Worktree and branch plan
These are suggested labels, not commands. Review paths and branch names before using them.
| Agent | Branch | Suggested path | Ownership |
|---|
Parallel software planning, without repository access
Turn one goal into isolated workstreams.
Design task dependencies, assign agents, reserve file ownership, estimate the critical path, and rehearse a safe merge order before anyone edits the same branch.
Try /goal, /status, /pause all, /resume agent-2, or /block agent-3 waiting on API schema
5workstreams
11hcritical path
2.1xideal speedup
$4.22estimated agent cost
1ownership conflicts
Dependency canvas
Edges are generated from each task's dependency field.
Goal event stream
Cross-turn status is simulated and kept with the plan.
Agent assignments
Edit ownership, dependencies, mode, model profile, and status.
Isolation is only the beginning
Plan the handoff before the merge.
A worktree prevents accidental edits on one working directory. It does not resolve shared files, hidden dependencies, stale context, or an unsafe merge sequence. Make those contracts explicit here.
File ownership collision detector
Exact path matches across active tasks are flagged. Directory-level conflicts still require human review.
Merge sequence and quality gates
Dependency order is paired with a review checklist. A green plan is still a proposal, not proof.
Saved plans and comparison
Snapshots are stored in this browser. Compare the latest two to see how decomposition changes risk and elapsed time.
Local snapshots
Latest comparison
Review signal
Start with contracts and shared files. The fastest parallel plan is usually the one that removes ambiguity before dispatch.
A field guide for parallel coding work
Separate directories do not create separate responsibilities.
Good parallelism comes from explicit boundaries, observable progress, and small, reviewable handoffs. Expand each topic for practical guidance.
What worktree isolation does and does not do
A Git worktree gives an agent a separate checkout tied to a branch, reducing working-directory interference. It does not prevent two branches from changing the same file or making incompatible assumptions. Reserve ownership and review diffs before merging.
How to model dependencies
Use dependencies for real prerequisites: schemas before clients, interfaces before implementations, and fixtures before tests. Avoid linking every task to every other task; that creates a serial plan disguised as a graph.
What belongs in a context handoff
Include the objective, constraints, owned files, expected interfaces, test commands, evidence, unresolved questions, and the exact base commit. A concise handoff is more durable than relying on conversational memory.
Why status, pause, and block states matter
Status makes invisible work observable. Pause protects budget when assumptions change. Block records a dependency rather than encouraging speculative edits. Resume only after the blocking contract or decision is explicit.
How to review and merge safely
Merge contracts and low-level dependencies first, then consumers, tests, and documentation. Rebase or update each branch, rerun focused and integration tests, inspect shared files, and keep rollback steps ready.
How to interpret the estimates
Time, token, and cost outputs are scenario math, not predictions. They ignore model variability, repository complexity, CI queueing, reviewer availability, and unexpected failures. Use them to compare plans, not to promise delivery dates.