pepa-pi-bot/dev/v0.2.0/STATUS.md

pepa v0.2.0 — status

Read this first if you're continuing the iteration in a new session. It's the single source of truth for "where we are, what shipped, what's next".

Live state snapshot lives in state/<host>/ (gitignored). Code lives on main. Anything in this folder is shared knowledge for future contributors / sessions.

Last update: 2026-05-27, after rc.3 merged (PR #22).

Current code: 0.2.0-rc.3. Live bot running on this version since 2026-05-27 15:32 (supervisor PID 31432, child PID rotates on hot-restart).

---

What v0.2.0 is

Major iteration over v0.1.x. The bot acquires a self-learning substrate: structured knowledge in SQLite, post-mortem analysis of every death by Pi, Russian chat narration, retrieval-augmented dispatch (lessons override planned actions), and a 30-min self-reflection loop where Pi asks the bot "are you in a loop?".

It also tightens the self-improvement pipeline: auto-patch now opens a PR for operator review instead of cherry-picking onto main (branch protection enforces — only operator can merge).

See docs/v0.2.0-self-learning.md for the original design doc.

---

Shipped (rc.1 → rc.3)

rc.1 — substrate ([PR #?]/9c2c56f)

• runtime/knowledge/{schema.sql,store.js,seed.js,lessons.js,index.js} — SQLite knowledge base at state/<host>/knowledge.db. 9 tables (recipes, mob_intel, block_intel, lessons, deaths, postmortems, poi, wiki_pages, chat_log, code_changes). Seeded from docs/minecraft-recipes.json + inline starter intel: 38 recipes, 15 mobs (creeper/zombie/etc. with verdict_no_weapon), 30 blocks, 12 starter lessons (creeper rule, "first wood→pickaxe→sword", etc.).
• runtime/coach/postmortem.js — bot.on('death') captures context (last skill, hostile, nearby journal, scenarios tail) → deaths table. Periodic drain (5min, ≤3 Pi calls/h) batches up to 8 unanalysed deaths and asks Pi for generalised lessons. Pi reply parsed into postmortems + lessons rows.
• runtime/coach/advice.js + reflex hooks — consultAdvice() runs before dispatch in curriculumReflex and defendReflex. Lessons with confidence ≥ 0.6 and matching avoid_skill swap the planned skill with prefer_skill (if it's in SAFE_OVERRIDES).
• runtime/persona/chatter.js — Russian narration in MC chat on skill transitions, threat spotted, day/night, respawn, milestone done. Rate-limited 8/h, min 75s gap, no-dup suppression.
• auto-patch.js switched to PR-open (no more direct cherry-pick to main). PEPA_AUTO_PATCH_MERGE=cherry-pick fallback for emergencies.
• Branch protection on main: 1 required approver, enforce_admins:false.

rc.2 — P0 hardening ([PR #21]/e84148d)

• PEPA_HEADLESS=1 — pi -p subprocesses (banter/coach/planner) no longer open a second MC connection. extensions/mineflayer-bridge.ts: session_start skips connect("startup") when env is set; pi-bridge sets it on every spawn.
• Test state isolation — runtime/config.js detects node test runner / PEPA_STATE_DIR and redirects stateDir to /tmp/pepa-test-state-<pid>/. npm test no longer pollutes live scenarios.jsonl / world-journal.jsonl / daily log.
• defendReflex outcome bug fix — reportAdviceOutcome was called with succeeded:false before the flee skill returned. Now wired through dispatchDefendFlee onComplete.
• Mode-name → skill-id translation in coach/advice.js. Pi-coach prefer values like "night_shelter", "self_preservation", "hunger" now map to survive.sleep / survive.flee / survive.eat.
• Self-reflection loop (runtime/coach/reflect.js). Every 30 min Pi gets asked "are you making progress, or in a loop?". Verdict + summary + new lessons. Writes to state/<host>/reflections/<ts>.md, lessons land in DB with source="pi-reflect". Rate-limited 2/h.

rc.3 — escape mechanics ([PR #22]/865aae1)

• survive.pillar-up — vertical escape skill. Places dirt/cobble/ planks under self, jumps onto it. Up to 8 steps per dispatch. NO pickaxe required.
• Wedged-emergency reflex — at top of curriculumReflex: if no hostile <6m AND position hasn't moved 16+ blocks in 60s AND placeable block in inv → force pillar-up. 2-min cooldown.
• consult() in fallback path — curriculumReflex wander/explore.far fallback now consults advice too. This closes the gap where Pi-coach lessons fired but the dispatcher's fallback bypassed them.
• POI on death — coach/postmortem.js calls recordPOI({kind:"danger"}) on every death. 6h expiry. POI table now populates as bot plays.
• SAFE_OVERRIDES extended: + survive.pillar-up, + village.choose-base.

