acstack 0.0.0-stage → 0.47.1

package/docs/terminal-workspace.md

@@ -0,0 +1,1780 @@
+# Terminal workspace: herdr, gh-dash, diffnav, lazygit
+
+This is an optional setup for running several coding agents on a remote machine, and reviewing their work without leaving the terminal.
+
+- **Your code lives on a machine you reach over the network**
+- You run agents on it
+- **A plain terminal session dies with its connection**, taking every running agent with it
+- **The tools worth reviewing code in assume a desktop**, meaning a clipboard, a browser, and a display
+- **That remote machine has none of them**
+    - A copy or an "open in browser" goes nowhere
+- **ACStack needs none of this setup**
+    - Skip the whole document if you work locally in an editor
+
+- Run herdr on the dev machine, a terminal multiplexer that holds many terminal sessions
+- herdr keeps every session running after you disconnect
+- Attach to it from your own machine with `herdr --remote`
+- A pane is one terminal window inside herdr
+- The server keeps every pane alive when your connection drops
+- The local client is what lets your own clipboard and browser reach the session, which an SSH shell cannot do
+- Run `/acstack:terminal-workspace` to install and configure all of it
+- One file, `~/.config/acstack/terminal-workspace.yaml`, turns each part on or off
+- The rest of this document explains what that sets up and why
+- Read it if you would rather configure it by hand or change the defaults
+
+```mermaid
+flowchart TD
+    subgraph yours["Your machine, where you type"]
+        EM["Terminal emulator<br/>(Windows Terminal, iTerm2, kitty)"]
+        CL["herdr client<br/>(herdr --remote)"]
+        OUT["Clipboard and browser<br/>(acstack-clip, acstack-open)"]
+    end
+    subgraph dev["Dev machine, where the code is"]
+        SRV["Terminals that survive disconnect<br/>(herdr server and its panes)"]
+        RD["Reading code and markdown<br/>(yazi, acstack-md)"]
+        PR["Diffs and pull requests<br/>(gh-dash, diffnav)"]
+        DB["A SQL editor on one key<br/>(harlequin, acstack-db)"]
+    end
+    EM -->|"keys you press"| CL
+    CL -->|"SSH"| SRV
+    SRV --> RD
+    SRV --> PR
+    SRV --> DB
+    RD -->|"clickable links and copied text<br/>(OSC 8, OSC 52)"| OUT
+    PR -->|"clickable links and copied text<br/>(OSC 8, OSC 52)"| OUT
+    OUT -->|"no browser or clipboard reachable"| FB["A link on screen for you to click<br/>(acstack-open link route)"]
+```
+
+| Part of the diagram | Section |
+| --- | --- |
+| herdr client | 4. How you attach |
+| herdr server and its panes | 5. The herdr server |
+| Terminal emulator | 6. Freeing your emulator's keys |
+| Reading code and markdown | 7. Reading code, markdown, and diagrams |
+| A SQL editor on one key | 8. Harlequin: a SQL editor on one key |
+| Diffs and pull requests | 9. Diffs and pull requests |
+| Clipboard and browser | 10. What crosses back to your machine |
+| A link on screen for you to click | 10.2. acstack-open: the browser |
+
+- Sections 1 and 2 cover what you get and the loop you run daily
+- Section 3 installs it
+- Section 11 explains why the parts work the way they do
+- Section 12 lists the things that cost real time to find
+
+## 1. What you get, and its parts
+
+*Five tools you drive day to day. herdr owns the terminals, gh-dash triages pull requests, diffnav reads diffs, lazygit commits them, harlequin queries the database. Five more tools and ACStack's own helpers sit behind them. Use this when your code lives on a dev machine, you drive more than one agent at a time, and you want a keyboard-only workflow.*
+
+- Skip it if you work locally in a graphical editor
+    - A desktop tool serves you better
+- **Agents that survive disconnect.** herdr runs a background server. Closing the terminal, dropping the network, or attaching from another machine never stops an agent mid-task
+- **One screen that shows who needs you.** herdr marks every pane working, blocked, idle, or done
+    - You stop hunting for the stuck one
+- **Review without a browser.** Pull request triage, diffs with a file tree, and comments, all from the terminal
+
+The five you drive day to day:
+
+| Tool | Role | Why this one |
+|---|---|---|
+| [herdr](https://herdr.dev) | Terminal multiplexer | Background server, per-pane agent state, one container of tabs and panes per git worktree |
+| [gh-dash](https://github.com/dlvhdr/gh-dash) | Pull request dashboard | Saved searches per section, custom actions, `gh` underneath |
+| [diffnav](https://github.com/dlvhdr/diffnav) | Diff reader | File tree beside the diff, the blast-radius view a plain pager lacks |
+| [lazygit](https://github.com/jesseduffield/lazygit) | Git TUI | Stage, commit, amend and push without leaving the keyboard, and LazyVim already binds it |
+| [harlequin](https://github.com/tconbeer/harlequin) | SQL editor | Query, results and schema tree in one window, with adapters for postgres and Snowflake |
+
+Five more tools do the work those five ask for.
+
+- You drive one of them yourself, tuicr, from the moment a review starts on gh-dash's `D` key
+
+| Tool | Role | Why this one |
+|---|---|---|
+| [delta](https://github.com/dandavison/delta) | Pager | Syntax highlighting for every diff, and gh-dash uses it internally |
+| [tuicr](https://github.com/agavra/tuicr) | Code review | Vim keys, and pull request review in the same tool gh-dash hands work to |
+| [yazi](https://github.com/sxyazi/yazi) | File manager | Live preview per file type, so markdown renders as the cursor moves |
+| [mermaid-ascii](https://github.com/AlexanderGrooff/mermaid-ascii) | Diagram renderer | Draws a mermaid diagram, meaning one written as text inside the file, as box-drawing art |
+| [glow](https://github.com/charmbracelet/glow) | Markdown pager | Readable markdown in the terminal |
+
+ACStack adds two helpers of its own beside those tools.
+
+- `acstack-md` reads markdown and draws its mermaid diagrams
+- `acstack-clip` and `acstack-open` carry your clipboard and your browser back across the link
+    - Section 10 covers that pair
+
+## 2. The daily loop
+
+*Six steps, from dispatching an agent to tearing its worktree down. Four of them start with one key.*
+
+Three terms, then the loop.
+
+- **A workspace** is one herdr container of tabs and panes, usually one per repository or worktree
+- **A worktree** is a second checkout of the same repository on its own branch
+- **Every key below** is one this setup binds (see sections 3 to 10)
+
+1. **Dispatch.** Press `alt+r` to create a worktree workspace. Or ask an agent, and the `new-worktree` skill registers one with herdr for you
+2. **Monitor.** Press `alt+s` for the sidebar, the strip listing every workspace and every agent.
+
+    A blocked agent needs an answer from you. A done agent has finished and you have not looked yet. Leave a working agent alone
+
+3. **Review locally.** Press `alt+f` to see what the agent's branch adds to dev. Press `alt+x` to compare something else. Run the `pre-commit-review` skill in the agent's own pane for a deeper pass
+4. **Ship.** The agent commits, pushes, and opens the pull request from its worktree
+5. **Review the pull request.** Press `alt+i` for the dashboard. Press `p` then `]` to reach Files Changed. Press `D` to review it in tuicr, or `S` to hand it to an agent
+6. **Tear down.** Run `herdr worktree remove --workspace <id>`. Then delete the branch
+
+## 3. Install
+
+*One command installs every tool and writes every config. The rest of this section is the same work done by hand.*
+
+The `/acstack:terminal-workspace` skill runs `assets/terminal-workspace/setup.sh` for you, and that script takes six commands.
+
+| Command | What it does |
+| --- | --- |
+| `setup.sh apply` | installs every enabled tool and writes its config |
+| `setup.sh status` | reports what this machine ended up with |
+| `setup.sh outdated` | reports which tools are behind and which components are new, and changes nothing |
+| `setup.sh upgrade [tool...]` | reinstalls what `outdated` reported, and adds any new component's config block to your file |
+| `setup.sh disable <component>` | removes one component's keys and config |
+| `setup.sh uninstall` | removes everything `apply` wrote, in every config it touched |
+
+What `outdated` answers in its exit code:
+
+- Exit 0 means nothing to do, and 1 means a missing, new, or stale finding
+- Exit 2 means it cannot read the config or the shipped example
+- It reads herdr's stable release manifest, so a machine on the preview channel sees herdr reported stale
+- Name a tool on `upgrade` to skip the version comparison, which is the only way to refresh `mermaid-ascii`
+    - `mermaid-ascii` has no version flag, so nothing can read the installed version
+
+What `uninstall` removes:
+
+- It strips the managed blocks from the herdr, tuicr, and yazi configs
+- It unsets the delta git settings
+- It deletes the helper scripts
+- It removes the keybindings it wrote into gh-dash
+- It deletes `~/.config/nvim/lua/plugins/acstack.lua`, a file it wrote whole
+- It takes back what the `pi` component added to pi (section 3.6)
+- Your own settings in those files survive, and so does every tool binary
+    - harlequin is the exception, because uv owns its whole environment rather than one file in `~/.local/bin`
+    - pi stays, because it holds your logins and sessions
+
+What the config holds:
+
+- The script reads `~/.config/acstack/terminal-workspace.yaml`, which the skill copies from `assets/terminal-workspace/config.example.yaml`
+- The config holds twelve components, each with its own `enabled` key
+    - You turn a part off there and run `apply` again
+- Eight components name a tool this setup installs: `herdr`, `gh_dash`, `diffnav`, `delta`, `tuicr`, `lazygit`, `yazi`, and `harlequin`
+- `neovim` names an editor you installed yourself, and writes one file into its config
+    - It starts off, and section 3.5 says why
+- `pi` names a coding agent, which npm installs, and ACStack then installs into
+    - It starts off too, and section 3.6 says why
+- The other two name a feature
+    - `markdown` covers `acstack-md` with mermaid-ascii and glow
+    - `remote` covers the clipboard and browser helpers
+- Config keys use underscores
+    - gh-dash's key is `gh_dash`
+
+Every setting has a default, so `apply` works with no config file and says so when it finds none.
+
+- The two gh-dash review keys used to be the exception
+    - That writer read `review` and `agent_review` straight from the yaml, with no default of its own
+    - A machine with no config had both pruned and neither written back
+    - `o` already defaulted
+    - It kept working, and the config looked healthy
+- Both keys now default like everything else, and an empty string still turns one off
+- A `gh_dash:` heading with nothing under it is what commenting the block out leaves behind
+- That now falls back to the defaults too, where it used to abort the gh-dash step with a traceback
+- The script edits between its own marker comments
+    - Hand-written config around it survives
+- gh-dash is the exception
+    - Python rewrites that config whole
+    - No comment survives the round trip
+    - The script therefore tracks its own keybindings there by name
+
+### 3.1. What `apply` installs, and how to do it by hand
+
+*`apply` installs eight tools into `~/.local/bin`, herdr through its own installer, gh-dash as a `gh` extension, and harlequin through uv. None of it needs root.*
+
+Two things to do first:
+
+- Put `~/.local/bin` on your `PATH`
+- Install the `gh` command
+
+Where each binary lands:
+
+- Every binary except herdr lands in `~/.local/bin`, or in `$ACSTACK_TW_BIN` when you set that
+- herdr's own installer always uses `~/.local/bin`
+
+By hand, herdr and gh-dash install like this:
+
+```bash
+# herdr
+curl -fsSL https://herdr.dev/install.sh | sh
+
+# gh-dash, as a gh extension
+gh extension install dlvhdr/gh-dash
+```
+
+Eight more are single static binaries.
