Guides
Fusion
One request, several agents: plan, build, test and review, with a fix round when the tests fail or the reviewer asks for changes. You see the result once it has been checked.
Run a Fusion
Open the Modes menu
In the composer, click + and choose Fusion. A Fusion pill appears in the composer.
Describe the outcome and press Enter
Each stage appears in the thread as it runs, with the agent that took it and why.
Answer any approvals
Every stage follows your access level and approvals, like any other run.
The stages
| Stage | Who | What it does |
|---|---|---|
| Plan | The best agent for planning | Reads the code and writes a plan. Changes nothing. |
| Build | The best agent for building | Implements the plan in the thread's worktree. Doesn't commit. |
| Test | Your computer | Runs the project's test command. Skipped if none is found. |
| Review | A different agent from Build, if one is available | Reads the diff and approves or asks for changes. |
| Fix | The builder, in the same session | Gets the failing tests or review notes. Then Test and Review run again. |
There are at most two fix rounds.
The test command
ThreadGrid detects common test commands: pnpm test, yarn test, npm test, cargo test, go test ./..., pytest and make test. JavaScript projects get their dependencies installed in the worktree first.
To set it yourself, go to Settings → Projects and fill in the project's Tests field.
How it ends
- Done: the last stage that ran succeeded.
- Needs you: the fix rounds ran out. Read the review notes and continue in the thread.
- Failed: an agent failed and no other agent could take the stage. Resume tries again.
- Interrupted: ThreadGrid quit mid-stage. Resume re-runs that stage.
If an agent errors out on a stage, ThreadGrid retries that stage once on a different agent. Stop cancels the running stage and the Fusion.
Pinning and Capacity Protection
Pin an agent in the composer and every agent stage runs on it, with no retry elsewhere. Otherwise each stage is routed on its own. An agent that Capacity Protection is preserving can still take the stages it's reserved for, planning and review by default.
Fix until green
A lighter option in the same menu: Fix until green builds, runs the tests, and fixes until they pass, for up to four rounds. There's no plan or review stage.
Not in the beta yet. Stages run one after another. Running tests and review at the same time, and starting on a cheaper model and stepping up when checks fail, are planned.
Related: Compare, Orchestrator.