Files
pepa-pi-bot/runtime/coach/postmortem.js
T
7d96e44804 v0.3.0: Maslow + Awareness — self-learning bot with needs ladder, event-driven reflex, and TimeWeb fast advisor (#27)
* v0.3.0-rc.1: live skill registry + fast advisor scaffold

Roots out the v0.2.x failure mode: Pi-extracted lessons routinely named
hallucinated skill ids (relocate.surface, choose.safe.surface,
survive.shelter, gather.visible_log, …). All 47 Pi-lessons in the live DB
had applied_count=0 because normalisePreferSkill couldn't find them.

Fix:
1. runtime/skill-registry.js — single source of truth derived from
   skills/index.js. Exports listSkillIds, isRegistered, and a
   prompt-ready block (skillRegistryPrompt) grouped by namespace.
2. Pi prompts (coach/postmortem, coach/reflect) embed the live registry
   with a "USE ONLY THESE, never invent" instruction. Lessons are
   filtered at write-time too — anything not in the registry and not a
   known mode name gets dropped.
3. coach/advice.js — normalisePreferSkill now returns null for unknown
   ids, hardening consult() against any hallucinations that slip
   through. Warn-logged for visibility.

Also lays the LLM substrate for the rest of v0.3.0:

- runtime/llm/provider.js — OpenAI-compatible chat client. Configured
  via PEPA_FAST_LLM_{BASE_URL,API_KEY,MODEL,TIMEOUT_MS}. Safe no-op
  unless API_KEY is set. Supports JSON-mode.
- runtime/coach/fast-advisor.js — tactical advisor tier (scaffold).
  Exposes advise() that asks the fast LLM what to do RIGHT NOW when
  the reflex is wedged/stuck. Rejects hallucinated skill ids using the
  registry. Rate-limited 6/h, 30s cooldown. Not auto-triggered yet —
  wired into reflex in rc.3 (awareness layer).

Tests: 279 green (+24 vs rc.3): 5 registry, 9 provider, 10 advisor.

See dev/v0.3.0/PLAN.md for the full iteration design (manifesto needs
ladder, event-driven awareness, skill pre-emption) and STATUS.md for
shipped/pending tracking.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* v0.3.0-rc.2: manifesto / needs ladder L0-L10

Adds an explicit hierarchical needs catalogue that the reflex consults
on every tick. The bot now pursues tangible intermediate goals (food,
wood tools, shelter, stone tools, ...) instead of inheriting whatever
the curriculum thought was "next".

Ladder:
  L0  alive          HP>5, food>0, not in lava, not panic-near hostile
  L1  food           ≥6 food items in inventory (or sated + any food)
  L2  tools_wood     wooden_pickaxe + wooden_axe + wooden_sword
  L3  shelter_basic  bed placed nearby or in inventory
  L4  tools_stone    stone-tier triplet
  L5  armor_basic    any chestplate (pursue=null until craft.leather-*
                     lands; ladder gracefully skips)
  L6  food_security  ≥16 food items
  L7  tools_iron     iron-tier triplet (pursue=gather.stone for now)
  L8  armor_iron     iron chestplate (pursue=null for now)
  L9  village_seed   bed + chest in nearby blocks
  L10 village_full   never detected, falls through to curriculum

Each need has detect(snapshot) → bool and pursue(snapshot) →
{skillId, args} | null. The ladder picks the LOWEST unsatisfied
pursuable need. Needs whose pursue is null get recorded as
blockedNeeds and the walk continues — no stalling on missing skills.

Wired into curriculumReflex: manifesto takes precedence over
curriculum.plan when it has a concrete suggestion. Tests can pass
ctx.disableManifesto=true to exercise the curriculum branch
in isolation (existing reflex tests keep passing this way).

Pi self-reflection prompt now includes
"activeNeed (Maslow ladder L0-L10): L2 tools_wood → gather.logs"
so Pi advises at the right level instead of giving generic guidance.

skillId returned by pursue() is validated against the live registry
(rc.1 plumbing) — manifesto cannot accidentally dispatch a
hallucinated skill name.

