@beeos-ai/device-agent 0.15.0 → 0.16.1

package/dist/agent-runtime/reflection.d.ts

@@ -1,100 +0,0 @@
-/**
- * agent-runtime/reflection — hierarchical reflection layer for AgentLoop.
- *
- * Mirrors the Python reference at
- *   /Users/heat/workspace/deviceagent_research/deviceagent/agent_runtime/loop/reflection.py
- *
- * Phase-1 reflector is heuristic-only (cheap):
- *  - **Action reflection** — 4 identical actions in a row → stall.
- *  - **Trajectory reflection** — ≥5 screenshots in the last 6 steps → stall
- *    (the model is staring at the screen instead of acting).
- *  - **Pre-action sanity** — block obviously-broken `nextAction` candidates
- *    *before* dispatch (zero-cost; mirrors mobile-use `enable_pre_reflection`).
- *  - **Model-based action reflection** is plumbed via `IActionReflector` —
- *    composed by AgentLoop after `act()` returns, gated by confidence /
- *    `actionResult` so the heuristic layer fires first.
- */
-import type { Action } from "@beeos-ai/device-common";
-import type { IActionReflector } from "./reflectors/action-reflector.js";
-import type { ITrajectoryReflector } from "./reflectors/trajectory-reflector.js";
-import type { IModelAdapter, StepHistory } from "./adapters/base.js";
-/**
- * Veto reason returned by {@link HierarchicalReflector.maybeBlockAction}.
- *  - `reason` — short human-readable string, copied into the next StepHistory
- *    rationale so the grounder sees *why* its previous attempt was blocked.
- *  - `code` — stable identifier for telemetry / metrics tagging.
- *
- * `no_progress_repeat` is the hash-aware successor to
- * `double_tap_same_coords` — emitted when {@link StepHistory.screenStateHashBefore}
- * equals {@link StepHistory.screenStateHashAfter} (i.e. the previous
- * action did not change the screen) and the next action is a near-repeat
- * of it. The legacy `double_tap_same_coords` code is preserved as the
- * fallback for pure-vision adapters that don't produce a `uiDump` hash.
- */
-export interface PreActionVeto {
-    code: "double_type_without_focus" | "double_tap_same_coords" | "no_progress_repeat" | "blind_swipe" | "vlm_emitted_shell";
-    reason: string;
-}
-export interface HierarchicalReflectorOptions {
-    /** Optional VLM adapter — reserved for future model-based reflection hooks. */
-    adapter?: IModelAdapter;
-    /**
-     * Optional model-based reflector composed after `act()` returns. AgentLoop
-     * calls `actionReflector?.shouldReflect(...)` first; the heuristic veto
-     * still runs before any model invocation.
-     */
-    actionReflector?: IActionReflector;
-    /**
-     * Optional trajectory reflector (Phase 2.2). AgentLoop fires it every N
-     * steps to assess global progress; bail-out verdicts (`off_track` /
-     * `stuck`) are converted into a reflection veto so the PlanExecutor can
-     * trigger a replan.
-     */
-    trajectoryReflector?: ITrajectoryReflector;
-}
-export declare class HierarchicalReflector {
-    private readonly adapter?;
-    readonly actionReflector?: IActionReflector;
-    readonly trajectoryReflector?: ITrajectoryReflector;
-    constructor(adapterOrOptions?: IModelAdapter | HierarchicalReflectorOptions);
-    /**
-     * Returns a non-empty veto reason when the loop should stop, or `null` if
-     * the run is healthy and may continue.
-     */
-    maybeVeto(history: StepHistory[], step: number, _maxSteps: number): string | null;
-    /**
-     * Pre-action sanity — block obviously-broken `nextAction` candidates
-     * **before** AgentLoop dispatches them to MCP. Returns a veto when the
-     * action looks like a known-bad pattern; the AgentLoop should then skip
-     * the dispatch and record the step with `actionResult: "skipped"`.
