Vendor Inter (variable + italic, OFL) under assets/fonts/inter/ and make it
the default application UI face, replacing Josefin Sans (which stays bundled
as a selectable option). pubspec font family + licenses.yaml entry per D-42.
Settings → Appearance gains a UI-font select (Inter / Josefin Sans). The root
DefaultTextStyle reads app.ui.font (kUiFontSettingKey) over the default and
re-applies live on settings change, so a pick takes effect immediately. Bump
the default UI weight to w400 — Inter reads better at Regular than Josefin's
Light.
The monospace picker is deferred to T-471: clideMonoFamily is a const at ~38
call sites (not inherited), so a live mono picker needs the family routed
through context first.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A Settings → Claude category sets per-user defaults for NEW sessions: model,
effort, and permission mode (generic schema selects persisting
app.claude.default*). The pane reads them at spawn — effort flows through the
existing --effort flag (SpawnSpec.effort); model and permission mode are sent
as control requests right after a fresh (non-resume, non-fork) session starts.
'default'/unset values are no-ops, leaving the CLI's own defaults.
The optional "settings changed — apply to current sessions?" prompt is filed
as T-470: it needs a custom control (a generic select can't prompt), so it's
a clean follow-up on the T-452 escape hatch.
Tests: defaultEffortFlag sentinel handling; the category contributes the three
default fields. The thin pane-side applySessionDefaults is covered by the gate.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Reflow lines `dart format` collapses — the T-451/T-452 test additions were
committed before `make format` ran, so the committed copies were unformatted.
No behavior change; keeps the push format gate green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Built-in extensions are always on and there's no third-party install path yet,
so the Extensions tab is a notice rather than a toggle list — a toggle list
would let users brick the app by disabling load-bearing builtins. It explains
that installing/enabling/disabling arrives with third-party (Lua) extensions
and points at the records that pick it up (D-16 / T-8).
Fleshes out the builtin.extensions-ui stub via the custom-control hatch. Adds
two renderer guards — empty section/field labels render no chrome — so a
notice-only category shows just its card.
Tests: contributes the Extensions category + notice control; the notice renders.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add the one bespoke control the schema engine defers to. New
SettingsControlContribution routes a WidgetBuilder into a kernel
SettingsControlRegistry under a customId; a SettingsFieldKind.custom field
names that id, and the renderer draws the registered widget full-width
(label on top, no scope tag — the control owns its own apply + scope).
The theme-picker extension uses it: an Appearance category whose theme field
is custom, backed by AppearanceThemeControl — base-theme chips + a
high-contrast toggle that apply live through ThemeController (persisted by
theme_persistence). Reuses the shared theme_families helpers.
Tests: control registry (register/dup/unregister), the renderer's custom-field
path, and the Appearance contribution + live theme apply.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The keybindings-ui extension contributes a Keymap SettingsCategory: a preset
select (Default / Vim / VS Code / JetBrains) reading the active preset from
kKeymapPresetSetting. Picking one routes through a new schema affordance —
SettingsField.applyCommandPrefix — running `keymap.preset.<value>`, which calls
KeymapService.setPreset to persist and reload the layered keymap live. The
prefix path keeps the engine generic: settings a subsystem applies via a
command (rather than a bare key write) declare the prefix; the scope tag and
current-value read still use the key.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Deny Bash(git add -A*), Bash(git add --all*), and Bash(git add .) so the
bulk-stage forms can't run; deny overrides the existing git add * allow. The
git-commit skill mandates specific-path staging, and a shared working tree
makes -A actively dangerous (it sweeps another agent's files into the commit).
