TODO

Next Up

Finish Squad View polish:
- tighten the left rail and team list
- reduce badge noise in the center header
- keep improving the chat transcript so it feels closer to a real terminal/chat hybrid
Improve supervisor-native recovery:
- make approval and failure flows read more like manager guidance
- keep reducing cases where users need to inspect raw pipeline details
Security Audit polish:
- manage the case where the orchestrator crashes mid-action and auditActionInFlight stays true (stale-lock detection or manual reset)
- decide whether the supervisor should nudge the user if all findings are handled but Deploy hasn't been clicked
- true rollback (e.g. git stash) when a scoped fix breaks tests, instead of just reverting the finding status

Host-owned policy service so policy and approvals don't live in agent-writable workspace
Stable approval IDs that survive project rename/copy
Tamper-resistant audit log for command/cwd/agent/decision/timestamp
Allowlisted execution path for common install/build/test ops (reduce raw Bash dependence)

Fast and Strict are the public supported modes today
The optional Security Audit (Agent E) is shipping — read-only OWASP-class pass with severity ranking, user-controlled fix loop, and explicit deploy gate
Sandboxed/isolated execution is not an active roadmap item. The Docker runner code remains for narrow cases but cannot be the default until Claude Code subscription auth in containers is reliable. If you need OS-level isolation, run The Dev Squad in a VM you own.
Manual mode still relies on Claude permission prompts rather than pipeline guardrails