Files
c910457817 v0.4.0 vNext — closed-loop world model + settlement contract (#29)
* feat(v0.4.0): vNext — closed-loop world model + settlement contract

Implements the vNext architecture from the research doc: demote the noisy
multi-rail planner in favour of a closed loop (world truth → invariant check)
plus a single utility-driven goal authority.

L1 services (fix no_drop / silent pathfinder hang first):
- InventoryLedger: diff-based "did I actually get it" verifier; acquire-food
  now confirms via ledger.gainedSince instead of the unreliable count/event.
- MotionService.gotoSafe: wall-clock timeout + progress watchdog +
  path_update(noPath/timeout) → structured {reached|stuck|timeout|nopath}.

L3 plan — unify the three competing rails (curriculum/manifesto/storyline):
- Settlement Contract: ordered M0–M9 milestones, each invariant-checked
  against an authoritative world view (early steps delegate to the proven
  curriculum; late game adds farming).
- InvariantChecker + predicate library; GoalManager selects the lowest unmet
  milestone via utility argmax (food-urgency preempts, DEPS-style).
- Wired into the scheduler: bot.js precomputes snapshot.contract; reflex.js
  consumes it in place of the storyline rail. Manifesto L0 still preempts.

Eval + robustness:
- Village Score (single 0..1 metric) on the snapshot + TUI "build" line.
- survive.dig-in skill + dusk_dig_in mode (exposed at night, no bed → cover).
- approach_block helper (GoalNear + lookAt, avoids GoalLookAtBlock #341).

+28 new tests (450 total green). LLM remains entirely off the tick path.

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

* feat(v0.4.0): finish vNext plan — anti-loop, skill-graph, worldDelta diff, flee→motion

Completes the remaining v0.4.0 plan items and one fix motivated by a live
in-game observation (flee hanging 30s against a persistent zombie).

- flee → MotionService.gotoSafe: structured {stuck|timeout|nopath} in ~4s with
  a blind-retreat fallback, instead of the observed 30s pathfinder hang + 3
  watchdog replans. Movements setup guarded so it is unit-testable.
- QW5 anti-loop (runtime/anti-loop.js): same skill failing >=3x in 5min →
  30min blacklist (reflex shouldSkip) + one-shot improvement_request
  (bot.js drainFired -> writeProposal).
- 4.1 closed-loop worldDelta: runSkill snapshots inventory before execute and
  attaches the real delta (_invObserved) to every successful result; opt-in
  skill.expectGain asserts the claimed gain or returns world_unchanged.
- 3.6 skill-graph (Plan4MC): declarative requires/produces for ~20 skills;
  prerequisitesMet/canRun/runnableFrontier; GoalManager annotates suggestions
  with blockedBy when prereqs are unmet.

+22 tests (472 total green). Live smoke confirmed dig-in works and no new
errors; flee loop is what this commit's flee migration addresses.

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>
2026-05-28 11:29:39 +03:00

71 lines
2.6 KiB
JavaScript

// Anti-loop detector (QW5). The scheduler already skips a (skill, situation)
// that has failed repeatedly (scenario-memory.shouldSkip), but that is keyed on
// a coarse situation hash and never escalates. This detector closes the loop
// the research describes: when the SAME skill fails N times inside a short
// window with no success in between, it (a) blacklists that skill for a cool-off
// and (b) emits a one-shot "fired" record the runtime turns into an
// improvement_request — so an operator/Codex gets a ticket instead of the bot
// silently thrashing (e.g. the live flee↔dig-in loop we observed).
//
// Pure + deterministic: inject `now` in tests. No disk, no bot.
export function createAntiLoop({
windowMs = 5 * 60_000,
threshold = 3,
blacklistMs = 30 * 60_000,
refireCooldownMs = 30 * 60_000,
} = {}) {
const state = new Map(); // key -> { fails: number[], blacklistUntil, lastFiredAt }
const firedQueue = [];
function keyOf(skillId, targetKey) {
return targetKey ? `${skillId}@${targetKey}` : skillId;
}
function get(key) {
let s = state.get(key);
// lastFiredAt = -Infinity so the FIRST loop always fires (a real epoch
// `now` minus 0 would otherwise be < refireCooldownMs early in uptime).
if (!s) { s = { fails: [], blacklistUntil: 0, lastFiredAt: Number.NEGATIVE_INFINITY }; state.set(key, s); }
return s;
}
function record({ skillId, ok, code = null, targetKey = null, detail = null, now = Date.now() }) {
if (!skillId) return { fired: false };
const key = keyOf(skillId, targetKey);
const s = get(key);
if (ok) { s.fails = []; return { fired: false }; }
s.fails.push(now);
s.fails = s.fails.filter((t) => now - t <= windowMs);
if (s.fails.length >= threshold) {
s.blacklistUntil = now + blacklistMs;
const count = s.fails.length;
s.fails = []; // reset streak so we don't blacklist-spam every further fail
if (now - s.lastFiredAt >= refireCooldownMs) {
s.lastFiredAt = now;
const fired = { key, skillId, targetKey, count, code, detail, ts: now, until: s.blacklistUntil };
firedQueue.push(fired);
return { fired: true, ...fired };
}
}
return { fired: false };
}
function shouldSkip(skillId, targetKey = null, now = Date.now()) {
const s = state.get(keyOf(skillId, targetKey));
return !!s && now < s.blacklistUntil;
}
// Returns and clears the queue of newly-fired loops (for improvement_requests).
function drainFired() {
return firedQueue.splice(0);
}
function snapshot() {
return { tracked: state.size, pendingFired: firedQueue.length };
}
return { record, shouldSkip, drainFired, snapshot, _state: () => state };
}