+
+- Download each project's Linux x86_64 release
+- Put the binary in `~/.local/bin`, then `chmod +x` it
+
+| Tool | GitHub repository | Release asset |
+|---|---|---|
+| delta | `dandavison/delta` | `x86_64-unknown-linux-gnu` |
+| diffnav | `dlvhdr/diffnav` | `Linux_x86_64` |
+| fzf | `junegunn/fzf` | `linux_amd64` |
+| lazygit | `jesseduffield/lazygit` | `linux_x86_64` |
+| tuicr | `agavra/tuicr` | `x86_64-unknown-linux-gnu` |
+| yazi | `sxyazi/yazi` | `x86_64-unknown-linux-gnu`, which also carries `ya` |
+| mermaid-ascii | `AlexanderGrooff/mermaid-ascii` | `Linux_x86_64` |
+| glow | `charmbracelet/glow` | `Linux_x86_64` |
+
+harlequin is the one tool uv installs, rather than a release tarball.
+
+- `uv tool install harlequin --with harlequin-postgres --with harlequin-snowflake` installs the editor and both adapters
+- A machine without `uv` skips the component, and `apply` prints the one command that installs uv
+
+Three more steps `apply` handles that no release tarball covers:
+
+- `ya pkg add yazi-rs/plugins:piper`, the yazi plugin that lets `acstack-md` draw the preview pane
+- `pip install --user markdown`, which `acstack-md --serve` needs
+- a copy of `mermaid.min.js` at `~/.local/share/acstack/mermaid.min.js`
+    - A served page fetches nothing from a content delivery network (CDN)
+
+### 3.2. Let herdr read agent state from the agent
+
+*A hook reports what an agent is doing, rather than herdr guessing from the screen.*
+
+```bash
+herdr integration install claude    # also: codex, opencode, copilot, and others
+```
+
+- Check what is available with `herdr integration status`
+- The hook applies to a session that starts after you install it
+- The `herdr.integration` key in the config picks which one `apply` installs
+
+### 3.3. The tuicr fork build is temporary
+
+*Two of the config keys for tuicr exist only in a fork build, which also carries one bug fix. Leave `build_from_fork: false` unless you want them.*
+
+Two upstream pull requests are still open:
+
+- [agavra/tuicr#607](https://github.com/agavra/tuicr/pull/607) by
+  [antonio2368](https://github.com/antonio2368) – configurable per-file
+  `+added -removed` counts in the tree, and the `show_file_line_stats` key
+
+- [agavra/tuicr#633](https://github.com/agavra/tuicr/pull/633), opened from ACStack's own
+  fork – move the file list boundary with `<leader>L` / `<leader>H`, and the
+  `file_list_width` key
+
+The fork carries a third change of its own, with no upstream pull request yet.
+
+- It fixes blank `+N -N` counts in pull request review mode, `tuicr pr <N>`
+- That mode has no version control system (VCS) backend of its own
+    - It borrowed the `File` type, which `--file <path>` uses, as a stand-in
+- The fork's whole-file gate hides counts for `--file` and `--all-files`, since every line there counts as added
+- That gate matched a review session too
+    - The counts stayed blank even with `show_file_line_stats` on
+- The fix gives a review session its own `PullRequest` VCS type
+    - The gate no longer matches it
+- Leave `build_from_fork: false` unless you want one of those three changes
+- The default installs the stock release and skips both config keys
+    - tuicr then never warns about an unknown key
+- Setting `build_from_fork: true` clones the fork and builds `local/daily` with cargo
+    - It also writes the two keys plus `no_update_check = true`
+        - `tuicr update` cannot replace the build
+    - `apply` replaces a value you set by hand on any of those keys rather than joining it
+        - TOML forbids two of one key, and tuicr ignores a config it cannot parse
+
+On a machine where you did set it true:
+
+- `setup.sh status` reports the state of both pull requests
+- After both merge, check whether anyone has submitted the counts fix above upstream too, then set `build_from_fork: false` again
+- Stock tuicr 0.23.0 has no counts feature at all
+    - It never had this bug or the fix
+- Switching back before that fix lands upstream brings the blank counts back to review mode
+- Re-run `apply` once you do switch
+- Check the merged key names first
+    - The review can rename them
+
+### 3.4. Pin tuicr's light or dark theme
+
+*tuicr works out its own theme by asking the terminal what colour it paints. Nothing answers under `herdr --remote`, so set `tuicr.appearance` on any machine you reach over the network.*
+
+- tuicr's own default asks the terminal for its background colour over OSC 11, an escape sequence a terminal answers with the colour it is painting
+- Under `herdr --remote` nothing answers, and tuicr falls back to `gsettings get org.gnome.desktop.interface color-scheme`
+- A headless machine answers that with `default`, which tuicr reads as light
+    - It then paints a light theme with a transparent background, leaving dark text on a black terminal
+- Set `tuicr.appearance` to `dark` or `light`, and tuicr skips the whole detection path
+- Leave it empty on a machine you sit at, where the terminal answers for itself
+
+### 3.5. LazyVim: a dot for unsaved changes, and copy over SSH
+
+*Set `neovim.enabled: true` to get both. LazyVim hides unsaved changes, and over SSH it cannot copy to your clipboard.*
+
+- LazyVim's `pretty_path` component sets `modified_sign = ""`, so the only signal is colour
+    - It sets `modified_hl = "MatchParen"`, which recolours the filename you are already looking at
+    - You cannot tell a saved buffer from an unsaved one at a glance
+- `apply` writes `~/.config/nvim/lua/plugins/acstack.lua`, which appends a lualine component to `lualine_c`
+    - LazyVim loads every file under `lua/plugins`, so a file of our own is a complete plugin spec
+    - The dot is `●` in catppuccin red, and `neovim.unsaved_dot_color` changes the colour
+- This is the only component that starts off, because a neovim config is one you built
+    - Every other component owns its files outright, or edits a marked region in a config that invites one
+- It writes one whole file and never edits one of yours
+    - `setup.sh disable neovim` deletes that file, and so does `uninstall`
+    - Neither one touches a `acstack.lua` you wrote by hand
+- `apply` skips a machine with no LazyVim config, because both parts of the file rely on LazyVim
+    - `apply` says so rather than writing a file that does nothing
+- `apply` leaves the dot out when a lualine override you wrote already marks a modified buffer, and names that file
+    - LazyVim loads every file under `lua/plugins`, so yours does exactly what this component does
+    - The statusline would then carry two dots side by side
+    - Neither file looks wrong on its own, and the only place the collision shows is the statusline
+    - Delete your own file to hand the dot to this component
+
+Copy over SSH:
+
+- Over SSH, LazyVim sets `clipboard` empty and neovim finds no provider, so `:%y+` fails with `clipboard: No provider`
+- The same file sets `vim.g.clipboard` to send every copy over OSC 52, which reaches the clipboard of the machine you sit at (section 10.1)
+    - It also sets `clipboard=unnamedplus` after LazyVim's own `VeryLazy` restore, so a plain `y` copies too
+    - `apply` leaves that option alone when a file of yours sets `clipboard`
+- `p` pastes what neovim copied last, and never asks the terminal
+    - Windows Terminal refuses OSC 52 reads, and neovim's own OSC 52 paste then waits ten seconds
+    - Paste text copied elsewhere with your terminal's paste key, `ctrl+shift+v` in Windows Terminal
+- The file does nothing on a machine with a display, or inside tmux, which each have a clipboard of their own
+- `apply` leaves the clipboard out when a file of yours already sets `vim.g.clipboard`, and names that file
+- `neovim.osc52_clipboard: false` turns it off
+- `setup.sh status` prints a `neovim clipboard:` line naming which provider neovim uses
+
+### 3.6. pi: ACStack in a second coding agent
+
+*Set `pi.enabled: true` to install pi, the coding agent, with ACStack and your pi settings. It starts off, because it writes into pi's own config.*
+
+What `apply` does:
+
+- It installs pi with `npm install -g @earendil-works/pi-coding-agent` when pi is missing
+- It installs ACStack into pi, pinned to the version of this plugin
+    - Re-running `apply` after a plugin update moves the pin forward, and never back
+- It installs `pi-subagents`, which ACStack on pi needs to launch a subagent
+- It installs `billion-context` by default for context compression in pi
+    - When the `bili` command is available, setup runs native plugin setup for pi
+    - Setup leaves Claude Code and Codex alone, and removes bili from them if an older setup added it
+    - Set `pi.billion_context: false` to leave it out, or to remove it later if this setup installed it
+- It installs `@juicesharp/rpiv-todo` and `pi-plugin-manager` by default
+    - Set `pi.todo: false` or `pi.plugin_manager: false` to leave either package out
+- It installs `@juicesharp/rpiv-ask-user-question` as pi's question tool by default
+    - ACStack uses it because decisions can include markdown, code and artifact previews
+    - It accepts freeform answers and notes, and needs no full-event telemetry environment variable
+    - This setup still keeps pi's install telemetry disabled through `pi.settings.install_telemetry: false`
+    - Set `pi.question_tool: false` to leave it out, or to remove it later if this setup installed it
+- It installs `pi-web-access` by default for Pi web search and web access
+    - The package is `pi-web-access` from `nicobailon/pi-web-access`
+    - It adds web search, URL fetching, GitHub repo cloning, PDF extraction and media understanding tools to pi
+    - This is separate from pi's anonymous install and update reporting; `pi.settings.install_telemetry: false` still controls that telemetry setting
+    - Set `pi.web_access: false` to leave it out, or to remove it later if this setup installed it
+- It can install `pi-lens` for diagnostics and code-intelligence tools
+    - It is off by default because it is a heavier dev-tool layer
+    - Set `pi.lens: true` to install it, or `pi.lens: false` to remove it later if this setup installed it
+- It installs `pi-slash-anywhere` by default, so slash commands work from any prompt
+    - Set `pi.slash_anywhere: false` to leave it out, or to remove it later if this setup installed it
+- It installs `pi-compact-tools` and runs `agentic-retro install-pi`, each behind its own toggle
+    - The agentic-retro step needs the agentic-retro plugin installed first, and `apply` says so when it is not
+- It writes the delegation line into pi's context file in `~/.pi/agent`, between HTML comment markers
+    - That is `AGENTS.md` by default, or the first of `AGENTS.override.md`, `AGENTS.md`, `AGENTS.MD`, `CLAUDE.md` and `CLAUDE.MD` that exists, which is the one pi reads
+    - `pi-subagents` delegates only when your request or your instructions allow it
+- It sets `"toolActivation": "eager"` in `~/.pi/agent/extensions/subagent/config.json`
+- It binds `"app.message.dequeue"` to `"ctrl+enter"` in `~/.pi/agent/keybindings.json`
+- It writes the four keys under `pi.settings` into `~/.pi/agent/settings.json`
+- It patches pi's bundled TUI so Windows Terminal Win32 input records paste as text
+    - `pi update` can replace the bundle, so `setup.sh upgrade pi` reapplies the patch
+
+How it treats pi's files:
+
+- pi's config is JSON, which has no comments to mark a managed region with
+- `apply` claims only the keys it writes, and every other key stays
+- `apply` installs and removes packages through `pi install` and `pi remove`, and never edits `packages` itself
+- `apply` touches nothing in a file that does not parse, and names it
+    - A broken `settings.json` stops the whole component, because `pi install` rewrites that file
+- `~/.config/acstack/terminal-workspace.pi.json` records what `apply` added
+    - `disable pi` and `uninstall` undo only what that file names
+    - A package or a key you had before `apply` stays yours
+    - A value of yours that `apply` replaced goes back, and so does an ACStack pin it moved forward
+    - A value you changed after `apply` stays as you set it
+- `apply` never touches `auth.json`
+    - Run `pi`, then `/login`, yourself
+- `setup.sh status` reports pi's version, its packages, the delegation line, `toolActivation`, and whether `auth.json` exists
+- `setup.sh outdated` and `setup.sh upgrade` include `billion-context`, `@juicesharp/rpiv-todo`, `pi-plugin-manager`, `@juicesharp/rpiv-ask-user-question`, `pi-web-access`, `pi-lens`, and `pi-slash-anywhere` like pi's other npm packages
+
+## 4. How you attach
+
+*herdr is a background server plus a terminal client. Where you put the client decides what the setup can do, and which key combinations reach herdr at all.*
+
+- The server owns the panes and keeps them running whatever happens to your connection
+- The client draws them
+- Those are two separate machines in this setup
+- Which one runs the client is the single decision the rest of this document depends on
+
+Two terms:
+
+- **A chord** is one key combination, such as `alt+f`
+- **A keymap** is the file that says which chord runs which command
+
+The chord family lives in the config as `herdr.chord_modifier`, so changing your mind costs one edit and one `setup.sh apply`, which rewrites the keymap and reloads it.