Tests: 315 green (was 279 on rc.1, +36 new):
- runtime/manifesto/needs.test.js — 24 tests (per-need detect/pursue,
  helper sums)
- runtime/manifesto/state.test.js — 10 tests (ladder walk, hostile
  takeover at L0, armor skipping, caching)
- runtime/reflex.test.js — 2 integration tests (manifesto overrides
  curriculum plan; well-fed bot pursues tools_stone)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* v0.3.0-rc.3: event-driven awareness + skill pre-emption

Adds a reactive layer on top of the polling reflex. The bot now
notices environmental shocks (forced moves, HP plunges, hostile
spawns) within ~100ms instead of waiting for the next DISPATCH tick,
and the in-flight skill is preempted so the next reflex cycle can
re-plan against the current world state.

This is the rc that wires the "rc.1 plumbing + rc.2 manifesto" into
a feedback loop:
  - awareness fires preempt → dispatch aborts
  - reflex tick re-evaluates → manifesto walks the ladder
  - new dispatch picks the right skill for the new world state

Pieces:

- runtime/awareness/events.js (new) — bot.on listeners:
  - move: single-tick Δposition ≥ 5 blocks → forced_move flag + preempt
  - health: HP drop ≥ 2 → health_plunge flag + preempt
  - entitySpawn: hostile mob within 12 blocks → hostile_added + preempt
  - blockUpdate: nearby block change → env_changed flag (no preempt,
    throttled 800ms; otherwise gather skills would self-preempt
    every dig)

- runtime/skills/index.js — RUNNER_CODES.PREEMPTED + raceWithAbort()
  wraps every execute() against ctx.abortSignal. Existing skills get
  preemption for free; they don't have to check the signal manually.

- runtime/bot.js:
  - dispatchAction creates a fresh AbortController per dispatch and
    stores it on reflexCtx.currentAbort
  - attachAwareness fires controller.abort() when something disrupts
    the active skill; runSkill returns code: "preempted" and the
    reflex moves on
  - reflexCtx.lastPreempt records the most recent shock

Tests: 332 green (was 315 on rc.2, +17 new):
- runtime/awareness/events.test.js — 12 tests (each event type +
  thresholds + throttling + passive-mob filter)
- runtime/skills/contract.test.js — 3 abortSignal tests
  (mid-flight, pre-armed, clean signal)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* chore(.env): add PEPA_FAST_LLM_* placeholders for v0.3.0 fast advisor

Empty values keep the fast-advisor tier disabled (safe no-op). Fill
in BASE_URL + API_KEY + MODEL to enable. TimeWeb-style endpoint
example included.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* chore(v0.3.0): rename fast-LLM env vars to TIMEWEB_* (match other projects)

Aligns with the user's other repos (proso) which use TIMEWEB_API_GROK /
TIMEWEB_URL_GROK. Single naming convention across projects avoids the
'which env var was it for this repo' mental tax.

  PEPA_FAST_LLM_BASE_URL → TIMEWEB_BASE_URL
  PEPA_FAST_LLM_API_KEY  → TIMEWEB_API_KEY
  PEPA_FAST_LLM_MODEL    → TIMEWEB_MODEL
  PEPA_FAST_LLM_TIMEOUT_MS → TIMEWEB_TIMEOUT_MS

Provider still works with any OpenAI-compatible endpoint — TimeWeb is
the default but the variable name doesn't lock us in. Tests + docs +
.env / .env.example updated.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* feat(scripts): TimeWeb smoke test + bump default LLM timeout to 20s

scripts/check-timeweb.js — three probes: plain text, JSON mode, full
fast-advisor stack (registry injection + skill validation). Loads .env,
prints {ok, latency, reply preview} for each. Doesn't touch bot state.

Bumped DEFAULT_TIMEOUT_MS 8s → 20s in runtime/llm/provider.js. TimeWeb's
hosted agent endpoint takes 5-15s for the fast-advisor prompt
(registry block + snapshot context), so 8s was producing spurious
timeouts. OpenAI direct returns much faster; env var TIMEWEB_TIMEOUT_MS
overrides if needed.

