Skip to content
Agents tracked: 486 Releases (7d): 757 Agent status: no reported issuesSince last limit reset: Claude 48d · Codex 10d For AI agents: Markdown Updated Nov 10, 2026
AgentGid
✦ Ask

Headroom changelog: what's new each month

Every stable Headroom release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 7 months covered; the current month updates daily.

November 2026 · so far

2 releases: 0.43.0 → 0.43.1

November brought SDK 2.x compatibility for MCP servers, new plugin support for Claude Code and other platforms, and numerous fixes to memory handling, caching, and request processing.

Highlights

  • MCP servers now port to SDK 2.x for improved compatibility.
  • New Headroom skill plugin available for Claude Code, Codex, Copilot and skills.sh.
  • Streaming chat completions now properly answer headroom_retrieve requests.
  • Model routing savings now visible in Headroom savings display.
  • Fixed critical timeout issue that was resetting the shared worker pool.

New

  • Headroom skill plugin for Claude Code, Codex, Copilot and skills.sh.
  • Model-routing savings tracking in proxy integration.

Improved

  • In-memory order queue now prevents re-stored entries from being evicted.
  • Memory re-indexing now replaces the HNSW vector instead of duplicating.
  • openClaw now accepts 12-character hashes in headroom_retrieve.
  • Model-keyed tokenizer, beacon and litellm caches are now bounded in proxy.
  • HTTP/1.1 client used for ChatGPT streaming and model metadata requests.
  • Relation key reading respects the memory_save schema advertisement.
  • Runtime discovery avoided in ORT workspace configuration.

Fixed

  • Image compression timeout no longer resets shared worker pool affecting concurrent requests.
  • DirectMem0Adapter.search_memories updated to work with current mem0 API.
  • Saved request metrics no longer lose cache optimization fields.
  • Damaged release PR envelopes now normalized before detection.

October 2026

3 releases: 0.40.0 → 0.42.0

October brought major improvements to compression diagnostics, savings tracking, and integrations with external IDEs. The month focused on making Headroom's cost-saving features more visible and actionable across dashboards and APIs.

Action needed
  • Removed CrewAI extra to eliminate chromadb from the lockfile; users relying on CrewAI integration should address this separately.

Highlights

  • Added per-message compression diagnostics and a new live-agent densify mode for lossless, cache-safe compression.
  • Introduced native integrations with Google Antigravity IDE and OpenCode 2.x plugin API.
  • Extended savings metrics across the dashboard, Rust proxy, and per-agent attribution in history checkpoints.
  • Made usage reporting opt-in by reading HEADROOM_LICENSE, and added CO₂ Saved card to the dashboard.
  • Enhanced router with max_output_tokens condition and improved turn hooks for rewriting OpenAI Responses requests.

New

  • Live-agent densify mode for lossless, cache-safe message compression.
  • Per-message compression diagnostics to track compression effectiveness.
  • Native Google Antigravity IDE integration.
  • OpenCode 2.x plugin API support.
  • Max_output_tokens condition added to ModelRoute.
  • Per-request counterfactual savings record.
  • CO₂ Saved card on dashboard and /stats API.
  • Headroom-snip Claude Code mod for animated compression visualization.

Improved

  • Savings metrics now have one canonical home across dashboard views.
  • Rust proxy now includes native spend and savings stats with dedicated dashboard.
  • Per-agent attribution tracking added to history checkpoints and rollups.
  • Turn hooks can now rewrite OpenAI Responses requests on both HTTP and WebSocket.
  • Router's compression cache is now bounded, with remaining data in SQLite.
  • OpenAI whitespace normalization now preserves indented lines.
  • Batch processing preserves live contexts during capacity cleanup.
  • Native Gemini continuation tools preserved during batch operations.

Fixed

  • Fixed gateway /v1/compress endpoint to report the same tool savings as the proxy path.
  • Fixed connectivity probes to route /api/hello requests to Anthropic correctly.
  • Fixed SourceCode and PlainText dispatch arms in live-zone.
  • Fixed native Gemini continuation tools handling in batch operations.
  • Fixed OpenAI whitespace normalization to preserve indented lines.

September 2026

3 releases: 0.38.0 → 0.39.1

September brought substantial improvements to cost tracking and savings visibility, along with new routing and compression capabilities. The month focused on making savings measurable and controllable across caching, routing, and compression.

Highlights

  • Dashboard now displays cost savings for the current session and all-time, with routing savings breakdown.
  • Compression now describes what it dropped rather than just the reduction amount, and supports a hard per-message veto via protect-messages hook.
  • Savings tracking now shows new-input savings basis separately and carries exact cache-read costs in history.
  • Proxy gateway accepts OpenAI Response shapes and gateway turns, improving compatibility.
  • Router can now elide dense machine-generated lines that compressors cannot shrink.

New

  • Cost Saved headline card on dashboard with per-row savings displayed as messages.
  • headroom-cache-ttl offline TTL estimator now available.
  • requiretools route condition for proxy model-router.
  • protectmessages hook to enforce hard per-message compression veto.
  • Cache-aware tool-schema pricing in cost calculations.
  • Lifetime output spend tracking for bill-share rate denominators.
  • Per-call config override support in compress() SDK method.
  • Transformations feed endpoint can omit message bodies for number-only pollers.

Improved

  • Beacon schema upgraded to v2 with routing and compression training signals and gzip transport.
  • Compression reporting now shows what was dropped, not just reduction percentage.
  • Routing statistics integrated into the proxy seam.
  • Host exclusion capability added to opencode transport plugin routing.
  • Savings history rollups carry exact cache-read cost information.
  • Savings display shows new-input basis alongside whole-wire figures.

Fixed

  • TPM limiter in 0.39.0 no longer refuses large-context requests permanently.

August 2026

9 releases: 0.34.0 → 0.37.0

August brought major improvements to compression and cost tracking, expanded model support across multiple providers, and significant reliability enhancements to streaming and proxy handling. The month featured both new capabilities and critical fixes across compression, proxy routing, and authentication.

Action needed
  • --no-ccr flag now disables both client and server-side CCR response handling.

Highlights

  • Session-aware compression sidecar mode with unified compression engine across proxy and sidecar paths.
  • Extended Claude Code support in VS Code and new PHP language support for code compression.
  • Deterministic runtime rollout controls and improved cost savings attribution across all integrations.
  • Enhanced Anthropic integration with proper 1M-context tier support and configurable fallback models.
  • Security improvements addressing multiple assessment findings and validation of upstream connections.

New

  • Session-aware compression sidecar mode (/v1/compress) with usage relay.
  • PHP support in CodeAwareCompressor.
  • Deterministic runtime rollout controls.
  • Extensions can now report cost savings and latency metrics.
  • Configurable fallback model for Claude via HEADROOM1MMODEL.
  • Kimi Code routing through the proxy.
  • Codex analysis failure surface in learning module.
  • Explicit Serena reconciliation for MCP.

Improved

  • Unified savings attribution across stats, perf, metrics, and dashboard.
  • StreamingCCRHandler now works on OpenAI streams.
  • Stream buffering optimized to only buffer when markers are redeemable.
  • Output token counting now based on stream text instead of wire size.
  • Tool-schema savings now attributed per model.
  • Proper handling of Codex additionaltools in proxy responses.
  • Self-limiting session state for compression-cache registry.
  • Configurable Serena pre-index stall budget.

Fixed

  • SSE ping timing on Bedrock streaming path.
  • Hermes toolcall bridge unwrapping in tool name map.
  • Proxy now returns 502 instead of 200 when upstream connect retries exhausted.
  • ChatGPT authentication detection from ID token claims.
  • Cross-turn dedup pointer handling on OpenAI chat streaming.
  • MIME type pinning for vendored static dashboard assets.

July 2026

5 releases: 0.29.0 → 0.33.0

July brought stability improvements across Headroom's proxy and agent systems, alongside new compression capabilities and expanded model support. The month focused on fixing edge cases in content routing, authentication, and memory management while adding Claude 5 pricing and enhanced savings metrics.

Highlights

  • Added lossless content compression for coding and general workloads with configurable compression strategies.
  • Introduced Claude 5 family pricing and updated current Anthropic rates.
  • Fixed critical issues with Copilot authentication token handling and memory adapter behavior.
  • Expanded compression endpoint functionality to work as a gateway or Kong sidecar.
  • Improved savings tracking with per-extension token metrics and tool-schema aggregation.

New

  • Coding and general workload personas with agent savings support.
  • Lossless-excluded content compression for grep, log, and JSON formats.
  • OpenAI Responses output shaping support.
  • Per-extension token savings metrics recording.
  • Copilot subscription backend support for Headroom models.
  • Compress endpoint usable as a gateway or Kong sidecar.

Improved

  • Content router now accepts any real compression level without minimum savings floor.
  • Proxy compression workers capped to CPU count for better resource efficiency.
  • Bedrock cache token usage now surfaced in non-streaming responses.
  • Codex AGENTS.md adapter made additive to prevent memory wiping.
  • Lifetime totals and session scope labels added to Headroom stats output.
  • Custom OpenAI gateway paths now preserved across operations.
  • Windows UTF-8 path handling and CLI shim fallback improved in learn module.