+
+### 4.1. From your own machine, with `herdr --remote`
+
+*The recommended path. The client runs beside your clipboard and browser, so both can reach the session.*
+
+Install herdr on the machine you type on as well as on the dev machine, then attach:
+
+```bash
+herdr --remote dev-box --remote-keybindings server
+```
+
+- The client connects over SSH, starts or attaches the server on the dev machine, and streams the session back
+- Panes and agents still run on the dev machine, and only the drawing happens locally
+- Section 4.4 lists four things worth setting up once
+- You can attach without any of them
+
+### 4.2. By SSH first
+
+*The fallback. Simpler, works from a phone, and gives up the local clipboard route.*
+
+```bash
+ssh you@dev-box
+herdr
+```
+
+- Everything runs on the dev machine now, the client included
+- This is the tmux-shaped path
+- Detach with `prefix+q`, where `prefix` is `ctrl+b`, the key you press before a herdr command
+- Then disconnect, SSH back, and run `herdr` again
+- Use this path from a phone SSH app, where there is no local herdr to run
+- Keep it as the fallback for when a remote attach misbehaves
+- Both paths attach to the same server
+    - You can move between them freely
+
+### 4.3. What differs, and why it decides your keymap
+
+*Each attach path can carry a different set of chords. Bind the family that survives both.*
+
+| | `herdr --remote` | `ssh`, then `herdr` |
+| --- | --- | --- |
+| Where the client runs | your machine | the dev machine |
+| Clipboard image paste into an agent | yes | no |
+| Which config supplies chords | the client's, unless `--remote-keybindings server` | the server's |
+| `alt+<key>` | works | works |
+| `ctrl+alt+<key>` | dead | works |
+| `ctrl+<digit>` | works | dead |
+| Works from a phone SSH app | no | yes |
+| Windows as the dev machine | not supported | not applicable |
+
+- We measured the three key rows on herdr 0.8.2 with a Windows client rather than inferring them
+- They are why this document binds plain `alt` throughout
+
+The two paths deliver keys by different routes.
+
+- Over SSH the emulator encodes a chord as an escape sequence, and herdr decodes it on the far side
+- Under `herdr --remote` the local client reads native key events instead
+- The two routes do not carry the same set
+- The `ctrl+alt` family survives the escape-sequence route and dies on the native one
+- `ctrl+<digit>` does the opposite
+
+- Either way herdr accepts the binding and then never sees the key
+- A chord that does not cross is therefore silently dead, and herdr reports no error
+- `alt+<key>` is the only family that crossed both
+- Test the wider family and use it, if you only ever attach one way
+- Bind `alt` and do not spend the time, if you use both or expect to
+- This affects only herdr's own chords
+    - gh-dash and diffnav read their own configs on the dev machine, where they run
+    - Their keys behave the same however you attached
+
+### 4.4. Worth setting up once, after you have attached
+
+*An SSH host block, a loaded key, the server keymap flag, and a shortcut you can tab-complete.*
+
+**An SSH host block**, so the target is a name rather than an address.
+
+- Anything you put in it carries through, port forwards included
+- herdr builds its own SSH config by including yours first
+
+```text
+Host dev-box
+  HostName 203.0.113.10
+  User you
+  LocalForward 6419 localhost:6419
+```
+
+**Key authentication through an agent.**
+
+- herdr's connection reuse is Unix-only
+    - A Windows client authenticates more than once while it sets the session up
+- With no key loaded that is a password prompt each time
+- Load it once per boot with `ssh-add`
+
+**`--remote-keybindings server`.**
+
+- By default herdr reads its chords from the config on the machine running the client
+- It then ignores the keymap on the dev machine
+- Every chord looks broken
+- This flag keeps the server config as the single source of truth
+- Section 4.3 covers why that matters more than it sounds
+
+**A shortcut, because you will type this many times a day.**
+
+- On Windows use a PowerShell function rather than an alias or a symlink
+- Both of those map one name to another
+    - The target arrives as an argument to herdr itself, and herdr rejects it
+
+```powershell
+function herdr-dev {
+  & "$env:LOCALAPPDATA\Programs\Herdr\bin\herdr.exe" `
+    --remote dev-box --remote-keybindings server @args
+}
+```
+
+- A hyphenated name tab-completes, where the second word of a two-word command never will
+- On macOS or Linux the same thing is a shell function in your profile
+- Add `--handoff` if you like
+    - It asks a running server to pass its live panes to a replacement rather than restarting them
+    - That matters the next time you upgrade herdr and your client meets an older server
+    - It does nothing on an ordinary attach
+        - It is safe to leave in the shortcut
+
+## 5. The herdr server
+
+*One config file, one keymap, and a numbering quirk that bites once you use worktrees.*
+
+### 5.1. Every key, and the three lists that answer to 1..9
+
+*Plain `alt` chords for everything you press often, and the prefix for anything that closes a pane.*
+
+Each tool lists its own keys.
+
+- Press `prefix+?` in herdr, then `/` to filter
+- Press `?` in gh-dash
+- Read diffnav's footer
+- Press `?` in lazygit and in tuicr
+- Every key worth memorising is on one page: [terminal-workspace-keys.md](terminal-workspace-keys.md)
+- Prefix is `ctrl+b`, and a chord without it is direct and needs no prefix
+- A popup floats over your layout and puts you back where you were when you close it
+
+| Key | Action |
+|---|---|
+| `alt+1`..`alt+9` | Focus agent by row |
+| `prefix+1`..`prefix+9` | Switch workspace by row |
+| `prefix+alt+1`..`prefix+alt+9` | Switch tab by position |
+| `alt+n` / `alt+u` | Next / previous agent |
+| `alt+left` / `alt+right` | Previous / next tab |
+| `alt+up` / `alt+down` | Previous / next workspace |
+| `alt+h` `alt+j` `alt+k` `alt+l` | Move between panes |
+| `alt+a` | Back to the last pane |
+| `alt+s` | Toggle sidebar |
+| `alt+g` / `alt+shift+g` | lazygit in a popup / in a new tab |
+| `prefix+t` | Goto picker |
+| `alt+z` | Zoom pane |
+| `alt+c` | New tab |
+| `alt+w` | New workspace |
+| `alt+r` | New worktree workspace |
+| `alt+e` / `alt+minus` | Split right / down |
+| `alt+y` / `alt+shift+y` | File tree in a popup / in a new tab |
+| `alt+f` / `alt+shift+f` | Diff of your branch against dev, in a popup / in a new tab |
+| `alt+x` / `alt+shift+x` | Pick what to compare, in a popup / in a new tab |
+| `alt+i` / `alt+shift+i` | Dashboard in a popup / in a new tab |
+| `prefix+q` | Detach, leaving everything running |
+| `prefix+x` / `prefix+shift+x` / `prefix+shift+d` | Close pane / tab / workspace |
+| `prefix+[` | Copy mode |
+| `prefix+w` | Navigate mode, a persistent movement surface |
+
+Three separate lists answer to `1..9`, and they are not the same list.
+
+- `alt+N` walks agents
+- `prefix+N` walks workspaces
+- `prefix+alt+N` walks the tabs of the workspace you are in
+- An agent row and a workspace row that share a number are a coincidence
+
+The `ctrl` and `shift` modifiers are not options for a fourth list, and which one fails depends on how you attached.
+
+- A `shift+<digit>` chord arrives as punctuation on both paths
+- A `ctrl+<digit>` chord reaches herdr under `herdr --remote` but not over SSH
+- A binding you make on one path is therefore silently dead on the other (see the table in section 4.3)
+- The three chords above, `alt+N`, `prefix+N`, and `prefix+alt+N`, are what both attach paths carry
+
+### 5.2. One config file, validated without a restart
+
+*Everything above lives in `~/.config/herdr/config.toml`. Three choices in that file are worth explaining.*
+
+- Validate with `herdr config check`
+- Apply without restarting with `herdr server reload-config`
+- Undo with `herdr config reset-keys`
+
+```toml
+[keys]
+# Jump straight to an agent by its sidebar row.
+focus_agent = "alt+1..9"
+
+# Panes, on the same vim letters as the prefix bindings.
+focus_pane_left  = ["prefix+h", "alt+h"]
+focus_pane_down  = ["prefix+j", "alt+j"]
+focus_pane_up    = ["prefix+k", "alt+k"]
+focus_pane_right = ["prefix+l", "alt+l"]
+
+# Agents, tabs, workspaces.
+next_agent = "alt+n"
+previous_agent = "alt+u"
+next_tab = ["prefix+n", "alt+right"]
+previous_tab = ["prefix+p", "alt+left"]
+next_workspace = ["alt+down"]
+previous_workspace = ["alt+up"]
+
+# By number. herdr leaves switch_workspace unbound and puts switch_tab on
+# prefix+1..9, so without these there is no way to reach a workspace by number.
