fix(gateway): strip port from openai-compat openaiCustomUrl (#1057)

[JOJOCrazy123]

Summary

Fixes #1057. When the openai-compat LLM provider is configured with an IP:port base URL (e.g. http://10.43.46.12:3000/v1), openaiCustomUrl was set to the full URL including the port. Higress derives the upstream host domain from openaiCustomUrl, so it extracted 10.43.46.12:3000 as the host domain — not a valid DNS hostname — and upstream DNS resolution failed, making the LLM endpoint unreachable.

The port is already provided separately via openaiCustomServicePort, so including it in openaiCustomUrl is both redundant and breaking.

Changes

• hiclaw-controller/internal/initializer/initializer.go: add customOpenAIURL() helper that strips the port (keeps scheme/host/path) and use it for openaiCustomUrl in both the openai-compat and custom-provider paths.
• manager/scripts/init/setup-higress.sh: build openaiCustomUrl from scheme + domain + path only (no port).
• hiclaw-controller/internal/initializer/initializer_test.go: unit tests for customOpenAIURL.
• changelog/current.md: record the fix.

Result: openaiCustomUrl becomes http://10.43.46.12/v1, host domain is 10.43.46.12, and the port is supplied only via openaiCustomServicePort.

Test plan

• bash -n manager/scripts/init/setup-higress.sh passes; parsing verified for IP:port+path, host:port, no-port, and no-path inputs.
• go test ./internal/initializer/ — could not run locally (no Go toolchain in this environment); expected to pass via CI.
• Manual: initialize with an openai-compat provider using an IP:port base URL and confirm the Higress provider "自定义 OpenAI 服务主机域" shows the host without the port.

Notes

The issue also mentions a secondary bug where the Go EnsureAIProvider/ensureServiceSource use POST-only (treating 409 as success) and cannot correct an already-misconfigured provider on re-run. That is a separate concern and intentionally left out of this narrow fix.

[github-actions]

❌ Integration Tests Failed (controller-cr-2 / mgr=copaw / wk=copaw)

Commit: 539460c
Workflow run: #1699

Test Results

No test output captured.

Debug Log (tail)

No debug logs available.

---

📦 Download full debug logs & test artifacts

[github-actions]

📊 CI Metrics Report

Summary

Metric	Current	Baseline	Change
LLM Calls	183	97	+86 ↑ +88.7%
Input Tokens	5628772	2719498	+2909274 ↑ +107.0%
Output Tokens	52922	16508	+36414 ↑ +220.6%
Total Tokens	5681694	2736006	+2945688 ↑ +107.7%

By Role

Role	Metric	Current	Baseline	Change
🧠 Manager	LLM Calls	111	72	+39 ↑ +54.2%
	Input Tokens	3802648	2092724	+1709924 ↑ +81.7%
	Output Tokens	24127	10274	+13853 ↑ +134.8%
	Total Tokens	3826775	2102998	+1723777 ↑ +82.0%
🔧 Workers	LLM Calls	72	25	+47 ↑ +188.0%
	Input Tokens	1826124	626774	+1199350 ↑ +191.4%
	Output Tokens	28795	6234	+22561 ↑ +361.9%
	Total Tokens	1854919	633008	+1221911 ↑ +193.0%

Per-Test Breakdown

Test	Mgr Calls	Wkr Calls	Δ Calls	Mgr In	Wkr In	Mgr Out	Wkr Out	Δ Tokens	Trend
02-create-worker	6	0	+1 ↑ +20.0%	168385	0	578	0	+31975 ↑ +23.3%	⚠️ regressed
03-assign-task	8	5	0 — 0%	221259	113392	1247	646	+26364 ↑ +8.5%	— unchanged
04-human-intervene	11	6	+3 ↑ +21.4%	276498	136008	1293	693	+78055 ↑ +23.2%	⚠️ regressed
05-heartbeat	6	5	+4 ↑ +57.1%	182641	111190	1019	741	+97387 ↑ +49.1%	⚠️ regressed
06-multi-worker	80	56	+78 ↑ +134.5%	2953865	1465534	19990	26715	+2711907 ↑ +154.6%	⚠️ regressed