---

Current state (2026-05-27 ~15:30, ~1.5h after rc.3)

Live DB:

deaths              28
postmortems         22  (coach analyzed all but 6 — backlog draining at 5min/batch)
lessons             34  (12 starter + 22 Pi-extracted)
lessons_applied     1   ← still very low; advice consult fires but most matches fail
poi                 22  (rc.3 fired recordPOI on every death since 14:17 deploy)
chat_log            0   (chatter narrates in MC chat but doesn't write to DB table yet)
reflections         2   (14:51, 15:22 — 30-min interval working)

The bot is alive but still struggling. Coach is producing high-quality lessons (see state/<host>/reflections/2026-05-27T12-21-58-656Z.md: "Я не продвинулся к цели… выбрать новое безопасное дневное место"), but:

1. Pi-coach + Pi-reflect routinely invent prefer_skill values that aren't registered skill ids (e.g. choose.safe.surface, relocate.daylight). normalisePreferSkill only knows mode names, not these. So even the right advice can't dispatch the right action.
2. The bot's home terrain (~608, 90) is genuinely bad — stone walls, no pickaxe to break out, hostiles spawn nearby. pillar-up can help but it needs a placeable block AND open sky above. In a closed cave it places dirt and hits a ceiling.
3. gather.stone returns missing_tool correctly (rc.0 was already fine here), but curriculum doesn't auto-route to craft.wooden-pickaxe — it just backs off and tries the same thing.

---

Known issues / followups (rc.4 candidates)

Ranked roughly by impact.

1. Pi-suggested prefer_skill is often a hallucinated action name

Symptom: reflections produce prefer: choose.safe.surface, prefer: relocate.daylight, prefer: defend.self — none are registered skills. SAFE_OVERRIDES check fails, lesson falls back to bare avoid (which dispatches nothing).

Fix options:

• (a) Include the full registered-skill list in the Pi prompt so it picks from valid IDs.
• (b) Add a fuzzy mapper: "choose|find|new + base|spot|place" → village.choose-base. "relocate|move + safe|day" → explore.far.
• (c) Both. (b) is a fallback safety net.

Where: runtime/coach/{postmortem.js, reflect.js} prompt builder + runtime/coach/advice.js normalisePreferSkill.

2. Tool-progression auto-craft

Symptom: gather.stone returns missing_tool → curriculum backs off → next tick tries gather.stone again → loop. Should auto-suggest craft.wooden-pickaxe. Same pattern for axe-required gathering.

Fix: In curriculum or in a new "tool-need" reflex, when a gather skill returns missing_tool, look up the tier requirement (already in block_intel.required_tool) and route to the matching craft skill if materials are available, else to the prerequisite gather skill.

Where: runtime/reflex.js or new runtime/coach/tool-progression.js.

3. village.choose-base should penalise danger POI

Symptom: bot may pick a new base that's adjacent to a danger POI (a recent death site). rc.3 records the POIs but choose-base scoring doesn't read them.

Fix: In runtime/base-site.js scoreCurrentPosition, query poiNearby({kind:"danger", x, z, radius: 32}) and subtract a penalty proportional to recent danger count.

4. Pillar-up perimeter sense

Symptom: in a closed cave, pillar-up places a block and hits the ceiling block above. Useless.

Fix: before starting pillar-up, scan a 3×3 above current head; if solid blocks in the column path, abort and request recovery.tunnel-out instead. Or walk to a column-clear cell within reach first.

5. chat_log table not populated

Symptom: 0 rows in chat_log despite bot narrating ~8 lines/hour.

Fix: runtime/persona/chatter.js sendChat() calls logChat() from knowledge index. Also wire the inbound side in bot.js bot.on("chat") handler.

6. Reflections don't write lessons to the DB until they parse

Symptom: confirmed working; just noting that the verdict + summary themselves aren't queryable. Currently only lessons table benefits from reflection.

Possible fix: add a reflections table mirroring the JSON output. Or run lessons-from-reflection through the same advice path so they get applied_count.

7. Re-narration spam

Symptom: persona narrated за дровами twice in 6 seconds after hot-restart. Cooldown is per-process state, lost on restart.

Fix: persist _lastNarrationAt and _narrationTimes in a small file in state/<host>/ so restarts inherit the cooldown.

---

Next session — quick start

1. Read this file first.
2. Check live DB for the latest counters:

   sqlite3 state/play.xmatic.team_25565/knowledge.db \
     "SELECT 'deaths', COUNT(*) FROM deaths
      UNION SELECT 'lessons', COUNT(*) FROM lessons
      UNION SELECT 'lessons_applied', COUNT(*) FROM lessons WHERE applied_count > 0
      UNION SELECT 'poi', COUNT(*) FROM poi;"
   

3. Read the latest reflection for current Pi assessment:

   ls -t state/play.xmatic.team_25565/reflections/ | head -1
   

4. Check bot is alive: ps aux | grep "runtime/(bot|supervisor)".
5. Pick a followup from §"Known issues" (1 and 2 are the highest impact: actually-applying-lessons + tool progression).

---

Workflow notes for next session

• main is protected — only the operator merges PRs. Auto-patch from the bot itself opens PRs via gh pr create.
• Tests: npm test (255+ green at last check). Tests run isolated under /tmp/pepa-test-state-<pid>/ and don't pollute live state.
• The bot supervisor watches runtime/**/*.js and hot-restarts on file change. After merging a PR, git pull locally and the supervisor picks up the new code.
• Stop bot: bash scripts/stop.sh. Start: npm run bot (inside tmux is fine).
• Memory in state/<host>/ is server-specific and gitignored.
• Don't bypass branch protection — the auto-mode classifier will block it.