-     *
-     * Rules (cover ~80% of common grounder mistakes — see plan Task 1.3):
-     *  1. Two consecutive `type` actions with no intervening `tap` /
-     *     `long_press` → focus probably wasn't established (mobile-use
-     *     `pre_reflection` rule).
-     *  2. Two consecutive `tap` actions that produced no screen change.
-     *     Hash-aware: when {@link StepHistory.screenStateHashBefore} ===
-     *     {@link StepHistory.screenStateHashAfter} on the previous step,
-     *     the action evidently did nothing, so a repeat (within 24 px) is
-     *     futile — emit `no_progress_repeat`. Pure-vision adapters with no
-     *     hash fall back to the legacy 8 px `double_tap_same_coords` rule.
-     *  3. Two consecutive `swipe` actions in the same direction without an
-     *     intervening `screenshot` → blind scrolling without re-grounding.
-     *  4. The grounder emitted `execute_command` (the VLM should never
-     *     produce shell). This is exactly the "Baidu image" failure mode
-     *     that motivated this rule set.
-     *
-     * Hash guard (applies to rules 1 / 2 / 3): when we have a `uiDump`-
-     * derived `screenStateHashBefore` AND a `screenStateHashAfter` for the
-     * previous step AND they differ, the previous action evidently
-     * produced screen progress, so we let the next action through even if
-     * it superficially repeats the prior one. This eliminates the "model
-     * tried again with a tiny perturbation but the screen actually
-     * advanced" false-positive that motivated the KISS hardening pass.
-     *
-     * Heuristics are conservative — when in doubt, return `null` (let the
-     * action through). Repeated stalls are still caught by `maybeVeto`.
-     */
-    maybeBlockAction(history: StepHistory[], nextAction: Action): PreActionVeto | null;
-}
-export declare function actionsEqual(a: Action, b: Action): boolean;

package/dist/agent-runtime/reflection.js

@@ -1,241 +0,0 @@
-/**
- * agent-runtime/reflection — hierarchical reflection layer for AgentLoop.
- *
- * Mirrors the Python reference at
- *   /Users/heat/workspace/deviceagent_research/deviceagent/agent_runtime/loop/reflection.py
- *
- * Phase-1 reflector is heuristic-only (cheap):
- *  - **Action reflection** — 4 identical actions in a row → stall.
- *  - **Trajectory reflection** — ≥5 screenshots in the last 6 steps → stall
- *    (the model is staring at the screen instead of acting).
- *  - **Pre-action sanity** — block obviously-broken `nextAction` candidates
- *    *before* dispatch (zero-cost; mirrors mobile-use `enable_pre_reflection`).
- *  - **Model-based action reflection** is plumbed via `IActionReflector` —
- *    composed by AgentLoop after `act()` returns, gated by confidence /
- *    `actionResult` so the heuristic layer fires first.
- */
-export class HierarchicalReflector {
-    adapter;
-    actionReflector;
-    trajectoryReflector;
-    constructor(adapterOrOptions) {
-        if (adapterOrOptions && "name" in adapterOrOptions) {
-            this.adapter = adapterOrOptions;
-        }
-        else {
-            const opts = adapterOrOptions ?? {};
-            this.adapter = opts.adapter;
-            this.actionReflector = opts.actionReflector;
-            this.trajectoryReflector = opts.trajectoryReflector;
-        }
-        void this.adapter; // reserved for future model-based reflection
-    }
-    /**
-     * Returns a non-empty veto reason when the loop should stop, or `null` if
-     * the run is healthy and may continue.
-     */
-    maybeVeto(history, step, _maxSteps) {
-        if (step <= 2 || history.length === 0)
-            return null;
-        const last = history[history.length - 1].action;
-        if (!last)
-            return null;
-        // Action reflection — 4 identical actions in a row.
