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>
170 lines
5.3 KiB
JavaScript
170 lines
5.3 KiB
JavaScript
// OpenAI-compatible chat client for the "fast advisor" tier.
|
|
//
|
|
// The original "coach" loop uses Pi via the CLI subprocess (5-15s latency,
|
|
// rate-limited to a few calls per hour). That's appropriate for deep
|
|
// post-mortem analytics but useless when the bot needs tactical advice
|
|
// right now ("I'm wedged in a pit, what should I do?").
|
|
//
|
|
// This provider opens a parallel path: any OpenAI-compatible HTTP endpoint
|
|
// (TimeWeb is the default — same env var convention as the user's other
|
|
// projects — but OpenAI direct, Groq, OpenRouter, and local Ollama with
|
|
// the OpenAI shim all work with the same plumbing) producing a structured
|
|
// JSON answer in ≤8 seconds.
|
|
//
|
|
// Configuration is strictly env-driven. The provider is a NO-OP unless
|
|
// TIMEWEB_API_KEY is set, so it's safe to ship the code disabled.
|
|
|
|
import { info, warn } from "../log.js";
|
|
|
|
const ENV = {
|
|
BASE_URL: "TIMEWEB_BASE_URL",
|
|
API_KEY: "TIMEWEB_API_KEY",
|
|
MODEL: "TIMEWEB_MODEL",
|
|
TIMEOUT_MS: "TIMEWEB_TIMEOUT_MS",
|
|
};
|
|
|
|
const DEFAULT_BASE_URL = "https://api.openai.com/v1";
|
|
// 20s default — TimeWeb's hosted agent endpoint takes 5-15s for the
|
|
// fast-advisor prompt (registry block + snapshot context). 8s was too
|
|
// tight and produced spurious timeouts in smoke tests. OpenAI direct
|
|
// returns much faster (<2s); the env var overrides if needed.
|
|
const DEFAULT_TIMEOUT_MS = 20000;
|
|
|
|
export function isAvailable() {
|
|
return !!process.env[ENV.API_KEY];
|
|
}
|
|
|
|
export function getConfig() {
|
|
return {
|
|
baseUrl: (process.env[ENV.BASE_URL] || DEFAULT_BASE_URL).replace(/\/+$/, ""),
|
|
apiKey: process.env[ENV.API_KEY] || null,
|
|
model: process.env[ENV.MODEL] || null,
|
|
timeoutMs: Number(process.env[ENV.TIMEOUT_MS]) || DEFAULT_TIMEOUT_MS,
|
|
};
|
|
}
|
|
|
|
/**
|
|
* complete({ system, user, json, model?, timeoutMs? })
|
|
* → { ok: true, text, raw, latencyMs } | { ok: false, code, detail, latencyMs }
|
|
*
|
|
* `json: true` requests JSON-mode (response_format) and returns the
|
|
* parsed object as `text`. If the provider doesn't honour JSON-mode the
|
|
* call still works but caller is responsible for parsing.
