708 Commits

Include renames

Author	SHA1	Message	Date
l0ng-aiandl0ng-ai	ce4a19ae89	fix(remote): let the supervisor's view of a link be the only one (#517) ... * feat(i18n): give the strip and the switcher the words for a link that is retrying The strip could count reconnect attempts but never say what any of them ran into, and the switcher had no word at all for a machine between links or for a workspace another client is holding. * fix(remote): let the supervisor's view of a link be the only one Four bugs met in this file's status pipeline, all of them a second opinion about a link that the supervisor already knew better. Take Back said it worked and did nothing (#488). Preempting a GUI client does not drop its control link — the server only hangs up on a `dedicated` one — so the pump's live branch marked the machine Attached and moved on, while the only site that ever sent `WorkspaceAttach` sat in the *reconnect* path, which a live link never reaches. The strip read Attached, `accepts_input` agreed, and every keystroke went into a socket the far end had handed to somebody else. So the live branch now attaches too. `reclaims_due` collects the workspaces the far end has not been told about — the ones a Take Back is reclaiming, and any this client has never spoken for over the link that is up now, because `connect_blocking` brings a link up and stops there and every switcher-initiated connect used to leave the daemon holding nothing. `attaching` keeps a reclaim that is on the wire from being sent again four times a second, `attach_sent` scopes "already told them" to one link, and `finish_reclaim` puts the takeover back — with the name that came with it, which is why `reclaiming` now carries it — when the far end refuses. While a reclaim is in flight the workspace reports Connecting: not Attached, because it is not, and not Preempted, because the button that would say so must not be clickable twice. A window's own failed connect no longer outranks the supervisor (#489). It was returned before `RemoteLinks` was consulted at all, and never compared against the window's own machine, so a failed connect to the GPU box replaced the strip of a window sitting happily on the build box. `resolve_status` takes both, and `clear_window_failures_for` retires what the windows still say once the supervisor gets through — precedence alone would not do, because the switcher reads `ConnectFlow` directly. And a reconnect says why (#498). `launch_attempt` set `Failed(e)` and the next tick overwrote it back to `Reconnecting`, 250ms later, so the reason existed for a quarter of a second. `MachineLink::last_error` outlives that. Retrying for ever stays: this is a laptop lid closed overnight, and giving up would be worse. * fix(switcher): draw a machine from what the supervisor knows, not from a table entry `link_state` read this window's `connect` and then asked whether a `HostLinks` entry existed. The pump drops that entry the moment a link dies, so a machine being retried right now rendered identically to one nobody has ever connected to — same grey dot, same "not connected", and, because the group body is only expanded when the link is not Offline, its workspaces vanished from the panel for the whole reconnect (#497). `Link::Reconnecting` is the state that was missing, and `link_from` builds it from `RemoteLinks::machine_status`, so the panel and the strip now read the same source. Disconnect and the machine menu treat it as engaged: a retry in flight is exactly what Disconnect is for. The route `launch_attempt` could not build wrote `LinkState::Failed(e)` where nothing the switcher read would ever find it. The group's error line now falls back to the supervisor's failure and to `last_error`, rather than having the supervisor write into `remote_host_errors` as well — one source, not two. A takeover leaves the link alone, so a machine whose workspace another client is holding drew as plain Connected with an "Open" badge on the row. It says so now, on the header and on the row it happened to. --------- Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-11 21:42:37 +08:00
l0ng-aiandl0ng-ai	bf9c57dec7	fix(ssh): let a rejected stored credential ask again (#519) ... * fix(ssh): let a rejected stored passphrase ask again (#486) Saving the wrong passphrase for an encrypted key locked that key out permanently. `passphrase_submit` wrote `SetKeyPassphrase` on the "remember" checkbox alone — before the daemon had tried the secret, since `apply_keychain_write` runs ahead of `respond_active` — and `try_identity_file` treated a stored passphrase as final: a decrypt failure with one went straight to "could not decrypt identity file", with no prompt and nothing in the UI that could let go of it. The daemon now says so. `AuthPromptKind::KeyPassphrase` grows a `rejected` flag, and a stored passphrase that does not open the file falls through to the interactive prompt carrying it, so the typed answer still gets its attempt. A passphrase the user typed this time keeps the hard failure — that is a wrong answer, not stale state. The sheet renders the warning line the password sheet already had, and a rejected prompt answered without "remember" now emits `DeleteKeyPassphrase`, mirroring the password idiom exactly. The flag is a `#[serde(default)]` field on a struct variant of an externally tagged enum, which is compatible in both directions: an older peer never sets it and serde ignores fields it does not know. So `PROTOCOL_VERSION` deliberately does not move — the remote-server handshake gates on it, and a bump would turn away older servers over a field they can safely ignore. `protocol.rs`'s compat test pins both directions. Also: deleting an SSH profile now drops the key-passphrase entries no other profile still references, which is what `delete_profile_confirmed`'s own comment already claimed to do but only ever did for the password. * fix(ssh): stop replaying a stale password at keyboard-interactive (#487) `try_keyboard_interactive` answered a password-shaped round from the keychain, marked the stored password spent whether or not it had been used, and returned on the first `Failure` — so the `MAX_ROUNDS` loop never got a second pass with the stored password withheld. The same dead secret went out on every reconnect and the user was never once asked to type a different one; `ki_submit` always emitted `KeychainWrite::None`, so nothing could clear it either. `collect_ki_answers` now reports where its answers came from, and only a round that actually sent the stored password spends it — which also fixes an OTP-then-password flow that was refusing the stored password for no reason, its first round having burned the allowance on a code. On a rejection whose last round came from the keychain, and where the server still offers the method, the request is started over with the stored password withheld, so the next round reaches the prompt. That retry is bounded twice over: the restart spends the stored password, so no second restart can qualify, and the round counter it shares with the info-request loop caps the method either way. The failure text now says which of the two was turned down. Scope, honestly: the only live scenario is auth mode Auto against a server offering keyboard-interactive but not password, with a stored password for that endpoint — a profile pinned to KeyboardInteractive gets `password: None` and always prompts, and Password never tries KI. Whether the symptom shows also depends on the server: OpenSSH ends a rejected kbdint request with USERAUTH_FAILURE (symptom holds), while a device that re-issues an InfoRequest in the same request already reached the prompt. `AuthPromptKind::KeyboardInteractive` grows a `#[serde(default)]` `stored_rejected`, same both-directions compatibility as `KeyPassphrase`'s `rejected` and the same reason `PROTOCOL_VERSION` stays put. The sheet shows the warning line and, on submit, forgets the rejected password. That needed an endpoint the KI prompt does not carry, which also fixed a bug next door: `raise_routed_auth` called `from_prompt(.., None, false)`, so every routed password write was keyed to port 22 regardless of the real port and the rejected self-heal could never fire there. `PendingAuth` now carries the endpoint and the auto-supplied flag, read straight off the route's `NativeSshSpec`. --------- Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-11 21:40:49 +08:00
l0ng-ai	d4643f0532	fix(ssh): offer keys in the order the user asked for them (#520) ... Every public key offered spends one of the server's `MaxAuthTries` — six by default — whether or not the server wants it, so the order decides who gets locked out when the budget runs dry. Offering the `~/.ssh` defaults before the agent, and offering them on top of a profile's own key rather than instead of it, spent the budget on the keys least likely to be accepted: three stale defaults in `~/.ssh` with the working key in a loaded agent could exhaust the attempts before the agent was reached, and a profile naming its own key came off worse than one naming none. Order the sources by how plainly the user asked for the key: 1. a key the profile names — "use this one" 2. the agent — "I loaded these" 3. the ~/.ssh defaults — nobody said anything, we are guessing Steps 1 and 3 are the same leg, because the defaults now stand in for the identity list only when the profile names no key of its own, the way `IdentityFile`'s default works in ssh_config — verified against the local OpenSSH: `ssh -G` lists five defaults with no IdentityFile and only the named key with one. WezTerm and Tabby both replace rather than append too (wezterm-ssh/src/config.rs:589, tabby-ssh/src/session/ssh.ts:194). The file-vs-agent order has no single convention to follow — WezTerm asks the agent first, Tabby reads files first, OpenSSH merges the two and trims the agent's extras with `IdentitiesOnly`, which tty7 does not parse. This order agrees with OpenSSH and Tabby when the user named a key and with WezTerm when they did not. Dedup by canonical path goes with it: the two lists are alternatives now, so there is nothing to dedup between them. `auth_steps` and `identity_offers` carry the two rules as pure functions, so both are unit tested instead of living inside the async round. The GUI's keychain preload follows the same rule, so both sides still key passphrases by the same strings. Closes #513.	2026-08-11 21:23:17 +08:00
