Local intelligence / observable work
Mission control.
Connecting to run records…
Showing all saved repositories.
Lifetime · all repositories · including demos
Tokens offloaded
—Gross worker usage · recorded at run endAccepted work—No recorded usage yet
Runs with usage— / —Input + output tokens
How this relates to savings
Recorded worker input + output tokens are a proxy for work offloaded from the orchestrator. These are not measured net frontier-token savings: tokenizers and execution paths differ, repeated input context is included, and orchestration/review tokens are not subtracted. All runs with complete recorded usage contribute, including rejected runs; accepted work is shown separately. Missing usage is excluded; providers that omit usage may have recorded zeros. Totals update when usage is written at run end.
Ready when you are
No runs yet.
Run a dirtywork task and its activity will appear here automatically.
dirtywork run --repo /path/to/repo "Your task"Selected run
Loading selected run…
Selected run—
Loading run records…
- Repository
- Not recorded
- Worktree
- Not recorded
- Branch
- Not recorded
Started
Activity
Verify
Review
Verification—
Tokens · in / out— / —
Changed files—
Review time—
Review this run
Add a review note (optional)
Start the timer before inspecting the files and checks below. Accept or Reject saves the review outcome, elapsed review time and optional note. You or your AI agent handle merging, retrying and cleanup separately.
File changes
Select a file to inspect added and removed lines. Live previews update after worker actions.
Verification evidence
Full task
Run timeline
Review history
Recorded separately from worker time
Select an event for its recorded time.
Activity stream
Latest 60 eventsAll repositories · including demos
Offloading fit
Real work only · all repositoriesLook for high acceptance and short reviews. Suggested types can be corrected on each run.
Start timed reviews on real tasks to compare review effort.
| Task type | Acceptance | Median review | Evidence |
|---|
How this is calculated
Demos and benchmarks are excluded. Acceptance rates count attempts with an accept or reject verdict. Unique task descriptions show how much work was repeated. Five unique reviewed and five unique timed task descriptions clear the limited-data label; this is a coverage guide. Review time includes checks and waiting after Start Review. These metrics do not measure net time saved.
From recorded to accepted
All repositories · including demosWorker comparison
All repositories · including demos| Model | Accepted / reviewed | Median worker time |
|---|
Different tasks and sample sizes; descriptive comparison.