Levelup Publish Agent Skill

levelup-publish

What this skill does

Compile accepted CDRs into actual artifacts in the team-ai-directives team AI directives and create a draft PR.

It is the implementation phase of the CDR lifecycle:

• Validate accepted CDRs against the signal gate
• Detect cross-CDR conflicts before publishing
• Generate context module files (rules, personas, examples, constitution)
• Generate eval goldenset files (goldset.md + goldset.json) for eval-type CDRs
• Generate skill artifacts (SKILL.md + .skills-entry.json)
• Update .skills.json manifest
• Update CDR.md index in team-ai-directives
• Create a branch, commit, and open a draft PR

This skill does not run until CDRs have been accepted via /levelup-clarify.

When to use

• After /levelup-clarify: Accepted CDRs need to be published
• Single skill build: Use --skill <name|CDR-id> to build one skill
• Context modules only: Use --context-only to skip skill generation

When NOT to use

• No accepted CDRs: Run /levelup-clarify first
• Uncommitted changes in team-ai-directives: Clean working tree first
• Discovering patterns: Use /levelup-init or /levelup-specify
• Reviewing CDRs: Use /levelup-clarify

Process

User Input

$ARGUMENTS

You MUST consider the user input before proceeding (if not empty).

Examples of User Input:

• "--ready" — Create ready PR instead of draft
• "--skip-skills" — Don't include skill CDRs
• "--context-only" — Only build context modules
• "--skill CDR-005" — Build only the skill from CDR-005
• "CDR-001 CDR-003" — Only implement specific CDRs
• Empty input: Implement all accepted CDRs as draft PR

Flags

• --ready: Create ready PR instead of draft
• --skip-skills: Skip skill-type CDRs
• --context-only: Build only context modules (skip all skills)
• --skill <name|CDR-id>: Build only one skill from a specific accepted skill CDR

Role & Context

You are acting as a Context Publisher — moving accepted CDRs from local drafts to team-ai-directives.

Your role involves:

• Validating that CDRs are accepted and ready
• Creating context module files from CDR content
• Creating skill artifacts from skill-type CDRs
• Managing Git operations (branch, commit, push, PR)

Outline

• Environment Setup (Phase 0): Resolve paths and list accepted CDRs
• Prerequisites Check (Phase 1): Ensure team-ai-directives is configured and clean
• Signal Gate Validation (Phase 2): Filter CDRs without concrete evidence
• Cross-CDR Conflict Check (Phase 3): Detect duplicate targets and rule conflicts
• Branch Preparation (Phase 4): Create branch in team-ai-directives
• Context Module Generation (Phase 5): Build rules/personas/examples/constitution
• Eval Goldenset Generation (Phase 6): Build goldset.md + goldset.json for eval CDRs
• Skill Generation (Phase 7): Build SKILL.md + .skills-entry.json
• CDR.md Update (Phase 8): Update team AI directives root CDR.md index

10. AGENTS.md Check (Phase 9): Create if missing

11. Commit and PR (Phase 10): Publish changes

12. Summary (Phase 11): Report results

Execution Steps

Phase 0: Environment Setup

Run:

scripts/bash/setup-levelup-publish.sh

Parse JSON for REPO_ROOT, TEAM_AI_DIRECTIVES, CDR_DRAFTS_DIR, ACCEPTED_CDRS, TD_CONFIGURED, TD_IS_GIT, TD_CLEAN.

If the setup script is unavailable or fails, resolve manually:

• REPO_ROOT — walk up from cwd to find .adlc/, or git rev-parse --show-toplevel, or use pwd.
• TEAM_AI_DIRECTIVES — TEAM_AI_DIRECTIVES env var, then .adlc/init-options.json → team_ai_directives, then REPO_ROOT/team-ai-directives.
• CDR_DRAFTS_DIR — REPO_ROOT/.adlc/drafts/cdr
• ACCEPTED_CDRS — grep -l '^### Status: \*\*Accepted\*\*' CDR_DRAFTS_DIR/CDR-*.md and extract IDs.
• TD_IS_GIT — git -C "$TEAM_AI_DIRECTIVES" rev-parse --is-inside-work-tree (exit 0 = true).
• TD_CLEAN — git -C "$TEAM_AI_DIRECTIVES" status --porcelain (empty = clean).

