sbs-context-audit

Files (10)

• SKILL.md
• agent-claude.md · reference/
• agent-codex.md · reference/
• __init__.py · scripts/adapters/
• base.py · scripts/adapters/
• claude.py · scripts/adapters/
• codex.py · scripts/adapters/
• analyze.py · scripts/
• capture.py · scripts/
• sweep.py · scripts/

SKILL.md (first lines)

# Context audit

Resolve this skill's directory first; every command below runs from it, and
`--cwd` is the project being measured, not the skill.

## Measure

Choose the measured agent with `--agent` (default: `claude`). Read
`reference/agent-<agent>.md` before interpreting results or suggesting cuts.

```bash
python3 <skill-directory>/scripts/capture.py --agent <agent> \
  --outdir /tmp/ca/base --cwd <project-directory>
python3 <skill-directory>/scripts/analyze.py --agent <agent> \
  --capture /tmp/ca/base
```

Show the report and let the user choose candidates. Use the agent reference to
identify load-bearing tools and valid disabling arguments.

Measure every proposed removal:

```bash
python3 <skill-directory>/scripts/sweep.py --agent <agent> \
  --cwd <project-directory> --workdir /tmp/ca \
  "baseline=" \
  "candidate=<agent-specific arguments>"
```

Everything after `=` is passed to the agent verbatim. Sweeps modify no settings.

Never infer total savings from schema cost alone: removing a tool can also remove
related prompt blocks. Use the measured scenario delta.

## Apply

Ask before writing. The user chooses scope; paths and setting syntax are in the
agent reference. State plainly when a choice affects every project.

Back up the target file outside a temporary directory and apply the change. Run
a fresh capture/analyze without candidate arguments into `/tmp/ca/after`; compare
it with `/tmp/ca/base`.

Describe the result as context headroom and first-call savings, not a latency win.

From balmasi/skill-or-be-skilled, MIT.

More from balmasi/skill-or-be-skilled