From forge
Use when a complete plan exists and tasks are ready for implementation. Dispatches fresh agents per task with individual verification. Phase: EXECUTION.
npx claudepluginhub caseyrtalbot/forge --plugin forgeThis skill uses the workspace's default tool permissions.
Orchestrate the implementation of a plan by dispatching subagents for each task, enforcing test-first discipline, and reviewing completed work before advancing. This skill is the execution coordinator -- it does not write code itself.
Verifies tests pass on completed feature branch, presents options to merge locally, create GitHub PR, keep as-is or discard; executes choice and cleans up worktree.
Guides root cause investigation for bugs, test failures, unexpected behavior, performance issues, and build failures before proposing fixes.
Writes implementation plans from specs for multi-step tasks, mapping files and breaking into TDD bite-sized steps before coding.
Orchestrate the implementation of a plan by dispatching subagents for each task, enforcing test-first discipline, and reviewing completed work before advancing. This skill is the execution coordinator -- it does not write code itself.
Do NOT implement tasks in the orchestrating context. Dispatch a fresh subagent (implementer agent) for each task. The orchestrator's job is coordination, not implementation. This separation prevents context pollution and ensures each task gets focused attention.digraph drive_execution {
"Read plan" [shape=box];
"Initialize workflow state" [shape=box];
"Select next task" [shape=box];
"Check dependencies met" [shape=diamond];
"Dispatch implementer" [shape=box];
"Implementer works\n(prove-first enforced)" [shape=box];
"Dispatch quality-auditor" [shape=box];
"Review passes?" [shape=diamond];
"Mark task complete" [shape=box];
"More tasks?" [shape=diamond];
"Invoke confirm-complete" [shape=doublecircle];
"Read plan" -> "Initialize workflow state";
"Initialize workflow state" -> "Select next task";
"Select next task" -> "Check dependencies met";
"Check dependencies met" -> "Select next task" [label="blocked, try another"];
"Check dependencies met" -> "Dispatch implementer" [label="ready"];
"Dispatch implementer" -> "Implementer works\n(prove-first enforced)";
"Implementer works\n(prove-first enforced)" -> "Dispatch quality-auditor";
"Dispatch quality-auditor" -> "Review passes?";
"Review passes?" -> "Dispatch implementer" [label="no, fix issues"];
"Review passes?" -> "Mark task complete" [label="yes"];
"Mark task complete" -> "More tasks?";
"More tasks?" -> "Select next task" [label="yes"];
"More tasks?" -> "Invoke confirm-complete" [label="no"];
}
.forge/forge-state.json with current phase, total tasks, and task statuses.forge/forge-state.jsonWhen the implementer reports back, handle based on status:
| Status | Meaning | Action |
|---|---|---|
| DONE | Task complete, tests pass | Dispatch quality-auditor for review |
| DONE_WITH_CONCERNS | Complete but has concerns | Review concerns, dispatch quality-auditor, note for distill-lessons |
| NEEDS_CONTEXT | Missing information | Provide context from spec/plan, re-dispatch same task |
| BLOCKED | Cannot proceed | Investigate blocker, update plan if needed, escalate to user if persistent |
When dispatching an implementer, provide exactly this context:
Task: [task number and title]
Description: [from plan]
Files to modify: [from plan]
Verification: [from plan]
Relevant spec sections:
[paste relevant sections from the design spec]
Constraints:
- Write tests BEFORE implementation code (prove-first discipline)
- Run verification command when done
- Commit working code with a descriptive message
"I'll implement this task myself instead of dispatching" The orchestrator must not implement. Mixing coordination and implementation pollutes the orchestrator's context and creates blind spots in review.
"Let me batch all tasks and review at the end" Review after every task, not after all tasks. Catching a design flaw in task 2 is cheap. Catching it after task 8 is expensive.
"The implementer is stuck, I'll take over" If the implementer is stuck, the task description is probably unclear. Improve the task description and re-dispatch. Taking over defeats the isolation benefit.
.forge/forge-state.json shows all tasks with status "complete"When all tasks are complete and individually verified, invoke confirm-complete for end-to-end verification.