[PSA-Team-source]

Summary

• Adds an UberLLM section to docs/local-openai-compatible-providers.md, after the LLMTR section. Docs only.
• Covers the base URL and key, using local/<catalog id> for every id (a bare qwen/ id goes to DashScope), and how the 64,000-token request relates to each model's output limit.
• Disclosure: I work on UberLLM. It is a paid, prepaid service with no free tier, and the section says so.
• Closes docs: document UberLLM OpenAI-compatible gateway setup #3375

Anti-slop triage

• Classification: docs-only
• Evidence: docs: document UberLLM OpenAI-compatible gateway setup #3375. Release binary v0.1.11 (claw-linux-x64, sha256 checked): claw --model local/deepseek/deepseek-v4-flash prompt "Reply exactly HELLO_WORLD_123" printed HELLO_WORLD_123, and so did local/moonshotai/kimi-k2.6 and local/qwen/qwen3.8-flash. A bare qwen/qwen3.8-flash failed with missing DashScope credentials, as the section says. A direct request with max_tokens: 64000 to aion-labs/aion-3-0-mini (output limit 32,768) answered normally.
• Non-destructive review result: merge candidate

Verification

• python .github/scripts/check_doc_source_of_truth.py and check_release_readiness.py pass. No Rust code changed, so cargo test wasn't run; I tested with the published release binary rather than a local build.
• git diff --check passes.
• The curl ... | jq model-list command in the section was run as written.

🤖 Generated with Claude Code

[Yeachan-Heo]

✓ Verified Claims

local/ prefix routing:

• Function strip_routing_prefix in rust/crates/api/src/providers/openai_compat.rs:1002 recognizes "local" and strips it.
• Remainder is returned verbatim: &model[pos + 1..]

Native route avoidance:

• Bare qwen/ models route to DashScope (confirmed in rust/crates/api/src/providers/mod.rs:288 check: if canonical.starts_with("qwen/"))
• Test at L1280 confirms: qwen-prefixed models resolve to DashScope auth, not UberLLM

Default max_tokens:

• max_tokens_for_model in rust/crates/api/src/providers/mod.rs:628 defaults to 64_000
• No OpenAI-compatible-specific override; default applies uniformly

Vendor prefixes verified against live UberLLM catalog:

• deepseek/deepseek-v4-flash exists, 128k max_output_tokens
• moonshotai/kimi-k2.6 exists, 65.5k max_output_tokens
• aion-labs/aion-3-0-mini exists, 32,768 max_output_tokens (example in doc)
• curl command confirmed functional and accurate (returns JSON with max_output_tokens field)

✓ Consistency & Quality

• Structure, length, and scope match precedent LLMTR section (merged 6ac59ab)
• Properly distinguishes UberLLM behavior (bounds output) vs LLMTR (rejects request)
• No duplication of generic OpenAI-compatible guidance from OmniRoute section
• Placeholder key format consistent: ull_... follows pattern
• No unverified claims or marketing language

✓ Validation

• python .github/scripts/check_doc_source_of_truth.py → passed
• git diff --check → no trailing whitespace

—
[repo owner's gaebal-gajae (clawdbot) 🦞]

[Yeachan-Heo]

Merged, thanks. The routing and token-limit details check out against the provider code, and the disclosure is clear.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]

[XinwangTao]

Thanks for merging this — the UberLLM section with the local/<catalog id> prefix guidance and the verified examples (deepseek/deepseek-v4-flash, moonshotai/kimi-k2.6, and the qwen3.8-flash DashScope fallback case) makes the setup easy to follow for new users.