[pattobrien]

Summary

Merges upstream herdrdev/herdr master at 3d9d2b18 (tags v0.8.2, v0.9.0, v0.9.1, v0.9.2, v0.9.3; 513 commits since the fork point 346411fa = v0.8.0) into patt/perf-graphics, then re-expresses the fork features that upstream's refactors broke.

patt/perf-graphics (134c42fb, v0.8.0-graphics.4)
└── cf614e08  merge upstream/master 3d9d2b18        # 15 conflicted paths resolved
    ├── cab7a8a1  perf: input lag while kitty graphics panes stream   # re-expressed on #4561/#4657 code
    ├── ebdf55fd  feat: client.sidebar.toggle api + herdr terminal sidebar toggle   # rewritten on the client.window_title pattern
    └── cca140a3  feat: mirror pane OSC 22 pointer shapes on the host cursor   # host half re-homed in the client shell

Upstream highlights in the range: the Herdr-specific pane.graphics.* API is gone in favour of native Kitty rendering (herdrdev#4561); shell/sidebar rendering and hit-testing moved into the client (herdrdev#3487); multi-client tab views (herdrdev#3526); SGR-pixels mouse reporting fixed (herdrdev#3295, v0.9.1); preview release channel; libghostty-vt extracted into crates/ghostty-vt (herdrdev#4661) and the vendored source now needs Zig 0.16.0.

Audit of fork-only commits

Fork commit	Topic	Class	Upstream citation	Action
e3aa7611 perf: fix second-long input lag while kitty graphics panes stream	graphics/perf	(c) still fork-only	No o=z upload compression upstream (flate2 only transitive); the one attempt, herdrdev#3934, was closed by the contributor gate. herdrdev#4561 added temp-file delivery but gates it to TERM_PROGRAM=ghostty; Kitty hosts still get inline base64. Unix client still shares one 256-slot channel for stdin and server frames (src/client/mod.rs:514); upstream fixed the same backpressure only on Windows. Related open issues herdrdev#3877, herdrdev#3783.	Re-expressed as cab7a8a1: write_kitty_upload in src/kitty_graphics.rs used by the inline branch of src/client/frame_output.rs; dedicated 2-slot server-frame channel consumed by the last arm of the biased select in src/client/events.rs::next_loop_event.
dd536e00 bench: add input-latency harness and v0.8.0 vs fork results	graphics/perf	(c) keep	No bench/ upstream; harness drives standard kitty escapes through the PTY (path herdrdev#4561 kept); experimental.kitty_graphics, experimental.allow_nested and the render.prof log line still exist.	Kept as-is. results-*.json and the README table are stale vs 0.9.3.
e930c6a1 fix: tolerate pre-release suffix when parsing versions	releases	(c) keep, drop candidate	Version::parse in src/update.rs is byte-identical upstream; Version::current() still panics on X.Y.Z-suffix. Upstream's supported path is HERDR_BUILD_CHANNEL/HERDR_BUILD_ID at build time (build_info::version(), .github/workflows/preview.yml:163).	Kept (clean auto-merge). Drop it if fork builds switch to HERDR_BUILD_CHANNEL=graphics HERDR_BUILD_ID=N with a plain Cargo.toml version.
fa653722 40a856ec 0f0aa005 134c42fb release: v0.8.0-graphics.1..4	releases	(d) superseded	Upstream Cargo.toml is 0.9.3.	Merge carries 0.9.3; the next fork release picks its own string.
21511f34 fix: report real pixel coordinates to SGR-pixels mouse panes	mouse coords	(a) resolved upstream	Issue herdrdev#3295 closed 2026-09-14, released v0.9.1: 1777e9b herdrdev#2523 (Position on the encoder), 3a822e8 herdrdev#3637 (MOUSE_FORMAT_SGR_PIXELS), 9631f8e herdrdev#4045, b0c05e6 herdrdev#4140. src/ghostty/mod.rs became crates/ghostty-vt/src/lib.rs (fff6c82 herdrdev#4661). Upstream test ghostty_mouse_sgr_pixels_preserves_exact_and_downgrades_cell_input asserts the cell-input downgrade the fork had replaced.	Dropped: upstream taken for src/pane/input.rs, the mouse hunks of src/pane/terminal.rs, and crates/ghostty-vt/src/lib.rs (git had carried cell_size_px across the rename).
ae7575eb feat: mirror pane OSC 22 pointer shapes on the host cursor	pointer shapes	(c) still fork-only	No OSC 22 / pointer-shape handling upstream (grep, log, gh search empty; herdrdev#116/herdrdev#117 are DECSCUSR text-cursor). Host half broken by 207be3c herdrdev#3487 (hit-testing and Mode moved to the client shell; src/app/input/mouse.rs deleted) and the frozen ServerMessage variant order.	Pane-side tracker kept in the merge; host half re-homed as cca140a3 via an endpoint.pane-pointer-shape.v1 EndpointControl message and a client-shell hover resolver (src/client/shell/pointer_shape.rs). PROTOCOL_VERSION untouched.
aef49f7d feat: add ui.sidebar.toggle api method and herdr sidebar toggle cli	sidebar toggle	(d) obsolete as written	sidebar_collapsed lives only in src/client/shell/* since 207be3c herdrdev#3487, so the fork's server-side flip no longer compiles. No ui.* namespace upstream; command.invoke only resolves custom commands. herdrdev#1623 closed NOT_PLANNED, discussion herdrdev#1624 open. docs/next/api/herdr-api.schema.json is generated (generated_protocol_schema_artifact_is_current).	Dropped from the merge; rewritten as ebdf55fd on the client.window_title.* intercept pattern: method client.sidebar.toggle, CLI herdr terminal sidebar toggle (what scripts/herdr-worktree-menu already calls), pushed to the foreground client as endpoint.sidebar-toggle.v1.
opentui_256_palette_query_burst_uses_host_snapshot (src/pane/terminal.rs)	palette queries	not a fork change	Upstream 503613c herdrdev#1759, already in the fork base; byte-identical at 346411fa and upstream/master. The only palette-adjacent upstream change, f83980d herdrdev#2162 (OSC 4 overrides), does not touch the query-burst path.	Upstream version intact.

