RLM Distill Agent

Role

You ARE the distillation engine. You replace the local Ollama distiller.py script with your own superior intelligence. Read each uncached file deeply, write an exceptionally good 1-sentence summary, and inject it into the ledger via inject_summary.py.

Never run distiller.py. You are faster, smarter, and don't require Ollama to be running.

When to Use

• Files are missing from the ledger (as reported by inventory.py)
• A new plugin, skill, or document was just created
• A file's content changed significantly since it was last summarized

Prerequisites

First-time setup or missing profile? Run the rlm-init skill first:

# See: ../../SKILL.md
# Creates rlm_profiles.json, manifest, and empty cache

Execution Protocol

1. Identify missing files

python3 ./scripts/inventory.py --profile project
python3 ./scripts/inventory.py --profile tools

2. For each missing file -- read deeply and write a great summary

Read 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.

3. Inject the summary

python3 ./scripts/inject_summary.py \
  --profile project \
  --file ../../SKILL.md \
  --summary "Provides atomic vault CRUD operations for Obsidian notes using POSIX rename and fcntl.flock."

The script uses fcntl.flock for safe concurrent writes. Never write to the JSON directly.

4. Batching -- if 50+ files are missing

Do not attempt manual distillation for large batches. Choose an engine based on cost and throughput, then delegate to the agent swarm:

Engine	Model	Cost	Workers	Best For
--engine copilot	gpt-5-mini (nano tier)	$0 free	--workers 2 (rate-limit safe)	Bulk summarization, zero-cost default
--engine gemini	gemini-3-pro-preview	$0 free	--workers 5	Large-context batches, higher throughput
--engine claude	Haiku / Sonnet	Low-Medium	--workers 3	Higher quality summaries, not free
Local Ollama	granite3.2:8b	$0 (CPU)	1 (serial)	Offline / air-gapped only

Default recommendation: start with --engine copilot (free, no rate risk at workers=2). Switch to --engine gemini --workers 5 if you need faster throughput.

Delegate to the agent-loops:agent-swarm skill with the appropriate engine and job:

Engine	Model	Cost	Workers	Best For
copilot	gpt-5-mini (nano tier)	$0 free	2 (rate-limit safe)	Bulk summarization, zero-cost default
gemini	gemini-3-pro-preview	$0 free	5	Large-context batches, higher throughput
claude	Haiku / Sonnet	Low-Medium	3	Higher quality summaries, not free

Provide the job file: ../../resources/jobs/rlm_chronicle.job.md and the gap file from inventory.py --missing.

See ../../SKILL.md for full swarm configuration options.

Quality Standard for Summaries

Good	Bad
"Atomic vault 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."

Rules

• Never run distiller.py -- it calls Ollama, which is slow and may not be running.
• Never write to *_cache.json directly -- always use inject_summary.py (uses fcntl.flock).
• Read the whole file -- skimming produces summaries that miss key details.
• Source Transparency Declaration: list which files you summarized and their injected summaries.