跳轉到

Changelog

Last 14 days of substantive changes. Older history lives in git log.

Format: YYYY-MM-DD · category · 1-line summary (commit-sha).

2026-07-01

  • content · Claude Sonnet 5 (released 2026-06-30) rolled through the repo, all first-party verified (platform.claude.com model docs + anthropic.com/news/claude-sonnet-5): claude-sonnet-4-6 → claude-sonnet-5 and Sonnet 4.6 → Sonnet 5 everywhere they name the current default Sonnet — Stage 1 model table + reading list + pricing example, glossary Context-Window / Computer-Use / Frontier-Model entries, Stage 8 Computer-Use row, setup-guide + examples/README model-picker, all walkthrough + stage-3/4 example CLI commands, the branches/for-developer Aider example (its two-generations-old claude-sonnet-4-20250514 snapshot also bumped to claude-sonnet-5), the repo CLAUDE.md picker, and the freshness-models.yml whitelist (47 content files, 43 model-ID swaps). Verified specs carried, none invented: 1M context (same as Opus 4.8), $3/$15 standard ($2/$10 intro through 2026-08-31), "best speed×intelligence" positioning; Sonnet 5 supersedes Sonnet 4.6 (now Legacy). Historical Sonnet 4.5 references (Stage 6 predecessor list, Stage 7 GAIA leaderboard) deliberately left intact. Tri-locale; anchor / zh-Hans / switcher gates + code-reviewer pass.
  • content · Stage 2 (Prompt Engineering) + glossary now spell out zero-shot / one-shot / few-shot in plain language — terms Exercise 2 leaned on but never defined. The glossary entry Few-shot / Zero-shot → Zero-shot / One-shot / Few-shot gains the previously-missing one-shot (exactly 1 example) plus a one-line framing (the three differ only in how many examples you show the LLM); Stage 2 Exercise 2 gets a 3-bullet inline explainer tying the 3-shot its code uses back to "few-shot". Also corrected a stray Traditional 分類 → 分类 in the zh-Hans exercise line. Tri-locale; anchor / zh-Hans / switcher gates + code-reviewer pass.

