framework/dist/control.d.ts

The (AI) Framework: turnkey, zero-config AI orchestration that wraps a coding-agent CLI (Claude Code) as a black box and takes you from an idea to a running app. Vite for AI.

import type { ChoiceBy } from './events.js';
import { type HandoffLevel } from './handoff-level.js';
/**
 * The dashboard-to-agent control channel (#344): the reverse of the event log.
 * Events flow run -> `.the-framework/events.jsonl` -> daemon -> browser; steering
 * flows browser -> daemon -> `.the-framework/control.jsonl` -> run. The daemon
 * appends a {@link ControlEntry} per Stop click / choice pick, and the agent tails
 * the file, aborting or resolving its parked gate. Same file-is-the-seam design
 * as the forward direction — no run<->daemon IPC.
 */
/** The control log filename under `.the-framework/`. */
export declare const CONTROL_FILE = "control.jsonl";
/** One steering instruction from the dashboard to the live agent. */
export type ControlEntry = 
/** Stop the agent (the daemon dashboard's Stop button). */
{
    kind: 'stop';
}
/** Resolve a parked choice gate: the pick for the pending {@link ChoiceRequest} id. */
 | {
    kind: 'choice';
    id: string;
    pick: string | string[];
    by: ChoiceBy;
}
/** A live-chat message the user sent to the running agent (#714). */
 | {
    kind: 'message';
    text: string;
}
/**
 * Move the end-of-session handoff (#1102): how far this session publishes itself when it
 * finishes — keep it local, push the branch, open a PR, merge it.
 *
 * One rung rather than a pair of booleans (B5): a surface offering checkboxes converts on its
 * side, so an impossible answer resolves *down* there instead of arriving here as "a PR with no
 * push" for this end to repair upward.
 *
 * Steering rather than an event because it is an instruction to the agent, and it has to reach a
 * run whose dashboard tab was opened after it started. The agent echoes what it applied back as an
 * event, which is what puts it on the meta the checkboxes read.
 */
 | {
    kind: 'handoff';
    level: HandoffLevel;
}
/**
 * The user's Merge action on a live session (#1391): arm the full publish ladder and record that
 * a human authorized the merge, so the merge gate (#1363) does not also demand the agent's
 * ready-for-merge signal — a human's word outranks it. The session still merges at its own end
 * (it ends itself once nothing needs a human, #1390); this is a pre-commitment, not an abort.
 */
 | {
    kind: 'merge';
};
/** The control log path for a workspace. */
export declare function controlPath(cwd: string): string;
/** Append one entry to the workspace's control log, creating it as needed. */
export declare function appendControl(cwd: string, entry: ControlEntry): Promise<void>;
/**
 * Truncate the control log. An agent calls this at start so a previous agent's picks
 * can never fire into this one (gate ids like `plan-approval` repeat across runs).
 */
export declare function resetControl(cwd: string): Promise<void>;
/** A live control tail. {@link close} stops watching (idempotent). */
export interface ControlWatcher {
    close(): void;
}
/**
 * Tail the workspace's control log, dispatching each well-formed entry as it is
 * appended. An `fs.watch` on `.the-framework/` plus a poll backstop, mirroring the
 * daemon's event tail (`fs.watch` is unreliable across platforms). Malformed or
 * unknown lines are skipped so a bad write can never crash an agent.
 */
export declare function watchControl(cwd: string, onEntry: (entry: ControlEntry) => void, pollMs?: number): ControlWatcher;
//# sourceMappingURL=control.d.ts.map