Native compilation is now default, so elisp is much faster than it was out of the box.
Tree-sitter support is quite smooth now, with automatic grammar installation.
Apparently very good Android support.
Canvas support is in master, people are building all kinds of crazy stuff[0] with this
Lots of nice built-in packages like modus themes, eglot, which-key. Editable grep/xref/occur results is also built-in now. New context menu things.
Load the "newcomer-presets" theme to enable stuff that newcomers tend to want enabled (tab bar, fancy completion, lots of little things).
[0] e.g. monadicsheep.org, and youtube.com / yibie/exc.el
(update) Supertag 0.4.0: lifestreams-style stream-agent, summary
I've been working on Supertag, a package that treats tagged Org headings as typed nodes (tags with fields, queries, backlinks, streams). 0.4.0 adds the piece I've wanted for a while, and I made a short video about the idea behind it.
The problem I kept hitting. Asking an LLM to make small edits to my notes ("I talked to X today, log it") works most of the time. Only most of the time, so I end up opening the note to check, almost every time. Checking costs about as much as doing it by hand.
What 0.4.0 does instead. You open a tag's Stream (a time-ordered list of its nodes) and describe in one sentence how it should work, e.g. "group people by how close we are, and give me a key to log that we talked today."* A model writes a small Elisp agent for that Stream: groups, a line per note, and keys that write properties. After that the agent runs with no model call. Pressing the key writes :LAST_CONTACT: 2026-10-26 into the property drawer, as plain text. `R` revises the agent with another sentence.
Some boundaries I care about, which I suspect this sub does too:
- The Org files are the only truth. The Stream is a view. Agents never create notes; they only read and write declared properties (and `SCHEDULED`/`DEADLINE` if you ask for dates). Delete an agent and nothing in your files changes.
- The model only runs when you ask: when designing or revising an agent. Nothing calls a model in the background.
- Generated code is checked before install. It is built from a fixed set of primitives; every field an action writes has to be declared; the view must place every note in a group, draw a single-note stream, and redraw after each action. Failures go back to the model, up to a limit. You can read the code with `C`; the sentence and its revisions are kept as comments.
- Bring your own model. Requests go through gptel (with a `supertag` preset) or Superchat. Local models via Ollama / LM Studio work.
The idea is borrowed from Lifestreams (Eric Freeman's 1997 Yale thesis): one stream of everything, substreams by query, and processors on top. Its processors had to be written by hand (a phone-call agent took ~150 lines of Tcl), which is why they never spread. An LLM can be that author, once.
Also in 0.4.0:
- `M-x supertag-view-stream` opens one stream of all nodes; `f` narrows with plain words, `tag:prj`, `KEY:VALUE` or a full query; `u` goes back.
- Discovery adds results by meaning after keyword matches (embeddings via Ollama or any OpenAI-compatible endpoint; off by default).
- Agents can indent tasks under their project, using the outline path.
- Archived notes (`ARCHIVE` tag, `org-archive-location`, denote `archive` keyword) are hidden from streams; `A` shows them.
- denote `#+IDENTIFIER:` and `:ID:` file headers are file nodes by default.
Repo: yibie/supertag
Release notes: github.com/yibie/supertag/releases/tag/v0.4.0
Requires Emacs 29.1+, Org 9.6+, textui 0.9.0. After upgrading, run `M-x supertag-sync-full-rescan` once.
The demo uses a made-up vault; the narration is a synthetic voice. Happy to hear where this breaks, and what you'd want an agent to do for one kind of your notes.
8 points · 4 comments
nope, just need the ROM. Here's a list of recent decompiled games via LLMs. yibie/awesome-rea
grid-table — a spreadsheet inside Emacs: Excel keys, dynamic arrays, flo...
grid-table is a spreadsheet for Emacs that behaves like Excel: the keys, the formulas, the error values and the file format. This tour uses one workbook, Northwind Tea Co., about 2,400 orders on three sheets, and shows what is new in 0.4.
