🧠 Language Anchoring — A Systematic Method for LLM Multilingual Adaptation

Key finding: The first active output from an LLM determines the thinking language for the entire session. Reading a language ≠ thinking in it.

Stage	Compliance	Intervention
Prompt constraints only (4 language rules)	~70%	Prompt text changes only
+ Translate 16 tool descriptions	~85%	Tool layer language alignment
+ Translate 50+ source file comments (fresh session)	~95%+	Code layer reinforcement
+ Anchoring instruction (7 lines)	~100%	Leverages autoregressive inertia

The anchoring instruction is minimal:

## Language Anchoring (Hard Requirement)
After receiving the user's first message, before executing any action or starting reasoning, you MUST first write a brief summary in the target language describing:
1. What the user is asking you to do
2. What steps you plan to take
3. What information is still missing
This output is your first response — it anchors the thinking language for the entire session.

Related academic work: Contextual Inertia (arXiv 2603.04783), Cognitive Inertia (arXiv 2503.01307) — these study "how to break harmful inertia". Our contribution is the positive utilization of the same mechanism.

📖 Discussion #1 — Language Anchoring RFC
💬 Discussion #2 — AI合作、信任、螺旋进化（中文）
🌐 Discussion #3 — How I Work with AI (English)

About This Fork

This is a personal fork of anomalyco/opencode. All fixes, optimizations, and features are implemented by AI — DeepSeek V4 Flash (thinking mode) / Sisyphus — under human supervision.

The upstream project is excellent. Windows and DeepSeek are not their priority. We handle them ourselves.

Implementation: DeepSeek V4 Flash (thinking mode) / Sisyphus (AI)
Review & direction: fkyah3 (human)
This fork is a live demonstration of what AI-built software looks like.
See fkyah3_dev/ for the full story.

What's Fixed

Fix	Area
`reasoning_content` loss on conversation replay — the root cause of API 400 errors with DeepSeek thinking mode	Core
OpenRouter `@openrouter/ai-sdk-provider` providerOptions key mismatch	Provider
Multi-window crash cascade — added circuit breaker to OMO tool-pair-validator	Stability
Per-session log isolation (no more truncated dev.log on restart)	DX
Windows CJK encoding — 3-layer GBK/UTF-8 fix for subprocess pipes	Win32
TUI plugin loading — Magic Context sidebar blank due to Zod strict mode	TUI
MC tool truncation — `[truncated]` → `[tool: N lines, MKB \| preview]` for compressed context	MC
DeepSeek V4 architecture research → config alignment: 400K context, 90% MC threshold, CSA-aligned compression	V4
Global session pool — all conversations visible from any directory (2.0)	Session
Full Chinese system prompt: Sisyphus, keyword-detector, system messages, environment info	i18n
Language Anchoring — systematic method for LLM multilingual adaptation (verified with dose-response data)	i18n

Quick Start

cd packages/opencode
bun run --conditions=browser src/index.ts

Model Configuration Reference

Flash (daily driver — all agents)

"deepseek-v4-flash": {
  "limit": { "context": 400000, "output": 393216 },
  "options": {
    "reasoningEffort": "max",
    "thinking": { "type": "enabled" }
  },
  "interleaved": { "field": "reasoning_content" }
}

reasoningEffort: high (default) or max. Per DeepSeek docs, agent tools auto-set to max anyway.
interleaved: Required for DeepSeek thinking mode via @ai-sdk/openai-compatible. Without it, reasoning_content is not forwarded.
thinking: Must be { type: "enabled" } for reasoning models.

Learn More

Link	Content
`fkyah3_dev/`	Fix details, analysis docs, issue tracking
`fkyah3_dev/internal/COMPLETION.md`	Completed work checklist & current state
Upstream issue #24104	reasoning_content root cause discussion

Upstream

Based on anomalyco/opencode (Apache 2.0). Changes merged periodically.

Crafted by AI. Curated by human. Built for the real world.