Smoke verified live (PR #27 branch):
  probe 1: 6.3s, plain prompt → "pepa hears you"
  probe 2: 5.4s, JSON mode → {"alive":true,"name":"pepa"}
  probe 3: 14.9s, advise() → action=switch_skill, skill=recovery.tunnel-out
           (correct registered skill, sensible rationale — registry
           injection successfully prevents hallucination)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* 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>

* feat(v0.3.0): paradigm shift — TimeWeb-only LLM + persistent advisor trail + improvement queue

This is the rc.4 batch the user requested:

  1. Emergency triggers (low HP + close hostile, lava-under-foot)
     bypass the long cooldown so the LLM is consulted BEFORE the bot
     dies, not after.
  2. Active manifesto need is now included in the advisor user prompt
     — the LLM picks suggestions that satisfy the bot's current
     concrete need (L2 tools_wood → "gather logs nearby" not
     "explore further").
  3. Every advisor recommendation is persisted to SQLite
     (advisor_recommendations table) with full token usage. The
     reflex marks 'applied=1' when it dispatches and updates
     outcome_ok/code when the dispatch completes. Ground truth for
     "is the LLM actually helping" lives in the DB, not in logs.
  4. Pi CLI is OUT of every background loop. coach/postmortem and
     coach/reflect now go through the same TimeWeb endpoint
     fast-advisor uses, via the shared coach/llm-call.js helper.
     Pi is reserved for manual operator commands.
  5. The LLM (postmortem, reflect, advisor) can flag "structural
     gaps" — missing skills/features the operator should implement.
     These land in the new improvement_requests table. Dedup by
     title bumps `votes` instead of inserting duplicates so the
     queue doesn't bloat. Operator views via
     `node scripts/list-improvements.js`.
  6. A deterministic trigger-tuner runs hourly: reads 24h of
     recommendation stats, flags triggers whose success rate is
     below 25% (sample ≥ 5) or whose prompts are expensive (>1000
     input tokens) with mediocre payoff. Improvements get
     source="tuner", category="tuning". No LLM call.

New files:
  runtime/coach/llm-call.js        — askAnalytical() helper
  runtime/coach/trigger-tuner.js   — stats → improvements
  runtime/coach/trigger-tuner.test.js
  scripts/list-improvements.js     — operator CLI

Schema additions:
  advisor_recommendations: id, ts, trigger_reason, planned_skill,
    recommended_skill, action, rationale, active_need, tokens_in,
    tokens_out, latency_ms, applied, outcome_ok, outcome_code, outcome_at
  improvement_requests: id, ts, source, category, title, description,
    context, priority, status, duplicate_of, votes, implemented_at, notes

Renamed env-var consumers:
  Pi-coach drainOnce({ askPi })   → drainOnce({ askAnalyticalFn? })
  Pi-reflect runOnce({ askPi })   → runOnce({ askAnalyticalFn? })
  bot.js attachCoach/attachReflect no longer pass askPi
  attachTuner() added to bot.js spawn handler
  lessons.source 'pi-coach'   → 'timeweb-coach'
  lessons.source 'pi-reflect' → 'timeweb-reflect'

Token cost measured live:
  ~705 input + 45 output = ~750 total per advisor call
  worst case @ 6 calls/hour rate cap = ~108K tokens/day
  OpenAI gpt-5-mini reference price: ~$0.60/month

Operator usage:
  node scripts/list-improvements.js                # open queue
  node scripts/list-improvements.js --stats        # advisor performance
  node scripts/list-improvements.js --done 17 "shipped in 0.3.1"
  node scripts/list-improvements.js --reject 18 "duplicate"

Tests: 360 green (was 332, +28 new).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

---------

Co-authored-by: Yuriy Mayatnikov <mayatnikov@me.com>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-27 19:37:56 +03:00

