Delegate to opencode

Delegate to opencode (CLI)

Run opencode run non-interactively for a second opinion, review, or a delegated edit worker. Model default: -m zai-coding-plan/glm-5.2 (verify the slug with opencode models | grep glm; don't hard-code temperature).

Happy path

1. Isolate and install agents. Copy the bundled agents (they are mode: all, required for direct --agent launches) into .opencode/agents/ of the checkout the run will use — opencode resolves --agent against --dir, so the agent must live there. Read-only review runs in the current checkout:

   slug="opencode-$(date +%Y%m%d-%H%M%S)"
   skill_dir="<directory containing this SKILL.md>"
   run_dir="/tmp/opencode-$slug"
   mkdir -p "$run_dir" .opencode/agents
   cp "$skill_dir/assets/reviewer.md" .opencode/agents/
   

   Edit work gets its own worktree instead (opencode never creates one):

   slug="opencode-$(date +%Y%m%d-%H%M%S)"
   skill_dir="<directory containing this SKILL.md>"
   worktree="../$(basename "$PWD")-$slug"
   git worktree add -b "agent/opencode/$slug" "$worktree" HEAD
   run_dir="$worktree/.agent-runs/$slug"
   mkdir -p "$run_dir" "$worktree/.opencode/agents"
   cp "$skill_dir/assets/editor.md" "$worktree/.opencode/agents/"
   

2. Write the brief to $run_dir/prompt.md: context, exact task, constraints, acceptance criteria, verification commands, required final output. It reaches opencode inline as the positional message — "$(cat prompt.md)". --file cannot carry the prompt (see Gotchas).

3. Launch. Three wrappers are mandatory: < /dev/null (stdin wedge), timeout (no stream timeout), and the output-cap env var (32k clamp):

   OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX=131072 \
   timeout --signal=TERM 2700 opencode run \
     --dir "$worktree" \
     --agent editor \
     -m zai-coding-plan/glm-5.2 \
     --format json \
     --title "$slug" \
     --auto \
     "$(cat "$run_dir/prompt.md")" \
     < /dev/null \
     > "$run_dir/events.jsonl"
   

   Reviewer: same command with --agent reviewer and --dir "$PWD". Stagger concurrent runs past two per machine (see Gotchas).

4. Harvest. Parse the final text event and sessionID from events.jsonl. Exit 0 does not mean an answer — always check the finish reason:

   jq -r 'select(.type=="step_finish") | .part.reason' "$run_dir/events.jsonl" \
     | grep -qx length && echo "TRUNCATED - raise OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX or shorten the brief"
   

   stop is healthy. Hung-run check: events.jsonl still 0 bytes after ~5 min, or mtime stale >10 min → kill and relaunch. Resume with opencode run --continue "..." < /dev/null or opencode run --session <id> "..." < /dev/null; add --fork to branch without mutating the original.

Gotchas

• Missing < /dev/null wedges the run at init — under a harness, stdin is a pipe that never closes and opencode parks on it before creating a session. Symptom: 0-byte events.jsonl from the start, log ends at init with no created id=ses_.... Check this before anything else.
• Silent provider stalls hang forever (no stream timeout; only explicit provider errors retry). The timeout wrapper is what turns a zombie into a clean failure.
• A third concurrent instance has hung at startup (zero log entries, blocked on the shared SQLite/WAL db under ~/.local/share/opencode/) — observed once on 1.17.13, so treat "stagger past two" as a conservative heuristic, not a documented limit.
• Output is clamped to 32k tokens including reasoning — a thinking-heavy brief can burn it all and emit nothing, with exit 0 and reason: "length". Set OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX to the model's real limit.output (models.dev/api.json); it can only raise toward the ceiling. On models without limit.input (glm-5.2 included) it shrinks usable context one-for-one — match the value to the model, don't set one number globally.
• --file never carries the prompt — it only attaches files, a positional message is still required, and the flag greedily swallows positionals that follow it (--file prompt.md "do X" dies with File not found: do X). Put the message BEFORE the flag, or inline with "$(cat prompt.md)".
• mode: subagent agents can't be launched with --agent — opencode silently falls back to the default agent, dropping your permission setup. Use mode: all or primary; verify with opencode agent list.
• --auto auto-approves but explicit deny rules still hold — edit: deny stays safe for reviewers.
• A dead mid-run worker loses the conversation, not the code — its edits are on disk in the worktree. Recover with a continue-brief to a fresh run: read the original prompt + git diff, judge the partial work critically, finish, verify.
• The shared log is ~/.local/share/opencode/log/opencode.log (UTC); grep run=/agent= to tell "never initialized" from "stalled mid-stream".
• Worktrees omit ignored files; copy only explicit prerequisites. If the worker needs uncommitted local changes, apply an explicit patch in the worktree — never checkpoint unrelated user WIP with git add -A.

Not possible

• No native worktree creation — manage git worktrees yourself.
• No stream/idle timeout — external timeout is the only guard.
• No opencode.json key for the output cap — env var only.
• No --format json on opencode agent list (text output).
• No prompt via --file alone — a positional message is always required.
• --dir with --attach is a path on the remote server, not local.

Inline config alternative

Instead of the markdown assets, declare the agents in opencode.json ("$schema": "https://opencode.ai/config.json"): mirror each asset's frontmatter — description, mode, model, permission — as agent.reviewer / agent.editor objects under the top-level agent key.

Evidence, exact failure modes, and the output-cap analysis: reference/gotchas.md.