+switch_workspace = "prefix+1..9"
+switch_tab = "prefix+alt+1..9"
+
+# Overlays and movement.
+toggle_sidebar = ["prefix+b", "alt+s"]
+# g belongs to git: alt+g is lazygit and alt+shift+g is its tab.
+goto = "prefix+t"
+zoom = ["prefix+z", "alt+z"]
+last_pane = "alt+a"
+
+# Creating things. Closing stays on the prefix so a mistyped chord cannot
+# kill a pane with an agent running in it.
+new_tab = ["prefix+c", "alt+c"]
+new_workspace = ["prefix+shift+n", "alt+w"]
+new_worktree = "alt+r"
+split_vertical = ["prefix+v", "alt+e"]
+split_horizontal = ["prefix+minus", "alt+minus"]
+
+# A popup floats over the layout and returns you where you were. Each one has
+# a tab variant on the shift chord, which goes through acstack-herdr-tab.
+# herdr runs a popup with its server's PATH. That PATH may lack ~/.local/bin.
+# Each popup goes through acstack-popup, which puts that folder back.
+[[keys.command]]
+key = "alt+f"
+type = "popup"
+command = "/home/you/.local/bin/acstack-popup acstack-diff"
+description = "diff: branch vs dev"
+width = "90%"
+height = "90%"
+
+[[keys.command]]
+key = "alt+shift+f"
+type = "shell"
+command = "acstack-herdr-tab diff acstack-diff"
+description = "diff: branch vs dev (tab)"
+
+[[keys.command]]
+key = "alt+g"
+type = "popup"
+command = "/home/you/.local/bin/acstack-popup lazygit"
+description = "git: stage, commit, push"
+width = "90%"
+height = "90%"
+
+[[keys.command]]
+key = "alt+shift+g"
+type = "shell"
+command = "acstack-herdr-tab lazygit lazygit"
+description = "git: stage, commit, push (tab)"
+
+[[keys.command]]
+key = "alt+x"
+type = "popup"
+command = "/home/you/.local/bin/acstack-popup acstack-diff pick"
+description = "diff: pick what to compare"
+width = "90%"
+height = "90%"
+
+[[keys.command]]
+key = "alt+shift+x"
+type = "shell"
+command = "acstack-herdr-tab diffpick acstack-diff pick"
+description = "diff: pick what to compare (tab)"
+
+[[keys.command]]
+key = "alt+i"
+type = "popup"
+command = "/home/you/.local/bin/acstack-popup acstack-dash"
+description = "PR dashboard"
+width = "95%"
+height = "95%"
+
+[[keys.command]]
+key = "alt+shift+i"
+type = "shell"
+command = "acstack-herdr-tab dash acstack-dash"
+description = "PR dashboard (tab)"
+
+[[keys.command]]
+key = "alt+y"
+type = "popup"
+command = "/home/you/.local/bin/acstack-popup yazi"
+description = "file tree"
+width = "95%"
+height = "95%"
+
+[[keys.command]]
+key = "alt+shift+y"
+type = "shell"
+command = "acstack-herdr-tab yazi yazi"
+description = "file tree (tab)"
+
+[worktrees]
+directory = "~/.herdr/worktrees"
+
+[ui]
+# A pane's scrollbar costs the pane a column that herdr does not subtract.
+pane_scrollbars = false
+show_agent_labels_on_pane_borders = true
+```
+
+- A `shell` command carries no `width` or `height`
+    - herdr rejects both on anything that is not a popup
+
+Four settings that are not keys sit alongside the block above. Each one is
+written only when the config names it, so a machine that says nothing about a
+setting keeps whatever herdr already does.
+
+| Config key | Writes | Default |
+| --- | --- | --- |
+| `herdr.agent_labels_on_pane_borders` | `show_agent_labels_on_pane_borders` in `[ui]` | `true`, and written either way |
+| `herdr.agent_panel_sort` | `agent_panel_sort` in `[ui]` | empty, so nothing is written |
+| `herdr.toast_delivery` | `delivery` in `[ui.toast]` | empty, so nothing is written |
+| `herdr.theme.name` and `herdr.theme.light` | `[theme]` | empty, so nothing is written |
+
+**One server, two devices, two backgrounds.**
+
+- `auto_switch` asks each attached client for its background colour over OSC 11, an escape sequence a terminal answers with the colour it is painting
+- herdr then picks `dark_name` or `light_name` from the answer, per client
+- A dark terminal on one device and a light one on another each get the right theme
+    - An e-ink panel is the case that makes this worth having
+- A client that answers nothing falls back to `dark_name`
+    - So `apply` writes `herdr.theme.name` into both `name` and `dark_name`
+- `herdr.theme.name` on its own writes the name and nothing else
+    - With no light theme, nothing has anything to switch to
+
+Three choices worth explaining.
+
+**Plain `alt` chords, not `ctrl+alt`.**
+
+- herdr's own keyboard page recommends `ctrl+alt`
+    - It is free across most terminals
+- Section 4.3 has the measurements showing it dies under `herdr --remote`
+- Follow herdr's advice only if you attach exactly one way and have tested it
+
+**Absolute `worktrees.directory`.**
+
+- A relative value resolves against the herdr session's base directory, not the repository you point at
+- Worktrees for a second repository then land inside the first
+- Use an absolute path unless you only ever work in one repository
+
+**`pane_scrollbars = false`, so a full-screen app gets the whole pane.**
+
+- herdr draws a border and a scrollbar inside the pane rect, then reports the full rect width to the program in the pane
+- A program that fills the pane then writes one column more than it has
+    - Every wrapped row starts one column left of the row above it
+- Measured on one machine: `herdr pane layout` reported a 132-column rect at `x=26`, the PTY and neovim's `&columns` both reported 131, and 130 were usable
+    - The arithmetic is herdr's, so it should hit anyone running `pane_scrollbars = true`, and that is unproven on a second machine
+- `alt+s` twice repairs the display, because toggling the sidebar recomputes the geometry from scratch
+- The cost is that no pane draws a scrollbar
+    - Scrolling still works, because the mouse wheel and copy mode never used it
+- Set `herdr.pane_scrollbars: true` in the workspace config to take the scrollbar back
+
+### 5.3. When the sidebar numbers stop matching
+
+*Creating a worktree moves the sidebar rows and not the stored order. One helper realigns them.*
+
+- Collapse the sidebar with `alt+s` and each workspace shows a number
+- That number is its position in herdr's stored list
+- The `prefix+N` chord uses the row's position in the grouped sidebar instead, where a worktree sits indented under its parent repository
+- They agree until you touch a worktree
+- A new one appends to the end of the stored list but appears mid-sidebar under its parent
+- Everything below it shifts
+    - The number you read stops being the number you press
+
+```bash
+acstack-herdr-renumber
+```
+
+- That rewrites the stored order to match the grouped order
+    - Both agree again
+- It prints the result and is safe to run repeatedly
+
+You rarely have to run it.
+
+- `apply` links a small herdr plugin, `acstack.herdr-numbers`, that runs the same helper on every event which can move a row
+- Run the helper by hand when you configured herdr yourself, or when `apply` reported that the plugin link failed
+- The `herdr.renumber_plugin: false` key turns the plugin off
+- The `acstack-herdr-` prefix is the convention for a helper that only works under herdr
+- A plain `acstack-` helper such as `acstack-diff` or `acstack-md` runs in any terminal
+    - The name tells you up front what a command depends on
+- Both forms are reachable through the CLI, which dispatches an unknown subcommand to `acstack-<name>` on your PATH
+
+Two things the helper cannot fix.
+
+- Only rows 1 to 9 are reachable
+    - A tenth workspace needs `prefix+t` or `alt+up` and `alt+down`
+- `prefix+N` stays positional
+    - It names a row rather than a workspace and still moves when the rows move
+
+## 6. Freeing your emulator's keys
+
+*You bind the keymap on the dev machine, but the emulator on the machine you type at intercepts the chords. Fix that where the emulator runs.*
+
+- Windows Terminal binds `alt+shift+d` to "duplicate pane" by default
+    - It never forwards the key
+        - A binding on it does nothing at all on the dev machine and the local tab splits instead
+    - That is why both diff keys sit on `f`: `alt+f` and `alt+shift+f`, and not `alt+d`
+- Other emulators claim other chords
+- This is true on both attach paths
+    - The emulator sees the key first either way
+
+To confirm an emulator is eating a chord rather than herdr ignoring it, run `cat -v` in any pane and press the key.
+
+- A chord that arrives prints an escape sequence such as `^[Z`
+- One that prints nothing never left your machine
+- [assets/terminal-workspace/client-keybindings.md](../assets/terminal-workspace/client-keybindings.md) lists every chord this setup uses and how to unbind them in the common emulators
+- Follow it yourself, or hand it to an agent running on that machine
+- Never try to edit a client-side emulator config from the dev machine
+- Never paste your local config onto the dev machine so that something there can rewrite it
+
+## 7. Reading code, markdown, and diagrams
+
+*A file tree with live preview. Markdown that draws its mermaid diagrams as text.*
+
+### 7.1. The file tree
+
+*yazi opens over your layout on `alt+y`, and every review popup works the same way.*
+
+- The file tree is yazi, a keyboard-driven file manager that previews whatever the cursor sits on
+- Press `alt+y` for yazi in a popup, which floats and leaves the tab layout alone
+- Press `alt+shift+y` for yazi in a new tab instead
+
+Every popup works this way: `alt+f` and `alt+shift+f` for diffnav, `alt+i` and `alt+shift+i` for gh-dash.
+
+- A popup is right for a glance
+- A tab is right for something you will come back to
+- Both come from the same command, and the tab variant goes through `acstack-herdr-tab`
+    - That helper creates the tab in the workspace and directory you pressed the key in
+    - Then it sends the command with `herdr pane run`
+    - Outside herdr it simply runs the command
+- This setup leaves `alt+t` alone on purpose
+    - It is Claude Code's thinking toggle
+- The one collision to know about is your terminal emulator claiming the same chords
+- You fix that where the emulator runs, per section 6
+- yazi previews each file type with its own command
+- It routes markdown to `acstack-md`
+    - A document therefore renders **as the cursor moves over it** rather than needing a keypress
+- The `Enter` key opens the same renderer full width, where more of a wide diagram fits than the preview pane allows
+- The `~` and `F1` keys both open yazi's help
+- Icons come from a Nerd Font, and without one every icon draws as an empty box
+- Install any Nerd Font and select it in your terminal
+- tuicr was the file tree before yazi
+- yazi then took that job, and tuicr kept the one it is better at, reading diffs and reviewing pull requests
+- It gets that work from gh-dash on the `D` key, which runs `tuicr pr <number>` in the local clone
+- The command `tuicr --file .` still browses a tree if you want it
+
+### 7.2. Reading markdown and mermaid
+
+*`acstack-md` draws a file with its diagrams, and two keys inside the pager change what you are looking at.*
+
+```bash
+acstack-md NOTES.md              # terminal, diagrams drawn as text
+acstack-md --numbered NOTES.md   # the source instead, with its line numbers
+acstack-md --diagram 2 NOTES.md  # one diagram alone, scrollable sideways
+acstack-md --serve NOTES.md      # browser, prints a clickable link
+acstack-md --browser NOTES.md    # the browser you are sitting at, wherever that is
+acstack-md --html NOTES.md       # that same page, as one document on stdout
+```
+
+`--numbered` answers the one question a rendering cannot, which is what line of the file this is.
