UNPKG

@phren/cli

Version:

Knowledge layer for AI agents - CLI, MCP server, and data layer

github.com/alaarab/phren

alaarab/phren

30 lines (25 loc) • 1.74 kB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
--- name: fanout description: Delegate independent, bounded implementation or review work to headless agent workers using phren fanout. --- Fan out when independent tasks can make progress alongside the lead's work. Give each writer a separate worktree and a concrete output to validate. A brief should name the goal, owned files, constraints, tests, and completion report. Include relevant findings and exact review issues; avoid sending the whole chat. Start every worker through `phren fanout run`, never with `opencode run`, `codex exec` or `claude -p` of your own (directly, in a loop, or from a script with nohup). `phren fanout run` writes the job manifest that ties the worker to your conversation; a worker without one is invisible to the phone, the conductor and `phren fanout list`, and nothing can resume, review or archive it. Enable Hook and fanout, then use `phren fanout run --tier narrow|wide|review --label LABEL --worktree PATH < brief.txt`. The command chooses an eligible provider using current usage, errors and concurrency. Use `phren fanout usage` to inspect it, and `--provider`/`--model` for an explicit choice subject to caps. A Codex agent that wants a Claude worker or reviewer runs `phren fanout run --provider claude --model opus` (add `--tier review` for a read-only one); it shows under the Codex conversation like any other worker. Use `phren fanout list` to inspect jobs, `phren fanout resume JOB < feedback.txt` to send a bounded follow-up, and `phren fanout review JOB` for a review round. Codex resumes inherit their original sandbox; ask a resumed author to review without edits. Use a fresh review-tier worker for an enforced read-only sandbox. Archive completed jobs with `phren fanout archive`.