Roadmap¶

What's planned, roughly in priority order. No promises on timelines.

Near term¶

Configurable quality loop -- let repos override the default of three quality-tool attempts.
Parallel audit runs -- run several kinds of audit against the same repo at once from the CLI. The API already runs different audits concurrently.
Cost reports -- every run records its cost in runs/; what's missing is a summary across runs, by repo, pipeline and model.

Untrusted input -- today Developer assumes the repos, issues and PRs it works on are trusted (Decision 014). Working on input from anyone needs:
- repo setup and quality commands run inside a sandbox, like the agent's own commands;
- tighter sandbox limits on reads and network access;
- Claude Code itself running in a container.
Webhook-driven execution -- trigger pipelines from GitHub webhooks instead of manual CLI invocation. The API server (developer serve) provides the HTTP layer; what's missing is webhook ingestion, the event-to-pipeline mapping, and the untrusted-input work above.
Persistent API tasks -- keep tasks across server restarts, with a queue, so the API can run unattended.

Web dashboard -- visibility into pipeline runs, metrics, and status without reading terminal output. The API server provides the backend (GET /tasks, GET /tasks/{id}, GET /repos), and runs/ holds each run's transcript; what's missing is the frontend.
Multi-repo audit campaigns -- run the same audit across all configured repos and aggregate findings.