delegate-to-opencode

A Claude Code skill for orchestrating the OpenCode CLI as a background build agent. Claude stays the orchestrator and reviewer; opencode does the heavy lifting; you get a reviewable diff.

It ships two things:

• SKILL.md — the operating manual the orchestrating agent reads: when to offload, how to split a task into cleanly-scoped subtasks, how to write a self-contained brief, how to launch, and how to verify and reintegrate what came back.
• scripts/run-opencode.sh — a wrapper around opencode run that makes headless runs loggable and resumable.

Works with any agent that can read a skill and run a shell command — Claude Code, Codex, or a plain terminal.

What the wrapper handles

• Auto-permissions. Runs use --auto so opencode is not asked to confirm tool calls. Pass --no-auto to require approval instead.
• Deterministic resume. The session ID is captured from the JSON event stream and written to $RUN_DIR/session_id, so follow-ups work even if you didn't note the ID. Pass --resume to continue a session, or --resume + --fork to branch it.
• Clean JSON. Raw output is filtered to NDJSON in events.jsonl; the final assistant message is reconstructed into last_message.txt.
• Briefs via stdin. Prompts go through stdin, so multi-line briefs need no shell quoting.
• Run artifacts. Every run leaves a directory behind rather than scrolling past you.

Install

Requires opencode (configured via opencode auth login), bash, and jq.

git clone https://github.com/luckeyfaraday/delegate-to-opencode.git
ln -s "$PWD/delegate-to-opencode" ~/.claude/skills/opencode

Then verify:

opencode models        # prints all models across configured providers
~/.claude/skills/opencode/scripts/run-opencode.sh --help

Usage

The agent normally drives this, but the wrapper stands alone:

# Offload a task; brief comes from stdin, run lands in $RUN_DIR
./scripts/run-opencode.sh -C ~/code/myrepo -m anthropic/claude-sonnet-4 -l migrate - < brief.md

# Follow up on the same session, with full memory of the prior run
./scripts/run-opencode.sh -C ~/code/myrepo --resume "$(cat $RUN_DIR/session_id)" - <<< 'Tests still fail: ...'

# Fork a session to explore a different fix
./scripts/run-opencode.sh -C ~/code/myrepo --resume "$(cat $RUN_DIR/session_id)" --fork - <<< 'Try a different approach: ...'

Options

Flag	Default	Notes
-C <dir>	cwd	OpenCode's working directory
-m <model>	config default	provider/model format; must be a real ID from opencode models
-v <variant>	provider default	Reasoning effort, e.g. high, max, minimal
-l <label>	run	Names the run directory and session title
--resume <id>	—	Continue a previous session
--fork	off	Fork the session on resume
--agent <name>	—	Use a specific opencode agent
-f <file>	—	Attach file(s) to the prompt; repeatable
--no-auto	off	Require permission approval instead of auto-approving

Environment: OPENCODE_ROUTER_MODEL, OPENCODE_ROUTER_VARIANT, OPENCODE_ROUTER_RUNS (run-directory root, default $TMPDIR/opencode-router).

Run artifacts

$RUN_DIR/
  prompt.md          the brief as sent
  events.jsonl       NDJSON event stream
  last_message.txt   opencode's final report, reconstructed from the stream
  meta.json          terminal event (end / error)
  session_id         for --resume
  stderr.log         diagnostics when a run exits non-zero
  raw.log            unprocessed output

Caveats

• Runs use --auto by default — opencode is not asked to confirm tool calls. Use --no-auto if you need approval gates.
• Model IDs use provider/model format and vary by configured providers. opencode models is the source of truth.
• Concurrent runs need separate checkouts (git worktree add). Never point two runs at one tree.

License

MIT — see LICENSE.