opencode-delegate skill

OpenCode Delegate

You are the orchestrator. This skill lets you hand a bounded coding task to a separate implementer — the OpenCode CLI — then review what it produced and land it yourself. You write the brief and own the judgment; OpenCode does the typing in its own session; you verify and commit.

Nothing here is specific to one orchestrating agent. The loop needs only the ability to run a shell command and read a file, so any agent with those two capabilities — Claude Code, OpenCode driving a sibling session, or a comparable one — can drive it. (It is designed for and run on Claude Code; treat other orchestrators as designed-for, not yet proven.)

When NOT to use this

• The task is small enough to just do inline — delegation overhead is not worth it.
• The opencode CLI is not installed or not authenticated (run opencode auth login).
• You want to write the code yourself, or you only need a review (use the plan agent via --read-only).

Prerequisites (check once)

opencode.ai) and opencode auth login.

1. opencode --version succeeds. If not, install (npm i -g opencode-ai, or the native installer from

opencode --version its version. The relay records the version it ran into result.json, so a stale binary is visible after the fact.

1. Confirm which opencode is on PATH. command -v opencode shows the active binary and

1. A model provider is authenticated — opencode auth list shows at least one credential.
2. You are in (or will point --cd at) the target git repository.

Choose the implementer model

OpenCode has no safe default — a bare opencode run errors — so a fresh run needs a model via --model or a fleet --lane that sets one (a resumed run inherits its session's model). Naming the model is the one decision a single-model backend like codex-delegate never had, and it has two owners:

per token (OpenRouter and the like); only the human knows which are their flat-rate subscriptions, and the CLI can't tell them apart. So the usable set is theirs — ideally stated once in the repo's AGENTS.md or their CLAUDE.md (e.g. "delegate mechanical work to opencode-go/…, hard logic to …").

• The human owns which models are allowed. opencode models lists hundreds of entries, most billed

model for a mechanical sweep (rename, migration, removal); a strong one for a subtle bug or a money/security path.

• You, the orchestrator, pick per task — from that set. Match the model to the brief: a cheap, fast

a surprise bill. Name the constraint to the human and let them choose.

• If no usable set is stated, ask — don't guess. Guessing from the catalog risks a metered model and

More depth: references/writing-the-brief.md.

The loop

Run these five steps per task. Steps 1, 4, and 5 are your judgment; 2 and 3 are mechanical.

1. Write the brief

OpenCode sees only the text you send plus what it can read from the working tree — no chat history, no shared context. Everything the task needs goes in the brief: the goal, the current state, what to change, what to leave untouched, the project's actual gate commands (discover them from the repo's AGENTS.md/CLAUDE.md/Makefile — do not assume), and a report contract. Tell OpenCode it will not commit (you will). Keep one task per brief. Full guidance and a template: references/writing-the-brief.md.

2. Dispatch

Send the brief to OpenCode with the bundled helper. It wraps opencode run, captures the run, and writes a structured result.json — so your only job is "run a command, read a file." ( below is this skill's installed directory — the folder containing this SKILL.md. Claude Code prints it as "Base directory for this skill" when the skill loads; on other orchestrators use that same directory — if unsure where it landed, run find ~ -name relay.mjs -path 'opencode-delegate' and substitute the directory above it.)

node "<skill-dir>/scripts/relay.mjs" --brief brief.txt --model <provider/model> --cd /path/to/repo
# --model (or a --lane that sets model) is required on a fresh run
# fleet lane from delegate-setup:           add --lane <name>  (dials apply; flags still win)
# read-only (review/diagnosis, no edits):   add --read-only   (uses the plan agent)
# continue the previous OpenCode session:   add --resume-last  (delta brief only; keeps the model)
# hard time limit (watchdog):               add --timeout 2h  (default: off; implementation runs routinely need 1-2h)
# see all options:                          node .../relay.mjs --help

The helper defaults to the write-capable build agent and writes its artifacts to a temp dir, so the repo under review stays clean. It never commits — see step 5. Mechanics, flags, and the result.json shape: references/dispatch-and-poll.md.

3. Wait for completion

The helper blocks until OpenCode finishes, so back it with whatever your orchestrator offers and resume when it returns:

the result file — … & in bash/zsh (including Git Bash/WSL), or your shell's equivalent (Start-Job in PowerShell, start /b in cmd). The run is done when result.json exists with a status. (A pre-run usage error — bad args or an empty brief — instead exits with code 2 and writes no result file, so check the exit code too. A missing opencode binary exits 127 but does write a result.json with status opencode_unavailable.)

• Claude Code: run the Bash call with runinbackground: true; you are notified on completion.
• Plain shell / other agents: run it in the foreground for short tasks, or background it and poll

Do not trust progress trackers over reality: a run is finished when result.json is written and the process has exited. Read the working tree, not a status line. The implementer's full report is the finalMessage field in result.json (also printed in full on stdout between the report markers).

4. Review — do not trust the self-report

OpenCode's result.json includes its own final message and any gate claims. Re-verify, don't accept:

"gates passed" on faith.

• Re-run the project's gates yourself (the test/lint/build commands from step 1). Never take

nothing less? touchedFiles in the result is your starting point.

• Read the diff against the brief: did OpenCode do what was asked, nothing more (scope creep) and

test-guard, etc. from guard-skills) — this skill produces the work; those skills judge it.

• Run the relevant guard skills on the diff if you have them installed (clean-code-guard,

• For schema/migration changes, round-trip them; for removals, grep for dangling references.

Full checklist: references/review-and-land.md.

5. Land it

The implementer edits the working tree; the orchestrator commits. Committing should be the act of the party that verified the work. Only after the gates pass and the diff holds:

review again.

• Commit the verified work yourself, with a clear message.
• If it needs changes, send a delta brief with --resume-last (don't restate the whole task) and

Autonomy model

OpenCode's autonomy is governed by the agent, not a sandbox enum:

equivalent of "let it implement."

• build (the relay default) — write-capable; edits files in the working dir headlessly. The

equivalent of "let it look but not edit."

• plan (via --read-only) — read-only; reviews and diagnoses without touching the tree. The

Permissions auto-approve by default: the relay passes --auto so a headless run never blocks on a prompt no one can answer. That is the point of unattended delegation — the orchestrator's diff review and the implementer sweep (step 4) are the safety net, not a per-action prompt. Pass --no-auto to honor the agent's own permission config instead (allow/ask/deny per action); pair it with an agent whose in-workspace permissions are set to allow, or a headless run can hang waiting on an ask. Read-only (plan) runs never get --auto — auto-approving would let the plan agent's ask-gated edit/bash permissions through and defeat "read-only," so a review can't be tricked into touching the tree.

Authorization model

Delegation is something the human opts into. Once they have ("run this queue", "proceed"), committing verified, gate-passing work is the agreed contract — that is the whole point. Two limits on that mandate: surface, don't absorb (report OpenCode's design decisions, defensible-but-unasked turns, and non-blocking nitpicks rather than silently keeping them) and stop for scope changes (if correct completion needs going beyond the brief, ask — don't expand the mandate yourself). The full treatment is in references/review-and-land.md.

References

execute blind: structure, XML blocks, the report contract, embedding the real gate commands.

• references/writing-the-brief.md — how to write a brief OpenCode can

result.json contract, backgrounding per orchestrator, and recovery when a run misbehaves.