The supported product is a safety-first investigation loop:
capture failure → collect deterministic evidence → explain → approve a repair → verify
The maintained command surface is ask, explain, fix, doctor, ui, config, models, and version. The TUI is limited to container/log inspection and command history.
- Preserve fail-closed tool approval, path safety, credential sanitization, and dry-run guarantees.
- Improve deterministic diagnosis and verification before expanding AI behavior.
- Keep local Ollama and optional cloud providers behind the same provider boundary.
- Add focused tests for safety-critical behavior and provider failure modes.
- Reject features that duplicate mature external tools or require permanent main-TUI surface.
The following are intentionally not part of the product: workflow YAML, runbook storage, long-term memory, PR/review/commit generation, generic code generation, test generation, export subsystems, an MCP binary, automatic infrastructure setup, diff rollback, pipeline visualization, and a multi-purpose infrastructure dashboard.
A new integration must provide evidence-based diagnosis or safe remediation that existing tools do not already offer. It must be optional, isolated from unrelated commands, useful without AI, testable without the external service, and absent from the main TUI until validated.