Codex uses the osc_title_stability idle strategy, but it draws its
progress in the pane body ('Working … esc to interrupt'), not the OSC
title. The title goes stable mid-turn, so ~2s later the classifier
declared codex idle while it was still working. Add a thinking-promoter
pattern ((?i)esc to interrupt) to the codex built-in preset; classify()
checks promoter regexes against the rendered screen before the
title-stability verdict, so codex stays in thinking until the turn's
in-progress footer actually disappears.
Resolves the [CODEX IDLE] TODO item.
handleConn processed requests serially, so a slow tool (e.g.
wait_for_pattern with a 300s timeout) monopolized the single per-agent
MCP connection and every queued call timed out behind it. Handle each
request in its own goroutine, serialize responses through a per-conn
write mutex (full response written atomically, partial writes handled),
copy the request line before handing it off (bufio reuses its buffer),
and wait on a WaitGroup before closing the conn so in-flight handlers
finish cleanly. Greeting stays sequential; notifications still get no
response.
Resolves the [MCP TIMEOUT] TODO item.
Grid snapshots pad every row to the full terminal width and leave the
bottom of the screen blank, so MCP grid reads carried a lot of dead
whitespace. Add normalizeGridText (CRLF/lone-CR to LF, right-trim each
line, collapse blank runs to a single blank, drop leading/trailing
blanks) and apply it to the grid branch of GetProcessOutput only.
Stream output, raw output, and WaitForPattern matching are untouched.
Resolves the terminal-read newline/token-waste TODO item.
Agent 'Close' (agent-close) sent a single SIGTERM via Session.Kill and
never escalated, so an agent that traps/ignores SIGTERM (e.g. opencode)
stayed in the running tab bar until the user closed it again. Add
Session.Terminate, which reuses terminateAndWait (SIGTERM, wait, then
SIGKILL) but preserves the session entry so the exited pane stays
readable, and route handleChildClose's agent path through it in a
goroutine to keep the UI responsive during the stop timeout.
Resolves the opencode double-close TODO item.
Mirrors the existing scratchpad_* tools end-to-end: catalog schema,
dispatch, ToolHost.ScratchpadDelete, and a host method that delegates to
scratchpad.Store.Delete and fires scratchpadsChanged() on success so the
sidebar refreshes. Missing-pad errors surface rather than being masked.
Resolves the [MCP SCRATCHPAD DELETE] TODO item.
Each agent tab now prefixes its label with the same one-rune idle
indicator the sidebar uses (✕ error, ? permission, ◐ thinking, ○ idle,
● working), so the state of every open agent is visible without
opening or focusing each tab. Tab redraws now fire on idle-state
changes in addition to sidebar redraws.
Stale timer bodies were re-delivered to the orchestrator pane after
the parent had already processed the sub-agent's reply and called
close_process. The timer registry held no link to the child
lifecycle, so timers owned by or watching the closed child lingered
until something triggered a fire — e.g. a trailing classifier tick
for the now-removed child.
Add an OnChildClosed hook to ChildEventListener, emit it from
Session.Close (and the terminal-corpse path in reapChild), and have
the timer manager prune the registry: cancel timers owned by the
closed child; remove the closed child from each timer's watched
list (cancel the timer outright when watched empties).
Natural exit deliberately does not route through this hook — the
classifier already emits an idle transition on exit which delivers
any legitimate "fire when sub-agent finishes" semantics exactly
once; cancelling on exit would swallow that.
Toasts now render three content rows with word-wrapped bodies. The
in-toast "Ctrl-N · N more" hint is replaced by a short
"Ctrl-N · dismiss" entry on the status strip that only appears
while a notification is live.
The box stops flickering while the focused child repaints its TUI:
the overlay is stitched onto the per-chunk PTY write under outMu
and bracketed by DECSET 2026 so supporting terminals buffer the
child's redraw and the box paint into a single frame.
Focused-section rows are now bare verbs (Rename, Close, Stop, Restart,
Delete, Edit) instead of repeating the focused name. The title bar
already carries the subject, and the row hint preserves fuzzy-search
matches like "close codex". Section banners are replaced by a single
blank spacer row so the verbs themselves carry the visual weight,
and the Open section no longer lists "Switch to <current>" for the
pane that's already focused.
Replaces the single-slot status-line flash with a top-right toast
stack over the focused pane. flashError, flashTransient, and
notifyAttention all push onto the same stack (cap 5, FIFO drop).
