snipCompact — head + tail trimming

snipCompact is the core compression algorithm inside ashlr__read. It preserves the top and tail of a file, elides the middle with a token-count marker, and returns the result in a single pass — no LLM required.

The algorithm

┌────────────────────────────────────────────────┐
│  HEAD  (first N lines — imports, signatures)   │
│  ...                                           │
│  [⋯ 312 lines elided · ~4,820 tokens saved ⋯] │
│  ...                                           │
│  TAIL  (last M lines — exports, returns)       │
└────────────────────────────────────────────────┘

Default budget: HEAD = 80 lines, TAIL = 40 lines. On code files (.ts, .py, .rs, etc.) line numbers are prepended so the model can cite file:line exactly.

The elision marker is a single line — never a block — so the model always knows the elision happened and how many tokens were saved.

Why head + tail

Most structural information in source files concentrates at two ends:

• Head: imports, type declarations, function signatures, docstrings.
• Tail: default exports, module.exports, summary comments, closing brackets.

The middle — implementation bodies — is what grows unboundedly with file size, and is also what the model least needs to see when orienting or searching for a symbol. If the model needs the middle, it calls ashlr__grep with a targeted query first.

Center truncation (keep the middle) or prefix truncation (keep only the head) both score worse on downstream grep-recall benchmarks. Head + tail is the minimal representation that preserves navigability.

The 2 KB threshold

ashlr__read only activates snipCompact on files ≥ 2 KB (approximately ≥ 500 tokens at the chars/4 heuristic). Below that, the file is returned verbatim — snipCompact's elision marker itself is ~20 tokens, so compressing a 30-line file would return more tokens than the original.

Practical effect: small configs, short utility modules, and stub files are always returned in full. The large-read savings apply to the ≥2 KB slice only. See savings math for the honest breakdown.

LLM summarization (≥ 16 KB)

For files above ~16 KB, ashlr__read escalates beyond snipCompact: it calls the configured LLM summarizer (Anthropic Haiku by default; local ONNX/Ollama fallback) and returns a structured prose summary with section headers instead of raw truncated text. The confidence badge in the output ([ashlr confidence: high/medium/low]) reflects summarization quality.

Pass bypassSummary: true to skip LLM summarization and get snipCompact output regardless of file size. Useful when you need raw text for regex matching or diffing.

Savings data

The cross-repo, cross-language headline (TypeScript + Python + Rust) is −57%. Large-file reads remain the highest-savings case, while small files are returned in full. See savings math for methodology.

Related

• ashlr__read — the tool that uses snipCompact
• Savings math — full accounting model
• /ashlr-benchmark — run snipCompact against your own repo