+
+- A review comment names a line number, and a rendered heading has none
+- `--numbered` therefore prints the source with a right-aligned gutter
+- The gutter is as wide as the largest number in it
+    - The source stays on one column however long the file
+- It pages through `less -S`, which chops rather than wraps
+- The arrow keys pan across a long line
+- The gutter keeps its column
+
+Two keys work while you are reading.
+
+- The `b` key opens the page in the browser you are sitting at and returns you to where you came from
+    - It uses the routes in section 7.4
+        - It is the same key whether you attached over `herdr --remote`, over SSH, or locally
+- The `#` key switches the document between the rendered view and its source with line numbers, and switches it back
+    - That is the moment you usually want a line number
+        - You do not have to quit back to the tree
+
+Both keys move.
+
+- `ACSTACK_MD_LINE_NUMBERS_KEY` moves `#`, and `ACSTACK_MD_BROWSER_KEY` moves `b`
+- In `less`, `b` is back-a-page
+    - The binding costs you nothing, and both `Ctrl-B` and `PageUp` still do that
+- Both bindings need `less` 582 or newer, and an older version leaves the two keys alone
+- Section 11.2 explains how they reach a running pager
+
+A wide diagram cannot fit a narrow pane.
+
+- `acstack-md` replaces anything wider than the terminal with a note
+- The note gives its size and the two ways to see it
+- The `--diagram N` flag prints that one diagram unwrapped through `less -S`, where the arrow keys scroll sideways
+- `ACSTACK_MD_PAD` tunes the spacing passed to `mermaid-ascii`, defaulting to `-x 2 -y 2`
+    - Tighter padding buys roughly a third of the height back and very little width
+- Section 11.3 explains why the note exists at all
+- Terminal mode replaces each mermaid fence with a box-drawing rendering
+- `mermaid-ascii` handles `graph`, `flowchart`, and `sequenceDiagram`
+- Anything else keeps its source visible with a note rather than disappearing
+
+Two things `acstack-md` normalises first, because `mermaid-ascii` handles
+neither:
+
+- `style`, `classDef`, `class`, `linkStyle`, and `click` lines, which it would
+  otherwise draw as if each were a node
+
+- node shapes other than `[square]` – `{decision}`, `((circle))`, `([stadium])`,
+  `[(database)]`, `{{hexagon}}` – whose syntax would otherwise leak into the label
+
+Serve mode renders the file with python-markdown and draws diagrams with a **locally vendored** mermaid.js, so the page fetches nothing from a CDN.
+
+- It binds `127.0.0.1` only
+- It re-reads the file on every request
+- It prints an OSC 8 hyperlink, the escape sequence that marks a piece of text as a link the terminal draws itself
+    - herdr then rebuilds that link into the frame it sends the client
+    - The bare URL prints underneath for a terminal without OSC 8
+- So ctrl+click opens the page in the browser on your own machine, as long as your SSH config forwards the port from there
+- The `acstack-md --serve` command takes that port from `$ACSTACK_MD_PORT` and falls back to 6419
+
+Only one server can hold that port at a time, and a second `--serve` names the process holding it and the command to stop it:
+
+```
+port 6419 is busy: [Errno 98] Address already in use
+  held by pid 680666: acstack-md-serve - /path/to/NOTES.md 6419 serve
+  stop it with:  kill 680666
+```
+
+- Use that pid rather than `pkill -f acstack-md-serve`, which also matches any shell whose own command line mentions the name, including the one you type it in
+
+### 7.3. Reading markdown from the file tree
+
+*Move the cursor onto a markdown file and yazi previews it rendered. `Enter` reads it full width.*
+
+The file tree draws markdown twice over.
+
+- Moving the cursor onto a `.md` file renders it straight into the preview pane, through the piper plugin running `acstack-md --preview`
+- The pane is narrow
+    - A wide diagram shows a placeholder there rather than a chopped drawing
+- `Enter` on the same file opens `acstack-md` full width, where the diagrams fit
+- A rule in yazi's `[opener]` config table puts that ahead of the editor
+    - Reading is the default
+- Press `O` to choose from the same menu instead, which offers **Read (acstack-md)**, **Read with line numbers**, then **Edit**
+- Nothing shims `$EDITOR`
+- Nothing changes your shell environment
+- `b` on a file hands it to the browser you are sitting at, by the routes in section 7.4
+- The key is free in yazi's file list
+    - Its only default binding is a word motion while you are typing into a prompt, which this does not touch
+- Move it with `yazi.browser_key`
+- `#` switches the preview pane between the rendered markdown and the source with line numbers, and switches it back
+    - Press it when you are about to quote a line in a review
+    - The pane redraws as `acstack-md --numbered` would print it, numbers down the left
+    - Every later `.md` file you move onto previews the same way until you press `#` again
+- The key is unbound in yazi's file list, and `yazi.line_numbers_key` moves it
+- The same key does the same job one level in, while you are reading a document full width after `Enter`
+    - That is the more common moment to want it
+    - The two are deliberately the same key
+        - There is nothing to remember about which of the two places you are standing in
+    - Section 7.2 describes the reader's half
+- It is a binding in `lesskey`, the key-binding config file for the `less` pager, rather than a yazi one
+    - So `yazi.line_numbers_key` does not move it
+- The switch is a file on disk, `$XDG_STATE_HOME/acstack/md-line-numbers`
+- `acstack-md --toggle-line-numbers` creates and removes it
+- That is what the key runs
+- Only the preview pane reads it (see section 11.5)
+- `setup.sh apply` reads a `yazi.toml` you wrote first and leaves alone anything you have set
+- Add `acstack-md` to your own `[opener]` table to read markdown with it
+- Add `show_hidden = true` to your own `mgr` table if you want hidden files shown
+- Section 11.4 says what it concedes and why
+- `D` sends the file under the cursor to the machine you sit at, over Tailscale Taildrop
+    - Taildrop is Tailscale's file send
+    - It beats `scp` back to a laptop, because it needs nothing installed at the other end and no inbound port there
+    - The file lands in that machine's Tailscale receive folder
+- The key appears only when `yazi.download_target` names a Tailscale node, and `yazi.download_key` moves it
+- Taildrop sends only between devices one Tailscale account owns on one tailnet
+    - On this machine `tailscale status` may report the target as `offline`
+        - That means you signed the other machine in to another tailnet, whatever its own Tailscale client shows
+    - A `502 Bad Gateway` from the send means the same thing
+    - Read `tailscale status` on both ends and compare the account column, because two nodes can share a hostname across two tailnets
+- One step has no config key, because it is machine-level
+    - `tailscale file cp` answers a non-root caller with `Access denied: file access denied` until the account owns the daemon
+    - Run `sudo tailscale set --operator=$USER` once
+
+Four keys and a cursor move cover the loop.
+
+1. Press `alt+y` for the tree.
+2. Move the cursor onto a markdown file to preview it.
+3. Press `#` to see its source with line numbers.
+4. Press `Enter` to read it full width with its diagrams drawn.
+5. Press `q` to go back to the tree.
+
+### 7.4. Getting the page to the browser you are sitting at
+
+*`--browser` works out where your browser is and picks a delivery that reaches it.*
+
+- `--browser` is the one to reach for
+    - The same key works however you attached
+    - It asks `acstack-open --why` where the browser is rather than deciding that a second time
+
+| Where the browser is | What `--browser` does |
+| --- | --- |
+| Behind the opener on your laptop | Posts the whole document to it. The opener stores it, serves it, and opens your default browser |
+| A desktop on this machine, or the Windows Subsystem for Linux (WSL) | Serves it and opens it |
+| The far end of the Playwriter bridge, with no opener | Hands the whole document down the Chrome DevTools Protocol (CDP) link, into the tab the extension is armed on |
+| Anywhere else, over SSH | Serves it and prints a clickable link |
+
+The [Playwriter bridge](../skills/bridge/SKILL.md) is the setup that lets a program on the dev machine drive Chrome on your laptop.
+
+- It works one tab at a time
+- Arming means clicking the extension's icon on the tab you are willing to hand over
+
+The opener is the route you want.
+
+- It is a small service on your laptop that takes a page and opens it in your default browser
+- Section 10.5 covers how it gets installed
+- There is nothing to arm and no extension
+- The browser it reaches is your default one rather than a dedicated Chrome profile
+- Read one document after another and each simply appears
+- Re-render a file you are already looking at, and the tab you have open updates in place, scroll position kept
+- No second tab joins the first
+
+The page itself decides that last part, rather than the opener remembering it.
+
+- Every page the opener serves polls it for its own version
+    - A tab that is still open says so by asking
+- Re-render that file and the opener sees a live tab and lets it reload itself
+- Close the tab and the asking stops
+    - The next render opens a fresh one
+- Remembering that it once opened something would get the closed-tab case wrong every time
+
+The bridge is the fallback, and it is the one that needs explaining.
+
+- Under `herdr --remote` the reverse SSH tunnel to your laptop runs the *other way*
+- Nothing on the laptop can open a port on the dev machine
+- A link to `localhost:6419` names the laptop's own localhost, where nothing is listening
+- So there is no link to hand over, only a document
+- That is what `--html` is for
+- `--html` prints the page `--serve` would have served, once, to stdout, and starts nothing
+- A document you can capture in a variable travels
+- A port does not
+- The `--browser` flag is `--html` plus the delivery
+- The two share one renderer
+    - The page cannot differ between them
+- They differ in exactly one place, which section 11.6 covers
+- On the bridge the page replaces what is in the armed tab
+- The helper opens no new tab
+- Arming the extension is how you nominate the tab this may take over
+    - That is the tab it takes over
+- `--browser` says so and stops rather than falling back to serving, when the bridge is the route and the push fails
+    - A link the laptop resolves to its own localhost is a wrong answer dressed as a working one
+
+We measured one delivery rather than estimating it.