Ctrl-N dismisses the most recent toast; empty stack falls through to
the focused PTY so readline / nano / emacs / opencode bindings keep
working. A new "Clear notifications" palette item empties the stack.
Removes the 2026-05-15 perf audit findings that have either shipped
(see CHANGELOG) or are tracked elsewhere, and replaces them with the
remaining palette-refinement notes: generic labels for focused
actions ("Close current agent") and a higher-level concern that the
palette has grown cluttered as features were added.
`mlugg/setup-zig` was chasing mirrors for ~4 minutes on every run
(see v0.0.1 / v0.0.2 logs) and `actions/setup-go` was spending
another ~4 minutes downloading Go before patterm started building.
mise already manages the project's zig pin; adding `go = "1.26.3"`
to `.mise.toml` (matching go.mod) lets `jdx/mise-action@v2` install
both with one cached step. Subsequent runs reuse the mise cache
instead of re-resolving mirror URLs and re-downloading toolchains.
Also adds an `actions/cache@v4` step for `~/.cache/go-build` and
`~/go/pkg/mod` keyed on `go.sum` so `go build` itself doesn't
re-pull modules every tag push.
`mlugg/setup-zig@v1` is deprecated and only knows the pre-0.14
tarball name (`zig-linux-x86_64-<ver>.tar.xz`), so every mirror —
and the official ziglang.org/builds — returned 404 for Zig 0.15.2
on both the v0.0.1 and v0.0.2 release runs. v2 uses the new
`zig-x86_64-linux-<ver>.tar.xz` layout that Zig switched to in
0.14+.
Also rolls the existing CHANGELOG `[Unreleased]` work into a
dated `[0.0.3]` section and adds the CI fix to its Fixed list.
Six-phase sweep: section headers (Focused / Open / Spawn / Quit) with
header-skip cursor; chip strip mirroring sw/sp/k macros, driven by
Tab; unified Spawn verbs across agent / process / terminal / custom;
dropped duplicate global Close list in favor of Ctrl-X inline close
on a Switch row plus the [Close] chip; scored matching (prefix >
word-boundary > substring > fuzzy) with matched-char highlighting;
title bar surfaces focus subject; rename forms split long subject
onto its own row; new Alt-1..9 quick-pick, Home/End, ? help overlay,
and Ctrl-R relaunch toggle inside the spawn-process form. Scroll
indicator and cursor/total counter round out the footer.
A sub-agent's reply to send_message lands in the caller's own pane
tagged [sub-agent:<name>], not in the sub-agent's output. The
descriptions for wait_for_pattern, send_message, both
timer_fire_when_idle_*, and the server-instructions preamble now
spell this out, along with the canonical send_message →
timer_fire_when_idle_any → read-own-pane pattern. help('readiness')
and help('coordination') updated to match. Previously agents reached
for wait_for_pattern on the sub-agent and deadlocked until timeout
because the reply had already been delivered to their own pane.
Typing into a focused child while its emulator viewport was
scrolled up left the keystroke heading to the PTY but the input
box invisible below the visible region — it looked like typing
did nothing. processStdin's flushForward now sets
pendingViewportBottom whenever bytes are actually injected, so
the existing post-loop handler snaps the viewport and repaints.
Wheel events and Ctrl-B paths are untouched: both are intercepted
before reaching forward, so wheel still scrolls into history and
Ctrl-B is still the explicit escape hatch. Only bytes that would
actually reach the child PTY trigger the snap.
Sidebar rows that overflow the rail width used to spill characters
into the main viewport. They now truncate with a trailing "…"
when unfocused (or when the focused name still fits). The focused
row whose name overflows runs a pause-scroll-pause marquee: 1 s
hold on the head, ~150 ms per cell scroll, 1 s hold on the tail,
snap back. The row's geometry never moves while it animates, so
nothing below shifts.
A dedicated 150 ms goroutine flips sidebarDirty only while a row
is actively animating; the chrome ticker does the actual repaint.
Idle is a single cheap wakeup. focus / spawn / exit / restart all
reset the marquee state so the new focused row starts from frame
zero. When the row's budget is tight, the trailing timer
indicator drops before the name ellipses since the name is the
only identifier the row carries.
clampVisible() is a defensive net inside write(): even if a row's
decoration size were mis-computed, it will not spill past the
sidebar band into the PTY area.
The top tab bar compared against focusedID, so stepping into a
sub-agent dropped the parent tab's highlight even though the user
was still inside that thread. activeAgentID already walks the
parent chain to the top-level root for the sidebar's agent tree
— reuse it for the tab strip too.
