Parcha-ai Cascade
Keep one living plan in .cascade/<project>.md or the existing project plan:
- Outcome and final acceptance check.
- Context needed to resume: workspace, constraints, decisions, source links.
- Tasks: ID, deliverable, dependencies, exit check, status, evidence.
- Current owners, blockers, and next actions.
Decompose. Split by verifiable outcomes or subsystems; detail only near-term work. Expand large tasks into children while retaining the parent's acceptance check. Convert chains into dependency graphs by keeping actual prerequisites and removing ordering that serves no dependency.
Execute. Pick tasks whose prerequisites passed. Build, verify, update the plan, repeat while
authorized work remains. Use todo, doing, done, blocked. Failed checks remain unfinished;
change approach when attempts stop producing information. Respect session limits and continue
independent work when a task is blocked.
Parallelize. Use available subagents for independent ready tasks. Give each worker context, owned files or worktree, deliverable, and exit check. Serialize overlapping edits. Workers return changes, evidence, and unresolved issues; the lead owns integration and the plan. Mirror the graph into native tasks when useful; otherwise the file suffices.
Verify. Mark done only when the exit check passes on the actual candidate. Record check, result, and revision or working-tree identity. Test relevant failure cases; establish a baseline for measured improvements. Never weaken acceptance to hide failure. Verify the assembled system against the original outcome after its required tasks pass.
Resume. Save progress at meaningful boundaries and before handoff. Compare the plan with actual files and workers; recheck stale evidence. Continue until the overall acceptance passes or no authorized work can advance. Leave concrete blockers and next actions when stopping.