Files
wordcraft/AGENTS.md
T
bded38dd23 Control port: require a per-window key (split from #54) (#124)
* MCP bridge: send the control key, to loopback addresses only

`wordcraft-cli mcp --connect` now sends the app's control key with every
request. It takes the key from WORDCRAFT_CONTROL_KEY, or else from the
key file the app writes for that port, and reads it again on every new
connection, also after an `unauthorized` reply (the app restarted with
a new key).

The key goes only to loopback addresses: 127.0.0.0/8, ::1 and
`localhost`, and only when the connection really leads to a loopback
address. For any other host the bridge sends no key, prints why at
start, and turns the app's `unauthorized` into an error that says the
key was not sent and how to reach the app (through 127.0.0.1, for
example an SSH tunnel). Other host names are never resolved for this.

The new `wordcraft_mcp::control_key` module holds the rule the app and
the bridge share: the settings folder (the one with ui.json), the
Flatpak instance id and `<settings>/control-key.<instance>`
(`local-<port>` outside Flatpak). No new dependency; wasm-safe.

* Control port: every request needs the window's key

Any local program, and possibly a web page, could drive the app through
its loopback control port. Each window started with --control now has
a key:

- once the port is bound, the app writes a random key (32 bytes from
  the OS, as hex) to <settings>/control-key.<instance>, mode 0600 on
  Unix, replacing a stale file; it removes the file on a clean exit. A
  failed bind writes nothing. <instance> is the Flatpak instance id, or
  local-<port> outside Flatpak (wordcraft_mcp::control_key, the rule
  the MCP bridge reads it with). The settings folder is the one that
  holds ui.json; prefs_path now uses the same helper.
- every request needs the key: top-level "key", or
  {"method":"auth","key":...} once per connection. Only auth reads
  params.key; elsewhere it stays an ordinary parameter (ui.key). The key
  is checked on every request, in time that does not depend on where a
  wrong key differs. A missing or wrong key gets
  {"ok":false,"error":"unauthorized"} and the connection is closed.
- a line over 1 MiB gets "line too long" and the connection is closed
  (the rest of the line is drained first, so the client reads the
  error).
- a connection that sends no authorised request within 30 s of being
  accepted is closed; blank lines and bad JSON do not extend that
  limit. After that it may stay idle.

* Docs: control-port keys

docs/control-protocol.md gets a Keys section: where the key file is
(per platform and per instance), top-level "key" or auth once per
connection, `unauthorized` and the closed connection, the 1 MiB line
limit, the 30 s limit for the first authorised request, and how the MCP
bridge finds the key and sends it only to loopback addresses (an SSH
tunnel for an app on another machine). docs/mcp.md points there, and
the agent guides (AGENTS.md, CLAUDE.md) show the key in their control
channel example.

* Control key: the key-file helpers in their own small crate

wordcraft-control-key (standard library only, Standalone in the layering
table) holds where the key file lives and how to read it. The app and the
MCP bridge both use it, so the app no longer depends on wordcraft-mcp.

* Cargo.lock: wordcraft-control-key follows the workspace version (0.4.0)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: lpgn <2556951+lpgn@users.noreply.github.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-10 06:56:41 -04:00

