task-coding-loop

Anveio 的头像作者 Anveio

通过验证获得信任。在会话开始时调用以建立可验证的检查点(环境、基线、完成)。信任来自关卡,而非声明。

📖 中文导读:通过验证获得信任。在会话开始时调用以建立可验证的检查点(环境、基线、完成)。信任来自关卡,而非声明。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: task-coding-loop description: Earn trust through verification. Invoke at session start to establish verifiable checkpoints (environment, baseline, completion). Trust comes from gates, not claims.

Coding Loop

Trust is earned, not assumed.

Context: Where Coding-Loop Fits

This skill operates within the Execute stage of RSID:

USER-DRIVEN:    Listen → Execute ← YOU ARE HERE → Reflect
AUTONOMOUS:     Ideate → Execute ← YOU ARE HERE → Reflect

After completing Execute (PR merged), invoke rsid skill to Reflect and record learnings.

The Principle

You have autonomy to change code. That autonomy requires accountability:

Claim Gate
"Environment is safe" devcontainer OR explicit host acknowledgment
"Starting state is clean" ./verify.sh --agent passes
"Changes are correct" ./verify.sh --agent passes
"Work is complete" PR merged + main verified

No gate, no trust.

Session Gates

  1. Start: Verify baseline green before writing code
  2. Design: Invoke coding-patterns before creating new packages (see Pre-Design Gate below)
  3. Implement: Verify after significant changes
  4. Pre-PR: Pass hard gates (see below) before creating PR
  5. Review: Spawn tsc-reviewer subagent (MANDATORY - do NOT self-review)
  6. Complete: PR merged (not just created) + main verified post-merge
  7. Reflect: Invoke rsid to record learnings in memory.yaml
  8. Stuck: After 3 failed attempts → escalate, don't loop

Hard Gates (Blocking)

These gates MUST be satisfied before PR creation:

0. Pre-Design Gate (for new packages)

Before designing new contract/port packages, invoke coding-patterns skill and answer:

  • Is this a capability (interface with methods) or data structure (pure types)?
  • If data structure: are operations in port as pure functions (not methods on contract)?
  • Does this follow existing package patterns in the monorepo?

If creating packages without this gate: Design will likely conflate data and operations. Stop, invoke skill, redesign.

1. Commit Atomicity Gate

# Check commit count and size
git log --oneline main..HEAD
git diff --stat main..HEAD
Commits Lines Changed Action Required
1 > 200 SPLIT: Use git rebase -i main per effective-git
Any > 500 total SPLIT: Break into logical units

If gate fails: Invoke effective-git skill and restructure before PR.

2. Subagent Review Gate

Before merge, the tsc-reviewer subagent MUST approve:

Spawn: "Use the tsc-reviewer agent to review this PR"

Loop until:
  - Verdict = APPROVED
  - Blockers = 0

Only then: gh pr merge

If subagent not spawned: Review is incomplete. Do not merge. If blockers remain: Address findings, re-spawn subagent, repeat.

3. Size-Based Skill Invocation

Change Size Required Skills
< 50 lines coding-loop (this)
50-200 lines + tsc-reviewer (agent for review)
> 200 lines + effective-git (restructure first)
> 10 files Consider human review escalation

Commands

./verify.sh --agent    # Minimal output: VERIFY:PASS or structured failure
git status -sb         # Know your state
gh pr diff <n>         # Review before merge (mandatory)

Slash Commands

Use these for accelerated workflows:

Command Purpose
/verify Run full pipeline or specific stage
/verify-lint [package] Quick lint check with optional scope
/lint-and-learn Auto-fix lint + extract lessons + create PR
/tsc-review <pr> Spawn TSC reviewer with auto-merge on APPROVED

Deeper Guidance

  • rsid → outer context, post-merge reflection
  • plan-writing → rigorous plans for distinguished engineers
  • coding-patterns → architecture patterns
  • tsc-reviewer (agent) → mandatory PR review before merge (contains all review criteria)
  • effective-git → commit discipline
  • verification-pipeline → debugging failures
  • devcontainer-sandboxing → environment safety
  • pull-request → PR authoring and review workflow