429 lines
14 KiB
JavaScript

// Death post-mortem coach.
//
// On every `bot.death` event we capture the surrounding context (last
// skill, last hostile, recent log/scenario tail, inventory before/after)
// and write a row into the `deaths` table of the knowledge DB.
//
// A separate slow loop drains `unanalysedDeaths()` and asks Pi to extract
// generalised lessons. Pi calls are rate-limited and deduped — many
// near-identical deaths produce ONE lesson, not 50.
//
// Lessons land in the `lessons` table and feed runtime/knowledge/recall()
// for future skill dispatch decisions.
//
// This file is import-safe: side-effect-free, attaches only when
// `attach(bot, ctx)` is called explicitly from bot.js.
import { existsSync, readFileSync } from "node:fs";
import { resolve } from "node:path";
import {
isAvailable as knowledgeAvailable,
insertDeath,
insertPostmortem,
markDeathAnalysed,
unanalysedDeaths,
record as recordLesson,
poiNearby,
recordPOI,
createImprovementRequest,
} from "../knowledge/index.js";
import { isRegistered, skillRegistryPrompt } from "../skill-registry.js";
import { isAvailable as llmAvailable } from "../llm/provider.js";
import { askAnalytical } from "./llm-call.js";
import { info, warn } from "../log.js";
const COACH_INTERVAL_MS = 5 * 60 * 1000; // 5 min between coach passes
const COACH_BATCH_MAX = 8; // up to 8 deaths per LLM call
const COACH_BUDGET_PER_HOUR = 3; // ≤ 3 analytical LLM calls/hour
const COACH_COOLDOWN_MS = 12 * 60 * 1000; // 12 min between calls
const RECENT_CHAT_TAIL = 6;
const SCENARIO_TAIL = 12;
let _attached = null;
let _llmCallTimes = [];
let _coachTimer = null;
let _lastInventory = null;
export function attach(bot, ctx = {}) {
if (_attached) {
warn("coach", "attach() called twice; ignoring second attach");
return;
}
if (!bot) return;
_attached = { bot, ctx };
// Snapshot inventory each tick (cheap) so death captures what was lost.
bot.on?.("playerCollect", () => { _lastInventory = snapshotInv(bot); });
bot.on?.("spawn", () => { _lastInventory = snapshotInv(bot); });
bot.on?.("death", () => {
try {
const death = captureDeath(bot, ctx);
if (!death) return;
const deathId = insertDeath(death);
info("coach", `death recorded id=${deathId ?? "-"} cause=${death.cause} hostile=${death.hostile ?? "?"} skill=${death.lastSkill ?? "?"}`);
// v0.2.0-rc.3 — mark this spot as a danger POI so spatial recall
// surfaces it next time the bot comes near. Expires after 6 hours
// so the danger doesn't outlive its relevance.
if (typeof death.x === "number" && typeof death.z === "number") {
recordPOI({
kind: "danger",
name: death.hostile ?? death.cause ?? "death",
x: death.x, y: death.y ?? 64, z: death.z,
expiresAt: Date.now() + 6 * 3600_000,
notes: `death id=${deathId} cause=${death.cause}`,
});
}
} catch (e) {
warn("coach", `captureDeath failed: ${e?.message ?? e}`);
}
});
// v0.3.0 — postmortem analysis runs through TimeWeb (the fast LLM
// provider). Pi CLI no longer drives this loop. The drain timer
// fires regardless of whether TimeWeb is configured; drainOnce()
// short-circuits when the LLM is unavailable.
if (!_coachTimer) {
_coachTimer = setInterval(() => {
drainOnce({ stateDir: ctx.stateDir }).catch((e) =>
warn("coach", `drain error: ${e?.message ?? e}`),
);
}, COACH_INTERVAL_MS);
_coachTimer.unref?.();
info("coach", `attached; drain every ${COACH_INTERVAL_MS / 1000}s${llmAvailable() ? " (TimeWeb)" : " (LLM disabled — deaths captured only)"}`);
}
}
export function detach() {
if (_coachTimer) {
clearInterval(_coachTimer);
_coachTimer = null;