Exact-match the bare dot rather than git add .* so git add .gitignore and
other dotfile paths still work.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The first real settings tab. The Claude extension contributes an Activity
SettingsCategory whose fold-level select binds to app.claude.activityFoldLevel;
the activity panes already rebuild off the settings notifier, so a pick applies
live. Proves the schema-driven engine end to end with a production setting —
opening Settings now shows a populated, working category.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A concurrent session's `git add -A` swept this turn's uncommitted doc
edits into unrelated feat(settings) commits (f56ad88, f643d6e). Add an
explicit-staging guardrail to the Git workflow section and route all
commits through the git-commit skill.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A search box (ClideFilterBox) atop the rail filters fields across every
registered category. While searching, the panel swaps to a results view that
groups the matching fields under category subheaders — rendered with the same
carded rows and editable inline — and each rail row shows its match count with
zero-match categories dimmed. Completes the settings-UI infra spine (T-444):
shell, engine, rail, scope tags, search.
Tests: search filters across categories, hides non-matches, and surfaces the
per-category rail count.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Each settings field gains a scope tag showing where its value lives — folder
= Project (.clide), globe = Always (~/.clide), circle-dashed = Default/unset —
colour-coded (statusSuccess / statusWarning / muted) with a tooltip. Tapping
opens a menu to move the value between the scopes the key supports, or reset
to default; the tag's menu replaces the interim reset button.
Backs it with scope-explicit SettingsStore access — rawAt / setAt / removeAt /
effectiveLayer / writableLayers — over the two storage files (app ~/.clide,
project .clide). ext.* keys layer project-over-app; app.*/project.* keys live
only in their prefix's file, so their menu offers that one scope + reset.
Tests: store scope ops (layering, reload, guards) and the tag (Default vs
All-clide rendering, menu reset).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The settings modal's left rail now lists the registered categories (icon +
title, data-driven from the SettingsRegistry) with an accent left-stripe +
surfaceHi selection per ui-design surface.md. Selecting a category drives the
modal's selection state and swaps the right panel; the rail scrolls when the
set exceeds the height. The cross-category search box atop the rail lands in
T-450.
Tests: rail lists categories and selecting one swaps the panel.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The core of the settings panel (epic T-444). Subsystems describe a category
as data — a SettingsCategory of carded SettingsSections of SettingsFields
(toggle / select / text / number / opens-external-file), each bound to a
SettingsStore key with help text, a default, and reset-to-default.
Registration is declarative: a new SettingsCategoryContribution carries the
category; the extension manager routes it into a new kernel SettingsRegistry
(exposed on KernelServices), which the panel reads via ClideKernel. Adding a
category is now pure data + a contribution — no widget code.
SettingsCategoryView renders a category into carded sections per ui-design
surface.md: panelHeader card fill, dividerColor border, inputs receding to
panelBackground; select reuses the anchored-overlay menu, text/number commit
on Enter or blur (numeric clamps to bounds). The modal panel now shows the
selected/first registered category, falling back to the empty state.
Tests: registry (sort / dedup / notify), contribution routing on activation,
renderer (render + toggle/select write-through + reset), modal-with-category.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Lay the foundation for the schema-driven settings UI (epic T-444). A new
`settings.open` command (⌘`,`, plus a File-menu and command-palette entry)
opens a centered Settings modal over the dimmed app via the dialog router,
built from the modalSurface* tokens (D-7, no Material). The shell frames the
two regions later tickets fill in — the category rail (T-447) and the
scrolling carded panel (T-448) — and dismisses on ✕, Esc, or barrier tap.
With no category registered yet it shows its empty state, which is the
correct runtime state.
Flesh out the `builtin.settings-ui` stub (was 0.0.0-stub) into a real
extension; ship its en-US i18n catalog. Relabel the theme picker's
`theme.pick` command title from "Settings…" to "Theme…" so the two no
longer collide in the palette (the picker folds into the new panel's
Appearance category in T-452).
Tests: command + ⌘`,` binding registered, shell renders, Esc and close
both dismiss.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Flag API-error conversation cards with a red statusError border (the card
already has an error status + bordered variant; just needs wiring). UI
tracker T-276.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
UI + monospace font pickers in the Appearance settings tab; bundle Inter
(SIL OFL) as the default UI font. Child of the settings-UI epic T-444,
renders alongside the theme picker (T-452).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
self.version is auto-synced from pubspec by make gen-build-info; pubspec
is already at 2.6.0 but the generated value had lagged at 2.5.0. Commit
the regenerated file so the working tree matches.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Both filed under the T-276 UI epic.
