My Brain Wiki

log-2026

31分钟阅读

2026-07-06 — Ingest: How we teach AI models (Lee Robinson / @leerob)

Source: @leerob (https://x.com/leerob/status/2080467752897146898) Type: Twitter Article (long-form), 875 likes, 1188 bookmarks, 66K views

New Files

  • sources/articles/leerob-how-we-teach-ai-models-2026.md — 篮球类比解释 pretraining → SFT → RL → alignment
  • wiki/concepts/model-alignment.md — Model Alignment 概念页：model spec、对齐流程、benchmark 评测、continual learning
  • wiki/people/lee-robinson.md — Lee Robinson 人物页（Cursor Model Behavior, 前 Vercel VP）

Updated Files

  • wiki/concepts/skill-file.md — 添加 model-alignment 反向链接（agent memory 是 continual learning 方案）
  • wiki/concepts/the-bitter-lesson.md — 添加 model-alignment 反向链接（scale vs alignment 张力）
  • index.md — 添加 2 个新页面条目，总页数 101→103

2026-07-06 — Ingest: Graph Engineering with Claude (Codez / @0xCodez)

Source: @0xCodez (https://x.com/0xCodez/status/2079165300625330317) Type: Twitter Article (long-form), 1129 likes, 528K views, 2914 bookmarks

New Files

  • sources/articles/codez-graph-engineering-claude-2026.md — 14 步图工程路线图原始长文
  • wiki/concepts/graph-engineering.md — Graph Engineering 概念页：节点/边/钻石拓扑/parallel/pipeline/verifier/loop-until-dry
  • wiki/people/codez.md — Codez 人物页

Updated Files

  • wiki/concepts/coding-agents.md — 添加 graph-engineering 反向链接
  • wiki/concepts/harness-engineering.md — 添加 graph-engineering 反向链接
  • wiki/concepts/loop-engineering.md — 添加 graph-engineering 反向链接
  • index.md — 添加 2 个新页面条目，总页数 99→101

2026-07-09 — Maintenance

Action: maintain

Fixes applied:

  • Broken source refs: 0 (already clean)
  • Cross-links: +1 on dan-koe.md (was 1 link, now 2)
  • Confidence: added to kanban.md, scrum.md (single-source pages, also fixed frontmatter from JSON to YAML format)
  • Source drift: fixed 1 hash on harnessx-agent-harness-foundry.md

Result: 0 issues remaining across all 4 phases

2026-07-09 — Ingest: KV Cache Management infographic (Akshay Pachaar)

Source: @akshay_pachaar (https://x.com/akshay_pachaar/status/2074502882812952666) Type: tweet + infographic, 163 likes, 89K views

New Files

  • sources/articles/akshay-lmcache-kv-cache-management-2026.md — 原始推文 + LMCache 架构摘要
  • wiki/projects/lmcache.md — LMCache 项目页

Updated Files

  • wiki/people/akshay-pachaar.md — 添加新推文引用和 LMCache 反向链接
  • wiki/concepts/kv-cache-and-prompt-caching.md — 添加 LMCache 反向链接
  • index.md — 添加 1 个新页面条目，总页数 98→99

2026-07-08 — Ingest: Harness Engineering for Self-Improvement (Lilian Weng)

Source: Lilian Weng, Lil’Log (https://lilianweng.github.io/posts/2026-07-04-harness/) Type: blog post, 28 min read, comprehensive survey

New Files

  • sources/articles/lilian-weng-harness-self-improvement-2026.md — 原始文章摘要
  • wiki/people/lilian-weng.md — Lilian Weng 人物页

Updated Files

  • wiki/concepts/harness-engineering.md — 添加 Lilian Weng 来源和反向链接
  • wiki/concepts/self-evolving-harness.md — 添加 Lilian Weng 来源和反向链接
  • wiki/concepts/meta-harness.md — 添加反向链接
  • wiki/concepts/autoresearch.md — 更新日期
  • wiki/concepts/harness.md — 添加反向链接
  • index.md — 添加 1 个新页面条目，总页数 97→98

2026-07-08 — Ingest: govctl (Tom Dörr tweet)

Source: Tom Dörr @tom_doerr (https://x.com/tom_doerr/status/2073089713771106703) Type: tweet + GitHub repo (https://github.com/govctl-org/govctl) Stats: 268 likes, 313 bookmarks, 17K views

New Files

  • sources/articles/govctl-tweet-2026.md — 原始推文 + GitHub README 摘要
  • wiki/projects/govctl.md — 项目页：governance-as-code CLI
  • wiki/people/tom-doerr.md — 人物页：Tom Dörr

Updated Files

  • wiki/concepts/coding-agents.md — 添加反向链接
  • wiki/concepts/harness-engineering.md — 添加反向链接
  • SCHEMA.md — 添加 governance-as-code tag
  • index.md — 添加 2 个新页面条目，总页数 95→97

2026-06-20 — Ingest: HarnessX

Source: arXiv:2606.14249 (https://arxiv.org/abs/2606.14249) Type: paper (Xiaomi Darwin Agent Team)

Files created:

  • sources/articles/harnessx-agent-harness-foundry.md — 原始来源（论文结构化摘要）
  • wiki/concepts/harnessx.md — HarnessX 概念页

Files updated:

  • wiki/concepts/self-evolving-harness.md — 添加 HarnessX 引用 + 证据时间线条目
  • wiki/concepts/meta-harness.md — 添加 HarnessX 交叉引用
  • wiki/concepts/autoresearch.md — 添加 HarnessX 交叉引用
  • index.md — 添加 harnessx 条目，总页数 92→93

2026-06-08 — Split: harness-engineering.md

Type: maintenance (page split)

Action: 拆分 280 行的 harness-engineering.md 为两个页面：

  • harness-engineering.md (115 行) — 核心概念：定义、关键原则、工程层级、Transient vs Persistent、模型共同演化
  • harness-engineering-case-studies.md (175 行) — 案例与框架：OpenAI Codex、Anthropic 多 Agent/Session Continuity、LangChain Bench、CAR Framework、HarnessCard、Meta-Harness、Sandbox、Visibility Gap

Files created:

  • wiki/concepts/harness-engineering-case-studies.md

Files updated:

  • wiki/concepts/harness-engineering.md — 移除案例/框架/sandbox 内容，保留核心概念
  • index.md — 添加 harness-engineering-case-studies 条目，总页数 84→85

2026-06-08 — Lint: Source 引用修复

Type: maintenance (lint)

Issues found & fixed:

  • 7 个页面的 sources: frontmatter 路径缺少 ../../ 前缀（新增页面遗漏）
    • wiki/people/alok-bishoyi.md
    • wiki/concepts/autoresearch.md
    • wiki/concepts/compound-engineering.md
    • wiki/concepts/curator.md
    • wiki/projects/every.md
    • wiki/projects/evo.md
    • wiki/people/kieran-klaassen.md

Lint results (post-fix):

  • Broken wikilinks: 0
  • Orphan pages: 0
  • Index completeness: 100%
  • Broken source refs: 0
  • Unknown tags: 0
  • Large pages: wiki/concepts/harness-engineering.md (280 lines — 建议后续拆分)

Note: log.md 已有 612 条，接近 500 条轮转阈值（实际超了，下次维护时轮转）

2026-05-28 — Ingest: Arvin Xu - Self-Evolving Harness

Source: https://x.com/arvin17x/status/2059489592097849698 Type: article (Twitter Article)

Files created:

  • sources/articles/self-evolving-harness-arvin-xu.md — 原始来源
  • wiki/concepts/self-evolving-harness.md — 核心概念：Harness 通过 Tracing 数据自动进化
  • wiki/concepts/agent-operation-tracing.md — 技术概念：Agent 可观测性基础设施
  • wiki/people/arvin-xu.md — 作者：LobeHub 创始人
  • wiki/projects/lobehub.md — 项目：开源 AI 聊天平台

Files updated:

Key Concepts Extracted:

  • Self-Evolving Harness: 模型+Harness=产品，Harness 应具备自我进化能力
  • Agent Operation Tracing: 单步执行 = trace event，Execution Snapshot 记录完整执行数据
  • Error Pattern 自动巡检: Agent 批量解析 tracing，9 轮后 Pattern 库饱和（31→104→新增 0），成功率 75%→95%+
  • 自进化四层次: L0 纯人工 → L1 Agent 辅助 → L2 Agent 主导 → L3 人监督 → L4 自主优化
  • Consumer Product 护城河: 不是模型，而是 Harness 中沉淀的 pattern 和 trajectory

2026-05-20 — Ingest: Hermes Agent Curator

Source: https://hermes-agent.nousresearch.com/docs/user-guide/features/curator Type: article (documentation)

Files created:

  • sources/articles/hermes-agent-curator.md — 原始来源
  • wiki/concepts/curator.md — 核心概念：Hermes Agent 的后台 skill 维护系统

Files updated:

  • wiki/concepts/skill-file.md — 添加 curator 交叉引用
  • wiki/concepts/harness-engineering.md — 添加 curator 交叉引用
  • wiki/concepts/thin-harness-fat-skills.md — 添加 curator 交叉引用
  • wiki/concepts/skillify.md — 添加 curator 交叉引用
  • index.md — 添加 curator 概念条目，更新页面计数

2026-05-12 — Ingest: The Bitter Lesson

Source: Rich Sutton, http://www.incompleteideas.net/IncIdeas/BitterLesson.html Type: article (essay)

Files created:

  • sources/articles/the-bitter-lesson.md — 原始来源
  • wiki/concepts/the-bitter-lesson.md — 核心概念：通用方法+大规模计算 > 人类知识驱动
  • wiki/people/rich-sutton.md — 作者实体页

Files updated:

  • index.md — +2 条目，页面总数 71→73

2026-05-10 — Ingest: MCP vs CLI was the wrong debate

Source: @akshay_pachaar X Article (https://x.com/akshay_pachaar/status/2053166970166772052) Type: article

Files created:

  • sources/articles/mcp-vs-cli-wrong-debate.md — 原始来源
  • wiki/concepts/code-mode.md — Code Mode 概念：agent 写代码调工具的运行时模式
  • wiki/concepts/mcp-model-context-protocol.md — MCP 协议概述及 context 开销分析
  • wiki/people/akshay-pachaar.md — 作者实体页

Files updated:

  • index.md — +3 条目，页面总数 68→71

2026-05-05 — Maintenance (follow-up)

Action: maintain (continued)

Fixes applied:

  • Cross-links: +31 wikilinks across 25 remaining pages (all pages now have ≥2 outbound links)
  • Confidence field: added confidence: medium to all 52 single-source pages
  • Source drift: fixed 4 remaining hashes (snarktank ralph x3, swe-atlas) + 1 placeholder (aparna)
  • Broken source refs: 0 (14 fixed in previous pass)

Final state:

  • Broken wikilinks: 0 ✅
  • Broken source refs: 0 ✅
  • Orphan pages: 0 ✅
  • Pages with <2 links: 0 ✅
  • Single-source no confidence: 0 ✅
  • Source drift: 0 ✅
  • Tags not in taxonomy: 0 ✅
  • Unused taxonomy tags: 7 (comparison, controversy, fine-tuning, lab, prediction, timeline, training — kept for future use)

2026-05-05 — Maintenance

Action: maintain

Scope: Full wiki maintenance pass (68 pages, 39 sources)

Fixes applied:

  • Broken source refs: 14 → 0 — Fixed incorrect relative paths across 14 pages:
    • 6 pages in wiki/concepts/ used ../sources/ instead of ../../sources/
    • 2 people pages with wrong raw/ or ../ prefixes
    • 1 filename typo (karpathy-llms-wiki.md → karpathy-llm-wiki.md)
    • 5 more with missing ../../ prefix
  • Cross-links: added 74 wikilinks across 34 pages (37 pages with <2 links → 25 remaining)
  • Source drift: 1 fixed (kv-cache-hidden-engine-jayanth.md sha256 corrected)

Remaining items (non-critical):

  • 25 pages with <2 outbound links (niche topics, low overlap with other pages)
  • 52 single-source pages without confidence field (informational, no action needed now)
  • 5 pre-existing source drift issues (placeholder hashes from early ingests)
  • 1 large page: harness-engineering.md (281 lines) — content-dense, acceptable
  • 7 unused taxonomy tags (comparison, controversy, fine-tuning, lab, prediction, timeline, training)