2026-06-30

  • content · Staleness audit Batch 1 (from the 2026-06-29 multi-agent repo audit): removed the phantom "Claude Mythos Preview" attribution on the Stage 7 WebArena benchmark row (→ "領先 model 未公布" — Mythos/Fable benchmarks were never published and access is suspended, so the cell contradicted the table's own caption); glossary Context-Window entry gains Grok 4.3 1M + Mistral Medium 3.5 256k for parity with the Frontier Model entry; cookbook "Claude 4.5+" → "Claude 4.8+"; A2A glossary entry refreshed to v1.0 (Linux Foundation governance, 150+ orgs, signed Agent Cards). All first-party verified. Tri-locale; gates pass.
  • content · Staleness audit Batch 2 (model-ID / recommendation refresh, all first-party verified): Stage 6 Path-2 reasoning list + observation line, Stage 8 example comment, and setup-guide updated Gemini 3.1 Pro → 3.5 Flash and added xAI Grok 4.3 (GA) to the strongest-reasoning options; GPT-5.5 deliberately kept in runnable example code since GPT-5.6 is still limited preview (not GA); setup-guide free-tier corrected to "GPT-5.5 Instant (rate-limited)"; Stage 4 OpenAI Agents SDK "April 2026 update" reframed to past-tense (the built-in-sandbox / 7-provider claim verified accurate), AG2 v0.2-vs-v0.4 note softened. Tri-locale; gates pass.
  • layout · Stage 1 model tables de-crammed: time-sensitive status/caveat text (Fable 5 suspension, license clauses, release dates, Arena rank) moved out of the data cells into one plain-language legend line per table; genuinely-old entries retired (GPT-5 / o-series, Gemini 3.1 Pro). Added an HTML-comment maintenance convention so the tables self-clean as models churn (new flagship = swap not append; resolved status = delete the legend line). Worst flagship cell dropped ~75 → ~33 chars. Tri-locale; plain-language; gates + code-reviewer (APPROVE) pass.
  • layout · Stage 2 Curated-Projects table de-crammed (same pattern): the two worst cells (dspy / NirDiamant, ~150-182 → ~38-74 units) trimmed to a short reason + ★ / license; framework-not-tutorial + NOASSERTION caveats moved to one legend line. Tri-locale; gates pass.
  • layout · Stage 1 Chinese-frontier table (the 7-provider / 7-col one, widest in the stage) split along the API-vs-open-weights line into two 6-col tables (① API-only: DeepSeek / Kimi / Hunyuan / MiniMax · ② open weights: Qwen / GLM / Yi); the License column folded into one plain legend per group. All 7 providers + license nuance preserved. Tri-locale; gates + column-count scan pass.
  • site · MkDocs-Material UI upgrade (verified with a local mkdocs build, clean across all 3 locales): code-copy buttons, instant/SPA-style navigation + progress, navigation.sections, search.share, content tabs/tooltips/annotate; :material-*: icon support; new docs/stylesheets/extra.css (indigo brand color, grid-card hover, tighter rounded tables). A custom card landing page is a planned follow-up (blocked on the README-vs-index i18n conflict).
  • site · Nav cleanup: the top nav was bilingual + inconsistent ("首頁 / Home", "Audience branches"…); switched to single-language source labels + i18n nav_translations so each locale shows only its own language (繁中 首頁 / en Home / 简中 首页), and dropped navigation.sections (over-expanded the sidebar). Verified with a local build, all 3 locales.
  • site · Custom card landing page shipped (resolves the earlier README-vs-index blocker): index.md / .en.md / .zh-Hans.md is now the trilingual home — hero + stat cards + track/stage grid-cards. The README moved to an /about/ page (staged as about.md so it no longer collides with index for the home slot), and mkdocs_hooks.py rewrites in-content README.md links to about at build time so they keep resolving (examples/README untouched). Verified locally: clean build, all 3 locale homes = landing, anchor / zh-Hans / switcher gates pass.
  • content · Staleness audit Batch 3 (new harness-engineering frames + adds, all first-party sourced): Stage 7 gains a "feedback loops, not a more perfect prompt" subsection — the 4 feedback timings (tool returns / mid-run steering / end-of-turn acceptance / outer loop), anchored on Anthropic's planner→generator→evaluator harness post; Stage 7.5 gains "Harnesses expire: Model-Harness-Fit + the Bitter Lesson" (Sutton 2019); deepagents (LangChain, LangGraph-based, MIT, v0.6.12) added to Stage 4 framework resources + a plain glossary "Deep Agent" entry. Written analogy-first for non-engineers, jargon glossed inline; Codex /goal folded into N1's outer-loop row (no separate section). Tri-locale; gates pass.

2026-06-29

  • content · Stage 1 model table + glossary (frontier + Context-Window entries) + scripts/freshness-models.yml whitelist refreshed with late-June-2026 frontier models, all first-party verified: GPT row gains GPT-5.6 (Sol / Terra / Luna, preview); Gemini row → 3.5 Flash (3.5 Pro in dev); glossary frontier adds xAI Grok 4.3 (GA) + Mistral Medium 3.5 (open weights, preview), relabeled by half-month (Fable 5 suspension note retained). Preview-vs-GA marked; no fabricated benchmark / context numbers. Tri-locale; anchor / zh-Hans / switcher gates pass.
  • content · Stage 6 reasoning-model table consistency follow-up: the "current (Jun 2026) frontier" intro + the GPT-5.5 and Gemini 3.1 Pro rows now flag that newer tiers exist (GPT-5.6 Sol / Terra / Luna preview; Gemini 3.5 Flash available, 3.5 Pro in dev). Existing verified rows and benchmarks (e.g. Gemini 3.1 Pro GPQA Diamond 94.3%) kept and correctly attributed; no preview-model benchmarks fabricated. Tri-locale; gates pass.