}
_attached = null;
}
function snapshotInv(bot) {
try {
const items = bot.inventory?.items?.() ?? [];
const dict = {};
for (const i of items) dict[i.name] = (dict[i.name] || 0) + i.count;
return dict;
} catch {
return null;
}
}
function diffInv(before, after) {
if (!before) return null;
const lost = [];
for (const [name, count] of Object.entries(before)) {
const now = after?.[name] ?? 0;
if (now < count) lost.push({ name, count: count - now });
}
return lost.length ? lost : null;
}
function captureDeath(bot, ctx) {
const pos = bot.entity?.position;
const lastInv = _lastInventory;
const nowInv = snapshotInv(bot);
const inventoryLost = diffInv(lastInv, nowInv);
const currentTask = readCurrentTask(ctx.stateDir);
const lastSkill = currentTask?.label ?? null;
const lastSkillCode = currentTask?.lastCode ?? null;
const hostile = closestHostileName(bot);
const cause = inferCause({ bot, hostile, lastSkill, lastSkillCode });
const recent = readRecentScenarios(ctx.stateDir, SCENARIO_TAIL);
const journalNearby = readJournalNearby(ctx.stateDir, pos, 32);
const chatTail = ctx.chatHistory?.recent?.(RECENT_CHAT_TAIL) ?? null;
const contextBlob = {
recentScenarios: recent,
journalNearby,
chatTail,
snapshot: {
pos,
hp: bot.health,
food: bot.food,
time: bot.time?.timeOfDay ?? null,
isRaining: !!bot.isRaining,
},
};
return {
ts: Date.now(),
x: pos?.x ?? null,
y: pos?.y ?? null,
z: pos?.z ?? null,
cause,
hostile,
lastSkill,
lastSkillCode,
hp: 0,
food: bot.food ?? null,
inventoryLost,
contextBlob,
};
}
function closestHostileName(bot) {
try {
const me = bot.entity?.position;
if (!me) return null;
let best = null;
let bestDist = Infinity;
for (const e of Object.values(bot.entities ?? {})) {
if (!e || e === bot.entity) continue;
if (e.type !== "hostile" && e.kind !== "Hostile mobs") continue;
const d = e.position?.distanceTo?.(me) ?? Infinity;
if (d < bestDist) {
best = e.name ?? e.mobType ?? null;
bestDist = d;
}
}
return best;
} catch {
return null;
}
}
function inferCause({ bot, hostile, lastSkill, lastSkillCode }) {
const y = bot.entity?.position?.y;
if (hostile) return "hostile";
if (typeof bot.food === "number" && bot.food <= 0) return "starvation";
if (typeof y === "number" && y < 30) return "fall";
if (lastSkillCode === "drowning") return "drowning";
if (lastSkillCode === "lava") return "lava";
return "unknown";
}
function readCurrentTask(stateDir) {
if (!stateDir) return null;
const f = resolve(stateDir, "current-task.json");
if (!existsSync(f)) return null;
try { return JSON.parse(readFileSync(f, "utf8")); } catch { return null; }
}
function readRecentScenarios(stateDir, n) {
if (!stateDir) return [];
const f = resolve(stateDir, "scenarios.jsonl");
if (!existsSync(f)) return [];
try {
const raw = readFileSync(f, "utf8");
const lines = raw.split("\n").filter(Boolean);
const tail = lines.slice(-n);
return tail.map((l) => {
try { return JSON.parse(l); } catch { return null; }
}).filter(Boolean);
} catch {
return [];
}
}
function readJournalNearby(stateDir, pos, radius) {
if (!stateDir || !pos) return [];
const f = resolve(stateDir, "world-journal.jsonl");
if (!existsSync(f)) return [];
try {
const raw = readFileSync(f, "utf8");
const lines = raw.split("\n").filter(Boolean).slice(-200);
const out = [];
for (const l of lines) {
let row;
try { row = JSON.parse(l); } catch { continue; }
const a = row.at;
if (!a) continue;
const dx = a.x - pos.x;
const dz = a.z - pos.z;
if (dx * dx + dz * dz <= radius * radius) out.push(row);
}
return out.slice(-20);
} catch {
return [];
}
}
/**