+
+- This document, 39KB of markdown, renders to 50KB of HTML in under 200ms
+- It reaches a laptop browser, diagram drawn, in about two seconds
+
+### 7.5. Why text and not inline images
+
+*No terminal you can reach from Windows or Android draws the graphics herdr emits.*
+
+- herdr embeds libghostty and emits the **kitty graphics protocol**, one of the two ways a terminal draws an inline image
+- It emits no sixel, the other way
+- Windows Terminal renders sixel and has never supported kitty, and [microsoft/terminal#8389](https://github.com/microsoft/terminal/issues/8389) is still open
+- No Google Play terminal supports either protocol
+- The intersection is empty
+    - An inline image never arrives no matter which terminal you pick
+- Text also suits e-ink, where the fast refresh modes an interactive terminal needs are the ones that discard the greyscale depth an image needs
+- `chafa` is worth adding by hand, with `apt install chafa`, if you want images drawn as text
+- It ships source-only so the setup script does not install it
+
+## 8. Harlequin: a SQL editor on one key
+
+*`alt+q` opens harlequin on a connection profile. `alt+shift+q` opens it in a tab.*
+
+- Query on the left, rows on the right, schema tree beside both
+- Profiles live in `~/.config/acstack/harlequin.toml`, which `acstack-db` writes once and never touches again
+- `acstack-db` with no argument launches the only profile, or opens an fzf menu when there is more than one
+- `acstack-db <profile>` skips the menu
+
+### 8.1. Why uv, and not a release binary
+
+- harlequin is a python program, so there is no static binary to download
+- `uv tool install harlequin --with harlequin-postgres --with harlequin-snowflake` builds it an environment of its own and links one command into `~/.local/bin`
+- A machine without `uv` skips the component, and `apply` names the one command that installs uv
+- Adding an adapter to `harlequin.adapters` and re-running `apply` reinstalls with the new set
+    - `uv tool install` on an already-installed tool does nothing, so `apply` compares the config against uv's own receipt and passes `--force` when they differ
+
+### 8.2. Postgres profiles
+
+- Keep the password out of the profile file: write the line `host:5432:db:user:password` in `~/.pgpass`, `chmod 600` it, and leave the password out of `conn_str`
+    - libpq reads that file on every connection, so harlequin never handles the secret
+- Put `read_only = true` on any production profile, which harlequin hands to the postgres adapter so every connection on that profile refuses a write
+
+```toml
+[profiles.prod]
+adapter = "postgres"
+conn_str = "postgresql://postgres@db.example.supabase.co:5432/postgres"
+read_only = true
+```
+
+### 8.3. Connecting to Snowflake
+
+- Put the credentials in `~/.snowflake/connections.toml` and name the connection from the profile, rather than writing a connection string into the acstack file
+- `externalbrowser` authentication does not work on a dev machine, because there is no browser for it to open
+    - Use a key pair, or a password
+- `--read-only` exists for postgres and not for snowflake
+
+### 8.4. Keys inside harlequin
+
+*A function key jumps straight to a pane. `tab` cycles them.*
+
+| Key | Does |
+|---|---|
+| `f2` | the editor |
+| `f5` | the results |
+| `f6` | the catalog |
+| `tab` / `shift+tab` | next / previous pane |
+| `ctrl+enter` or `ctrl+j` | run the query |
+| `f4` | format the query |
+| `f8` | query history |
+| `ctrl+e` | export the results |
+| `f9` or `ctrl+b` | hide the catalog sidebar |
+| `f10` | full-screen the focused pane, and again to restore it |
+| `ctrl+p` | the command palette |
+| `f1` | every key |
+| `ctrl+q` | quit |
+
+- `tab` cycles the panes from outside the editor, and inside it `tab` indents
+- In the catalog, the schema tree, `space` expands the node under the cursor and `ctrl+enter` puts its name in the editor at the cursor
+    - Plain `enter` selects the node and leaves the editor alone
+- The footer writes `ctrl+q` as `^q`, and `f1` opens the full list
+- Custom keys go in a `[[keymaps.<name>]]` block in the profile file, named from `keymap_name` on a profile, which harlequin's own documentation covers
+
+## 9. Diffs and pull requests
+
+*Diffs with a file tree, pull request triage with saved searches, and the two helpers that make one key always show something useful.*
+
+### 9.1. Five git settings that route every diff through delta
+
+*One pager for `git diff`, `git show`, and gh-dash alike.*
+
+```bash
+git config --global core.pager delta
+git config --global interactive.diffFilter "delta --color-only"
+git config --global delta.navigate true
+git config --global delta.line-numbers true
+git config --global merge.conflictstyle zdiff3
+```
+
+- This does not change what agents see
+    - The git command only pages to a terminal
+    - A caller that pipes or captures the output still gets plain text
+
+### 9.2. Saved searches become dashboard tabs
+
+*Any GitHub search string is a section, and two keys hand a pull request to tuicr or to an agent.*
+
+`~/.config/gh-dash/config.yml`.
+
+```yaml
+prSections:
+  - title: Mine
+    filters: is:open author:@me
+  - title: Needs review
+    filters: is:open review-requested:@me
+  - title: Failing
+    filters: is:open author:@me status:failure
+
+issuesSections:
+  - title: Mine
+    filters: is:open author:@me
+  - title: Assigned
+    filters: is:open assignee:@me
+
+# Local clones. Required for checkout and for {{.RepoPath}} in keybindings.
+repoPaths:
+  <you>/<repo>: ~/code/<repo>
+
+keybindings:
+  prs:
+    - key: D
+      name: review (tuicr)
+      command: >
+        cd {{.RepoPath}} && tuicr pr {{.PrNumber}}
+    - key: S
+      name: agent review
+      command: >
+        cd {{.RepoPath}} && claude "/code-review {{.PrNumber}}"
+```
+
+- Anything you can type into GitHub's search box becomes a tab
+- Avoid binding `R`, the built-in refresh-all
+
+### 9.3. The diff and dashboard helpers
+
+*One key shows what your branch adds to dev. Another picks any comparison. A third opens the dashboard, scoped to where you are standing.*
+
+- `alt+f` runs `acstack-diff`, which compares the branch you are on against `origin/dev`, committed work only
+- A repository with no dev branch falls back to the default branch, read from `origin/HEAD`
+- The comparison is `git diff <base>...HEAD`, three dots
+    - You see the commits your branch added since the two diverged, and never the ones the base picked up meanwhile
+- A pane often sits in `<root>/<repo>/`, the parent directory herdr groups a repository's worktree workspaces under, rather than in a checkout
+    - The helper then offers a numbered list of the checkouts it finds there
+
+Every launch opens with a banner naming both sides, because diffnav renders whatever precedes the first `diff --git` line:
+
+```
+COMPARING  origin/dev  ==>  deployment-map
+base     origin/dev  a84216b  16 hours ago  Merge pull request #44 from
+compare  deployment-map  549e94b  68 minutes ago  chore: integration edits
+showing  commits on deployment-map that origin/dev does not have
+command  git diff origin/dev...HEAD
+```
+
+- Base first, then compare
+    - That is the order `git diff` takes its arguments
+    - It is also the order GitHub labels its two compare pickers
+- The `command` line is real git syntax
+    - You can paste it into a shell
+- The comparison also reaches diffnav's status bar, which stays visible while you walk the file tree
+- That works because diffnav prints the line following a `commit <sha>` header as the commit subject
+    - The helper synthesises that header for a plain diff
+- A single-commit view already carries git's own header and gets no synthetic one
+- No banner line may start with a space, because diffnav strips leading whitespace
+    - Any indentation collapses and the columns stop lining up
+
+#### A file diffnav cannot draw
+
+*One committed source map used to abort the whole diff. The helper drops that file and names it.*
+
+diffnav hands each file to delta, and delta dies on `SIGABRT` when one line runs to hundreds of kilobytes. diffnav reports its child's death and quits:
+
+```
+Error: signal: aborted (core dumped)
+2026/08/24 02:04:25 FATA signal: aborted (core dumped)
+```
+
+- A committed build artifact is what does it
+    - ACStack's own `cli/dist/index.js.map` carries a single 982,428-character line
+    - `alt+f` on `main` aborted in 1.7 seconds and showed none of the other 51 files
+
+So the file goes, and not the diff. The helper drops any file whose patch holds a line over 20,000 characters, and names it in the banner:
+
+```
+skipped  cli/dist/index.js.map  (one line is 982428 chars, and diffnav aborts on it)
+```
+
+- 20,000 is far above any line a person writes and far below where delta gives out
+- `ACSTACK_DIFF_LINE_LIMIT` changes it
+- The helper refuses anything that is not a positive whole number
+- awk compares each line length against that value
+    - A value like `abc` makes every comparison a string comparison
+    - A string comparison is never true here
+    - A typo would otherwise switch the guard off and bring the abort back, with nothing on screen to explain it
+- Where the helper drops every changed file, it says so
+    - The old code said "No difference between these two"
+    - That claimed nothing had changed in a repository that had
+
+#### Picking a different comparison
+
+*`alt+x` opens an fzf menu of seven comparisons. Every row reads left to right as base, then compare.*
+
+```
+dev          ==>  my-branch                 committed work only
+dev          ==>  my-branch + uncommitted   committed work plus what is not committed
+main         ==>  my-branch                 committed work only
+main         ==>  my-branch + uncommitted   committed work plus what is not committed
+HEAD         ==>  my uncommitted changes    staged and unstaged changes only
+its parent   ==>  one commit                pick a commit from this branch's history
+any branch   ==>  any branch                pick the base, then the branch to compare
+```
+
+- The dev rows appear only where a dev branch exists, and the branch named is the one you are on
+- Branch lists come from `git for-each-ref` over local and remote branches, newest commit first, with `git log` in the preview pane
+- The commit list is the last 300 commits, with `git show --stat` in the preview
+- Picking "any branch" asks twice, prompting `source (compare against)` and then `change (the new work)`
+- `apply` installs fzf alongside diffnav, and without fzf the picker says so and falls back to the automatic diff
+
+`acstack-dash` adds a section for whichever repository you are standing in, then hands a generated config to gh-dash:
+
+```bash
+#!/usr/bin/env bash
+set -uo pipefail
+BASE="$HOME/.config/gh-dash/config.yml"
+REPO="$(gh repo view --json nameWithOwner -q .nameWithOwner 2>/dev/null)"
+[ -z "$REPO" ] && exec gh dash "$@"
+GEN="$(mktemp)"; trap 'rm -f "$GEN"' EXIT
+REPO="$REPO" python3 - "$BASE" "$GEN" <<'PY'
+import os, sys, yaml
+cfg = yaml.safe_load(open(sys.argv[1])) or {}
+repo = os.environ["REPO"]; short = repo.split("/")[-1]
+s = [x for x in cfg.get("prSections", []) if x.get("title") != short]
+s.insert(0, {"title": short, "filters": f"repo:{repo} is:open"})
+cfg["prSections"] = s
+yaml.safe_dump(cfg, open(sys.argv[2], "w"), sort_keys=False)
+PY
+gh dash --config "$GEN" "$@"
+```
+
+- Because a herdr popup inherits the focused pane's directory, `alt+i` from an agent's worktree opens the dashboard already scoped to that repository
+
+### 9.4. acstack-gh: why an action failed
+
+*gh-dash throws gh's stderr away, so a helper on `$PATH` turns a refusal into a notification.*
+
+- gh-dash shells out to `gh` for everything it does to a pull request, and throws the command's stderr away
+- GitHub refusing one therefore arrives as `exit status 1` in the footer, for two seconds
+- Approving your own pull request is the case you meet daily
+    - GitHub always refuses that
+- The input box closes, nothing else happens
+- The dashboard looks like it ignored the key
+
+`acstack-dash` answers that without patching gh-dash.
+
+- It links `acstack-gh` into a directory of its own at the front of `$PATH`, under the name `gh`
+- gh-dash therefore finds it before the real one
+- The real `gh` still does the work and still decides the exit code
+- The only thing added is a notification carrying gh's own words when a `pr` or `issue` action fails:
+
+```
+gh pr review failed
+Can not approve your own pull request.
