v0.2.0-rc.2: P0 hardening — Pi headless, test isolation, advice fixes #21

2026-05-27T13:57:09+03:00

halofourteen commented

2026-05-27 13:57:09 +03:00

(Migrated from github.com)

Summary

P0 hardening patch on top of rc.1, addressing four real bugs spotted from
live observation + adding the self-reflection loop the operator asked
for.

Bugs fixed

Pi subprocess opens second MC connection. When the hybrid runtime
spawns pi -p \"...\" for banter / coach / planner, the loaded
extensions/mineflayer-bridge.ts calls connect(\"startup\") and
races for the nickname. PEPA_HEADLESS=1 env (set by
runtime/pi-bridge.js) now makes the bridge skip its connect.
Tests poison live state. runtime/log.js, scenario-memory.js,
world-journal.js, and knowledge/store.js all key off
config.stateDir. Under npm test that resolved to the live
state/<host>/, so test rows ended up in real scenarios.jsonl.
config.js now detects node test context (or PEPA_STATE_DIR
override) and redirects to /tmp/pepa-test-state-<pid>/. ~550
historical test rows cleaned from live state in this commit.
defendReflex marks lesson failed too early. Combat advice override
used to call reportAdviceOutcome({succeeded:false}) before the
flee skill returned. The creeper-rule lesson got
applied_count=2, succeeded_count=0 not because the lesson is bad
but because we marked it bad pre-emptively. Now dispatchDefendFlee
carries lessonId and reports the actual flee outcome.
Pi-coach mode-name leak. Coach lessons sometimes set
prefer_skill: \"night_shelter\" or self_preservation (those are
mode names, not skill ids), so they failed SAFE_OVERRIDES and
silently fell to avoid instead of override.
normalisePreferSkill in coach/advice.js translates mode names
and underscore-shapes to canonical skill ids.

New behavior

Self-reflection loop (runtime/coach/reflect.js). Every 30 min
Pi gets asked: "Are you making progress, or in a loop? What should
you do?" Reply parsed into verdict + summary + next-action + new
lessons. Reflection written to state/<host>/reflections/<ts>.md;
lessons land in knowledge.lessons with source=\"pi-reflect\".
Rate-limited 2 calls/hour.

Test plan

npm test — 246/246 green (was 237; +9 advice/reflect coverage).
Verified empirically: after the config fix, npm test adds 0
rows to the live scenarios.jsonl.
extensions/mineflayer-bridge.ts typecheck — diff only adds an
early-return guard.

Followups (rc.3 candidates)

Wedged-pit emergency policy: at low HP + no food + no movement, force
village.choose-base or survive.sleep instead of explore.far.
Promote village.choose-base early in the curriculum when bot's been
stuck in same chunk > N minutes.
Tool-progression auto-craft: if gather.stone returns missing_tool,
curriculum should auto-dispatch craft.wooden-pickaxe.

## Summary P0 hardening patch on top of rc.1, addressing four real bugs spotted from live observation + adding the self-reflection loop the operator asked for. ### Bugs fixed 1. **Pi subprocess opens second MC connection.** When the hybrid runtime spawns `pi -p \"...\"` for banter / coach / planner, the loaded `extensions/mineflayer-bridge.ts` calls `connect(\"startup\")` and races for the nickname. `PEPA_HEADLESS=1` env (set by `runtime/pi-bridge.js`) now makes the bridge skip its connect. 2. **Tests poison live state.** `runtime/log.js`, `scenario-memory.js`, `world-journal.js`, and `knowledge/store.js` all key off `config.stateDir`. Under `npm test` that resolved to the live `state/<host>/`, so test rows ended up in real `scenarios.jsonl`. `config.js` now detects node test context (or `PEPA_STATE_DIR` override) and redirects to `/tmp/pepa-test-state-<pid>/`. ~550 historical test rows cleaned from live state in this commit. 3. **defendReflex marks lesson failed too early.** Combat advice override used to call `reportAdviceOutcome({succeeded:false})` *before* the flee skill returned. The creeper-rule lesson got `applied_count=2, succeeded_count=0` not because the lesson is bad but because we marked it bad pre-emptively. Now `dispatchDefendFlee` carries `lessonId` and reports the actual flee outcome. 4. **Pi-coach mode-name leak.** Coach lessons sometimes set `prefer_skill: \"night_shelter\"` or `self_preservation` (those are mode names, not skill ids), so they failed `SAFE_OVERRIDES` and silently fell to `avoid` instead of `override`. `normalisePreferSkill` in `coach/advice.js` translates mode names and underscore-shapes to canonical skill ids. ### New behavior 5. **Self-reflection loop** (`runtime/coach/reflect.js`). Every 30 min Pi gets asked: \"Are you making progress, or in a loop? What should you do?\" Reply parsed into verdict + summary + next-action + new lessons. Reflection written to `state/<host>/reflections/<ts>.md`; lessons land in `knowledge.lessons` with `source=\"pi-reflect\"`. Rate-limited 2 calls/hour. ## Test plan - [x] `npm test` — 246/246 green (was 237; +9 advice/reflect coverage). - [x] Verified empirically: after the config fix, `npm test` adds 0 rows to the live `scenarios.jsonl`. - [x] `extensions/mineflayer-bridge.ts` typecheck — diff only adds an early-return guard. ## Followups (rc.3 candidates) - Wedged-pit emergency policy: at low HP + no food + no movement, force `village.choose-base` or `survive.sleep` instead of `explore.far`. - Promote `village.choose-base` early in the curriculum when bot's been stuck in same chunk > N minutes. - Tool-progression auto-craft: if `gather.stone` returns `missing_tool`, curriculum should auto-dispatch `craft.wooden-pickaxe`.

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: xmatic/pepa-pi-bot#21