minhtran3124/agent-harness/subagent-driven-development
- community
Checking that the SKILL.md file is still in the repository.
Use to execute an approved multi-task PLAN.md: waves of isolated implementer subagents with per-task review, separate-session resume, and final delivery/correctness/intent gates before shipping.
Install
npx vskill@latest install minhtran3124/agent-harness/subagent-driven-developmentSKILL.md
Subagent-Driven Development
Execute a valid plan with isolated implementation and review contexts. A task may advance only after its Verify command and both per-task reviews are green.
Resume first
For resume <slug>, run the resume authority before editing — source tree first, deployed runtime second:
The fallback fires ONLY when the source file is absent — never on a non-zero exit — so a fail-closed exit 3 is not silently re-answered from the gitignored, possibly-stale deployed copy. Do not discard stderr. A default consumer install owns .claude/runtime/resume_decision.py, not a root runtime/; its sibling run_state.py resolves because Python puts the script's own directory on sys.path. Treat ANY non-zero exit as stop, and require schema_version == 1 before acting on action. Branch only on the structured action/reason_code: execute-plan, resume-repair, resume-review-chain, wait, stop, or rebuild.
On execute-plan, first re-run every cursor.checks_to_rerun item (each claimed-complete task's exact Verify) and dispatch NO pending task until those claims verify; a failed check routes to the repair path. Apply required_transition ONLY after the session confirms the external condition it names. Do not reconstruct the FSM from prose or initialize a missing historical run during execution. Read references/resume.md only when the returned action requires a manual transition or repair.
Preflight
Before the first task, read rules/plan-format.md, rules/wave-parallelism.md, and rules/auto-correct-scope.md. Stop unless:
- Every parsed task has Files, Action, Verify, and Done. New contract plans additionally pass
python3 scripts/check_plan_contract.py <PLAN.md>.
- Same-wave task file sets do not overlap and dependencies are ordered across waves.
- Every Verify is one automated, exit-code-checkable command.
- The plan meets the repository's plan threshold.
Ensure branch isolation, set the plan status: active, and transition the durable run to implementing when it exists. A missing durable run remains honestly untracked.
Execute waves
- Before dispatch, generate a deterministic brief using
skills/subagent-driven-development/scripts/task_brief.py; pass its path to the implementer. Fill [LANE] from SUMMARY. The implementer reports inline; the controller writes it verbatim to .harness-state/sdd/report-<task>.md per rules/orchestration.md. Record non-none deltas in SUMMARY. Never paste history or report contents into prompts.
- Generate one explicit
BASE..HEADpackage withreview_package.py(BASE = the commit
before this task's first commit, never the branch base); pass brief/report/package paths to task-reviewer, whose model is selected by the runtime entry binding. One reviewer returns both spec_verdict and quality_verdict using task-reviewer-prompt.md.
cannot_verifygets one focused context or test-runner retry. If still unknown, escalate; never
treat it as a pass. Fix all Critical/Important findings in one dispatch, then re-review both verdicts. Record Minor findings in SUMMARY and the progress ledger; provide their roll-up to the final review chain.
NEEDS_CONTEXTreceives missing context;BLOCKEDis re-dispatched with changed context/model,
split, or escalated when the plan itself is wrong. Repeated verification failure or blast-radius escape is an escalation signal. Between dispatches, tell the user in a sentence which task is starting and what came back (verdicts, blockers); keep detail in SUMMARY and the ledger.
After all tasks pass, transition to verifying when tracked and read references/review-chain.md for the final sequence.
Receipt and ship gate
Write .review-receipt.json at the reviewed HEAD. Handoff requires both a passing python3 scripts/verify_summary.py --check <slug> (including Success Criteria coverage) and a passing receipt with zero blocking findings. Any post-review code commit invalidates the receipt: re-run the affected review; never hand-edit its SHA.
Security
- Security scanPassed, 100 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingskills/subagent-driven-development/SKILL.md
Popularity
- Repository stars
- 7
- 7-day trend score
- 10
Stars and forks are for the whole GitHub repository, not this skill alone. Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/minhtran3124/agent-harness/subagent-driven-development)