Changelog

What shipped

Repellox SDK and API version history, newest releases first.

v0.9.0

Context tracking and cross-turn aggregation

  • Added context_id threading: reuse one context ID across turns for cross-turn risk aggregation in the audit log.
  • Context fingerprinting tracks mutations across pipeline nodes and emits context_drift_score in inspection results.
  • Added inspect/batch for multiple items in one HTTP call, up to 50 items per batch.
  • Python SDK now supports repellox.Client.inspect_batch(), reducing latency for RAG pipelines inspecting multiple retrieved chunks.
  • Dashboard audit log export now supports JSON and NDJSON. CSV export was removed; use API.
  • Fixed streaming inspection dropping the last partial chunk in some tokenizer configurations.
v0.8.0

LlamaIndex wrapper and output inspection

  • Added repellox.wrap_llamaindex() for LlamaIndex ReActAgent, QueryEngine, and SubQuestionQueryEngine, same LangChain wrapper pattern.
  • Output inspection now scans responses when inspect_output=True, detecting exfiltration patterns before they reach the caller.
  • Added threat class exfiltration_attempt for output patterns consistent with context window leakage.
  • Configure confidence thresholds per threat class with ThresholdPolicy, rather than globally only.
  • TypeScript SDK preview: install with npm install repellox@preview. Supports OpenAI Agents SDK and raw fetch-based agents; API matches Python SDK.
v0.7.0

Tool authorization and jailbreak detection updates

  • Configure tool authorization policies in the dashboard or SDK, including allowed names, parameter constraints, and destination domains per project.
  • Added threat class unauthorized_tool_call for tools outside the allowlist or calls violating parameter policy.
  • Jailbreak detection gained 14 signatures for encoding obfuscation variants: Base64, Unicode normalization, and homoglyph substitution.
  • p99 inspection latency fell from 4.2ms to 1.9ms through inference kernel optimization; p50 remains 0.8ms.
  • Dashboard threat timeline shows blocked events over time with inspection detail drill-down.
  • Fixed monitor mode incorrectly counting inspections against quota, retroactively for affected accounts.
v0.6.0

LangChain wrapper and public beta

  • Public beta launched. SDK on PyPI as repellox.
  • repellox.wrap() for LangChain AgentExecutor intercepts user inputs and retrieved documents before each model call.
  • Core threat classes: direct_injection, indirect_injection, jailbreak, and context_poisoning.
  • Three modes: strict, monitor, and permissive.
  • Dashboard has real-time audit log, threat breakdown by class, and daily inspection volume chart.
  • Dev tier free with 5k inspections/mo, Protect $149/mo with 500k, and Scale $499/mo with 5M inspections.