Files
pepa-pi-bot/runtime/coach/advisor-trigger.js
T
mayatnikovandClaude Opus 4.7 bc381b2a4b feat(v0.3.0): auto-trigger fast-advisor + token usage tracking
Closes the awareness → LLM → action loop that the rc.1/2/3 sequence
left as a followup. When the bot is wedged, looping, or just suffered
a preempt-then-retry, the reflex fires advise() in the background;
when the recommendation lands it overrides the next dispatch.

Async by design: advise() takes 5-15s on TimeWeb's hosted endpoint —
too slow for a synchronous reflex tick. tickAdvisor() is fire-and-
forget, the result lands on ctx.advisorRecommendation, and the *next*
tick reads and consumes it. Recommendations age out after 60s.

Components:

- runtime/coach/advisor-trigger.js — policy + async fire path
  - tickAdvisor(ctx, {plannedSkillId}) checks three triggers:
    1. wedged > 60s (no significant move)
    2. last 4+ dispatches are the same skill AND it's planned again
    3. preempt within last 30s + same skill being retried
  - 90s trigger cooldown, single-in-flight guard
  - consumeFreshRecommendation(ctx) reads/clears the cache
- runtime/reflex.js — curriculumReflex calls tickAdvisor() every tick
  and consumes a fresh recommendation BEFORE dispatching. ctx flag
  disableAdvisor=true for tests.
- runtime/bot.js — dispatchAction maintains a rolling 8-slot
  reflexCtx.recentSkillIds for the loop-detection trigger.

Token usage:

- runtime/llm/provider.js — normaliseUsage() reads OpenAI/TimeWeb-
  style {prompt_tokens, completion_tokens, total_tokens} from the
  response. Returned on every complete() result and logged at info
  level as "in=Nt/out=Mt".
- runtime/coach/fast-advisor.js — getUsageSnapshot() aggregates
  total tokens across all calls in the session.

Measured on live TimeWeb endpoint (gpt-5.4-mini agent):
  per call: ~705 input + 45 output = ~750 tokens
  rate limit: 6 calls/hour
  worst case at full budget: ~108K tokens/day
  estimated cost (OpenAI gpt-5-mini reference price): ~$0.60/month

Well within any reasonable budget — model can run hot 24/7.

Smoke verified: scripts/check-timeweb.js probe 4 produces
  trigger fired: true (wedged_90s)
  recommendation: recovery.tunnel-out
  rationale: "Stuck wedged for 90s; exploration is failing."
  latency: 5302ms

Tests: 345 green (was 332, +13 advisor-trigger).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-27 18:44:30 +03:00