l0ng-aiandl0ng-ai	16ef93693f	fix(settings): confirm before forgetting a password other profiles share (#510) ... "Forget Password" was a bare menu item: one click deleted the keychain entry, with no confirmation and nothing said about who else it took down. The entry is keyed by `user@host:port`, so two profiles that reach the same endpoint — one direct, one through a jump host — share exactly one secret, and forgetting from either row signed both of them out. The notification even worded itself by endpoint while the action hung off a single profile's menu. It now asks first, the way deleting a profile does, and when the endpoint is shared the dialog names the blast radius instead of leaving it to turn up at the next connect on a host nobody touched. Deleting a profile stays conservative on purpose — the menu that could remove the secret is about to disappear — so the two paths keep their different policies. What they no longer keep is two copies of the "is this endpoint shared" question: `profiles_sharing_endpoint` is now the one place that answers it, and it has the test. Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-11 21:22:52 +08:00
l0ng-aiandl0ng-ai	01969ef6bb	fix(settings): say what an ssh config import added, updated and could not keep (#515) ... Importing from ~/.ssh/config was silent in three ways. A missing or unreadable file did nothing; a file of nothing but `Host *` and `Match` did nothing; and a successful import did nothing visible either, so the only way to learn what had happened was to go count the host list. Options tty7 has no field for — IdentityAgent, CertificateFile, AddKeysToAgent and the rest — were dropped without a word. They still are, because there is nowhere to put them, but the import now names them and the hosts that set them instead of pretending they were kept. Parsing keeps each keyword's original spelling alongside the lowercased form it matches on, and `option_is_supported` is the one list both the resolver and the report read, so the two cannot drift. Ignored options are grouped per Host block rather than per resolved alias: a keyword under a two-alias `Host` line is one omission, not two, and `Host *` noise stays out of the report entirely. `merge_imported` now returns added/updated/unchanged, comparing the six fields it writes before it writes them — so re-importing an unedited file reports six hosts unchanged rather than six updated. Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-11 21:22:36 +08:00
l0ng-aiandl0ng-ai	9c9a18f410	fix(settings): refuse a half-filled SSH profile instead of saving it (#511) ... The SSH profile form saved whatever was in it. An empty host wrote a profile that renders as a blank row in the host list and hands `TcpStream::connect` an empty name — and Connect had no gate at all, so the first thing the user heard about it was a DNS error naming nothing. A jump host with a typo in it resolved to `None` and saved as a direct connection, with the field blank the next time the form opened. And `parse_host_port` was `parse().unwrap_or(0)`, so `proxy.example.com`, `proxy.example.com:` and `proxy.example.com:88O` all saved a proxy on port 0, which the socket layer got to explain (#492, #493, #494). The rules now live in `validate_ssh_draft`, a plain function over a plain-String snapshot of the form, which returns both the profile the form would save and what is wrong with it. Both, always: the Escape prompt asks whether the form differs from the config, and handing back only the errors would make a brand-new invalid profile compare equal to the nothing on disk — Escape would throw the typing away without asking. Only the host is required. A name is not, because the list already falls back to the address and every host imported from ~/.ssh/config arrives without one. A blank port still means 22, but a non-empty one has to be a port, so "0", "abc" and "70000" are refused rather than saved as written or quietly rewritten. A proxy address with no port takes the scheme's default (1080 / 8080) and `host_port_text` writes that back into the field, so the number it picked is visible; a colon with nothing usable after it is an error. `map_proxy` also stops treating port 0 as a proxy, because configs written before this are already on disk. Each complaint prints under the field it is about, and Save and Connect are disabled while any of them stands; a section holding one unfolds so the disabled button always has a visible reason. The "needs a host" line waits until the name/host/port/user group has something in it — every field notifies per keystroke, so otherwise a new host would be told off before anyone had typed a character. Consequence: on a pristine new form Save is now disabled where it used to be enabled. Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-11 21:22:19 +08:00
l0ng-aiandl0ng-ai	efe345174b	fix(ssh): stop a new host-key algorithm from reading as a compromise (#516) ... * fix(ssh): stop a new host-key algorithm from reading as a compromise A host that grows an ed25519 key beside the ssh-rsa one it has always had raised the full man-in-the-middle sheet — red border, fingerprint diff, a "type yes" field — because `check_in_str` folded "known by another algorithm" into `HostKeyStatus::Changed`. OpenSSH treats a key of an algorithm the host has no entry for as simply unknown, and saves the alarm for a key that contradicts one on file. `ChangedAlgorithm` splits the two apart, with `Changed` keeping precedence so a same-algorithm mismatch still screams however many other-algorithm lines sit beside it. The dialog was only half of it. Negotiation started from russh's default order, which leads with ed25519, so a host known only by ssh-rsa was *asked about on every single connection* — and an attacker could pick an algorithm the user had no entry for to trade the alarm for the mild confirmation. `build_preferred` now orders the host-key list the way OpenSSH's `order_hostkeyalgs()` does: what is already on file goes first, nothing is dropped, and a pinned `HostKeyAlgorithms` is left alone. It matches on key type, so all three RSA spellings travel together rather than pinning the host to SHA-1 signatures. The prompt reuses `AuthPromptKind::HostKeyUnknown` with an added optional field rather than gaining a variant: the enum is externally tagged and crosses both the daemon/GUI and the GUI/tty7-server boundaries, where a new variant is a hard decode failure on an older peer and a new field is not. Also fixes a defect the issue did not mention: overriding a genuinely changed key appended the new line without removing the old one, and since any same-algorithm match answers `Known`, the superseded — possibly attacker's — key stayed trusted forever, silently. The superseded line is now dropped first, and only lines naming this one host are touched, so a wildcard or `@revoked` entry is never collateral. * fix(ssh): make the Override button on a changed host key actually override `host_key_changed_decision` returns `accept: false` for anything but "yes", which is byte-for-byte what Abort sends — and the button had no disabled state and closed the sheet unconditionally. So clicking Override with an empty field rejected the key and dismissed the prompt, indistinguishable from having aborted, with nothing said. Enter on the input had the same trap. Override is now dead until the word is there, which is what the line above the field has been claiming all along, and Enter on a half-typed answer leaves the sheet up instead of quietly deciding. `changed_confirmed` is the single predicate behind both, so the button and the decision cannot disagree about what "yes" means. `host_key_changed_decision`'s `false` branch stays as defence in depth. Both input subscriptions also notify on `Change`, or the enabled flag would go stale between keystrokes, and a hint appears once the field holds something that is not "yes". Abort is untouched: still primary, still last. --------- Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-11 21:22:03 +08:00
l0ng-aiandl0ng-ai	c91f0cf660	fix(forwards): say why a forward was refused, and never lose the rule it replaces (#514) ... The managed-forward form in the Forwards panel used to bare-`return` from Add whenever the fields did not make a rule — an unparseable bind port, a target port of 0, an empty target host for a Local or Remote rule. The button did nothing, said nothing, and left the form exactly as it was. It now collects the fields through the same conditions the settings sheet applies (`ForwardRuleForm::collect`), keeps Add disabled while there is nothing to add, and shows the settings sheet's own "needs both" / "needs a listen port" line under the form once it has been touched. Saving an edit removed the old forward and then added the new one, so a rule that could not be started took a working forward with it. Remove-then-add is still the order — the ordinary edit keeps the bind port, and adding first would collide with the very rule being replaced — but the panel now looks at the entry the add appended, and on `ForwardStatus::Error` it takes that entry back out, puts the old rule back, and keeps the form open with the reason under it. `mf_editing` carries the whole `ManagedForward` rather than an id so there is something to put back, and it is re-pointed at the restored entry, which comes back under a new id. Adjacent, and the reason a failed request could not be told from a successful one: `ForwardRoute::add`/`remove` and their `RemoteTerminal` fallbacks turned every transport error into an empty `Vec`, which the caller then assigned straight into the panel's list — so one unreachable round trip blanked a panel full of live forwards. They return `Option` now, and the panel only takes a list the far side actually sent. Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-11 21:21:47 +08:00
