hoelee-blog/docs/linkedin/04-slashing-agent-context-with-a-code-graph.md

LinkedIn Draft & Consultant Review: Slashing Agent Context and Token Churn With an AST Code Graph

文档性质：从 blog.hoelee.com 改写的 LinkedIn 待发布技术动态。 发布状态：待顾问审阅 (未在 LinkedIn 发布) 源文章：https://blog.hoelee.com/posts/slashing-agent-context-with-a-code-graph/ 系统标识：slashing-agent-context-with-a-code-graph | 分类：ai 字数：1975 chars (LinkedIn 上限 3,000 chars)

---

一、 核心背景与铁律约束提示（供顾问审阅对照）

1. 求职画像契合（Senior 全栈/DevOps/AI 架构师）：严禁流露“做生意的老板/小公司创始人”倾向；一切商业实战均转换为“生产系统架构、高可用交付与自动化运维”视角。
2. 纯洁人设与零禁用词：绝对禁止 homelab, home lab, 家用, 小主机, 玩具, 折腾, 玩票 等词汇。必须使用 the infrastructure I manage, my self-hosted production nodes, production host。
3. 第一人称单数：统一使用单数 I / 我，杜绝虚张声势的复数 we。
4. 数据绝对保真：所有数字、性能比率、吞吐提升、故障现象严禁凭空捏造。
5. OPSEC 安全脱敏：无内部绝对路径（如 D:/dev），无真实私有内网 IP，公网 IP 必须转为文档保留地址。

---

二、 LinkedIn 拟发布动态文案

When you prompt an AI coding agent to modify a 50,000-line codebase, its default instinct is brute force: grep directories and recursively dump dozens of files into context.

The result is instant token bankruptcy:
• Reading 25 files to trace one dependency consumes 40,000+ tokens in a single turn.
• High-context saturation triggers cognitive degradation — models start hallucinating method signatures and missing subtle type contracts.

I tested three conventional fixes, and all of them failed under real development conditions:
1. Ripgrep keyword filtering: Zero semantic awareness. A search for `processOrder` matches comments, test mocks, and logging across 60 files. The agent still had to open 15 files just to find the implementation.
2. Vector RAG embeddings: Excellent for fuzzy concepts, useless for deterministic code trees. Vector similarity regularly misses internal helper functions (like `_verifyNonce`), leading agents to generate patches with nonexistent signatures.
3. File outline skeletons: Lacks parameter types and caller hierarchies, forcing speculative full-file reads anyway.

The durable fix: A pre-indexed AST Code Knowledge Graph.

Before the agent touches the code, an external parser extracts the repository's Abstract Syntax Tree into a symbol graph of callers, callees, definitions, and interface implementations.
• When the agent needs to inspect a function, it queries the graph and receives an exact 140-token node (signature, parameter types, caller list).
• Instead of dumping 3,000 lines of code across 5 files, the agent targets its read exclusively to the specific 25 lines that require mutation.

Across 20 production tickets, AST code graphs reduced agent context overhead by ~80% and eliminated 100% of signature hallucination errors.

Are you still using vector search for code intelligence, or have you transitioned to deterministic AST graphs?

Write-up: https://blog.hoelee.com/posts/slashing-agent-context-with-a-code-graph/

---

三、 外部顾问审阅评估反馈单

• 审阅结论：[ ] 通过 / [ ] 需微调 / [ ] 需重构 / [ ] 建议暂缓
• 1. 职场形象评估（是否符合高级技术雇佣人设，无老板/卖货感）：
• 2. 技术真实性与逻辑推导（因果链条与数据是否严谨）：
• 3. 语言地道度与人设去油（是否纯正地道、零业余玩具感）：
• 4. 优化修改建议（如有优化建议或修改文案，请在此批注）：