Skillscope

SkillScope For Codex

Use this skill when the user asks to run /skillscope, capture Codex skill coverage, inspect which skill instructions were followed, or open the SkillScope browser view for the current Codex session.

User Entry

The user entry is the Codex slash prompt:

/skillscope

That slash prompt should call the repository launcher:

npm --prefix  run codex -- --project-cwd "$(pwd)"

Capture Contract

• Identify the Codex rollout JSONL for the active or requested session.
• Prefer the session whose session_meta.payload.cwd matches the current project.
• Identify SKILL.md, CLAUDE.md, AGENTS.md, or *.skill.md files that the trace proves were read or used.
• If no loaded skill can be proven from tool traces, fall back to project skill discovery and mark those files as discovered rather than confirmed loaded.
• Preserve task prompt, result metadata, model, cwd, agent version, session ID, and inferred session segments when available.
• Generate a skillscope.capture.json bundle before running analysis.
• Run SkillScope analysis with the generated bundle, not with separately guessed skill and trace files.
• Open the browser viewer after capture unless SKILLSCOPE_NO_OPEN=1 or --no-open is set.

Commands For Codex To Run

From any user project, the slash prompt should run:

npm --prefix  run codex -- --project-cwd "$(pwd)"

If the active session cannot be inferred, ask for an explicit trace path and run:

npm --prefix  run codex -- \
  --project-cwd "$(pwd)" \
  --trace 

Optional precise inputs:

npm --prefix  run codex -- \
  --project-cwd "$(pwd)" \
  --trace  \
  --skill  \
  --task  \
  --result 

Analysis Boundary

• The current default analyzer is heuristic: constraint extraction, trace normalization, keyword overlap, and structured tool signals.
• Do not describe heuristic findings as AI judgment.
• The launcher is responsible for provenance: which session and which skill artifacts are being compared.
• The browser "启动分析" action can run an agent-native judge after the user selects a skill-use window.
• Agent judge runs must follow skills/skillscope-analyzer/SKILL.md: judge only the provided local artifacts and cite concrete trace event IDs.
• Do not call a separate model API for SkillScope judge. Codex captures should use Codex; Claude Code captures should use Claude Code.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: mathieu0905
• Source: mathieu0905/skilllens
• License: MIT

Install and usage instructions live in the source repository linked above.