feat(model): add generic OpenAI-compatible LLM backend

[nankingjing]

Summary

Adds a generic openai_compatible model backend
(skillopt/model/openai_compatible_backend.py) so SkillOpt can talk to any
service that exposes an OpenAI-compatible /chat/completions endpoint using
nothing more than a base_url and an api_key.

A single backend therefore unlocks a large family of providers, including:

• DeepSeek (https://api.deepseek.com/v1)
• Groq (https://api.groq.com/openai/v1)
• Together AI (https://api.together.xyz/v1)
• Ollama (http://localhost:11434/v1)
• vLLM / SGLang / TGI self-hosted servers
• LiteLLM proxy (http://localhost:4000)
• OpenRouter, Fireworks, Mistral, xAI Grok, Perplexity, and OpenAI itself

What's included

• New backend module built on the official openai SDK, mirroring the
  callable surface of the existing chat backends (qwen_backend,
  minimax_backend): chat_optimizer, chat_target,
  chat_optimizer_messages, chat_target_messages, plus token tracking,
  configure_openai_compatible(...), and the standard setter functions.
• Per-role config (optimizer/target) via configure_openai_compatible(...)
  or environment variables (OPENAI_COMPATIBLE_BASE_URL,
  OPENAI_COMPATIBLE_API_KEY, OPENAI_COMPATIBLE_MODEL, plus
  _TEMPERATURE/_MAX_TOKENS/_TIMEOUT_SECONDS, with OPTIMIZER_*/TARGET_*
  overrides). The API key is optional so keyless local servers (Ollama, vLLM)
  work out of the box.
• count_tokens() using tiktoken when available, with a character-based
  estimate as a fallback for non-OpenAI models.
• Tool / function calling support through the *_messages(..., tools=...)
  entry points, returning the shared CompatAssistantMessage objects.
• Retries with exponential backoff and usage recorded through the shared
  TokenTracker.

Registry / routing wiring

• common.py — registered openai_compatible in _BACKEND_DEFAULT_MODELS and
  _BACKEND_ALIASES (aliases: openai_compatible, openai_compatible_chat,
  openai-compatible, compat).
• backend_config.py — accepted as a valid optimizer/target chat backend.
• model/__init__.py — routed through chat_optimizer/chat_target and their
  _messages variants, added to token-summary aggregation and the
  reset/setter helpers, exposed configure_openai_compatible, and wired into
  the legacy set_backend("openai_compatible") shorthand.

Usage

import skillopt.model as model

model.set_backend("openai_compatible")
model.configure_openai_compatible(
    base_url="https://api.deepseek.com/v1",
    api_key="sk-...",
    model="deepseek-chat",
)

or purely via env vars:

export TARGET_BACKEND=openai_compatible
export OPENAI_COMPATIBLE_BASE_URL="https://api.groq.com/openai/v1"
export OPENAI_COMPATIBLE_API_KEY="gsk_..."
export OPENAI_COMPATIBLE_MODEL="llama-3.3-70b-versatile"

Docs

docs/guide/new-backend.md now opens with a section describing this built-in
backend so contributors reach for it before writing a bespoke one.

Testing

• python -m py_compile on all touched modules.
• Local import of skillopt.model and backend selection
  (set_backend/get_backend_name/get_optimizer_backend).
• Functional test with a mocked OpenAI client: verified routing, that the
  configured model/max_tokens are sent, CompatAssistantMessage return
  values, token-summary aggregation, and count_tokens() on both the tiktoken
  and character-fallback paths.

Note: the existing docs/guide/new-backend.md template describes an
aspirational ModelBackend/async generate() interface that does not match
the current module-based backends. This backend intentionally follows the
actual in-tree pattern so it integrates and runs today; the added docs
section reflects the real interface.

[nankingjing]

@microsoft-github-policy-service agree

[Yif-Yang]

Thanks — a generic OpenAI-compatible backend is a meaningful usability improvement, and we would like to land it through this original PR so the contribution remains attributed to you.

The PR now conflicts with newer backend registry/configuration changes. Could you please rebase onto the latest main and resolve the conflicts as a union rather than replacing the current backend lists?

Acceptance points:

• Keep all current Azure/OpenAI, Claude, Qwen, MiniMax, Codex-exec, and Claude-code-exec paths.
• Add openai_compatible for both optimizer and target routing.
• Preserve role-specific endpoint/key/model configuration.
• Ensure get_token_summary() aggregates each backend exactly once.
• Keep imports lazy enough that users of other backends do not require OpenAI-compatible credentials.

We tested a local union resolution with 47 model-focused tests plus the full suite passing. Please add or retain focused routing/config/token tests when rebasing, then ping us for a prompt re-review. We prefer merging this PR rather than replacing it, to preserve your authorship and PR history.

[nankingjing]

Rebased onto the latest main and resolved the conflicts as a union, preserving all existing Azure/OpenAI, Claude, Qwen, MiniMax, Codex-exec, and Claude-code-exec paths.

Changes made:

• Kept every existing backend dispatch and token tracker path in skillopt/model/__init__.py.
• Added openai_compatible routing for both optimizer and target chat paths (and updated unsupported-backend error messages).
• Retained role-specific endpoint/key/model configuration via configure_openai_compatible.
• Kept openai_compatible lazy: clients are built only when first requested, and the backend imports only the shared OpenAI SDK.
• Added tests/test_openai_compatible_backend.py covering role-specific config, optimizer/target routing, client laziness, and token-summary aggregation.

Validation:

• 23 model-focused tests passed (new test file + test_qwen_backend.py + test_handoff_backend.py).
• Broader unit test run: 242 passed, 3 skipped, 5 failed. The 5 failures are in test_run_sleep_fallback.py and reproduce on a clean checkout of main in this Windows environment (path/codec issues unrelated to this backend change).

Let me know if you’d like any adjustments before re-review.

[nankingjing]

The branch is already rebased on the latest main (0 commits behind) and remains mergeable. Here is a fresh verification:

Union resolution — All existing paths preserved:

• Azure/OpenAI, Claude, Qwen, MiniMax, Codex-exec, Claude-code-exec — all intact
• openai_compatible added for both optimizer and target routing in chat_optimizer, chat_target, and *_messages variants
• Role-specific config via configure_openai_compatible(...) with optimizer/target overrides
• get_token_summary() aggregates all 5 backends exactly once each
• Lazy imports preserved — openai_compatible_backend only imports the shared openai SDK

Test results:

• 23 focused model tests (new test_openai_compatible_backend.py + test_qwen_backend.py + test_handoff_backend.py) — all pass
• Full suite: 261 passed, 2 skipped, 1 failed (pre-existing test_alfworld_paths.py path separator issue on Windows, unrelated)

Ready for re-review.