DeepSeek Harness rc.8: multimodal image inputs, Claude Code/Codex profile bundles, SQLite storage format break
Release candidate 8 adds native multimodal support (image input in /goal, /plan, etc.), on-demand Claude Code and Codex subagent profiles with non-interactive permission modes, and persistent PowerShell on Windows PTY. The SQLite storage backend is rewritten for performance — the format is incompatible, requiring a migration. A brand usage policy is formalised.
PUBLISHED2026-08-19
OBSERVED2026-08-21
AGE6d
SOURCES1
- Multimodal: DeepSeek adapter can now send native image requests;
/goal,/planand other commands accept image+text input;@menu references files and sessions. - Subagent profiles: Claude Code and Codex subagents installable on demand as Profile Bundles; Codex supports non-interactive permission modes and named instances.
- Windows PTY: persistent PowerShell sessions, enabled by default in the Minimal preset.
- Fixes: oversized images no longer break model requests; cancelled streamed generations preserve displayed response prefix in follow-ups and forks; custom OpenAI-compatible gateway request-format differences fixed; reasoning content no longer goes missing.
- UI improvements:
~abbreviation for home directories, narrow composer layout, feedback UI, sidebar search focus, workflow panel improvements, model-selector interaction, retry on local file open failure. - Tool calls:
web_searchsupports concurrent queries; subagentreportDeliverywakes parent tasks promptly. - Startup: reduced dependency download size;
dsh webopens browser automatically. - Fork performance: large-history session forks are significantly faster.
- SQLite backend: read/write/fork performance improved and storage size reduced — incompatible format (breaking change).
- Brand: "DeepSeek Harness" registered trademark; usage guidelines published.
- Python SDK: covers all four built-in Agent Presets plus
rg/glob search and MCP stdio tool dependencies.
COMMUNITY
No curated reactions recorded for this event. Facts and takes are kept in separate layers — community context is added by hand, never blended into the record above.