-        const recent = history
-            .slice(-4)
-            .map((h) => h.action)
-            .filter((a) => a !== null);
-        if (recent.length === 4 && recent.every((a) => actionsEqual(a, recent[0]))) {
-            return `stalled: action ${recent[0].type} repeated 4 consecutive times`;
-        }
-        // [已弃用] screenshot-heavy stall 检测：连续6步中有5个 screenshot 动作时终止。
-        // 当前所有适配器（autoglm、general 等）截图均由 AgentLoop 被动触发，
-        // 模型不会主动发出 screenshot 动作，此规则永远不会触发，故弃用。
-        return null;
-    }
-    /**
-     * Pre-action sanity — block obviously-broken `nextAction` candidates
-     * **before** AgentLoop dispatches them to MCP. Returns a veto when the
-     * action looks like a known-bad pattern; the AgentLoop should then skip
-     * the dispatch and record the step with `actionResult: "skipped"`.
-     *
-     * Rules (cover ~80% of common grounder mistakes — see plan Task 1.3):
-     *  1. Two consecutive `type` actions with no intervening `tap` /
-     *     `long_press` → focus probably wasn't established (mobile-use
-     *     `pre_reflection` rule).
-     *  2. Two consecutive `tap` actions that produced no screen change.
-     *     Hash-aware: when {@link StepHistory.screenStateHashBefore} ===
-     *     {@link StepHistory.screenStateHashAfter} on the previous step,
-     *     the action evidently did nothing, so a repeat (within 24 px) is
-     *     futile — emit `no_progress_repeat`. Pure-vision adapters with no
-     *     hash fall back to the legacy 8 px `double_tap_same_coords` rule.
-     *  3. Two consecutive `swipe` actions in the same direction without an
-     *     intervening `screenshot` → blind scrolling without re-grounding.
-     *  4. The grounder emitted `execute_command` (the VLM should never
-     *     produce shell). This is exactly the "Baidu image" failure mode
-     *     that motivated this rule set.
-     *
-     * Hash guard (applies to rules 1 / 2 / 3): when we have a `uiDump`-
-     * derived `screenStateHashBefore` AND a `screenStateHashAfter` for the
-     * previous step AND they differ, the previous action evidently
-     * produced screen progress, so we let the next action through even if
-     * it superficially repeats the prior one. This eliminates the "model
-     * tried again with a tiny perturbation but the screen actually
-     * advanced" false-positive that motivated the KISS hardening pass.
-     *
-     * Heuristics are conservative — when in doubt, return `null` (let the
-     * action through). Repeated stalls are still caught by `maybeVeto`.
-     */
-    maybeBlockAction(history, nextAction) {
-        if (!nextAction)
-            return null;
-        // Rule 4 — VLM emitted shell. Always block, regardless of hash data.
-        if (nextAction.type === "execute_command") {
-            return {
-                code: "vlm_emitted_shell",
-                reason: "grounder emitted execute_command — VLMs should never produce shell; planner should route shell via tools",
-            };
-        }
-        if (history.length === 0)
-            return null;
-        const prev = lastNonNullStep(history);
-        if (!prev)
-            return null;
-        // Hash guard — if we KNOW the previous action moved the screen, skip
-        // every "no progress" check below. Absent hash data we fall back to
-        // the legacy coordinate / direction heuristics so pure-vision
-        // adapters (Open-AutoGLM, UI-TARS) keep working.
-        const haveHash = prev.entry.screenStateHashBefore !== undefined &&
-            prev.entry.screenStateHashAfter !== undefined;
-        const screenChanged = haveHash && prev.entry.screenStateHashBefore !== prev.entry.screenStateHashAfter;
-        const screenUnchanged = haveHash && prev.entry.screenStateHashBefore === prev.entry.screenStateHashAfter;
-        // Rule 1 — double-type without re-focus. If the previous `type`
-        // visibly altered the screen (e.g. text actually appeared), don't
-        // block; the grounder may legitimately want to keep typing.
-        // Rule 1 — 仅当两次 type 内容完全相同时才触发（重复卡死）。
-        // 原逻辑依赖 !screenChanged，无 hash 时永远为 true，会误杀合法的追加输入。
-        if (nextAction.type === "type" &&
-            prev.action.type === "type" &&
-            nextAction.payload.text ===
-                prev.action.payload.text) {
-            return {
-                code: "double_type_without_focus",
-                reason: "two consecutive `type` actions with identical text — likely a stuck loop, focus probably wasn't established",
-            };
-        }
-        // Rule 2 — repeated tap.
