Plugins

Predicts your next message after each completed turn and shows it as ghost text in the composer: press → to accept, or keep typing and finish it word by word with Ctrl+→, Esc to dismiss. The prediction model follows the session by default or can be picked in the settings card, which also has an on/off switch and a speed test.

Codex-style paste collapse for the DeepSeek Harness composer: a paste of more than 1000 characters becomes a `[Pasted Content N chars]` placeholder, a second paste of the same size is numbered `#2`, and the host expands it back to the full text before the model reads it, so the model still receives every character. A pill on the top edge of the composer shows the preview, the sent message shows the same pill, and clicking either opens the whole paste in a side panel with a copy button. Pure in memory: nothing is written to disk and no token is saved.

Adds a session-scoped context window selector next to the reasoning-effort row in the DeepSeek Harness model popup: a pill showing the effective window opens a dropdown of presets below the model default, the choice applies to the next request of that session only, and a new session falls back to the profile default. It writes no config file and appends nothing to the session log.

Codex-style reasoning-effort card for the composer: takes over the model selector and shows the current level, the model row and a segmented slider over the model's own adapter-provided levels, with a model list, a reset, and a purple aurora with twinkling particles at the maximum level.

Composer model selector grouped by provider: pick a provider card, then drill into that provider to choose a model, with cross-provider search and Reasoning Effort kept.

Makes the provider groups in the composer model menu collapsible, keeping only the provider of the current model expanded.

Live preview inside the sidebar session hover card: what the session is doing right now (running tool with elapsed time, thinking, last tool duration), an attachment thumbnail with click-to-enlarge, the latest user prompt and assistant reply with injected system scaffolding stripped out, a recent-tool icon strip, turn count plus token totals plus model name on one line, and a 12-item settings page (11 switches plus a slider) persisted to the profile.

Unofficial Claude UI-inspired model and reasoning-effort selector using the current model's available levels, with animated purple pixel effects.

Switch providers from the left column and choose a model from the right; hover, focus or click a provider to update the model list.

Windows only; needs PowerShell and the bundled install.ps1. Adds an interface menu to the left of the model selector with three actions that run immediately: reload the interface, restart DSH, and close DSH. The Windows build ships no built-in reload or restart entry, so this adds one.

A temp note in the DSH right sidebar with session, workspace, and global scopes, browser-local autosave, one-click copy, and one-click LLM polish through the default model.

Adds a sticky effort slider inside the composer model menu to select reasoning levels, with generic stops for custom models.

Links DeepSeek Harness to VS Code: approvals answered in native notifications, ask_user_question answered in a sidebar card, an ide_context tool fed by the active file, diagnostics and selection, selection to composer, and a sidebar with the live transcript, workspace-scoped session picker, model and permission controls.

Model quick picker for DeepSeek Harness: a button left of the model dropdown opens a full model-selection dialog with provider grouping, live search, provider filter, capability tags read from the user's own input config, reasoning badges, recent-model shortcuts and a reasoning-effort picker; the choice writes back to the native dropdown immediately.

Adds a ✨ button and an /enhance command to the composer that rewrite the draft into a structured prompt, flag open questions below the input, follow the current session model, and offer an undo.

Responsive card grid for the DeepSeek Harness composer model picker, grouped by provider. Host-only: injects a stylesheet through webserver/index-inject without rebuilding the client bundle, and supports both the stock picker and the dsh-peakrate menu.

DSH guard against self-repetition in streamed model output: detects a degeneration loop in reasoning or visible text, aborts the runaway turn and wakes the agent with a corrective notice.

A social deduction game played against your own model configurations. A table has N seats (six by default, one of them yours) and some of the rest are played by agents bound to different provider, model, temperature, system prompt and disguise settings; you interrogate them and cross-check the answers to work out which seats are not human. Seat identity and each agent model configuration stay hidden in the host state until a seat is judged, then the disguise parameters are revealed. Every round exports a per-seat conversation JSONL plus the judgment record, and the same seed with the same move script replays byte-identical, so runs with different models or disguise strengths can be diffed.

Two games inside DeepSeek Harness where the model supplies intent and a local engine owns the rules: a twin-brain snake that periodically draws a strategy card, and Chinese chess where the model only comments on a move picked from locally generated legal candidates. Both keep running when the model is slow, offline or out of quota.

Connects the DSHM Android client to computer-hosted DeepSeek Harness sessions for remote chat, model switching, and approvals with device pairing and revocation.

Deterministic prompt-injection interception: tool results from file reads, repo searches and web fetches are scanned before they enter model context; high-severity findings quarantine-wrap or block, and credential-material files quarantine on the path alone.

System-prompt surface drift gate: every section any plugin or preset contributes to the assembled system prompt is hashed and baselined; a silent rewrite of what the model is told fails loud.

Context firewall for DeepSeek Harness: credentials in successful plain-text tool results become opaque ⟦cloak:category:n⟧ placeholders before they reach the model, with originals kept only in a process-memory, agent-scoped vault. 36 built-in signature families (shared with the cloakpack git pre-push guard), key-name rules for unknown formats, custom regex rules, dry-run mode, an audit log, and the /cloak command.

Watches a turn's thinking and output streams and cancels the turn when the model starts repeating itself. Four detectors: a verbatim cycle, a counting loop whose lines differ only by a number, repeated lines separated by other text, and a user-supplied phrase list. A catch raises a manual-dismiss alarm card in the global overlay with an optional tone, and a Monitor tab holds the switches, thresholds and hit history. Detection is local string comparison with no model calls, so it costs no tokens.