If TD_IS_GIT is false, Phase 10 (branch/commit/PR) cannot run. Offer to git init the team AI directives or write files directly without git.

Phase 1: Prerequisites Check

Verify Team Directives configured:

Team AI directives repository not configured.
Run: team-setup
Or set: export TEAM_AI_DIRECTIVES=/path/to/team-ai-directives

Check Working Tree:

If TD_CLEAN is false:

team-ai-directives has uncommitted changes.
Please commit or stash changes before running /levelup-publish.

Check Accepted CDRs:

If ACCEPTED_CDRS is empty:

No accepted CDRs found.
Run /levelup-clarify to accept CDRs first.

Phase 2: Signal Gate Validation

For each accepted CDR, evaluate it against these four criteria:

• Team-wide applicability: Does this pattern apply to multiple projects/teams, or is it specific to one project? If the CDR's context or evidence only references a single project's internals with no generalizable lesson → SKIP (reason: "project-specific").

• Evidence quality: Does the CDR reference concrete file paths, commit SHAs, or test cases? If the evidence section is empty or vague ("various files", "general practice") → SKIP (reason: "no evidence").

• Uniqueness: Does this duplicate an existing directive in team-ai-directives? Check context_modules/rules/, context_modules/examples/, and CDR.md for overlapping content. If it overlaps → SKIP (reason: "duplicate").

• High value: Is this a genuinely useful pattern, or a nice-to-have minor convenience? If the CDR explicitly states "low value" or "minor convenience", or the pattern is trivial (e.g., "use semicolons") → SKIP (reason: "low value").

Skip CDRs that fail any criterion. Skipped CDRs remain in local drafts.

Report:

## Signal Gate Validation

**Passing**: N | **Skipped**: M

### Skipped CDRs

| CDR | Reason |
|---|---|
| CDR-003 | No evidence |
| CDR-005 | Project-specific |

Phase 3: Cross-CDR Conflict Check

Check for:

• Duplicate Targets: Multiple CDRs targeting the same module path
• Rule Conflicts: Same concern, different implementations
• Unresolved Inconsistencies: CDRs with type "Inconsistency" not marked Resolved

If conflicts found:

Cross-CDR conflicts detected. Resolve via /levelup-clarify before implementing.

Phase 4: Branch Preparation

Create branch in team-ai-directives (skip if TD_IS_GIT=false — git operations are handled in Phase 10):

cd "$TEAM_AI_DIRECTIVES"
git checkout -b "levelup/$(basename "$REPO_ROOT")" main 2>/dev/null || git checkout -b "levelup/$(basename "$REPO_ROOT")"

Phase 5: Context Module Generation

For each accepted non-skill CDR, create/update the target file.

Extracting fields from CDRs: CDRs use single-line field format (### Field: value). Extract values by parsing the line after the ### prefix:

• title — from the ## CDR-NNN: heading
• description — from ### Descriptor: line
• id — from ## CDR-NNN heading (e.g., CDR-001)
• cdr_ref — same as id
• domain — from ### Domain: line (default: general)
• context-type — from ### Context Type: line (lowercased; default: rule)
• created / modified / verified — from ### Date: line (use today's date for modified/verified if not present)
• evidence — from ### Feature Implementation Evidence or ### Evidence section body
• {Content from CDR} — from ### Context section body
• OKF fields (resource, tags, timestamp) are derived: resource = relative path from context type/domain/file, tags = context type, timestamp = ISO 8601 datetime

Rules:

---
type: Rule
title: {title}
description: {description}
resource: ./context_modules/rules/{domain}/{file}.md
tags: [{context-type}]
timestamp: {today}T00:00:00Z
id: {id}
cdr_ref: {cdr_ref}
created: {created}
modified: {modified}
verified: {verified}
age_days: 0
evidence:
{evidence}
---

> ⚠️ **Memory Verification**
> This directive is 0 days old. Before applying:
> - [ ] Pattern still exists in current codebase
> - [ ] Rule is actively followed by team
> - [ ] No conflicting rules introduced

# {Title}

{Content from CDR}

## Source

Contributed from: {project-name}
CDR: {cdr_ref}

Personas and Examples follow similar templates with appropriate type.

