feat(prompts): inject external knowledge into codegen

[ceilf6]

Linked Issue Or Context

Closes #363.

Summary

• Reuses the existing EXTERNAL_KNOWLEDGE_PROTOCOL in code-generation prompts.
• Injects the protocol into generateCodeForFile and generateModifiedCode so executor codegen follows the same external-knowledge discipline already added to planner prompts.
• Adds focused prompt-capture tests for both codegen paths, including the reverse constraint that stable known APIs do not need web_fetch.

Impact Scope

Prompt-only change in packages/core/src/llm/code-generation.ts plus focused tests in packages/core/src/llm/code-generation.test.ts. No schema, tool routing, executor contract, planner, or web_fetch runtime changes.

GitNexus Impact Summary

• Risk level: LOW
• Critical skeleton changes: none
• GitNexus impact: impact(generateCodeForFile, upstream) and impact(generateModifiedCode, upstream) both returned LOW with direct=0, processes_affected=0. Additional call-site inspection shows the changed codegen functions are reached through LLMService thin wrappers from executor create_file/apply_patch codegen paths; signatures and return contracts are unchanged.
• Verification: gitnexus detect_changes compare against develop reported changed_count=2, changed_files=2, affected_count=0, risk_level=low; changed symbols were generateCodeForFile and generateModifiedCode only.

Verification

• pnpm agent:bootstrap
• pnpm quality:predev (initially blocked by unrelated local AGENTS.md/CLAUDE.md/.claude skill edits; passed after isolating feat(prompts): inject external-knowledge discipline into code-generation prompts (learned from CLAUDE-FABLE-5) #363 scope through quality:local)
• pnpm --filter @frontagent/core test -- code-generation.test.ts
• pnpm quality:precommit
• pnpm build:verify (rerun after one transient esbuild service stop in the first quality:local attempt)
• pnpm quality:local
• gitnexus detect_changes compare develop

Checklist

• I have linked an issue or explained why this PR stands alone.
• I have kept the diff focused on the stated change.
• I have run pnpm quality:precommit, or explained why it could not run.
• I have run pnpm quality:local for critical skeleton changes, or explained why it could not run.
• I have updated docs or tests when behavior, public APIs, or Harness contracts changed.
• For critical skeleton changes, I have filled the GitNexus impact summary with concrete results.

[github-actions]

🛡️ ceilf6/repo-guard

代码评审报告: feat(prompts): inject external knowledge into codegen

风险等级: 中
处理建议: 批准
决策摘要: ** 可以合并；变更按 Issue #363 的最小范围复用现有 EXTERNAL_KNOWLEDGE_PROTOCOL 注入两条 codegen system prompt，未改变签名、schema、工具路由或执行契约。

级联分析

• 变更符号: 原始模型未提供结构化级联字段。
• 受影响流程: 原始模型未提供结构化级联字段。
• 变更集外调用方: unknown
• 置信度: degraded

问题发现

1. [低] 未发现阻塞性问题
   • 证据: diff 仅新增 EXTERNAL_KNOWLEDGE_PROTOCOL import，并插入到 generateCodeForFile 与 generateModifiedCode 的 system prompt 末尾；测试直接捕获 generateText 的 system 参数并断言完整协议与关键反向约束短语存在。
   • 受影响调用方/流程: create/modify codegen 路径会收到更完整的 system prompt，但函数输入输出契约保持不变。
   • 最小可行修复: 无需合并前修改；若后续协议常量内容变动，现有测试会继续验证 codegen 侧复用同一常量而非复制漂移。

行级发现

• 无明确变更行归属。

Karpathy 评审

• 假设: 模型输出需要归一化为固定 Markdown 契约。
• 简洁性: 已提取 summary、finding、evidence 与 fix；原始 prose 不再附在评论中，避免占用下游解析与代理上下文。
• 变更范围: 原始模型未提供结构化范围字段。
• 验证: 需要查看 CI、测试或人工 CR 证据补强合并信心。

缺失覆盖

• 输出未命中 Repo Guard Markdown 契约；建议补充真实模型质量评估覆盖。