openclaw - 💡(How to fix) Fix [Bug] imageModel=codex/gpt-5.4 fails with "Failed to extract accountId from token"; media-understanding also rejects inbound images with "Model does not support images" [1 participants]

Bug Description

Inbound image understanding is broken on this setup due to two observable failures on the OpenClaw image path:

1. media-understanding rejects upstream

   • [media-understanding] image: failed (0/1) reason=Model does not support images

2. explicit image tool call fails on imageModel

   • Image model failed (codex/gpt-5.4): Failed to extract accountId from token

Net effect:

• text chat works
• images visibly arrive in Discord
• the agent can acknowledge that an image was received
• but it cannot read screenshot text or describe image contents

Environment

• OpenClaw: 2026.4.21 (f788c88)
• macOS: 15.7.4 (24G517)
• Channel: Discord
• Session type: direct DM
• Auth: OAuth

Relevant config:

{
  "primary": "openai-codex/gpt-5.4",
  "imageModel": "codex/gpt-5.4",
  "fallbacks": [
    "gk-chat/gemma-4-26b",
    "gk-chat/gpt-5.4-share",
    "gk-chat/qwen3.5-122b-self"
  ]
}

Important distinction:

• working text model: openai-codex/gpt-5.4
• failing image model: codex/gpt-5.4

Steps to Reproduce

1. Set:
   • primary = openai-codex/gpt-5.4
   • imageModel = codex/gpt-5.4
2. Start an agent in a Discord DM.
3. Send an image containing readable text (for example, a screenshot).
4. Ask the agent to read or describe the image text.

Expected Behavior

Either:

• OpenClaw should successfully process the image and allow screenshot/image text recognition

or:

• if imageModel=codex/gpt-5.4 is unsupported on this auth/runtime path, OpenClaw should fail fast with an explicit configuration/runtime error instead of failing later with Failed to extract accountId from token

Actual Behavior

• inbound image is received
• auto media-understanding rejects with Model does not support images
• explicit image tool call fails with Failed to extract accountId from token
• result: the agent knows an image was sent, but cannot read it

Logs / Evidence

2026-04-22T11:07:44.231+08:00 [media-understanding] image: failed (0/1) reason=Model does not support images
2026-04-22T11:07:54.129+08:00 [tools] image failed: Image model failed (codex/gpt-5.4): Failed to extract accountId from token

2026-04-22T14:46:02.775+08:00 [media-understanding] image: failed (0/1) reason=Model does not support images
2026-04-22T14:46:26.438+08:00 [tools] image failed: Image model failed (codex/gpt-5.4): Failed to extract accountId from token

2026-04-22T18:09:38.705+08:00 [media-understanding] image: failed (0/1) reason=Model does not support images
2026-04-22T18:09:57.031+08:00 [tools] image failed: Image model failed (codex/gpt-5.4): Failed to extract accountId from token

Related supporting pattern:

[agent/embedded] ... provider=codex error=Failed to extract accountId from token
[model-fallback/decision] ... candidate=codex/gpt-5.4 ... detail=Failed to extract accountId from token

Relation to Existing Issues

Related history:

• #65431 — [Bug] image tool: Model does not support images despite input: ["text", "image"] in config (OPEN)
• #23452 — [Bug] Vision/Image Recognition Broken Across Multiple Channels (CLOSED)
• #27793 — [Bug]: image tool fails on 2026.2.24 with openai-codex/gpt-5.3-codex: {"detail":"Instructions are required"} (CLOSED)

This looks like the same image/vision failure family resurfacing on a new route:

• upstream media-understanding rejection (Model does not support images)
• plus codex/gpt-5.4 imageModel failure with Failed to extract accountId from token

Prior reports #23452 / #27793 were closed, but the pattern recurs here, and #65431 appears to be the current open tracking issue for the upstream image rejection side.