Oxy Instance Skill Evaluator

Oxy Instance Skill Evaluator

You evaluate the output of one of the 4 oxy instance-building skills against its rubric and propose specific, actionable edits to the skill's SKILL.md file.

When You Activate

Activate when the user says:

• "Evaluate the [skill name] output"
• "Score the semantic layer / workflow / ETL / app builder results"
• "What needs to be fixed in the [skill] skill?"
• Or when /oxy:eval-and-improve invokes you for a specific skill

Evaluation Workflow

Step 1: Identify What to Evaluate

Ask the user (or read from context) which skill just ran and where the output files are. Expected outputs per skill:

• oxy-semantic-layer: semantics/views/*.view.yml, semantics/topics/*.topic.yml
• oxy-workflow-builder: workflows/*.workflow.yml, agents/*.agent.yml, example_sql/*.sql
• oxy-etl-builder: etl/sources/<provider>/, etl/runners/<provider>_<entity>.py
• oxy-app-builder: apps/*.app.yml or *.app.yml

Step 2: Read the Rubric

Read the rubric for the skill being evaluated:

• Semantic layer: eval/rubrics/semantic-layer.md
• Workflow builder: eval/rubrics/workflow-builder.md
• ETL builder: eval/rubrics/etl-builder.md
• App builder: eval/rubrics/app-builder.md

If eval/rubrics/ is not in the current directory, look for it relative to the skills plugin directory (where this SKILL.md lives: ../../../eval/rubrics/).

Step 3: Inspect Output Files

Read every generated file. For each rubric item, check the file content. Do NOT skip any must-pass items. Verify:

For YAML files: Use Read to inspect structure. Check field names exactly — a field named query: is different from sql_query:.

For Python files: Use Bash to run python -m py_compile <file> and check exit code.

For runtime checks: Use Bash to run the verification commands listed in the rubric:

oxy validate --file=<file>
oxy run <workflow>.workflow.yml --dry-run
oxy build

Step 4: Score Each Item

For each rubric item, record:

• ✅ PASS or ❌ FAIL
• Brief evidence (one line: what you found that confirms pass or fail)

Present results grouped by MUST-PASS first, then SHOULD-PASS:

## Skill: oxy-workflow-builder
### MUST-PASS (X/10 passed)
✅ M1 — tasks: array present at top level (confirmed in workflows/daily_report.workflow.yml:4)
❌ M2 — type: sql found instead of type: execute_sql (line 12)
...

### SHOULD-PASS (X/5 passed)
✅ Q1 — semantic layer checked before SQL (saw grep commands in Claude output)
❌ Q4 — {% set %} syntax used instead of variables: block defaults

Step 5: Identify Root Causes in SKILL.md

For each failed item, identify which part of the skill's SKILL.md caused it. Common patterns:

• Missing documentation → Claude didn't know the correct syntax
• Conflicting instructions → SKILL.md says two different things
• No example provided → Claude fell back to general knowledge
• Wrong example in template → Claude followed the wrong template

To find root causes:

1. Read the skill's current SKILL.md
2. Search for the section related to the failing item
3. Identify the gap or error

Step 6: Propose SKILL.md Edits

For each failed must-pass item, propose a specific edit to the skill's SKILL.md. Format proposals as:

### Fix for M2 (type: execute_sql)
File: skills/oxy-workflow-builder/SKILL.md
Section: "Workflow File Structure"
Issue: Template shows correct structure but no explicit warning about wrong values.
Proposed addition after line ~185:

> ⚠️ CRITICAL: The task type must be `execute_sql` exactly.
> Do NOT use `type: sql`, `type: execute`, or any other variant.
> These will pass `oxy validate` but fail at runtime.

Be specific: include the file, section, and exact text to add or change.

Step 7: Apply Fixes (With Confirmation)

After presenting all proposed fixes, ask the user: "Should I apply these changes to the SKILL.md files now?"

If yes:

1. Apply each change using Edit tool
2. Re-run the failing verification commands to confirm fixes work
3. Report final pass/fail counts

Output Format

Always end with a summary table:

## Summary: oxy-[skill-name]

| Category | Passed | Failed | Total |
|----------|--------|--------|-------|
| Must-Pass | X | Y | Z |
| Should-Pass | X | Y | Z |

Skill ready for production: YES / NO (must-pass all green)

Top 3 fixes needed:
1. [M#] — one line description
2. [M#] — one line description
3. [Q#] — one line description

Retrieval Tool Reference

The correct syntax for the retrieval tool in agent files (commonly missing):

tools:
  - type: execute_sql
    database: <db_name>
  - type: retrieval
    src:
      - example_sql/*.sql
      - workflows/*.workflow.yml
    key_var: OPENAI_API_KEY   # embedding model API key env var

Required fields: type, src Optional: name (defaults to "retrieval"), key_var, embed_model, top_k, db_path

Notes

• --dry-run only works for SQL files (oxy run query.sql --dry-run). For workflow files, --dry-run is silently ignored. The only true verification for workflows is running them. Wrong field names (like type: sql instead of type: execute_sql) only surface at runtime.
• If oxy build or oxy run commands fail due to missing OXY_DATABASE_URL, note this as an environment issue, not a skill failure. Use oxy ≥ 0.5.27 or set the env var.
• Skill activation reliability (whether skill name appeared in Claude output) is informational. Always verify in a fresh session if uncertain whether the skill actually ran.