Trends

⚠️ 4 test(s) regressed (more LLM calls)

---

Generated by HiClaw CI on 2026-07-19 08:36:24 UTC

---

📦 Download debug logs & test artifacts

[JOJOCrazy123]

CI note: the only red check (controller-cr-2 / test-21) is unrelated to this PR

The sole failing check is integration-tests (controller-cr-2, SHARD_D_TESTS, copaw, copaw) → test-21-team-project-dag. All other shards and build jobs are green.

This PR only touches the gateway openai-compat URL handling (manager/.../initializer.go, manager/scripts/init/setup-higress.sh, changelog). It does not touch the CoPaw team-controller / Leader coordination path that test-21 exercises.

test-21 is a non-deterministic LLM-behavior E2E test. In this run it failed because the CoPaw Leader posted an interim planning message in the Leader DM ("I'll start by reviewing the team-coordination... to plan and coordinate this project.") instead of replying NO_REPLY and putting the first assignment in the Team Room — which trips the fail-fast heuristic (MAX_DM_ONLY_POLLS=1). That's a prompt-adherence flake in the runtime, not something this change can affect.

Importantly, the same shard is also failing on main: latest push run https://github.com/agentscope-ai/AgentTeams/actions/runs/29649250704 (controller-cr-2, SHARD_D_TESTS red). So this is a pre-existing flaky failure, not a regression from #1061.

Could a maintainer re-run the failed job? Happy to rebase if that helps. Thanks!

[JOJOCrazy123]

Hi maintainers, this PR is ready for review, but the workflow is still awaiting maintainer approval. Could someone please approve the workflow and review the PR when convenient? Thank you!

[shiyiyue1102]

Thanks for the focused fix. The IPv4 host:port case is handled, but customOpenAIURL currently reconstructs the URL from Scheme, Hostname, and the decoded Path, so it changes more than the port: it drops RawQuery, decodes escaped path segments, and emits unbracketed IPv6 literals. I reproduced all three cases locally.

Could you preserve the parsed URL and remove only the port, then add regression cases for query strings, escaped paths, and IPv6 literals? The existing controller unit and envtest integration suites pass locally; the red test-21 shard also fails on main and appears unrelated.

---

感谢这个聚焦的修复。当前 IPv4 host:port 场景已经处理，但 customOpenAIURL 通过 Scheme、Hostname 和解码后的 Path 重新拼装 URL，实际修改的不只是端口：它会丢失 RawQuery、解码转义路径，并生成没有方括号的 IPv6 URL。我已在本地复现这三种情况。

请保留解析后 URL 的其他字段，仅移除端口，并补充 query、转义路径和 IPv6 字面量的回归用例。现有 Controller 单元测试和 envtest 集成测试在本地均通过；红色的 test-21 分片在 main 上也失败，判断与本 PR 无关。

[JOJOCrazy123]

Thanks for the careful review — you're right on all three counts.

I've reworked customOpenAIURL so it no longer rebuilds the URL from Scheme/Hostname/Path. It now keeps the parsed *url.URL and removes only the port (sets u.Host to the hostname, re-bracketing IPv6 literals) then returns u.String(). As a result:

• Query strings (RawQuery) are preserved.
• Escaped path segments are no longer decoded (via EscapedPath() in String()).
• IPv6 literals stay bracketed, e.g. http://[2001:db8::1]:3000/v1 → http://[2001:db8::1]/v1.

Added regression cases for query strings, escaped paths (%2F), and IPv6 (with and without a port).

I also merged latest main to resolve the conflict from the hiclaw-controller → agentteams-controller rename (the helper and its test moved with it). Pushed as 718c83d.

Agreed the red test-21 shard is unrelated — it also fails on main.