build
Implementation engine. Parallel agents in worktrees from plan.
arbazkhan971 · Static checks on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add arbazkhan971/godmode --skill buildAbout this skill
name: build description: Implementation engine. Parallel agents in worktrees from plan.
Activate When
/godmode:build, "build", "implement", "create"- Plan exists with unimplemented tasks
Lessons Integration
Before planning: read .godmode/lessons.md for known build pitfalls.
After session: append lessons about task decomposition or merge strategies that worked/failed.
Input
Read .godmode/plan.yaml. Missing → /godmode:plan first. Skip plan only for changes touching ≤2 files.
Session Resume
On start: check .godmode/session-state.json. If resuming (stop_reason is null), load completed tasks and pending tasks from saved state. Skip completed tasks, continue from next pending.
After each round: atomically save state (completed task IDs, pending task IDs, current round, agent results) to .godmode/session-state.json.
On completion: set stop_reason in the state file.
The Loop
tasks = load_plan()
WHILE tasks remain:
round = pick_tasks_with_no_unmet_deps(tasks, completed)[:5]
FOR each task: Agent("Implement: {task.title}\nFiles: {task.files}\nDone when: {task.done_when}\nStack: {stack}", isolation: "worktree")
FOR each completed agent:
merge worktree → conflict: `git merge --abort`, discard, retry narrower scope → test fail: `/godmode:fix` (max 2) or `git revert HEAD`
VERIFY: build_cmd && lint_cmd && test_cmd → fail: /godmode:fix (max 3). All 3 must pass.
Append `.godmode/build-log.tsv`: round, task_id, agent_time_ms, status(merged/reverted/conflict). Print `Round {N}: {done}/{total}`
Output Format
Print: Build: {completed}/{total} tasks in {rounds} rounds. Merged: {merged}. Reverted: {reverted}. Conflicts: {conflicts}.
Deterministic Merge Order
Agents complete in parallel but MERGE in dispatch order (not completion order). This ensures deterministic state for test runs.
Conflict Resolution Protocol
Merge conflict → discard agent's work immediately (don't resolve). Re-queue task in NEXT round with note: "Conflict with {agent}:{file}. Narrow scope." Log: round, task_id, agent_id, status=conflict, conflicted_file.
Revert Budget
After merge+test failure: max 2 re-attempts.
- Attempt 1: re-queue to fix specific test failures.
- Attempt 2: re-queue with broader scope.
- Attempt 3+: discard task, log "reverted_3x", move to backlog.
Scope Enforcement
IF tests fail after merge: revert agent's work. IF agent needs file outside scope: report NEEDS_CONTEXT.
Agent discovers it needs file X not in task.files: → STOP, report NEEDS_CONTEXT to orchestrator. → Orchestrator amends scope or discards task. → Agent does NOT touch files outside scope.
Keep/Discard Discipline
After EACH agent merge:
KEEP if: build_cmd && lint_cmd && test_cmd all pass after merge
DISCARD if: any check fails OR merge conflict
On discard: git reset --hard HEAD~1. Re-queue task in next round with narrower scope.
Never keep a merge that breaks any guard check.
Learning from Discards
After task DISCARD: check build-failures.tsv for patterns.
If >3 merge_conflict: reduce to 2 agents per round.
If >3 test_regression: add pre-merge test run per agent before merge.
Generalization Rule
Every change must help a CLASS of tasks, not one specific task. "If this test file was deleted, would this code change still be correct?" If NO → rethink.
Reflective Diagnosis on Failure
On 3+ task failures in a round:
Diagnose: read failed diffs and test output.
Write: "Tasks are failing because {shared pattern}.
The plan needs {specific adjustment}."
Apply diagnosis before next round.
Stop Conditions
STOP when FIRST of:
- target_reached: all tasks in plan completed and guards pass
- budget_exhausted: max rounds reached (tasks * 2)
- diminishing_returns: 3 consecutive rounds with 0 tasks merged
- stuck: >5 consecutive discards across rounds
Hard Rules
- Inherits Default Activations per
SKILL.md §14. Principles prelude (Think/Simplicity/Surgical/Goal-driven), pre-commit audit (agents/builder.md step 10a), terse/stdio/tokens, DispatchContext validation, Progressive Disclosure routing, discard cost hierarchy, and coordination patterns all fire by default. Do NOT require explicit flags. - One task per agent — never batch multiple tasks. One commit per task.
- Agent may only modify files listed in task.files. Touching other files = immediate discard.
- After every merge:
build_cmd && lint_cmd && test_cmd. Any non-zero exit = stop round, fix before next merge. - Max 5 agents per round. Dependency order always. No unplanned refactoring, no TODOs, no stubs.
- Every new function gets a test. Every merge gets a TSV row. No exceptions.
- Never ask to continue. Loop autonomously until all tasks complete or stop conditions trigger.
Workflow
- Load plan from
.godmode/plan.yaml— missing plan means run/godmode:planfirst. - Pick up to 5 tasks with no unmet dependencies, dispatch each as a parallel agent in a worktree.
- Merge completed agents in dispatch order — conflict or test failure triggers discard and re-queue.
- Run
build_cmd && lint_cmd && test_cmdafter every merge — pass to keep, fail to revert. - Repeat rounds until all tasks complete or stop conditions trigger.
# Run build and measure output
npm run build 2>&1
npx bundlesize --config bundlesize.config.json
Rules
- One task per agent. Commit message:
feat({module}): {task.title}. One commit per task. - Agent may only modify files listed in task.files. Touching other files = discard.
- After every merge:
build_cmd && lint_cmd && test_cmd. Any non-zero exit = stop round, fix before next merge. - Max 5 agents per round. Dependency order always.
- Build exactly what the plan says. No unplanned refactoring, no TODOs, no stubs, no
// placeholder. - Every new function gets a test. Every merge gets a TSV row. No exceptions.
- Each agent receives: task.title, task.files, task.done_when