Code, install notes and the full key reference:
yibie/grid-table
What you'll see:
- The new view, drawn with ETAF, next to the old one
- Excel keys: typing, RET, TAB, F2, fill, copy and paste with formats, undo
- XLOOKUP, SUMIFS, FILTER, dynamic arrays that spill, LET and LAMBDA
- One edit recalculates only what depends on it
- Tracing where a number comes from, with error values that behave like Excel's
- Floating charts over the cells
- Results checked against Microsoft's own function examples
- Export to .xlsx, edit in LibreOffice, open it again in Emacs
- A workbook that reads another workbook's values
- A new formula function in three lines of Lisp
- Inside Org notes, and in a terminal
I hope there is a day, when somebody introducing Emacs to others, he can add one more thing: Emacs can runs Excel, too.
84 points · 37 comments
@yebyen @helixml If it's the Mac side, look at @yibie's strata-mlx. I've got it running full time on my M5 Pro and have an image support PR open on it. yibie/strata-mlx
3 likes · 27 views
@yuanhao 哈哈哈哈，很开心strata-mlx在这里发挥了作用。 欢迎大家来把玩：yibie/strata-mlx
1 like · 79 views
Clonaram o Photoshop com IA. A Lição Que Todo Programador Precisa Aprender
🚀 Curso "Você Já Sabe Programar. Agora Aprenda a Construir Software": construasoftware.com.br 🎓 Edunio, minha startup: edunio.com.br 🌎 Speak Online Internacional (parceria): speakonlineinternational.s.gy
Um desenvolvedor está reconstruindo Photoshop, Premiere, Illustrator e outros programas com inteligência artificial, de graça e com código aberto no GitHub. Só que a parte mais interessante dessa história não é a reconstrução: é uma promessa de compatibilidade que estava na documentação e não se confirmou quando alguém da comunidade testou de outro jeito.
Neste vídeo eu explico a diferença entre copiar a interface e reproduzir o comportamento, o que é engenharia reversa, como projetos de jogos usam comparação byte a byte como critério, o caso dos 437 arquivos PSD e por que "o resultado de um teste só demonstra aquilo que aquele teste verificou". No final, as 4 habilidades que ficam mais importantes quando a IA escreve código e um exercício prático pra você fazer.
⏱️ Capítulos 0:00 Clonaram o Photoshop com IA 1:02 Quem está por trás: ArtCraft, PhotoCraft, FilmCraft… 1:26 "Clonar" não é reproduzir tudo 2:03 Interface × comportamento 2:51 Engenharia reversa 3:19 Exemplo: a busca que funciona offline 3:45 Jogos e a comparação byte a byte 5:02 A promessa: 134 de 135 arquivos 5:37 O teste da comunidade: 437 arquivos PSD 6:22 Cuidado com a interpretação 7:02 A correção no pull request 7:41 O que um teste realmente demonstra 8:00 O sistema de agendamento que "funcionava" 9:01 Você ainda precisa aprender programação? 9:31 Habilidade 1: decompor problemas 10:00 Habilidade 2: comportamento esperado 10:28 Habilidade 3: critérios de aceitação e testes 11:34 Habilidade 4: decisões de engenharia 12:05 Exercício: a função de desfazer 13:16 Duas formas de olhar pra essa notícia 14:25 E você, o que acha?
🔎 Fontes • TechRadar — One dev is recreating 7 Adobe apps in Rust for free: techradar.com • ArtCraft no GitHub: github.com/storytold • PhotoCraft (README e status do projeto): storytold/photocraft • Issue #200 (teste com 437 PSDs): photocraft#200 • Pull request #236: photocraft#236 • Matching decompilation / engenharia reversa: yibie/awesome-rea
Se esse vídeo te fez pensar, manda pra um amigo programador.
