Routines

Code review routines

Review a git branch or hunt for bugs using the same graph-aware context as chat, run as background jobs.

Experimental demo — not part of the dokkai core. Core = the retrieval pipeline, MCP server, chat and graph. Code review, bug hunt and the playbooks/skills library are all routines; they may be removed in a future release, or promoted to core, depending on adoption.

Routines run as background jobs on the same shared job system as ingestion. Launch one with POST /routines/runs and stream live per-stage progress over the existing job-events SSE endpoint — no separate routines SSE route.

Review progresses diff → context → playbooks → skills → analyze → summarize.
Bug hunt progresses resolve → playbooks → hunt → findings → summarize.

Model resolution — no hardcoded default

A run resolves its model from the launch payload's model/provider, then the active LLM config. If neither supplies a usable provider+model, the launch fails synchronously with 400.

Model recommendation. Review and bug hunt are harder than chat. 3B-class local models are not reliable for them; use qwen2.5-coder:14b or larger — ollama pull qwen2.5-coder:14b.

Findings are kept, not dropped

Each finding carries an anchored boolean. When the model's line anchor can't be verified against real code, the finding is kept with anchored: false (treat its line numbers as approximate) — never silently dropped.

Routines UI

The Frontend UI has a dedicated /routines screen (a beta nav badge): a launch card, a run history list with severity pills, a live step timeline, a findings browser with a per-finding diff view, and a library tab for playbooks and skills.

← PreviousInstructions profile & watchdog Next →Code review