fix: Ollama API compatibility — allow bare model names and smart openai/ prefix handling on wire#3167
fix: Ollama API compatibility — allow bare model names and smart openai/ prefix handling on wire#3167GeorgijeSav wants to merge 1 commit into
Conversation
…ai/ prefix handling on wire fix: Ollama API compatibility — allow bare model names and smart openai/ prefix handling on wire
There was a problem hiding this comment.
Pull request overview
Note
Copilot was unable to run its full agentic suite in this review.
This PR adjusts model ID handling for OpenAI-compatible endpoints, allowing “bare” model names for local gateways and refining when the openai/ routing prefix is stripped vs preserved.
Changes:
- Add loopback base-URL detection to allow bare model IDs in the CLI when
OPENAI_BASE_URLtargets localhost/loopback. - Update OpenAI-compat provider logic to strip
openai/only for default OpenAI and loopback endpoints, preserving it for non-local gateways. - Expand/adjust unit and integration tests to cover the new routing/validation behavior.
Reviewed changes
Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.
| File | Description |
|---|---|
| rust/crates/rusty-claude-cli/src/main.rs | Adds local base URL detection and updates model syntax validation + tests to allow bare model names when pointing at loopback. |
| rust/crates/api/src/providers/openai_compat.rs | Refines openai/ prefix stripping decisions based on normalized base URL and loopback host detection; adds focused tests. |
| rust/crates/api/tests/openai_compat_integration.rs | Updates integration expectations and adds coverage for non-routing slash-containing model IDs. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| fn is_local_base_url(url: &str) -> bool { | ||
| let rest = match url.split_once("://") { | ||
| Some((_, r)) => r, | ||
| None => return false, | ||
| }; |
| fn normalize_base_url(url: &str) -> &str { | ||
| let url = url.trim_end_matches('/'); | ||
| url.strip_suffix("/chat/completions") | ||
| .map(|s| s.trim_end_matches('/')) | ||
| .unwrap_or(url) | ||
| } | ||
|
|
||
| /// Extract the host (without port) from a URL string. | ||
| /// Returns an empty string if the URL cannot be parsed. | ||
| fn url_host(url: &str) -> &str { | ||
| // Strip scheme ("https://", "http://", etc.) | ||
| let rest = match url.split_once("://") { | ||
| Some((_, r)) => r, | ||
| None => return "", | ||
| }; |
| #[test] | ||
| fn wire_model_strips_openai_prefix_for_custom_base_url() { |
|
Note, my corporate LLM API server do not present itself as localhost, and still need to accept model names without slashes in them. The assumtion that this only is relevant for localhost/127.0.0.1/::1 is not correct in my case. Also, it is worth noting that 127.0.0.2 and all other IPv4 addresses starting with 127 are also localhost. |
|
To be simplified why not completely remove function validate_model_syntax? |
| .eq_ignore_ascii_case(normalize_base_url(config.default_base_url)); | ||
| let host = url_host(base_url); | ||
| let is_local_url = | ||
| host.eq_ignore_ascii_case("localhost") || matches!(host, "127.0.0.1" | "::1"); |
There was a problem hiding this comment.
Should not be limited to localhost IP.
I may understand that local is implied to be really local (as one machine) but there is the case where the open endpoint can be at a totally :
- different IP address locally (in example the Nividia DGX Spark as a inference only server) at 10.0.0.0/8, 127.x.x.x ,172.16/12, 192.168/16
- an custom url with a DNS (example : vllm-prod-inference.github/v1)
- public IP address or URL (with a token as a payload for instance)
fix: Ollama API compatibility — allow bare model names and smart
openai/prefix handling on wireFixes: ultraworkers/claw-code#3123
PR: GeorgijeSav/claw-code#1
Related: #3051, #3044
Summary
Ollama and other local OpenAI-compatible providers expose bare model names such as:
without a
provider/prefix.This caused two cascading failures reported in #3123:
validate_model_syntaxrejected bare model names outright because it requiredprovider/modelformat.to satisfy validation still failed because the full slug was forwarded on the wire, producing:
from the local provider.
This PR fixes both layers.
Changes
validate_model_syntaxBare model names (without
/) are now accepted whenOPENAI_BASE_URLresolves to a loopback host:localhost127.0.0.1::1Non-local gateways (e.g. OpenRouter) still require
provider/modelformat.wire_model_for_base_urlThe
openai/prefix is now stripped for:The prefix is preserved for custom non-local gateways.
Example:
continues to route correctly through providers such as OpenRouter.
Tests
Added:
Updated:
Anti-slop triage
Classification
actionable-fixEvidence
ultraworkers/claw-code#3123
Reproduction:
Result:
and:
Result:
Both failure modes are now covered and fixed by tests.
Non-destructive review result
merge candidateChanges are additive and guarded behind loopback URL detection.
Behaviour for non-local gateways remains unchanged.
No existing tests were removed.
Regression coverage includes:
/chat/completionssuffix handlingVerification
cargo test --workspacepassesgit diff --checkpassesResolution gate
Resolves ultraworkers/claw-code#3123.
Verification evidence:
test_bare_model_name_passes_when_openai_base_url_setwire_model_strips_openai_prefix_for_custom_base_urlAll pass successfully.
Merge status