l0ng-aiandl0ng-ai	f2fe829cb6	fix(sftp,forwards,files): tell a failure apart from an empty result (#518) ... * fix(sftp): stop a failed poll from reading as an empty transfer list A transfer poll that could not reach the daemon answered with an empty `Vec`, which is indistinguishable from "every transfer is gone": the tray disappeared and every upload the panel was waiting on counted as landed, so a spurious "the upload finished" refresh fired. Over a link that is down that is the permanent answer, not a blink. `SftpRoute::transfer_list` and `RemoteTerminal::sftp_transfer_list` now report the failure the way `sftp_list` already does. A failed poll keeps the jobs the panel last saw, settles nothing, and says so in the transfer tray through a new `jobs_error` — kept apart from `SftpPanelState::error`, which blanks the directory listing a poll knows nothing about. * fix(forwards): let a forward whose loop has exited say so `ForwardEntry.status` was written once when the forward was set up and never touched again, so a local or dynamic forward went on reporting `Listening` after its accept loop had already exited. The pane outlives a dead SSH transport on purpose, the daemon keeps it while a subscriber is attached, and the panel re-polls every 2s — so the stale `Listening` is not a blink but the permanent answer. `nc` to the port gets accepted once and refused thereafter while the panel still shows it as live. The status is now an `Arc<Mutex<ForwardStatus>>` shared with the task, and both break arms record why they left: the listening socket closed, or the SSH connection went away. `ForwardStatus::Error` carries it rather than a new variant, because the enum crosses the protocol to `tty7-server` builds that would not know one. `find_auto_local` reads the live status too, so a loopback link is no longer reused after its forward has stopped serving. A remote forward has no accept loop of its own — the far end opens the channels — so it keeps whatever the `tcpip-forward` request answered. * fix(files): tell a failed search from an empty one, and name the file a write failed on Two ways the file tree answered a failure with something that reads as a result. A search was `unwrap_or_default()`ed inside the worker, so a host that refused the walk left `hits` empty and the column printed "Nothing matches {query}" — byte-identical to a genuine zero-hit search. The worker now reports `(ok, hits)` the way `spawn_load` already reports a listing, and a failed walk draws a `SearchFailed` note in the danger colour, the same distinction `FileTreeState.unreadable` draws for a directory. Creating and renaming pushed the bare `io::Error`, so the toast was literally "Permission denied (os error 13)" — neither which file nor what was being done to it. Both now go through `HostOps::notify_err` like delete and drop-copy already do, naming the file; a rename names the name it is leaving, which is the one still on screen. --------- Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-11 21:20:39 +08:00
Hongwei Qin	364e24af87	feat(ssh): probe ~/.ssh default identity keys (#507) ... A connection with no explicit IdentityFile used to offer the server nothing at all when the agent was unavailable — the default on Windows, where the OpenSSH Authentication Agent service ships disabled — and then reported "no public key was accepted", for keys it had never sent. Offer the `~/.ssh` defaults (id_ed25519, id_ecdsa, id_rsa) after the explicit identities and before the agent, from one shared candidate list in `core::ssh_profile` so the GUI and the daemon key `key_passphrases` by the same strings. Candidates are deduped against the explicit list by canonical path, comparing the expanded paths the reader actually opens. A discovered key that is encrypted is used only when its passphrase is already cached, never prompted for; explicit keys keep prompting. Files that do not exist are skipped in silence, a `.pub` is never offered as a private key, and the failure text now separates "the server turned these down" from "nothing usable was found". The tests build their ed25519 fixture at run time from a fixed seed rather than embedding a PEM blob, so the tree carries no private key. Closes #484.	2026-08-11 20:03:02 +08:00
l0ng-aiandl0ng-ai	72060d8257	fix(ui): stop washing out the workspace discs in the switcher (#483) ... The monogram disc beside every workspace row was drawn at opacity 0.55 unless that row was the current workspace, and the initial inside it is already the foreground at 0.65 — so the letter landed at about 0.36 and went illegible on exactly the rows the panel exists to let you pick between. Only one row in the list is ever the current workspace; the other however-many all got the dimmed treatment. The liveness dot is painted on the wrapper rather than inside the disc, so it never dimmed with it: a washed-out grey blob with a full-strength green dot stuck to its corner, which is what made the column look dirty rather than quiet. Drop the dimming. Nothing is lost by it — the current workspace already carries a "this window" badge, a medium-weight name and the selected background, so the disc was saying a fourth time what three louder things had said. `current` was the only reason the helper took that argument, so it goes too. Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-11 11:26:27 +08:00
l0ng-aiandl0ng-ai	c4645aeea8	feat(terminal): draw a scrollback scrollbar down the right edge of a pane (#480) ... A pane's scroll position lives in alacritty's `display_offset` — rows of scrollback, not pixels of laid-out content — so it has no `ScrollHandle` to hand a scrollbar. `TerminalScrollHandle` implements gpui-component's `ScrollbarHandle` over the grid instead, which lets the pane draw the same `Scrollbar` the sidebar and every list already use: same theme, same `Scrolling` show mode, same fade-out. The bar never touches the terminal. `set_offset` only records the row it wants; `sync_scrollbar` applies that on the next render — clearing the sub-line remainder and cancelling an in-flight smooth scroll on the way — and reports back where the grid actually ended up. Scrollback piling up at the live edge is deliberately not reported: the bar shows itself whenever the offset it reads changed, so a pane printing a build log would otherwise hold a thumb on screen for as long as the output ran. Every other change passes through, including the history shrinking, which is a cleared scrollback rather than growth. Closes #432 Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-11 10:54:12 +08:00
l0ng-ai	9735a490b5	Merge pull request #479 from l0ng-ai/perf/wsl-tab-open-cost ... perf(wsl): stop re-proving the distro on every new tab	2026-08-11 10:35:59 +08:00
l0ng-ai	f676fb96de	perf(wsl): stop asking twice whether a distro is ready ... Every pane on a WSL workspace ran `ensure_wsl_server` from the client before it even connected to the daemon — and then the daemon ran the very same probe inside `router::open_link` before opening the link. Two full rounds of five serial `wsl.exe` calls, to learn one fact. The client's copy bought nothing. It threw the answer away and kept only the error, which the route ack reports just as well; and the consent question for a first install still finds its way here, because the daemon runs its probe under `RouteSetup::blocking`, which installs the relay that turns that question into a frame on this connection. Measured on a distro that was already running and connected: 800ms to open a tab, down to 440ms. Issue #454 is the same code on a machine where one `wsl.exe` round trip takes 3.3s, where the duplicate was costing 15s a tab.	2026-08-11 09:36:51 +08:00
l0ng-aiandl0ng-ai	af3928da1e	fix(tree-sync): pay back a remote window's owed tree pull (#472) ... * fix(tree-sync): pay back a remote window's owed tree pull A window opening onto a remote workspace is empty until `hydrate` pulls the machine's tree and rebuilds its tabs from it, and it has to be: an empty window diffs into "close every tab", so `sync_window` holds anything back until the pull lands. When the pull fails, `owe_rehydration` records the debt and returns, on the promise in its own doc comment that the next sync settles it — "which is what a reconnect does through `on_link_up`". `on_link_up` is called for `HostId::LOCAL` and nowhere else. On a remote host the debt was only ever settled by a reconnect completing, by an edit in the window, or by restarting the app. So a pull that failed while the link stayed up was never noticed again: no reconnect, and an empty window has nothing in it to edit. The window sat on the home page with every tab and every shell still on the machine, and only a restart brought them back. Two ways to fail a pull with a healthy link, both routine. A `MachineGet` can overrun its ten seconds on a slow link. And a `WorkspaceCreate` can lose its race with `start_prime`, which runs the same create from the other side of the same window opening — that one fires on every remote workspace opened, and is only invisible because the workspace it usually lands on is empty anyway. So: arm a backed-off retry when the debt is taken on, drive it through `sync_window` where the rules about whether a window may still adopt the machine's layout already live, and stop treating a lost create as a failure — read the tree again and hydrate from what is really there. `on_link_up` is also wired to a remote link coming up, which is what the comment always claimed: a link the switcher connects finishes no attempt, so nothing told its windows the machine could be reached. * fix(tree-sync): end the backoff with the run of failures, and stop shouting Review follow-ups on the owed-pull retry. The attempt count paces the retry, so it has to mean "failures in a row", but it was only cleared when a hydration landed. A debt abandoned rather than paid — a `Replace` dropped because the user filled the window in themselves — and a prime that landed both left it standing, so the next first failure waited the 30s cap on an outage that was already over. It is now cleared wherever the run ends. A window left open on a machine that is really gone retries forever by design, which meant a warn and an info every ~45s for as long as it stayed open. Once the backoff settles at its cap those lines stop being events and become a fact about the machine, so they step down to debug. The retry is exactly as persistent; only the volume drops. Also: report the create's own refusal when the reread finds the workspace still missing, and say at debug that the reread happened at all — the race recovery was silent, so the extra round trip was invisible when reading a log. And correct the comment on the window-gone guard: closing a window drops its whole `WsState` through `forget`, debt and all, so nothing is parked for the next opener. Tests: the count ends with the run at all three sites, the level steps down at the settle point, and the armed retry is driven through a real timer (advance_clock) into the window-gone guard — the first coverage of the retry actually firing rather than of the predicate it consults. --------- Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-10 23:47:29 +08:00
ARNO	12df66fb0f	fix(ui): dim panes by blending terminal colors toward the window background (#464) ... * fix(ui): dim panes by blending terminal colors toward the window background * fix merge * fix(ui): tighten the pane-dim blend	2026-08-10 23:07:37 +08:00
