A lean terminal coding agent forked from opencode and rebuilt around China-focused model routing, smart compaction, sliding output budgets, RMB cost tracking, LSP diagnostics, and Playwright browser testing.
Copy any supported document, paste it into the OpenChinaCode input box, and it gets OCR-scanned into structured text instantly.
Ask in natural language or use slash commands. OpenChinaCode routes image and video generation through Volcengine Ark with automatic download to local files.
Up to 2K resolution, multiple aspect ratios, reference-image support, and PNG/JPEG output.
4 to 15 second clips, 720p, first/last frame control, reference images, and generated audio.
/image-generate and
/video-generate with an optional prompt. Missing details trigger follow-up questions, not
guesses.
Dual-layer routing: base subagent routes plus optional extra LLM dispatch.
Profile judge + active-task extraction + raw recent tail.
Model-aware max tokens with overflow-safe triggers.
Paste any image and any model can see it — vision for all.
OpenChinaCode keeps the terminal-first opencode workflow, then narrows the default model surface to GLM, Kimi, and DeepSeek with model-aware request transforms, task routing, compaction, testing, and RMB cost display.
Architecture, complex planning, and heavy refactors favor GLM-5.2 variants for deep reasoning.
Quick review, implement, explore, summarize, and compaction subagents use fast Kimi K3 routes when the task fits.
Debug and quick exploration routes are tuned for DeepSeek speed, with sliding max-token behavior.
Plan, architecture, refactor, review, implement, explore, debug, test_fix, summarize, compaction, and visual_check each have quick/medium/complex routes.
Manual or automatic compaction keeps a general summary, extracts the active task at higher granularity,
and can retain raw recent turns with /compact keep N.
A structured document — objective, decisions with rejected alternatives, pitfalls, verified facts — is
merged by a judge at every compaction and injected into each turn, so months-long sessions stay coherent.
View with /memory.
/auto-maxtokens uses official model windows, task signals, and overflow checks so complex
coding turns can request large outputs without needless compression.
GLM, Kimi, and DeepSeek use direct official OpenAI-compatible APIs with provider-specific max-token, reasoning, sampling, and tool-call behavior.
/test-mcp, /browser-check, and /integration-test wire Playwright
MCP and test reports into the agent workflow.
The TUI shows RMB cost, model-aware context usage, route details, LSP diagnostics, and compaction debug stages so behavior is inspectable.