#programação #inteligenciaartificial #photoshop
144 likes · 20 comments · 3,869 views
原来 2 天前就有用户提议 Strata 开发者期待可以添加 strata-mlx 的支持。
经过 2 天的研究后，他的结论就是「No need, mac support with one day!」
Strata#1968
这德性，以前不是穷要饭的吧？这么小气。
我管你 need 不 need 的，关我屁事，我本来只是出于研究的兴趣，才发起这个项目。
我研究的东西多了去了，不光 strata 那么一个。
不信可以看我的 Github 主页：github.com/yibie
12 likes · 2 reposts · 3,220 views
这些用 AI 反编译老游戏的项目，真正的分水岭是有人先把「什么算对」定死：要么与原 ROM 逐字节相同，要么行为测试过关；闸门定死之后，连便宜模型都干得可靠。
一个常驻的 autoresearch agent 把这些收成持续更新的索引，现在 17 条、四类：yibie/awesome-rea
用 AI 把老游戏反编译、重编译、移植到别的平台，这条线这几天一直在出结果。但让它成立的关键，是有人先回答了另一个问题：什么叫做对。
GoldenEye 007 的 PC 端口是个好例子。干这件事的是两个 coding agent——一个本地 Qwen3.8-27B 跑在单张 RTX 5090 上，一个 Claude Code——43 天、1005 个 commit。它们的验收方式是与原 ROM 逐字节对比，再配一份 porting-notes.md：每查清一类复发 bug，就必须回写进去。机械的 ABI 改动免签，动到游戏逻辑要写正式 sign-off 申请。
momo5502（Maurice Heumann）那篇更彻底：600 到 700 亿 token 反编译一款第一人称射击，99% 的函数已重建、83% 字节级精确，验收就是字节匹配 harness。最有价值的副产品是成本反过来的那一面——既然验收是机器判定的，他们把上下文压缩阈值从 90% 降到 42%（反编译的上下文大部分是易腐信息），并发现以前用便宜模型效果极差，现在便宜模型也能可靠干活。严格验收降低了成本，而不是抬高了成本。
另外几个：
· Super Mario 64 DS 的反编译现在到 11,385 / 11,389 个函数字节匹配。每个函数一张 ticket，agent 分角色（协调者、生产者、侦查、可读性审查、独立字节校验），字节校验由单独的 agent 做。2,900 多个 PR 来自 19 个人，最难的那批函数"每个几美分"。
· Xenosaga Ep I：一个人一周做到 7,645 个函数里的 3,669 个，另配一个 decomp-compat-kit 做独立字节校验。
· BattleShip（N64 大乱斗）和 ssbu-recomp（Switch 大乱斗）把验收换成了"能不能玩"——行为与可玩性，外加交叉编译到 Switch 上跑 60fps 这样的硬指标。
· 工具侧开始被真实项目点名：Hex-Rays 官方放出 ida-mcp（1,056 星），mrexodia 的 ida-pro-mcp 已有 12,707 星，momo5502 那篇里用的是官方 ida-mcp。
反面的声音比正面更值得看。
Chris Lewis 用 84 天把 Snowboard Kids 反编译到 100%，而续作花了 596 天。但他自己写得很清楚：那是续作、他本人已有两年经验、约 4.8% 的匹配提交需要人类专家介入——"把差距全归给 AI 是严重过度简化"。他还有一个更具体的观察：指望脚本用 m2c 批量自动匹配，1,830 个函数里只成了 17 个（0.93%），长尾必须人机一起磨。
Hacker News 上反编译社区的抵触也很真实：接受一个用 Claude 的 PR 会永久污染项目；这些所谓原生端口还不如模拟器，玩到第二关就崩。对立面的反驳同样清楚：反编译的输出是可证明的，要么编译出原版二进制，要么不是。
法律边界也绕不开：原生端口默认不含原游戏资产，matching decompilation 是净室重实现，而"反编译任何东西"这类工具在 EULA 与 DMCA 上的风险是真实的。
我把这条线交给了一个常驻的 autoresearch agent 去做索引：
yibie/awesome-rea
它现在 17 条、分四类：字节级验证的反编译、行为验收的移植、工具与 MCP、实践与讨论。每条一句话，并回答同一个问题——这个项目怎么定义"完成"。
它有一条硬规则：看不出验收方式的项目，不进主分类。所以两个真实且热门的高星项目被它拒收了——Sly Cooper 的反编译（273 星）和可玩端口 ProjectCane（245 星），因为仓库里找不到任何 AI 参与的声明。一个 AI 反编译列表主动拒掉高星项目，这比它收了多少条更能说明它在筛什么。
原文：yibie/awesome-rea
#反编译 #Agent #awesome
4 likes · 1 repost · 547 views
My org-attach config, only place files go (paste / drag / download), with names I can actually read
My notes had two image piles: org-download threw everything into one folder, and org-attach had files called WechatIMG7604.jpg or some 64-char hash. Every time I added a file I had to think about where it should go. Got tired of it.