165 lines
5.8 KiB
JavaScript

// Auto-trigger policy for the fast tactical advisor.
//
// The advisor is too slow for synchronous use inside a reflex tick
// (5-15s via TimeWeb's hosted agent endpoint). The strategy here is
// asynchronous: when conditions warrant tactical advice, fire-and-forget
// an advise() call; when the result eventually arrives, cache it on
// ctx.advisorRecommendation. The next reflex tick reads that cache and
// can substitute the recommended skill before dispatching.
//
// Triggers (any one, AND-ed with the not-recently-asked cooldown):
//
// 1. Wedged > 60s — bot's position hasn't shifted ≥16 blocks in over
// a minute (already tracked by reflex.js as lastSignificantMoveAt)
// 2. Last 4+ dispatches were the same skill — clear loop signal
// 3. Last awareness preempt was very recent AND followed by same
// skill being dispatched again — env-shock-blind retry
//
// The recommendation has a TTL (60s). After that it's stale and the
// reflex falls back to manifesto / curriculum. This keeps the system
// reactive — advice ages out, fresh data drives fresh advice.
import { advise, isAvailable as advisorAvailable } from "./fast-advisor.js";
import { isRegistered } from "../skill-registry.js";
import { info, warn } from "../log.js";
const TRIGGER_COOLDOWN_MS = 90_000;
const RECOMMENDATION_TTL_MS = 60_000;
const WEDGED_THRESHOLD_MS = 60_000;
const REPEAT_THRESHOLD = 4;
const PREEMPT_WINDOW_MS = 30_000;
let _lastTriggerAt = 0;
let _inFlight = false;
export function _resetForTest() {
_lastTriggerAt = 0;
_inFlight = false;
}
export function getTriggerState() {
return {
lastTriggerAt: _lastTriggerAt,
inFlight: _inFlight,
};
}
/**
* tickAdvisor(ctx) → maybe-fires advise() in background.
*
* Called from reflex AFTER it has chosen a plannedSkillId but BEFORE
* dispatching. Does NOT block — the in-flight call resolves later and
* writes ctx.advisorRecommendation. The caller decides whether to
* consume a fresh recommendation on this tick or wait for the next.
*/
export function tickAdvisor(ctx, { plannedSkillId } = {}) {
if (!advisorAvailable()) return { fired: false, reason: "disabled" };
if (_inFlight) return { fired: false, reason: "in_flight" };
const now = Date.now();
if (now - _lastTriggerAt < TRIGGER_COOLDOWN_MS) {
return { fired: false, reason: "cooldown" };
}
// Drop a recommendation that's already aged out.
if (ctx.advisorRecommendation && now - ctx.advisorRecommendation.at > RECOMMENDATION_TTL_MS) {
ctx.advisorRecommendation = null;
}
const reason = detectTrigger(ctx, now, plannedSkillId);
if (!reason) return { fired: false, reason: "no_trigger" };
_lastTriggerAt = now;
_inFlight = true;
const snapshot = ctx.snapshot ?? null;
const recentSkillIds = (ctx.recentSkillIds ?? []).slice(-8);
info("advisor-trigger", `firing because ${reason} (planned=${plannedSkillId ?? "?"})`);
// Fire-and-forget. The promise's resolution writes ctx.advisorRecommendation.
advise({ snapshot, reason, recentSkillIds, lessonsTail: ctx.recentLessons ?? [], force: true })
.then((result) => {
_inFlight = false;
if (result.ok && result.action === "switch_skill" && isRegistered(result.skillId)) {
ctx.advisorRecommendation = {
at: Date.now(),
skillId: result.skillId,
action: "switch_skill",
rationale: result.rationale,
triggerReason: reason,
latencyMs: result.latencyMs,
usage: result.usage ?? null,
};
info("advisor-trigger", `recommendation cached: ${result.skillId} (${result.latencyMs}ms, in=${result.usage?.in ?? "?"}t/out=${result.usage?.out ?? "?"}t)`);
} else if (result.ok && (result.action === "wait" || result.action === "continue")) {
ctx.advisorRecommendation = {
at: Date.now(),
action: result.action,
rationale: result.rationale,
triggerReason: reason,
latencyMs: result.latencyMs,
usage: result.usage ?? null,
};
info("advisor-trigger", `recommendation: ${result.action} (${result.latencyMs}ms)`);
} else if (!result.ok) {
warn("advisor-trigger", `advise failed: ${result.code} (${result.detail})`);
}
})
.catch((e) => {
_inFlight = false;
warn("advisor-trigger", `advise threw: ${e?.message ?? e}`);
});
return { fired: true, reason };
}
function detectTrigger(ctx, now, plannedSkillId) {
// 1. Wedged > threshold
if (ctx.lastSignificantMoveAt && (now - ctx.lastSignificantMoveAt) > WEDGED_THRESHOLD_MS) {
return `wedged_${Math.round((now - ctx.lastSignificantMoveAt) / 1000)}s`;
}
// 2. Repeat-skill loop
const recent = ctx.recentSkillIds ?? [];
if (recent.length >= REPEAT_THRESHOLD) {
const tail = recent.slice(-REPEAT_THRESHOLD);
const allSame = tail.every((id) => id === tail[0]);
if (allSame && plannedSkillId === tail[0]) {
return `repeat_${REPEAT_THRESHOLD}_${tail[0]}`;
}
}
// 3. Recent preempt followed by same skill again
if (ctx.lastPreempt && now - ctx.lastPreempt.at < PREEMPT_WINDOW_MS) {
const lastDispatched = recent[recent.length - 1];
if (lastDispatched && lastDispatched === plannedSkillId) {
return `preempt_retry_${ctx.lastPreempt.reason}`;
}
}
return null;
}
/**
* consumeFreshRecommendation(ctx) → { skillId, action, rationale } | null
*
* Returns a recommendation if one is currently cached and fresh, and
* clears it so subsequent ticks don't re-apply the same advice.
*/
export function consumeFreshRecommendation(ctx) {
const rec = ctx.advisorRecommendation;
if (!rec) return null;
if (Date.now() - rec.at > RECOMMENDATION_TTL_MS) {
ctx.advisorRecommendation = null;
return null;
}
if (rec.action !== "switch_skill" || !rec.skillId) {
// 'continue' / 'wait' don't replace dispatch; surface for telemetry only
return null;
}
ctx.advisorRecommendation = null;
return rec;
}
// Test exports
export const __testing = {
TRIGGER_COOLDOWN_MS, RECOMMENDATION_TTL_MS, WEDGED_THRESHOLD_MS,
REPEAT_THRESHOLD, PREEMPT_WINDOW_MS, detectTrigger,
};