Plugins

Model sampling sliders (temperature / maxTokens) in the composer: override every model request for all providers via the agent/request hook, with hot/persist modes.

Right-side file-tree sidebar that opens workspace files, plus a DeepSeek account balance meter in the composer stats line.

Clippy is back and this time he can actually help. An office assistant pet that reacts to real agent state, jumps when the turn completes, and opens a classic "illegal operation" dialog when a turn fails.

Session cost estimate in the DSH Web composer dock (tokenUsage × configurable price table, one-click official-price refresh).

Select text in a conversation, then quote it into the composer or reply in a new window.

Drive9 workspace and durable evidence components for DeepSeek Harness

Conversation-style UI/UX visual review plugin for DeepSeek Harness: vision_review / vision_ask tools backed by any OpenAI-compatible multimodal endpoint (e.g. agnes-2.5-flash)

Local AI code-review gate: review a workspace/branch/commit diff via the managed ocr.js, returning structured findings with token stats; gate mode fails on findings.

Take control of your tmux panes: list/send-keys/capture, run long jobs in a pane with watch mode, and approval-gated destructive commands.

Code knowledge graph bridge for codebase-memory-mcp: semantic symbol search (BM25), source snippets, architecture overview with Leiden communities, call/data-flow/cross-service traces, graph-augmented grep.

Advanced memoria tools: session context injection (profile + recall), recent decisions, health report, namespace allowlist, memory relation graph, and entity search; companion to dsh-memoria.

Post-compaction premise-drift guard: injects a one-shot notice when a compaction summary drops a critical literal anchor.

Bridge into a global Markdown knowledge base shared with the Codex kb.cmd CLI: kb_add/kb_search/kb_show/kb_timeline tools with byte-compatible frontmatter.

File-backed working memory: memorize/recall key premises verbatim in a session notes file so they survive context compaction losslessly.