mempalace.write_sanitizer

Source: mempalace/write_sanitizer.py

write_sanitizer.py — Observation-grade input hygiene for the write path (#40).

Counterpart to query_sanitizer.py (read path). The fork is local-only and the write surface is not adversarial — Claude Code sessions are the only producer — so this layer is deliberately observation-grade:

• Strip control characters (except newline/tab/carriage-return) and NUL bytes.
• Collapse runs of 3+ blank lines down to 2 (cosmetic; preserves paragraph breaks).
• Truncate at MAX_CONTENT_LENGTH (1 MiB) with an explicit flag, so a runaway writer never blocks a save but always leaves a trail.
• Reject empty content — the only hard error, because storing an empty drawer is never useful and the existing sanitize_content already raises here.

Returns a result dict (cleaned, was_sanitized, flags) so callers can record the sanitization signal in metadata or WAL without changing the write contract. Callers that want strict validation should continue to call mempalace.config.sanitize_content afterwards.

Per issue #40 this is flag-grade, not gate-grade — adversarial content doesn't reach the local palace unless the user types it, so we observe and record rather than block.

Functions

sanitize_write_content

def sanitize_write_content(value: str) -> dict

Observation-grade pass over drawer/diary content before write.

Returns: dict with: cleaned (str): the sanitized content (may equal input if no changes) was_sanitized (bool): True if any cleaning was applied flags (list[str]): which sanitizers fired, e.g. ["control_chars_stripped", "whitespace_normalized", "truncated"] original_length (int): length of the input cleaned_length (int): length of the output error (str|None): set if the input is unsalvageable (empty)

On unsalvageable input the function returns cleaned="" with error set; callers should treat that as a write rejection.

sanitize_write_name

def sanitize_write_name(value: str, field_name: str = 'name') -> dict

Observation-grade pass over wing/room/agent names.

Strips control characters and truncates at MAX_NAME_LENGTH. The strict shape check (sanitize_name in config.py) still runs after this and rejects path traversal, bad characters, etc. — this layer only handles control-char hygiene and length capping.

Returns the same dict shape as sanitize_write_content.