2026-05-05 — Lint

Action: lint

Results (68 pages, 445 log lines):

  • Broken wikilinks: 0 ✅
  • Malformed frontmatter: 0 ✅
  • Orphan pages: 0 ✅
  • Missing fields: 0 ✅
  • Index completeness: 0 missing, 0 stale ✅
  • Tags in taxonomy: 0 missing ✅
  • Large pages (>200 lines): 1 (harness-engineering.md: 281 lines)
  • Source drift: 5 pre-existing (snarktank ralph x3, aparna, swe-atlas) + 1 fixed (kv-cache-hidden-engine-jayanth.md)
  • Contested/low-confidence: 0 ✅

Fixes applied: Corrected sha256 for kv-cache-hidden-engine-jayanth.md Remaining: 5 pre-existing source drift (placeholder hashes from earlier ingests), 1 large page (acceptable)

2026-05-05 — Ingest: Jayanth Sanku - KV Cache Hidden Engine

Action: ingest Scope: sources/articles/ + wiki/concepts/ (update) Source: Twitter Notes (x.com/JayanthSanku01/status/2050963464915743150) Changes:

  • Saved source: sources/articles/kv-cache-hidden-engine-jayanth.md
  • Updated wiki/concepts/kv-cache-and-prompt-caching.md: added Trade-offs, Optimizations, Real-World Use Cases sections
  • Key additions: Paged KV Cache, Quantized KV Cache, Eviction Strategies, Flash Attention

2026-04-23 — Ingest: He et al. - Harness Engineering (CAR & HarnessCard)

Action: ingest

Scope: sources/articles/ + wiki/concepts/ + wiki/people/

Source: Preprints.org paper doi:10.20944/preprints202603.1756.v2

Changes:

  • Saved source paper: sources/articles/harness-engineering-car-harnesscard.md
  • Created/updated wiki pages:
    • wiki/concepts/car-framework.md (新概念 - Control, Agency, Runtime 分解)
    • wiki/concepts/harnesscard.md (新概念 - 轻量级报告规范)
    • wiki/concepts/harness-engineering.md (更新 - 添加 CAR、HarnessCard、Visibility Gap)
    • wiki/people/chaoyue-he.md (新人物)
  • Updated index.md with new entries

Key Concepts Extracted:

  • CAR Framework: H = ⟨C, A, R⟩ — Control, Agency, Runtime 形式化分解 harness 层
  • HarnessCard: 9 required + 2 recommended fields for reporting agent configurations
  • Visibility Gap: 63 篇工作审计，学术论文与工程笔记之间的可见性鸿沟
  • Key argument: 很多 “agent gains” 是 harness-sensitive 而非纯模型驱动
  • Engineering evolution: Software → Prompt → Context → Harness Engineering

2026-04-23 — Ingest: Garry Tan - Thin Harness, Fat Skills

Action: ingest

Scope: sources/articles/thin-harness-fat-skills.md + wiki/

Source: https://github.com/garrytan/gbrain/blob/master/docs/ethos/THIN_HARNESS_FAT_SKILLS.md

Changes:

  • Saved source article with frontmatter
  • Created/updated wiki pages:
    • wiki/concepts/thin-harness-fat-skills.md (核心架构原则)
    • wiki/concepts/skill-file.md (新概念)
    • wiki/concepts/harness.md (新概念)
    • wiki/concepts/resolver.md (更新)
    • wiki/concepts/latent-vs-deterministic.md (新概念)
    • wiki/concepts/diarization.md (新概念)
    • wiki/people/garry-tan.md (更新)
    • wiki/projects/gbrain.md (更新)
  • Updated index.md with new entries

Key Concepts Extracted:

  • 五定义：Skill File, Harness, Resolver, Latent vs Deterministic, Diarization
  • 三层架构：Fat Skills → Thin Harness → Your App
  • Skill 如方法调用：相同流程，不同参数
  • Resolver 路由上下文
  • 自学习循环：retrieve → diarize → rewrite skill

2026-04-23 — Source Redownload & Full Reformat

Action: redownload + reformat

Scope: sources/articles/2046876981711769720.md

Changes:

  • Redownloaded from https://x.com/garrytan/status/2046876981711769720
  • Full Markdown reformat: proper headings, lists, code blocks, blockquotes
  • Removed all Twitter UI elements and engagement metrics
  • Added structured sections matching article’s logical flow
  • Clean References section at end

2026-04-23 — Source Cleanup

Action: format_fix

Scope: sources/articles/2046876981711769720.md

Changes:

  • Removed Twitter engagement metrics (likes, retweets, replies, views)
  • Removed Twitter UI elements (“Want to publish your own Article?”, timestamp, etc.)
  • Consolidated author line to single format: “Garry Tan (@garrytan)”
  • Added References section for project links
  • Restored accidentally deleted paragraphs during cleanup

2026-04-20 — Maintenance Pass #2

Action: maintain

Scope: 40 wiki pages + 62 source files

Findings & Fixes:

  1. Orphan fixed (1):
  • wiki/people/gleb-rodionov.md → reasoning-shift.md 添加入链
  1. Tag 统一 (5 files):
  • llm → llms（复数规范）: chaofa-yuan, llm-wiki-pattern, kv-cache-and-prompt-caching, reasoning-shift, claude-opus-4-7
  1. Updated 日期同步: 5个修改页更新为 2026-04-20

Stats:

  • Orphans: 0（after fix）
  • Tag issues: 0（after fix）
  • Stale pages: 0
  • Broken source refs: 0
  • Large pages (>500 lines): 0
  • Index: complete (40/40)

2026-04-20 — Ingest: Aparna Dhinakaran

Action: ingest

Source: Sandboxes Are the Servers of the Harness Era

Changes:

  1. Created source: sources/articles/2045925288908046570.md
  2. Updated wiki: wiki/concepts/harness-engineering.md — added “Sandbox: The Server of the Harness Era” section
  3. Created wiki: wiki/people/aparna-dhinakaran.md — new person page
  4. Created wiki: wiki/concepts/sandbox.md — new concept page
  5. Updated index: Added new source, person, and concept entries
  • Broken source refs: 0
  • Large pages (>500 lines): 0
  • Index: complete (40/40)

