ypollak2/llm-router

v5.2.0 — Audit Remediation, Observability, and Release Automation (2026-04-14)

Added

• Structured routing observability — the hot path now emits structlog events plus optional OpenTelemetry spans for classify_complexity, score_all_models, build_chain, route_and_call, and individual provider_call attempts.
• Versioned classifier prompt assets — the classifier system prompt now lives in src/llm_router/prompts/classifier_v1.txt, making prompt revisions visible in diffs and release history.
• Classifier eval harness (scripts/eval_classifier.py) — ships with a 100-example golden set so prompt or model-chain changes can be measured instead of guessed.
• Automated release workflow (scripts/release.py) — synchronizes package/plugin versions, verifies changelog metadata, builds artifacts, and drives the commit/push/publish/tag flow from one command.

Changed

• Routing fallbacks now refresh live provider pressure per attempt — long fallback walks stop as soon as a provider exhausts its budget, instead of using stale pressure from the start of the chain.
• Hook IPC is now atomic and retry-friendly — route/session files are written with atomic replace semantics, readers retry transient partial JSON, and obvious implementation prompts fast-path straight to llm_code.
• Integration coverage is back in CI — tests/test_integration.py now uses mocked end-to-end coverage that is safe for CI and included in the default pipeline again.
• Spend aggregation is configurable — multi-source spend can now be combined with either max (default, overlap-safe) or sum (LLM_ROUTER_SPEND_AGGREGATION=sum) for independent traffic channels.
• README and release docs are aligned with the current product — the docs now describe the tracing stack, classifier eval flow, dashboard token auth, and release automation.

Fixed

• Dynamic + static fallback chains no longer waste slots on duplicates — repeated model IDs are removed while preserving the intended order.
• Background benchmark refresh lock ownership (benchmarks.py) — the worker now releases the specific lock instance it acquired, preventing the release unlocked lock warning seen in concurrent test/hot-reload scenarios.
• Hook regression blind spots — a larger golden prompt matrix and source-hook-based tests now cover the real shipped hook path instead of stale copies.

Technical Notes

• Full verification baseline for this release is uv run pytest tests/ -q --ignore=tests/test_agno_integration.py and uv run ruff check src/ tests/.
• This release continues the audit-remediation track and bundles the reliability, observability, and release-discipline tasks targeted for v5.2.0.