Now anything that enters an Org buffer becomes an attachment of the current heading and gets a link inserted. The nice part is how little it took, because everything already ends up in org-attach-attach:
(setq org-yank-dnd-method 'attach
org-yank-image-save-method 'attach
org-download-method 'attach)
Then one :around advice on org-attach-attach does the rest:
- junk names (hashes, IMG_1234, Screenshot…) become 20261010-083713-heading-title.png; names that already mean something are left alone
- files are moved instead of copied, so ~/Downloads doesn't fill up with leftovers
s-v pastes an image as an attachment if the clipboard has one, otherwise it's a normal yank. This was the annoying bit. "Clipboard has an image" isn't enough: browsers put the URL next to the image, and spreadsheets put a rendered picture next to the cells you copied. I ended up with "image + a single line that looks like a URL or image filename = image, anything else = text". Works for the apps I use, probably not for all of them.
There's also a transient on C-c C-a (attach newest download, rename and fix the links, Quick Look, reveal in Finder).
Full file: gist.github.com
Caveats: the clipboard part is macOS only (needs pngpaste), and since it moves files, don't point it at a folder you want to keep intact.
How do you all tell "image" from "text" on paste? Curious if there's a better rule.
9 points
Mentions
Native compilation is now default, so elisp is much faster than it was out of the box. Tree-sitter support is quite smooth now, with automatic grammar installation. Apparently very good Android support. Canvas support is in master, people are building all kinds of crazy stuff[0] with this Lots of nice built-in packages like modus themes, eglot, which-key. Editable grep/xref/occur results is also built-in now. New context menu things. Load the "newcomer-presets" theme to enable stuff that newcomers tend to want enabled (tab bar, fancy completion, lots of little things). [0] e.g. monadicsheep.org, and youtube.com / yibie/exc.el
(update) Supertag 0.4.0: lifestreams-style stream-agent, summary I've been working on Supertag, a package that treats tagged Org headings as typed nodes (tags with fields, queries, backlinks, streams). 0.4.0 adds the piece I've wanted for a while, and I made a short video about the idea behind it. The problem I kept hitting. Asking an LLM to make small edits to my notes ("I talked to X today, log it") works most of the time. Only most of the time, so I end up opening the note to check, almost every time. Checking costs about as much as doing it by hand. What 0.4.0 does instead. You open a tag's Stream (a time-ordered list of its nodes) and describe in one sentence how it should work, e.g. "group people by how close we are, and give me a key to log that we talked today."