Plugins

Vision for text-only DeepSeek via Doubao Web by default (zero-cost, no API key — drives your logged-in Chrome through a Windows CDP bridge), with Antigravity IDE quota (flash/pro) or Gemini fallback; auto detail escalation, vision evidence memory with compaction rehydration, content-hash cache, and a bilingual client panel.

Configurable DSH usage statistics: a sidebar text/card widget plus a detail dialog with per-provider and per-model breakdowns, cache-hit rates, token trends, and latency details (TTFT and duration).

Per-conversation x-opencode-session header injection for OpenCode Go in DeepSeek Harness (DSH). Fixes 400 MissingSessionID, keeps prompt-cache routing optimal, scoped strictly to opencode.ai, runtime-switchable without restart.

A slim port of Headroom: compresses tool output, cuts up to 20% of context, without affecting the model's context cache or agent performance.

Performance profiling and LLM cache diagnostics for DeepSeek Harness — context engineering, latency profiling, and cache behavior reports.

Local OCR fallback for text-only routes: when the session model declares it cannot accept images, the attached image is cached locally and its path injected so the model can call ocr_image — PP-OCRv5 + ONNX Runtime on CPU, no API key, images never leave the machine. Silent when the model can see images.

Cost dashboard for DeepSeek Harness: aggregates model input/output/cache tokens across all local sessions, prices them with a builtin table (including DeepSeek peak/off-peak), and shows daily trends plus per-model and per-session tables.

Environment housekeeper for the agent: toolchain inventory (node/pnpm/git/gh/ffmpeg/browsers), scratch/cache scan with whitelist-guarded one-click cleanup, and the machine rules file (AGENTS.md) editor - all in the Web GUI settings.

Right-hand widget rail for the DSH Web UI: live session stats (turns, LLM/tool time, TTFT, speed, cache, tokens) and OpenCode Go quota via a same-origin host proxy; extensible widget registry for more platform usage widgets, visuals, utility actions and external integrations.

Activity heatmap in the DSH Web sidebar: GitHub-style grid of daily commits, token usage, and estimated spend, with a today stats line for all-session token totals, cache hit rate, and per-model auto-priced cost.

Lazy MCP gateway for DeepSeek Harness: one stable proxy tool instead of N tool schemas, servers connect on first use and idle out, metadata cached to disk.

DSH Desktop enhancement plugin: it puts verifiable facts on the table. Constraints are injected only when relevant, so idle chat costs no extra tokens; injection is layered (session-stable text in the system prompt, volatile content as a tail snapshot) and the prompt cache stays valid. Discipline: each turn is classified (question / research / discussion / diagnosis / execution) with boundaries enforced - no file edits on question turns, no unsolicited fixes on diagnosis turns; citing code lines not opened this session, negative claims about unseen symbols, requirements that do not match the deliverable, and dressing up “I did not check” as “waiting for your decision” are all called out; replies are written for humans - a code, field name or id must be explained in plain words the first time it appears. Memory and knowledge: 75%/90% context warnings plus an exported session recap (say "continue" in a new session); project knowledge accumulated across sessions by working directory, auto-captured and deletable by id. Inspectable artifacts: task contract / change record (auto-logged, unverified items flagged at delivery) / hypothesis record (conclusions must pass an evidence gate) / design decisions. Dashboard and personas: routing, trigger hits and block assembly are written to lume-metrics.jsonl and queryable via lume_metrics; personas are distilled from chat logs, novels, scripts or character sheets, with long-term and temporary memory, a memory star map and persona card export/import.

Shows every session theme in the DSH Web left sidebar immediately on page load by warming the session projection cache at startup; no need to click into a conversation first.

Full-type session lifecycle management: one-shot subagents archived on completion, idle continuable subagents / main sessions archived (recoverable), a capacity cap, projection-cache cleanup, and a hot-reload settings panel.

Weak-network adaptor: transparent long-backoff model-stream retries (10-min cap), local response replay cache, heartbeat auto-reconnect, and degraded-mode token economy, with a bilingual (zh/en) settings page.

Settings usage dashboard: tokens, requests, output tokens, and cache-hit rate from session logs, with a stacked chart and a breakdown table by provider, model, or workspace.

Shows DeepSeek balance and estimated daily spend, merges external vision-call JSONL records, and adds a per-session dashboard with model statistics, TTFT, cache rate, call logs, cost estimates, filters, and CSV export.

Native LLM-provider vision bridge: images pasted in the chat are described by a vision model (Qwen3-VL via pi-ai/llama.cpp) and the text description is fed to text-only DeepSeek for the reply — image admission, routing and compaction all run through harness-native mechanisms, with an LRU description cache and 503 retry.

Model pricing for the DSH settings page: ~7,250 models / 213 providers with the harness catalog overlaid, honest lowest-listed comparison, session costs priced by the actual route, cache-leak attribution, and a verified promotion feed.

Prompt-cache diagnostics dashboard showing provider-reported token usage, adjacent logical-input changes, and Call TTFT.

Version-token-validated in-process result cache for the DSH read tool: repeated reads of unchanged files skip file I/O and line rendering, with a readcache tool for stats and clearing.

Composer-dock balance/usage readout for DeepSeek, Kimi Coding, OpenCode Go, Zhipu, MiniMax, OpenRouter and OpenAI Codex — provider-based, 2s polling, 5min cache.

Vision toolkit for text-only DeepSeek: model-invokable `vision` tool, wrapper adapters for deepseek/opencode-go (v4 flash/pro), Antigravity IDE quota (default, flash/pro) / any OpenAI-compatible VLM / Gemini / local Ollama channels, evidence memory with compaction rehydration, content-hash cache, and a bilingual client panel.

Renders the built-in cache-hit percentage with three decimals and widens the composer stats line to prevent truncation.