openclaw - 💡(How to fix) Fix [Bug]: DeepInfra provider not functional across OpenClaw versions (.23 and .29) [1 comments, 2 participants]

Bug type

Crash (process/app exits or hangs)

Beta release blocker

No

Summary

DeepInfra provider fails to discover or route models correctly; channel modules missing in .29, model discovery broken in .23.

Steps to reproduce

For .29 (channel modules missing):

1. Install or upgrade to OpenClaw 2026.4.29 on Oracle Ubuntu ARM64 (linux 6.8.0-1047-oracle, node 24.14.1)

2. Run openclaw status

3. Observe channels section shows empty — Telegram, Discord, Slack, Nostr all fail to load with errors like:
   failed to load bundled channel telegram: Cannot find module 'grammy' failed to load bundled channel discord: Cannot find module
   'discord-api-types/v10' failed to load bundled channel slack: Cannot find module '@slack/web-api'
   failed to load bundled channel nostr: Cannot find module 'nostr-tools'

4. Rebooting the gateway does not resolve the issue.

For .23 (DeepInfra model discovery):

1. Install OpenClaw 2026.4.23
2. Configure DeepInfra provider with baseUrl https://api.deepinfra.com/v1/openai, api openai-completions, and valid API key
3. Add DeepInfra auth profile via auth-profiles.json with token
4. Run openclaw models list
5. Observe only minimax-portal/MiniMax-M2.7 and minimax-portal/MiniMax-M2.7-highspeed appear — no DeepInfra models are
   listed despite API key being valid
6. Direct API test confirms the key works: curl -s "https://api.deepinfra.com/v1/openai/chat/completions" with the key returns
   valid responses for deepseek-ai/DeepSeek-V4-Flash
7. The LCM (Lossless Context Management compaction) system also fails to use
   configured DeepInfra models — logs show deepinfra/moonshotai/Kimi-K2.6 (override) regardless of config, because the env file (gateway.systemd.env) was apparently cached with an old value

Expected behavior

For .29: All bundled channel modules (grammy for Telegram, discord.js for Discord, etc.) should be present and loadable so channels work out of the box.

For .23: DeepInfra models should appear in openclaw models list when properly configured. The provider uses OpenAI-compatible API
(https://api.deepinfra.com/v1/openai/chat/completions) which should work with the openai-completions api type.

Actual behavior

For .29: Channels fail to initialize. Telegram is completely non-functional. This is a regression from .23 where channels worked.

For .23: DeepInfra provider is configured but invisible to the model
discovery system. Models like deepseek-ai/DeepSeek-V4-Flash work via direct API calls but don't appear in OpenClaw's model list or routing. The LCM model override (LCM_SUMMARY_MODEL env var) was being read from
gateway.systemd.env with a stale value (deepinfra/moonshotai/Kimi-K2.6) that took precedence over the config file.

OpenClaw version

2026.4.23 and 2026.4.29

Operating system

Oracle Ubuntu 24.04 ARM64 (linux 6.8.0-1047-oracle)

Install method

npm global (npm install -g openclaw@<version>)

Model

DeepSeek-V4-Flash (via DeepInfra), MiniMax-M2.7 (via minimax-portal — working)

Provider / routing chain

Working: openclaw → minimax-portal → MiniMax-M2.7 - Broken: openclaw → deepinfra → deepseek-ai/DeepSeek-V4-Flash (direct API works, OpenClaw integration broken)

Additional provider/model setup details

Config in ~/.openclaw/openclaw.json:

"models": {
  "providers": {
 "deepinfra": { 
   "baseUrl": "https://api.deepinfra.com/v1/openai",
   "api": "openai-completions",                     
   "authHeader": true,
   "models": []                                                               
 }
  }                                                                           
}                                                         

Auth profile in ~/.openclaw/auth-profiles.json:
{
  "deepinfra:default": {
    "provider": "deepinfra",
    "mode": "api_key",      
    "credentials": { "token": "<valid-key>" }
  }                                          
}  
 
The DeepInfra API key is valid — tested directly:
curl -s "https://api.deepinfra.com/v1/openai/chat/completions" \              
  -H "Authorization: Bearer <key>" \                            
  -H "Content-Type: application/json" \                                       
  -d '{"model": "deepseek-ai/DeepSeek-V4-Flash", "messages": [{"role": "user",
 "content": "hi"}], "max_tokens": 10}'                                        
# Returns: "Hello! How can I help you today?" ✅

### Logs, screenshots, and evidence

```shell
openclaw status on .29 (channels missing):                
Channels
┌──────────┬─────────┬────────┬────────────────────────────────┐
│ Channel  │ Enabled │ State  │ Detail                          │
├──────────┼─────────┼────────┼────────────────────────────────┤
└──────────┴─────────┴────────┴────────────────────────────────┘              
 
Errors from openclaw status:                                                  
[channels] failed to load bundled channel telegram: Cannot find module
'grammy'                                                              
[channels] failed to load bundled channel discord: Cannot find module
'discord-api-types/v10'
[channels] failed to load bundled channel slack: Cannot find module           
'@slack/web-api'
[channels] failed to load bundled channel nostr: Cannot find module           
'nostr-tools'                                             

openclaw models list on .23 (no DeepInfra models):
Model                                      Input      Ctx      Local Auth
Tags                                                                     
minimax-portal/MiniMax-M2.7                text       195k     no    yes      
default,configured,alias:minimax-m2.7
minimax-portal/MiniMax-M2.7-highspeed      text       195k     no    yes      
configured,alias:minimax-m2.7-highspeed                   
                                                                              
openclaw models list --provider deepinfra on .23:         
No models found.

Gateway log showing LCM using wrong model on .23:
[lcm] Compaction summarization model: deepinfra/moonshotai/Kimi-K2.6
(override)

The LCM model was resolved from LCM_SUMMARY_MODEL in
/home/ubuntu/.openclaw/gateway.systemd.env, which had a stale value. After
updating the env file to LCM_SUMMARY_MODEL=minimax-portal/MiniMax-M2.7 and
restarting the systemd service, LCM correctly shows:
[lcm] Compaction summarization model: minimax-portal/MiniMax-M2.7 (override)

Impact and severity

Affected users/systems/channels:

• Users on OpenClaw .29 who rely on Telegram, Discord, Slack, or Nostr channels — all channels fail to load
• Users on .23 who want to use DeepInfra as a model provider — DeepInfra models don't appear in model discovery

Severity:

• .29: Critical — all channel-based communication is broken, agent cannot receive or send messages via any channel
• .23: Annoying — DeepInfra models work via direct API but OpenClaw integration is broken, forcing users to use alternative providers

Frequency:

• .29: 100% — all channel loads fail on clean install
• .23: 100% — DeepInfra model discovery consistently returns no models

Consequence:

• .29: Agent is completely non-functional for multi-channel setups; users must downgrade to .23
• .23: DeepInfra cannot be used as a model provider despite having valid API credentials and working API access

Additional information

Regression note:

• Working version for channels: 2026.4.23
• Broken version for channels: 2026.4.29
• DeepInfra model discovery was apparently never functional in the tested versions (.23 and .29), not a regression

Workaround for .29 channels: Downgrade to 2026.4.23 where channel modules load correctly

Workaround for .23 DeepInfra:

1. Use minimax-portal/MiniMax-M2.7 instead (functional)
2. Or use OpenRouter which works but is more expensive
3. The LCM model issue can be resolved by ensuring
   LCM_SUMMARY_MODEL=minimax-portal/MiniMax-M2.7 in gateway.systemd.env and
   restarting via systemctl --user restart openclaw-gateway

Provider API compatibility:

• DeepInfra uses OpenAI-compatible endpoint:
  https://api.deepinfra.com/v1/openai/chat/completions
• Model IDs on DeepInfra follow the pattern deepseek-ai/DeepSeek-V4-Flash (no prefix needed when calling the openai-compatible endpoint)
• Direct curl tests confirm API key is valid and models like
  deepseek-ai/DeepSeek-V4-Flash are available

Last known good: 2026.4.23 (for channel functionality)
First known bad (channels): 2026.4.29
DeepInfra model discovery: NOT_ENOUGH_INFO — may have never worked in these versions