Plugins

Runs the upstream Headroom compression toolkit (headroom-ai, Apache-2.0) inside DeepSeek Harness over MCP: headroom_compress, headroom_retrieve and headroom_stats are exposed as mcp__headroom__* tools backed by the upstream server, installed on demand by uvx. Compression is lossy but reversible — originals are fetched back by hash from a local CCR store. Also ships a registrar that adds DeepSeek Harness to Headroom's own MCP registry, and a model-header sanitizer for the upstream proxy.

Session deletion for DeepSeek Harness: a sidebar menu item, a workspace-grouped batch cleanup panel and a session_delete tool (dry-run by default). Deletes session logs, projection cache, workspace-registry references and derived search indexes, and cascades to all subagent sessions.

Chunked compaction for oversized sessions: takes over only when the live context exceeds 85% of the model window (or the request exceeds 4 MB), summarizes the history in 1.5 MB chunks and merges them; smaller sessions keep the built-in compaction.

Delete sessions for real from the dsh sidebar (local logs and projection cache included, subagent and fork children cascade) and manage workspaces and ungrouped sessions, with an optional daily cleanup of ungrouped sessions.

Adds a permanent Delete session action to the sidebar session menu, behind a confirmation dialog that reports the log size, running turns, subagents, background jobs and reminders bound to the session; the id is then released from the live, archive and workspace registries, so the row stays gone on refresh instead of folding back in.

Turn rail on the left edge of the conversation: one tick per user message, hover to preview, click to jump (auto-loads earlier pages when the turn is unloaded), plus an active-turn highlight and a scroll-position indicator.

Lightweight crash-recovery for dsh: after a restart, it detects sessions whose last turn was killed mid-flight and re-injects a continuation prompt that first demands the agent verify real-world state before acting, so side-effecting operations are not blindly repeated. It only reacts to the genuine crash marker (turn/end reason 'interrupted', synthesized on cold read by session repair), so it never misfires on user cancels, rate limits or model errors - which is why it needs no error classification, backoff or configuration UI. Unlike general-purpose auto-continue plugins, it also covers the case where a subagent was interrupted mid-run: the parent turn has already ended normally while it waits, so a naive check sees nothing to resume and the whole subagent tree dies silently; this plugin detects the interrupted child and wakes the parent instead. Single host file, zero external dependencies, seven config constants.

Deletes a whole conversation from the sidebar session menu, and opens a right-click menu on a chat text selection with Copy, Quote (insert a Markdown quote block into the composer) and Ask: Ask sends the selection plus your question to one reused continuable subagent (label `ask`, found again through the subagent catalog after a restart) and streams the answer into the right sidebar.

Delete a DSH conversation from its sidebar menu or the conversation header, optionally cascading into that conversation's associated memory nodes when you tick the box (the confirmation dialog shows the hit count before you commit); deletes land in a plugin recycle bin with restore, every action is appended to an audit log, and batch deletes are rate-limited with per-conversation rollback.

Prepares context summaries in the background and applies them at safe step boundaries, with native compaction fallback and a per-conversation history of summaries, application status, timings, and token usage.

Indexes session transcripts into the Akasha session layer as turns end, records successful compactions, and adds a per-turn rhythm line to the runtime context. Companions: Hindsight (repo memory), OpenViking (user memory), and the built-in session search (see repo README). Setup: clone the repo, copy core/ aside, then replace <AKASHA_DIR> in cordis.patch.yml with its absolute path (or set the AKASHA_DIR env var); plugins fail loudly while the placeholder remains.

Collects the sessions that need you into one dock at the bottom of the web GUI sidebar: running sessions, sessions waiting on an approval, plan review or answer, and finished runs whose result has not been opened yet. Rows are sorted by last update and clicking one switches the main view to that session; it reuses the shipped useSessions and useSessionStatus hooks rather than adding a data source, polls nothing, stores nothing and modifies no session. The header dot pulses while anything is running or waiting and settles to a static dot when every row is only a finished-but-unread reminder; with no such sessions the dock collapses to one quiet line, and a collapsed 56px sidebar shows a 36px badge with the count instead of the list. It registers in the sidebar footer slot, so it is part of the layout and does not overlay the conversation list, leaving the existing footer entries below it.

Registers /compact_queue, a compaction command that does not drop the request when the session is busy: it compacts immediately when idle, otherwise queues the request in that session and runs it at the session's next idle window, with a 2-second fallback tick beside the agent status event so the trigger is not missed, a 30-minute wait cap, and /compact_queue cancel to withdraw a queued request or abort a running one. Each session keeps its own queue state from a single command registration, completion and failure arrive as a collapsed notice that does not wake the model, and every ManualCompactionError code except busy is terminal so a failure is not retried. Host half only; the core /compact command is left untouched.

Persona manager for DeepSeek Harness — a JSON-backed persona library, a per-session binding so each session can run a different persona, an optional default persona for sessions that have no binding, and switching that applies either at once or from the next session; the Web UI adds a Model personas settings page and a persona chip next to the composer, and three model tools let the model list, switch and create personas itself.

Mounts a persona kept in Markdown as a system-prompt section, so editing ~/.dsh/persona/*.md changes how the agent speaks; it reloads 150ms after a save instead of requiring a restart, and project-level .dsh/persona/ files override the global ones.

Describe the persona you want and the agent writes it: nickname, bio, taglines and links come from the host model, the avatar comes from your local asset library (or a path you give), then it fills your profile page and hands the picture to the site's own uploader. Dry-run by default - it only touches profile fields, never passwords or security settings.

Connect OpenAI Codex app-server as a DSH main model with ChatGPT sign-in, HTTP transport, automatic official Codex latest runtime updates, on-demand live Codex model catalog, native tools, settings, selectable questions and automatic chat replies, on-demand local and connector skills, Windows Computer Use recovery, context management, quota and Token usage history.

Answers model calls by driving your own logged-in chatgpt.com session in a local Chrome, so inference spends the ChatGPT web quota rather than Codex quota, and adds a JSON tool-call protocol so the web model can run DSH tools.

Native multi-platform accounts and model routing with staggered check-ins, event notifications, and quota forecasts. Its background worker needs a standalone Node >=22.23.3 with node:sqlite (enforced by engines.node); Qwen Office coding additionally needs Python with cryptography and wasmtime.

Adds per-member model and reasoning-effort routing to DeepSeek Desktop Agent Teams, plus leader-initiated release of a teammate subagent run instance.

Exposes the accounts in a local CC Switch database to DeepSeek Harness: every account (Claude Code, Codex, OpenCode, Grok Build) becomes a model category in the picker, and the SQLite database is re-read on every model listing and request, so switching accounts in CC Switch needs no DSH restart. Requires CC Switch and the DSH desktop app.

Tiers free-channel models into three 1M fallback chains plus a custom chain; orders them by measured throughput, drops models below throughput or error-rate thresholds, and rotates accounts. Requires an OpenAI-compatible gateway (default :8326).

LocalAI provider for language models, embeddings, reranking, speech, and image generation with LocalAI-specific controls in DSH. Requires Firefox2100/dsh-ai-providers#runtime to be installed and enabled.

Jina provider for embeddings and reranking in DSH. Requires Firefox2100/dsh-ai-providers#runtime to be installed and enabled.