l0ng-aiandl0ng-ai	0f5e63701e	fix(ui): let the overlay scrollbars fade out again (#471) ... * fix(search): wash a match in the accent, at a strength the theme can afford A search hit was washed from the terminal palette's selection colour at a fixed 1.45:1 against the background, so it read as a weaker selection on a grid that is already grey on grey — and 1.45:1 is under what a hairline is worth, spread over a whole cell. Two changes. The tint is now the theme's accent (`ActiveAccent`, already floored at 3:1 by `legible_accent`), which is the one colour the terminal surface has nothing else in. And the strength is derived per theme instead of fixed: the wash is opaque with the glyph drawn on top, so what it may spend is the theme's own text-contrast budget. A palette with 21:1 between text and background can afford a wash you cannot miss; one with 6.6:1 cannot, and a single constant has to be safe for the second. The current match drops its caret-coloured outline. That existed because a fill 2.1:1 off the background could not say "this one" on its own; now that it sits at the top of the theme's budget, the outline is the same colour saying the same thing twice. * fix(ui): let the overlay scrollbars fade out again macOS reports should_auto_hide_scrollbars() = false for anyone with a mouse plugged in, and apply_theme turned that into ScrollbarShow::Always for every list in the app. That preference is about legacy scrollbars, which take a gutter out of the layout; ours are overlay bars painted on top of the content, so Always parked an opaque bar over the switcher's tab column for as long as the panel stayed open, with nothing to fade it. Pin scrollbar_show to Scrolling instead, so every list fades its bar out after it stops scrolling. --------- Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-10 18:36:21 +08:00
l0ng-aiandl0ng-ai	99a388331c	fix(search): wash a match in the accent, at a strength the theme can afford (#470) ... A search hit was washed from the terminal palette's selection colour at a fixed 1.45:1 against the background, so it read as a weaker selection on a grid that is already grey on grey — and 1.45:1 is under what a hairline is worth, spread over a whole cell. Two changes. The tint is now the theme's accent (`ActiveAccent`, already floored at 3:1 by `legible_accent`), which is the one colour the terminal surface has nothing else in. And the strength is derived per theme instead of fixed: the wash is opaque with the glyph drawn on top, so what it may spend is the theme's own text-contrast budget. A palette with 21:1 between text and background can afford a wash you cannot miss; one with 6.6:1 cannot, and a single constant has to be safe for the second. The current match drops its caret-coloured outline. That existed because a fill 2.1:1 off the background could not say "this one" on its own; now that it sits at the top of the theme's budget, the outline is the same colour saying the same thing twice. Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-10 18:26:18 +08:00
l0ng-aiandl0ng-ai	fafcaa0685	feat(splits): draw the pane grip as Ghostty draws its own (#463) ... The bar a pane was picked up by grew and recoloured under the pointer, and showed the moment the pointer was anywhere in the pane at all. It was loud in the wrong places: a mark on top of the terminal wherever the mouse happened to rest, and a target that moved while being reached for. Cut to the shape Ghostty gives its grab handle instead: * three dots, 80x12 of reach around them, and nothing between the two states but ink — 0.3 in the band, 0.8 on the grip itself. * the dots are asked for by the pane's top fifth (floored at 24px), not by the whole pane, so the terminal is left alone everywhere else. * the target is there for as long as the pane can be moved, and only the dots come and go, so a pointer going straight for the top of a pane can press the grip on the frame it arrives. * a 150ms fade in, so the dots read as arriving rather than blinking. The fading a pane is under is now worn by the terminal it holds rather than by the pane, which keeps the grip legible on the very panes `dim_inactive_panes` fades — the ones being reached for. Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-10 17:24:40 +08:00
l0ng-ai	a55340ed7f	fix(daemon): sweep a dead daemon's leavings on the writer's tick, not at startup ... Review follow-ups on this branch. `history::sweep` still ran at startup, three lines under a new comment explaining why sweeping there is wrong. The reasoning transfers exactly, and worse than by analogy: a restore carries the dead pane's commands to its successor via `history::carry`, so sweeping before the window can ask deletes the file the request is about. Same shape as the scrollback bug, one file over. Both sweeps now run on the writer's tick off one shared id set, and the writer is named for what it does. `pane_attachable` lost its only caller when the restore path moved to `pane_free_for`, leaving a function kept alive by the test asserting on it. The attach site does not need to predict the listing: it tries the attach, and a pane that is gone falls through to the fresh spawn on its own. Gone, with its tests folded into `pane_free_for`'s. `restored_screen` now drops the snapshot in both directions. Keeping the file when it decoded to nothing left it to be re-read and re-rejected by every later restore, and swept never, for a pane the tree still names. Also: the module doc still said scrollback was off unless asked for, which is what this branch reverses; and #449 landed the whole feature with no CHANGELOG entry, so nothing told anyone that pane output now lives on disk.	2026-08-10 16:36:35 +08:00
l0ng-ai	852d3178c8	style: rustfmt	2026-08-10 16:10:58 +08:00
l0ng-ai	477d82524f	feat(daemon): keep every pane's screen, without asking ... `persist_scrollback` is gone, and with it the switch, its three translations and the branches that read it. Keeping a capped tail of each pane's output is now what the daemon does, not something it can be asked to do. This reverses the call made when the feature landed. The argument for off-by-default was that the ring holds whatever the pane printed — echoed tokens, `env` output, an agent's transcript — and that writing that down should be the user's decision to make. What the argument missed is when the decision gets made: the moment anyone learns they wanted this is the moment a daemon has already died, and by then the setting could only be turned on for next time. A feature whose entire purpose is to survive an event nobody schedules cannot be opt-in. The cost is real and does not go away: pane output now lives at `<config>/scrollback/*.bin` on every machine, 0600 on unix and behind the config directory's ACL on Windows, capped at 256 KiB per pane and dropped as soon as no window can still ask for it. Old configs naming the key still parse — nothing in `Config` refuses unknown fields — so the key simply stops meaning anything.	2026-08-10 16:06:54 +08:00
l0ng-ai	412bfcfc90	fix(session): let a dead pane keep its id so its screen can be asked for ... `pane_attachable` answered one question and was used for two. Deciding whether to attach needs to know the pane is alive; deciding which dead pane a fresh one replaces needs only its id — and that is the case the stored screen exists for. Using the first answer for the second was self-defeating. After the daemon restarts, every pane the window held is missing from the new daemon's listing, so the id was ruled out, so `restore_pane` was `None`, so the window spawned a pane that had never heard of a predecessor. No attach was tried, no restore was requested, and the screen the daemon still had on disk was swept a tick later without anyone reading it. The setting was on, the snapshot was written, the daemon was ready to hand it over, and nothing ever asked. Ownership still rules an id out, because another workspace's pane is neither ours to attach to nor ours to show. Liveness no longer does: the attach is still tried first and still gives way to a fresh spawn when the pane really is gone, which is the arrangement the tree path already argues for at length — `live` is a hint about what to show, never the judge of what to destroy.	2026-08-10 16:06:54 +08:00
l0ng-ai	3093babddd	fix(pane): say when a restore is not asked for ... Dropping the request here produced a blank pane, which is also what a pane with nothing stored looks like and what a daemon that refused would produce. Three causes and one appearance, with nothing anywhere to tell them apart — the filter was silent, so reading the source was the only way to find out which had happened.	2026-08-10 16:06:54 +08:00
l0ng-ai	c138be687a	fix(daemon): keep a pane's shell and its screen across a restart ... Two things a pane lost when the background service stopped and started, both of them things the tree was the only possible place to keep. **The shell.** `PaneRecord` and `PaneSeed` carried a pane's cwd, its ssh spec and its agent, but never what it was running. A window rebuilding a dead pane from the tree therefore had nothing to pass and spawned on whatever the default shell is now — so a restart turned a bash pane into a PowerShell one, quietly and in place. The daemon resolves the override against the config at spawn time and is the only party that knows the answer, so it keeps it and reports it; the seed carries it too, for the panes a window spawned itself. A handoff carries it in the blob, because nothing on the far side of an `execve` can work out the command line of a child it never spawned. **The screen.** The startup sweep ran before the endpoint was listening, which is the one moment nothing can answer the question it asks: the registry is empty and the windows that know which screens are still wanted cannot say so yet. A tree that failed to parse made it worse — `read_machine` quarantines it and returns an empty `Machine`, so one bad file took every pane's stored screen with it. The sweep now happens only on the periodic pass, a tick later, with the registry filled in and the tree caught up; nothing is serving a request in between. Turning the setting *off* still clears the directory at once, because there the promptness is the whole promise. Two smaller ones alongside it: `restorable_pane_ids` now counts the tree's pane list and not only the panes some tab currently stands on — the two disagree while a window is between layouts, and being wrong costs a file swept a tick late in one direction and somebody's terminal in the other. And `restored_screen` drops the snapshot file *after* deciding it was not empty, so a snapshot holding nothing is no longer consumed by the request it could not answer. The restore path had no end-to-end test, which is how this shipped: the unit tests cover the file, not whether a window that reattaches is shown anything. The new one runs a real daemon, puts a marker on a real pane, stops the daemon, starts another, and reads the wire.	2026-08-10 16:06:54 +08:00