Bundles the in-flight work into the second tagged release. See
CHANGELOG.md `[0.0.2] - 2026-05-15` for the full per-change list.
Highlights:
- libghostty-vt was building in zig's silent Debug default, capping
the full pipeline at 34-63 fps. Makefile now defaults to
ReleaseFast (.mise.toml pins zig 0.15.2 so the build is
reproducible). End-to-end pipeline now runs at 930-2030 fps —
27-32× faster, with 7-16× headroom over a 120 fps target.
- --debug[=DIR] and --profile[=DIR] flags capture full PTY logs,
pprof data, and per-hot-path metrics (chunks/sec, mean/max
latencies, cache hit rates) for offline analysis. Nothing
pollutes stdout/stderr.
- ASCII-video benchmark suite (8-colour / truecolor / Bad-Apple
patterns at 30/60/120 fps) plus a renderer microbenchmark set
for stable A/B comparisons across changes.
- Click-and-drag text selection from alt-screen TUIs (codex) now
works — host mouse mode follows the focused child's screen side
instead of being permanently armed.
- Long claude session resume + codex steady-state rendering pay
less per chunk: drawSidebar deferred to the chrome ticker,
emulator.Title CGO poll gated on a containsOSC scan.
- Vendor-TUI orientation: MCP initialize.instructions, the
spawn_agent tool description, and help('spawning') all spell
out the anti-patterns (shell-out, perl-into-socket) that
produced codex's stray top-level tabs.
Added .mise.toml pinning zig = "0.15.2" (the minimum the vendored
Ghostty commit requires) and taught the Makefile to resolve zig
through mise when available, falling back to PATH. Contributors run
`mise install` once and `make deps` just works.
Re-ran the pipeline benchmarks after rebuilding libghostty-vt with
ReleaseFast (same hardware, AMD Ryzen 7 7800X3D):
Debug ReleaseFast speedup
Pipeline 8-colour @120fps 63 fps 2030 fps 32x
Pipeline truecolor @120fps 34 fps 931 fps 27x
Emulator-only truecolor 34 fps 2051 fps 60x
7-16x headroom over 120 fps for the heaviest workload (truecolor
full-screen redraws). Static library size 33 MiB -> 13 MiB.
TODO.md baseline numbers updated to reflect post-fix throughput;
the "Debug-mode lib" finding is folded into the result it produced
rather than left as an open item.
Added a full ASCII-video benchmark suite that hammers the renderer
with 30 KiB / 70 KiB full-screen frames at 30, 60, and 120 fps
targets — both renderer-only and full-pipeline (em.Write + renderer
+ stdout). Each stream benchmark reports µs/frame, fps_ceiling, and
percent of the per-frame budget consumed.
The pipeline benchmarks revealed we were missing 120 fps by a wide
margin (190%-350% of budget at 120fps, 60-90 fps ceiling). Isolating
em.Write confirmed libghostty-vt is the bottleneck — 16-29 ms per
truecolor frame, library file at 33 MiB.
Root cause: the Makefile invoked `zig build` with no
-Doptimize, and Zig's standardOptimizeOption defaults to Debug. So
the shipped libghostty-vt was unoptimised. Fixed by pinning
ReleaseFast in the Makefile (override via GHOSTTY_VT_OPTIMIZE for
debug builds of the upstream lib).
Existing checkouts need `make clean-deps && make deps` to pick up
the rebuild.
Live metrics (--profile):
- New metricsTracker instruments OnPTYOut, viewport renderer,
stdout writes, libghostty-vt Write/Title CGO calls, sidebar /
tabbar / status draws (with cache-hit accounting), snapshot
replays, and the chrome ticker (so we can see ticker fires that
did nothing).
- Writes metrics.jsonl (one snapshot per second) and metrics.json
+ summary.txt on exit, alongside the existing pprof files.
- All record* methods are nil-safe so disabled paths pay only a
cheap nil check; counters are atomic so the per-PTY-chunk hot
path stays lock-free.
Benchmark suite (go test -bench=.):
- Three workload fixtures — plain ASCII, SGR-styled lines, and a
ratatui-style cursor-shuffling burst — plus a containsOSC
microbenchmark. Reports ns/op, MB/s, allocs/op, B/op.
