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>
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>
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>
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>
mayatnikovandClaude Opus 4.7