rlm-distill-agent — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited rlm-distill-agent (Agent Skill) and scored it 87/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 3 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 3 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
This skill requires Python 3.8+ and standard library only. No external packages needed.
To install this skill's dependencies:
pip-compile ./requirements.in
pip install -r ./requirements.txtSee ./requirements.txt for the dependency lockfile (currently empty — standard library only).
You ARE the distillation engine. Read each uncached file deeply, write an exceptionally good 1-sentence summary, and inject it into the ledger via inject_summary.py.
inventory.py)First-time setup or missing profile? Run the rlm-init skill first:
# See: ../SKILL.md
# Creates rlm_profiles.json, manifest, and empty cachepython ./scripts/inventory.py --profile project
python ./scripts/inventory.py --profile toolsRead the entire file with view_file. Do not skim.
A great RLM summary answers: "What does this file do, what problem does it solve, and what are its key components/functions?" in one dense sentence.
python ./scripts/inject_summary.py \
--profile project \
--file ../SKILL.md \
--summary "Provides atomic file CRUD operations for markdown notes using POSIX rename and fcntl.flock."The script handles atomic writes safely. Never write to the Markdown files manually.
Do not attempt manual distillation for large batches. Choose an engine based on the user's CLI context and cost profile, then delegate to the agent swarm:
CRITICAL: Determine User's CLI Context First! Before blindly using --engine copilot, determine which agent CLI the user is running (Claude Code, GitHub Copilot CLI, or Google Gemini CLI). You can often tell from the terminal process or simply by asking the user which AI CLI they have access to.
| User's CLI Tool | Recommended Engine Flag | Cost Profile | Workers |
|---|---|---|---|
| GitHub Copilot CLI | --engine copilot (gpt-5-mini nano tier) | $0 free | --workers 2 (rate-limit safe) |
| Google Gemini CLI | --engine gemini (gemini-3-flash-preview) | $0 free | --workers 5 (high throughput) |
| Claude Code | --engine claude (Haiku / Sonnet) | Low-Medium | --workers 3 |
Default Protocol: Ask the user: "I noticed we have over 50 files to distill. Do you have access to Copilot CLI or Gemini CLI for zero-cost batch processing, or should I use Claude Code?"
Then, run the swarm job based on their answer. For example, if they use Gemini:
python ./scripts/swarm_run.py --engine gemini --workers 5 --files-from rlm_distill_tasks_project.mdProvide a job file describing the summarization task and the gap file from inventory.py --missing.
See SKILL.md for full swarm configuration options.
| Good | Bad |
|---|---|
| "Atomic file CRUD using POSIX rename + flock, preserving YAML frontmatter via ruamel.yaml." | "This file handles file operations." |
| "3-phase search skill: RLM ledger -> ChromaDB -> grep, escalating from O(1) to exact match." | "Searches for things in the codebase." |
inject_summary.py.~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.