- Initial baseline numbers added to TODO under the perf-audit
section, alongside two new findings (renderer allocs ~1 per 4
bytes on styled chunks; styled throughput tops out near
90 MB/s) those benchmarks surfaced.
Codebase sweep for perf issues outside the per-PTY-chunk path that
recent CHANGELOG work already covered. Ten findings under a new
"Perf Audit (auto-generated)" section in TODO.md — anchored to
file:line, classified MEDIUM/LOW, with a sketched fix per entry.
None landed as code changes; review pending.
- Add --debug[=DIR] / --profile[=DIR] flags that write run artefacts
(patterm.log, events.jsonl, per-child raw PTY captures, CPU + heap
+ goroutine pprof) to a dir without polluting stdout/stderr.
- Strengthen vendor-TUI orientation in three places (MCP
initialize.instructions, the spawn_agent tool description, and
help('spawning')) to head off codex's habits of poking the Unix
socket via perl and shelling out to launch peers — both bypass
caller identity and produce orphaned top-level tabs.
- Fix click-and-drag text selection from alt-screen TUIs. Host SGR
mouse reporting now follows the focused child's screen side
instead of being permanently armed; alt-screen TUIs that need
mouse re-enable it themselves and the toggle is forwarded.
- Move drawSidebar() off the per-PTY-chunk hot path. Long claude
session resume was paying a full sidebar rebuild for every
scrolled chunk; the chrome ticker now drains a dirty flag at 60 Hz.
- Gate the per-chunk Title() CGO poll on a containsOSC scan so
codex/ratatui's many SGR-only chunks no longer pay a CGO call each.
- Palette's per-child "Kill <name>" action is now labelled "Close <name>"
(action kind unchanged; still SIGTERM). Matches the existing "Close
agent: …" context entry and reads less violent for a graceful term.
- New "New Terminal" palette entry spawns a bare interactive $SHELL pane
via LaunchTerminal (kind=terminal). Replaces the default "shell"
process preset that was seeded on first run.
- Exited KindTerminal entries are now dropped from the session in
reapChild — terminals have no restart path, so leaving them behind as
greyed rows in the Processes sidebar was just clutter. processList
also filters defensively.
Classifies every running child as idle/working/thinking/permission/error
using one of three pluggable strategies (output_activity,
osc_title_stability, osc_title_status) plus optional regex promoters
applied to the tail of recent output. State and last-match reason are
exposed via MCP on ProcessInfo and get_process_status. Per-preset
configuration lives on a new preset.IdleDetection block with bundled
defaults for the first-party claude/codex/opencode presets.
OSC title plumbing is exposed as Emulator.Title(), polled from the
session pump after each emulator write so title-change activity feeds
into the classifier without an extra cgo callback.
The MCP timer surface expands to match Solo: timer_set,
timer_fire_when_idle_any/all, timer_cancel, timer_pause, timer_resume,
timer_list. timer_wait is now a thin wrapper that shares the same
manager so it shows up in timer_list while pending. Timer bodies are
delivered to the owner process through the existing
InjectAsOrchestrator path. Top-level (non-agent) callers can attach
timers to a specific process via owner_process_id; omitting it grants
universal cancel/pause/resume/list privileges.
The sidebar gains a state glyph per process row and appends a
nearest-timer indicator when one is pending or paused.
Tests: idle_test.go covers the classify() pure function across the
three strategies and regex promotion; timers_test.go covers the
manager. Harness scenarios cover output_activity, osc_title_stability,
osc_title_status, and regex promotion, plus timer_set delivery,
cancel, pause/resume, idle_any-on-transition, idle_all-pending, and
idle_all-already-satisfied. A new wait_until_mcp harness step type
polls an MCP method until an assertion holds.
When opened with Ctrl-K, the palette now prepends entries for whatever
is currently focused:
- Focused scratchpad: Delete / Rename (inline form) / Edit (fire-and-
forget zed launch with stdio detached so the TUI is not suspended).
- Focused agent: Rename (inline form) / Close.
- Focused process: Rename / Delete (drops the entry; SIGKILL if alive)
/ Stop (SIGTERM, keep entry) / Restart (same argv).
The rename UX is a single-field inline form that mirrors the existing
spawn-process form, so the modal-input contract is unchanged.
scratchpad.Store grows Delete / Rename / Path so the palette can act
on a pad file by name. focusedPad is plumbed onto uiState ahead of the
scratchpad-focus UI work; until that lands it stays empty and the
scratchpad-context entries simply never surface.
Tested with palette_context_test.go and a new rename_process_via_palette
harness scenario.
