Ask opencode

ask-opencode

Use this skill to call local OpenCode as an external helper process from an AI agent session.

Default rule

Prefer read-only delegation unless the user explicitly wants OpenCode to edit files.

Command

Run the bundled wrapper by resolving the script path relative to this SKILL.md; do not assume $HOME or any fixed install root:

skill_dir="<directory containing this SKILL.md>"
python3 "$skill_dir/scripts/ask_opencode.py" [options] -- "prompt"

The wrapper calls:

opencode run --format json --dir <dir> ... <prompt>

prints only OpenCode text output, and deletes the created OpenCode session by default so one-shot delegation does not pollute OpenCode history.

History behavior

Default behavior is ephemeral: after extracting output, the wrapper deletes the created OpenCode session with opencode session delete <sessionID> and removes exact session-id residue such as session_diff/<sessionID>.json under OpenCode state/cache roots.

Only pass --keep-session when the user explicitly wants the OpenCode conversation to remain visible in OpenCode history.

On Ctrl-C or timeout, the wrapper terminates the OpenCode child process group and still attempts session cleanup from partial JSON output or the generated unique title.

Model selection

In examples below, skill_dir means the directory containing this SKILL.md. The wrapper supports explicit OpenCode model selection:

python3 "$skill_dir/scripts/ask_opencode.py" \
  --model deepseek/deepseek-v4-flash \
  -- "<prompt>"

Use any model string accepted by opencode run --model, usually provider/model. Examples from the local OpenCode config include:

• deepseek/deepseek-v4-flash for cheap/fast helper work
• deepseek/deepseek-v4-pro for harder review, architecture, or debugging

If the user names a model, always pass --model <provider/model>.

Modes

Read-only diagnosis / review / exploration

Use this for independent opinions, code review, root-cause guesses, repo mapping, and plan critique:

python3 "$skill_dir/scripts/ask_opencode.py" \
  --mode readonly \
  --model deepseek/deepseek-v4-flash \
  --dir "$PWD" \
  -- "只读检查当前仓库：<task>。不要修改文件。输出结论和证据路径。"

Default: no explicit OpenCode agent. The wrapper passes --dangerously-skip-permissions because opencode run is non-interactive; keep the prompt read-only. Pass --agent <primary-agent> only when you know that agent works with opencode run.

Edit mode

Use only when the user explicitly allows OpenCode to modify files. Keep scope narrow:

python3 "$skill_dir/scripts/ask_opencode.py" \
  --mode edit \
  --model deepseek/deepseek-v4-flash \
  --dir "$PWD" \
  -- "只允许修改 <files>。完成 <task>。最后列出改动和验证。"

Default: no explicit OpenCode agent; passes --dangerously-skip-permissions so OpenCode can complete the scoped edit. Pass --agent <primary-agent> only when needed.

Good prompts

• Choose --model explicitly when model quality/cost matters.
• State read-only vs edit clearly.
• Restrict target files or directories.
• Ask for exact file paths and concise evidence.
• For edit mode, tell OpenCode to avoid unrelated cleanup.

Guardrails

• Do not run edit mode for broad or destructive tasks.
• Do not let Codex and OpenCode edit the same files concurrently.
• After edit mode, Codex must inspect git diff and run the relevant verification itself before claiming completion.
• If OpenCode output is empty, rerun with --raw-json only for debugging.
• Do not pass --keep-session for subagent-style one-shot calls unless the user asks to preserve OpenCode history.
• On Ctrl-C/timeout, expect exit code 130/124 and no persistent OpenCode session when cleanup succeeds.