opencode-result-handling

name	opencode-result-handling
description	How Claude should process and present OpenCode worker results — validate, synthesize, present. Token-efficient.
user-invocable	false

OpenCode Result Handling

Workers return results via SendMessage. Claude validates and presents them — never regenerates from scratch.

---

Core principle

Workers do the heavy analysis. Claude validates and synthesizes, spending ~200 tokens max on commentary. This is where the token savings compound.

---

When you receive a worker result via SendMessage

1. Sanity check — does it actually answer the question asked?
2. File references — spot-check 1-2 if the response cites files/lines you haven't seen
3. Code correctness — are code suggestions syntactically valid for this language?
4. Completeness — anything obvious missing?

If all 4 pass → present with a one-line "✓ validated" note and stop.

If issues found → add a brief "Claude notes:" section at the end, flag what's wrong.

---

Presenting results

• Keep file paths and line numbers exactly as the worker reported them
• Present code snippets as-is unless obviously wrong
• Preserve severity ratings (CRITICAL / HIGH / MEDIUM / LOW)
• Note if a fallback model was used: (fallback: X, primary Y unavailable)

Don't:

• Re-explain what the worker already explained clearly
• Rewrite correct code the worker provided
• Add hedging ("I think", "it seems") — be direct
• Auto-fix issues from a review — ask the user which ones to fix first

---

If the worker fails or returns empty

• Say so briefly: "OpenCode worker failed to return a result."
• Do NOT generate a substitute answer to fill the gap
• Suggest: try again, or handle this task directly in Claude if it's simple enough

---

After a code review

STOP after presenting. Do not auto-fix. Ask:

"Which of these issues would you like me to fix?"

Wait for the user to choose before touching any files.

---

Validation cost budget

Spend at most ~200 tokens on validation. If the worker's output is solid, just say so in one line.