lopi-context

name	lopi-context
description	Full lopi project context — complete phase plan, current health metrics, standing rules for agents, crate details. Auto-load when working on lopi sprint planning, architecture decisions, or Phase 5+ work.
user-invocable	false

lopi — Full Project Context

Phase Plan

Phase	Status	What Shipped
1 — MVP Core	✅ v0.1.0	Cargo workspace, all 8 crates, CLI verbs
2 — N Parallel Agents + Live Dashboard	✅ v0.4.0	AgentPool, EventBus, ratatui TUI, web dashboard, WebSocket
3 — Remote Control + Pattern Mining	✅ v0.3.0	Telegram bot, GitHub webhook HMAC, pattern miner
4 — Scheduled Tasks + Repo Profiles	✅ v0.5.0	cron scheduler, RepoProfile, lopi watch --remote
5 — Self-Improvement Engine	🔲 Next	Pattern learning, failure post-mortem, adaptive retry
6 — GitHub Webhooks + CI Integration	🔲 Planned	lopi-webhook end-to-end, GitHub App mode
7 — Production Web UI	🔲 Planned	React/Svelte frontend, auth, mobile-responsive
8 — Native Mobile App	🔲 Planned	React Native, push notifications, voice input
9 — Intelligence + Evolution	🔲 Ongoing	Multi-agent roles, goal decomposition, embedding memory

Current Health (Phase 4 / v0.5.0)

• Tests: 46 passing, 0 failing
• Build: Clean (0 warnings)
• Crates: 9 (lopi-core, lopi-git, lopi-agent, lopi-memory, lopi-orchestrator, lopi-ui, lopi-remote, lopi-webhook, lopi-toon)
• CLI: run, watch, tail, dock, sail, schedules
• API: GET /api/tasks, POST /api/tasks, GET /api/tasks/:id, GET /api/patterns, GET /api/health, GET /ws/tasks

Phase 5 — Self-Improvement Engine (Next Sprint)

Priority items:

1. Pattern learning: before running a new task, query similar past tasks → suggest constraints that worked → pre-load into system prompt
2. lopi learn CLI — show pattern library, success rates, top constraints
3. Failure post-mortem: when all retries fail, run a "post-mortem" Claude session → analyze error log → generate constraint/approach suggestion → store as pattern
4. Adaptive retry: if attempt N failed with error type X, adjust prompt strategy for attempt N+1 (pass error + suggest different approach)
5. Self-modification loop (guarded): allow_self_modify = true in config + same git isolation + PR workflow applies
6. Scoring evolution: score weights configurable and tunable based on which metrics correlate with user-approved vs rejected PRs

Standing Rules for Agents on lopi

• Stay inside crates/ and src/ — never touch .github/, infra/, root Cargo.lock deliberately
• cargo build must stay green — if it goes red, fix before doing anything else
• No unwrap() outside tests — use anyhow::Result and ?
• No silent failure — if a fallback path swallows an error, log via tracing::warn!
• Async everywhere — Tokio is the only runtime; no blocking I/O on async paths (use spawn_blocking)

Key Architecture Decisions (Load-Bearing)

• EventBus<TaskStatus> uses tokio::broadcast — subscriber lag drops old events; this is intentional
• AgentPool uses Arc<Semaphore> for bounded concurrency — no global mutex on the task queue
• lopi-memory uses sqlx with SQLite + WAL mode — concurrent readers are fine, one writer at a time
• Pattern mining: keyword fingerprint extraction via find_similar_patterns(goal) — not embedding-based yet (Phase 9)
• GitHub webhook HMAC: constant-time comparison via subtle::ConstantTimeEq — do not simplify to ==