pi-redactor

Pi extension that redacts sensitive strings from user input, tool results, and context before the LLM sees them.

Installation

pi install npm:pi-redactor

Or from a local path:

pi install /path/to/pi-redactor

How It Works

The extension intercepts messages at three points:

1. input — rewrites your message before it reaches the LLM.
2. tool_result — redacts tool output (file reads, command output, etc.) before it enters the conversation.
3. context — scans the full message history immediately before each LLM call, catching secrets in historical context, assistant echoes, or custom messages.

All configured patterns are applied using a single-pass regex with longest-match-first semantics. The transformed text is sent instead of the original.

The LLM never sees the original sensitive strings.

If redaction fails at runtime (e.g., a malformed pattern), the extension fails closed: user messages are blocked, tool results are replaced with an error, and context is emptied. An error notification is displayed in each case. Disable the redactor with /redact off to bypass.

Commands

Command	Description
/redact	Show help and usage
/redact add <string>	Add pattern; replaces with [REDACTED]
/redact add <string> as <label>	Add pattern with custom replacement label
/redact remove <string>	Remove pattern by its original string
/redact list	Show all active patterns
/redact clear	Remove all patterns (prompts for confirmation)
/redact on	Enable redaction
/redact off	Disable redaction
/redact limit <n>	Set max pattern count (default 100, max 1000)

Parsing Rule

The delimiter ` as ` splits original from label. The last occurrence of ` as ` is used, so:

/redact add John Smith as CEO as [PERSON]

Redacts John Smith as CEO → [PERSON].

Example

/redact add John Smith as [CLIENT]
/redact add acct_12345

Input: "Contact John Smith about acct_12345."

LLM receives: "Contact [CLIENT] about [REDACTED]."

Notification: 🔒 Redacted 2 occurrence(s) from your message

Behavior

• Case-insensitive: john smith matches John Smith.
• Longest match wins: If both secret and secretkey are patterns, the input secretkey matches the longer pattern. Patterns do not apply sequentially.
• Match order is deterministic: Results are identical regardless of the order patterns were added.
• Fail-closed on error: If redaction fails, the content is blocked — it never reaches the LLM unredacted. User messages return handled, tool results are replaced with an error, and context is emptied.
• Disabled on degraded config: If the config file is corrupted or patterns could not be loaded, the redactor starts disabled. Run /redact list to review, then /redact on to re-enable.
• Default limit: 100 patterns. Configurable up to 1000 via /redact limit <n>. Adding a pattern beyond the current limit produces an error.
• Maximum 1000 characters per pattern (original and replacement).
• Stale pattern warnings. At session start, a warning is shown if any patterns were added more than 90 days ago.

Status Bar

State	Display
Enabled, N patterns (N > 0)	🔒 Redactor: N pattern(s)
Enabled, 0 patterns	🔒 Redactor: no patterns
Disabled	🔓 Redactor: off

Configuration

Patterns persist to a platform-specific state directory that is outside of typical version-controlled dotfile paths:

Platform	Path
Linux	~/.local/state/pi-redactor/config.json
macOS	~/.local/state/pi-redactor/config.json
Windows	%LOCALAPPDATA%\pi-redactor\config.json

The XDG_STATE_HOME environment variable is respected on Linux/macOS. On Windows, LOCALAPPDATA is used.

On-Disk Format

The file uses an envelope format with integrity checking:

{
  "version": 1,
  "configVersion": 3,
  "checksum": "<sha256-hex-of-serialized-data>",
  "data": {
    "enabled": true,
    "maxPatterns": 100,
    "patterns": [
      {
        "original": "John Smith",
        "replacement": "[CLIENT]",
        "createdAt": 1710000000000
      }
    ]
  }
}

Field	Description
version	Envelope schema version. Currently 1.
configVersion	Monotonically incrementing counter for concurrency control.
checksum	SHA-256 hex digest of JSON.stringify(data). Used for tamper detection.
data.enabled	Whether redaction is active.
data.maxPatterns	Maximum number of patterns allowed. Default 100, max 1000.
data.patterns[].original	The original string to redact.
data.patterns[].replacement	The replacement label.
data.patterns[].createdAt	Unix timestamp in milliseconds when the pattern was added.

Integrity and Concurrency

• Checksum verification. On load, the checksum is recomputed and compared. A mismatch triggers a tamper warning and disables the redactor until the user reviews.
• Optimistic concurrency control. Each save increments configVersion. If another pi session modified the file since this session last read it, the save fails with an error instructing the user to reload.
• Atomic writes. Saves write to a temporary file and rename, preventing corruption from interrupted writes.
• Degraded config recovery. If the config file is corrupted, unreadable, or has dropped patterns, the extension starts with the redactor disabled and displays a warning. The user must explicitly re-enable after reviewing.

Limitations

• No expiration or automatic cleanup of stored patterns.
• Performance may degrade as pattern count approaches the 1000 ceiling due to regex alternation size. Untested above 50 patterns at maximum character length.
• /redact list displays the raw original strings in the local UI. These are the sensitive values being redacted. Do not use this command where the terminal output may be captured or shared.
• Pattern matching uses toLowerCase() without Unicode normalization. Visually identical strings in different normal forms (NFC vs NFD) are treated as distinct patterns (e.g., é as U+00E9 vs U+0065 U+0301).
• Signed thinking blocks are not redacted. The context handler scans unsigned thinking blocks, but blocks with a thinkingSignature are passed through unchanged. Modifying the thinking text would invalidate the cryptographic signature used by LLM providers for multi-turn thought chain continuity. Secrets that appeared in signed thinking blocks (from turns before a redaction pattern was configured) will persist in the conversation context.
• Redacted thinking blocks are opaque. Blocks with redacted: true are encrypted payloads with no meaningful text to scan.
• ToolCall arguments are not redacted. The context handler does not scan type: "toolCall" block arguments. The corresponding ToolResultMessage content is redacted.
• TOCTOU window in config saves. Concurrent saves within milliseconds from separate processes can silently overwrite each other. This is bounded by the single-process Pi event model and is not expected to occur in normal usage.

Development Disclosure

This package was developed predominantly using an AI coding agent. All code, tests, and documentation were reviewed, validated, and approved by me.

License

MIT

This site is open source. Improve this page.