+```
+
+- `acstack-gh` passes `gh dash` itself straight through
+- It also passes through every subcommand that is not an action, `repo view` and `api` among them
+- Those fail for reasons a notification cannot help with
+
+### 9.5. gh-dash keys
+
+*Vim movement, `[` and `]` for the preview tabs, and `D` or `S` to hand the pull request on.*
+
+| Key | Action |
+|---|---|
+| `j` / `k` | Move between rows |
+| `h` / `l` | Previous / next section |
+| `p` | Toggle the preview pane |
+| `[` / `]` | Previous / next preview tab: Overview, Activity, Commits, Checks, Files Changed |
+| `PageDown` / `PageUp` | Scroll the preview |
+| `ctrl+d` / `ctrl+u` | Scroll the preview, vim style |
+| `e` | Expand the description |
+| `d` | Built-in diff |
+| `D` | Open the pull request in tuicr and review it there |
+| `S` | Hand the pull request to an agent |
+| `C` or `space` | Check the branch out locally |
+| `w` | Watch checks |
+| `c` / `v` / `m` | Comment / approve / merge |
+| `y` / `Y` | Copy number / URL |
+| `r` / `R` | Refresh section / all |
+| `/` · `?` · `q` | Search · help · quit |
+
+### 9.6. diffnav keys
+
+*Vim movement in the tree, and `Tab` to cross between the tree and the diff.*
+
+| Key | Action |
+|---|---|
+| `j` / `k` | Move in the file tree |
+| `n` / `N` | Next / previous file |
+| `ctrl+d` / `ctrl+u` | Scroll the diff half a page |
+| `ctrl+e` / `ctrl+y` | Scroll the diff one line |
+| `Tab` | Switch focus between tree and diff |
+| `t` | Search and jump to a file |
+| `e` | Toggle the file tree |
+| `s` | Side-by-side or unified |
+| `o` | Open the file in Neovim, in a new herdr tab |
+| `q` | Quit |
+
+`o` opens a tab rather than taking over the pane, and the reason is the popup:
+
+- diffnav runs `$EDITOR` inside its own pane, and a herdr popup cannot host a second full-screen app
+- Neovim never drew in the `alt+f` and `alt+x` popups, so the key looked dead
+- `acstack-edit` stands in as `$EDITOR` and hands the file to `acstack-herdr-tab`, the same helper the tab keys use
+    - The tab opens in the workspace and directory you pressed the key in
+    - The tab carries the file's name, not its whole path
+    - The popup keeps the diff, so close it to reach the tab
+    - Outside herdr the helper runs your real editor in place
+
+## 10. What crosses back to your machine
+
+*A dev machine has no display and no clipboard of its own. The `acstack-clip` and `acstack-open` helpers carry each one back across the link.*
+
+Two gaps, and three gh-dash keys fall into them:
+
+- `o`, open on GitHub, fails with `exit status 1`. gh-dash opens URLs through `$BROWSER`, falling back to `xdg-open`, and `xdg-open` with no `$DISPLAY` exits 1
+- `y` and `Y`, copy the URL and the number, fail with `Failed copying to clipboard`. The gh-dash tool copies through a Go library that shells out to `xclip`, `xsel`, `wl-copy` or `termux-clipboard-set`. A bare dev machine has none of them installed
+- An install would not make them work
+- Neither is a gh-dash bug, because the clipboard and the browser are on the machine you are typing at, several hops away
+- The two helpers below carry each one back across
+
+### 10.1. acstack-clip: the clipboard
+
+*OSC 52 crosses SSH for free, so a copy on the dev machine lands on your own clipboard.*
+
+- OSC 52 is the escape sequence that asks a terminal to put text on its own clipboard
+- It is bytes in the terminal stream
+    - It crosses SSH for free
+- herdr forwards it from a pane to whatever terminal hosts it
+- Windows Terminal, iTerm2, kitty, and Ghostty all act on it
+
+```bash
+acstack-clip "some text"      # copy the arguments
+git rev-parse HEAD | acstack-clip
+acstack-clip --out            # print what was copied last
+```
+
+- Reading back is not symmetrical
+- Windows Terminal refuses OSC 52 clipboard *reads* on purpose
+    - A program that can read your clipboard without asking is a security hole
+- So `--out` replays a local cache, and not the real clipboard
+- To reach a program that only knows how to shell out, `setup.sh apply` also writes an `xclip` onto `$PATH` backed by `acstack-clip`
+- That is what makes gh-dash's `y` and `Y` work unchanged, with no rebinding and no flicker
+- The `remote.clipboard_shim` key turns it off, and `setup.sh uninstall` removes it
+- The script skips it anyway on a machine with a real `xclip`, or with a display for one to talk to
+
+### 10.2. acstack-open: the browser
+
+*Seven routes, tried in order, ending in a link you can click from any terminal.*
+
+- `o` is a gh-dash keybinding rather than a `$BROWSER` setting
+- That is deliberate
+- The dashboard is still on screen when gh-dash runs `$BROWSER`
+- It discards that command's output
+- A route that needs to say anything, or to hand you a link to click, therefore has nowhere to put it
+- As a keybinding gh-dash steps aside and gives `acstack-open` the terminal
+
+It tries, in order:
+
+1. `$ACSTACK_OPEN_CMD`, if you set one. The escape hatch
+2. `xdg-open`, when this machine has a display after all
+3. `wslview` or `explorer.exe`, when the Windows half of the machine holds the browser
+4. The opener on your laptop, which puts the page in your default browser with nothing to click. See section 10.5
+5. Chrome on your laptop through the [Playwriter bridge](../skills/bridge/SKILL.md), but only after it proves the browser is really reachable that way. See section 11.7
+6. A clickable link. The terminal you are sitting at draws the URL as an OSC 8 hyperlink, so ctrl+click reaches your own browser. The URL goes on your clipboard too
+7. With no terminal to draw on either, the URL still goes on the clipboard, and the command reports failure
+
+- The link route needs nothing installed in between
+    - That is why it is the last one that can still work
+- A terminal that ignores OSC 8 still shows the bare URL, and its own URL detection can catch that
+- Route 7 reports failure on purpose
+    - Silent success is what left gh-dash claiming it had opened a page that never opened
+
+The opener sits ahead of the bridge because the two are not competing for the same job.
+
+- The bridge exists so an *agent* can drive a browser
+    - It attaches one tab at a time, only after somebody clicks the extension icon
+    - It does so in a dedicated Chrome profile
+- The opener exists so it can show a *person* a page
+    - It needs no click at all
+    - It reaches the browser you actually use
+- Both can be up at once
+- Each keeps its own job
+- `setup.sh status` prints which route a machine will take, and the last line of `~/.cache/acstack/open.log` says what the last press actually did
+
+### 10.3. Under `herdr --remote`
+
+*We wrote both helpers for this shape, and they need no change. The clipboard crosses, and the browser falls to the link route, which is the one built for it.*
+
+- `herdr --remote <target>` runs the server, and therefore every pane process, on the dev machine
+- The client is a thin attach
+    - It sends input and draws what the server sends back
+
+`acstack-clip` works, and we measured that rather than inferring it.
+
+- A pane's OSC 52 write reaches the clipboard on the machine you attached from
+- So `acstack-clip "some text"` on the dev machine pastes on your own
+- Nothing in herdr's own documentation promises this
+    - It mentions OSC 7 and OSC 8, and never OSC 52
+- We tested it on herdr 0.8.2 and it crossed
+- Terminals differ in what they act on
+    - Copy something with `acstack-clip` after your first attach and check that it pasted
+- Clipboard *images* travel the other way, and only on this path
+- Copy a screenshot on your machine, focus an agent pane, and herdr stages the image on the dev machine and pastes its path
+- An SSH shell cannot do that at all
+    - That is the single strongest reason to prefer a remote attach
+- `acstack-open` runs on the dev machine
+    - That is the honest answer for routes 1 to 5
+- An override, a display, WSL, the opener or a bridge tunnel all have to be reachable *from there*
+- The opener and the bridge are reachable
+    - The laptop opens reverse tunnels that carry them back
+- So on a machine set up for either, that is the route it takes
+- With neither, it falls to the link route, the one built for this
+- In that route herdr tracks hyperlinks per cell and re-emits them when it renders
+    - The client therefore draws the link rather than shipping raw bytes
+    - ctrl+click opens the browser on the machine you attached from
+- The link route also degrades further than the others
+- Even with no OSC 8 and no OSC 52 at all, the URL is on screen as plain text, which drag-select copies
+- That is why the link is its own text rather than a label over it
+
+One trap belongs to your SSH config rather than to herdr.
+
+- The `acstack-md --serve` command prints the URL it is listening on
+- That is the port *on the dev machine*, meaning `$ACSTACK_MD_PORT` or 6419
+- The printed link is wrong from where you are sitting, if your host block forwards it to a different local port
+- You want the local number instead
+- Forwarding `6419` to `6419` avoids the question entirely
+
+### 10.4. On a machine with none of this
+
+*The link route needs only a terminal. One environment variable buys back a real one-key open.*
+
+The link route works over SSH, through herdr, and under `herdr --remote`. To get a real one-key open instead, give `acstack-open` something that can do it:
+
+```bash
+export ACSTACK_OPEN_CMD="ssh laptop open"   # or any command taking a URL
+```
+
+### 10.5. The opener: a page in your own browser, with nothing to click
+
+*A small service on your laptop. It takes a page from the dev machine, stores it, serves it back, and opens your default browser on it.*
+
+- The dev machine has no browser and no way to reach yours
+- The bridge solved that for agents, and not for reading
+    - It needs a tab armed by hand before every session
+    - The tab it drives lives in a dedicated Chrome profile rather than your default browser
+- The opener is the answer for reading
+- It is a separate service on purpose, per [ADR 0006](adr/0006-document-opener-service.md)
+
+What it does is deliberately small:
+
+| The dev machine sends | The opener does |
+| --- | --- |
+| A URL | Hands it to your default browser |
+| A rendered document | Stores it, serves it at `http://127.0.0.1:19989/doc/<id>`, opens that |
+| A vendored `mermaid.min.js`, once | Keeps it, and answers `/mermaid.js` off it from then on |
+| A request to restart the relay or the tunnel | Restarts that scheduled task |
+
+- The relay is the process on your laptop that the Playwriter extension connects to
+- Restarting it, and restarting the tunnel, are the two recovery actions the dev machine could never perform
+- They now go through the opener instead of arriving as a block for you to paste into PowerShell
+- Arming the extension is still yours
+    - It is a click inside a third-party extension and nothing on either machine can press it
+
+Installing the opener is the same command that registers the bridge, which now registers the opener too:
+
+```powershell
+cd $env:USERPROFILE\playwriter-bridge
+.\register-tasks.ps1 -VMs @("vm1.example.com")
+```
+
+- That generates a shared secret and stores it at `%LOCALAPPDATA%\playwriter-bridge\opener.token`
+- It copies the secret to each dev machine at `~/.config/acstack/opener.token`, over the same ssh the tunnel uses
+- Every request from the dev machine carries it
+- Loopback binding alone would not be enough
+    - The reverse tunnel makes the port reachable by anything running on the dev machine
+    - This is a service that puts pages on your screen
+
+The opener rides the same machinery as the bridge.