2026-06-24

  • catalog · Added DeusData/codebase-memory-mcp (★ 13.5k, MIT) to §5 Dev Collaboration — a code-intelligence MCP that indexes a codebase into a queryable knowledge graph (query structure / symbols / call paths instead of grep+read). Plain, non-marketing description (notes the re-index-after-edits + verify-load-bearing-claims caveats); tri-locale; §5 TOC count 7→9 (also corrects a pre-existing off-by-one); gates pass.

2026-06-13

  • catalog · Added 12 high-confidence repos (all gh-verified stars/license, none previously listed): microsoft/agent-framework (Stage 4); getzep/graphiti + lancedb/lancedb (Stage 6); comet-ml/opik, pydantic/logfire, NVIDIA-NeMo/Guardrails, BoundaryML/baml (Stage 7, incl. new Safety/Guardrails + Structured-Output rows); bytedance/UI-TARS-desktop + trycua/cua (Stage 8, new Computer Use Agent Stack); awslabs/mcp + ComposioHQ/composio (MCP/Skills catalog §6 / §12); microsoft/mcp-for-beginners (Stage 5.2 reading list). Tri-locale; per-section counts updated; anchor / zh-Hans / switcher gates pass.
  • content · Stage 5 — plain-language orientation box in 5.1 (Claude Code = terminal agent for devs; Claude Cowork = desktop agent for non-coders; OpenAI parallels = Codex CLI / ChatGPT agent) so beginners see Claude Code is one shape among several; plus first-use plain glosses for heavy terms (harness / orchestration / scaffolding / control plane). Tri-locale; Cowork + ChatGPT agent verified first-party; anchor / zh-Hans / switcher gates pass.
  • content · Reframed Claude Fable 5 across the roadmap after Anthropic suspended all access to Fable 5 + Mythos 5 on 2026-06-12 (US government export-control directive; status · statement; no restoration timeline). Documentation tables (CLAUDE.md / examples/ / stages 01·06·07·07.5·08 / glossary) now mark Fable 5 as suspended and currently unavailable, with Opus 4.8 as the current top usable Claude tier; recommendation pick-lists (Path-2 reasoning chooser, Computer Use vendor table, OmniParser / browser-use swap-lists) drop Fable 5 so no reader is pointed at an inaccessible model. Tri-locale; anchor / zh-Hans-localize / language-switcher gates all pass. Suspension verified against two first-party sources, no fabricated facts.
  • docs · CITATION.cff version 2026.05.19 → 2026.06.13 (was stale vs the recent content batches).

2026-06-12

  • content · Claude Fable 5 (Mythos-class, claude-fable-5, GA 2026-06-09) added as the new top Claude tier across the trilingual roadmap — model tables in CLAUDE.md / examples/ / stages 01·06·07·07.5·08 + glossary frontier entry; Opus 4.8 reframed as Opus-class flagship + Fable 5 safeguard-fallback. No fabricated context-window or benchmark numbers (Anthropic published none — marked "not yet published"). Also fixed a pre-existing claude-opus-4-7 → claude-opus-4-8 inconsistency (12980b3).
  • content · Stage 5 — new 5.6 Dynamic Workflows section after 5.5 Subagents (ecosystem-level intro + cross-link to the 7.5 deep-dive, no duplication); old 5.6 Source → 5.7, old 5.7 SDK → 5.8, all in-file refs + 7-Layer-map ranges + cross-file anchors (glossary / stages 03·06·07) relinked, tri-locale (5044008).
  • catalog · 1weiho/open-slide (★4.9k, MIT) added to §2 as an agent-native slide framework — ships Claude Code Skills, distinct from Stage 4 orchestration frameworks; tri-locale (7d3fd5d).
  • docs · MCP/Skills catalog count made drift-proof — stale 62 → robust 65+, category count reconciled to 15, across 33 files / all locales (3782dd4). Propagated the 7→8 stage reality into design notes / style-guide / reader docs (39d397a) and fixed outreach-draft count drift (25785f0).
  • outreach · send-day copy-paste packages playbook for awesome-list submissions (afd7a76).
  • content · per-chapter improvement audit (12-agent fan-out + skeptical filter) → 5 gap-fills, all tri-locale: Stage 3 lethal-trifecta security callout + MCP router note + glossary (f3bde60); Stage 1 next-token / sampling mental-model box (1bd171f); Stage 5 Hooks (L3 control layer) subsection (9d2897f); Stage 7 Loop Engineering note + glossary (eb8e64c).
  • catalog · new Web Search / Retrieval category (exa-mcp + tavily-mcp) + Context7 in Dev-Collaboration; category count 15→16 (b1718d3).
  • content · improvement-audit medium batch (6 more tri-locale gap-fills): Stage 3 structured outputs / JSON-mode (93006a8); Stage 2 reasoning-vs-CoT + Stage 5 MCP-in-2026 (Registry / FastMCP / security) + Stage 8 accessibility-tree & Playwright-MCP (ea0633e); Stage 6 RAG ingest-parsing + embedding-model selection + Stage 7 OTel GenAI conventions / pass^k·τ²-bench / MAST (2fcfc6b).

