Today · Manager runs

Manager runs

The in-app manager agent. Give it a goal and an ordered plan of action steps, and RepoOps drives the loop end to end: execute each step through the gated action seam, verify the result, and escalate anything that needs a human.

Where to find it

  • Localhost: /manager-runs.html?repo=<id>
  • API: GET /api/actions/goal-run/list?repo=<id> (start a run with POST /api/actions/goal-run?repo=<id>)
  • Keyboard: ⌘ K then manager runs
  • Sidebar: Today → Manager runs

What it does for you

Turn a goal into a plan, then run it step by step.Write a goal and add ordered steps, each a verb, an input JSON, and an optional confidence from 0 to 1. RepoOps runs them in order and shows live progress: a per-step dot (pending, done, failed, escalated) and a count of how many are done.
Every step goes through the same gate as a hand-run action.There are no new verbs and no side door. A held or low-confidence step routes to the Actions queue and pauses the run. Approve it there, then hit Resume on the run to continue.
Let a model draft the plan, or supply it yourself.Deterministic by default: you write the steps. Opt in and a Claude model drafts the ordered plan from the goal. That incurs a real model cost, is capped per day, and may only use verbs from the planner allowlist your operator set. LLM-planned steps pause for human approval before they act unless an operator turned that off. Either way, every step still runs through the policy, approval, and audit gate.

Where run state lives

Run state is append-only JSONL under the aggregator's own .claude/brain/manager-runs/<repoId>.jsonl, not a mirror, so it survives mirror resets and restarts. Run-level events also land in the audit ledger.

Read more