Repository navigation
Tags
Tags
feat(cards): the built-in cards sound new - Submarine, Bottle, Sosumi… … and Blow (#151) ok plays Submarine, warn plays Bottle, fail plays Sosumi and info plays Blow. Agent cards that need you take the info sound, so they play Blow. All four are macOS system sounds, so drip ships no audio files. The README's custom preset example moves from Submarine to Hero, so it still sounds unlike ok.
feat(web): the screen shows the bare notch while the engine loads, an… …d a message only if it fails (#148) * perf(web): the fonts are kept for good by the browser, and Cloudflare can send them ahead of the page build.sh names the two fonts by their content, as it names the engine, and writes _headers: the named files are immutable, and each page carries Link headers for its fonts, which Cloudflare replays as 103 Early Hints. The build fails when a page does not preload a font or still asks for one under its old name. * feat(web): the screen shows the bare notch while the engine loads, and a message only if it fails
feat(drip): a card leaves the moment its time is up, and comes in the… … moment it has a place (#122) A card waited for the stack to come to rest before it could leave or come in. Now nothing waits. - A card's place is a sum of springs. Each change adds a spring that starts from rest, so a card can leave or come in while the stack moves. What moves a card also moves all that hangs from it. - A new card drops out of whatever is last in the stack, a card on its way out included. - Cards due at the same moment leave as one. The round is one frame at 30 Hz, not 0.25 s, so neighbours 0.1 s apart leave apart. - The overflow drop gives a card a place the moment one is free. - A drop is drawn against the shape its anchor has in that same frame. Two cards that come out together no longer show a clipped edge where they part, and their content stays hidden until they part. - The title fades in without the 4-point rise. - New MovingStackTests fly stacks at 120 Hz and check every rule on every frame. CI on Linux flies every frame.
fix(drip): draw card icons at full resolution (#97) * fix(drip): draw card icons at full resolution An app's icon file holds pictures up to 2048 px but reports 32 pt, so drawing the whole image took the 64 px picture and stretched it to the card's 56 px. Draw the smallest picture with at least twice the pixels the icon needs, and shrink it with high-quality interpolation. Vector pictures draw as before. An SF Symbol's palette tint also reset its point size, so the symbol was drawn at 15 pt and stretched to 28 pt. Apply the palette on top of the symbol's own configuration. * test(drip): re-record snapshot references for the sharper card icons
feat(drip)!: --on-click and DRIP_FORWARD_ENV replace the built-in Gho… …stty, herdr and WezTerm click paths (#92) A click on a card now brings its app forward and selects its tmux pane, as before. For anything else, --on-click CMD on drip, drip run, drip hook and drip follow runs CMD with /bin/sh -c and a small versioned JSON payload on stdin (v, kind, title, id, app, machine, env). - DRIP_FORWARD_ENV names the variables a payload carries, from the local shell or, for drip follow, from the agent's session on the other machine, which replaces the hook's fixed list. Names, values and count are checked on both sides. - Only the Mac side sets the command. A queued event cannot set a command, path, app or pane. - History never keeps or replays the click command. - Removed: the Ghostty raise and Space yield, open after a failed raise, the herdr machine lookup, --herdr-machine, and herdr and WezTerm pane selection. Recipes in docs/recipes.md do the same with --on-click. - A card from another machine no longer asks which pane shows there: it shows as if you were away. - The zsh block and drip hook no longer read an orchestrator's task variable. - docs/contract.md lists what scripts can rely on and what stays private. BREAKING CHANGE: --herdr-machine is gone, and clicks no longer select herdr or WezTerm panes or raise a Ghostty window on another Space without an --on-click command.