Conflicts resolved

15 conflicted paths in cf614e08:

• Cargo.toml, Cargo.lock: upstream (0.9.3, workspace with crates/ghostty-vt).
• src/pane/input.rs, src/kitty_graphics.rs, src/client/mod.rs, src/server/headless.rs, src/protocol/wire.rs, src/api/schema.rs, src/api/schema/response.rs, src/api/server.rs, docs/next/api/herdr-api.schema.json: upstream; the fork hunks were either superseded or re-expressed in the three follow-up commits.
• src/app/input/mouse.rs: deleted (upstream removed it).
• src/events.rs, src/app/actions.rs: kept PanePointerShape, dropped PrefixInputSource (removed upstream).
• src/pane/terminal.rs: upstream for the mouse-encoder region and the renamed tests; kept process_pty_bytes_surfaces_pointer_shape_changes; removed the fork's two SGR-pixel tests.
• Auto-merged hunks reverted to upstream because their commits were dropped: src/app/api.rs, src/cli.rs, src/app/state.rs, src/app/mod.rs, crates/ghostty-vt/src/lib.rs.
• Cherry-picking cca140a3 after ebdf55fd conflicted additively in src/client/endpoint/control.rs, src/protocol/endpoint.rs, src/client/mod.rs (both added an endpoint-control kind, decoder arm and client handler); both sides kept.

Evidence

• Before: patt/perf-graphics at 134c42fb builds against v0.8.0; git merge upstream/master stops with 15 conflicted paths, and the fork's SGR-pixels and ui.sidebar.toggle hunks no longer compile against the client-rendered shell.
  After: cca140a3 on this branch:
  • cargo fmt --check: clean.
  • cargo clippy --all-targets --locked -- -D warnings on the repo-pinned toolchain (1.96.1 via rust-toolchain.toml; RUSTUP_TOOLCHAIN unset): clean.
  • Unit tests, run per top-level module in isolated processes (cargo-nextest is not installed and under plain cargo test upstream's platform::unix_common test resets SIGPIPE in-process, after which server::client_transport::...write_failure kills the whole binary; nextest's per-test processes hide this): 3,593/3,593 bin tests pass, 72/72 lib tests pass. Touched modules re-run on the final tree: client 668, server 309, protocol 156, kitty_graphics 46, api 314, cli 107, pane 381, terminal 286, update 78, app 480, events 3, terminal_effects 3; 0 failures.
  • Integration tests (cargo test --test '*' -- --test-threads=1): 17 binaries, 140 tests, 0 failures (api_ping 13, client_mode 27, cross_area 10, detach_reattach 13, live_handoff 23, machine_setup 11, multi_client 9, remote_attach 1, server_headless 18, session_delete 6, session_restore 6, broken_pipe 3).
  • cargo build --release --locked: ok.
  • Smoke test in a throwaway herdr --session merge-smoke (isolated XDG_CONFIG_HOME/XDG_STATE_HOME, HERDR_* scrubbed, kitty-replying pty host) against the release binary:

server_started: true
workspace_list: {"code":0,"workspaces":1}
plugin_link: ok (tests/fixtures/plugin-smoke, id example.smoke)
plugin_pane_open --placement split --no-focus: ok, pane.focused=false
focused pane before/after: w1:p1 -> w1:p1
OSC 10;? / OSC 11;? from a pane: ]10;rgb:cdcd/d6d6/f4f4  ]11;rgb:1e1e/1e1e/2e2e
session_stop: ok, session_delete: ok

New tests added by the three feature commits: kitty_graphics::tests::{large_raw_uploads_are_deflated_and_round_trip, png_and_small_uploads_stay_raw, incompressible_raw_uploads_stay_raw}, client::events::tests::input_is_delivered_before_queued_server_frames, api::schema::tests::client_sidebar_toggle_request_round_trips, client::endpoint::control::tests::{sidebar_toggle_optional_control_round_trips, pane_pointer_shape_optional_control_round_trips}, client::shell::tests::chrome_context::toggle_sidebar_flips_the_collapsed_state_and_requests_a_resize, server::headless::tests::{sidebar_toggle_api_reaches_the_foreground_client, sidebar_toggle_api_reports_no_foreground_client}, client::shell::tests::pointer_shape::*.

Note: the plugin pane did not show up in herdr pane list (count stayed at 1) even though plugin pane open returned a pane object with focused: false; the stock 0.9.3 binary was not checked for the same listing behaviour. The --no-focus claim rests on the response and the unchanged focused pane id.

Note: the fork's OSC 22 host-cursor mirroring and the input-latency fix were verified by unit tests only; no interactive latency measurement was repeated on this branch.

Merge Danger

Door: two-way

Plain merge commit plus three feature commits on a fork branch; git revert -m 1 cf614e08 or resetting patt/perf-graphics to 134c42fb restores the previous release line. Upstream master on the fork was fast-forwarded separately (e7c96206..3d9d2b18).

Blast Radius: fork-wide

• Every upstream behaviour change from v0.8.0 to v0.9.3 lands at once, including the removal of the pane.graphics.* API (feat!: replace the custom graphics api with optimized native kitty rendering herdrdev/herdr#4561) and the client-rendered shell (refactor: render the shell in the client herdrdev/herdr#3487).
• Cargo.toml is now 0.9.3; a fork binary built from this branch reports 0.9.3 and the stable updater will offer herdrdev's 0.9.4 when it ships. Decide between the suffixed version (keep e930c6a1) and HERDR_BUILD_CHANNEL=graphics before the next fork release.
• Building now requires Zig 0.16.0 (vendor/libghostty-vt/build.zig.zon, upstream CI); mise install zig@0.16.0 and ZIG=~/.local/share/mise/installs/zig/0.16.0/bin/zig cargo build --release.
• patt/api-first-read-poll (7a8a5ad7) now conflicts with master in src/api/server.rs: git merge-tree shows upstream's ssh_agent_registration_lasts_only_for_the_api_connection test (3602c75 fix: refresh forwarded ssh agents after reconnect herdrdev/herdr#4443) landing where the cherry-pick adds initial_request_line_arriving_after_accept_is_read_without_poll_delay; 18 upstream commits touched that file (fix: keep api listener alive after transient accept errors herdrdev/herdr#4659 accept-error handling, fix: keep the server alive when thread spawns fail herdrdev/herdr#4958/fix: keep the server running on hangup and log why it stops herdrdev/herdr#4965 server liveness), so that PR needs a rebase and a re-check that the read-loop change still applies.

Follow-ups

• Separate PR to drop e930c6a1 once fork releases use HERDR_BUILD_CHANNEL/HERDR_BUILD_ID.
• Regenerate bench/results-*.json against this branch; the fork numbers predate fix: keep offscreen kitty images loaded instead of resending them on scroll herdrdev/herdr#4686/perf: send scrolling pane output as row shifts and skip blank cells herdrdev/herdr#4711.
• ja/zh-cn drafts for the new client.sidebar.toggle docs were not written.

https://claude.ai/code/session_01ApRKSwFTijvcbtt8EG9Erx

[pattobrien]

Superseded by #3: same tree, based on master so the diff is only the fork delta.