T-458: the editor records neither the original EOL style nor the
trailing-newline state when it reads a file (registry.dart), and only
normalizes on save when .editorconfig asks. Any layer that rewrites
newlines therefore produces cross-platform commit churn with no
.gitattributes backstop. Ticket captures the record-on-read /
preserve-on-save / gitattributes plan.
T-459: closing the last editor panel in the main column's top split
does not collapse it. Suspect the asymmetric top/bottom split logic in
slot_host.dart and closeEditor() not moving the active tab off
editor.active. Root cause not yet confirmed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Bump pubspec 2.5.0 -> 2.6.0 and move the Unreleased entries under a dated
2.6.0 heading. Minor (not patch) — the cut includes new features (vim
ex-line + gt/gT, crash-survivable logging, watchdog) alongside the fixes.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The push race created two T-440s: the remote tree-sitter.dll task and our
local Playwright-e2e follow-up. Kept the already-shared remote one as T-440;
renumbered our local one to T-443 via pql ticket relabel (record_id stable).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Closes out the T-403 cross-pane vim layer. gt/gT bind to the existing
workspace.tab.next/previous commands (also on ctrl+pagedown/up for every
preset), resolved by the focused editor's matcher or a pane's PaneKeyNav —
bare-g sequences stay editor/pane-local (T-404's global matcher only engages
on modified-chord prefixes), so no global-matcher surgery.
- vim.yaml: g t -> command:workspace.tab.next, g shift+t -> .previous
(vim.normal); shares the `g` prefix with `g g` (docStart / nav.top),
distinguished by the final chord.
- PaneKeyNav now EXECUTES non-editor.vim.* command intents (e.g.
workspace.tab.*) instead of swallowing all command intents, so gt/gT work
from a focused pane; editor.vim.* buffer edits stay blocked in panes.
Tests: vim-preset resolution (gt/gT bind, gg intact), PaneKeyNav executes
the command from a pane, and editor.vim.* is never run from a pane. make
test green; analyze + format clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
User opted to skip the live GUI review; improvement tickets will be filed
if anything needs polish. Completes all built children of T-403; only
T-405 part 2 (gt/gT) remains.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Under the Vim preset, `:` opens a transient one-line ex overlay running a
fixed v1 table; ZZ runs :wq directly. Completes the last built child of the
T-403 cross-pane vim layer (T-405 part 2 gt/gT still open).
- ExLineController + parseExCommand grammar + editor-targeted executors
(lib/kernel/src/ex_line.dart); the overlay (lib/widgets/src/ex_line_overlay
.dart) reuses the quick-open chrome, mounts in the root_shell Stack, and
publishes the exline.open scope flag. Unknown commands flash + stay open;
with no active buffer every command no-ops (2026-06-13 decision).
- :q closes the active tab via editor.close on its id — the registry promotes
the next buffer and the split self-collapses on the last (2026-06-12
decision); :w/:wq/:x/ZZ save (+close) the active buffer.
- :e <path> seeds quick-open (new QuickOpenController.open(seed:)); :N adds the
editor.goto-line IPC/CLI verb (reuses _offsetForLine). Goto needs caret sync:
EditorController now handles editor.selection-changed and the editor view
moves the caret on a selection-only change.
- `:` and ZZ are typed intents; the editor matcher and PaneKeyNav now bubble
unhandled typed intents to the app-root Actions, so they fire from any focus.
vim.yaml binds `:`, ZZ (shift+z shift+z), and Esc-dismiss.
Tests: parser/controller/executors, editor.goto-line daemon tests,
selection-changed (controller + view), full overlay widget test. make test
green; analyze + format clean.
