v0.2.2

Full Changelog: v0.2.1...v0.2.2

v0.2.1

Headline: Shipping StackIntercept v0.2.1 — A Self-Hosted Financial Firewall for Your AI Stack

If you're running high-concurrency LLM agents or data pipelines, your API bills scale linearly with your traffic. Most of those tokens are spent re-processing identical or highly similar user intents, multi-turn system prompts, and repetitive structural extraction tasks.

StackIntercept v0.2.1 is an open-source, single-binary, self-hosted OpenAI-compatible proxy built from scratch in Rust. It drops directly into your infrastructure (via a simple base_url configuration update) to protect your applications from redundant token consumption and runaway API costs.

Key Capabilities in v0.2.1:

• O(1) Exact Caching Matrix: Blistering fast, memory-mapped key lookup hashes incoming payloads to intercept identical requests instantly.

• Bounded Local Semantic Cache: Runs a local open-source embedding model (bge-small-en-v1.5) directly on the host CPU. Conceptual query hits bypass upstream endpoints in under 4ms.

• Safe, Opt-In Budget Routing: Configurable, in-flight prompt assessment seamlessly down-routes simple conversational tokens to more efficient alternative providers (like DeepSeek V3/R1) with zero-overhead SSE stream model-name masking.

• Deterministic Security Gating: Includes deep multi-turn chat history context hashing and custom fallback overrides. Automatically blocks routing optimizations if unsafe parameter mutations—such as tool schemas, complex multi-modal fields, or non-zero temperatures—are verified.

• Hardened Self-Hosted Persistence: Memory spaces are strictly bounded via FIFO and TTL eviction sweeps. State directories flash asynchronously to a binary cache snapshot on disk to handle unexpected process restarts or host updates gracefully.

• Lock-Free Operational Metrics API: Query /admin/metrics for real-time cache distribution tracking, upstream connection diagnostic flags, and precise system telemetries.

New in v0.2.1 (from v0.2.0):

• TOML config file support (stack-intercept.toml)
• Admin routes: /admin/metrics, /admin/cache (GET/DELETE), per-key eviction
• Atomic request-level metrics (exact hits, semantic hits, misses, routing stats)
• Configurable cache sizing and TTL from config/env
• Disk persistence with debounced snapshots
• Secrets masked in startup logs
• 51 Python integration checks, 32 Rust unit tests

Stop burning cash blindly on redundant provider calls. Measure your own cache hit rate and estimated avoided calls before changing application code.

Get the v0.2.1 release binaries and deployment specs: https://github.com/sidsri14/stack-intercept

v0.2.0

Full Changelog: v0.1.1...v0.2.0

v0.1.1

Full Changelog: v0.1.0...v0.1.1

StackIntercept v0.1.0

[0.1.0] - 2026-06-28

Added

• Exact cache (SHA256-based, in-memory)
• Semantic cache (BGE-small-en-v1.5 embeddings, opt-in)
• Dynamic model routing (opt-in, safe-by-default)
• Transparent route headers on all responses
• Tenant-level cache isolation
• Mock-server integration tests (84 checks)
• 60-second demo script
• Latency benchmark script

Safety

• Routing is opt-in (default: off)
• Unsafe features block routing (tools, structured output, temp > 0, multimodal)
• High-reasoning keyword classifier (30+ patterns)
• Explicit model requirement detection
• Cache namespace isolation for routed vs passthrough responses
• Fallback API key leakage prevention
• Bearer prefix normalization