Consult opencode

Consult OpenCode

Why this exists

OpenCode runs a different model (and provider) than you do — by default the deepseek/deepseek-v4-flash model. When you ask it the same question, agreement is a strong signal you're on solid ground, and disagreement surfaces a blind spot you couldn't see from inside your own reasoning. This is the same principle behind adversarial multi-model debate: divergence between genuinely different reasoning systems is information. A single model agreeing with itself is not.

Use it to triangulate a hard decision, pressure-test a design before building, get a review of a diff or a plan, or break a tie when you're stuck between two interpretations. It is not for trivial lookups or things you're already sure of — each call spends real tokens and a few seconds, so spend it where an independent perspective actually changes the risk.

How to consult OpenCode

1. Write a self-contained prompt — OpenCode sees none of your conversation

This is the most important step. OpenCode starts cold: it does not see this chat, the files you've read, or the decision history. Whatever it needs to reason well, you must put in the prompt. Write it to a file (avoids shell-quoting pain) and include:

• Context — the relevant facts, code, constraints, and the options on the table. Enough that OpenCode can reason, not so much that you bury the question.
• The specific question — one clear ask, with the concrete alternatives if it's a fork.
• A standing instruction to disagree if warranted — e.g. "You are an independent reviewer. Reason from the facts below and give a decisive recommendation. Do not just agree — if you disagree, say so and why." You want its real opinion, so present the context fairly and don't lead it to your conclusion.

Write the prompt to a temp file, e.g. /tmp/opencode-query.md.

2. Run it through the wrapper script

bash <skill-dir>/scripts/ask-opencode.sh /tmp/opencode-query.md

The script encapsulates the exact non-interactive invocation (run subcommand, read-only plan agent, model selection, stdin piping) so it works first time. Read the header comments in scripts/ask-opencode.sh for what each choice does and why it matters. Optional overrides via env var:

OPENCODE_MODEL=anthropic/claude-opus-4-8 bash <skill-dir>/scripts/ask-opencode.sh /tmp/opencode-query.md
# OPENCODE_AGENT=build   # only if OpenCode must write/run; default plan is read-only and right for advice

If you ever need the raw command (no script), it is:

OPENCODE_CONFIG_CONTENT='{"model":"deepseek/deepseek-v4-flash"}' \
  opencode run --agent plan -m deepseek/deepseek-v4-flash < /tmp/opencode-query.md

The model is forced two ways on purpose. -m drives the run session and keeps the banner honest; OPENCODE_CONFIG_CONTENT is a robust, subcommand-agnostic lock that also overrides a target repo's own opencode.json. Both are verified on opencode 1.17.9 to override a repo's pinned model, so the consult genuinely runs on the model you chose — not whatever the repo you happen to be in has configured. (The env var matters most because OpenCode's acp subcommand rejects -m; the run subcommand we use accepts it.)

3. Report back, and surface divergence honestly

Relay OpenCode's verdict, then say plainly where it agrees and where it diverges from your own view — don't quietly fold its answer into yours, and don't quietly ignore it. If OpenCode contradicts a conclusion you'd already reached from evidence, don't just switch sides: name the conflict and reconcile it ("I found X, OpenCode argues Y, here's which constraint breaks the tie"). Cross-model agreement is the product; so is a well-explained disagreement. Clean up the temp prompt file when done.

Interpreting the result

• Both agree → strong signal. State it; it's earned confidence, not just repetition.
• OpenCode disagrees → the valuable case. Weigh its reasoning on the merits. If it has a point you missed, adapt. If you have primary evidence it underweighted, hold your ground and say why. Either way the user sees both sides.
• Treat OpenCode as a peer reviewer, not an oracle. It can be confidently wrong; so can you.

Troubleshooting

• opencode: command not found / wrong binary — verify with opencode --version. The binary installs to ~/.opencode/bin/opencode; make sure it's on your PATH.
• Model not available / auth error — confirm the provider is configured with opencode models (lists available provider/model ids) and opencode auth list (shows authenticated providers). The default model is deepseek/deepseek-v4-flash.
• Wrong model actually ran — confirm with --print-logs and grep the llm runtime selected line (opencode run ... --print-logs 2>&1 | grep 'runtime selected'). Without an override, a repo's opencode.json governs the model; the script's -m + OPENCODE_CONFIG_CONTENT both override that, so the model you pass wins.
• OpenCode refuses to write/run something — that's the read-only plan agent working as intended; it can read repo files in the cwd but not mutate them. Set OPENCODE_AGENT=build only if the consult genuinely requires writing or running commands.
• A read failed / "auto-rejecting" line in the output — OpenCode tried to touch a path outside the cwd, which is an "ask" permission; in non-interactive run mode it auto-rejects and continues rather than hanging. If it actually needed that file, include the content directly in your prompt instead.