2026-04-20 — Maintenance Pass

Action: maintain

Scope: Full audit of 38 wiki pages + 62 source files

Findings & Fixes:

  1. Orphans fixed (4):
  • wiki/projects/ddia.md → 添加入链：claude-code-architecture.md related 到它
  • wiki/concepts/agile.md → 添加入链：claude-code-workflow.md related 到它
  • wiki/concepts/kanban.md / wiki/concepts/scrum.md → 已通过 agile.md 间接连通（agile related 到它们，现在 agile 有入链）
  1. Missing people pages created (2):
  • wiki/people/martin-kleppmann.md — DDIA 作者
  • wiki/people/gleb-rodionov.md — Reasoning Shift 论文作者
  1. index.md 更新: 新增 martin-kleppmann、gleb-rodionov 两个 people 条目

Stats:

  • Pages audited: 38 wiki + 62 sources
  • Orphans fixed: 4
  • New people pages: 2
  • Stale pages: 0
  • Tag issues: 0
  • Broken source refs: 0
  • Index: complete (40/40 pages listed)

2026-04-20 — Ingest: DDIA 第二版

Action: ingest

Source: https://ddia.vonng.com (设计数据密集型应用第二版，中文译本，Martin Kleppmann 著，冯若航译)

Files created:

  • sources/books/ddia-preface.md (序言)
  • sources/books/ddia-ch01.md 到 ddia-ch14.md (14章)
  • wiki/projects/ddia.md (综合知识页)

Script: scripts/fetch_ddia.py (curl + BeautifulSoup，静态 HTML 直接解析)

Stats:

  • 15 个源文件，577,730 字符
  • 1 个 wiki 综合页（14章精华 + 概念索引表）
  • index.md: Books + Projects 各添加1条
  • 无新增 people 页（作者 Martin Kleppmann 未在库中）

2026-04-17 — Maintenance Pass

Action: maintain

Scope: Full audit of 33 wiki pages

Findings & Fixes:

  1. Broken link fixed: wiki/people/thariq-shihipar.md had ../concepts/claude-code-workflow.md (wrong path). Fixed to ../projects/claude-code-workflow.md. Also added ../concepts/claude-code-session-management.md cross-ref.

  2. Cross-references added (low-connectivity pages):

  • claude-opus-4-7.md ↔ openai-codex-2026.md (mutual competitor reference)
  • claude-code-architecture.md → claude-code-session-management.md (natural extension)
  • harness-engineering.md → synthesis/agentic-rag-as-harness.md (synthesis page now reachable)
  1. Updated updated: dates on 5 modified pages.

Stats:

  • Pages audited: 33 wiki + 40 sources
  • Orphans: 0 (after fixes)
  • Stale pages: 0
  • Tag issues: 0
  • Broken source refs: 0
  • Index: complete (33/33 pages listed)

Log

Append-only record of operations.

  • 2026-04-07 | init | Created knowledge base structure and schema

  • 2026-04-07 | ingest | Added Karpathy LLM Wiki article → created source, wiki/concepts/llm-wiki-pattern, wiki/people/andrej-karpathy

  • 2026-04-07 | ingest | Added Garry Tan GBrain article → created source, updated llm-wiki-pattern, created wiki/people/garry-tan

  • 2026-04-07 | ingest | Added Böckeler’s Harness Engineering article → created source, wiki/concepts/harness-engineering, wiki/concepts/coding-agents, wiki/people/martin-fowler, wiki/people/birgitta-bockeler; updated llm-wiki-pattern cross-refs

  • 2026-04-07 | query+create | “What is Ashby’s Law” → created wiki/concepts/ashbys-law, updated harness-engineering cross-refs

  • 2026-04-07 | ingest | Added OpenAI’s Harness Engineering article → created source, updated wiki/concepts/harness-engineering with Codex case study, updated wiki/concepts/coding-agents

  • 2026-04-07 | ingest | Added Anthropic’s Harness Design article → created source, updated harness-engineering with GAN-inspired multi-agent pattern, updated coding-agents with context anxiety and harness simplification insights

  • 2026-04-07 | maintain | Audit pass: fixed 1 broken ref (removed non-existent personal-knowledge-management.md), fixed 1 orphan (linked birgitta-bockeler from harness-engineering), added missing cross-refs (people ↔ concepts)

  • 2026-04-07 | ingest | Added Justin Young’s Effective Harnesses article → updated harness-engineering with session continuity pattern, updated coding-agents with failure modes and file-based handoff

  • 2026-04-07 | maintain | Audit pass: fixed tag inconsistency (ai → ai-agents in karpathy), added missing cross-refs (martin-fowler → harness-engineering, ashbys-law)

  • 2026-04-07 | maintain | Found gap: maintain workflow missed Justin Young (no entity coverage check). Created wiki/people/justin-young. Updated CLAUDE.md maintain workflow to include entity coverage check

  • 2026-04-07 | ingest | Added claude-howto guide → created source, created wiki/projects/claude-code-workflow (first project page!), mapped Claude Code features to harness engineering concepts

  • 2026-04-07 | maintain | Audit pass: fixed 2 orphans (linked claude-code-workflow and justin-young from concept pages), created 2 missing author pages (prithvi-rajasekaran, luongnv89), added missing cross-ref (coding-agents → ashbys-law)

  • 2026-04-07 | ingest | Added Meta-Harness paper (arXiv:2603.28052) → created source, wiki/concepts/meta-harness, wiki/people/chelsea-finn, wiki/people/omar-khattab; updated harness-engineering and coding-agents with Meta-Harness cross-refs

  • 2026-04-07 | ingest | Added Jagtap’s Meta-Harness library article → created source, wiki/people/shashikant-jagtap; updated meta-harness with open-source implementation section, updated harness-engineering and coding-agents timelines

  • 2026-04-07 | maintain | Audit pass: added shashikant-jagtap to harness-engineering related, added missing source refs to coding-agents (meta-harness-optimization, meta-harness-library-jagtap), linked luongnv89 from claude-code-workflow (was orphan)

  • 2026-04-07 | ingest | Added Anthropic’s “Building Effective Agents” (Schluntz & Zhang) → created source, wiki/concepts/agentic-patterns (6 composable patterns), wiki/people/erik-schluntz, wiki/people/barry-zhang; updated coding-agents and harness-engineering cross-refs

  • 2026-04-07 | maintain | Audit pass: fixed tag inconsistency (architecture → architecture-pattern in agentic-patterns). No orphans, no broken refs, all entity coverage OK. 19 wiki pages, 10 sources.

