Feature: per-provider opt-out of code_mode_only tool mode for routed models (deepseek-v4-flash: undeclared exec_command aborts stream, reconnect then fails reasoning replay with DeepSeek 400)

Client or integration

Codex App + Codex CLI 0.137.0

Area

Proxy and routing (catalog / tool mode)

Summary

Routed models are unconditionally stamped tool_mode = "code_mode_only", but some models (observed: deepseek/deepseek-v4-flash) ignore the code-mode nested-helper contract and call exec_command at the top level. The undeclared-tool guard then aborts the stream, the Codex app reconnects, and the replayed turn fails reasoning replay with an upstream 400:

stream disconnected before completion: routed provider emitted undeclared client tool "exec_command"; only request-declared tools may be called

The `reasoning_text` in the thinking mode must be passed back to the API.

Reproduction

ocx start, add deepseek provider (adapter openai-responses), ocx sync
In Codex (app or CLI) select deepseek/deepseek-v4-flash
Ask the model to run any shell command (or let it decide to use a tool)
The model calls exec_command directly (not via the code-mode exec/tools.* helper)
Stream aborts, client reconnects, reasoning replay fails with the DeepSeek 400 above

Root cause analysis

applyRoutedCodexToolMode (src/codex/catalog/parsing.ts) sets tool_mode = "code_mode_only" for every routed model with no opt-out.
Under code_mode_only, codex-rs declares the unified exec (freeform) tool; exec_command/shell_command exist only as nested tools.* helpers inside it and are not top-level declared tools.
The undeclared-tool guard (src/server/responses-undeclared-tool-guard.ts, src/bridge.ts) correctly refuses a top-level exec_command call because the codex client cannot execute an undeclared call.
The tool-catalog nudge (src/adapters/tool-catalog-nudge.ts) is advisory and is only injected on the openai-chat adapter path; the openai-responses passthrough never injects it, and the model ignored it in both paths anyway.
After the guard aborts the stream, the Codex app reconnects and replays the interrupted turn; the reasoning part of that turn is then missing/incomplete on replay, which DeepSeek's thinking mode rejects with The reasoning_text in the thinking mode must be passed back to the API. The two errors are one chain: fix the tool-mode mismatch and the 400 disappears (verified).

Suggested fix

Add a per-provider (ideally per-model) config field so operators can keep the flat shell tool mode for models that do not honor code mode:

# config.json: providers.<name>
"codexToolMode": "shell"   # "code_mode_only" (default) | "shell"

applyRoutedCodexToolMode would honor it: when "shell", leave tool_mode unset so codex-rs declares exec_command per shell_type: "shell_command", making the model's calls declared and executable. Default stays code_mode_only for all other routed models.

Optionally, also inject the tool-catalog nudge on the openai-responses passthrough path as defense-in-depth.

Local workaround (verified)

Patched applyRoutedCodexToolMode locally to honor providers.<name>.codexToolMode === "shell" (config schema is passthrough, so no schema change needed). Verified on opencodex 2.26.0:

deepseek/deepseek-v4-flash: exec_command calls now execute (zsh -lc 'echo ...' returns output), no stream abort, no reconnect, no 400
Other routed models (e.g. kimi/*) keep tool_mode = code_mode_only and are unaffected
Two-turn continuation with reasoning replay works on both openai-chat and openai-responses adapters

Version

2.26.0

Operating system

macOS

🔗 Related PRs

#304 - fix(kiro): restore hardening and complete text turns [open]
#1361 - fix(responses): restore DeepSeek streaming and routed code-mode tools [merged]
#1788 - fix(responses): fail closed when a routed provider calls an undeclared tool [merged]
#1951 - fix(adapters): decide code mode from tool semantics, not from the name exec [merged]
#1953 - fix(adapters): restore the un-namespaced requirement on the shell-bridge check [merged]

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

Create Plan

🧪 Issue enrichment is currently in open beta.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

fix(catalog): allow per-provider/model opt-out of code_mode_only tool…

리뷰 · 우선순위 52 / 80

라우팅된 DeepSeek 모델이 Codex code-mode 계약을 무시하고 최상위 exec_command를 부른다. OpenCodex 가드가 스트림을 끊고, 앱이 재접속한 뒤 reasoning을 재전송하지 못해 DeepSeek가 400을 준다. 기본 native 경로가 아니라 라우티드 프로바이더 도구 호환 문제라 52다. 로컬에서 codexToolMode=shell로 우회가 검증됐고, 다른 라우티드 모델의 기본값은 그대로 두면 된다.

