caveman-compress

caveman-compress — shrink memory files

Rewrite long markdown documents in caveman style, byte-preserving everything that must not change.

When to fire

User intent triggers (any of):

• "/caveman-compress <file>"
• "compress AGENTS.md", "shrink this README", "make CLAUDE.md smaller"
• "trim the docs", "reduce the runbook"

Preconditions

• Target file is markdown (.md).
• Caller has read+write access to the file.
• Working tree is clean for the target file (otherwise refuse and ask the user to stash or commit first — accidentally overwriting uncommitted edits would be hard to recover).

Steps

1. Confirm the target. Echo <file>: <bytes> bytes, <lines> lines to the user. If the file is < 200 bytes, refuse — not worth a roundtrip.

2. Backup. Copy the original to <file>.original.md. Refuse if a backup already exists with content different from the current file — that indicates an earlier compression session was not finalized.

3. Invoke the compressor:

   python -m scripts.caveman compress <file> --mode full
   

   Modes: lite, full (default), ultra. The compressor calls the LLM with the caveman ruleset and a strict preservation contract.

4. Validate the output. The compressor enforces, and fails the run if violated:

   • Every fenced code block (``` ... ```) is byte-identical to the original.
   • Every heading (#, ##, ###, …) text and level is identical (order preserved).
   • Every URL (https?://...) appears in the output exactly once and unchanged.
   • Every file path (/-containing tokens) appears in the output exactly once and unchanged.
   • Every bash command line (lines starting with $ or inside bash fenced blocks) is byte-identical.

5. Retry on validation failure. Up to 2 retry passes with targeted patches — only the violated tokens are sent back to the LLM with the instruction to restore them. After 2 retries, restore from <file>.original.md and exit 1.

6. Report. Print:

   <file>: <orig_bytes> → <new_bytes> ({pct}% saved)
   backup: <file>.original.md
   

Output shape

After completion, summarise to the user:

• Original size, compressed size, percent saved.
• Backup path.
• Whether any retries were needed.
• A two-line diff sample (first preserved heading, first prose paragraph compressed).

Trust boundary

• Never overwrite the original without writing a backup first.
• Never run a third retry — bail out and restore.
• If the LLM call errors out (network, quota), do NOT touch the original file. The compressor handles this; this skill only orchestrates and reports.

Guardrails

• Do not compress files that aren't markdown (refuse with Only .md supported).
• Do not compress files inside .git/, node_modules/, __pycache__/, or dist/.
• Do not compress files larger than 100 KB without explicit --force from the user — a single LLM call may not fit them and partial compression risks data loss.

See also

• scripts/caveman/compress.py — the compressor implementation.
• docs/runbooks/caveman-toggle.md — how compression fits the broader caveman feature.