  • 2026-04-07 | ingest | Added rosa’s “从Bash工具开始理解Agent” → created source, wiki/concepts/tool-use-as-meta-ability, wiki/people/rosa; updated coding-agents and agentic-patterns cross-refs

  • 2026-04-07 | maintain | Audit pass: fixed tag inconsistency (tools → tool-use in coding-agents, claude-code-workflow), added missing cross-ref (tool-use-as-meta-ability → rosa). No orphans, no broken refs, entity coverage OK. 21 wiki pages, 11 sources.

  • 2026-04-07 | ingest | Added Chaofa Yuan’s harness engineering article → created source, wiki/people/chaofa-yuan; updated harness-engineering with engineering hierarchy, transient/persistent distinction, co-evolution, LangChain case study; updated coding-agents timeline

  • 2026-04-07 | ingest | Added Chaofa Yuan’s KV Cache & Prompt Caching article → created source, wiki/concepts/kv-cache-and-prompt-caching (new topic: LLM inference optimization); updated wiki/people/chaofa-yuan with second article

  • 2026-04-07 | ingest | Added rosa’s “基于Openclaw的官方文档理解其架构设计” → created source, updated wiki/people/rosa (added second article), added openclaw as emerging concept (gateway, agentic-loop, skills, mcp, memory, heartbeat)

  • 2026-04-07 | ingest | Added Chaofa Yuan’s “Agent 系统中的 Prompt Caching 设计（上）” → created source, updated wiki/people/chaofa-yuan (third article), updated kv-cache-and-prompt-caching with Claude Code 4-layer architecture and tool management patterns

  • 2026-04-07 | ingest | Added Chaofa Yuan’s “Agent 系统中的 Prompt Caching 设计（下）” → created source, updated wiki/people/chaofa-yuan (fourth article), updated kv-cache-and-prompt-caching with context rot, compaction, sub-agent patterns

  • 2026-04-07 | maintain | Audit pass: fixed 8 orphans (linked all people to concept pages), added kv-cache-and-prompt-caching to rosa and chaofa-yuan related, added luongnv89 to coding-agents, fixed tag inconsistency (trailing spaces), updated claude-code-workflow related

  • 2026-04-07 | ingest | Added “爱马仕：一家做了 189 年 Harness 的公司” → created source, created wiki/people/jinse-chuanshuo-dacongrming, updated harness-engineering with new metaphor

  • 2026-04-08: Maintenance pass

  • Orphan pages: ✅ None found

  • Stale pages (90+ days): ✅ None found

  • Broken refs: ✅ None found

  • Tag consistency: ✅ All lowercase/hyphenated

  • Entity coverage: ✅ All source authors have wiki pages (including new entries)

  • Index sync: ✅ Fixed missing entry (金色传说大聪明)

  • 2026-04-08 | maintain | Audit pass: added hermes-harness-metaphor.md to harness-engineering sources (was missing), updated rosa.md with OpenClaw article content in Key Points and Timeline, added rosa-openclaw-architecture.md to coding-agents sources with timeline entry. 17 sources, 8 concepts, 15 people, 1 project. No orphans, no broken refs.

  • 2026-04-08 | ingest | Added Chaofa Yuan’s “RAG 进化之路” → created source, created wiki/concepts/agentic-rag (new concept: RAG evolution); updated chaofa-yuan (5th article), added cross-refs to agentic-patterns and coding-agents. 18 sources, 9 concepts.

  • 2026-04-08 | maintain | Audit pass: no orphans, no broken refs, no stale pages, entity coverage complete. Fixed 4 tag inconsistencies: agent → ai-agents (3 files), prompt-cache → prompt-caching (2 files), tools → tool-use (1 file). 18 sources, 9 concepts, 15 people, 1 project.

  • 2026-04-08 | query+create | “将 Agentic RAG 做成 harness” → created wiki/synthesis/agentic-rag-as-harness (first synthesis page!). Cross-cutting analysis: harness 化的 RAG 检索轨迹可桥接工具驱动与 RL 驱动两条路径.

  • 2026-04-08: Ingested Agency Agents (msitarzewski/agency-agents)

  • Created source: sources/articles/agency-agents.md

  • Created wiki: wiki/people/msitarzewski.md

  • Updated index.md with both entries

  • 2026-04-10 | maintain | Audit pass results:

  • Orphan pages: Fixed 3 orphans — linked msitarzewski from coding-agents, agentic-rag-as-harness from agentic-rag, crypto-morningstar-integration from claude-code-workflow (mutual link)

  • Stale pages: ✅ None (all pages <90 days old)

  • Contradictions: ✅ None found

  • Tag consistency: Fixed 2 files using ai-agent (singular) → ai-agents (plural, standard). All tags now lowercase/hyphenated.

  • Entity coverage: ✅ All source authors have wiki pages (false positives from multi-author and parenthetical fields in parser)

  • Index sync: ✅ All 30 wiki pages appear in index.md

  • Updated updated: dates on 6 modified pages.

  • 2026-04-10 | ingest | Added “Claude Code from Source” full book (18 chapters) by Alejandro Balderas. Source: https://claude-code-from-source.com/. Created 18 source files (sources/books/claude-code-from-source-ch01..ch18.md), wiki/projects/claude-code-architecture.md (main book page), wiki/people/alejandro-balderas.md. Updated 4 concept pages (coding-agents, agentic-patterns, kv-cache-and-prompt-caching, tool-use-as-meta-ability) with cross-refs and timeline entries. Updated index.md with Books section, new project, and new person. Total: 37 sources (18 book chapters + 19 articles), 9 concepts, 17 people, 3 projects, 1 synthesis.