* One pass: take up to COACH_BATCH_MAX unanalysed deaths, summarise them
* for Pi, parse the JSON reply, write lessons + postmortems.
*
* Rate-limited: at most COACH_PI_BUDGET_PER_HOUR calls/hour, with
* COACH_COOLDOWN_MS gap between calls.
*/
export async function drainOnce({ stateDir, force = false, askAnalyticalFn = askAnalytical } = {}) {
if (!knowledgeAvailable()) return { ok: false, reason: "knowledge unavailable" };
if (!llmAvailable()) return { ok: false, reason: "llm not configured" };
const now = Date.now();
const hourAgo = now - 60 * 60 * 1000;
_llmCallTimes = _llmCallTimes.filter((t) => t > hourAgo);
if (!force && _llmCallTimes.length >= COACH_BUDGET_PER_HOUR) {
return { ok: false, reason: "hourly budget exhausted", calls: _llmCallTimes.length };
}
if (!force && _llmCallTimes.length > 0 && now - _llmCallTimes[_llmCallTimes.length - 1] < COACH_COOLDOWN_MS) {
return { ok: false, reason: "cooldown" };
}
const pending = unanalysedDeaths({ limit: COACH_BATCH_MAX });
if (pending.length === 0) return { ok: true, analysed: 0 };
const { system, user } = buildPrompt(pending);
_llmCallTimes.push(now);
const parsed = await askAnalyticalFn({ system, user, json: true });
if (!parsed) return { ok: false, reason: "no reply" };
const reply = typeof parsed === "string" ? parsed : JSON.stringify(parsed);
let lessonsCount = 0;
let rejectedPreferCount = 0;
for (const item of asArray(parsed.lessons ?? parsed)) {
if (!item || !item.lesson) continue;
// Skill ids referenced by Pi must be in the live registry.
// Mode names (e.g. "night_shelter") are tolerated at write time and
// translated at consult time by advice.js#normalisePreferSkill.
let preferSkill = item.prefer_skill ?? null;
if (preferSkill && !isRegistered(preferSkill) && !isLikelyModeName(preferSkill)) {
rejectedPreferCount += 1;
preferSkill = null;
}
let avoidSkill = item.avoid_skill ?? null;
if (avoidSkill && !isRegistered(avoidSkill) && !isLikelyModeName(avoidSkill)) {
avoidSkill = null;
}
recordLesson({
text: item.lesson,
category: item.category ?? "survival",
triggerSkill: item.trigger_skill ?? null,
triggerHostile: item.trigger_hostile ?? null,
triggerSituation: item.trigger_situation ?? null,
avoidSkill,
preferSkill,
confidence: clamp(Number(item.confidence) || 0.6, 0.1, 0.95),
source: "pi-coach",
sourceRef: item.source_ref ?? null,
});
lessonsCount += 1;
}
if (rejectedPreferCount > 0) {
warn("coach", `dropped prefer_skill from ${rejectedPreferCount} lessons (not in registry)`);
}
// Write one postmortem per death; if grouped, share the same lesson.
const groupLesson = parsed.lessons?.[0]?.lesson ?? parsed.lesson ?? null;
for (const d of pending) {
insertPostmortem({
deathId: d.id,
cause: parsed.cause ?? d.cause,
lesson: groupLesson,
nextAction: parsed.next_action ?? null,
rawResponse: reply.slice(0, 4000),
source: "timeweb",
});
markDeathAnalysed(d.id);
}
// v0.3.0 — record any improvement requests the LLM flagged. The
// LLM is encouraged to do this when the deaths point to a missing
// skill or feature; the operator reads scripts/list-improvements.js
// and decides what to implement.
let improvementsCount = 0;
for (const imp of asArray(parsed.improvements ?? [])) {
if (!imp?.title) continue;
createImprovementRequest({
source: "postmortem",
category: imp.category ?? "skill",
title: String(imp.title).slice(0, 120),
description: imp.description ?? null,
context: { death_ids: pending.map((d) => d.id), cause: parsed.cause },
priority: imp.priority ?? 3,
});
improvementsCount += 1;
}
