bot-coordinator: Use opencode for second-opinion reviews and well-defined work

[jmarrero-bot]

Builds on #10 (the skill renderer). Until #10 merges, this PR also shows its commit; review only the last one, daa8414.

All of the coordinator's workers and reviewers were Claude subagents. This lets it hand some work to opencode on a GitHub Copilot model: a different model family for second opinions, and a cheaper hand for work that needs no design choices.

• bot/coordinator/bot-opencode runs one task with opencode run --auto in a given directory, prints the result and logs the whole run under ~/.local/state/bot-opencode/. The model is BOT_OPENCODE_MODEL (github-copilot/gpt-6.1-sol) or --model.
• The coordinator container gets the Copilot login (the jmarrero-bot-copilot-auth podman secret that botd's opencode agent already uses) and the default model.
• Operator's notes: bot/coordinator/operator-notes.md, installed as ~/.config/bot-harness/notes.md. render-skills.py now appends it to the coordinator skill, and it says when to use opencode:
  • Reviews: every forge PR gets an opencode review besides the reviewer subagent, following the reviewer preamble and posted on the PR. When they disagree, the stricter verdict stands until resolved.
  • Well-defined work: dependency bumps, mechanical refactors, a test for a described case, a clear compiler, clippy or CI fix, and docs for a stated change go to opencode with the worker preamble. Design, ambiguous asks, unknown-cause debugging and anything touching sign-off, credentials or the sandbox stay with Claude.
  • opencode's results get the same review as any worker's. It runs in the background so its exit wakes the session, at most two at once. If it fails, the item goes to a Claude worker.

Testing

• bot/test_render.py checks that the notes end the rendered coordinator skill; all bot tests pass.
• On trashcan, inside the coordinator container: bot-opencode --name smoke read install: Add --var-mount-spec and DPS /var discovery (prototype on #2521) jmarrero-forge/bootc#1 with the bot's gh login and summarized it correctly (including that fork CI is off). It exited 0 and logged the run. The coordinator's rendered skill has the notes, and bot-poll is running.

Generated-by: https://github.com/jmarrero-bot#llms

🤖 Generated with Claude Code