Skip to content

/lc-verify

Read-only audit. Checks that astra.yaml, the code, and the materialized results all agree.

Source: claude/lightcone/skills/lc-verify/SKILL.md.

Allowed tools

Read, Glob, Grep,
Bash(astra:*), Bash(lc:*), Bash(python:*), Bash(ls:*),
AskUserQuestion

No Write, no Edit. The skill cannot modify the project.

What it checks (per universe; default baseline)

  1. Spec validation — astra validate astra.yaml. Fix and iterate until clean.
  2. Materialization status — lc status --universe <U>. Every output should be ok. Anything stale, missing, or alias that's not expected gets flagged.
  3. Decision-code alignment — the core value. For every decision in astra.yaml, confirm the code accepts it as a parameter rather than hardcoding the value. Cross-checks astra info --decisions against argparse usage in scripts/.
  4. Results match spec — for every output, verify the result files exist and look well-formed. For type: metric outputs, check that each JSON file parses and contains a {"value": …} entry.

Report format

| Check                    | Status |
|--------------------------|--------|
| Spec validation          | ✓/✗    |
| Materialization (N/N)    | ✓/✗    |
| Decision-code alignment  | ✓/⚠/✗  |
| Results match spec (N/N) | ✓/✗    |

The skill lists each finding with file paths and line numbers, and suggests concrete fixes when something fails.

Hard rules

  • Read-only — never modifies files.
  • One universe at a time.
  • Never skips the decision-code alignment check.
  • Always reads actual result files; never infers from code.
  • /lc-build — fix anything /lc-verify flags.
  • lc verify — the deeper, hash-based audit on the CLI side. They complement each other: the skill checks spec-vs-code-vs-results alignment; the CLI checks data integrity.