TaskForge
Orchestrator that turns a request into a task graph, isolates each task in its own Git worktree and runs them in parallel waves, with human review and merge at the end. The planner is read-only, and no agent can publish on its own.
- Context
- Meta-tooling that turns a request into a task DAG and runs agents in parallel waves, one Git worktree per task.
- Pressure
- An agent writing to a shared repository is a risk: a solo publish has to be impossible, not merely unlikely.
- Decision
- Split planning from execution, keep the planner read-only, isolate every task in its own branch and worktree, and leave review and merge to a person.
- Evidence
- Guards for commitless branches and for histories with no common ancestor; process skills with built-in gates — approval before execution, 80% minimum coverage and reversible migrations; four interfaces over one engine, sharing plans on disk.
- Impact
- The thesis stopped being "run agents" and became mapping how the team works, then attaching agents only where they fit.