fix(higress): strip provider prefix from extra-provider model names

[nattiini45]

What this fixes

OpenAI-compatible upstreams (Ollama Cloud, MiMo, Vanchin/StreamLake) reject model names carrying the <provider>/ routing prefix with a 404 — they expect the bare upstream name (e.g. gpt-oss:120b, not ollama/gpt-oss:120b). The extra-provider AI routes created by setup-higress.sh matched on the prefixed name for routing but left modelMapping empty, so the prefixed name was forwarded upstream unchanged. This is the OpenClaw bug #67435 / #72353 pattern applied to custom openai-type providers.

The fix

Each extra-provider AI route now carries a regex modelMapping:

"modelMapping": {"~^<provider>/(.+)$":"$1"}

This strips the <provider>/ prefix before the upstream call, while the route's modelPredicate still matches the prefixed name for routing. Applied to both the POST (new route) and PUT (upgrade existing route) paths in setup-higress.sh.

How it was verified

• Higress ai-proxy source (provider.go:getMappedModel): supports ~regex keys with $1 capture via Go's regexp.ReplaceAllString. Confirmed by reading the source.
• Simulated with real Go regexp semantics (not Python's \1):
  • ollama/gpt-oss:120b → gpt-oss:120b ✅
  • ollama/glm-5.2 → glm-5.2 ✅
  • ollama/kimi-k2.7-code → kimi-k2.7-code ✅
  • mimo/mimo-v2.5-pro → mimo-v2.5-pro ✅
  • vanchin/kat-coder-pro-v2 → kat-coder-pro-v2 ✅
  • vanchin/kat-coder-pro-v2.5 → kat-coder-pro-v2.5 ✅
• bash -n setup-higress.sh — syntax valid
• go build ./... — compiles clean
• go vet ./... — clean
• go test ./internal/agentconfig/... — passes
• Generated route JSON — validated

Model catalog (also fixed)

Updated the 3 source spots (known-models.json, manager-openclaw.json.tmpl, agentconfig/generator.go presets + lists) with the real provider model IDs and correct context windows / maxTokens. The prior provisional entries used ollama/gpt-oss:120b-cloud (the -cloud suffix doesn't exist upstream) and mimo/MiMo-V2.5-Pro (capitalized; MiMo's API wants lowercase mimo-v2.5-pro) — both would 404 even after this fix.

New entries:

• ollama/gpt-oss:120b, ollama/glm-5.2, ollama/glm-5.1, ollama/kimi-k2.7-code
• mimo/mimo-v2.5, mimo/mimo-v2.5-pro
• vanchin/kat-coder-pro-v2 (reasoning:false, 32k max), vanchin/kat-coder-pro-v2.5 (reasoning:false, 64k max)

Files changed (4)

• manager/scripts/init/setup-higress.sh — modelMapping regex on POST + PUT paths
• manager/configs/known-models.json — real model IDs, correct ctx/max
• manager/configs/manager-openclaw.json.tmpl — same
• hiclaw-controller/internal/agentconfig/generator.go — presets + lists

go vet clean, agentconfig tests pass.

🤖 This PR was generated by Rumi (Caspar's AI assistant) after live-verifying all 4 upstream providers against the actual APIs and confirming the root cause against the Higress source code.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Overview

Clean fix that correctly strips the <provider>/ routing prefix from model names before forwarding to OpenAI-compatible upstreams. The regex ~^<provider>/(.+)$ → $1 is verified correct for Go's regexp.ReplaceAllString. Both POST (new route) and PUT (upgrade) paths are covered. Model catalog entries are consistent across all 3 catalog files (generator.go, known-models.json, manager-openclaw.json.tmpl).

Files Reviewed (4 files)

• hiclaw-controller/internal/agentconfig/generator.go — model presets + lists updated
• manager/configs/known-models.json — model catalog JSON
• manager/configs/manager-openclaw.json.tmpl — model catalog template
• manager/scripts/init/setup-higress.sh — modelMapping regex on POST + PUT paths

---

_{Reviewed by mimo-v2.5-pro · Input: 65.3K · Output: 15K · Cached: 386.9K}