Multilingual Triggers

Output language: Respond automatically in the user's conversation language.

C31 Context Engineering

The core pain point of the C31 system: a 200K context window, but with poor information organization = effective capacity of only ~30K.

This is not a capacity problem — it is an information structure problem.

Core Principles

1. Layered loading — Not all information is worth loading
2. On-demand retrieval — GBrain searches only when needed, not on every call
3. Active discarding — Expired, completed, or irrelevant information is cleaned up promptly
4. Anti-rationalization — Do not retain information "just in case"

Information Layering Architecture (Four-Layer Model)

Hard rule: L4 is never automatically injected into context. It must be explicitly retrieved via GBrain search or memory_search.

Rules Files Management

On-Demand Loading Strategy

AGENTS.md uses modular splits (core.md, decision-boundary.md, communication.md, workflow.md).

Loading rules:

• Every session → load core.md (always)
• Involves decisions/permissions → load decision-boundary.md
• Involves message format/platform → load communication.md
• Involves scheduling/cron/heartbeat → load workflow.md
• Sub-files not relevant to the current task are not loaded

SKILL Trigger Management

• Read the corresponding SKILL.md only when user input matches a trigger phrase
• Do not "preload all skills"
• Subagents only receive the SKILL.md content they actually need

Context Packing

Packing Principles

1. Summary over full text — 10 pages of conversation → 3-line summary + key conclusions
2. Structured over narrative — Use tables, lists, YAML frontmatter instead of long paragraphs
3. Reference over copy — "See PLAN.md Phase 3" is better than copying Phase 3 in full
4. State over history — "Current state: X" is better than "how we got from Y to X"

Packing Operations

At session start:

1. Read session_state.json (if present)
2. Read active project files (PROJECT.md + STATE.md + current PLAN.md)
3. Assess context usage → if >35%, skip non-critical files
4. If >50%, ask user whether to /clear

On task switch:

1. Save current task state to STATE.md
2. Unload old task's L2 files
3. Load new task's L2 files
4. Retain L1 immediate layer (conversation continuity)

Information Discard Strategy (Active Forgetting)

Discard Decision Tree

Does this information affect the current decision?
├── Yes → Retain (L1/L2)
└── No → Is this information long-term knowledge?
    ├── Yes → Write to memory/solutions/ or GBrain; discard from context
    └── No → Discard immediately

Specific Discard Rules

Anti-Rationalization Table

When you're about to say "keep it just in case," consult this table:

GBrain Integration Rules

Retrieval Trigger Conditions

Must search GBrain/memory when:

1. User mentions "said before", "did before", "wasn't this already done"
2. Before solving a new problem, check if a solution already exists
3. Involves past project decisions or technology choices
4. User asks "why did I choose this"
5. Before debugging a bug, search for similar existing records

Do NOT search when:

1. User is asking about general knowledge (unless related to a historical project)
2. Current task is completely independent with no historical connection
3. Context already contains the required information

Post-Retrieval Processing

1. memory_search returns results
2. Assess relevance: directly relevant / indirectly relevant / unrelated
3. Directly relevant → extract key info, inject summary into context
4. Indirectly relevant → record reference; do not inject full text
5. Unrelated → ignore; do not retain because it "might be useful"

Context Budget Management Integration

Connects with the threshold system in AGENTS.md:

Execution Flow

At Session Start

1. Load session_state.json (restore active projects)
2. Read L3 Rules layer (core.md + on-demand sub-files)
3. Read L2 Working layer (current project STATE.md + PLAN.md)
4. Calculate context usage
5. If >35%: report status, suggest completing current task before starting a new one
6. Do not load any L4 memory (unless session_state.json explicitly requires it)

During Task Execution

1. When historical information is needed → trigger memory_search or gbrain think
2. After task completion → update STATE.md, clean up expired intermediate files in L2
3. When a solution worth recording is found → trigger compound flow

Before Session End (user says "that's it for today" / long idle period)

1. Save session_state.json (active projects, todos, pending decisions)
2. Summarize key decisions from this turn to STATE.md
3. Clean up all temporary / intermediate artifacts
4. Confirm that L4-worthy information has been compounded

Anti-Rationalization Commitment

"I commit not to hoard information out of anxiety. I trust the retrieval system. I trust that when information is needed, it will come back."

Discarding takes more courage than retaining. A clean context beats 100K of redundant information.

Trigger Phrases

• "context", "上下文", "信息组织", "变蠢"
• "context engineering", "信息太多"
• "AI怎么变笨了", "你好像忘了"
• "之前不是说过", "上下文不够了"
• "清理上下文", "精简上下文"
• "为什么你不记得", "搜索一下记忆"

Output Format

When asked to check context health:

📊 Context Health Report
Usage: X% (X / 200K tokens)
L1 Immediate: [summary]
L2 Working: [active projects/files]
L3 Rules: [loaded rules]
L4 Memory: [not loaded, retrievable]
Recommendation: [maintain / clean up / suggest new session]