Evan Hu
afcb260554
All plugins are stdlib-only Python with no external dependencies. - auto-summarizer: extractive conversation summary for long context compression, persists per-agent summaries to disk - conversation-logger: JSONL audit logs per agent with ISO 8601 timestamps, auto-creates log directory tree - guardrails: safety filter detecting PII (email, phone, SSN, CC), prompt injection patterns, and credential exposure via regex - keyword-memory: extracts entities (emails, URLs, dates, technical terms like camelCase/snake_case/dotted identifiers) as memories - sentiment-tracker: keyword-based sentiment scoring with intensifiers and negation handling, only injects context for non-neutral sentiment - todo-tracker: detects action items via 7 task patterns, tracks completion, deduplicates, persists per-agent with 20-item FIFO limit
28 lines
1.3 KiB
Markdown
28 lines
1.3 KiB
Markdown
# guardrails
|
|
|
|
Safety filter plugin that detects potentially harmful content patterns in user messages and injects warning memories into agent context. Uses only Python stdlib regex -- no external dependencies.
|
|
|
|
## Detection Categories
|
|
|
|
| Category | Examples | Memory Tag |
|
|
|----------|----------|------------|
|
|
| PII | Email addresses, phone numbers, SSNs, credit card numbers | `[guardrails:pii]` |
|
|
| Prompt injection | "ignore previous instructions", "you are now", "system prompt:" | `[guardrails:injection]` |
|
|
| Credentials | `password=`, `api_key=`, `secret=`, `token=`, PEM private keys | `[guardrails:credential]` |
|
|
|
|
## Hooks
|
|
|
|
| Hook | Script | Description |
|
|
|------|--------|-------------|
|
|
| ingest | `hooks/ingest.py` | Scans user messages for harmful patterns and returns warning memories |
|
|
|
|
## How It Works
|
|
|
|
When a user message arrives, the ingest hook runs all pattern checks against it. For each detected issue a memory is returned with the category tag and a recommendation for the agent (e.g. "avoid echoing PII", "maintain original instructions"). If nothing is detected the plugin returns an empty memories list.
|
|
|
|
All patterns use word boundaries and anchoring to minimise false positives on casual conversation.
|
|
|
|
## Usage
|
|
|
|
Installed automatically when enabled in agent configuration.
|