2026-05-31

  • tooling · pruned the ops-metric scripts that don't touch stars or URL validity (strategic-review action #1, scoped down per maintainer): removed scripts/snapshot-traffic.py (GitHub traffic snapshots), scripts/refresh-outreach-status.py (outreach-matrix drift), scripts/check-catalog-staleness.py (dormant-entry pinger), and the docs/traffic/ snapshot dir. KEPT the weekly stars + URL auto-update (weekly-catalog-refresh.yml + lint.yml's star-drift job) — the maintainer values the weekly cadence for star-count refresh and link-rot checking. All correctness + trilingual-parity guards intact (anchor / link-rot / mirror-sync / stage-template / banned-words / overclaim / zh-Hans-localize).

2026-05-26

  • ci · lint.yml overclaim check expanded (P3-G from audit) — promoted from case-sensitive exact-phrase to case-insensitive (grep -Fi), broadened scope to include tracks/ examples/ resources/ (which the previous narrower scope missed — letting 5 uppercase Production-grade H2 headers in examples/ slip through the earlier sweep). Strict-blocking list now includes all style-guide §3 phrases (首選 / 首选 / 唯一選擇 / 唯一选择 / 業界最佳 / 业界最佳 / 業界最強 / 全世界最好的 / 最緊迫 / the most canonical) plus English equivalents (production-grade / world-class / best-in-class / cutting-edge / state-of-the-art / industry-leading). Corpus pre-cleaned across tri-locale before flipping to strict.
  • content · overclaim residue swept across tri-locale (18 file edits) before the lint flip — 3 × ## Production-grade … H2 headers in examples/stage-{6,7}/ normalized to ## Production-ready …; 3 × inline 首選 in stages/05 / stages/06 / tracks/cli/A1 softened per style-guide §3; tracks/cli/A1 最完整的中文社群資源 → 中文社群資源豐富 (marketing → factual).
  • tooling · scripts/snapshot-traffic.py shipped — captures weekly 14-day traffic window (views / clones / referrers / paths + point-in-time totals) to docs/traffic/snapshots/YYYY-MM-DD.json so historical trend survives the GitHub API's 14-day visibility limit. Each file ~5 KB. First snapshot included (docs/traffic/snapshots/2026-05-26.json).
  • tooling · scripts/refresh-outreach-status.py shipped — reads .github/channel-partners.md, extracts PR URLs, queries gh pr view, reports drift between recorded status and live PR state (merged / closed / ghosted / approved). Report-only (text / markdown / json), --check for CI. Closes P2-F from the 2026-05-25 audit; P2-E closed by snapshot-traffic.