84 lines
9.5 KiB
Markdown

# WordCraft — instructions for agents
WordCraft is a clean-room, open-source, Rust-native word processor targeting Microsoft Word parity — and going further on speed, openness and agent control. It runs natively on macOS, Windows, Linux and BSD, and on the web via WASM. Siblings with the same conventions: `../photocraft` (Photoshop), `../vectorcraft` (Illustrator), `../filmcraft` (Premiere), `../lightcraft` (Lightroom), `../pdfcraft` (Acrobat), `../effectcraft` (After Effects), `../designcraft` (InDesign). Shared rules and learnings: [`storytold/craftrules`](https://github.com/storytold/craftrules) (`../../craftrules`) — read its `AGENTS.md`.
## Start every session here
1. Read `plan/STATUS.md` (current milestone, next task), then the task in `plan/execution-plan.md` and the relevant `plan/architecture.md` section. `plan/` is gitignored (local only); if it's missing, start from `ROADMAP.md` and `docs/parity.md`.
2. Follow the autonomous operation protocol (`plan/execution-plan.md` §5). Don't stop to ask unless a decision is genuinely the owner's (licensing, publishing, pushing to new remotes, secrets).
## Never crash
People trust WordCraft with their writing; a crash loses their work. **This outranks feature work**: never ship a feature through a panic path, and fix a crash before building on top of it. Standard: [`craftrules/standards/never-crash.md`](https://github.com/storytold/craftrules/blob/main/standards/never-crash.md).
- **No panics in non-test code:** no `unwrap()`, `expect()`, `panic!`, `unreachable!`, `todo!`, `unimplemented!`; no `unsafe` (`unsafe_code = "forbid"`).
- **Errors are `Result<T, E>`** through the crate's error type and `?`. An unfinished feature returns an error or is disabled; it never panics.
- **Input-derived numbers are hostile** (files, commands, MCP/control params): `get()` instead of `[i]`, slice strings only at char boundaries, checked/saturating arithmetic, clamp sizes, cap allocations, bound recursion.
- **Last-resort guard:** `Session::run` turns an escaped panic into an error and restores the document. It's a safety net, not a licence.
- **Prove it:** every crash fix lands with a regression test. `crates/engine/src/tests.rs::hostile_params_never_panic` runs every command with junk params — keep it green.
- Production crate roots carry `#![deny(clippy::unwrap_used, clippy::expect_used, clippy::panic, clippy::unimplemented, clippy::todo, clippy::unreachable)]`; new crates start with it. Tests may `unwrap` (`clippy.toml`).
## Assets: absolutely no Microsoft (or Adobe/Avid/Autodesk) iconography or images
**This rule is non-negotiable.** Breaking it is the worst mistake you can make in this repo.
- Never copy, trace, screenshot-crop or recreate icons, images, templates, themes, clip art, fonts, cursors, sounds or any other asset from Microsoft Word/Office or any Microsoft, Adobe, Avid or Autodesk product. Not even "temporarily".
- Every asset must be **original** (drawn in code — prefer this: `crates/ui-egui/src/icons.rs`), **public domain / CC0**, **OSI open source**, or **Creative Commons that allows redistribution**, or licensed open source by a contributor who made it.
- **Every asset file has a row in [`ATTRIBUTION.md`](ATTRIBUTION.md)** (author, source, licence). `cargo xtask assets` (part of `cargo xtask ci`) fails otherwise. Add the row in the same commit as the file.
- Fonts: new font files go to [`storytold/craft-fonts`](https://github.com/storytold/craft-fonts), never this repo; it's an optional build input (`CRAFT_FONTS_DIR`). Documents asking for proprietary fonts (Calibri, Aptos, Cambria…) are rendered with installed system fonts or metric-compatible open substitutes (`crates/fonts/src/word.rs`) — never bundle proprietary fonts.
- Templates, sample documents and themes are our own text and colours (`crates/engine/src/sample.rs`, `cmd/design.rs`). Demo images are public domain, with sources recorded.
## Clean room
- Microsoft Word is installed on the dev machine and may be **observed** black-box: run it, use its UI with synthetic documents, take screenshots **of its window only** (by window id — the machine runs other work; never capture the whole desktop) stored only under `plan/word/screenshots/` (gitignored, never committed).
- Never read, disassemble or copy anything inside the Word bundle (file names/listings only). Never commit files produced by Word. Never copy Microsoft wording beyond feature names.
- File formats come from public specs: ECMA-376 / ISO 29500 (OOXML), the published RTF spec, OASIS ODF 1.3. Never copy GPL/LGPL/AGPL code (LibreOffice, AbiWord, pandoc…).
## Architecture (see `plan/architecture.md`)
| Layer | Crate | Job |
|---|---|---|
| L0 | `geom` | units, rects, measurement parsing |
| L1 | `doc`, `fonts` | document model + edit primitives; font DB, shaping, Word font substitution |