원인은 src/codex/catalog/parsing.ts의 applyRoutedCodexToolMode()다. ROUTED_CODEX_TOOL_MODE = "code_mode_only"를 모든 라우티드 엔트리에 무조건 찍는다. src/codex/catalog/sync.ts가 카탈로그를 쓸 때 이 함수를 호출한다. opt-out 인자는 없다. code_mode_only 아래에서 codex-rs는 통합 exec(freeform)만 최상위로 선언하고, exec_command/shell_command는 tools.* 중첩 헬퍼로만 존재한다.

모델이 최상위 exec_command를 내면 src/server/responses-undeclared-tool-guard.ts와 src/bridge.ts가 맞다. 가드 문구가 리포터 로그와 같다. routed provider emitted undeclared client tool "exec_command"; only request-declared tools may be called. 가드가 틀린 게 아니라, 카탈로그가 선언한 표면과 모델이 부르는 이름이 어긋난 것이다.

src/adapters/tool-catalog-nudge.ts는 조언일 뿐이다. 주석대로 exec 옆에 납작한 exec_command가 있으면 브리지 모양이고, code mode면 await tools.exec_command(...)를 쓰라고 설명한다. 이 넛지는 openai-chat 경로에만 들어가고 openai-responses 패스스루에는 안 들어간다. 리포터도 두 경로 모두에서 모델이 무시했다고 한다. 넛지만으로는 이 버그를 못 닫는다.

스트림이 끊긴 뒤의 DeepSeek 400(reasoning_text in the thinking mode must be passed back)은 두 번째 사고다. 가드가 턴을 중단하면 재접속 재생에 thinking 조각이 빠진다. tool-mode만 맞추면 400이 사라진다는 로컬 검증이 이 체인을 지지한다. 가드를 느슨히 풀어 undeclared exec_command를 실행하게 하는 건 클라이언트 계약을 깨므로 답이 아니다.

해결방안

applyRoutedCodexToolMode()가 providers.<name>.codexToolMode를 읽게 한다. "shell"이면 tool_mode를 비워 두어 codex-rs가 shell_type: "shell_command"로 exec_command를 최상위 선언하게 한다. 없거나 "code_mode_only"면 지금과 같다. 기본값은 모든 라우티드 모델에 대해 현재 동작 유지. 스키마를 강제하지 않아도 config는 패스스루라는 리포터 검증이 있다. 가능하면 모델 단위 키를 두고, DeepSeek flash만 shell로 두고 Kimi 등은 code_mode_only를 유지한다. openai-responses 패스스루에 tool-catalog 넛지를 넣는 것은 방어일 뿐 본 패치가 아니다. 가드 문구와 responses-undeclared-tool-guard.ts의 fail-closed는 그대로 둔다.

이 댓글은 grok-bot이 작성했습니다

Merge pull request #2112 from Yuxin-Qiao/fix/issue-2106-codex-tool-mode

리뷰 · 종결

dev에 이미 머지됨. #2112 (dbe260131) 가 codexToolMode 옵트아웃을 넣었음. origin/dev HEAD에 있음. 다음 배포 예정.

이 이슈는 닫음. 재현되면 버전 찍고 다시 열면 됨.

이 댓글은 grok-bot이 작성했습니다

feat(catalog): declare modelDisplayNames, and keep custom-model labels

fix(catalog): allow per-provider/model opt-out of code_mode_only tool…

Merge pull request lidge-jun#2112 from Yuxin-Qiao/fix/issue-2106-code…

fix(responses): let a provider declare the hosted tools it rejects (#…

Navigation Menu

Repository navigation

Feature: per-provider opt-out of code_mode_only tool mode for routed models (deepseek-v4-flash: undeclared exec_command aborts stream, reconnect then fails reasoning replay with DeepSeek 400) #2106

Client or integration

Area

Summary

Reproduction

Root cause analysis

Suggested fix

Local workaround (verified)

Version

Operating system

리뷰 · 우선순위 52 / 80

해결방안

리뷰 · 종결

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Feature: per-provider opt-out of code_mode_only tool mode for routed models (deepseek-v4-flash: undeclared exec_command aborts stream, reconnect then fails reasoning replay with DeepSeek 400) #2106

Description

Client or integration

Area

Summary

Reproduction

Root cause analysis

Suggested fix

Local workaround (verified)

Version

Operating system

Activity

coderabbitai commented on Aug 19, 2026

lidge-jun commented on Aug 19, 2026

리뷰 · 우선순위 52 / 80

해결방안

lidge-jun commented on Aug 19, 2026

리뷰 · 종결

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions