Hashline Edits

Hashline is mew’s line-anchored editing system. When the model edits a file, it uses line numbers instead of fragile string matching — and every edit is stamped with a file hash so stale edits fail safely instead of corrupting the wrong content.

What it does

• Line-anchored edits: the model references exact line numbers from read output, not string patterns that might match multiple locations.
• Staleness detection: every edit carries a hash of the file content at the time it was read. If the file changed since then, the edit is rejected — the model must read the file again.
• Block-aware operations: for supported languages (Rust, TypeScript, Python, Go, Markdown), the model can target an entire function or block without counting closing braces.
• In-session recovery: if the file drifted during the session (another edit shifted lines), mew can remap the anchors automatically.

How the model uses it

1. The model calls read on a file. The output includes a [path#hash] header and numbered lines.
2. The model calls edit_hashline with line numbers from the read output and the hash.
3. If the hash matches the current file, the edit applies. If not, the edit is rejected with a clear error.

You don’t need to learn the patch format — the model handles it automatically. If you see a “hash mismatch” error, it means the file changed and the model needs to re-read it.

Disabling hashline

Some models don’t follow line-numbered formats reliably. You can disable hashline for a specific provider in your config:

[providers.my-provider]
shape = "openai"
base_url = "https://api.example.com/v1"
credential_ref = "example"
disable_hashline = true

When disabled, the agent falls back to string-replace edits and full-file writes.

For the patch format syntax and internal architecture, see Hashline Internals.