l0ng-ai	2dc6a88af6	merge: main into the Source Control branch ... Conflicts were the two streams touching the same seams, resolved by taking the newer decision on each side: - main's interface font scale (rems tokens) wins in right_panel.rs; the SCM panel keeps its local px steps until it moves onto that scale, and the now-unused PANEL_TEXT constants are gone. - main's l10n_keys! macro (idents only) means the key list carries no doc comments any more; our SCM keys fold into it, and PanelUntracked stays deleted — its only caller was the panel this branch replaced. - main's Command::localized palette style carries our Git group; ORDER keeps main's visibility and our width. - main's ansi_seed/clear_ink refactor in presets.rs carries the lane colours: lanes() now clears through the same helper semantics uses. - file_tree keeps both: main's drag-and-drop targets and this branch's git decorations per row. - diff_overlay keeps both: main's sidebar-count write-back on snapshot install and this branch's epoch read and untracked preview. - main's window.prompt SSH-close confirmation supersedes the bespoke modal our branch still carried; main's tile-glyph revert stands. - main's two new guards are satisfied: the fourteen SCM actions carry authored names on the Keybindings page (their palette wording, plus a new CmdGitToggleGraph), ja translates ScmDetached, and CmdGroupGit joins the kept-in-English list — Git is a name. 2571 tests, 0 failures.	2026-08-10 13:25:30 +08:00
l0ng-ai	feb027da1f	feat(scm): show an untracked file's content when its row is opened ... Focusing an untracked file in the diff overlay used to fall through to the names-only "Untracked files (N)" card — git has no patch for a file it does not know, and `--no-index` needs a null device whose spelling is platform business. The overlay now reads the file's own bytes (lazily, only the focused file, 4 MiB cap) and synthesizes the card a parsed added-file patch would produce: every line an addition, new-side numbers, true counts past the single-file budget, git's own NUL-in-the-first-8000-bytes binary rule. A fresh snapshot clears the preview so an edit shows up on the same cadence a tracked file's does; a failed read says so instead of showing an empty file. Found in manual acceptance of the panel.	2026-08-10 13:06:15 +08:00
l0ng-ai	1df43b72b5	feat(files): copy dropped files into the folder they were dropped on (#458) ... * feat(files): copy dropped files into the folder they were dropped on The Files panel has only ever been a drag *source* — a row dragged into a terminal inserts its path. Nothing on the tree ever registered a drop, so a file dragged in from the desktop did nothing at all, not even a highlight. Closes #453. The drop is the whole gesture: files land where the cursor was, not somewhere a dialog asks about afterwards. A folder row takes them itself, a file row stands in for the folder holding it — "next to this one" — and the space the rows do not cover belongs to the top of the tree. The placeholder inside an empty folder takes a drop too; it is the only thing drawn there, and letting it fall through to the root would put files somewhere the cursor never was. A row under the cursor wins over the column, which is what gpui's innermost-first dispatch already does. The copy itself goes through the `Host` the tree is listing, so a remote workspace reads here and writes there. Locally it is `fs::copy`, which is what keeps the executable bit that `write_file` would drop; remotely the bytes ride one control frame, and a file too big for that is refused with the advice to use SFTP rather than half-sent. Names already taken are asked about before anything is written, and the answer governs the whole drop — a half-done copy would have to be undone to honour a "no". Replacing a folder replaces it rather than merging into it. A drag let go where it started is a miss, not an error, so it says nothing. * fix(sftp): list the directory again once an upload lands An upload is written to `<name>.tty7-upload-<hex>` and renamed into place at the very end. The browser listed the directory the moment the transfer was handed to the daemon, so it caught that temporary name — and nothing ever listed again, so a finished upload sat on screen as a file with a hash glued to its name until the directory was navigated by hand. The premature listing is gone, and the panel now remembers the job ids it started: once one stops running — done, failed, cancelled, or dropped off the job list entirely — the directory is listed once more. Two uploads in flight settle independently, so the second one finishing does not depend on the first. * docs(changelog): note the SFTP upload listing fix * ci(host-boundary): allow the source side of a file drop, and stop scanning two files as empty The Files panel now copies dropped files in, and what the desktop hands over is by construction a path on the desktop's own machine: reading it is a local read even when the tree being dropped on is remote. The destination side goes through `Host`, and the one `std::fs::copy` that touches a destination sits inside a branch already gated on `host.id().is_local()`. While adding that entry: `attr` starts unset, which awk reads as 0, so a file whose first line is `mod something` matched `attr == NR - 1` and cut its body at line 0. `head -n -1` then errored and the file was scanned as empty — `src/terminal/mod.rs` and `src/ui/tray/mod.rs` both open that way, and the guard had been blind to both. Neither contains a violation, so seeing them is free.	2026-08-10 12:41:17 +08:00
l0ng-ai	58d7ef5838	fix(scm): close out the review's minor findings across the data and UI layers ... The second pass over the branch review: every remaining finding verified against the code, the real ones fixed. Data layer: - A truncated log parse is never called complete: RecordSplitter drops an overlong record whole and reports the count (delivered cut short, a commit body cut mid-way reads as the real message), parse_log carries a truncated flag past MAX_LOG_BYTES, and load_page only says "end of history" when the parse read everything git returned. - Every scope pins symbolic revs to shas before walking, so a commit landing between two pages can no longer shift where page two starts under Head and Refs scopes; unresolvable names read as "no history" rather than as a load failure. --parents was doing nothing and is gone; edge sort is stable so a merge's Outs keep first-parent order. - The lane model's central invariant now names the join case — a merge whose second parent already has a lane reserved sends its Out onto that lane, one line below the cut, not two — with a golden test for the commonest merge topology of all, which no golden covered. - DiffSource revs get the same could-be-an-option guard log already had; C-quoted paths decode the full escape set (a tab decoded to a literal t broke the :(literal) re-probe); rename from/to lines override the ambiguous diff --git header; combined-diff line numbers follow the sides rather than the colour, so a " +" line no longer drifts every number below it. - A rename's old path stays out of the per-file decoration map, where it outranked a file re-created at that path; ignored records decorate as Ignored, not Modified; checkout <branch> gains the trailing -- that keeps a stale name from falling back to a worktree-clobbering path checkout; unstage before the first commit takes -f (worktree- safe with --cached); batches split by bytes as well as count for Windows' 32K command line; a deadline expiry reports Timeout, not "git could not be run"; error details keep both streams. - probe_status distinguishes "not a repository" from "could not ask": a dropped link keeps the cached status (stale beats blank) and rests 10s instead of erasing the panel, while a definitive not-a-repo also drops the cwd→root mappings so the panel stops drawing Loading for a repository that is gone. Probe and watch work are wrapped against panics that would wedge their in-flight bookkeeping forever, watch landings check the wipe counter, superseded probes relaunch through the debounce, and a refused network slot says so instead of eating the click. UI: - Reset --hard confirms with its own words (commits fall off the branch), not the discard dialog's; a merge commit whose prefilled message the user cleared is committable again; the disabled commit button distinguishes "nothing to commit" from "write a message". - Selection highlight matches on the diff source too, so a file staged and edited again no longer lights both of its rows for one overlay. - The graph materializes only the rows in the viewport window (5000 flex children per frame was most of a frame), row clicks carry the page Arc and an index instead of a deep Commit clone per row per frame, filter results are cached per (page, query), and a selected merge ring's hole matches the selection band under it. - A failed commit_files read says the list could not be read instead of "0 files changed"; the STAGED chip and the graph's relative times go through the i18n table; the keys-awaiting-a-caller list is pruned to the seven that still are; the orphaned PanelUntracked key is gone; the zh commit placeholder reads naturally. 2398 tests, 0 failures. Known flake: daemon::singleton's second-claim test, untouched by this branch, fails ~1 in 3 full parallel runs and passes alone.	2026-08-10 12:31:58 +08:00
l0ng-aiandl0ng-ai	5f1ee966ec	fix(windows): give the pane grip and drags in flight a cursor (#455) ... Win32 ships neither an open- nor a closed-hand cursor, and gpui's Windows backend answers both with the plain arrow. The pane drag grip asked for `cursor_grab()` and so read as ordinary background there, and the pointing hand the sidebar's group header had worked around it with was dropped again the moment a drag began, since the active drag cursor is `ClosedHand`. Lift that workaround into `reorder::cursor_grab` so the grip and the group header share one answer, and pick the held cursor per platform too. Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-10 12:12:09 +08:00
l0ng-ai	a764d92132	fix(scm): sequence compound verbs, cap graph paging, back off failed loads ... Review findings on this branch, all in the seams between async operations: - Commit-and-push, commit-and-sync, sync and discard-all dispatched both halves into the worker pool at once, so a push could resolve the branch tip before the commit (or pull) it was waiting for and quietly send the old one. Compound verbs now carry a ScmFollowUp that the first half's landing closure starts on success only; a refused commit, a failed pull or a cancelled confirmation drops the follow-up with it. - Push sent `git push <remote> <branch>` with the branch taken from the upstream's name — a bare name means a *local* branch, so `feat` tracking `origin/main` pushed stale local `main`. The refspec is now `HEAD:<branch>`, and the branch is validated with the full branch check since a `:` would smuggle a second refspec in. - `scm.committing` was armed before the amend confirmation and never disarmed on failure, so a cancelled prompt (or a hook rejection) plus any later unrelated HEAD move cleared a message that was never committed. It is armed at dispatch and disarmed when the commit errors. - Discard-all fed staged-only paths to `checkout --`, where a staged deletion sank the whole batch as an unmatched pathspec. Only unstaged paths go in, one confirmation covers both halves, and the two gits no longer run concurrently. - One "load more" click at 5000 commits grew `requested` past what `load_page` clamps to, so the freshness check never passed again and every frame refetched the full page. Growth stops at the cap, the button hides there, and a failing `git log` is remembered per key instead of being retried from every render. - A repository switch now drops the previous repository's page before anything can draw it or grow from it — a stale row's context menu used to build ops for the new repo with the old repo's rev. - A watch that failed to open was retried at frame rate, one host round trip per render; it now rests for WATCH_RETRY between attempts. - Non-network writes on a remote host ran under the interactive 20-second deadline while the server ran the job to completion, so a slow pre-commit hook was reported failed and then landed anyway. Every write now goes through git_with_deadline, 120s for local verbs.	2026-08-10 11:24:09 +08:00