2026-05-25

  • tooling · scripts/check-catalog-staleness.py shipped — queries gh api repos/<owner>/<repo> for pushed_at + archived, flags catalog entries dormant >= N months (default 12) or archived. Report-only (text / markdown / json). Initial run on the 247-repo catalog surfaced 17 stale entries: 5 archived (incl. langchain-ai/langserve archived 2026-05-05 still cited as live, RooCodeInc/Roo-Code archived 2026-05-15 in setup-guide) + 12 dormant (oldest: microsoft/prompt-engine 37 mo).
  • i18n · Stage 1 + Stage 2 mirror schema resync — ## 🎯 Curated Projects regenerated from canonical (en hand-translated · zh-Hans via opencc tw2s + zh-hans-localize vocab); −358 lines of stale H3-card format replaced with compact-table parity to canonical. Also normalized 5 Stage 1 .zh-Hans H2 titles back to canonical wording + emoji. Eliminates the forward-schema drift across all 8 stages.

2026-05-19

  • catalog · microsoft/ai-agents-for-beginners added to Stage 3 選讀/進階補充 as a parallel beginner course (explicitly not a substitute for the stage's hands-on practice), tri-locale (2d83f72, 94f2d73).

2026-05-18

  • catalog · Kimi-K2 + GLM-4.5 added to §11 中文圈專用 — neutral schema, gh-verified Stars/License, tri-locale (fd81f31, ad80845).
  • ci · weekly catalog-refresh PR now guarded auto-merge: sanity guard (star-token-only diff, ≤150 lines, anchors pass) → squash-merge, else label needs-manual-review (3dc6ecd).

2026-05-17

  • docs · per-track Capstone + 4-level self-assess rubric (CAPSTONE), tri-locale (dbf1ef3, a31dde5).
  • docs · Pages unified — mkdocs at /, mdBook at /book/, one workflow; README's GitHub-only switcher stripped from rendered site (5e59c7c, 001d765).
  • docs · README positioning reframed (trilingual, English fully maintained); stale exercise-folder count corrected 27 → 23 (b4bb862, 24a87fe).
  • outreach · English-audience launch drafts — HN / Reddit / newsletters / awesome-lists (b8f365b).

2026-05-16

  • governance · CoC + SECURITY + CITATION.cff + issue-template config added, tri-locale mirrors (9aa2963, 84bc58f).
  • docs · public ROADMAP.md + learner PROGRESS.md tracker added, tri-locale (e5cc310, 3e628e9).
  • docs · GitHub Pages site (mkdocs-material, trilingual) + live docs-site badge (498932c, ea4530f).
  • i18n · zh-Hans mainland-localization pass + Lint gate blocking Taiwan-vocab/「」 drift (7f73b8a, 805ae57).
  • visuals · final ASCII concept blocks replaced with generated PNGs — 10/10 complete (tri-locale) (21a2bbf).
  • ci · actions bumped off deprecated Node20 ahead of June 2026 forced migration (c6a8c19).
  • outreach · CONTRIBUTORS — @demo112 (#14) + @Rain120 (#18) (7040738).

2026-05-15

  • content · Stage 1 §主流 LLM 家族對比 (US 3 + China 7 + Western-OSS 4 + decision tree + benchmark + caveat) (8f578bf).
  • content · Stage 5 §7-Layer Architecture Map (Claude primitives × 3 engineering disciplines) + embedded figures (5f99bbb, 1e5a12b).
  • content · subagent teaching deepened — dispatch who/how/what, vs Skill/Slash-Command disambiguation, advanced doc + figures (009ddf9, 21c555b, e8a919e).
  • content · 5 audience branches tableized (使用情境 / 流程 / Tier ladder) + academic-style polish, tri-locale (184015b, 6b7e5f6).
  • i18n · 97 broken outbound mirror anchors fixed + anchor-checker now enforces mirror files (e1991a6, ab3a6d0).

2026-05-14

  • content · NEW Stage 7.5 — Advanced Agentic Concepts (OpenAI Harness Engineering 5 principles, Why→What→How map, work-boundary diagram) (4a6bf18, e2c1d11).
  • content · Track A3 §6 advanced-concept playbooks for daily CLI work (876a457).
  • visuals · § (513×) and 🔄 (24×) symbols stripped across all user-facing docs; concept diagrams embedded as PNG × 3 locales (29eb774, d04c224).
  • catalog · 4 Anthropic-related resources added across stages (0af7fbc).
  • ci · weekly catalog-refresh workflow + --apply flag (dc91a8b).

