieved" is never mistaken for a mechanically-verified one (G1, reward-hack honesty). */
export type GoalVerificationKind = "mechanical" | "self_report_only";
export interface GoalSpec {
    /** The goal (injected as iteration 1's objective + GOAL_COMPLETION_GUIDANCE via the trusted goalMode flag). */
    objective: string;
    /**
     * 🔴 The CORE completion gate (REQUIRED). The LLM self-report (gate 1) is necessary but NOT sufficient: the
     * goal stops as `achieved` only when gate 1 AND `doneCheck.done`. Provide a PURE read-check (it may read the
     * session/working tree but must have no side effects — it can be cancelled/timed out). A throw/reject
     * fail-closes to `failed`/`goal.donecheck_error` (never treated as achieved). Core ships no default —
     * supplying it forces the caller to confront the reward-hack red line.
     */
    doneCheck: (state: GoalTurnState, signal: AbortSignal) => Promise<GoalVerdict>;
    /** Honest declaration of `doneCheck`'s substance (§3.2). `self_report_only` ⇒ surfaced on the result. */
    verificationKind: GoalVerificationKind;
    /** Hard iteration cap (≥1) — exceeding without the double gate → `max_iterations` (NOT achieved). */
    maxIterations: number;
    /** Per-iteration base spec (model/tools/toolPolicy/principal/env/outputSchema/…). `objective` is overridden
     *  per iteration; `sessionId` is minted to thread the whole goal if absent; `signal` is set by the engine. */
    taskSpec: Omit<TaskSpec, "objective">;
    /** Cumulative token ceiling (own + nested); reaching it stops with `budget`. */
    budgetTokens?: number;
    /** Whole-goal wall-clock ceiling (ms); reaching it stops with `budget`. */
    totalTimeoutMs?: number;
    /** Cancels the whole goal (folded into each iteration's signal + the doneCheck). */
    signal?: AbortSignal;
    /**
     * design/73 §1 — aggregation key of the `TaskOutcome` emitted at the goal's terminal state (only when
     * `verificationKind:"mechanical"` — the red line). Default: `goal:<sha256(objective) first 16 hex>`,
     * deterministic per objective. Set it when the "same task" spans differently-worded objectives.
     */
    taskSignature?: string;
}
export type GoalStatus = "achieved" | "max_iterations" | "budget" | "blocked" | "suspended" | "needs_review" | "failed" | "aborted";
/**
 * #111 — which ceiling ended a `status:"budget"` run. The status set stays CLOSED (a consumer
 * switching on it keeps compiling); this is the terminating FACTOR beside it, because the two
 * ceilings ask the caller for opposite corrections: `"tokens"` = the allowance is spent (narrow the
 * objective / raise `budgetTokens`), `"walltime"` = `totalTimeoutMs` elapsed (give it more time).
 */
export type GoalBudgetCause = "tokens" | "walltime";
export interface GoalResult {
    readonly status: GoalStatus;
    /** Last iteration's result; absent on a preflight-terminate (a stop BEFORE iteration 1 ran). */
    readonly result?: TaskResult;
    readonly iterations: number;
    /** The last `doneCheck` verdict (present once a doneCheck ran). */
    readonly lastVerdict?: GoalVerdict;
    /** The caller's declared verification kind (always present — replaces the ambiguous "verifiedBy"). */
    readonly verificationKind: GoalVerificationKind;
    /** 🔴 G1 explicit: runGoal NEVER accepts the artifact. Always false. `achieved` = "stopped + awaiting
     *  review"; the caller decides acceptance (via a Gate-1 boundary), never runGoal. */
    readonly accepted: false;
    readonly cumulativeTokens: number;
    /** Carried up on suspended/needs_review so the caller can resume the underlying task. */
    readonly checkpointToken?: CheckpointToken;
    /** Set on `failed` (incl. `goal.donecheck_error`). */
    readonly errorCode?: string;
    /** #111 — present IFF `status === "budget"` (the KEY is absent otherwise, so presence alone reads
     *  as "stopped on a ceiling"). See {@link GoalBudgetCause}. */
    readonly budgetCause?: GoalBudgetCause;
}
/** Reserved name of the goal-completion signal tool (injected by `runGoal` when there is no outputSchema). */
export declare const DECLARE_DONE_TOOL_NAME = "DeclareDone";
/**
 * Run a goal to a double-gated completion (or a bound). Iterates `runner.runTask` on ONE session: each
 * iteration the LLM works and signals completion (declare_done / submit_output); `doneCheck` then verifies.
 * Only `declare_done` AND `doneCheck.done` → `achieved` (surfaced, NEVER auto-accepted, G1). A rejected
 * doneCheck feeds its reason back (fenced) and continues. Bounded by maxIterations / budget / wall-clock /
 * cancel; a durable pause (suspended/needs_review) is a hard boundary that surfaces the checkpointToken.
 */
export declare function runGoal(runner: RunnerSelfSeat, spec: GoalSpec): Promise<GoalResult>;