Constitution:

• For Constitution Creation CDRs: create context_modules/constitution.md
• For Constitution Amendment CDRs: append to existing constitution

Phase 6: Eval Goldenset Generation

For each accepted eval-type CDR, generate goldenset files in team-ai-directives/evals/.

Extracting fields from eval CDRs: Parse the CDR's single-line fields:

• directive_id — from ### Paired Directive CDR: line (e.g., CDR-001)
• descriptor — from ### Descriptor: line
• pass_cases — from ### Pass Cases section body
• fail_cases — from ### Fail Cases section body
• adversarial_cases — from ### Adversarial Cases section body

Step 1: Create evals directory

mkdir -p "$TEAM_AI_DIRECTIVES/evals/{directive_id}"

Step 2: Write goldset.md

Write {TEAM_AI_DIRECTIVES}/evals/{directive_id}/goldset.md:

---
type: Eval
title: {title from CDR}
description: {descriptor from CDR}
resource: ./evals/{directive_id}/goldset.md
tags: [eval]
timestamp: {today}T00:00:00Z
id: {eval CDR id}
cdr_ref: {eval CDR id}
paired_directive: {directive_id}
created: {date from CDR}
modified: {today}
verified: {today}
age_days: 0
---

# Goldset: {Title}

## Directive Under Test

- **CDR**: {directive_id}
- **Path**: {target module of paired directive CDR}

## Pass Cases

{pass cases from CDR — each with scenario, input, output, why-it-passes}

## Fail Cases

{fail cases from CDR — each with scenario, input, output, why-it-fails, correction}

## Adversarial Cases

{adversarial cases from CDR — each with scenario, expected}

Step 3: Write goldset.json

Write {TEAM_AI_DIRECTIVES}/evals/{directive_id}/goldset.json — machine-readable version for grader consumption:

{
  "id": "{eval CDR id}",
  "paired_directive": "{directive_id}",
  "title": "{title}",
  "description": "{descriptor}",
  "cases": [
    {
      "id": "PASS-001",
      "type": "pass",
      "scenario": "...",
      "input_context": "...",
      "expected_output": "...",
      "actual_output": "...",
      "reason": "..."
    },
    {
      "id": "FAIL-001",
      "type": "fail",
      "scenario": "...",
      "input_context": "...",
      "expected_output": "...",
      "actual_output": "...",
      "reason": "...",
      "correction": "..."
    }
  ]
}

Step 4: Report

Eval goldenset published: evals/{directive_id}/goldset.md
Eval goldenset JSON: evals/{directive_id}/goldset.json

Phase 7: Skill Generation

Skip if --skip-skills or if all skill CDRs excluded.

For skill-type CDRs (or when --skill <name|CDR-id> is specified):

• Generate skills/{name}/SKILL.md:

---
name: {name}
description: {description from CDR}
disable-model-invocation: true
---

# {name}

## What this skill does

{Summary}

## When to use

- {Trigger 1}
- {Trigger 2}

## Steps

1. {Step 1}
2. {Step 2}

## Example

{Minimal example}

## Verification

{How to verify}

## Related

- CDRs: {cdr_ref}

• Generate skills/{name}/.skills-entry.json:

{
  "name": "{name}",
  "description": "{description}",
  "version": "1.0.0",
  "cdr_ref": "{cdr_ref}"
}

• Register in .skills.json:

{
  "local:./skills/{name}": {
    "version": "1.0.0",
    "description": "{description}",
    "categories": ["..."]
  }
}

• Update AGENTS.md Skills section with the new skill.

Phase 8: CDR.md Update

Create/update {TEAM_AI_DIRECTIVES}/CDR.md with accepted CDRs:

# Context Directive Records

## CDR Index

| ID | Target Module | Type | Status | Created | Verified | Age | Descriptor |
|---|---|---|---|---|---|---|---|
| CDR-001 | context_modules/rules/... | Rule | Accepted | ... | ... | 0 | ... |

Phase 9: AGENTS.md Check

If AGENTS.md is missing, create it from a template.

