superlazy-subagents

name	superlazy-subagents
description	Silent subagent-driven plan execution. Use when executing an implementation plan with independent tasks and subagent support is available.

Superlazy Subagents

SUPERLAZY mode: output ≤3 lines (questions / user-actions / ✅❌); minimal code, full quality; do the work, never narrate it.

Loop (per task, no user check-ins between tasks)

1. Extract the task's text to a brief FILE. Dispatch a fresh implementer subagent: one line of project context + brief path + interfaces/decisions from earlier tasks + report-file path. NEVER paste session history or prior-task summaries; never make it read the whole plan.
2. Implementer questions → answer, re-dispatch. BLOCKED → more context, stronger model, or split the task — never blind retry.
3. Review the task: generate diff as a FILE (git diff BASE..HEAD, BASE recorded before dispatch — not HEAD~1), dispatch a reviewer with brief + report + diff paths + verbatim spec constraints. Never pre-judge findings ("don't flag X") in the prompt.
4. Critical/Important findings → one fix subagent with ALL findings, then re-review. Minor → ledger.
5. Append to ledger file (.superlazy/progress.md): Task N: complete (commits X..Y). Ledger survives compaction — trust it + git log over memory; NEVER re-dispatch a task it marks complete.
6. All tasks done → one final whole-branch review (merge-base diff file, most capable model). Findings → ONE fixer for the full list. Then ✅.

Cost rules

• Cheapest model that fits, stated explicitly per dispatch: transcription-from-plan/1-file fixes → cheapest; multi-file integration or prose-spec implementers and reviewers → mid-tier; architecture + final review → most capable. Turn count beats token price — too-cheap models take 2-3× turns.
• Artifacts move as files (brief, report, diff), never pasted text.
• Every subagent prompt embeds the superlazy output contract: write findings to the report file, reply ≤1 line. Verbose subagent replies are billed output.
• Never dispatch implementers in parallel (conflicts). Never skip re-review after fixes.