Also files T-441 (drop bold from the ticket-id card label) and T-442
(sub-agent renders as 3 cards instead of one bundle) under the T-276 UI epic.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Split the Windows-DLL slice out of T-25 (cross-OS libtree-sitter
build). The FFI loader already resolves tree-sitter.dll on Windows,
but no Windows binary is vendored — only the Linux .so — so
tree-sitter highlighting silently falls back to plain text on
Windows. T-440 is the narrow task to build and vendor that DLL now;
CI reproducibility and the macOS .dylib stay on the T-25 umbrella.
Co-Authored-By: Claude <noreply@anthropic.com>
Acceptance gap confirmed closed: a make-install build launched from the
Linux application menu (not a terminal) resolves the login-shell PATH —
pql/git/claude/PTY children find user-installed tools. macOS Homebrew
path remains code-correct; a separate bug will be filed if it surfaces.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Live-confirmed on CLI 2.1.177: /clear clears the primary pane to an
empty conversation without exiting code 1. The await-death teardown fix
(081678e) holds in the running app.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Desktop/dock-launched clide inherits a sparse PATH (no ~/.local/bin, brew,
nvm, …), so pql/git/claude and PTY tools went missing. T-347 fixed only the
toolchain/pql path on Linux; env.dart's expander was still macOS-only and
claude/PTY/git used the raw PATH — the breakage recurred per spawn site
because there were three divergent expanders.
Consolidate into one resolver (lib/src/env/shell_env.dart):
- primeLoginShellPath(): probe the user's real login shell once at startup
(`$SHELL -l -c`, sentinel-framed, bounded timeout, graceful fallback to the
process PATH). Captures the user's actual PATH, not a hardcoded guess.
- expandToolPath(): the canonical merge (moved from toolchain_paths, which
re-exports it for its tests) — unions the well-known user/local bin dirs.
- resolvedToolPath(): currentSearchPath() + expandToolPath, the single call
every spawn site uses.
Routed through it: PTY children (registry.dart now overrides PATH), git
(env.dart → operations.dart), the toolchain probe (toolchain_paths), and
hosted claude (agent_bootstrap). Primed in main.dart's !kIsWeb boot. Deleted
the macOS-only env.dart copy and the cli_install copy.
Tests: new shell_env_test (probe + every fallback + merge); env_test and
cli_install_test updated to the consolidated surface. analyze clean, web wasm
build still green, make test green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
`flutter build web --wasm` had been broken since the tree-sitter/PTY dart:ffi
pivot. Per D-100 (resolving Q-50: keep the web "happy accident" alive), every
native binding now sits behind a `dart.library.ffi` conditional import with a
graceful web stub. Desktop builds are unchanged — no fidelity loss; the web
target degrades (no terminal, native git, or syntax highlighting).
Discriminator is `dart.library.ffi`, not `dart.library.io` — dart2wasm provides
dart:io, so FFI is the only blocker.
Fences:
- PTY: pty_session → pty_backend_io / pty_backend_web (stub throws).
- tree-sitter: pure types → syntax_result.dart; tree_sitter_service is now a
facade over _ffi/_stub; tree_sitter_boot_io/stub fences TreeSitterLib.init().
- watchdog: watchdog_windows_stub (all -1 sampler).
- claude ABI probe: native_abi_io/stub (was `dart:ffi show Abi`).
- testmode fd-check: fd_check_io/stub.
Also dart2js-safe: the 64-bit FNV literals in session_naming.dart + paths.dart
(the dual JS fallback rejected them) — split into 32-bit halves, dropped a
no-op 64-bit mask. Desktop/wasm hash values unchanged.
CI: added a `web-wasm` job (flutter build web --wasm) so the fence can't rot.
Two FFI-constructing tree-sitter tests import _ffi.dart directly (the analyzer
resolves the conditional facade to the stub branch).
Verified: `flutter build web --wasm` → built; `flutter analyze` clean;
`make test` green. Full Playwright e2e harness wiring is the tracked follow-on.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Regenerated governance/README.md index (pql decisions sync): D-32
retitled, D-100 added, Q-50 moved to Resolved. Belongs with 4bba2ef.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- D-32 amended: Gitea-primary/not-activated → GitHub Actions, active
(Linux test + Windows ConPTY/soak + release; web-WASM e2e withheld).
