Step-by-step supervision
Hooks record agent actions, protect important files, flag scope changes and send feedback while the agent works. The Stop gate checks tests before it lets the agent finish.
REPLUME / ROADMAP
Where Replume is going
A supervisor beside your coding agent: observe the work, intervene while there is time to correct it, then check the result and show the evidence. Statuses describe the current product and work in progress, not a release date.
Today
Model review
A second model reviews the final change alongside code rules; reviewing selected risky steps during the run comes next. Work is under way on using your own Claude or Codex CLI and, later, a Replume-provided model. The default-on final review belongs to the 0.2.2 candidate, not the published 0.2.1.
Isolated checks
Run tests on the committed change in Docker when it is ready, with a separate working copy fallback. The fallback isolates the Git checkout, not the machine.
User-like runtime checks
Run the committed app in Docker, try a CLI command or localhost request, and capture web errors and screenshots. A manual runtime probe exists; reliable automatic checks across project types and scenario-driven behavior are being built.
PR checks
Run verification against the PR base in CI and put an evidence-backed report where the team reviews. A local prototype exists; the full GitHub Actions path is still being validated.
Team dashboard
A shared place to follow tasks, review reports, and see recurring failures across agents and repositories.
Evidence reports
Each session separates PASS, FAIL and INCOMPLETE, with checks, changed paths, intervention history, and measured time and tokens. Sharing and PR comments are opt-in.
HOW WE MEASURE
Same task, two runs
We compare the same task, starting code, agent and model with and without Replume. We keep the prompt and environment as close as possible and record the complete outcome, including failures and incomplete runs.
- Scenarios: run repeatable tasks with stated checks, keep the task-start state and test output, and inspect what changed.
- Manual checks: a person reviews whether the requested behavior actually works, including behavior tests cannot establish.
- False alarms: review every block or steer and mark when Replume interrupted valid work or missed a real problem.
We will publish paired results and the method after repeat runs. The current page has no performance figure.
The goal is one continuous path from the original task to a result a developer can inspect. See how the current flow works.