headroom-proxy-0.39.1-1.1 RPM for noarch

From OpenSuSE Tumbleweed for noarch

Runtime extra matching pip install headroom-ai[proxy]: FastAPI/uvicorn
HTTP+MCP proxy, Magika content routing, ONNX Runtime, and sqlite-vec
(--memory). Stock "headroom proxy" and "headroom wrap" refuse to start
without this set.

Provides

• headroom-proxy

Requires

• headroom = 0.39.1
• python313-fastapi >= 0.100.0
• python313-h2 >= 3.0
• python313-httpx >= 0.24.0
• python313-magika >= 0.6.0
• python313-mcp >= 1.28.1
• python313-onnxruntime >= 1.24.0
• python313-openai >= 2.14.0
• python313-orjson >= 3.9.14
• python313-sqlite-vec >= 0.1.6
• python313-transformers >= 5.5.0
• python313-uvicorn >= 0.23.0
• python313-watchdog >= 4.0.0
• python313-websockets >= 13.0
• python313-zstandard >= 0.20.0
• rpmlib(CompressedFileNames) <= 3.0.4-1
• rpmlib(FileDigests) <= 4.6.0-1
• rpmlib(PayloadFilesHavePrefix) <= 4.0-1
• rpmlib(PayloadIsZstd) <= 5.4.18-1

License

Apache-2.0 AND MPL-2.0

Changelog

* Mon Sep 28 2026 Martin Pluskal <mpluskal@suse.com>
  - Update to 0.39.1:
    * Proxy: the token-per-minute limit is opt-in again (it defaults
      to unlimited), so a stock "headroom proxy" is no longer a 100k
      tokens/minute throttle; --tpm / HEADROOM_TPM still enforce one
    * A request larger than the bucket is admitted once the bucket is
      full and then charged in full, so a large-context session no
      longer gets a 429 it can never satisfy
    * Neither crates/ nor Cargo.lock changed, so the vendored crate
      tree and the licence set derived from it are unchanged
* Sat Sep 26 2026 Martin Pluskal <mpluskal@suse.com>
  - Update to 0.39.0:
    * New protect_messages hook vetoes compression of a given
      message; wrap now handles 1m context in VS Code Claude
    * Compression reports what it dropped, not just the byte count,
      and dropped scalar-array items get a retrievable sentinel
    * Image blocks are no longer dropped from /v1/messages requests
    * Provider cache stops being busted by prefix transforms, keeps
      cache_control and the forwarded OpenAI chat prefix, and is
      partitioned per upstream
    * Fixed-width command output is detected as tabular and kept out
      of Kompress; timestamped log rows no longer route as grep
      matches; .xls cells no longer fabricate digits above 2^53
    * Kompress bounds its inference deadline to the request, keeps
      the boolean connectives holding a condition together, and
      preserves null and empty strings in lossless CSV compaction
    * Proxy: rate limits scoped per api-key, configured TPM limits
      enforced, urlopen ALPN pinned to http/1.1, and the validated
      upstream address pinned so DNS rebinding cannot move it
    * Savings rollups carry exact cache-read cost and lifetime output
      spend, and report negative savings instead of flooring at zero
    * Memory: extract_system_prompt is concatenative, unresolved
      tools fail closed, None message content no longer crashes
    * The Copilot OAuth token file is written to a private temp and
      renamed into place, so a refused chmod can no longer leave the
      refresh token world-readable
    * cargo_vendor update=true moved tiktoken-rs, fancy-regex and
      friends; licence set re-derived against the new tree, unchanged
    * Many more fixes and improvements; see upstream's release notes
      for the full list
* Tue Sep 22 2026 Martin Pluskal <mpluskal@suse.com>
  - Update to 0.38.0:
    * New headroom-cache-ttl command: offline estimator that turns
      recorded cache observations into a learned TTL table for
      providers that do not publish a cache TTL
    * Bound decompressed request bodies and cap the
      /admin/runtime-env request body at 64 KiB, closing two proxy
      denial-of-service paths
    * New headroom telemetry command prints what the opt-out
      beacon sends; the beacon payload widens to cache, session
      and configuration counters
    * Match the Bearer auth scheme case-insensitively (RFC 7235)
    * Proxy accepts the OpenAI Responses shape on the gateway turn
      and enforces budget limits on the OpenAI and Gemini routes
    * Compression preserves Bash control flow and pauses AST
      compression per language after repeated invalid-syntax
      discards; file-read tool output stays byte-exact
    * Dashboard gains a Cost Saved headline card, routing savings
      and keyboard-operable controls
    * Write per-port proxy logs so concurrent instances stop
      truncating each other
    * Vendored rustls updated to 0.23.45 for RUSTSEC-2026-0285 and
      tree-sitter to 0.27.0
    * Many more fixes and improvements; see upstream's release
      notes for the full list
  - Follow upstream and raise the litellm requirement to 1.96.2,
    now unconditional on the Python version
* Tue Sep 08 2026 Martin Pluskal <mpluskal@suse.com>
  - Initial package 0.37.0: core CLI built with maturin (pyo3
    headroom._core)
    * Requires distro ast-grep rather than the ast-grep-cli PyPI
      wheel; binaries.resolve() searches PATH
    * ExclusiveArch rust_tier1_arches; crates vendored with
      cargo_vendor mode=manual
    * headroom-proxy subpackage is headroom-ai[proxy]: FastAPI,
      uvicorn, magika, onnxruntime, sqlite-vec, and the rest of
      the CLI gate (ensure_proxy_dependencies)
    * headroom-proxy is noarch (Requires-only extra)

Files

Generated by rpm2html 1.8.1

Fabrice Bellet, Fri Oct 9 23:56:32 2026