Phase 10: Commit and PR

Verify files created, then follow the git decision tree:

Case A: team-ai-directives IS a git repo (TD_IS_GIT=true)

• Create branch (use main if it exists, otherwise HEAD):

cd "$TEAM_AI_DIRECTIVES"
git checkout -b "levelup/$(basename "$REPO_ROOT")" main 2>/dev/null || git checkout -b "levelup/$(basename "$REPO_ROOT")"

• Commit:

git add -A
git commit -m "Add context modules from $(basename "$REPO_ROOT")

CDRs implemented:
- CDR-001: ...
- CDR-002: ...
"

• Check for a remote:

git remote get-url origin 2>/dev/null

4a. If remote "origin" exists AND gh CLI is available — push and create draft PR:

git push -u origin "levelup/$(basename "$REPO_ROOT")"
gh pr create --draft --title "Add context modules from $(basename "$REPO_ROOT")" --body "..."

4b. If remote "origin" exists but gh is NOT available — push only, tell user to open PR manually:

git push -u origin "levelup/$(basename "$REPO_ROOT")"

Report: "Pushed to origin/levelup/{project-name}. Open a PR manually at your Git host."

4c. If no remote exists — commit locally only. Report: "Committed to local branch levelup/{project-name}. Add a remote (git remote add origin <url>) and push when ready."

Case B: team-ai-directives is NOT a git repo (TD_IS_GIT=false)

Offer the user two options:

• git init + commit — initialize git, create an initial commit with the scaffold, then commit the new context modules:

cd "$TEAM_AI_DIRECTIVES"
git init
git add -A
git commit -m "Initial team-ai-directives scaffold"
git checkout -b "levelup/$(basename "$REPO_ROOT")"
git add -A
git commit -m "Add context modules from $(basename "$REPO_ROOT")"

Then follow steps 3–4 above (remote check).

• Write files only — skip all git operations. Files are written directly to the working tree. Report: "Files written to {TEAM_AI_DIRECTIVES}. Initialize git and commit when ready."

Phase 11: Summary

## LevelUp Implement Summary

**Project**: {project-name}
**Branch**: levelup/{project-name}
**CDRs Implemented**: N
**CDRs Skipped (Signal Gate)**: M

### Artifacts Created

| Type | Count |
|---|---|
| Rules | N |
| Personas | N |
| Examples | N |
| Skills | N |
| Constitution Changes | N |
| Evals | N |

### PR Details

**URL**: {PR-URL}
**Status**: Draft

### Next Steps

1. Review PR
2. Merge when approved
3. Run `/team-repair` after merge to re-index and validate

Key Rules

Implement Only Accepted CDRs

• Status must be Accepted
• Discovered/Proposed/Rejected CDRs are skipped

Signal Gate

• Strict mode: skip CDRs without concrete evidence
• Skipped CDRs remain in local drafts for refinement

Cross-CDR Validation

• Stop on unresolved conflicts
• Require /levelup-clarify to resolve

Memory Engineering

• Published files include YAML frontmatter with created, modified, verified, age_days
• CDR.md index tracks freshness

Context Modules Before CDR.md

• CRITICAL: Create ALL context module and skill files before updating CDR.md
• Do not create CDR.md first and skip the actual modules

Workflow Guidance & Transitions

After /levelup-publish

After PR is merged in team-ai-directives, run /team-repair to:

• Re-index CDR.md, .skills.json, and AGENTS.md
• Run conflict scanning
• Refresh verification timestamps

Complete CDR Lifecycle

/levelup-init or /levelup-specify
    ↓
/levelup-clarify
    ↓
/levelup-publish
    ↓
PR merged
    ↓
/team-repair --validate

Next Steps

After implementation, monitor the PR for review. Once merged, run /team-repair to validate the updated team AI directives.

Verification

• All accepted CDRs passed signal gate or were explicitly skipped
• Context module files created in {TEAM_AI_DIRECTIVES}/context_modules/
• Eval goldenset files created in {TEAM_AI_DIRECTIVES}/evals/ (for eval-type CDRs)
• Skill directories created in {TEAM_AI_DIRECTIVES}/skills/ (unless skipped)
• .skills.json updated with new skills
• CDR.md updated at {TEAM_AI_DIRECTIVES}/CDR.md
• Draft PR created with proper description
• team-ai-directives working tree is clean after commit

Context

$ARGUMENTS