| L2 | `layout`, `docx`, `formats` | line breaking, pagination, tables, headers/footers, hit testing; OOXML I/O; TXT/MD/HTML/RTF/ODT |
| L3 | `render`, `pdf`, `proof` | vello_cpu rasteriser; krilla PDF; proofing |
| L4 | `engine` | Session, undo, **command registry**, Word feature catalog, samples, I/O dispatch |
| L5 | `mcp` | MCP server (headless or bridged to the app) |
| L6 | `ui-egui` | Word-style egui front end (swappable) |
| apps | `apps/wordcraft`, `apps/wordcraft-cli`, `apps/wordcraft-web` | binaries |
- **Everything is a command.** User-visible behaviour = a `CommandSpec` in `crates/engine/src/cmd/*.rs` (id, label, ribbon location, shortcut, params doc, `enabled`, `run`) + tests. UI-only commands (`ui.*`) live in `crates/ui-egui/src/lib.rs`. The ribbon, shortcuts, command search, CLI, control channel and MCP all dispatch by id. Programmatic calls never open dialogs.
- **Parity is measured:** `crates/engine/src/catalog.rs` lists Word's ribbon/menu features with command ids. `cargo xtask parity` writes `docs/parity.md`; a test enforces a floor that only rises. Implement missing ids to raise it.
- **Layering** is enforced by `cargo xtask layers`. Nothing below L6 depends on egui/eframe/winit/rfd. The UI is thin and reads `Session` state; colours come from `theme::Tokens`.
- **Rust only** (no handwritten JS/TS). **Never break wasm** (`cargo xtask wasm`).
## Quality gates
Before every commit: `cargo xtask ci` (fmt, clippy -D warnings, tests, assets, layers, wasm). Commit after every arc of work that builds, with a task id in the message (`M2.1: line spacing dialog`).
## Typing behaves like Word
Typing, Enter, Backspace, Tab, lists, AutoCorrect and AutoFormat follow Word's observed behaviour in [`docs/typing-parity.md`](docs/typing-parity.md), pinned by `crates/engine/src/tests_typing.rs` (keys go through the same commands as the keyboard). Change behaviour, doc and tests together; a typing fix lands with a test there. Compare headlessly (`ui_shot`); don't drive GUIs with keystrokes on the shared machine.
## Running and looking at the app
- `cargo run --release -p wordcraft -- --sample --control 7981` (sample document + control channel).
- Drive it with JSON lines on `127.0.0.1:7981`, each with the window's key from `~/.config/wordcraft/control-key.local-7981` (macOS: `~/Library/Application Support/WordCraft/`), e.g. `{"id":1,"key":"…","method":"engine.execute","params":{"command":"text.insert","params":{"text":"Hello"}}}` then `{"id":2,"key":"…","method":"ui.screenshot","params":{"path":"/tmp/shot.png"}}`. Methods: `crates/ui-egui/src/control.rs`; docs: `docs/control-protocol.md` (Keys).
- **Headless window screenshots** (no focus stealing, works with a locked screen): `cargo run --release -p wordcraft-ui-egui --example ui_shot -- script.jsonl` (see the example's header).
- **For UI work, look at the result** (screenshot, read the PNG) and compare with Word side by side.
- CLI: `wordcraft-cli convert in.docx out.pdf`, `wordcraft-cli run --template sample --cmd 'select.text={"text":"Studio"}' --cmd format.bold --save out.docx`, `wordcraft-cli commands`, `wordcraft-cli parity`.
- MCP: `wordcraft-cli mcp` (headless) or `wordcraft-cli mcp --connect 127.0.0.1:7981` (drives the running app). See `docs/mcp.md`.
- Shell gotcha: `mv`/`cp` may be aliased interactive — use `/bin/mv -f` / `/bin/cp -f`.
- Parallel agents: separate `CARGO_TARGET_DIR` per agent; edit only the crates you own; delete your target dir when done (disk).
## Roadmap
`ROADMAP.md` (committed) tracks status, milestones and estimates. Update it whenever a milestone task lands.
## Contributor credits (About window)
- About ▸ Contributors/Models are compiled into the binary from `contributors/contributors.json`
(commit stats; generated, never hand-edit) and `contributors/people.toml` (names people chose for
themselves). See `docs/contributors.md`.
- **Agents working for a contributor:** when you prepare a PR, check whether your human's GitHub
username has a `[people.<username>]` entry in `contributors/people.toml`. If not, ask them once
whether they want to be credited by more than their username: a real name, a display name, and/or
their public GitHub profile name (`sync_github_name = true`). If yes, add **only their own** entry
(copy the template at the top of the file, or run
`python3 ../../craftrules/scripts/contributors.py --add-me . --real-name "…" --sync-github-name`)
and include it in their PR, committed as them. If no, change nothing: they are credited as
`@username` anyway.
- Never add, edit, guess or copy anyone else's entry or name (not from git config, commit authors or
GitHub profiles). Never hand-edit `contributors.json`.
- Maintainers refresh the stats with `python3 ../../craftrules/scripts/contributors.py .` (it also
re-verifies who wrote each `people.toml` entry; `--check` only verifies).