* A model writes a small Elisp agent for that Stream: groups, a line per note, and keys that write properties. After that the agent runs with no model call. Pressing the key writes :LAST_CONTACT: 2026-10-26 into the property drawer, as plain text. `R` revises the agent with another sentence. Some boundaries I care about, which I suspect this sub does too: - The Org files are the only truth. The Stream is a view. Agents never create notes; they only read and write declared properties (and `SCHEDULED`/`DEADLINE` if you ask for dates). Delete an agent and nothing in your files changes. - The model only runs when you ask: when designing or revising an agent. Nothing calls a model in the background. - Generated code is checked before install. It is built from a fixed set of primitives; every field an action writes has to be declared; the view must place every note in a group, draw a single-note stream, and redraw after each action. Failures go back to the model, up to a limit. You can read the code with `C`; the sentence and its revisions are kept as comments. - Bring your own model. Requests go through gptel (with a `supertag` preset) or Superchat. Local models via Ollama / LM Studio work. The idea is borrowed from Lifestreams (Eric Freeman's 1997 Yale thesis): one stream of everything, substreams by query, and processors on top. Its processors had to be written by hand (a phone-call agent took ~150 lines of Tcl), which is why they never spread. An LLM can be that author, once. Also in 0.4.0: - `M-x supertag-view-stream` opens one stream of all nodes; `f` narrows with plain words, `tag:prj`, `KEY:VALUE` or a full query; `u` goes back. - Discovery adds results by meaning after keyword matches (embeddings via Ollama or any OpenAI-compatible endpoint; off by default). - Agents can indent tasks under their project, using the outline path. - Archived notes (`ARCHIVE` tag, `org-archive-location`, denote `archive` keyword) are hidden from streams; `A` shows them. - denote `#+IDENTIFIER:` and `:ID:` file headers are file nodes by default. Repo: yibie/supertag Release notes: github.com/yibie/supertag/releases/tag/v0.4.0 Requires Emacs 29.1+, Org 9.6+, textui 0.9.0. After upgrading, run `M-x supertag-sync-full-rescan` once. The demo uses a made-up vault; the narration is a synthetic voice. Happy to hear where this breaks, and what you'd want an agent to do for one kind of your notes.
nope, just need the ROM. Here's a list of recent decompiled games via LLMs. yibie/awesome-rea
grid-table — a spreadsheet inside Emacs: Excel keys, dynamic arrays, flo... grid-table is a spreadsheet for Emacs that behaves like Excel: the keys, the formulas, the error values and the file format. This tour uses one workbook, Northwind Tea Co., about 2,400 orders on three sheets, and shows what is new in 0.4. Code, install notes and the full key reference: yibie/grid-table What you'll see: - The new view, drawn with ETAF, next to the old one - Excel keys: typing, RET, TAB, F2, fill, copy and paste with formats, undo - XLOOKUP, SUMIFS, FILTER, dynamic arrays that spill, LET and LAMBDA - One edit recalculates only what depends on it - Tracing where a number comes from, with error values that behave like Excel's - Floating charts over the cells - Results checked against Microsoft's own function examples - Export to .xlsx, edit in LibreOffice, open it again in Emacs - A workbook that reads another workbook's values - A new formula function in three lines of Lisp - Inside Org notes, and in a terminal I hope there is a day, when somebody introducing Emacs to others, he can add one more thing: Emacs can runs Excel, too.
