Description
When Honcho (plastic-labs/openclaw-honcho) is the active memory provider, all exec and memory_search tool output renders as image attachments in the conversation transcript instead of plain text. This affects both the model-side rendering and the Discord chat display.
Impact
Makes shell commands, search results, and CLI tools unusable for text-based workflows. The model can still read the content (JSON/metadata is returned correctly) but the rendered chat display shows an image attachment instead of text, making it impossible to copy/paste results or read them from the UI.
Reproduction Steps
- Run OpenClaw with plastic-labs/honcho as the active memory provider
- Execute exec() with a shell command (e.g., ls -la, echo test, python3 script.py)
- Output returns as an image attachment instead of text
- Also reproducible with memory_search() and read tools (intermittently for read)
Affected Tools
- exec() - shell command output always renders as image
- memory_search() - search result JSON renders as image
- gh CLI output via exec renders as image
- read tool - renders as image intermittently, especially for large content
- session_status - also renders as image
Not Affected
- write, edit, message tools - unaffected (likely because these produce confirmation metadata, not raw output)
Expected Behavior
Plain text rendering in the conversation transcript - the same way it works when Honcho is not the active memory provider.
Additional Context
- JSON data is returned correctly to the model's context - this is purely a rendering/presentation layer issue
- Does not affect other memory providers (e.g., built-in memory)
- First observed 2026-07-05 around 19:30 UTC
- The issue worsened over the session: small outputs (single word) rendered fine at first, then eventually even echo test returned as an image
Environment
- OpenClaw runtime (latest from git)
- Honcho memory provider (plastic-labs/openclaw-honcho)
- Discord chat surface
Description
When Honcho (plastic-labs/openclaw-honcho) is the active memory provider, all exec and memory_search tool output renders as image attachments in the conversation transcript instead of plain text. This affects both the model-side rendering and the Discord chat display.
Impact
Makes shell commands, search results, and CLI tools unusable for text-based workflows. The model can still read the content (JSON/metadata is returned correctly) but the rendered chat display shows an image attachment instead of text, making it impossible to copy/paste results or read them from the UI.
Reproduction Steps
Affected Tools
Not Affected
Expected Behavior
Plain text rendering in the conversation transcript - the same way it works when Honcho is not the active memory provider.
Additional Context
Environment