The staged Gitea pipeline was never activated and is gone.
- Q-50 resolved → D-100: FENCE the web/WASM target. Every dart:ffi
importer goes behind a conditional-import facade with a web stub so
`flutter build web --wasm` compiles; desktop fidelity untouched. Keeps
the web "happy accident" alive as a hopeful future target per user.
- T-438 filed for the fence implementation (12 ffi importers + CI wasm
compile gate + re-enable e2e/ui targets).
- T-384 closed (done): Gitea premise OBE, scripts repointed (2026-06-12),
D-32 reconciled; the dead e2e targets delegated to T-438.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
/clear tore the session down and respawned on the same deterministic
--session-id BEFORE the old claude process had actually exited. The
orchestrator's close() ran conversation.dispose() unawaited and kill()
only sent SIGTERM without awaiting exitCode, so the respawn raced a
still-alive holder of the id — claude 2.1.177 rejects it as "Session ID
… is already in use" and exits 1.
Root cause confirmed from clide's own crash log + isolated probes against
2.1.177: the id frees the instant the holder dies (SIGTERM cleans the new
~/.claude/sessions/<pid>.json registry), so awaiting real death is the
fix — and it preserves T-268's deterministic-id continuity (chosen over
minting a fresh id, which would change the continuity model).
- stream_json_session: kill() awaits exitCode (SIGTERM → 2s → SIGKILL →
await); dispose() idempotent (shared cached future); new
SessionEnd.reason getter (last non-empty stderr line, capped).
- session_orchestrator: close() awaits session.dispose() so teardown
returns only once the process is truly dead, before clear + respawn.
- claude_pane: surface end.reason in the status line — no more opaque
"code 1".
- session_naming: correct the stale clearSessionTranscript doc (real
sidecar is the shared memory/ dir) + the await-death precondition.
- tests: close() blocks until process exit; SessionEnd.reason.
CLI 2.1.177 re-probe (folded-in scope): sessions/ registry characterized
(PID-keyed, cleaned on exit); init cache auto-refreshes; advertised
slash_commands show no routing-table drift. No further code change needed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The init-cache/routing-table refresh and sessions/ registry
characterization are now explicit deliverables of T-437, not a
separate follow-up, with updated acceptance criteria.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Regression of T-268's /clear mechanism. /clear deletes the transcript and
respawns with the same deterministic --session-id, but claude 2.1.177 now
tracks session ids in ~/.claude/sessions/<pid>.json (+history.jsonl) beyond
the per-project transcript clide purges, so the id reads as in-use and
claude exits 1 at startup validation. Codebase only probed <=2.1.175.
Hypothesis is strong but unconfirmed: the pane shows an opaque "code 1" and
swallows claude's stderr — capturing it is fix step 1. Filed high with
ranked fixes (mint fresh id on clear; or clean the registry; surface stderr).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
T-425 had a pre-existing breakdown (T-426-430) I didn't check for and
re-filed as T-432-436, which is what got implemented. Cancelled T-426-430 as
duplicates (each notes its implemented twin) and closed the epic — the
crash-survivable logging / observability work is complete.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The boot-time toggle (CLIDE_LOG / app.log.level) existed; this makes it
adjustable at runtime, from the two surfaces that fit — D-6 parity — and NOT a
Claude-composer slash command (log verbosity isn't a Claude-session concept).
- Output dock Level chip: was a view-only filter; now also drives the kernel
Logger.minLevel and persists app.log.level, so the choice is real (changes
what's captured) and survives restart. Initialized from the logger's current
level so it reflects a CLI change. A null callback keeps it a pure view
filter (tests / no kernel).
- `clide log level [<level>]` (lib/src/daemon/log_commands.dart): no arg reports
the level + vocabulary; a valid level sets the live logger and persists;
unknown → bad_arg (code 64), logger untouched. Wired in buildDispatcher with
the kernel logger + settings (captured post-boot).