@yebyen @helixml If it's the Mac side, look at @yibie's strata-mlx. I've got it running full time on my M5 Pro and have an image support PR open on it. yibie/strata-mlx
@yuanhao 哈哈哈哈，很开心strata-mlx在这里发挥了作用。 欢迎大家来把玩：yibie/strata-mlx
Clonaram o Photoshop com IA. A Lição Que Todo Programador Precisa Aprender 🚀 Curso "Você Já Sabe Programar. Agora Aprenda a Construir Software": construasoftware.com.br 🎓 Edunio, minha startup: edunio.com.br 🌎 Speak Online Internacional (parceria): speakonlineinternational.s.gy Um desenvolvedor está reconstruindo Photoshop, Premiere, Illustrator e outros programas com inteligência artificial, de graça e com código aberto no GitHub. Só que a parte mais interessante dessa história não é a reconstrução: é uma promessa de compatibilidade que estava na documentação e não se confirmou quando alguém da comunidade testou de outro jeito. Neste vídeo eu explico a diferença entre copiar a interface e reproduzir o comportamento, o que é engenharia reversa, como projetos de jogos usam comparação byte a byte como critério, o caso dos 437 arquivos PSD e por que "o resultado de um teste só demonstra aquilo que aquele teste verificou". No final, as 4 habilidades que ficam mais importantes quando a IA escreve código e um exercício prático pra você fazer. ⏱️ Capítulos 0:00 Clonaram o Photoshop com IA 1:02 Quem está por trás: ArtCraft, PhotoCraft, FilmCraft… 1:26 "Clonar" não é reproduzir tudo 2:03 Interface × comportamento 2:51 Engenharia reversa 3:19 Exemplo: a busca que funciona offline 3:45 Jogos e a comparação byte a byte 5:02 A promessa: 134 de 135 arquivos 5:37 O teste da comunidade: 437 arquivos PSD 6:22 Cuidado com a interpretação 7:02 A correção no pull request 7:41 O que um teste realmente demonstra 8:00 O sistema de agendamento que "funcionava" 9:01 Você ainda precisa aprender programação? 9:31 Habilidade 1: decompor problemas 10:00 Habilidade 2: comportamento esperado 10:28 Habilidade 3: critérios de aceitação e testes 11:34 Habilidade 4: decisões de engenharia 12:05 Exercício: a função de desfazer 13:16 Duas formas de olhar pra essa notícia 14:25 E você, o que acha? 🔎 Fontes • TechRadar — One dev is recreating 7 Adobe apps in Rust for free: techradar.com • ArtCraft no GitHub: github.com/storytold • PhotoCraft (README e status do projeto): storytold/photocraft • Issue #200 (teste com 437 PSDs): photocraft#200 • Pull request #236: photocraft#236 • Matching decompilation / engenharia reversa: yibie/awesome-rea Se esse vídeo te fez pensar, manda pra um amigo programador. #programação #inteligenciaartificial #photoshop
原来 2 天前就有用户提议 Strata 开发者期待可以添加 strata-mlx 的支持。 经过 2 天的研究后，他的结论就是「No need, mac support with one day!」 Strata#1968 这德性，以前不是穷要饭的吧？这么小气。 我管你 need 不 need 的，关我屁事，我本来只是出于研究的兴趣，才发起这个项目。 我研究的东西多了去了，不光 strata 那么一个。 不信可以看我的 Github 主页：github.com/yibie
