levelup-clarify
Review, accept, reject, or defer Context Directive Records (CDRs) discovered by levelup-init or proposed by levelup-specify. Interactive one-CDR-at-a-time workflow.
tikalk · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add tikalk/adlc-team-skills --skill levelup-clarify技能说明
name: levelup-clarify description: Review, accept, reject, or defer Context Directive Records (CDRs) discovered by levelup-init or proposed by levelup-specify. Interactive one-CDR-at-a-time workflow. disable-model-invocation: true
levelup-clarify
What this skill does
Review pending CDRs (status Discovered or Proposed) and decide their fate: Accepted, Rejected, or Deferred.
This is the quality gate for all contributions to team-ai-directives:
- Validate that patterns are team-wide (not project-specific)
- Check for duplicates against existing team-ai-directives
- Ensure CDRs have clear context, decision, and evidence
- Update CDR statuses in
{REPO_ROOT}/.adlc/drafts/cdr/CDR-{NNN}.md - Regenerate
{REPO_ROOT}/.adlc/drafts/cdr/cdr.mdindex
This is an interactive command. Present exactly one CDR per interaction and wait for user input.
When to use
- After
/levelup-init: Validate brownfield discoveries - After
/levelup-specify: Review proposed feature learnings - After
/team-repairfound conflicts: Resolve conflict CDRs created by repair - Periodic review: Clean up stale pending CDRs
When NOT to use
- No pending CDRs: If no CDRs have status Discovered/Proposed, there is nothing to clarify
- Direct editing: Do not use this skill to bypass the review workflow
- Routine health checks: Use
/team-repairfor team AI directives maintenance
Process
User Input
$ARGUMENTS
You MUST consider the user input before proceeding (if not empty).
Examples of User Input:
"CDR-001 CDR-003"— Focus on specific CDRs"rules"— Clarify only rule-type CDRs"all"— Clarify all pending CDRs- Empty input: Clarify all CDRs with status "Discovered" or "Proposed"
Flags
--all: Clarify all pending CDRs (same as empty input)--type TYPE: Filter by context type (rules, personas, examples, skills, constitution, evals)--limit N: Limit to N clarifications per session (default: 5)--no-evals-gate: Disable the evals regression gate (default: gate is ON)
Role & Context
You are acting as a Context Validator reviewing discovered patterns. Your role involves:
- Validating that patterns are still relevant
- Clarifying scope (team-wide vs project-specific)
- Checking against existing team-ai-directives for overlap
- Refining CDR content through targeted questions
CDR Quality Checklist
Each CDR should have:
- Clear context explaining the pattern
- Explicit decision statement
- Evidence from codebase or feature
- Target module path well-formed
- Status is accurate
- No conflicts with existing directives
- Team-wide applicability
Outline
- Load Pending CDRs (Phase 1): Parse CDR files with status Discovered/Proposed
- Pre-Validation (Phase 2): Skip CDRs missing required sections
- Evals Regression Gate (Phase 2a): Run existing goldensets before accepting (default ON)
- Gap Identification (Phase 3): List clarification needs
- Sequential Clarification (Phase 4): One CDR per interaction
- Update CDRs (Phase 5): Write status and clarification metadata after each decision
- Regenerate Index (Phase 6): Update
cdr.md - Summary (Phase 7): Present results
Execution Steps
Phase 1: Load Pending CDRs
Run setup script:
scripts/bash/setup-levelup-clarify.sh
Read all {REPO_ROOT}/.adlc/drafts/cdr/CDR-*.md files and filter:
- Include:
### Status: **Discovered**or### Status: **Proposed** - Skip:
### Status: **Accepted**,**Rejected**,**Deprecated**
If the setup script is unavailable or fails, resolve manually:
REPO_ROOT— walk up from cwd to find.adlc/, orgit rev-parse --show-toplevel.CDR_DRAFTS_DIR—REPO_ROOT/.adlc/drafts/cdrPENDING_COUNT—grep -l '^### Status: \*\*(Discovered|Proposed)\*\*' CDR_DRAFTS_DIR/CDR-*.md | wc -l
If user specified specific CDR IDs or types, filter accordingly.
If no pending CDRs:
No pending CDRs found.
Run /levelup-init or /levelup-specify to create CDRs first.
Phase 2: Pre-Validation
For each pending CDR, check required sections:
### Context### Decision### Evidence### Target Module
Skip invalid CDRs and report:
## Skipped CDRs
| CDR | Issue | Action |
|---|---|---|
| CDR-XXX | Missing Evidence | Add evidence and re-run clarify |
Phase 2a: Evals Regression Gate
Default: ON. Use --no-evals-gate to disable.
When reviewing an eval CDR or a directive CDR that has a paired eval, run the existing goldensets in team-ai-directives/evals/ to check if accepting this CDR would break existing directive compliance tests.
For eval CDRs: Validate eval quality:
- Every case has a concrete evidence reference (session description or code file:line)
- Pass/fail is binary (no ambiguous cases)
- Eval CDR references its paired directive CDR
- Fail cases have a correction (what should have been done)
For directive CDRs with paired evals: Run regression check:
- Read existing goldensets from
{TEAM_AI_DIRECTIVES}/evals/ - If no existing goldensets → skip gate (no regression possible)
- If goldensets exist for the same directive area:
- Run the existing goldenset cases against the agent via LLM calls
- If accepting this CDR would modify the directive, check if existing eval cases would still pass
- If existing evals would fail with the new directive → mark CDR as
Blocked (Evals)
Blocked CDRs remain in pending status with a note:
### Clarification
- **Date**: [YYYY-MM-DD]
- **Action**: Blocked (Evals)
- **Reason**: Accepting this CDR would break N existing goldenset cases in evals/{directive-id}/
Report:
## Evals Regression Gate
| CDR | Gate Result | Details |
|---|---|---|
| CDR-001 | PASS | No existing goldensets in scope |
| CDR-002 | BLOCKED | 3 existing cases would fail with mod
阅读完整来源 ↗