feat(drip): approve an agent's permission from its card (#87) * feat(drip): approve an agent's permission from its card With approve = true in the config, or DRIP_APPROVE=1, the card for a Claude Code or Codex PermissionRequest has Allow and Deny, and a click prints the agent's decision on the hook's stdout. Nothing clicked in 9 minutes prints nothing, so the terminal asks as before, inside the 10-minute hook timeout both agents use. An answer in the terminal clears the card: Claude Code ends the hook with SIGTERM on a deny, which drip turns into a dismiss and an exit within a second; an allow clears at PostToolUse. Codex holds its prompt until the hook returns, so its card asks only while you are not watching its terminal, and goes as soon as the terminal comes forward. Off by default. Remote, Gemini CLI, locked-screen and paused cards only tell, as before. * fix(drip): allow from the card only for a press of Allow A permission card that the stack dropped before it ever showed, because a newer card of the session replaced it or a dismiss came for it, left the drip exiting 0, and the hook read that as Allow. The stack now notes such a card as missed, and the hook prints Allow only when CardActions recorded a press of Allow, and Deny only for a press of Deny. Every other ending leaves the request to the terminal.
feat(drip): a card for any slow zsh command, from drip setup zsh (#86) * feat(drip): a card for any slow zsh command, from drip setup zsh * fix(drip): find the zsh block's markers without Foundation's range(of:range:), which the wasm build lacks * fix(web): the page's engine answers drip setup zsh and the hidden finished command * fix(drip): the zsh block keeps its start time when the rc file is sourced again, ignores the user's shell options and a bad threshold * fix(drip): cards show a command's time in whole seconds
fix(drip): let cards whose time runs out together leave one at a time (… …#85) * fix(drip): let cards whose time runs out together leave one at a time Three cards whose countdowns ended within one exit of each other all started leaving at once: the bottom card's drop was still rising when the top card left, and the middle card, held only until the drop going into it had merged, left at the very moment the top card's drop let go of its place, so it never moved up and retracted from the middle with a gap above it. A card whose time is up now leaves only from a stack at rest, once the last card's drop has gone and the others have settled into their places, the one longest past its time first. A card taken away on purpose still goes at once. The overflow tests that pinned two same-tick countdowns leaving in one step now expect one at a time; their guards against a stuck or doubled card are unchanged. * no-mistakes(review): Let timed-out cards leave before the next card enters
fix(drip): drip-remote-hook reads its whole stdin payload even when i… …t queues nothing (#84) The hook returned before reading stdin for an agent it does not know, or when it could not make the queue's directory or temp file, and exited. An agent still writing the payload then got a broken pipe: a payload past the pipe buffer failed every time, a small one whenever the hook ended first. That was the RemoteTests theHookNeverFailsTheAgentsTurn flake on Linux CI (EPIPE writing "{}" after the hook for "cursor" had exited). The hook now reads the rest of stdin after queueing, unless the payload came as an argument or stdin is a terminal. The test sends payloads past the pipe buffer, so it fails every run without the fix.
fix(drip): watch the mouse only for cards that take clicks (#83) * fix(drip): watch the mouse only for a lone card that takes clicks IslandPresenter started the 60 Hz pointer-gate timer and the system-wide mouse-move monitors for every lone card, though a card that takes no clicks never needs them. Start them only while the shown card is clickable: in start, and in replace when a replace makes the card clickable; stop them when a replace makes it non-clickable. The panel ignores the mouse from the start, so a card that takes no clicks lets every click through, as before. StackPresenter is unchanged: its 60 Hz tick drives the stack's countdowns and motion anyway, and a stack can gain a clickable card at any time. * test(drip): pin two countdowns that end on the same tick Two cards whose time left is equal expire in one step of the stack's clock. Pin what drip does then: in a full stack with an overflow drop, both leave in that one step, oldest first, the cards waiting grow out of the drop in the order they came and none is shown twice or stuck; in a stack of two with a card queued, both leave in one step and the queued card comes in. Passes as is, so the tie exposed no bug. * chore(drip): keep scripts/sim-overflow.sh as a permanent tool Its TEMPORARY header said to remove it once the overflow drop's motion was tuned. The motion has shipped, but nothing else in the repo replays the overflow drop on a real screen: record-demo.sh does not, and mac-e2e.sh and the tests do not cover it at the notch. Keeping it is cheap and it is the way to film the drop or look at its motion after a change, so make it permanent: drop the header and document it with the other scripts in the README. * no-mistakes(document): Rewrap sim-overflow.sh header comment
PreviousNext