opencode-cli-runtime

Opencode Runtime

Use this skill only inside the opencode:opencode-review subagent (and, in future plans, opencode:opencode-rescue).

Primary helper for free-form prompt forwarding (the subagent's main mode):

• node "${CLAUDE_PLUGIN_ROOT}/scripts/buddy.mjs" prompt --prompt-file <path> — REQUIRED form for the subagent. The prompt body must be written to a temp file via a quoted-delimiter heredoc; never inline the prompt text into the bash command line. See agents/opencode-review.md for the exact pattern.

Secondary helper for git-diff convenience review (rarely used by the subagent — /opencode:review covers that):

• node "${CLAUDE_PLUGIN_ROOT}/scripts/buddy.mjs" review "<flag-style args>"

Tertiary helper for write-capable task delegation (used by opencode:opencode-run subagent):

• node "${CLAUDE_PLUGIN_ROOT}/scripts/buddy.mjs" run --task-file <path> [--model <provider/model>] [--yolo] [--background] — same --task-file-under-allowed-dir contract as the prompt route. --yolo opts into --dangerously-skip-permissions (opencode writes without prompting); without it, opencode's own prompts apply and may block the subagent (in non-interactive contexts the companion now hard-rejects without --yolo).

When to use which:

• prompt — review / Q&A / no file modifications expected.
• run — explicit task delegation that may modify files.
• review — git-diff convenience for the user-facing slash command (rarely used by subagents).

Other input modes for prompt:

• ... prompt <positional words> — joined by space; only safe when the prompt is known to contain no shell metacharacters. The subagent never uses this; reserved for ad-hoc CLI testing.
• --stdin is explicitly NOT supported in plan 000 (security review deferred — it would enable arbitrary file reads via shell redirection).

Execution rules:

• The review subagent is a forwarder, not an orchestrator. Its only job is to invoke ONE companion subcommand once and return that stdout unchanged.
• Prefer prompt for forwarded review requests from the orchestrator. The orchestrator constructs the full review prompt (including any references to specific files, focus questions, or expected output format).
• Use review only when the orchestrator explicitly says "review the working-tree diff" or "review the branch diff" without supplying its own prompt text.
• Do not call setup from the review subagent — /opencode:setup is a user-facing command.

Output:

• Return the stdout of the companion command verbatim.
• The orchestrator parses the trailing verdict: line for routing decisions; do not reformat or strip it.
• Do not paraphrase, summarize, or add commentary before or after it.
• If the Bash call fails or opencode cannot be invoked, return the script's stderr verbatim.

Trailer behavior — important contract difference between the two routes:

• review route: the prompt explicitly asks the model for a trailer. Missing trailer → verdict: needs-attention (parse error) is always printed.
• prompt route: the orchestrator may or may not have asked for a trailer. If a trailer is present, the script prints both the text and a parsed verdict line. If no trailer is present, the script prints the text only — no verdict line is synthesized. Orchestrators that need a verdict signal must include hybrid-output instructions in the prompt body (typically a fenced JSON block with verdict and blockers).