  • 2026-04-17 | ingest | Added Anthropic “Introducing Claude Opus 4.7” article

  • Created source: sources/articles/claude-opus-4-7.md

  • Created wiki: wiki/concepts/claude-opus-4-7.md (new concept)

  • Updated index.md with new concept entry

  • Related to: coding-agents, agentic-patterns

  • Key findings: 13% improvement on coding benchmark, xhigh effort level, higher resolution vision (2,576px), updated tokenizer (1.0-1.35x more tokens)

  • 2026-04-17 | ingest | Added OpenAI “Codex for (almost) everything” article (via Wayback)

  • Created source: sources/articles/codex-update-2026.md

  • Created wiki: wiki/concepts/openai-codex-2026.md (new concept)

  • Updated index.md with new concept entry

  • Related to: coding-agents, agentic-patterns

  • Key findings: Background computer use, multi-agent parallel work, memory preview, 90+ new plugins, in-app browser

  • 2026-04-17 | ingest | Added Anthropic “Using Claude Code: session management and 1M context” article

  • Created source: sources/articles/claude-code-session-management.md

  • Created wiki: wiki/concepts/claude-code-session-management.md (new concept)

  • Created wiki: wiki/people/thariq-shihipar.md (new person)

  • Updated index.md with new concept and person

  • Related to: coding-agents, harness-engineering, claude-code-workflow, justin-young, thariq-shihipar

  • Key findings: Context rot, 1M token, 5 tools (continue/rewind/clear/compact/subagent), decision table

  • 2026-04-17 | ingest | Added Microsoft Learn “什么是敏捷？” article

  • Created source: sources/articles/what-is-agile.md

  • Created wiki: wiki/concepts/agile.md (new concept)

  • Updated index.md with new source and concept

  • Related to: devops, software-development, methodology

  • Key findings: 敏捷宣言四个价值观、敏捷框架(Scrum/看板) vs 敏捷实践(规划扑克/持续集成)、常见误解澄清

  • 2026-04-17 | ingest | Added Microsoft Learn 敏捷系列文章 (5篇)

  • Sources created:

  • sources/articles/what-is-agile-development.md

  • sources/articles/what-is-scrum.md

  • sources/articles/what-is-kanban.md

  • sources/articles/adopting-agile.md

  • sources/articles/building-productive-teams.md

  • Wiki concepts created:

  • wiki/concepts/scrum.md

  • wiki/concepts/kanban.md

  • Updated wiki/concepts/agile.md with new content from all 5 articles

  • Updated index.md with all new sources and concepts

  • Key findings: 敏捷开发三要素、Scrum 框架、看板四原则、敏捷文化、双机组系统 (F-Crew/C-Crew)

  • 2026-04-17 | maintain | Audit pass:

  • Broken source paths fixed: agile.md, scrum.md, kanban.md had ../sources/ (resolves to wiki/sources/) — fixed to ../../sources/ (correct root-relative path)

  • Index sync: Added 3 missing source entries (claude-opus-4-7.md, codex-update-2026.md, claude-code-session-management.md) to Sources section

  • Cross-refs added: coding-agents → claude-opus-4-7, openai-codex-2026, claude-code-session-management; harness-engineering → claude-code-session-management; justin-young → claude-code-session-management

  • Orphans: 0 (agile/scrum/kanban form a self-contained cluster)

  • Stale pages: 0 (all <90 days)

  • Tag consistency: ✅

  • Entity coverage: ✅ (Microsoft Learn articles have corporate authorship, no individual author pages needed)

  • Stats: 36 wiki pages, 47 sources. Updated updated: dates on 4 modified pages.

  • 2026-04-20 | ingest | Added arXiv paper “Reasoning Shift: How Context Silently Shortens LLM Reasoning” (Rodionov, 2026)

  • Source created: sources/articles/reasoning-shift-rodionov.md

  • Wiki concept created: wiki/concepts/reasoning-shift.md

  • Updated related pages: wiki/concepts/claude-code-session-management.md, wiki/concepts/harness-engineering.md

  • Updated index.md with new source and concept entries

  • Key findings: 推理型 LLM 在非隔离上下文条件下推理链最多压缩 50%，自我验证行为（double-checking）显著减少。对简单问题影响小，对复杂任务性能下降。与 context rot、harness engineering 子代理策略高度相关。

  • 2026-04-22 | ingest | Added Garry Tan’s “How to really stop your agents from making the same mistakes” article

  • Created source: sources/articles/2046876981711769720.md

  • Created wiki concepts: wiki/concepts/skillify.md, wiki/concepts/thin-harness-fat-skills.md, wiki/concepts/resolver.md

  • Created wiki project: wiki/projects/gbrain.md

  • Updated wiki people: wiki/people/garry-tan.md

  • Updated index.md with new source, concepts, and project entries

  • Related to: ai-agents, reliability, testing, skills, langchain, gbrain, harness-engineering

  • Key findings: Skillify practice (10-step checklist), thin harness/fat skills architecture, resolver routing, critique of LangChain for tools without workflow

  • 2026-04-22 | maintain | Audit pass:

    • 孤儿子检查: 2个误报页面 (kanban.md, scrum.md 被 agile.md 引用，不是真正的孤儿)
    • 标签一致性: ✅ 良好
    • 实体覆盖: ✅ 所有个人作者都有 people 页面
    • 索引同步: ✅ 所有页面都在 index.md 中
    • 断裂引用: ✅ 没有断裂引用
    • 统计: 46 wiki 页面, 64 sources. 无实际问题需要修复。

[2026-05-02] ingest | snarktank/ralph

  • 来源: https://github.com/snarktank/ralph
  • 保存原始内容:
    • raw/articles/snarktank-ralph-github-2026.md
    • raw/articles/snarktank-ralph-ralph-sh-2026.md
    • raw/articles/snarktank-ralph-prompt-md-2026.md
  • 创建页面:
    • entities/ralph.md — 自主 AI 编码 Agent 循环，bash 脚本驱动，PRD 定义任务
    • concepts/agent-loop-pattern.md — Agent 循环迭代模式综述，对比长 context 策略
  • 更新:
    • index.md (+2 pages)
    • concepts/agent-context-management.md (添加 ralph 交叉引用)

[2026-05-02] ingest | snarktank/ralph (correction)

  • 修正: entities/ralph.md → wiki/people/ralph.md (匹配现有 index 结构)
  • 创建附加页面:
    • wiki/people/ryan-carson.md — Ralph 作者
    • wiki/people/geoffrey-huntley.md — Ralph 模式原始提出者

