Add compression for OpenAI Responses API (/v1/responses)

Fix tool_result ordering: no stray user message between tool_calls and tool

Fix zstd streaming decompression, add zstandard dep; bump to 0.5.14

Fix streaming tool calls, compressed request bodies, beacon field names; bump to 0.5.13

Switch telemetry to proxy_telemetry_v2 table, bump to 0.5.12

Bump to 0.5.11: fix cost calculation, Windows path decoding, unused imports

Fix cost calculation: don't mix cache savings into Headroom savings

fix: remove unused imports in test_langgraph.py (ruff F401)

Fix Windows drive letter path decoding in headroom learn (fixes #69)

Merge pull request #60 from gglucass/feat/add-headroom-sdk

Merge pull request #68 from KunalLohtia/feat/langgraph-compress-tool-messages

feat: add LangGraph compress_tool_messages node for ToolMessage compression

docs: add OpenClaw plugin to README, integration guide, index, and TS SDK docs

fix: split Supabase anon key to avoid GitGuardian false positive

Rename plugin to headroom-openclaw (npm scope not available)

Bump to 0.5.10 (0.5.9 had partial PyPI upload)

Bump to 0.5.9

Merge feat/openclaw-plugin: OpenClaw ContextEngine plugin + telemetry fix + cost tracker

OpenClaw plugin fixes, telemetry fix, cost tracker improvement, tokenBudget support

feat: add OpenClaw ContextEngine plugin (@headroom-ai/openclaw)

Revert version bump: table schema fix resolves telemetry for all versions >= 0.5.6

Bump to 0.5.9: fix telemetry beacon (restore auth header, requires table schema update)

fix: restore Authorization header in telemetry beacon

fix: remove Bearer auth header for Supabase publishable key (telemetry broken since 0.5.5)

style: fix ruff format in server.py

Bump to 0.5.8: fix Windows Unicode crash, fix Bedrock Claude 4.6 model IDs

Fix Windows Unicode crash and Bedrock Claude 4.6 model IDs

Merge pull request #62 from matgasser/fix/extract-user-query-closure-scope

fix: move extract_user_query import out of conditional scope

feat: allow overriding proxy telemetry sdk via HEADROOM_SDK

Fix Vercel AI SDK v6 format: use input/output instead of args/result

Make compress() format-aware: auto-detect OpenAI, Anthropic, Vercel, Gemini

Fix RTK binary path on Windows (.exe extension)

Add TypeScript SDK (headroom-ai npm package)

Fix CI test, bump to 0.5.7

Fix API integration test: use tool output not user message for tags

Remove LLMLingua: Kompress is the sole text compressor

Protect workflow XML tags from text compression

Add notebook for langchain-ai/how_to_fix_your_context PR

Use realistic verbose RAG chunks in demo — triggers Kompress within-item

Diversity-aware SmartCrusher: keep unique items, compress text within

Pass user query context through proxy pipeline (9 call sites)

Fix context-blind compression: pass user query to SmartCrusher relevance scorer

Revamp docs site: new theme, comprehensive landing page, complete nav

Fix docs to match implementation: remove false claims, add Strands guide

Fix test failures: update hash test for MD5, handle unicode surrogates

Reduce proxy latency, enrich telemetry, protect Bash output, bump to 0.5.6

Merge pull request #57 from Softizy/fix/forward-ratelimit-headers-streaming

Dedupe telemetry beacon: upsert per session instead of appending

test: verify ratelimit headers are forwarded in streaming responses