info("coach", `drain: analysed ${pending.length} deaths → ${lessonsCount} lessons, ${improvementsCount} improvement requests`);
return { ok: true, analysed: pending.length, lessons: lessonsCount, improvements: improvementsCount };
}
// Mode names from runtime/modes.js (advice.js#MODE_TO_SKILL) — we accept
// these at write time because advice.js maps them to real skills at consult.
const KNOWN_MODE_NAMES = new Set([
"self_preservation", "night_shelter", "hunger", "shelter",
"flee", "sleep", "eat", "tunnel_out", "tunnel-out", "explore", "wander",
]);
function isLikelyModeName(s) {
if (!s || typeof s !== "string") return false;
return KNOWN_MODE_NAMES.has(s.toLowerCase().trim());
}
function buildPrompt(deaths) {
const summary = deaths.map((d) => {
const ctx = safeParse(d.context_blob);
const tail = ctx?.recentScenarios ?? [];
const tailFmt = tail.slice(-6).map((s) => ` - ${s.skillId} ${s.code}`).join("\n");
return [
`death id=${d.id} ts=${new Date(d.ts).toISOString()}`,
` position: (${Math.round(d.x ?? 0)}, ${Math.round(d.y ?? 0)}, ${Math.round(d.z ?? 0)})`,
` cause: ${d.cause}`,
` hostile: ${d.hostile ?? "(none)"}`,
` last skill: ${d.last_skill ?? "(none)"} (code: ${d.last_skill_code ?? "?"})`,
` hp at death: 0 food: ${d.food_at_death ?? "?"}`,
tailFmt ? ` recent dispatches:\n${tailFmt}` : null,
].filter(Boolean).join("\n");
}).join("\n\n");
const system = [
"You are reviewing recent deaths of an autonomous Minecraft survival bot (pepa).",
"The bot is trying to gather wood, craft tools, build a small village, and survive nights.",
"Your job: extract 1-3 short, generalised lessons + flag any missing-skill gaps.",
"",
skillRegistryPrompt({ limit: 1800 }),
"",
"Reply with ONE JSON object (no markdown fences):",
'{ "cause": "<short>", "next_action": "<one-sentence directive>",',
' "lessons": [',
' { "lesson": "...", "category": "combat|pathing|crafting|survival|social",',
' "trigger_skill": "<skill id or null>",',
' "trigger_hostile": "<mob name or null>",',
' "avoid_skill": "<registered skill id to NOT dispatch, or null>",',
' "prefer_skill": "<registered skill id to use instead, or null>",',
' "confidence": 0.7 } ],',
' "improvements": [',
' { "title": "<≤80 chars: what skill/feature is missing>",',
' "description": "<why current registry doesn\'t cover this; concrete example>",',
' "category": "skill|tuning|perception|planning|social|other",',
' "priority": 1 } ] }',
"",
"Keep each lesson under 30 words. Be specific.",
"CRITICAL: avoid_skill and prefer_skill MUST be one of the registered ids above, or null.",
"Use 'improvements' ONLY when a death is plausibly caused by the bot lacking a skill that doesn't exist in the registry (e.g. 'no skill to craft iron armor'). Skip it otherwise.",
].join("\n");
const user = `DEATHS:\n${summary}`;
return { system, user };
}
function extractJson(text) {
if (!text) return null;
// Try to find a JSON object somewhere in the reply.
const cleaned = text.trim().replace(/^```(?:json)?/, "").replace(/```$/, "").trim();
try { return JSON.parse(cleaned); } catch {}
const m = cleaned.match(/\{[\s\S]*\}/);
if (!m) return null;
try { return JSON.parse(m[0]); } catch { return null; }
}
function asArray(v) {
if (Array.isArray(v)) return v;
if (v && typeof v === "object") return [v];
return [];
}
function clamp(v, lo, hi) { return Math.max(lo, Math.min(hi, v)); }
function safeParse(s) { try { return JSON.parse(s); } catch { return null; } }
// Test-only exports
export const __testing = { captureDeath, buildPrompt, extractJson, inferCause, isLikelyModeName, KNOWN_MODE_NAMES };