[2026-05-02] merge | 本地 wiki 与 my-brain GitHub 仓库合并

  • 来源: https://github.com/xiaoquisme/my-brain
  • 策略: 用 my-brain 仓库替换 ~/wiki，将本地独有页面合并入
  • 本地新增页面 (11):
    • people/arize-alyx.md — Arize AI & Alyx Agent
    • people/scale-ai.md — Scale AI
    • concepts/agent-context-management.md — Agent context 管理收敛模式
    • concepts/ai-coding-benchmark.md — AI 编程基准综述
    • concepts/codebase-qna.md — 代码库问答评测形式
    • projects/claude-code-harness.md — Claude Code harness
    • projects/harbor.md — Harbor 任务运行框架
    • projects/letta-code.md — Letta Code harness
    • projects/openclaw.md — OpenClaw harness
    • projects/pi-mono.md — Pi harness
    • projects/swe-atlas.md — SWE-Atlas 评测基准
  • 新增来源文件 (2):
    • sources/articles/aparna-context-management-agent-harnesses-2026.md
    • sources/articles/swe-atlas-github-2026.md
  • 更新: SCHEMA.md (新建), index.md (68 pages)
  • 备份: ~/wiki-backup-20260502

[2026-05-02] ingest | snarktank/ralph

  • 来源: https://github.com/snarktank/ralph
  • 保存原始内容:
    • sources/articles/snarktank-ralph-github-2026.md (README)
    • sources/articles/snarktank-ralph-ralph-sh-2026.md (核心 bash 循环脚本)
    • sources/articles/snarktank-ralph-prompt-md-2026.md (Amp agent 指令模板)
  • 创建页面:
    • wiki/people/ralph.md — 自主 AI 编码 Agent 循环，113 行 bash 脚本驱动，PRD 定义任务，支持 Amp + Claude Code
    • wiki/people/ryan-carson.md — Ralph 作者，snarktank 仓库维护者
    • wiki/people/geoffrey-huntley.md — Ralph 模式原始提出者，文件系统做 Agent 记忆的理念
    • wiki/concepts/agent-loop-pattern.md — Agent 循环迭代模式综述：每轮新实例 + 文件系统记忆 + PRD 驱动，对比长 context 策略
  • 更新:
    • index.md (+4 wiki pages, +3 source entries, page count 68)
    • concepts/agent-context-management.md (添加 ralph 交叉引用)

[2026-05-02] restructure | index.md & SCHEMA.md 重建

  • 问题: index.md 仅有 11 条旧路径引用（entities/, concepts/），实际文件结构为 wiki/concepts/、wiki/people/、wiki/projects/、wiki/synthesis/
  • 操作:
    • 重建 index.md: 68 个页面全部按实际路径列出（30 concepts, 27 people, 10 projects, 1 synthesis）
    • 更新 SCHEMA.md: 反映实际目录结构（sources/ + wiki/ 四个子目录），更新 tag taxonomy（从 20 个扩展到匹配实际使用的 99 个标签，按 6 大类分组）
  • 统计: 68 wiki 页面, 63 source 文件（34 articles + 29 book chapters）

[2026-05-02] lint | Wiki 全面维护

发现问题

  • 断链: 0
  • 孤儿页面: 56 (无入站链接)
  • Index 完整性: 完美 (68/68)
  • Frontmatter 问题: 61 页面缺少 type 字段, 4 文件 frontmatter 格式错误
  • 页面超 200 行: 1 (harness-engineering.md: 275 行)
  • Tag 分类外: 62 个 tag 不在 SCHEMA.md 分类中
  • 未使用分类 tag: 7 个
  • 来源漂移: 0
  • 矛盾页面: 0
  • 单来源无 confidence: 49 页面

修复操作

  • 修复 4 个 malformed frontmatter (claude-code-harness, letta-code, openclaw, pi-mono): related 字段 YAML 格式错误 → 修正为标准列表
  • 为 60 个页面添加 type 字段 (从目录推断: concept/entity/synthesis)
  • SCHEMA.md tag 分类扩展: 从 ~30 个扩展到 105 个, 新增 11 个分类 (Agent 架构, 模型实例, 知识与认知, 系统理论, 数据等)
  • methodology → software-development 统一重命名 (3 页面)
  • 为 56 个孤儿页面添加 112 条交叉引用, 更新 51 个目标页面
  • 所有 68 页面现在均有入站链接

待关注

  • 1 页面超 200 行 (harness-engineering.md) — 候选拆分
  • 49 页面单来源无 confidence 设置 — 建议后续补充或降低 confidence
  • log.md 408 行, 接近 500 行轮转阈值

[2026-06-07] ingest | Evo Autoresearch + LawBench SOTA

来源：Alok Bishoyi Twitter Article (https://x.com/alokbishoyi97/status/2063193533930942605)

新建文件

  • sources/articles/evo-autoresearch-lawbench-2026.md — 原始文章
  • wiki/projects/evo.md — Evo 项目页：开源 autoresearch orchestrator，LawBench 0.776 SOTA
  • wiki/people/alok-bishoyi.md — Alok Bishoyi 人物页
  • wiki/concepts/autoresearch.md — Autoresearch 概念页：base model × harness × verifiable loop

更新文件

  • wiki/concepts/self-evolving-harness.md — 添加反向链接到 autoresearch 和 evo
  • wiki/concepts/meta-harness.md — 添加反向链接到 autoresearch 和 evo
  • index.md — 添加 3 个新页面条目，总页数 78→81

[2026-06-07] ingest | Compound Engineering (Every.to)

来源：Kieran Klaassen, Every.to (https://every.to/guides/compound-engineering)

新建文件

  • sources/articles/compound-engineering-every-2026.md — 原始文章（约 40K 字符）
  • wiki/concepts/compound-engineering.md — 核心概念：七步循环、50/50 规则、五级成熟度、40+ 审查 agent
  • wiki/people/kieran-klaassen.md — 作者页面
  • wiki/projects/every.md — Every 公司页面（6 个产品）

更新文件

  • wiki/concepts/skill-file.md — 添加反向链接
  • wiki/concepts/skillify.md — 添加反向链接
  • wiki/concepts/thin-harness-fat-skills.md — 添加反向链接
  • index.md — 添加 3 个新页面条目，总页数 81→84

[2026-06-09] ingest | 控制论与智能体编码中的”人在环上”

