v0.5.0: Experimental positioning and documentation overhaul

What changed

SubLLM's public documentation and package metadata now reflect its actual scope: an experimental, personal-use gateway — not a product pitch or commercial proxy.

Documentation

• Reframe project positioning — tagline, motivation, feature bullets, usage guidance, and terms section rewritten from assertive product pitch to conservative experimental framing
• Remove internal planning docs — docs/DESIGN.md (architecture, roadmap) and docs/RESEARCH.md (latency analysis, feature parity) deleted; content remains in git history
• Add experimental scope disclaimer — first bullet in the supported contract section marks the package as experimental and personal-use, applied in both README and code-generated docs source (docs.py)
• Align benchmark table formatting — consistent column alignment across Auth Check, Completion, Multi-turn, and Cross-provider Handoff tables

Package metadata

• Description updated to match experimental positioning across pyproject.toml, __init__.py module docstring, and README tagline
• Keywords changed from proxy, subscription to cli, gateway
• Telemetry resource version updated from hardcoded 0.4.0 to 0.5.0

What this release is NOT

No functional changes to providers, router, types, server, CLI, or test suite. All behavioral code is identical to v0.4.0. This release is purely documentation, metadata, and positioning.

Full diff: v0.4.0...v0.5.0

v0.4.0 — SDK-Only Claude Code Provider

Breaking Changes

• ClaudeCodeProvider no longer accepts the use_sdk parameter
• The subllm[sdk] optional extra is removed
• claude-agent-sdk is now a required dependency

Performance

• 42–76% latency reduction across all single-request workloads
• Warm call: −75.6%, Cold start: −66.8%, Multi-turn: −64.5%

Changes

• perf(claude)!: remove CLI subprocess path, SDK-only inference
• build(deps): add twine to dev dependencies for PyPI publishing

Full Changelog: v0.3.0...v0.4.0

v0.3.0 — First PyPI Release

SubLLM is now available on PyPI: pip install subllm

Route OpenAI-compatible API calls through subscription-authenticated coding agents — no API keys required.

Highlights

• 3 providers: Claude Code, OpenAI Codex CLI, Gemini CLI
• OpenAI-compatible proxy server — drop-in replacement for any OpenAI SDK client
• CLI with auth checking (subllm auth) and model listing (subllm models)
• Streaming and batch support with configurable concurrency
• PEP 561 typed package — full inline type annotations with py.typed marker

Supported Models

Provider	Models
Claude Code	claude-opus-4-6, claude-sonnet-4-5, claude-haiku-3-5
Codex CLI	gpt-4.1, gpt-5.2-codex, gpt-5.2, gpt-5-mini
Gemini CLI	gemini-2.5-pro, gemini-2.5-flash

Install

pip install subllm              # core library
pip install subllm[server]      # with proxy server

Quick Start

import subllm

response = await subllm.completion(
  model="claude-code/claude-sonnet-4-5",
  messages=[{"role": "user", "content": "Hello"}],
)

Links

- PyPI: https://pypi.org/project/subllm/0.3.0/
- Full README: https://github.com/Mathews-Tom/SubLLM#readme