minhtran3124/agent-harness/subagent-driven-development

Trust tier 2
  • community
Source check pending

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-development

SKILL.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:

if [ -f runtime/resume_decision.py ]; then python3 runtime/resume_decision.py --slug <slug>; else python3 .claude/runtime/resume_decision.py --slug <slug>; fi

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:

  1. Every parsed task has Files, Action, Verify, and Done. New contract plans additionally pass

python3 scripts/check_plan_contract.py <PLAN.md>.

  1. Same-wave task file sets do not overlap and dependencies are ordered across waves.
  2. Every Verify is one automated, exit-code-checkable command.
  3. 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..HEAD package with review_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_verify gets 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_CONTEXT receives missing context; BLOCKED is 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

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
vskill VERIFIED badge
[![vskill VERIFIED](https://verified-skill.com/api/v1/skills/minhtran3124/agent-harness/subagent-driven-development/badge)](https://verified-skill.com/skills/minhtran3124/agent-harness/subagent-driven-development)

Something wrong with this skill or its page? Tell us.