+
+- A scheduled task from `register-tasks.ps1` runs it, and a supervisor restarts it
+- The `sync.ps1` script reconciles the deployment on every Claude Code launch
+- Your laptop opens a reverse SSH tunnel
+- The opener gets its **own** tunnel task, `Playwriter-OpenerTunnel-VM<N>`, which carries port 19989
+- The bridge keeps port 19988 on a task of its own
+- Both tasks run the same `tunnel.ps1` script
+- They stay apart because ssh runs with `ExitOnForwardFailure=yes`
+    - One wedged port fails the whole connection
+    - Sharing one connection would let a stuck opener port take the bridge down with it
+- A document outlives the session that rendered it, which lets a page still work after the dev machine has gone away
+- The opener prunes a stored document after a week
+
+## 11. Design notes
+
+*Why the parts work the way they do. Read this when you are changing the setup or debugging it, and skip it otherwise.*
+
+### 11.1. Why the markdown helpers are separate executables
+
+*Node's startup would roughly double the cost of a preview that runs on every cursor move.*
+
+- The `acstack-md` and `acstack-clip` helpers are their own executables rather than subcommands of the `acstack` CLI
+- The `acstack-md --preview` command runs on every cursor move in the file manager, and Node's startup would roughly double it
+- The CLI dispatches to them anyway, the way git does
+- So `acstack md` runs `acstack-md`, either form works
+- A config can keep the fast one
+
+### 11.2. How `b` and `#` reach a running pager
+
+*A `lesskey` binding whose `quit` action carries an exit status, which `acstack-md` reads and acts on.*
+
+- The `less` pager has no action that runs a fixed command
+- Its one shell escape would hand over the rendered temporary copy rather than the file you asked for
+- So each key quits `less` with an exit status of its own, and `acstack-md` reads that status and does the work
+- It needs `less` 582 or newer, and an older version ignores the bindings and leaves both keys alone
+- `setup.sh apply` writes the `#` binding as `\#` in the `lesskey` file
+- `lesskey` reads a line starting with `#` as a comment
+    - It drops a bare binding without a word
+    - The key then does nothing but ring the terminal bell
+- Only a leading `#` needs the escape
+    - A backslash means something of its own to `lesskey`, and `\b` would bind backspace rather than the letter
+
+### 11.3. Why a wide diagram becomes a note
+
+*Nothing can shrink a diagram. Wrapping box-drawing art destroys it.*
+
+- A diagram's width comes from its node labels
+    - Nothing can shrink a wide one into a narrow pane
+- Wrapping box-drawing art destroys it instead of fitting it
+- A note that names the size is the honest answer, and a badly drawn diagram is not
+- A wide diagram still appears in place when it fits
+- glow wraps whatever it renders
+    - `acstack-md` holds the drawing back
+    - It runs glow on the prose, then splices the full-width art into the output
+- The pager is `less -S`
+    - glow already wrapped the prose to the pane
+        - Only the diagram lines chop, and the arrow keys pan across them
+
+### 11.4. What yazi's config merge concedes
+
+*Declaring anything twice would make yazi reject the whole file, so `apply` gives up whatever you already set.*
+
+- `setup.sh apply` reads your `yazi.toml` first and leaves alone anything you have set
+    - That covers your `mgr` settings, your markdown opener, your `plugin.prepend_previewers`, and your `open.prepend_rules`
+    - Whichever of the four it skipped, it says so
+- Declaring any of them a second time would make yazi reject the whole config and fall back to presets
+    - It concedes them instead
+- What it cannot read is a `yazi.toml` that does not parse
+    - yazi is then already ignoring the file in favour of presets
+        - Fix the error and re-run `apply`
+- The `b` and `#` bindings live in `keymap.toml` rather than `yazi.toml`, and `apply` writes them as `[[mgr.prepend_keymap]]`
+- That spelling is an array of tables
+    - It stacks with a binding you have already written the same way
+- The one spelling it cannot sit beside is the inline `prepend_keymap = [...]` under `[mgr]`
+    - That is a single key, and TOML forbids declaring it twice
+    - The script detects that form, gives the bindings up rather than cost you the whole keymap, and says so
+- The `[opener]` templates say `%s` rather than `"$@"`
+- yazi runs a template as `sh -c '<template>'` with nothing after it
+    - `$0` is `sh` and `$@` is empty
+- A template written with `"$@"` hands the helper no file at all
+- `%s` is the placeholder yazi substitutes, already quoted
+- Measured on yazi 26.8.15
+- The `keymap.toml` bindings say `%h` for the same reason, and never `%*`
+    - `%h` is the hovered file
+    - yazi quotes `%h` itself, so write it bare. Quotes around it end up inside the path.
+    - `%*` belongs to `[opener]`, and a keybinding never expands it
+        - The two characters go through untouched
+        - The command reports them back as a filename, such as `open %*: no such file or directory`
+    - So a binding that carries a file uses `%h`, and takes one file at a time
+- The `#` key runs two actions rather than one, flipping the flag, then forcing yazi to draw the pane again
+- Without the redraw the pane keeps showing whatever it drew before you pressed the key
+- The switch looks broken until you move the cursor
+- The flip blocks, though it has nothing to say
+    - A detached one races the redraw and loses about half the time
+
+One collision belongs to herdr rather than yazi.
+
+- herdr hosts a `type = "shell"` command in a real pane for as long as the process runs
+- So `acstack-herdr-tab` creates the tab and then hands the wait to a detached child
+- It returns in about 100ms rather than three and a half seconds
+- Without that a stray pane sits in the current tab
+
+### 11.5. Why the line-number switch lives in a file
+
+*The key and the preview pane are two processes that never meet, so the switch has to sit on disk between them.*
+
+- The state file is `$XDG_STATE_HOME/acstack/md-line-numbers`, or `~/.local/state/acstack/md-line-numbers` when that variable is unset
+- Its presence is the whole setting
+    - `acstack-md --toggle-line-numbers` only has to create or remove it
+- Only the preview pane reads that file
+    - Every other route asks for a view by name
+- The full-width read has **Read with line numbers** as its own entry under `O`
+- The `--diagram N` flag outranks the file too, since asking for one drawing is a different question from which view the pane is on
+
+### 11.6. Why `--html` names a CDN and `--serve` does not
+
+*A document standing on its own has no server behind it to answer for `/mermaid.js`.*
+
+- `--serve` answers for `/mermaid.js` off the vendored copy
+    - Its page fetches nothing from a CDN
+- A document standing on its own has no server behind it
+    - `--html` names the CDN instead
+- The vendored file is 3.5MB, and inlining it would make the page offline-proof and far too big to hand anywhere
+
+A document bound for the opener is the third case.
+
+- Here `--browser` hands it over once, like `--html`
+- The document does end up behind a server, the opener's
+    - It asks for `/mermaid.js` too
+- The 3.5MB goes up once, the first time the opener admits it has no copy
+- Every document after that draws its diagrams without reaching a CDN at all
+
+### 11.7. Why the bridge and the opener must prove themselves
+
+*`agent-browser` launches a headless Chrome when it cannot attach. It then reports success to nobody.*
+
+- `agent-browser` launches a headless Chrome on the dev machine when it cannot attach to the endpoint you gave it
+- That Chrome navigates perfectly happily, reports success, and shows nobody anything
+- The page opens on the dev machine, several hops from the screen you are looking at
+- Nothing about the relay answering on port 19988 rules that out either
+    - Ours answered `/json/version` while refusing every CDP connection with `Multiple extensions connected. Specify extensionId.`
+    - Every open landed in a headless Chrome for hours without one error message
+- So the bridge route asks the relay's `/json/list` what is on the far end, and takes the route only when something answers
+- The Playwriter extension attaches per tab, and `/json/list` is its own answer to that question
+- An empty `[]` means nobody has armed it on a tab
+    - There is no browser to drive however healthy the tunnel underneath looks
+
+The opener proves itself the same way and for the same reason.
+
+- The `acstack-open` helper asks it for `/health`, carrying the shared token, and takes the route only if the opener answers
+- A token sitting on disk proves nothing about a service being up
+- A service being up proves nothing without the token it is going to demand
+- Success on the bridge is likewise not an exit status
+- The pushed script ends with the page title
+    - The browser answers with what it is now holding
+- The `--browser` flag only reports success when that answer is the file you asked for
+- A command that exited 0 is not a page that arrived
+- The helper opens no new tab either
+- That is the second thing we measured rather than assumed
+- CDP creates a tab in the **background**
+- Nothing on the dev machine can bring it to the front
+- The document would land in it, the helper would report success
+- You would never see it
+
+## 12. Traps
+
+*Each of these cost real time to find.*
+
+- **`ctrl+b` is both herdr's prefix and Claude Code's "background this task"**
+    - Press it twice inside a Claude pane to reach Claude, or rebind herdr's prefix
+- **Never submit a prompt to a blocked agent**
+    - A blocked agent waits on a permission prompt
+    - Injected text answers that prompt instead of giving an instruction
+    - Wait for idle
+- **`--cwd` for `herdr worktree create` must be the main checkout**, and never a nested worktree
+    - herdr stores it as the workspace's repository root
+    - It also groups worktree workspaces under it in the sidebar
+- **Read the created path from the command output**
+    - Never assume where a worktree landed
+        - The configured root decides
+- **Worktree workspaces nest, and plain workspaces do not**
+    - A workspace made with `alt+w` always sits at the top level, whatever directory it points at
+- **Sidebar actions act on the selected workspace**, and not on the focused pane
+    - Open the sidebar and select before creating a worktree or closing a workspace
+- **In-process teammates are invisible to herdr**
+    - A Claude teammate shares its parent's pane and session
+        - It never appears as its own agent
+    - Two agents in one worktree means two real sessions
+- **gh-dash never says why an action failed**
+    - It discards gh's stderr
+    - A refusal shows as `exit status 1` for two seconds
+    - `acstack-gh`, which `acstack-dash` puts on `$PATH` as `gh`, turns that into a notification quoting gh
+    - Approving your own pull request is the one you will hit
+- **A dev machine has no clipboard and no browser**
+    - The `o`, `y`, and `Y` keys in gh-dash all fail there until `apply` installs `acstack-clip` and `acstack-open`
+    - The `setup.sh status` command says which browser route a machine ended up with
+- **Snowflake's browser login cannot work on a dev machine**
+    - `externalbrowser` authentication opens a browser, and that machine has none
+    - Use a key pair or a password, held in `~/.snowflake/connections.toml`
+    - `--read-only` guards a postgres connection only, so a snowflake profile has no such switch
+- **diffnav aborts on one very long line**
+    - delta dies on `SIGABRT`, and diffnav quits with `FATA signal: aborted (core dumped)`, showing none of the diff
+    - `acstack-diff` drops any file whose patch holds a line over 20,000 characters and names it in the banner
+    - A committed source map is the usual culprit
+        - The file is worth keeping out of git in the first place