Coli — Scheduled Task Supervisor for pi

中文文档

Coli is a pi-coding-agent extension for deferred execution + multi-round supervision:

• The working agent does the task with full tools/context.
• A different supervisor model evaluates each round (done / continue / fail).
• Supervision is bounded by maxRounds (default: 5, range: 1-20).

---

Install

pi install git:github.com/cocolinfff/coli

Then run /reload (or restart pi).

Quick Start (Usage First)

1) Configure a supervisor model (required once)

/coli setup model

This auto-discovers available models and prefers any model whose display name contains [coli].

You can also set one explicitly:

/coli setup model anthropic/claude-haiku-4-5

2) Schedule a task

/coli schedule in 2 minutes Check and fix TypeScript errors under src/

Example output:

✓ Coli task scheduled [a1b2c3d4]
Trigger: 14:32:00 (1m 58s)
Task: Check and fix TypeScript errors under src/
Supervisor: anthropic/claude-haiku-4-5
Max rounds: 5

After the trigger time, Coli injects the task into the current session as a follow-up message and starts supervision on each agent_end event.

---

Command Reference

Command	Description
/coli help	Show help
/coli setup	Show current config
/coli setup model	Auto-discover supervisor model (prefers [coli])
/coli setup model <provider>/<modelId>	Set supervisor model explicitly
/coli setup rounds <n>	Set default max rounds (1-20)
/coli setup reset	Reset config
/coli schedule <when> <task>	Schedule a supervised task
/coli list	List all tasks
/coli cancel <id>	Cancel one task

Time formats for /coli schedule

Format	Example
in X minutes	in 10 minutes
in X hours	in 2 hours
in X seconds	in 30 seconds
at HH:MM	at 15:30
Milliseconds delay	300000

---

How It Works

User / LLM -> schedule task (delay + supervisor) -> timer fires
                                                  |
                                                  v
                                     Working agent executes task
                                                  |
                                               agent_end
                                                  |
                                                  v
                                   Supervisor evaluates this round
                               -> done | continue | fail decision

Model selection priority

1. Explicit config from /coli setup model <provider>/<id>
2. First available model whose name includes [coli]
3. First available model
4. Current session model (last fallback)

Safety and limits

• Supervisor is evaluation-only (no file I/O, no command execution).
• Hard round cap (maxRounds) prevents infinite loops.
• Task timers are session-bound and in-memory.
• On session_shutdown, active timers are cleaned up.

---

Practical Examples

Delayed review

/coli schedule in 5 minutes Review all changes under src/auth/, find security/logic issues, and fix them

Nightly test-and-fix

/coli schedule at 02:00 Run npm test, analyze failures, fix code, rerun until all tests pass

Manual follow-up

After refactoring, run /coli schedule yourself for post-check validation.

---

Limitations

• Timer precision is second-level, not exact millisecond real-time.
• If pi stops, in-memory timers are lost.
• Supervision uses one-shot model calls and depends on provider/API availability.

License

MIT