agentfootprint/dist/core/agent/delivery/rules.js

The explainable agent framework — backtrack a wrong answer to the exact context that caused it (evidence, not guesses). Built on footprintjs.

"use strict";
/**
 * delivery/rules — where a delivered injection may go, and whether it is
 * already there.
 *
 * Pattern: Pure functions over the window (no scope, no I/O, no provider).
 * Role:    core/ layer. The counterpart to `window/turns.ts`: that module
 *          decides what may LEAVE the window, this one decides what may
 *          ENTER it. Both answer with a refusal that names itself, and both
 *          exist so the rule lives in exactly one place instead of being
 *          re-derived at each call site.
 * Emits:   N/A.
 *
 * ── The one rule worth arguing about ─────────────────────────────────
 * Collision is judged on the EFFECTIVE wire role, where `'tool'` counts as
 * `'user'`. That is deliberately more conservative than any single provider
 * requires: Anthropic coalesces `tool_result`s into a user turn (so a user
 * message after tool results breaks its strict alternation), while OpenAI
 * keeps `role: 'tool'` and would accept one. Judging per-provider would make
 * the SAME declaration deliver on one wire and defer on another, with
 * nothing in the recording to tell them apart — which is the exact
 * provider-dependent truth this whole feature exists to avoid. So the floor
 * is the rule, uniformly.
 *
 * The consequence is real and stated rather than hidden: inside a tool-using
 * agent loop the window ends on the user's turn (iteration 1) or on tool
 * results (every iteration after), so a `'user'`-role injection typically
 * never gets a slot. `'assistant'` does, and `'system'` does on the providers
 * that carry it.
 *
 * Testable on its own — see `test/core/delivery/delivery-rules.test.ts`.
 */
Object.defineProperty(exports, "__esModule", { value: true });
exports.tailWireRole = exports.refusalForPlacement = exports.keysInWindow = exports.deliveryKey = exports.effectiveWireRole = void 0;
const helpers_js_1 = require("../../slots/helpers.js");
/**
 * The role a message OCCUPIES on the wire, for sequencing purposes.
 *
 * A `tool` message is folded into `user` because that is what the strictest
 * provider does with it. Everything else is itself.
 */
function effectiveWireRole(role) {
    return role === 'tool' ? 'user' : role;
}
exports.effectiveWireRole = effectiveWireRole;
/**
 * The stable identity of one deliverable message: who declared it, and what
 * it says.
 *
 * Position within the injection's own list is deliberately NOT part of the
 * key — it cannot be recovered from a message already sitting in the window,
 * and the key has to mean the same thing whether it is read from the run's
 * ledger or recovered from a marker on a restored history. Content identity
 * is what matters: a memory recall whose text changed between iterations is a
 * NEW delivery (it is new content), an always-on fact repeating verbatim is
 * the same one and is delivered once, and one injection declaring the same
 * sentence twice says it once.
 */
function deliveryKey(injectionId, role, content) {
    return `${injectionId}:${(0, helpers_js_1.fnv1a)(`${role}:${content}`)}`;
}
exports.deliveryKey = deliveryKey;
/**
 * Every delivery key already present as a marker in the window.
 *
 * Unioned with the run's ledger before deciding what to deliver, and that
 * union is what makes replay safe: `agent.resumeOnError` rebuilds the
 * conversation but starts a fresh scope, so the ledger is empty while the
 * delivered messages are right there in the restored history. Reading the
 * window itself is the only way to know that.
 */
function keysInWindow(history) {
    const keys = new Set();
    for (const msg of history) {
        const by = msg.injectedBy;
        if (by === undefined)
            continue;
        keys.add(deliveryKey(by.injectionId, msg.role, msg.content));
    }
    return keys;
}
exports.keysInWindow = keysInWindow;
/**
 * Why a message may NOT be appended to the end of `history` right now, or
 * `undefined` when it may.
 *
 * Two refusals, in the order a reader cares about them:
 *   • `unanswered-tool-call` — the tail is an assistant turn whose tool calls
 *     have no results yet. Appending here would sit BETWEEN a `tool_use` and
 *     its `tool_result`, which is the one split the window family refuses in
 *     the other direction too.
 *   • `role-collision` — the turn at the end already speaks in this effective
 *     role, and providers reject two in a row.
 */
function refusalForPlacement(history, role) {
    if (history.length === 0)
        return undefined;
    const answered = new Set();
    for (const msg of history) {
        if (msg.role === 'tool' && msg.toolCallId)
            answered.add(msg.toolCallId);
    }
    for (const msg of history) {
        for (const call of msg.toolCalls ?? []) {
            if (!answered.has(call.id))
                return 'unanswered-tool-call';
        }
    }
    const tail = history[history.length - 1];
    if (effectiveWireRole(tail.role) === effectiveWireRole(role))
        return 'role-collision';
    return undefined;
}
exports.refusalForPlacement = refusalForPlacement;
/** The effective role of the turn a delivery would land behind, for the note. */
function tailWireRole(history) {
    const tail = history[history.length - 1];
    return tail === undefined ? undefined : effectiveWireRole(tail.role);
}
exports.tailWireRole = tailWireRole;
//# sourceMappingURL=rules.js.map