这些用 AI 反编译老游戏的项目，真正的分水岭是有人先把「什么算对」定死：要么与原 ROM 逐字节相同，要么行为测试过关；闸门定死之后，连便宜模型都干得可靠。 一个常驻的 autoresearch agent 把这些收成持续更新的索引，现在 17 条、四类：yibie/awesome-rea 用 AI 把老游戏反编译、重编译、移植到别的平台，这条线这几天一直在出结果。但让它成立的关键，是有人先回答了另一个问题：什么叫做对。 GoldenEye 007 的 PC 端口是个好例子。干这件事的是两个 coding agent——一个本地 Qwen3.8-27B 跑在单张 RTX 5090 上，一个 Claude Code——43 天、1005 个 commit。它们的验收方式是与原 ROM 逐字节对比，再配一份 porting-notes.md：每查清一类复发 bug，就必须回写进去。机械的 ABI 改动免签，动到游戏逻辑要写正式 sign-off 申请。 momo5502（Maurice Heumann）那篇更彻底：600 到 700 亿 token 反编译一款第一人称射击，99% 的函数已重建、83% 字节级精确，验收就是字节匹配 harness。最有价值的副产品是成本反过来的那一面——既然验收是机器判定的，他们把上下文压缩阈值从 90% 降到 42%（反编译的上下文大部分是易腐信息），并发现以前用便宜模型效果极差，现在便宜模型也能可靠干活。严格验收降低了成本，而不是抬高了成本。 另外几个： · Super Mario 64 DS 的反编译现在到 11,385 / 11,389 个函数字节匹配。每个函数一张 ticket，agent 分角色（协调者、生产者、侦查、可读性审查、独立字节校验），字节校验由单独的 agent 做。2,900 多个 PR 来自 19 个人，最难的那批函数"每个几美分"。 · Xenosaga Ep I：一个人一周做到 7,645 个函数里的 3,669 个，另配一个 decomp-compat-kit 做独立字节校验。 · BattleShip（N64 大乱斗）和 ssbu-recomp（Switch 大乱斗）把验收换成了"能不能玩"——行为与可玩性，外加交叉编译到 Switch 上跑 60fps 这样的硬指标。 · 工具侧开始被真实项目点名：Hex-Rays 官方放出 ida-mcp（1,056 星），mrexodia 的 ida-pro-mcp 已有 12,707 星，momo5502 那篇里用的是官方 ida-mcp。 反面的声音比正面更值得看。 Chris Lewis 用 84 天把 Snowboard Kids 反编译到 100%，而续作花了 596 天。但他自己写得很清楚：那是续作、他本人已有两年经验、约 4.8% 的匹配提交需要人类专家介入——"把差距全归给 AI 是严重过度简化"。他还有一个更具体的观察：指望脚本用 m2c 批量自动匹配，1,830 个函数里只成了 17 个（0.93%），长尾必须人机一起磨。 Hacker News 上反编译社区的抵触也很真实：接受一个用 Claude 的 PR 会永久污染项目；这些所谓原生端口还不如模拟器，玩到第二关就崩。对立面的反驳同样清楚：反编译的输出是可证明的，要么编译出原版二进制，要么不是。 法律边界也绕不开：原生端口默认不含原游戏资产，matching decompilation 是净室重实现，而"反编译任何东西"这类工具在 EULA 与 DMCA 上的风险是真实的。 我把这条线交给了一个常驻的 autoresearch agent 去做索引： yibie/awesome-rea 它现在 17 条、分四类：字节级验证的反编译、行为验收的移植、工具与 MCP、实践与讨论。每条一句话，并回答同一个问题——这个项目怎么定义"完成"。 它有一条硬规则：看不出验收方式的项目，不进主分类。所以两个真实且热门的高星项目被它拒收了——Sly Cooper 的反编译（273 星）和可玩端口 ProjectCane（245 星），因为仓库里找不到任何 AI 参与的声明。一个 AI 反编译列表主动拒掉高星项目，这比它收了多少条更能说明它在筛什么。 原文：yibie/awesome-rea #反编译 #Agent #awesome
My org-attach config, only place files go (paste / drag / download), with names I can actually read My notes had two image piles: org-download threw everything into one folder, and org-attach had files called WechatIMG7604.jpg or some 64-char hash. Every time I added a file I had to think about where it should go. Got tired of it. Now anything that enters an Org buffer becomes an attachment of the current heading and gets a link inserted. The nice part is how little it took, because everything already ends up in org-attach-attach: (setq org-yank-dnd-method 'attach org-yank-image-save-method 'attach org-download-method 'attach) Then one :around advice on org-attach-attach does the rest: - junk names (hashes, IMG_1234, Screenshot…) become 20261010-083713-heading-title.png; names that already mean something are left alone - files are moved instead of copied, so ~/Downloads doesn't fill up with leftovers s-v pastes an image as an attachment if the clipboard has one, otherwise it's a normal yank. This was the annoying bit. "Clipboard has an image" isn't enough: browsers put the URL next to the image, and spreadsheets put a rendered picture next to the cells you copied. I ended up with "image + a single line that looks like a URL or image filename = image, anything else = text". Works for the apps I use, probably not for all of them. There's also a transient on C-c C-a (attach newest download, rename and fix the links, Quick Look, reveal in Finder). Full file: gist.github.com Caveats: the clipboard part is macOS only (needs pngpaste), and since it moves files, don't point it at a folder you want to keep intact. How do you all tell "image" from "text" on paste? Curious if there's a better rule.