l0ng-aiandl0ng-ai	30b16c65b5	fix(settings): keep one restart button for the stale background server (#452) ... The in-place-update notice carried its own Restart server button while the Server section right below it carried an identical one, both calling restart_daemon. Move the notice into the Server section: the stale build line sits under the header and its explanation replaces the generic one, so the single button that ends every running pane is the only one on the page. Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-10 11:22:11 +08:00
l0ng-aiandl0ng-ai	88bf9a5da5	feat(daemon): upgrade in place, keep pane screens across a crash, and give panes their own history (#449) ... * feat(daemon): keep a pane's screen across a death nobody chose A daemon that crashes, is `kill -9`'d, or goes down with the machine takes every pane's replay ring with it, and the window comes back to a row of blank shells. The processes cannot be saved that way — nothing written to a file brings a process back — but the picture can. The daemon now keeps a capped tail of each pane's ring under the config directory, and a client whose `Attach` found nothing can ask, on the `Spawn` that replaces it, for the dead pane's screen. The new pane opens showing it, under a rule that says the shell below is new. - periodic and dirty-only: a ring that has not moved is not rewritten, so an idle machine does no IO at all. Write-through would be an enormous amount of write amplification for a few seconds of freshness. - capped at 256 KiB per pane, far below the ring's 8 MiB: the value of scrollback decays with distance from the bottom, and every byte here is a byte of someone's terminal on disk. - off by default. The ring holds whatever the pane printed, including echoed tokens, `env` output and agent transcripts; in memory that dies with the daemon, and writing it down is the whole feature and the whole cost. Files are 0600, and turning the setting off deletes what was kept. - dropped by relevance, not by calendar: a pane the user closed, or one no workspace names any more, has its file removed on the next sweep. Restored bytes are replayed at the geometry they were written at, and are preceded by resets — leave the alternate screen, show the cursor, restore autowrap, clear SGR — because a snapshot is cut at the front and can begin in the middle of any of them. * feat(daemon): upgrade the daemon in place instead of killing every shell Picking up a new build meant stopping the daemon, and stopping the daemon means every pane dies: the pty master is a descriptor this process holds, so when the process goes the slave side raises SIGHUP and takes the shell, the agent and the half-finished command with it. That is why the update path leaves the old daemon serving and Settings has to offer the restart as a thing you schedule for a quiet moment. `execve` does not have that problem. It replaces the image and keeps the process: same pid, same children, same descriptors, same file locks. The daemon now rewrites itself that way on `ClientMsg::Handoff` — it writes what it knows about each pane into a blob, clears FD_CLOEXEC on the pty masters, the blob and the singleton lock, and execs the new binary, which picks the panes back up on the other side. - **the seat travels on the command line, not in the blob.** The lock is still held by this process, so the new image must adopt the descriptor rather than ask for the lock again — asking would be refused by its own lock and it would stand down in favour of itself. A daemon that loses its panes is a bad afternoon; a daemon that exits leaves the machine with nothing serving, so that one fact has to survive an unreadable blob. - **the blob is unlinked before it is written.** It holds every pane's ring, which is the output `scrollback` makes people opt into storing; a handoff must not be a back door for writing it to disk. - **the exec is the last step.** Everything is staged first, so any failure before it costs a log line and the daemon carries on serving — which is what lets callers treat a failed handoff as "fall back to a restart" without having lost anything on the way. Native SSH panes cannot cross — their session is cipher state in memory, not a descriptor — so they are hung up first and the far end sees a clean close. Windows has neither execve nor a transferable ConPTY handle, so it keeps the stop/start path; the dialogs there still promise what they always did, and the new copy is shown only where it is true. Also retries flock on EINTR: a signal landing mid-call said nothing about the lock, but was reported as "could not be evaluated", which starts a second daemon beside the first — the split machine singleton exists to prevent. The end-to-end test sets a variable in the shell, hands over, and reads it back. Nothing but the original process can answer that, and the daemon's instance id changing while its pid does not is what says an exec really happened. * feat(shell): give each pane its own history when asked Two panes running zsh with `share_history` are appending to one file and reading each other's lines back, which is either the feature or the problem depending on what the panes are for. Someone with a pane per task wants Up to walk that task's commands, not an interleaving of four. Each pane can now have its own history file instead. It is seeded from the shell's real history, so a new pane is not blank, and what the pane added is appended back when it closes, so nothing typed is lost — a per-pane history that evaporated would be a way of losing commands, not of organising them. The seeding is done by the shell, not the daemon, and that is the only reason it works: `HISTFILE` belongs to the user's rc file and can point anywhere, long after the pane's environment was decided. tty7's snippet is appended to the rc it wraps, so it runs after that decision and is the one place the real path is known — it copies the tail, records how much it copied, and repoints. Both shells load history after their startup files, so the switch lands before the first line is read. The daemon's half is a filename, a rename when a restored pane inherits its predecessor's file, a merge on close, and a sweep for the panes a killed daemon never got to retire. Off by default: shared history is what a terminal has always done, and someone who did not ask for the change would experience it as their history mysteriously forgetting the other window. bash and zsh only — fish and PowerShell do not keep a HISTFILE, and a shell launched with the user's own arguments gets no snippet to repoint anything in. * fix(daemon): store pane screens on the shutdown a restart actually uses The periodic writer covers a death nobody prepares for and the SIGTERM path covers a signal, but the restart the app itself performs goes through ClientMsg::Shutdown — which killed every pty without taking a copy first. That is the one shutdown where the panes are expected back. * fix(daemon): leave nothing dangerous behind when a handoff fails or lands Review findings on the in-place upgrade and per-pane history: - A failed exec now puts back everything it had staged: FD_CLOEXEC on the seat and every pty master (a child inheriting the seat keeps the flock held past the daemon's death, so no future daemon could seat itself), and the SIGPIPE disposition plus this thread's signal mask, both of which Command::exec resets on its way to the attempt — without this, the still-serving daemon dies on the first client that hangs up mid-write. - The adopting image restores close-on-exec on the seat and on every adopted master, so children it spawns later cannot hold a pty open past its pane, or the seat past the daemon. - The target binary is checked before the handoff gives anything up: native-SSH panes are hung up on the promise that this process is about to be replaced, and an exec that was never going to work must not collect on it. - The integration snippets raise HISTSIZE/HISTFILESIZE (bash) and SAVEHIST/HISTSIZE (zsh) for the pane's private history file. At their defaults the exit rewrite truncates the file below its own seed mark, which the merge-back rightly reads as "replaced under us" — silently losing the pane's commands for anyone with more history than the caps. - The restart dialog's promise now binds the action: where the copy said "nothing is interrupted", a failed handoff is reported instead of silently traded for the restart that kills every pane. - The scrollback writer checks the ring's mark before cloning it, so an idle pane no longer costs a full ring copy under the state lock every tick. Each behavioural fix carries a test that fails without it; the history truncation one was verified to fail with the snippet change removed. --------- Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-10 10:59:02 +08:00
l0ng-ai	44f0683d0a	fix(sidebar): cut labels on grapheme clusters, not on chars (#450) ... The sidebar's elision measures against real glyph widths but slices by `char`, so it can satisfy every width check and still hand back a torn cluster. Scanning budgets from 30px to 200px over emoji fixtures, 48 widths produced output no font can render as intended: "release-…\u{200d}👩\u{200d}👧" a joiner with nothing in front of it "lon…\u{fe0f}" the variation selector lands on the ellipsis "abcdef…🇳ghijklmnopqr" half a flag, which renders as a bare N A tab title carrying an emoji is not exotic — plenty of TUIs and coding agents put one there — and the second case is the same U+FE0F this repo already carries an alacritty patch for. `elide_keep_edges`, the tail-only fallback, and `short_title`'s 40-glyph clamp now index grapheme clusters. `elide_path_keep_tail` cuts on `/` and was already safe. Widths are unchanged: clusters are measured the same way chars were, so every existing elision test still passes on the same fixtures. `unicode-segmentation` is already in the tree via gpui; pinning it here adds one line to Cargo.lock and no new code. Tests assert the property rather than the symptom: whatever survives on either side of the ellipsis has to be a cluster-aligned prefix and suffix of the input. That catches any tear, not just the three shapes found here. Written first, confirmed failing on all three cut sites, and green after.	2026-08-10 10:38:04 +08:00
