feat: add OpenRouter as LLM and embedding provider

[oglenyaboss]

Summary

• Add OpenRouter as a first-class LLM and embedding provider, enabling access to 200+ models through a single OPENROUTER_API_KEY
• Both repowise init (doc generation) and repowise serve (chat/search) work with OpenRouter out of the box
• No new pip dependency — uses the existing openai package (OpenAI-compatible API)

LLM Provider

• OpenRouterProvider with generate() + stream_chat() support
• Default model: anthropic/claude-sonnet-4.6
• Sets recommended HTTP-Referer and X-Title headers for OpenRouter dashboard tracking
• Rate limits: 60 RPM / 200K TPM (conservative defaults, user-overridable)

Embedding Provider

• OpenRouterEmbedder for semantic search and chat RAG
• Default model: google/gemini-embedding-001 (768 dims)
• One API key covers both LLM and embeddings

Integration Points

• Registered in both LLM and embedding registries (lazy import)
• CLI auto-detection from OPENROUTER_API_KEY env var
• Interactive provider/embedder selection in repowise init and repowise serve
• Server provider catalog updated for web UI

Known Limitations

• Cost tracking disabled for OpenRouter: since it proxies 200+ models with varying prices, the fallback pricing would show inflated numbers. Users should check the OpenRouter dashboard for actual costs. A future PR could fetch real pricing via the /api/v1/models endpoint.
• Chat and embeddings not fully tested — LLM generation was tested end-to-end on a real 800+ file project using qwen/qwen3.6-plus via OpenRouter. Chat and embedding functionality was not integration-tested but follows the same OpenAI-compatible patterns as the existing OpenAI provider.

Test plan

• 13 unit tests for OpenRouterProvider (construction, generation, error mapping, headers)
• Registry test updated (builtin count 6 → 7, openrouter in list)
• Integration test added (skipped without OPENROUTER_API_KEY)
• Manual end-to-end test: repowise init with OpenRouter on a real project (328 wiki pages generated successfully)
• Chat via web UI with OpenRouter as active provider
• Embedding search with OpenRouterEmbedder

[RaghavChamadiya]

This is really thorough, probably the most complete provider PR we've gotten. LLM + embedder + CLI + server catalog all done, and the honest known limitations section is appreciated.

One thing to fix before merge:

1. cost_tracker dead code (blocking): the constructor accepts cost_tracker but hardcodes it to None, and then there's an if self._cost_tracker block in generate() that can never execute. Either remove the parameter entirely or add a log warning that cost tracking is disabled for OpenRouter. Just don't want dead code sitting there confusing people.

Also flagging for awareness (not blocking):

2. stream_chat duplication: the streaming logic is identical to OpenAI's, about 95 lines copy-pasted. Would be nice to extract a shared base eventually, but not blocking this PR on it.

3. Tests: no coverage for stream_chat or OpenRouterEmbedder. Even basic construction + mock tests would help, but also not blocking.

Should be a quick fix, happy to merge after (1).

[oglenyaboss]

Thanks for the review!

1. cost_tracker dead code — fixed in c1c1773. Removed the cost_tracker parameter and the unreachable if self._cost_tracker block entirely. The constructor now accepts **_kwargs so it won't break if the pipeline passes cost_tracker= at instantiation time.

2. stream_chat duplication — agreed, the shared base extraction makes sense. Happy to tackle that in a follow-up PR if you'd like.

3. Tests — added in 57d40c7:

• stream_chat: text deltas, tool calls, and rate limit error (3 tests)
• OpenRouterEmbedder: construction, dimensions, embedding, and base URL verification (12 tests)

[swati510]

Nice clean provider addition, follows the existing shape well. A few things worth looking at before merge:

1. Silent dimension corruption risk in OpenRouterEmbedder: _DIMS falls back to 768 for any model not in the dict, but the actual model might be 1024 or 3072 dims. If a user picks e.g. cohere/embed-english-v3, they'll get 768-dim stored vectors that don't match the model's real output, vector store silently corrupted. Safer to raise ValueError("unknown embedding model %s, add to _DIMS") instead of falling back.

2. OpenAI param compat: you're using max_completion_tokens, which is the newer OpenAI SDK name. OpenRouter proxies 200+ models and not all of them accept that param (some only take max_tokens). Worth smoke-testing against a few of the listed defaults (anthropic/claude-sonnet-4.6, google/gemini-, meta-llama/) to confirm it works, and falling back to max_tokens if not.

3. The OpenRouterProvider constructor accepts **_kwargs: Any and silently drops them. If the registry later starts passing e.g. cost_tracker=..., rate_limiter=..., or tier=... (now that minimax/zai use tier), those will just vanish. Drop the **_kwargs catchall or convert it to explicit params like every other provider.

Embedder dimensions fallback is the one that could bite users badly, the other two are mostly hygiene.

[swati510]

Silent 768-dim fallback for unknown models is a footgun. If a user picks a model with different real dimensions (e.g. 1024), stored vectors won't match the model output and the vector store is corrupted without any error. Raise ValueError here instead, force users to add new models to _DIMS explicitly.

[swati510]

Silently swallowing unknown kwargs means any future change to how the registry constructs providers (e.g. passing cost_tracker, rate_limiter, tier) will fail silently for this provider. Drop the catchall and declare the params you accept explicitly.

[oglenyaboss]

Thanks for the catches! Fixed all three in 1f9f67a:

1. Silent dimension fallback → ValueError. Moved the check to __init__ rather than the dimensions property — fails at construction time so misconfig surfaces at startup instead of at the first vector-store sizing call. Error message names the bad model and lists the known ones so the fix is obvious.

2. **_kwargs catchall → explicit cost_tracker param. Dropped the catchall and added cost_tracker: CostTracker | None = None to match the other providers' signatures. Accepted but unused (still no cost tracking for OpenRouter — docstring updated to say so). Any future kwarg from the registry will now blow up loudly instead of vanishing. Added a test_rejects_unknown_kwargs to lock that in.

3. max_completion_tokens → max_tokens. Verified via OpenRouter's API docs that max_tokens is the standard parameter — it's what's in their main Request schema and all their Python/Swift examples. max_completion_tokens is an OpenAI-specific newer name. Switched both generate() and stream_chat().

Tests updated: replaced the unknown-model-defaults test with a construction-failure test, switched the kwargs assertion to max_tokens, and added the cost_tracker positive + unknown-kwarg negative tests. 104/104 provider+persistence tests pass.