Fixed

  • Anthropic passthrough tool order now preserved.
  • MiniMax-M3 provider prefix and pre-registration handling resolved.
  • Verbosity baselines now aggregated across projects instead of overwriting.
  • Magika backend skipped on x86 CPUs without AVX2 support.
  • ONNX embedding backend gated behind AVX2 to prevent SIGILL errors.
  • TypeScript export duplication and comment displacement in code generation.

June 2026

7 releases: 0.22.4 → 0.28.0

June brought major expansions to Headroom's provider support and dashboard capabilities, with significant improvements to compression, caching, and diagnostic tooling. The month focused on reliability, performance visibility, and better integration with multiple AI platforms.

Action needed
  • Release tag format changed to vX.Y.Z (dropped release-please component prefix).
  • Gunicorn moved to [proxy-prod] extra with Windows guard added.

Highlights

  • Added GitHub Copilot subscription mode integration and OAuth2 support through Headroom.
  • Introduced light mode for dashboard with per-model savings breakdown and cost tracking.
  • Added diagnostic tooling including headroom doctor setup command and update notifications.
  • Expanded provider support to include Vertex AI, Cortex Code, OpenCode, and Mistral Vibe.
  • Enhanced compression safety with circuit breakers and detection of re-served tool results as waste signals.

New

  • GitHub Copilot subscription authentication and deterministic token handoff through Headroom.
  • OAuth2 client-credentials upstream-auth proxy extension for custom gateways.
  • Vertex AI and Cortex Code (Snowflake) proxy routing support.
  • OpenCode first-class support with wrap, learn, and mcp install commands.
  • Dashboard light mode with per-model savings and expected-vs-actual cost charts.
  • Compression extraction with Rust knob exposure and traffic audits.
  • Token throughput measurement (tokens/sec) through the proxy.
  • Perl language support in code-aware compressor.

Improved

  • CLI help text, validation, and exception handling comprehensive improvements.
  • Dashboard now displays compression-vs-cache net impact in Prefix Cache panel.
  • Compression safety rails added: error-output protection, pipeline circuit breaker, and library inflation guard.
  • Prefix cache scoping by workspace to prevent cross-project leaks.
  • Agent usage stats now surface on dashboard.
  • Output-token reduction through verbosity shaper and per-user learning with counterfactual savings.
  • Adversarial-input robustness grid added for compressor evaluation.
  • Policy cache decay optimization from idle time near TTL expiration.

Fixed

  • Corrected tiktoken encoding for unknown GPT-4 model snapshots.
  • Fixed cross-project data leaks in proactive expansion scoping.
  • Restored green lint gate on main branch.
  • Fixed generic endpoint for non-subscription OAuth with Copilot.
  • Resolved wrapper breadth issues across cline, continue, goose, and openhands.
  • Fixed subscription tokens saved data plane wiring.

May 2026

45 releases: 0.20.16 → 0.22.2

May brought major architectural improvements to Headroom's proxy layer, streaming support across multiple cloud providers, and significant performance optimizations. The month focused on consolidating request handling, enhancing compression policies, and expanding multi-cloud capabilities.

Highlights

  • Native support for AWS Bedrock and Google Vertex AI with authentication and streaming.
  • Automatic prompt caching configuration for OpenAI and Anthropic to reduce latency.
  • Per-authentication-mode compression policies for fine-grained control over payload optimization.
  • Memory system isolation so projects can no longer share stored preferences.
  • Streamlined request outcome tracking and compression decision visibility in telemetry.

New

  • Native Bedrock InvokeModel route with SigV4 authentication and binary EventStream support.
  • Native Vertex AI publisher path with Application Default Credentials.
  • Automatic cache control placement for Anthropic requests.
  • Automatic prompt cache key injection for OpenAI requests.
  • Volatile content detector with customer warnings.
  • Cache-bust drift detector telemetry.
  • MCP server auto-registration for Headroom in Claude and Codex.
  • Exclude tools flag and environment variable to selectively disable tools.

Improved

  • Bedrock observability integration with authentication mode support.
  • Tool array and schema key sorting for consistent ordering.
  • Compression latency visibility to identify performance bottlenecks.
  • Kompress inference speed optimization.
  • Codex WebSocket compression via inline PyO3 calls.
  • Request outcome funnel consolidating multiple streaming finalizers.
  • Cache read/write recording on backend-routed streaming requests.
  • Platform compatibility with glibc 2.38 and improved build portability across architectures.

Fixed

  • Duplicate --code-aware flag breaking proxy import.
  • UTF-8 slice panic in Bedrock header preview by using floor char boundary.
  • Codex configuration requiring OpenAI auth and missing base URL injection.
  • SSE usage recovery from truncated final events.
  • Per-project memory isolation preventing cross-project preference bleeding.
  • Regex-based preference extraction and system reminder noise in memory filtering.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Headroom: pricing, features and alternatives · All changelogs