ARNOandl0ng-ai	1fd99ca7c6	feat(sidebar): tail-first label elision and a hover card for hidden details (#446) ... * feat(sidebar): tail-first label elision and a hover card for hidden details ff * fix(sidebar): keep the head of a non-path label, and let the card expand what the row hid The elision landed with four gaps between what a row hides and what the hover card can give back. A renamed tab is elided like anything else, but `sidebar_info` still carried the old tooltip's guard and bailed out on any non-empty `tab.name`, so a long custom name was shortened by the row and the card refused to spell it out. An empty terminal title took the opposite path: the row falls back to `Shell 3`, the card compared that placeholder against the empty string it came from, found them different, and opened on a row that had hidden nothing. Both came from deriving the same strings twice; the row now hands `sidebar_info` what it rendered next to what it rendered it from, and the card is decided by comparison alone. `elide_path_keep_tail` was applied to every title, but a title is not always a path — `npm run dev`, or a name someone typed. Dropping the head of those says less than the truncation this replaced, so `elide_label` picks the rule: tail-first for a path, both edges otherwise. `elide_keep_edges` promised both ends and delivered neither on a token with no break in reach: the head ran to the 12-glyph cap, `head…` alone overran the budget, and it fell through to a bare tail. It now tries shorter heads before giving up, and never trades the whole tail away for a longer head (`feature/…` became `fea…thing`). Also: measure the active row at the MEDIUM weight it actually renders at; subtract the list's own padding from the text budget, so a label that "just fits" is not handed back to CSS truncation; rejoin a path with the separator it arrived with, instead of spelling one tab `C:\Users\dev\app` while it fits and `C:/…/app` once it does not; count the gap between the two diff counts rather than the space standing in for it; let the card wrap instead of truncating the one string it promised in full; and read the remote host off the same leaf the title came from. The elision tests shape through gpui's `NoopTextSystem`, where every glyph is one em — deterministic across the three CI targets, but blind to the proportional and CJK widths this exists for. Said so where the fixtures are built, rather than implying the pixels are real. --------- Co-authored-by: l0ng-ai <ysdpk123@gmail.com>	2026-08-09 22:13:01 +08:00
l0ng-aiandl0ng-ai	86799220ca	feat(splits): rearrange a tab's panes by dragging one onto the layout (#445) ... * feat(splits): rearrange a tab's panes by dragging one onto the layout Hovering a pane floats a small grip along its top edge; dragging it picks the pane up and puts it somewhere else in the same tab. Three landings, resolved from where the pointer is: * a pane's edge — split that pane and take the side dropped on * a pane's middle — trade the two panes' places * the band along the outside of the tab — sit beside everything else as a full-width or full-height band, which is the only way to say "make this a full-height column" in one gesture from the middle of a 2x2 The landing is highlighted while the drag is in flight, and is offered only once the tree agrees the drop changes something, so the highlight is never a promise the drop does not keep. * pane: move_leaf / move_leaf_to_edge / swap_leaves, each built on a clone and installed only when the layout really differs * pane_drag: the pointer-to-landing geometry, the drag state, and the grip * tree_sync: reconcile a tab that kept its panes but changed shape with a single PaneMove instead of closing and rebuilding the tab * feat(splits): drop a pane beside its neighbours, not on top of one Trying the drag out on real layouts turned up three ways the drop model asked for more precision than it should have. A drop on a pane's side always halved that pane, so putting a new column into a row of columns was only reachable at the very edge of the window, where the band rule took over. A side facing a neighbour in the same row or column now joins that run: the newcomer takes an equal share and the others give it up in proportion, keeping whatever relative sizes they were dragged to. A side facing across the run has no run to join and still halves the pane it landed on. The band along the tab's edge was a flat 26px, which on any real window is a hair's breadth. It is now measured against the pane it is read in — a sixth of it, floored at 32px and capped at 120 — and only counts on a side that faces the window rather than another pane. Landing there takes an even share of the columns that side already has instead of half the tab, so a third column is a third and not a half. The highlight is no longer drawn from the rule. The drop is carried out on a deep copy and the dragged pane's new rectangle is measured off it, so the preview and the result cannot disagree; the copy is deep because sharing a run out writes ratios the live tree's splits hold in common. Also: the grip is a quiet 22x3 bar that grows to 40x5 under a fixed 56x10 target (it needs an id of its own, or gpui settles its size before the group-hover is known), and every rearrangeable pane keeps an 8px strip clear above its grid so the grip never sits on the first row. * fix(splits): pin a drop to the pane it was offered against Review follow-ups on the pane drag. A drop zone named its target by position in the tab's leaf order, but it is read on one frame and carried out on the next: a pane closing in between shifts every index after it, and the drop lands beside a pane the user never aimed at. The zone now carries the target itself once the frame that drew it has resolved it, so a target that has gone refuses the drop instead of sliding it sideways. Alongside it: * `Pane` is no longer `Clone`. The two copies it can be asked for differ in whether they share their splits' sizes, which is not a difference to leave to whichever one `.clone()` happens to mean; `shallow_clone` is now named and private, next to `deep_clone`. * `edge_landing` no longer hands back a share that only a test read. The test reads it off the split the landing produced instead, which is the number the drop actually lands. * A test pins the invariant the drop zones rest on: `leaf_rects` comes back in the order `leaves` does. * Drop a doc comment that had landed on `close_focused` describing a different method, and an `Option` in `drop_pane` that was wrapped only to be unwrapped two lines later. * The changelog claimed every rearranged tab now syncs as one `PaneMove`. Only a drop beside a single pane does; a drop beside a whole group is not something `PaneMove` can name, and still takes the rebuild. Both entries move under `Unreleased` — v26.8.2 was tagged before either landed. --------- Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-09 22:04:24 +08:00
l0ng-ai	c2b6ba2ec0	Merge pull request #442 from l0ng-ai/fix/issues-426-430 ... fix: small-window Settings layout, vim caret on a raw pty, wheel zoom, theme preview and cross-locale palette search	2026-08-09 20:39:16 +08:00
l0ng-ai	2f31a11df0	fix(settings): size the nav floor for the longest label in any locale ... SidebarMenuItem clips its label rather than eliding it, so a 140pt floor that fits English cut a glyph in half elsewhere: zh-CN lost the right half of the last character of 窗口与标签页. Size the floor for ja-JP ウィンドウとタブ, the widest of the three, which costs the page 35pt at the narrowest window and keeps every nav label whole.	2026-08-09 19:20:15 +07:00
l0ng-ai	3dcffe6f4e	fix(windows): grow a remembered bound back up to the minimum size ... window_min_size governs what a drag may do to a window, not the bounds it opens with, so a remembered bound walked straight under the declared 720pt minimum — the reported settings window measured 641. Clamp the restored size on the way in, keeping the origin.	2026-08-09 19:20:15 +07:00
l0ng-aiandl0ng-ai	c13efbce32	fix(terminal): draw a multiline history entry on one menu row (#436) ... A command composed in the inline editor keeps its newlines when `submit_command` pushes it into the in-session history, even though `history::append` and the zsh/bash/fish parsers all refuse to carry an entry across lines. gpui breaks text on `\n` whatever `white_space` says, so one such entry painted a dozen lines inside a one-line row and covered the whole reverse-search menu. Fold line breaks to a visible `↵` when drawing — one char for one char, so the fuzzy matcher's highlight positions still line up — and leave the stored entry untouched, so recalling and running it are unchanged. Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-09 20:19:53 +08:00
l0ng-aiandl0ng-ai	8dbc7efa1a	fix(cli): stamp panes with the workspace that holds them, not the client's name (#425) ... * fix(cli): stamp panes with the workspace that holds them, not the client's name A pane's owner names the workspace allowed to attach to it. The CLI wrote a literal "tty7-cli" there for every pane it made, so a window opening on a CLI-built workspace found none of them attachable: it spawned a fresh shell for each tab, orphaned the live ones, and — because the tree still carried each pane's agent session — greeted the user with a failing `claude --resume <id>` in every one of them. Both spawn paths now pass the workspace id, and restore treats an owner that parses as no workspace as no claim at all, so panes already stamped by an older CLI attach instead of stranding. * fix(cli): let the OWNER column speak only when it disagrees with WS Now that a pane's owner is the id of the workspace holding it, printing both spells the same id twice on every row of `pane ls --all` — and buries the rows that matter. The column now shows a dash when the two agree, so what is left is exactly what is worth reading: a pane its holder may not attach to, and an orphan still naming where it belongs. --------- Co-authored-by: l0ng-ai <24760907+l0ng-ai@users.noreply.github.com>	2026-08-09 20:08:03 +08:00
