CLI hangs after completed generation run; usage shows completed but no assistant response

Commencez par les points d’entrée de génération utilisés par commandcode -p et par le commandcode --model interactif, puis comparez leur gestion des réponses en streaming avec les paramètres documentés du proxy HTTP. Reproduisez le problème avec le prompt Reply with exactly the word TEST et examinez les fichiers d’historique JSONL. C’est terminé lorsqu’une exécution terminée du dashboard affiche la réponse de l’assistant et écrit localement une entrée de l’assistant.

Summary

Command Code CLI sends requests successfully and the usage dashboard shows completed runs/token usage, but the CLI never displays or saves the assistant response. The request appears to hang indefinitely in both Windows PowerShell and WSL.

Environment

• Command Code CLI: 0.32.3
• Account shown by commandcode status: rikitikitavi2012-debug
• Windows: PowerShell, installed via npm global wrapper in C:\Users\User\AppData\Roaming\npm
• WSL: Ubuntu 24.04 on WSL2
• Node tested:
  • Windows Node from C:\Program Files\nodejs\node.exe
  • WSL Node v24.12.0
  • WSL Node v22.22.1

Proxy setup

The issue reproduces with the normal proxy setup used on this machine:

• Windows:
  • SOCKS tunnel: ssh -D 1080 ...
  • HTTP bridge: http-proxy-to-socks listening on 127.0.0.1:8001
  • CLI tested with:
    • HTTP_PROXY=http://127.0.0.1:8001
    • HTTPS_PROXY=http://127.0.0.1:8001
• WSL:
  • HTTP_PROXY=http://127.0.0.1:8118
  • HTTPS_PROXY=http://127.0.0.1:8118
  • privoxy forwards to SOCKS5 127.0.0.1:1080

What works

These commands work in both Windows and WSL:

commandcode status
commandcode --list-models
commandcode info

The dashboard also shows completed runs and consumed credits/tokens after attempted prompts. Example from the dashboard:

• Status: COMPLETED
• Model: moonshotai/Kimi-K2.5
• Mode: agent
• Input/output tokens are counted
• Cost is consumed

Manual unauthenticated POST checks through the proxy also reach the API and return the expected 401 response:

curl -X POST https://api.commandcode.ai/alpha/generate \
  -H 'content-type: application/json' \
  --data '{}'

Response:

{"success":false,"error":{"code":"UNAUTHORIZED","status":401,"message":"Invalid 'Authorization' header or token."}}

What fails

Generation hangs and produces no visible output:

commandcode -p --model moonshotai/Kimi-K2.5 "Reply with exactly the word TEST"

Also reproduced with:

commandcode --model moonshotai/Kimi-K2.5

Then sending a first message such as:

привет

The UI shows a running state such as reviewing/mapping/ruminating, but no assistant response appears.

Local history result

After a failed interactive prompt, local Command Code history contains only the user message and no assistant message.

Windows example:

C:\Users\User\.commandcode\projects\c-users-user\417de20a-54a8-410f-9cc1-595140c811fc.jsonl

Contains:

{"role":"user","content":[{"type":"text","text":"привет"}]}

No role: "assistant" entry appears.

WSL showed the same behavior in ~/.commandcode/projects/.../*.jsonl.

Expected behavior

If the dashboard shows the run as completed, the CLI should display the assistant response and save it to local history.

Actual behavior

The backend appears to complete the run, but the CLI never receives/displays/saves the final assistant response.

Notes

This does not look like a simple auth/account/model issue because status, --list-models, and the usage dashboard all work. It looks like either a CLI streaming/response handling bug or an incompatibility with the proxy chain used for long-running generation responses.

Please advise how to debug the streaming response path or whether there is a known issue/workaround for proxied environments.