-        //   * with hash data: only block when the prior tap produced ZERO
-        //     screen change (`no_progress_repeat`). Tolerance widens to 24 px
-        //     because the model often nudges the tap by a few pixels on
-        //     retry without changing semantics.
-        //   * without hash data: legacy 8 px `double_tap_same_coords` rule.
-        if (nextAction.type === "tap" &&
-            prev.action.type === "tap" &&
-            isPointTarget(nextAction.target) &&
-            isPointTarget(prev.action.target)) {
-            const dx = Math.abs(nextAction.target.x - prev.action.target.x);
-            const dy = Math.abs(nextAction.target.y - prev.action.target.y);
-            if (haveHash) {
-                if (screenUnchanged && dx <= 24 && dy <= 24) {
-                    return {
-                        code: "no_progress_repeat",
-                        reason: `repeated tap near (${prev.action.target.x},${prev.action.target.y}) — previous tap produced no screen change`,
-                    };
-                }
-                // hash known + screen progressed (or coords too far apart) → allow
-            }
-            else {
-                // legacy fallback for pure-vision adapters — 此阶段暂不开启（无 hash 数据时对 autoglm 等纯视觉适配器过于激进）
-                // if (dx <= 8 && dy <= 8) {
-                //   return {
-                //     code: "double_tap_same_coords",
-                //     reason: `two consecutive tap on (${prev.action.target.x},${prev.action.target.y}) → looks like a misclassified double_tap`,
-                //   };
-                // }
-            }
-        }
-        // [已弃用] Rule 3 — blind_swipe：连续同向 swipe 且中间无 screenshot 时触发。
-        // 依赖 !screenChanged，无 hash 时永远为 true，会误杀合法的连续滚动操作。
-        // AgentLoop 被动截图不计入 sawScreenshotBetween，导致条件永远满足，故弃用。
-        // if (
-        //   nextAction.type === "swipe" &&
-        //   prev.action.type === "swipe" &&
-        //   isBoxTarget(nextAction.target) &&
-        //   isBoxTarget(prev.action.target) &&
-        //   !screenChanged
-        // ) {
-        //   const sameDir =
-        //     sign(nextAction.target.x2 - nextAction.target.x1) ===
-        //       sign(prev.action.target.x2 - prev.action.target.x1) &&
-        //     sign(nextAction.target.y2 - nextAction.target.y1) ===
-        //       sign(prev.action.target.y2 - prev.action.target.y1);
-        //   const sawScreenshotBetween = anyActionBetween(
-        //     history,
-        //     prev.idx,
-        //     history.length,
-        //     (a) => a.type === "screenshot",
-        //   );
-        //   if (sameDir && !sawScreenshotBetween) {
-        //     return {
-        //       code: "blind_swipe",
-        //       reason:
-        //         "swipe in the same direction as the previous step with no intervening screenshot or scroll progress — re-ground before scrolling further",
-        //     };
-        //   }
-        // }
-        return null;
-    }
-}
-/* -------------------------------------------------------------------------- */
-/* helpers                                                                    */
-/* -------------------------------------------------------------------------- */
-export function actionsEqual(a, b) {
-    if (a.type !== b.type)
-        return false;
-    if (!targetsEqual(a.target, b.target))
-        return false;
-    return jsonEquals(a.payload, b.payload);
-}
-function targetsEqual(a, b) {
-    if (!a && !b)
-        return true;
-    if (!a || !b)
-        return false;
-    if (a.kind !== b.kind)
-        return false;
-    return jsonEquals(a, b);
-}
-function jsonEquals(a, b) {
-    if (Object.is(a, b))
-        return true;
-    if (typeof a !== "object" || typeof b !== "object" || a === null || b === null)
-        return false;
-    return JSON.stringify(a) === JSON.stringify(b);
-}
-function isPointTarget(t) {
-    return !!t && t.kind === "point";
-}
-function isBoxTarget(t) {
-    return !!t && t.kind === "box";
-}
-function sign(n) {
-    return n > 0 ? 1 : n < 0 ? -1 : 0;
-}
-function lastNonNullStep(history) {
-    for (let i = history.length - 1; i >= 0; i--) {
-        const entry = history[i];
-        const a = entry.action;
-        if (a)
-            return { idx: i, entry, action: a };
-    }
-    return null;
-}
-function anyActionBetween(history, startExclusive, endExclusive, pred) {
-    for (let i = startExclusive + 1; i < endExclusive; i++) {
-        const a = history[i].action;
-        if (a && pred(a))
-            return true;
-    }
-    return false;
-}
-//# sourceMappingURL=reflection.js.map

package/dist/agent-runtime/reflection.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"reflection.js","sourceRoot":"","sources":["../../src/agent-runtime/reflection.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAiDH,MAAM,OAAO,qBAAqB;IACf,OAAO,CAAiB;IAChC,eAAe,CAAoB;IACnC,mBAAmB,CAAwB;IAEpD,YAAY,gBAA+D;QACzE,IAAI,gBAAgB,IAAI,MAAM,IAAK,gBAAkC,EAAE,CAAC;YACtE,IAAI,CAAC,OAAO,GAAG,gBAAiC,CAAC;QACnD,CAAC;aAAM,CAAC;YACN,MAAM,IAAI,GAAI,gBAA6D,IAAI,EAAE,CAAC;YAClF,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC;YAC5B,IAAI,CAAC,eAAe,GAAG,IAAI,CAAC,eAAe,CAAC;YAC5C,IAAI,CAAC,mBAAmB,GAAG,IAAI,CAAC,mBAAmB,CAAC;QACtD,CAAC;QACD,KAAK,IAAI,CAAC,OAAO,CAAC,CAAC,6CAA6C;IAClE,CAAC;IAED;;;OAGG;IACH,SAAS,CAAC,OAAsB,EAAE,IAAY,EAAE,SAAiB;QAC/D,IAAI,IAAI,IAAI,CAAC,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;YAAE,OAAO,IAAI,CAAC;QAEnD,MAAM,IAAI,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC;QAChD,IAAI,CAAC,IAAI;YAAE,OAAO,IAAI,CAAC;QAEvB,oDAAoD;QACpD,MAAM,MAAM,GAAG,OAAO;aACnB,KAAK,CAAC,CAAC,CAAC,CAAC;aACT,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,MAAM,CAAC;aACpB,MAAM,CAAC,CAAC,CAAC,EAAe,EAAE,CAAC,CAAC,KAAK,IAAI,CAAC,CAAC;QAC1C,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC,IAAI,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,YAAY,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;YAC3E,OAAO,mBAAmB,MAAM,CAAC,CAAC,CAAC,CAAC,IAAI,+BAA+B,CAAC;QAC1E,CAAC;QAED,6DAA6D;QAC7D,iDAAiD;QACjD,wCAAwC;QAExC,OAAO,IAAI,CAAC;IACd,CAAC;IAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;OAgCG;IACH,gBAAgB,CAAC,OAAsB,EAAE,UAAkB;QACzD,IAAI,CAAC,UAAU;YAAE,OAAO,IAAI,CAAC;QAE7B,qEAAqE;QACrE,IAAI,UAAU,CAAC,IAAI,KAAK,iBAAiB,EAAE,CAAC;YAC1C,OAAO;gBACL,IAAI,EAAE,mBAAmB;gBACzB,MAAM,EACJ,0GAA0G;aAC7G,CAAC;QACJ,CAAC;QAED,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;YAAE,OAAO,IAAI,CAAC;QACtC,MAAM,IAAI,GAAG,eAAe,CAAC,OAAO,CAAC,CAAC;QACtC,IAAI,CAAC,IAAI;YAAE,OAAO,IAAI,CAAC;QAEvB,qEAAqE;QACrE,oEAAoE;QACpE,8DAA8D;QAC9D,iDAAiD;QACjD,MAAM,QAAQ,GACZ,IAAI,CAAC,KAAK,CAAC,qBAAqB,KAAK,SAAS;YAC9C,IAAI,CAAC,KAAK,CAAC,oBAAoB,KAAK,SAAS,CAAC;QAChD,MAAM,aAAa,GACjB,QAAQ,IAAI,IAAI,CAAC,KAAK,CAAC,qBAAqB,KAAK,IAAI,CAAC,KAAK,CAAC,oBAAoB,CAAC;QACnF,MAAM,eAAe,GACnB,QAAQ,IAAI,IAAI,CAAC,KAAK,CAAC,qBAAqB,KAAK,IAAI,CAAC,KAAK,CAAC,oBAAoB,CAAC;QAEnF,gEAAgE;QAChE,kEAAkE;QAClE,4DAA4D;QAC5D,uCAAuC;QACvC,oDAAoD;QACpD,IACE,UAAU,CAAC,IAAI,KAAK,MAAM;YAC1B,IAAI,CAAC,MAAM,CAAC,IAAI,KAAK,MAAM;YAC1B,UAAU,CAAC,OAA8B,CAAC,IAAI;gBAC5C,IAAI,CAAC,MAAM,CAAC,OAA8B,CAAC,IAAI,EAClD,CAAC;YACD,OAAO;gBACL,IAAI,EAAE,2BAA2B;gBACjC,MAAM,EACJ,6GAA6G;aAChH,CAAC;QACJ,CAAC;QAED,yBAAyB;QACzB,kEAAkE;QAClE,sEAAsE;QACtE,gEAAgE;QAChE,wCAAwC;QACxC,oEAAoE;QACpE,IACE,UAAU,CAAC,IAAI,KAAK,KAAK;YACzB,IAAI,CAAC,MAAM,CAAC,IAAI,KAAK,KAAK;YAC1B,aAAa,CAAC,UAAU,CAAC,MAAM,CAAC;YAChC,aAAa,CAAC,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,EACjC,CAAC;YACD,MAAM,EAAE,GAAG,IAAI,CAAC,GAAG,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;YAChE,MAAM,EAAE,GAAG,IAAI,CAAC,GAAG,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;YAChE,IAAI,QAAQ,EAAE,CAAC;gBACb,IAAI,eAAe,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,EAAE,CAAC;oBAC5C,OAAO;wBACL,IAAI,EAAE,oBAAoB;wBAC1B,MAAM,EAAE,sBAAsB,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,IAAI,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,4CAA4C;qBACvH,CAAC;gBACJ,CAAC;gBACD,mEAAmE;YACrE,CAAC;iBAAM,CAAC;gBACN,sFAAsF;gBACtF,4BAA4B;gBAC5B,aAAa;gBACb,sCAAsC;gBACtC,kIAAkI;gBAClI,OAAO;gBACP,IAAI;YACN,CAAC;QACH,CAAC;QAED,6DAA6D;QAC7D,mDAAmD;QACnD,uDAAuD;QACvD,OAAO;QACP,mCAAmC;QACnC,oCAAoC;QACpC,sCAAsC;QACtC,uCAAuC;QACvC,mBAAmB;QACnB,MAAM;QACN,oBAAoB;QACpB,4DAA4D;QAC5D,+DAA+D;QAC/D,4DAA4D;QAC5D,6DAA6D;QAC7D,mDAAmD;QACnD,eAAe;QACf,gBAAgB;QAChB,sBAAsB;QACtB,sCAAsC;QACtC,OAAO;QACP,4CAA4C;QAC5C,eAAe;QACf,6BAA6B;QAC7B,gBAAgB;QAChB,qJAAqJ;QACrJ,SAAS;QACT,MAAM;QACN,IAAI;QAEJ,OAAO,IAAI,CAAC;IACd,CAAC;CACF;AAED,gFAAgF;AAChF,gFAAgF;AAChF,gFAAgF;AAEhF,MAAM,UAAU,YAAY,CAAC,CAAS,EAAE,CAAS;IAC/C,IAAI,CAAC,CAAC,IAAI,KAAK,CAAC,CAAC,IAAI;QAAE,OAAO,KAAK,CAAC;IACpC,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC,MAAM,CAAC;QAAE,OAAO,KAAK,CAAC;IACpD,OAAO,UAAU,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC,OAAO,CAAC,CAAC;AAC1C,CAAC;AAED,SAAS,YAAY,CAAC,CAAmB,EAAE,CAAmB;IAC5D,IAAI,CAAC,CAAC,IAAI,CAAC,CAAC;QAAE,OAAO,IAAI,CAAC;IAC1B,IAAI,CAAC,CAAC,IAAI,CAAC,CAAC;QAAE,OAAO,KAAK,CAAC;IAC3B,IAAI,CAAC,CAAC,IAAI,KAAK,CAAC,CAAC,IAAI;QAAE,OAAO,KAAK,CAAC;IACpC,OAAO,UAAU,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;AAC1B,CAAC;AAED,SAAS,UAAU,CAAC,CAAU,EAAE,CAAU;IACxC,IAAI,MAAM,CAAC,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC;QAAE,OAAO,IAAI,CAAC;IACjC,IAAI,OAAO,CAAC,KAAK,QAAQ,IAAI,OAAO,CAAC,KAAK,QAAQ,IAAI,CAAC,KAAK,IAAI,IAAI,CAAC,KAAK,IAAI;QAAE,OAAO,KAAK,CAAC;IAC7F,OAAO,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,KAAK,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC;AACjD,CAAC;AAED,SAAS,aAAa,CAAC,CAAmB;IACxC,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,IAAI,KAAK,OAAO,CAAC;AACnC,CAAC;AAED,SAAS,WAAW,CAClB,CAAmB;IAEnB,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,IAAI,KAAK,KAAK,CAAC;AACjC,CAAC;AAED,SAAS,IAAI,CAAC,CAAS;IACrB,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;AACpC,CAAC;AAED,SAAS,eAAe,CACtB,OAAsB;IAEtB,KAAK,IAAI,CAAC,GAAG,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC7C,MAAM,KAAK,GAAG,OAAO,CAAC,CAAC,CAAC,CAAC;QACzB,MAAM,CAAC,GAAG,KAAK,CAAC,MAAM,CAAC;QACvB,IAAI,CAAC;YAAE,OAAO,EAAE,GAAG,EAAE,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,CAAC,EAAE,CAAC;IAC7C,CAAC;IACD,OAAO,IAAI,CAAC;AACd,CAAC;AAED,SAAS,gBAAgB,CACvB,OAAsB,EACtB,cAAsB,EACtB,YAAoB,EACpB,IAA4B;IAE5B,KAAK,IAAI,CAAC,GAAG,cAAc,GAAG,CAAC,EAAE,CAAC,GAAG,YAAY,EAAE,CAAC,EAAE,EAAE,CAAC;QACvD,MAAM,CAAC,GAAG,OAAO,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC;QAC5B,IAAI,CAAC,IAAI,IAAI,CAAC,CAAC,CAAC;YAAE,OAAO,IAAI,CAAC;IAChC,CAAC;IACD,OAAO,KAAK,CAAC;AACf,CAAC"}

package/dist/agent-runtime/reflectors/action-reflector.d.ts

@@ -1,97 +0,0 @@
-/**
- * agent-runtime/reflectors/action-reflector — model-based action reflection.
- *
- * Phase 1.4 of the device-agent reflector upgrade plan
- * ([../../../../../../.cursor/plans/device-agent-reflector-skills_d5599350.plan.md]).
- *
- * Goal: catch *the same step* the grounder just emitted when it's likely
- * wrong (low confidence, or the MCP `act` failed) and feed back a textual
- * verdict the next predict call can read. Concretely:
- *   - heuristic veto (`HierarchicalReflector.maybeVeto`) runs first and
- *     terminates the loop with a hard stop on stalls;
- *   - if the heuristics are quiet, we ask the action reflector whether to
- *     take a closer look. The default {@link LLMActionReflector} fires on
- *     `actionResult === "act_failed"` OR on low-confidence predictions
- *     OR (as a defensive fallback) on a small whitelist of high-risk verbs
- *     so we still help even when the adapter doesn't expose logprobs.
- *   - the reflector verdict is **advisory**: a `passed: false` result is
- *     spliced into the next StepHistory rationale ("act_failed: …") so the
- *     grounder retries with that context. We do NOT terminate the loop —
- *     stalls are still the heuristic layer's job.
- *
- * The cost model is documented in
- * `agents/device-agent/docs/reflectors.md` (Phase 2.4). Per-task budget
- * with default config is ≤ 4 cheap text glm-4-flash calls / task, ~600-1200
- * input tokens / 50 output tokens each.
- */
-import OpenAI from "openai";
-import type { Action } from "@beeos-ai/device-common";
-import type { ModelUsage, Prediction, StepHistory } from "../adapters/base.js";
-/** Verbs that warrant a reflection on Branch C of the logprobs spike. */
-export declare const HIGH_RISK_VERBS: ReadonlySet<Action["type"]>;
-export type ActionReflectorGating = "logprob" | "logprob-degraded" | "whitelist";
-export interface ActionReflectCtx {
-    /** Goal of the surrounding task. */
-    goal: string;
-    /** Up to N previous step entries (most recent last). */
-    history: StepHistory[];
-    /** The prediction we want to reflect on. */
-    prediction: Prediction;
-    /** Last observation summary (no base64 — text only). */
-    observationSummary: string;
-}
-export interface ActionReflectVerdict {
-    /**
-     * `true`  — reflector agrees with the action; nothing to do.
-     * `false` — reflector flagged the action; the AgentLoop should record
-     *           `act_failed: <reason>` and let the grounder retry.
-     */
-    passed: boolean;
-    reason: string;
-    usage: ModelUsage;
-}
-export interface IActionReflector {
-    /**
-     * Should the loop invoke {@link reflect} on this prediction? Cheap to
-     * call — only returns true on the small subset that warrants the LLM
-     * round-trip.
-     */
-    shouldReflect(history: StepHistory[], lastPrediction: Prediction, actionResult: StepHistory["actionResult"]): boolean;
-    /** Run the actual reflection. */
-    reflect(ctx: ActionReflectCtx, signal?: AbortSignal): Promise<ActionReflectVerdict>;
-    /** Optional cleanup. */
-    close?(): Promise<void>;
-}
-export interface LLMActionReflectorOptions {
-    /** Override `process.env.PLANNER_BASE_URL` (planner endpoint reused). */
-    baseUrl?: string;
-    apiKey?: string;
-    /** Defaults to `glm-4-flash`. */
-    modelName?: string;
-    /** Confidence threshold; predictions below this fire the reflector. Defaults 0.6. */
-    confidenceThreshold?: number;
-    /**
-     * Spike branch (see {@link ActionReflectorGating}). Defaults to "whitelist"
-     * which is the safe fallback when the AutoGLM-Phone-9B endpoint doesn't
-     * expose logprobs.
-     */
-    gating?: ActionReflectorGating;
-    /** Sampling temperature. Defaults to 0. */
-    temperature?: number;
-    /** Max output tokens. Defaults to 80. */
-    maxTokens?: number;
-    /** Reused OpenAI client for tests. */
-    client?: OpenAI;
-}
-export declare class LLMActionReflector implements IActionReflector {
-    readonly name = "llm-action-reflector";
-    private readonly client;
-    private readonly modelName;
-    private readonly confidenceThreshold;
-    private readonly gating;
-    private readonly maxTokens;
-    private readonly temperature;
-    constructor(opts?: LLMActionReflectorOptions);
-    shouldReflect(_history: StepHistory[], lastPrediction: Prediction, actionResult: StepHistory["actionResult"]): boolean;
-    reflect(ctx: ActionReflectCtx, signal?: AbortSignal): Promise<ActionReflectVerdict>;
-}