|
|
*/
|
|
export async function complete({
|
|
system,
|
|
user,
|
|
json = false,
|
|
model,
|
|
timeoutMs,
|
|
} = {}) {
|
|
const cfg = getConfig();
|
|
if (!cfg.apiKey) {
|
|
return { ok: false, code: "not_configured", detail: `set ${ENV.API_KEY}`, latencyMs: 0 };
|
|
}
|
|
const useModel = model || cfg.model;
|
|
if (!useModel) {
|
|
return { ok: false, code: "no_model", detail: `set ${ENV.MODEL} env or pass model arg`, latencyMs: 0 };
|
|
}
|
|
|
|
const body = {
|
|
model: useModel,
|
|
messages: [
|
|
system ? { role: "system", content: system } : null,
|
|
{ role: "user", content: user ?? "" },
|
|
].filter(Boolean),
|
|
temperature: 0.3,
|
|
};
|
|
if (json) {
|
|
body.response_format = { type: "json_object" };
|
|
}
|
|
|
|
const url = `${cfg.baseUrl}/chat/completions`;
|
|
const startedAt = Date.now();
|
|
const controller = new AbortController();
|
|
const t = setTimeout(() => controller.abort(), timeoutMs ?? cfg.timeoutMs);
|
|
|
|
let resp;
|
|
try {
|
|
resp = await fetch(url, {
|
|
method: "POST",
|
|
headers: {
|
|
"Content-Type": "application/json",
|
|
Authorization: `Bearer ${cfg.apiKey}`,
|
|
},
|
|
body: JSON.stringify(body),
|
|
signal: controller.signal,
|
|
});
|
|
} catch (e) {
|
|
clearTimeout(t);
|
|
const latency = Date.now() - startedAt;
|
|
const aborted = e?.name === "AbortError";
|
|
return {
|
|
ok: false,
|
|
code: aborted ? "timeout" : "network_error",
|
|
detail: e?.message ?? String(e),
|
|
latencyMs: latency,
|
|
};
|
|
}
|
|
clearTimeout(t);
|
|
|
|
const latencyMs = Date.now() - startedAt;
|
|
if (!resp.ok) {
|
|
let body;
|
|
try { body = await resp.text(); } catch { body = "<no body>"; }
|
|
warn("llm", `${useModel} ${resp.status}: ${body.slice(0, 200)}`);
|
|
return {
|
|
ok: false,
|
|
code: `http_${resp.status}`,
|
|
detail: body.slice(0, 500),
|
|
latencyMs,
|
|
};
|
|
}
|
|
|
|
let payload;
|
|
try {
|
|
payload = await resp.json();
|
|
} catch (e) {
|
|
return { ok: false, code: "bad_json", detail: e?.message ?? "parse error", latencyMs };
|
|
}
|
|
|
|
const text = payload?.choices?.[0]?.message?.content;
|
|
if (typeof text !== "string") {
|
|
return { ok: false, code: "no_content", detail: "no choices[0].message.content", latencyMs };
|
|
}
|
|
|
|
let parsed = text;
|
|
if (json) {
|
|
parsed = tryParseJson(text);
|
|
if (parsed === null) {
|
|
return { ok: false, code: "bad_json", detail: text.slice(0, 200), latencyMs };
|
|
}
|
|
}
|
|
|
|
// usage shape per OpenAI / TimeWeb / most compat endpoints:
|
|
// { prompt_tokens, completion_tokens, total_tokens }
|
|
const usage = normaliseUsage(payload?.usage);
|
|
info("llm", `${useModel} ok (${latencyMs}ms, ${text.length}ch, in=${usage.in}/out=${usage.out}t)`);
|
|
return { ok: true, text: parsed, raw: text, latencyMs, usage };
|
|
}
|
|
|
|
function normaliseUsage(u) {
|
|
if (!u || typeof u !== "object") return { in: 0, out: 0, total: 0 };
|
|
const inT = Number(u.prompt_tokens ?? u.input_tokens ?? 0) || 0;
|
|
const outT = Number(u.completion_tokens ?? u.output_tokens ?? 0) || 0;
|
|
const total = Number(u.total_tokens ?? inT + outT) || (inT + outT);
|
|
return { in: inT, out: outT, total };
|
|
}
|
|
|
|
function tryParseJson(text) {
|
|
if (!text) return null;
|
|
const trimmed = text.trim().replace(/^```(?:json)?/i, "").replace(/```$/, "").trim();
|
|
try { return JSON.parse(trimmed); } catch {}
|
|
const m = trimmed.match(/\{[\s\S]*\}/);
|
|
if (!m) return null;
|
|
try { return JSON.parse(m[0]); } catch { return null; }
|
|
}
|
|
|
|
// Test exports
|
|
export const __testing = { ENV, DEFAULT_BASE_URL, DEFAULT_TIMEOUT_MS, tryParseJson, normaliseUsage };
|