2026-05-13

  • content · Stage 4/6/7 verified + merged to main (cdb0ae3); Stage 8 NEW — Agent Interfaces, §1-15 across 3 commits A/B/C (b83c894, 6c87a2f, 069406f).
  • content · curation positioning crystallized — exercises reframed foundational/illustrative; repo = curation hub + simple cases, depth → hello-agents (00dc046, 0206dbc).
  • content · 精選 Projects consolidated to single 適合誰 tables across Stages 0-8 + Track A (fd94d80, 19a14a8).
  • content · Stage 5 expanded (§5.1-5.6: Claude Code basics, MCP/Plugin/Skill 定位, §5.5 Subagents, Harness Internals) (2c3f1dd, f7de4e7).
  • content · Stage 6 RAG-first restructure + GraphRAG / Contextual Retrieval / Hybrid Search; 2026 frontier-model refresh (f00e2c2, acbc9dc).
  • ci · 4 checks added — anchor validator, mirror-sync reminder, 2026 freshness, stage-template enforce (a14c809, 4491e6e).
  • i18n · 8-stage tri-locale mirror catch-up via Codex + Gemini delegation; 37 legacy anchors fixed, validator → strict (8b39c75, 706d257).
  • catalog · whale (DeepSeek terminal) + a-stock-data added to Chinese ecosystem (#14) (3d375bd).

2026-05-12

  • content · examples/ bootstrapped — Stage 1 (6) + 2 (4) + 3 (6) + 4 (5) + 6 (5) + 7 (5) inline starters + folder examples, tri-locale (c1fcaa7, 8051861, 7d2c1b7).
  • content · dual-path examples — Ollama (default, cost-driven) alongside Anthropic; per-stage budget + LLM recommendation list (bc37ad8, 3fa5410).
  • content · tool-calling-tutor — installable Claude Code skill + Stage 5 §5.3 meta-example (3584669).
  • i18n · diagrams renamed .zh-Hans.png per BCP 47 / W3C convention (78797a3).

2026-05-11

  • accessibility · resources/setup-guide.md (3 langs) — addresses the dev-fluency assumption gap that subagent audit flagged across 5 non-dev branches. 5 sections covering API key registration, Python install, hello-world, Claude Code first auth, SKILL.md primer (3c88b2b). Plus 15 branch-top callouts on all 5 audience branches. resources/README.{en,zh-Hans}.md created for trilingual parity.
  • accessibility · README — promoted setup-guide pointer to top of Quick Start across all 3 langs (ad47706). Was buried in Related Resources where non-dev visitors hit technical walls before discovering it.
  • accessibility · setup-guide opens with a 4-tier on-ramp (Web / Desktop / CLI / API) + official download URLs for Claude.ai, ChatGPT, Gemini, Le Chat, Claude Desktop, ChatGPT Desktop, LM Studio (3c89952). Replaces the abstract "decide two things" intro so non-dev readers see "just use claude.ai for free" as the first option, not "register API key → install Python".
  • accessibility · setup-guide adds a 3rd tier between Desktop and CLI: IDE with built-in AI (Cursor, Windsurf, Cline, Continue, Roo Code, Zed, GitHub Copilot) with download URLs (7e14093). Distinguishes "AI sidekick while you write code" from "agent runs autonomous task in terminal".

