rabbit-rewrites

Model-backed rewriting

Detection in this plugin needs no model. scan.py is pure Python and runs on a Pi today. Rewriting is the part that needs one, and this is the path that uses a small local model for it rather than a frontier model.

Paths. {baseDir} below expands to this skill's own directory, the one holding this file. On a host that leaves the placeholder literal, resolve each path relative to that directory.

The design rests on three core principles: targeted chunking, persistent settings reuse, and gated execution.

1. Targeted chunking and context

Rather than sending entire documents (which exceed small model context windows and cause hallucinated edits), the engine chunks flagged prose into focused, contextual units:

• Sentence units with local context: A tell sitting in a sentence is chunked alongside its preceding sentence context. The model receives the target sentence, the specific problem to remove, and surrounding context to resolve pronoun referents, antecedents, and narrative tone without being asked to rewrite the context itself.
• Passage and block units: Structural and density tells (uniformity, tier2-cluster, tier3-density) are chunked at the full paragraph level, sized against the endpoint's input budget.
• Unit merging: Multiple findings landing within the same sentence are merged into a single rewrite unit, preventing colliding edits from invalidating the comparison baseline.

A 10,000-word draft with 40 findings becomes 40 focused 150- to 350-token requests, keeping token budgets tight while giving the model enough context to preserve flow.

2. Settings: Looked at and saved first

Always look for existing settings before prompting or reconfiguring. The engine checks configuration sources in this order:

1. .rabbit-model in the directory beside the document or the current working directory.
2. .rabbit-model at the repository root.
3. Environment variables: $RABBIT_MODEL_BASE_URL, $RABBIT_MODEL_NAME, $RABBIT_MODEL_API_KEY.

Save user configuration to .rabbit-model so settings persist and are automatically reused on every run:

{
  "base_url": "http://127.0.0.1:8080/v1",
  "model": "qwen2.5-3b-instruct",
  "context_tokens": 4096,
  "max_output_tokens": 640,
  "temperature": 0.2,
  "disable_thinking": true
}

For hosted endpoints (e.g. OpenRouter), specify the environment variable holding the key (api_key_env) rather than committing raw keys:

{
  "base_url": "https://openrouter.ai/api/v1",
  "model": "qwen/qwen3-4b-instruct",
  "api_key_env": "OPENROUTER_API_KEY"
}

3. Recommended models and download locations

Small local models (1.5B to 8B parameters) perform reliably when guided by strict prompts and the verification gate. Detailed download sources, commands, and quantization notes live in references/models.md.

Suggested storage paths

• Local project directory: ./models/
• User model cache: ~/.cache/llama.cpp/ or ~/.local/share/models/
• LM Studio / Ollama cache: ~/.cache/lm-studio/models/ or ~/.ollama/models/

Serving an endpoint

# llama-server
llama-server -m ./models/qwen2.5-3b-instruct-q4_k_m.gguf --port 8080 -c 4096 --flash-attn

# Ollama (endpoint at http://127.0.0.1:11434/v1)
ollama serve

4. System prompt and editing rules

The rewriting system prompt blends plain-language simplification with targeted de-slopping:

You are a copy editor. You rewrite one short passage at a time into clear, natural, and plain language.

Remove these machine-writing characteristics:
- Weird subject and verb combinations, and roundabout pseudo-epiphanies.
- Objects performing action verbs (keep actions for humans, groups, or agents; avoid phrases like "this file carries..." or "the module names...").
- Self-praise, AI hedges, and conversational padding.
- Distracting beats and unnecessary complexity.

Absolute rules:
- Keep every number, date, name, file path, URL, and quoted phrase exactly as written.
- Keep all markdown unchanged: code spans, links, list markers, headings, and emphasis.
- Leave fenced code blocks completely unchanged.
- Never use an em dash (—).
- Do not add facts, opinions, examples, or a closing summary sentence.
- Do not change the meaning. Say the same thing in plainer, everyday words.
- Reply ONLY with the rewritten passage and nothing else: no preamble, no commentary, no labels, no surrounding quotes, no code fence.

Style presets

• Default (Plain Language): Everyday vocabulary, active verbs, concise sentences.
• TL;DR: Short summary retaining every key fact, cutting length by half.
• 5-Year-Old (5y): Maximum simplicity, short conversational phrasing.

Thinking / reasoning tokens: Disabled by default ("disable_thinking": true, sending reasoning_effort: "none" and enable_thinking: false). Reasoning blocks consume the output token budget before the model emits the rewrite.

5. The gating mechanism

A small model is never trusted blindly:

1. Fact preservation: Every candidate must pass verify.py validation (numbers, dates, paths, code spans, quotes, and links must match exactly).
2. Rescan validation: A rescan verifies that the flagged tell is gone and no new tells were introduced.
3. Retry loop: Rejected completions are retried with the rejection reason appended. If all attempts fail, the original text is preserved untouched.

Run it

Deterministic fixes first:

python3 {baseDir}/scripts/scan.py draft.md --apply-safe --write

Dry run / plan what would be sent:

python3 {baseDir}/scripts/scan.py draft.md --apply-model --model-plan

Apply model rewrites:

python3 {baseDir}/scripts/scan.py draft.md --apply-model
python3 {baseDir}/scripts/scan.py draft.md --apply-model --write
python3 {baseDir}/scripts/scan.py draft.md --apply-model --stdout | diff draft.md -

--model-limit N stops after N passages. --model-attempts N controls retry attempts per passage.

Benchmarking models

Measure actual pass rates using the fixed 12-passage test battery:

python3 {baseDir}/scripts/bench.py --model-endpoint http://127.0.0.1:8080/v1 --model-name qwen2.5-3b
python3 {baseDir}/scripts/bench.py --repeat 3 --json > qwen2.5-3b.json

What this will not do

• Will not match a voice profile: Matching a personal voice fingerprint requires broader context and a capable frontier model. Use rabbit-writes for voice conversion.
• Will not touch code, tables, headings, or lists: Only prose blocks are processed.
• Will not alter safety band findings: Any concealed instructions or prompt injections are rejected prior to sending.

Where things live