Tested: the command (get/set/case-insensitive/bad-arg, live + persist) and the
controller (initial level, callback fires, same-level no-op, filter-only when
unwired). Coverage gate 95.05%.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The T-436 soak run came back green but produced no conpty-kill-crumbs artifact:
IsolateCrumbFile opened its append handle without creating the parent dir, and
the standalone orphan probe points it at a fresh CLIDE_LOG_DIR that nothing
else had made — so openSync failed, the crumb file was disabled, and no crumbs
were written. In the app this was masked because FileLogSink already creates
logDirectory() at boot. Create the parent ourselves (no-op when it exists).
Verified by the bundle-smoke artifact, which DID capture a real watchdog sample
(threads=31, handles=25, rssMB=175) because the release app makes the dir.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Closes the observability loop: the log + breadcrumb + watchdog files are now
collected by CI so a wedged run leaves downloadable evidence instead of
nothing.
- logDirectory(): CLIDE_LOG_DIR overrides the per-platform default, so CI can
point the logs at an uploadable workspace dir (and tests at a temp dir).
Now takes an injectable env map; tested.
- test_app.dart: when CLIDE_LOG_DIR is set, the testmode harness tees its
logger to a FileLogSink + spawns the watchdog (off by default — normal
run-testmode keeps the stderr-only path, no isolate). _say breadcrumbs each
test into the file.
- conpty_orphan_probe.dart: with CLIDE_LOG_DIR set it passes a verbose PtyLog,
so when soak-conpty-kill.ps1 force-kills the parent, the reader/waiter
isolates' LAST crumb is fsynced to disk — naming what the wedged isolate was
doing at the instant of death.
- bundle-smoke job: runs the real release app with CLIDE_LOG=debug +
CLIDE_LOG_DIR, uploads clide-logs (watchdog heartbeat/sample + FileLogSink)
in an always() step.
- windows-soak kill-probe job: sets CLIDE_LOG_DIR, uploads the FFI crumbs.
Coverage gate 95.08%.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
watchdog.dart doesn't import pty_log.dart (only references the pattern), so the
[IsolateCrumbFile] doc references couldn't resolve and the dart-doc gate fails
on the warning. Plain-code-font instead of a doc link — no import added just for
a cross-reference.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A main-isolate Timer would freeze WITH the main isolate and tell us nothing,
so the watchdog runs in its own isolate: it fsyncs a heartbeat every ~500ms
(so the last on-disk heartbeat bounds a freeze to ~500ms) and every ~2s samples
this process's thread / handle-or-fd / child-host / RSS counts. A monotonically
climbing child or thread count is the leak signature the soak couldn't
reproduce on CI but a real freeze would show. Output is JSON-lines in
clide-watchdog.log, bounded by the same truncate-on-cap scheme as the crumb
files.
- watchdog.dart (Flutter-free, tested): ResourceSample, ResourceSampler
(forPlatform), PosixResourceSampler (/proc/self: Threads, fd count, task
children, ProcessInfo.currentRss), WatchdogFile (bounded fsynced JSON-lines),
runWatchdog (the loop, bounded by maxTicks for tests), watchdogEntry (the
sendable Isolate.spawn entry).
- watchdog_windows.dart (coverage:ignore — Win32 FFI, validated only at runtime
on Windows): one CreateToolhelp32Snapshot for thread + conhost/OpenConsole
child count, GetProcessHandleCount, ProcessInfo.currentRss. Exhaustively
defensive: any failure yields a -1 field, snapshot handle always closed, never
throws.
- main.dart: spawn the watchdog at boot (desktop only), non-fatal. Per-line
fsync means the OS reaping the isolate at exit loses nothing.
Tests: ResourceSample.toJson, Posix sampler against real /proc, WatchdogFile
(JSON shape, cap, disabled), runWatchdog (immediate baseline tick). Coverage
gate 95.08%.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
