rjv-opencode-agentslisted
Install: claude install-skill rjvim/ai-skills
# opencode agents: launch, watch, recover
opencode is the cheapest worker (DeepSeek v4.1 Flash, one retry on
MiniMax M3). It is also the one most likely to run long or die
silently. The launcher never waits on it in the foreground: every run
goes through the **watcher**, and you keep working while it reports.
Which tasks go to opencode, and the 2 / 5 minute boxes, are owned by
`rjv-subagents-claude` (Codex: `rjv-subagents-codex`). This skill owns
what happens after you decide.
## 1. Pick the mode
| The task needs... | Mode | Where opencode runs |
|---|---|---|
| only the text in the prompt | `text` | a fresh empty directory the script makes |
| to read a repo (review, recon) | `repo` | the repo, with `--auto` |
**opencode reads files and runs shell commands in its working
directory even without `--auto`** (probe, 2026-11-04: it read a file
and wrote another with no approval). So `text` mode always runs in an
empty directory, and anything that touches secrets, customer data or
production stays on Claude, as `rjv-subagents-claude` §3 says.
In `repo` mode the brief says read-only unless edits are the task. The
watcher fingerprints the git tree before and after. `tree=CHANGED` on a
read-only brief: discard the result, list the paths from `git status
--porcelain` in your report, and never commit them. Revert only when the
user says so.
## 2. Launch through the watcher
Write the brief to a file, then start the watcher. It runs on its own;
you never sit in a foreground wait