Opencode Sidecar

Defaults:

• host: 127.0.0.1
• port: 4096

2. Send one bounded task:

./opencode-sidecar-run.sh /path/to/repo "Search for auth config loading and summarize the entrypoints."

3. Consume only the returned summary. Do not pull raw intermediate context back into the main agent unless necessary.

Output contract

The runner already constrains the prompt. Expect output to be:

• read-only
• at most 8 concise bullets
• explicit about uncertainty
• path-oriented, with line numbers when possible

Design your task prompt so the answer can stay short.

Good prompt shapes:

• Find the entrypoints for X and summarize them in 5 bullets.
• Read these logs and identify the first failing step.
• Compare these two payload shapes and list material mismatches only.

Bad prompt shapes:

• Fix this bug.
• Refactor this subsystem.
• Investigate everything related to auth.

Session behavior

• The runner uses one reusable OpenCode session per target directory.
• Session IDs are cached in ./state/.
• If a cached session is stale, the runner recreates it automatically.
• ./state/ is only local sidecar session state. It is not repo data.

Files

• ./start-opencode-sidecar.sh starts the headless OpenCode server
• ./opencode-sidecar-run.sh sends bounded read-only tasks to the server

Quality rules

• Keep delegation narrow and concrete.
• Prefer findings, summaries, and mappings over prose.
• Ask for exact paths and line numbers when useful.
• Keep final integration, decisions, and edits in the primary agent.