feat: add local/ routing prefix for model IDs with embedded slashes

[petterreinholdtsen]

Summary

Some backends (e.g., gpt.uio.no running Ollama) have model IDs containing slashes like Qwen/Qwen3.6-27B-FP8. The existing routing prefix logic would strip 'Qwen/' as a provider prefix, sending only the remainder on the wire and causing 403 Forbidden from the server.

The new local/ prefix is an escape hatch: it strips just 'local/' and sends everything after it verbatim, preserving embedded slashes in model IDs.

Usage: --model local/Qwen/Qwen3.6-27B-FP8

This patch was developed with assistance from OpenCode and local Qwen 3.6 API server.

Anti-slop triage

I do not undersnad these fields:

• Classification:
• Evidence:
• Non-destructive review result:

Verification

I do not understand the two first of these check list points.

• Targeted tests/docs checks ran, or the gap is explicitly recorded.
• git diff --check passes.
• No live secrets, tokens, private logs, or unrelated generated churn are included.

Resolution gate

I do not understand these check list points:

• If this PR resolves an issue, the issue number and fix evidence are linked.
• If this PR should not merge, the rejection/defer rationale is evidence-backed and does not rely on vibes.
• I did not merge/close remote PRs or issues from an automation lane without owner approval.

I believe this is related to #3036, but unsure if it fix it.

[petterreinholdtsen]

I asked opencode to keep notes on what it was solving, and here is what it ended up with for this patch. I've removed secrets and private DNS names.

Crash: "Access denied to model" on corporate API server (routing prefix stripping) [SOLVED]

• Error: api returned 403 Forbidden: Access denied to model: Qwen3.6-27B-FP8
• Root cause: Corporate API server has model IDs containing slashes like Qwen/Qwen3.6-27B-FP8. The existing routing prefix logic in wire_model_for_base_url() stripped Qwen/ as a known provider prefix, sending only Qwen3.6-27B-FP8 on the wire — which the server doesn't recognize (403). Additionally, validate_model_syntax() rejected 3+ parts when split by /, so even if you tried to work around it, the CLI would refuse the model name.
• Key files:
  • rust/crates/api/src/providers/openai_compat.rs:920-933 — strip_routing_prefix() now accepts "local" as a routing prefix
  • rust/crates/api/src/providers/openai_compat.rs:964-975 — wire_model_for_base_url() handles "local/" escape hatch
  • rust/crates/rusty-claude-cli/src/main.rs:1654-1658 — validate_model_syntax() now allows embedded slashes for local/ prefix

What was done

• Added "local" to the routing prefixes in both strip_routing_prefix() and wire_model_for_base_url(). The local/ prefix is an escape hatch: it strips just local/ and sends everything after it verbatim on the wire, preserving embedded slashes in model IDs.
• Updated validate_model_syntax() to allow more than 2 parts when split by / for models starting with local/, so local/Qwen/Qwen3.6-27B-FP8 passes validation.
• Use: --model local/Qwen/Qwen3.6-27B-FP8 sends the full ID on the wire.

Test command

echo "Say hello." | OPENAI_BASE_URL=https://corporate.example.com/api/v1/chat/completions OPENAI_API_KEY=secret ~/src/claw-code-upstream/rust/target/debug/claw --model local/Qwen/Qwen3.6-27B-FP8

[1716775457damn]

The local/ escape hatch is a clean solution for model IDs with embedded slashes. Much better than trying to guess which / is a provider separator vs part of the model name. This should also help with the broader Ollama compatibility issues in #3123.

[Noobzik]

One more test case then it should be good for the maintainer to approve
@code-yeongyu

[Noobzik]

I think there should be one more case where you have local/some_random_model:7b that should assert to local/some_random_model:7b

Overwise, that look's good for me

[1716775457damn]

Nice approach. The local/ prefix convention is clean and doesn't conflict with existing provider routing. One thought: since the model ID after local/ is used as-is, servers with query-string authentication (e.g. http://localhost:1234/v1?key=xxx) could have their key leaked into model_id. Consider either stripping query params, or documenting that local/ expects just the host:port and paths, not query strings.