来源：Dirk Lässig, 微信公众号 (https://mp.weixin.qq.com/s/OqbXqE8Go-vsLTX-O-JwbA)

新建文件

  • sources/articles/cybernetics-human-on-the-loop-2026.md — 原始文章
  • wiki/concepts/human-on-the-loop.md — 核心概念：人类退居元维度引导 Agent 系统
  • wiki/concepts/attenuation-amplification.md — 衰减与放大：HotL 的两个核心控制机制
  • wiki/concepts/gemba-go-see.md — 现地现物：精益管理在 Agent SDLC 中的校准实践

更新文件

  • wiki/concepts/ashbys-law.md — 添加反向链接到 human-on-the-loop、attenuation-amplification
  • wiki/concepts/harness-engineering.md — 添加反向链接到三个新概念页
  • index.md — 添加 3 个新页面条目，总页数 85→88

[2026-06-09] lint | 3 broken source refs fixed

Issues found and fixed:

  • 3 broken source refs in new pages (missing ../../ prefix)
    • wiki/concepts/human-on-the-loop.md
    • wiki/concepts/attenuation-amplification.md
    • wiki/concepts/gemba-go-see.md

Lint results (post-fix):

  • Broken wikilinks: 0
  • Orphan pages: 0
  • Index completeness: 100%
  • Broken source refs: 0
  • Unknown tags: 0
  • Large pages: 1 (harness-engineering-case-studies: 207 lines)

[2026-06-09] ingest | Loop Engineering (Addy Osmani)

来源：Addy Osmani (https://addyosmani.com/blog/loop-engineering/)

新建文件

  • sources/articles/loop-engineering-osmani-2026.md — 原始文章
  • wiki/concepts/loop-engineering.md — 核心概念：五原语循环架构（Automations, Worktrees, Skills, Plugins, Sub-agents）
  • wiki/people/addy-osmani.md — Addy Osmani 人物页
  • wiki/concepts/factory-model.md — Factory Model 概念页

更新文件

  • wiki/concepts/harness-engineering.md — 添加反向链接
  • wiki/concepts/human-on-the-loop.md — 添加反向链接
  • wiki/concepts/skill-file.md — 添加反向链接
  • wiki/concepts/compound-engineering.md — 添加相关页面 section
  • index.md — 添加 3 个新页面条目，总页数 88→91

[2026-06-09] ingest | Stop Using /init for AGENTS.md (Addy Osmani)

来源：Addy Osmani (https://addyosmani.com/blog/agents-md/)

新建文件

  • sources/articles/agents-md-osmani-2026.md — 原始文章
  • wiki/concepts/agents-md-best-practices.md — AGENTS.md 最佳实践：只写不可发现信息、三层架构、诊断工具心智模型

更新文件

  • wiki/concepts/agent-context-management.md — 添加反向链接
  • wiki/concepts/harness-engineering.md — 添加反向链接
  • wiki/people/addy-osmani.md — 添加新文章引用
  • index.md — 添加 1 个新页面条目，总页数 91→92

[2026-06-30] ingest | Human 3.0 – Dan Koe (Twitter Article)

来源：Dan Koe @thedankoe (https://x.com/thedankoe/status/2023779299367809063) Twitter Article, 4M views, 12.5K likes, 33.6K bookmarks

新建文件

  • sources/articles/human-3.0-dan-koe-2026.md — 原始文章（Twitter Article 长文）
  • wiki/concepts/human-3.0.md — 核心概念：四象限 × 三层级 × 三阶段个人发展框架，AI 作为 Glitch
  • wiki/people/dan-koe.md — Dan Koe 人物页

更新文件

  • SCHEMA.md — 添加个人发展 tag 分类（personal-development, philosophy, human-potential, psychology）
  • wiki/concepts/the-bitter-lesson.md — 添加反向链接
  • wiki/concepts/skill-file.md — 添加反向链接
  • wiki/concepts/llm-wiki-pattern.md — 添加反向链接
  • index.md — 添加 2 个新页面条目，总页数 93→95

[2026-07-13] update | Compound Engineering — source drift detected & full re-ingest

来源: https://every.to/guides/compound-engineering

Drift 检测

源文章自 2026-06-07 首次收录后有重大更新:

  • 主循环从 7 步简化为 4 步 (Plan > Work > Review > Compound)
  • 时间分配新增 80/20 规则 (features) + 保留 50/50 (system improvement)
  • 命令从 /ce-* 更名为 /workflows:*
  • Review agent 从泛称 40+ 细化为 14 个具名专家
  • 新增章节: Getting Started, Design in Code, Vibe Coding, Team Collaboration, User Research, Copy, Marketing
  • Beliefs 从 10-8 条, 新增 Emotional Reactions 小节

更新文件

  • sources/articles/compound-engineering-every-2026.md — 完整重写, 更新 sha256
  • wiki/concepts/compound-engineering.md — 重写核心内容 (四步循环 + 新实践领域)
  • wiki/projects/every.md — 更新 plugin 统计数据和证据时间线
  • wiki/people/kieran-klaassen.md — 更新核心贡献描述和证据时间线
  • index.md — 更新摘要描述 (七步 to 四步)
  • wiki/concepts/compound-engineering.md — 清理重复链接

2026-07-27 — Ingest: Context Engineering for Claude 5 (Thariq / @trq212)

Source: @trq212 (https://x.com/trq212/status/2080710971228918066) Type: Twitter Article (long-form), 14965 likes, 30875 bookmarks, 3.9M views

New Files

  • sources/articles/thariq-context-engineering-claude5-2026.md — Anthropic 删除 80% 系统 prompt 的经验总结
  • wiki/concepts/context-engineering.md — Context Engineering 概念页：六条旧规则→新规则演进

Updated Files

  • wiki/people/thariq-shihipar.md — 补充 context engineering 文章，更新人物描述
  • wiki/concepts/claude-code-session-management.md — 添加 context engineering 演进内容和反向链接
  • wiki/concepts/agent-context-management.md — 添加 context engineering shift 章节和反向链接
  • wiki/concepts/agents-md-best-practices.md — 添加 Anthropic 实践验证章节
  • wiki/concepts/skill-file.md — 添加 progressive disclosure 实践验证
  • index.md — 添加 1 个新页面条目，总页数 103→104

关系图谱