l0ng-ai	150f6ff76f	fix(settings): stop the page being the only column that gives width back ... The settings nav, the SSH host list and the theme panel were fixed widths that never yielded, so the page they frame absorbed every shortfall. In a 641pt window — the one the report came from — that ran all the way down. On SSH, 220 of nav and 280 of host list left the detail panel 141pt and its empty state painted a couple of hundred points past the right edge of the window. On Appearance with the theme panel open the page got about 125pt, and a Chinese description came out one character per line. The columns are now allocated against the window instead of asserted: each list is handed its full width, then gives back a share of whatever is missing until the page reaches 420, and no list goes below the width at which it stops being itself. Below the width where the nav, the panel and a readable page cannot all fit in one row, the theme panel stops being a column and lays itself over the page — it is a temporary layer over one choice, and Escape already closed it first. The floor the page keeps is derived rather than picked: it is what the narrowest window in the wild leaves the SSH page, the one that spends a second list, once both lists stand on their own floors. It is a target for the allocator and not a `min_w` — a floor a flex row cannot honour does not push its siblings back, it overflows, and overflow here means content painted off the window, which is the failure being fixed. What makes the floor liveable instead is that the wide controls can now shrink into it. The thresholds that decide when a row stacks are widths a label needs, so they follow the interface font size — at 24pt every label is half as wide again while the slider beside it is still 240px. The rows that were hand-rolled rather than built by `settings_row` get the same treatment: the keybinding preset and prefix rows stack at the same width, every binding row lets its label wrap and its key caps wrap to a second line, the theme card drops its preview and stops pushing "change theme" off the card, the SSH quick-connect field shrinks instead of running past the pane it sits in, and a port-forwarding rule takes two lines — or three, at the width the report came from — rather than one that does not fit.	2026-08-09 18:48:06 +07:00
l0ng-ai	14c092b0d1	feat(palette): preview themes live while the picker is open ... Picking a theme from the command palette closed it, so finding out what a theme looks like meant reopening the palette and retyping the search for every single one. The theme picker now applies whatever row is highlighted straight to the running window and stays open: Return persists the pick, Escape or any other way of closing the palette puts the previous theme back. A preview only touches the in-memory config, so arrowing through the list never writes config.json. The picker opens on the theme already in use, so opening it changes nothing by itself.	2026-08-09 18:39:31 +07:00
l0ng-ai	54960be6bf	feat(palette): match commands in every locale, not just the shown one ... The palette matched a query against the label it was rendering and nothing else, so a window running in Chinese answered "no matching commands" to `theme` — the English wording of that row did not exist anywhere the filter could see it. Each entry built from a locale key now carries the same key as every other locale words it, plus the stable command id, as hidden search aliases. They are built once with the entry, cost nothing per keystroke, and are never rendered: the row keeps showing its localized label, and an alias hit scores just below the same hit on that label so a visible match still comes first.	2026-08-09 18:39:31 +07:00
l0ng-ai	68d6b4b062	feat(terminal): zoom the font with the platform modifier and the wheel ... Holding Cmd (Ctrl off macOS) and scrolling over a terminal now resizes the font instead of the scrollback, which is what you reach for when showing a pane to someone else. A wheel detent is one step whatever the platform bills it as, and a trackpad accumulates until the fingers have travelled three lines, so a flick does not run the font end to end. Steps go out as the existing IncreaseFontSize/DecreaseFontSize actions, so the clamp and the saved setting stay in one place.	2026-08-09 18:39:31 +07:00
l0ng-ai	44d33031e1	fix(terminal): stop repairing a parked cursor on a raw pty (#430) ... Typing `:wq!` in vim wrote it onto the row being edited instead of the command line, and left the cursor there. The command line is not what moved: the cursor is. `ParkedCursorRepair` (#362) exists for conhost, which brackets every frame it paints with `?25l` … `?25h` and, on the frames where it did not paint the cursor, commits the show wherever the last erase or write left it. The repair pairs the hide with its show, calls the show parked when the run moved the cursor around to paint but did not end on a move, and puts the cursor back on the cell the hide caught it on. The reader ran it on every platform. Off Windows there is no conhost in between and the application owns the cursor. Captured from vim 9.1 on a raw macOS pty, opening the command line is \x1b[?25l \x1b[11;10H: \x1b[1;1H \x1b[11;1H \x1b[K \x1b[11;1H: \x1b[?25h — a run full of moves that ends on the `:` it wrote, which is exactly the shape the scanner calls parked. So the repair dragged the cursor off the command line and back onto the file, and vim, which echoes the following keystrokes as bare bytes with no positioning of their own, wrote `wq!` over the text. Gate the repair on `cfg!(windows)`, the way `conpty_resize` already is: the artifact is ConPTY's, and on a raw pty the cell a frame leaves the cursor on is the cell the application meant. This also keeps the scanner off the client's output path entirely on Unix, where it used to walk every batch. A macOS or Linux client attached to a *remote* Windows daemon loses the repair with it — the same limit `conpty_resize` has — which costs a stray caret there and buys back an unshredded screen on every local pane.	2026-08-09 18:39:31 +07:00
l0ng-ai	62b922f2c2	Merge origin/main into integration/polish ... main dropped the client-side command-mark store (#404) while this branch had just started reading it: the close confirmation names the command it is about to end, and the mark was the only place that text existed on the client. Keep both. The OSC 133 tokenizer main left in place already sees every mark, so the command line now rides alongside `zle_reading` and `shell_vi_mode` as one shared string — set on `C`, cleared on `B` and on a `C` that carries no line — instead of a store with a list, a lock and a cap. `busy()` reads that. The rest: - settings.rs takes main's opaque overlay surface and background layers, keeping this branch's no-match note and scrolled body. The inner `.bg()` goes, per main's reason: the root already paints it, and a second fill hides the theme image. - i18n keeps this branch's `every_key_is_translated_in_every_locale`, which walks `L10nKey::ALL` in all three locales, over main's hand-listed zh coverage test it replaced. It immediately caught three of main's new backdrop keys reading English in ja — Mica, Mica Alt and Acrylic, which is what Japanese Windows calls them, so they join the allowlist with that reason. - app.rs keeps both sides' tests and drops both sides' now-dead imports: `window_background` (main deleted the function) and `humanize_action` (this branch's keybinding note uses `keymap::action_entry` instead). Verified: `sleep 300` then ⌘W asks about "sleep 300"; ⌘W after it ends closes without asking.	2026-08-09 16:38:15 +07:00
l0ng-ai	5c65e2b08f	test(git): pin the git config the fixtures assume, and one path spelling ... Two Windows-only failures, both of them the tests asserting on the runner's git rather than on the code. `core.autocrlf` is `true` by default in Git for Windows, so a file written as `one\n`, committed, and restored by `checkout --` comes back as `one\r\n`. The `-c` list the helpers already pass would not have fixed it: the checkout in that round trip is `run_op`, production code running its own git with no overrides. So the pins go into `<repo>/.git/config` right after `init`, where repository config outranks the system config that carries the default. The other is the same path-spelling mismatch fixed earlier in `git_data.rs`: `git rev-parse` answers with forward slashes and no extended-length prefix even on Windows, while `fs::canonicalize` returns `\\?\C:\…`. Both name the same directory and the Win32 APIs take either, so the production path is right and only the comparison needs one spelling. Fixed as classes rather than as instances. `core/git/mod.rs` gains a `test_support` module holding the `-c` list, the repo-config pin and the path normaliser, and `status.rs`, `log.rs`, `ops.rs` and `diff.rs` all read from it — `ops.rs` had no config pins at all and `diff.rs` was missing gpgsign. `git_data.rs` keeps its own copy because a `#[cfg(test)]` item does not exist in the `tty7-core` the binary crate links against; a comment says so and points at the other copy. The normaliser has its own test over literal `\\?\C:\…`, `C:\…`, git's `C:/…` and a unix path, and the line-ending fix was reproduced locally by pointing `GIT_CONFIG_SYSTEM` at a config with `core.autocrlf = true`: that panics exactly as CI did with the pins reverted, and passes with them. One latent hazard hardened while here — `status.rs`'s scratch directory had no pid in its name, unlike its sibling, so a leftover that resisted removal would have been silently reused as a fixture.	2026-08-09 17:02:29 +08:00
l0ng-ai	0106430ecd	merge: main into the Source Control branch ... The one conflict is an import list in `diff_overlay.rs`: this branch added `SharedString` for the unified view's row labels, main added `Background` and `Hsla` for the window backdrop work. Both sides are still used, so the resolution is the union. Worth recording why this merge happened when it did. `main` moving is not normally urgent — branch protection dropped its strict check, so a branch behind main still merges — but a *conflicting* branch is different: GitHub cannot compute `refs/pull/N/merge`, and every workflow that triggers on `pull_request` silently stops running. Three pushes in a row registered no CI at all on #424 while other PRs kept going green, which reads as a GitHub incident and is really just an unresolved conflict.	2026-08-09 16:20:19 +08:00
l0ng-ai	d0d5e149c7	fix(git): compare watch directories in one spelling of a path ... `scm_watch_dirs` hands back what `git rev-parse` answered, and git writes forward slashes and no extended-length prefix even on Windows. The test's expected side is built from `fs::canonicalize`, which on Windows returns `\\?\C:\…` — so the two named the same directory and compared unequal, and the Windows job failed on a path the watcher would have been perfectly happy with. Both spellings reach the same directory through the Win32 file APIs, so the watcher is right to pass git's answer straight to `Host::watch` and nothing changes outside the test. The five assertions now go through a `one_spelling` helper and keep their exact-equality teeth; on unix it is a no-op, which is why this was invisible until Windows CI ran the branch for the first time.	2026-08-09 16:08:46 +08:00