2026-05-10

  • funnel · Stage 1 → Stage 2 callouts added across 3 langs to address visible drop in traffic/popular/paths (0ee2a3a)
  • outreach · 3 awesome-list targets backfilled into channel-partners matrix from launch-checklist: travisvn/awesome-claude-skills, WangRongsheng/awesome-LLM-resources, AiHubCN/Awesome-Chinese-LLM (90a6ad1)
  • outreach · PR #6135 to punkpeye/awesome-mcp-servers — addressed bot name-check, replied to non-applicable glama-check + emoji-check (81a7313)
  • content · Cookbook Recipe 6 — Local-LLM × CLI Agent walkthrough (5855852). Bridges Stage 1 (local LLM) + Stage 5 (CLI agent) end-to-end. Explicitly notes Claude Code does not support local LLM as backend; routes readers to OpenCode / goose / Aider / Hermes instead. Stage 5 + cli-agents-guide also gain matching pointers.
  • catalog · Hermes Agent (NousResearch/hermes-agent ★142k) added as 7th major CLI agent across cli-agents-guide, tracks/cli/A1, and 5 dependent files (698f13a). Differentiator: cloud-VM-native, model-neutral (200+ LLMs via OpenRouter / NIM / GLM / Kimi / etc.), self-improving skill loop.
  • i18n · *.zh-CN.md → *.zh-Hans.md migration per BCP 47 / W3C compliance (21b653d). 25 files renamed, ~270 markdown lines updated, tooling (sync-language-switchers.py, lint.yml, generate-stage5-stack.py) migrated. Thanks @xfq (W3C i18n lead) for flagging in #9. Added to CONTRIBUTORS (868691d).
  • visuals · English README hero (banner.en.png), Learning Map (learning-map.en.png), and Branch Decision Tree (branch-decision-tree.en.png) refreshed to ChatGPT-rendered versions (c7edff8, 4be6b88, 6c03c58).

2026-05-09

  • outreach · Day 1 PR sent: punkpeye/awesome-mcp-servers#6135, adding awesome-agentic-ai-zh to ## Tutorials (a0dc4d5). Plan revised after upstream audit caught hesreallyhim/awesome-claude-code mid-reorg (Day 2 = issue not PR) (708259c).
  • outreach · 8 channel-partner pitch templates created in .github/outreach/ plus tracking matrix .github/channel-partners.md (2f63745). Targets: Datawhale, liyupi, HuggingFace, LangChain (kyrolabs), awesome-claude-code, awesome-mcp-servers, Zhipu, Moonshot.
  • catalog · 11 中文圈專用 expanded from 2 → 7 entries: QwenLM/Qwen-Agent, coze-dev/coze-studio, coze-dev/coze-loop, liaokongVFX/LangChain-Chinese-Getting-Started-Guide, chatchat-space/Langchain-Chatchat (4809039).
  • funnel · Stage 0 → Stage 1 callouts added (3dfe761).
  • ci · zh-Hans companion files excluded from zh-TW banned-word audit (closes #7) (3acc3f2).

2026-05-08

  • content · for-teacher branch expanded with 3-tier teacher AI use-case framework (Chen 2020, Mittal 2024) via @scott0127 PR #6 (cd1cad4).
  • content · Stage 6 unit guide: memory + RAG overview via @scott0127 PR #5.
  • content · Branch decision tree (zh-Hans) added, English banner added, for-developer branch thickened 56 → 138 lines × 3 langs.

2026-05-07

  • catalog · 3 user-flagged gaps filled: safishamsi/graphify, pbakaus/impeccable, netease-youdao/LobsterAI + context-engineering and harness-engineering coverage.
  • content · resources/cookbook.md added with 5 (now 6) step-by-step recipes covering Skill / MCP / Office / NotebookLM / Zotero / Local-LLM workflows.

2026-05-06

  • launch · Repo announced to bilingual community. Star count: 0 → 519 in week one.
  • content · learning-map.png polished, README hero banner placement finalized.

Conventions

  • Each commit SHA is clickable: https://github.com/WenyuChiou/awesome-agentic-ai-zh/commit/<sha>
  • Categories: content (stages/branches/tracks) · docs (project meta-docs: README/ROADMAP/PROGRESS/CAPSTONE/Pages site) · governance (CoC/SECURITY/CITATION/issue templates) · accessibility (on-ramp/setup friction) · catalog (mcp-skills-catalog entries) · funnel (cross-stage navigation) · visuals (diagrams/banners) · i18n (translation/locale) · outreach (channel partners) · ci (workflows/lint) · launch (one-time events)
  • Maintained manually; not auto-generated. Updated alongside substantive commits.