Agent Audit: Full Pipeline Verification

Overview

This skill performs a comprehensive audit of the agent pipeline to detect inconsistencies, missing files, untranslated content, and structural drift across all three layers:

source/ + others/  (Spanish/Staging)
resources/languages/en/  (English/Production)
.opencode/agents/  (OpenCode/Deploy)

Run this skill before committing after any agent changes, or periodically to ensure the pipeline is healthy.

The Audit Process

Check 1: Agent Inventory Parity

Goal: Every agent must exist in all three layers.

Procedure:

1. •Scan source/ directories → list of source agents
2. •Scan others/ directories → list of others agents
3. •Combine into master agent list
4. •For each agent, verify existence in:
   • •resources/languages/en/source/<Agent>/AGENTS.md (or en/others/ for third-party)
   • •.opencode/agents/<agent-kebab>.md

Report format:

AGENT INVENTORY PARITY
| Agent           | Spanish | English | OpenCode | Status |
|-----------------|---------|---------|----------|--------|
| Yupi Dupi       | OK      | OK      | OK       | PASS   |
| Carmen Marin    | OK      | MISSING | MISSING  | FAIL   |

Failure actions:

• •Missing English → Run agent-sync skill
• •Missing OpenCode → Run agent-to-opencode skill
• •Missing Spanish → This is the source of truth; something was deleted or never created

Check 2: File Structure Parity

Goal: Each agent must have the same set of files in Spanish and English.

Procedure:

For each agent:

1. •List files in source/<Agent>/ (or others/<Agent>/)
2. •List files in resources/languages/en/source/<Agent>/ (or en/others/)
3. •Compare file lists

Report format:

FILE STRUCTURE PARITY
| Agent      | File       | Spanish | English | Status |
|------------|------------|---------|---------|--------|
| Yupi Dupi  | AGENTS.md  | OK      | OK      | PASS   |
| Yupi Dupi  | SOUL.md    | OK      | OK      | PASS   |
| Yupi Dupi  | README.md  | OK      | OK      | PASS   |
| Diana      | AGENTS.md  | OK      | OK      | PASS   |
| Diana      | README.md  | MISSING | MISSING | WARN   |

Status meanings:

• •PASS - File exists in both languages
• •FAIL - File exists in Spanish but missing in English (needs sync)
• •WARN - File missing in both (not a sync issue, but maybe incomplete agent)
• •ORPHAN - File exists in English but not in Spanish (shouldn't happen)

Check 3: Section Structure Consistency

Goal: Translated files must have the same section structure as originals.

Procedure:

For each AGENTS.md / SOUL.md pair (Spanish vs English):

1. •Extract all H2 headers (## Section) from the Spanish file
2. •Extract all H2 headers from the English file
3. •Map Spanish headers to expected English headers using the translation table
4. •Report any mismatches

Section header mapping:

Report format:

SECTION STRUCTURE
| Agent      | File      | Spanish Sections | English Sections | Status |
|------------|-----------|-----------------|-----------------|--------|
| Yupi Dupi  | SOUL.md   | 6               | 6               | PASS   |
| Carmen     | AGENTS.md | 8               | 8               | PASS   |

Check 4: Content Freshness

Goal: Detect when Spanish source has been modified but English hasn't been updated.

Procedure:

For each file pair, compare:

1. •Number of bullet points in each section
2. •Number of lines in each section
3. •If Spanish has significantly more content → English is stale

Heuristic: If a Spanish section has 2+ more lines than the English equivalent, flag it for review.

Check 5: OpenCode Completeness

Goal: Every OpenCode agent file must have valid frontmatter and content.

Procedure:

For each file in .opencode/agents/:

1. •Parse YAML frontmatter
2. •Verify required fields exist:
   • • description (non-empty, English)
   • • mode (must be subagent)
   • • tools (must have write, edit)
   • • permission.bash (must be set)
   • • color (valid hex or theme color)
3. •Verify content body:
   • • Has at least 3 H2 sections
   • • No Spanish text detected (check for common Spanish words: "Personalidad", "Comportamiento", "Reglas", "Idioma", "Filosofía")
   • • Content length > 500 characters

Report format:

OPENCODE VALIDATION
| File              | description | mode | tools | permission | color | content | Status |
|-------------------|-------------|------|-------|------------|-------|---------|--------|
| yupi-dupi.md      | OK          | OK   | OK    | OK         | OK    | OK      | PASS   |
| carmen-marin.md   | OK          | OK   | OK    | OK         | OK    | OK      | PASS   |

Check 6: Orphan Detection

Goal: Detect files that exist in one layer but not others.

Procedure:

1. •List all agent names from .opencode/agents/ (reverse kebab-case to get display name)
2. •Verify each has a corresponding Spanish source
3. •Flag any OpenCode agent without a Spanish source as an orphan

Also check:

• •English directories without Spanish counterparts
• •Spanish directories not in the known locations (source/ or others/)

Full Audit Summary Template

===================================
AGENT PIPELINE AUDIT REPORT
Date: YYYY-MM-DD
===================================

LAYER COUNTS:
  Spanish (staging):  N agents
  English (production): N agents
  OpenCode (deploy):  N agents

CHECK 1 - INVENTORY PARITY:     PASS/FAIL (N issues)
CHECK 2 - FILE STRUCTURE:       PASS/FAIL (N issues)
CHECK 3 - SECTION STRUCTURE:    PASS/FAIL (N issues)
CHECK 4 - CONTENT FRESHNESS:    PASS/WARN (N stale)
CHECK 5 - OPENCODE VALIDATION:  PASS/FAIL (N issues)
CHECK 6 - ORPHAN DETECTION:     PASS/FAIL (N orphans)

OVERALL: PASS / NEEDS ATTENTION

RECOMMENDED ACTIONS:
  1. Run agent-sync for: <list>
  2. Run agent-to-opencode for: <list>
  3. Review stale translations: <list>
===================================

When to Run

• •Before every commit that touches agent files
• •After running agent-sync to verify it worked
• •After running agent-to-opencode to verify output
• •Periodically (weekly recommended) to catch drift
• •When adding a new agent to verify full pipeline coverage

Key Principles

• •Non-destructive - This skill only reads and reports; it never modifies files
• •Actionable output - Every failure must include which skill to run to fix it
• •Complete coverage - Checks all three layers, not just pairwise
• •Fast - Only reads headers and metadata, not full content comparison (unless Check 4)