lean-tea-meta — Gemini/LMStudio PM for Claude Code

A polling meta-agent that watches N zellij panes (each running claude / Claude Code) and decides when to nudge each one with a new instruction.

• Cheap classifier every stall — LMStudio + a local Gemma reads the pane snapshot, decides continue / instruct / ask_user. Bounded LLM spend even over multi-week kernel-tuning sessions.
• Heavy audit on demand — /review AGENT_ID re-reads the whole memo log + a bigger pane snapshot and asks Gemini 2.5 Pro (reviewBackend) for a concrete redirection.
• Multi-agent — each managed agent runs its own polling loop via IO.asTask, all state behind IO.Ref snapshots so the controller (TUI / REPL) reads them cheaply.
• Persistent — JSON config for the agent roster + LLM backends, per-agent JSONL memo log for --resume.

Rendered UI:

┌────────────────────────────────────────────────────────────┐
│  meta_orchestrator — 3 agent(s)                            │
│ ┌────────────────────────────────────────────────────────┐ │
│ │ kernel-bwd       running   pane=term_18  poll=42  [t+…]│ │
│ │ port-cuda        stable    pane=term_22  poll=17  [t+…]│ │
│ │ mem-audit        asking-… pane=term_26  poll=3   [t+…] │ │
│ └────────────────────────────────────────────────────────┘ │
│                                                            │
│  > /review kernel-bwd_                                     │
└────────────────────────────────────────────────────────────┘

Stack

• lean-tea — LeanTea.Cloud.Gemini, LeanTea.Llm.Openai, LeanTea.Tui
• zellij action dump-screen / write-chars — pane I/O via CLI
• FNV-1a hash of pane text — cheap stall detection

Run

lake update
lake build meta_orchestrator
./.lake/build/bin/meta_orchestrator --config meta_orchestrator.config.json

> /add kernel-bwd terminal_18 Optimise FlashAttn v3 bwd SM90 2x throughput
> /add port-cuda  terminal_22 Port the rest of the kernels from CUDA to CK
> /list
> /save

Required env: GEMINI_API_KEY (only if reviewBackend is a gemini kind; issue one at https://aistudio.google.com/apikey).

Config file shape

{
  "agents": [
    {"id": "kernel-bwd", "pane": "terminal_18",
     "goal": "Optimise FlashAttn v3 bwd SM90 2x throughput",
     "stallSec": 30, "pollSec": 5, "enabled": true}
  ],
  "logDir": "./logs",
  "decideBackend": {
    "type": "openaiCompat",
    "baseUrl": "http://127.0.0.1:11211/v1",
    "model": "local-model"
  },
  "reviewBackend": {
    "type": "gemini",
    "model": "gemini-2.5-pro"
  }
}

Slash commands

command	meaning
/list	show every agent + its current status
/add ID PANE GOAL...	register and start an agent
/stop ID	cooperatively stop the polling loop
/start ID	resume a stopped/disabled agent
/remove ID	drop the agent from the config
/reply ID TEXT...	inject a free-text reply (use when the decide backend said ask_user)
/review ID	run a heavy-weight audit via reviewBackend
/save [PATH]	persist the config to PATH
/load PATH	replace the in-memory config
/quit	stop every agent, save the config, exit

Layout

MetaOrchestrator/
  Zellij.lean    — dump-screen / write-chars / submit CLI wrappers
  Llm.lean       — Backend union (openaiCompat | gemini) + JSON codec
  Director.lean  — decide (JSON verdict) + review (free-form audit) prompts
  Config.lean    — ManagedAgent + Config records + JSON codec
  Runtime.lean   — per-agent polling loop (IO.asTask + IO.Ref)
  Tui.lean       — full-screen TUI on LeanTea.Tui
  Main.lean      — controller — boots agents, TUI or REPL