Skip to content

[Bug] AI 对话切换 Claude Code 必定 400「不支持的执行器」:assistantEngineClaude 与 executorModelCatalogs 键名不一致 #986

Open

Description

@tanjunxiaoge

摘要

在 AI 对话里把执行器切到 Claude Code,页面提示「不支持的执行器」,模型下拉为空,消息发不出去。CodeBuddy / Codex / DeepSeek Harness 均正常。

对应接口:GET /api/assistant/models?executor=claude-code

环境

  • LoopForge v0.3.8(ghcr.io/tencent/loopforge:0.3.8,revision 39f4fe76f9d757fff5fdba21da0e68dab8e68d9f)
  • Linux aarch64 / Docker Compose 自托管,AGENTFLOW_LOCAL_SANDBOX=docker

复现

curl -s -o /dev/null -w '%{http_code} %{size_download}\n' \
  -b 'agentflow_user=<session>' \
  'http://127.0.0.1:8090/api/assistant/models?executor=claude-code'
# 实际:400 33        ← {"error":"不支持的执行器"}

对照其他 id:

?executor=claude      → 200 1376
?executor=codex       → 200 874
?executor=dsh         → 200 293
?executor=claude-code → 400 33      ← 只有这一个挂

服务端日志(59 = 33 字节响应 gzip 后):

[HTTP] GET /api/assistant/models -> 400 (59 bytes, 1ms)

根因

internal/httpapi/assistant_engines.go

assistantEngineClaude = "claude-code"   // 引擎 id

internal/httpapi/executor_models.go

return map[string]executorModelCatalog{
    "codebuddy": {...}, "codex": {...},
    "claude": {...},   // ← 键是 "claude",不是 "claude-code"
    ...
}

internal/httpapi/assistant_models.go 的早返回分支直接拿引擎 id 查这张表:

id := assistantExecutor(raw)
catalog, ok := executorModelCatalogs()[id]
if !ok {
    c.JSON(400, gin.H{"error": "不支持的执行器"})
    return
}
...
Family: map[string]string{"codex": "gpt", "claude": "claude"}[id]   // 同样取不到

⇒ executorModelCatalogs()["claude-code"] 恒 miss。

为什么是 bug 而不是设计

  1. 同仓库已有针对该不一致的显式映射,说明作者知道两套 id 需要对齐,只是漏在了 models handler:
// assistant_engines.go
func assistantEngineBuiltinModels(engine string) []assistantDiscoveredModel {
    catalogKey := engine
    if engine == assistantEngineClaude {
        catalogKey = "claude"          // ← 这里转了一次
    }
    for _, model := range executorModelCatalogs()[catalogKey].Models {
  1. 执行器资源侧(web/app/agents.js)白名单用的也是 claude:
const allowed = new Set(['codebuddy','codex','claude','opencode','cursor','agy','kimi','pi','dsh','zcode']);

即"executor 资源 id = claude"与"对话引擎 id = claude-code"是两个命名空间,却被同一个 map 复用。

  1. assistantCLIEngines() 下发的引擎列表里明确是 claude-code(实测 /api/assistant/models 返回 engines: ['codebuddy','codex','claude-code',...])。前端据此选中 claude-code 再回查模型 → 必然 400。

影响面

逐项比对,只有 claude-code 不匹配,其余 9 个(codebuddy / codex / opencode / cursor / agy / kimi / pi / dsh / zcode)正常。
即 v0.3.8 下 Claude Code 完全无法在 AI 对话里使用。同问题在 v0.3.9-rc.1 仍存在(已比对两处源码)。

建议修复(最小改动)

归一化键名,不要动共享 map(其他接口依赖 claude 键):

id := assistantExecutor(raw)
catalogKey := id
if id == assistantEngineClaude {
    catalogKey = "claude"
}
catalog, ok := executorModelCatalogs()[catalogKey]
if !ok {
    c.JSON(400, gin.H{"error": "不支持的执行器"})
    return
}
family := map[string]string{"codex": "gpt", "claude": "claude"}[catalogKey]

更彻底的做法是把这层映射抽成一个函数复用(现在 assistantEngineBuiltinModels 里已有一份拷贝,再抄一份会继续漂)。

Activity

  1. cocofhu00 commented on Nov 17, 2026

    @cocofhu00
    Contributor

    感谢反馈!理解你的意思是:在 v0.3.8 的 AI 对话里把执行器切到 Claude Code 时,GET /api/assistant/models?executor=claude-code 返回 400「不支持的执行器」,模型下拉为空、消息发不出去,而 CodeBuddy、Codex、DeepSeek Harness 正常。

    已经看到,会按现有 Issue 排期查看。版本、复现步骤和对照结果都很完整,谢谢。

  2. Dylan5237 commented on Nov 17, 2026

    @Dylan5237

    Confirmed on the published ghcr.io/tencent/loopforge:0.3.8 image (x86_64), and thanks for the very thorough root-cause analysis — it matches what I reproduced exactly.

    Two additional data points that may help with triage/prioritization:

    1. Impact is less severe than "messages can't be sent" — chat actually still works.
    In my deployment the Main Bot conversation with executor claude-code completes normally: assistant_generations.status=completed, a full assistant reply is returned, token usage is recorded in assistant_sessions.payload.messages[].usage, and MCP tool calls (e.g. GitHub issue comments) execute fine. The 400 only surfaces as a red "不支持的执行器" banner in the UI, from the frontend's model-catalog probe. So this is UX-visible rather than a hard functional blocker for chat (SOP runs / role pickers that query the catalog may be more affected).

    Verified endpoint matrix with an authenticated session:

    ?executor=claude-code  -> 400 {"error":"不支持的执行器"}
    ?executor=claude       -> 200
    ?executor=codex        -> 200
    ?executor=dsh          -> 200
    

    2. Verified local hotfix for anyone on 0.3.8 who can't wait for #999:
    Adding the alias in executorModelCatalogs()'s map (instead of only patching the handler) keeps all callers consistent — the early-return handler, /api/executor-models, and assistantEngineBuiltinModels all read the same map:

    catalogs := map[string]executorModelCatalog{ /* ...existing entries... */ }
    if claude, ok := catalogs["claude"]; ok {
        catalogs["claude-code"] = claude   // chat engine id -> executor resource id
    }
    return catalogs

    I rebuilt v0.3.8 with this 6-line change (build order matters: run bash scripts/build-chatbot.sh first to generate web/native-chat, otherwise go build fails with embed: pattern all:native-chat: no matching files found; then CGO_ENABLED=0 go build -trimpath -ldflags="-s -w" ./cmd/agentflow), mounted the binary over /app/bin/agentflow, and recreated only the platform container. Re-test: ?executor=claude-code now returns 200 and the banner is gone; chat, credentials and token usage unaffected. Long-term the shared mapping helper suggested in the report is cleaner. Hope #999 lands soon.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions