The PR/issue cache only seeds next launch's sidebar, but every flush rewrote the whole
~200 KB orca-github-cache.json. Store one row per entry in orca-github-cache.db, write only
changed rows, and save at most every three minutes plus at quit and explicit flush barriers.
An existing sidecar is imported and removed; older builds read a missing sidecar as empty.
* test(perf): render and identity contracts for agent status, sidebar, tab strip and chat
Twelve operation, render and identity contracts at the production seams of
the hottest UI paths, each driven at N and 10N with one hardcoded bound and
shown to fail with its regression restored. Registers them in the contract
list and fixes the perf workflow's path filter, which had drifted to 12 of
19 contracts. Removes test-only production exports nothing reads.
* test(perf): prove composer typing lands; drop dead rebuild counter
Review: the composer contract never checked the draft reached the composer,
and its transcript control could not fail; the live-index rebuild counter
and a bucket constant had no readers after the test exports went.
* chore(ratchets): count sync fs calls on Electron main and setInterval calls
Sync fs on main and fixed-cadence intervals that never settle were two of the
most common perf fix classes since September, and no gate counted them. Both
ratchets grandfather today's uses per file and fail when a file's count grows.
* chore(ratchets): count every sync fs export, src/shared and node:timers intervals
Review: the allow-list missed descriptor/copy APIs, src/shared helpers run on
main too, and setInterval imported from node:timers escaped the timer count.
* chore(ratchets): keep hosted setInterval calls beside a named timer import
Review: a named node:timers import zeroed window/globalThis calls too, a
destructured require was counted twice, and collector edits didn't select
the timer ratchet in check:changed.
* chore(ratchets): charge calls through an alias, not the alias binding
Decider ruling B on review finding C6: binding a plain alias of a gated call
counted as a use, so an alias refactor failed every collector-based ratchet.
Conditional/fallback aliases and exported bindings stay charged.
* chore(ratchets): count bare setInterval only when nothing local shadows it
Review: the timer ratchet kept its own narrower binding resolver, so alias
and namespace-destructure forms of an imported setInterval counted twice.
Module bindings are the shared collector's; a bare call is the global only
when the file declares no setInterval binding.
* chore(ratchets): resolve bare setInterval calls per scope
Decider ruling A: the file-wide shadow rule let an unrelated parameter or type
declaration named setInterval hide every global poll in the file. Each bare
call now resolves through a single-file TypeScript checker (noLib, noResolve)
and counts when it binds to nothing declared in the file.
* chore(ratchets): write the timer baseline exactly as --prune does
* chore(ratchets): charge global timer aliases and bracket access
CodeRabbit: `const schedule = setInterval; schedule(poll)` and
`window['setInterval'](poll)` counted zero. Count value references to the
global timer (per scope) instead of only direct calls.
* chore(ratchets): charge shorthand timer escapes; skip writes, typeof and JSX tags
Review: `{ setInterval }` of the global counted 0, while assignments to it,
parenthesized typeof probes and a <setInterval/> tag counted 1.
* chore(ratchets): charge compound assignments to the global timer
Review: `||=` and `??=` read setInterval before writing it, so only a
plain `=` target is write-only.
* chore(ratchets): regenerate baselines for files main moved
main's SSH reorganisation moved counted files under src/main/ssh and
src/main/managed-server; counts are unchanged except one sync fs call main
removed.
* Open agent setup with the current model highlighted
* Start setup on the current model only when it is a rendered, non-Arguments row; remount once defaults settle
* Return the setup highlight to the current model when a model search is cleared
* Scope setup highlight navigation to one event and scroll the restored model into view
* Scope setup highlight navigation to the event's phases and reveal the start model when it arrives
A cell that streams hundreds of MiB of output used to grow the renderer
heap without limit and freeze the window. Each cell now keeps the first
2 MiB of stream text, 16 MiB of total output and 2000 outputs, then adds
one notice and drops later output. The first error and the first result
each get a separate bounded allowance, so the error summary still shows.
A display update that does not fit is rejected and the previous display
is kept. clear_output, with or without wait, resets the budget.
markRunCommitted and forgetFinishedRuns move to ipynb-kernel-runs.ts so
ipynb-kernel-session.ts stays under the max-lines lint cap.
Replaces #27721.
Co-authored-by: m4air <m4air@m4airs-MacBook-Air.local>
Co-authored-by: brennanb2025 <79079362+brennanb2025@users.noreply.github.com>
* refactor(main): organize SSH transport and managed server owners
* Remove orphan deploy readiness fixture and shrink test baselines
* refactor(wsl): import lock records from their defining owner
* refactor(wsl): use the canonical deferred update contract
Chat file links share the terminal's row builder, which added the row
whenever the source tab had a split or neighbor. Open file already places
the file beside the chat, so the extra row was noise. Terminal links keep it.
* perf(ai-vault): skip re-rendering unchanged Session History rows
* fix(ai-vault): keep memoized history rows current with host capabilities and time
* fix(ai-vault): recompute chat-resume eligibility when host agent facts change
* refactor(ai-vault): simplify the history row memo comparator
* fix(ai-vault): drop the hook-deps suppression React Compiler bails on
* Let native chat startup follow provider readiness
* Remove obsolete startup deadline expectation
* Preserve startup retry when Claude exits before initialize
* Join startup cleanup with the observed Claude exit
* Model the observed Claude exit in the startup fixture
* Wake Pi startup on fatal session failure
* Keep Pi refusal assertion typed as an unknown result
* Keep startup tests tied to the fixed regressions
Unread rows show the desktop's bell, swipe right toggles read, swipe
left clears a finished agent, and the long-press and options menus get
Mark read/unread, Unread only, Mark all read and Clear finished. Writes
go through agentActivity.setRead/clear and show optimistically until the
desktop's list confirms them.
Workspace and Project add older in-scope sessions by matching agent
folder names to the scope. Several real names were missed: Claude's
truncated long-path buckets, Windows paths in another case, drive and
filesystem roots, the other WSL share spelling, and Pi buckets under
PI_CODING_AGENT_DIR's sessions/ folder. An unreadable folder now shows
up as a scan issue instead of looking empty.
* fix(ai-vault): keep search results on screen while typing in Session History
* perf(ai-vault): fold session search text once and keep resolver maps stable on empty replies
* fix(ai-vault): cut search-index disk writes from merge rewrites and idle vacuums
The session-search index sustained tens of GB of daily writes on busy
multi-agent days (#36666): ~650 KB/s of file-backed pages dirtied by the
Orca Helper process that owns the index, against a ~214 MB database.
Two write paths compound there:
1. FTS5 merges whole segment levels on every steady append. With the
default automerge=4, an index of this size rewrites several bytes of
deep segments for each byte appended, and every merge page is then
copied from WAL into the database by autocheckpoint.
2. Every full purge ran `incremental_vacuum(2000)` even when it deleted
nothing, moving free pages around the file. Full purges back every
freshness wait (`reconcile({ full: true })`) and the periodic sweep,
so idle sweeps themselves became a write source. A measured loop
(one busy transcript appending 100 KB per 20 s cycle, sweeps every
15 cycles) shows sweeps dropping from 428-736 KB to 132-144 KB per
pass and merge spikes from ~1 MB to 432-800 KB (-28% total WAL bytes
at 10 MB index scale; the merge saving grows with index size).
Fix:
- Set FTS5 `automerge` to 8 on every open (new and existing databases,
no schema rebuild): merging wider rewrites roughly half the
deep-level pages on this insert-heavy, rarely-deleted index, at a
small read fan-in cost.
- Gate the trailing `incremental_vacuum` in retention drains on actual
deletion work. Drains scheduled after real deletions (retention,
replacement, retirement) keep reclaiming pages; recurring unlimited
purges with nothing to delete leave the freelist for page reuse
instead of shuffling it.
Read paths, cache semantics, and the on-disk schema are unchanged.
Fixes#36666
* fix(ai-vault): vacuum the search index only when free pages pile up
FTS merges free pages on every write and later writes reuse them, so a
vacuum after every drain batch and sweep only moved pages around. Gate it
on the free share of the file instead of on whether a purge deleted
anything, which also covers drains after whole-session replaces.
---------
Co-authored-by: ggbdpq <ggbdpq@gmail.com>
* feat(agents): send paired phones the desktop Agents list's verdict per agent
While a phone is paired, the desktop window publishes its own Agents-list
classification for each pane (listed, status, unread, clearable, title,
worker parent) and worktree.ps attaches it to each agent row, so the phone
lists agents exactly as the sidebar does. agentActivity.setRead/clear apply
through the sidebar's own Activity actions and resolve once the new
verdicts are stored.
* feat(mobile): add an Agents tab to Home
One row per agent across every paired desktop, grouped by status, with
a desktop filter, search, group-by, nested workers and a long-press
menu. Feed's per-workspace Agents section becomes a short Needs you
list, and the needs-you badge moves to the Agents tab.
While a phone is paired, the desktop window publishes its own Agents-list
classification for each pane (listed, status, unread, clearable, title,
worker parent) and worktree.ps attaches it to each agent row, so the phone
lists agents exactly as the sidebar does. agentActivity.setRead/clear apply
through the sidebar's own Activity actions and resolve once the new
verdicts are stored.
* feat(runtime): tell phones when git status changes
* fix(runtime): tell phones about HEAD and packed-refs moves too
* fix(runtime): don't let a phone's status re-read join a read begun before the change
* perf(runtime): limit status-read invalidation to the checkouts an outside change touched
* fix(runtime): match a checkout's status keys under any execution host
* fix(runtime): invalidate globally while a primary checkout is unread
* test(mobile): re-record the Changes and settings goldens on the rebased main
* fix(runtime): invalidate globally while a listed checkout has no cached path
* test: follow the moved paired-client harness
* refactor(mobile): let the host list's catalog read hold the event router
* fix(runtime): send a terminal's first screen before closing its stream on exit
A PTY exit while terminal.subscribe was still publishing its first snapshot
ended the stream at once: the phone got end with no screen. The exit now
waits for that publication, the exited PTY's model stays readable (off the
live map) until it finishes, and queued layout work settles on exit so it
can't hold the stream open.
* fix(runtime): skip the renderer seed when the PTY exits while its screen is read
* fix(github): stop PR pages from hanging when the rate-limit cache expires
Slot holders await assertRateLimitBudget -> getRateLimit while holding a
gh semaphore slot, and the /rate_limit probe queued for a slot behind
them. With all four slots held by budget waiters nothing ever released.
The probe is quota-exempt and deduped, so it no longer takes a slot.
* fix(github): bound rate-limit probe concurrency without the gh semaphore
Forced getRateLimit calls now join an in-flight probe, and all /rate_limit
probes share a small probe-only limiter that acquires nothing else.
Fold the deadlock regression test into rate-limit.test.ts as a direct await.
* fix(github): make forced rate-limit probes read quota after the call
Forced callers (403 reset refinement) share one follow-up probe queued
behind the running one instead of joining a probe that may predate the
403, so a stale positive budget cannot clear a fresh breaker block.
* fix(github): hand forced rate-limit follow-ups off synchronously
A pending forced follow-up could outlive its probe's settlement, so a
later forced caller joined a probe that predated it. Settlement now
detaches the shared deferred follow-up and starts its probe in one
synchronous step. Run rate-limit.test.ts on the Node runtime, where the
V8 scheduling that opens the race is reproducible.
* Open search-result agent setup as a hover submenu
The + menu's search-result agent rows opened setup through a separate
click popover. Route them through the same native submenu the regular
agent rows use, delete the popover path, and keep hover from moving
focus off the omnibox input.
* Keep omnibox typing in the search box when agent setup opens or closes
Hovering a search row's › no longer moves focus into the setup panel, and
Escape, ArrowLeft or leaving the open row return focus to the search input.
Also stop dimming disabled setup rows twice.
* Give omnibox agent setup pointer grace before another row closes it
Crossing an ordinary result on the way to the panel no longer closes the
setup; another row closes it after 300 ms unless the pointer reaches the
panel or returns to its row, like Radix submenus.
* Keep a hover-opened omnibox setup from taking focus after a keystroke
Typing can shift a row under a resting pointer and open its setup while
Radix is in keyboard mode; its content autofocus now returns to the
search input until the pointer moves into the panel.
* Open omnibox agent setup on pointer movement, and keep panel typing on leave
Per the decider ruling: a row shifting under a resting pointer no longer
opens its setup (Radix sub-triggers also open on pointermove), and leaving
the row after clicking › keeps focus in the setup editor instead of moving
typing back to the search box.
* Return omnibox focus to the search box when a focused setup closes on its own
Clicking › and then moving to another row let the hover grace close the
panel while it held focus, dropping focus to the page; closing it now
returns typing to the search input.
* Test that a caller-closed setup returns focus to the omnibox input
* Render omnibox search tests inside the + menu so agent setup has its menu
* Keep omnibox setup wiring React Compiler friendly
Build each row's setup handlers at the call site from plain hook state, and
configure the launched option inside the handler, so nothing that reads refs
runs during render.
* Open the branch picker from the branch name under the chat composer
* Copy the current branch from its row in the composer's branch picker
* Keep the composer copy button clickable and out of the list's keyboard, and keep a twin copy's status out of the picker
* Read a twin copy's status privately in the branch picker and drop refreshes that land after its twin is selected
* Block on a twin's returned read error, re-read after the picker closes itself, and distrust a twin's leftover status on becoming owner
* Build the environment name selector outside the component so the React compiler can lower the picker
* Trust a copy's status again once its picker refresh settles, even when an identical newer poll superseded it
* Read a newly owning copy's own status until the inherited slot changes, instead of trusting a settled refresh
* Route the composer picker through the copy's owner and block while this chat's turn is active
* Keep the composer picker's agent gate on the shared working check, as in Source Control
* Show CI result lines on the phone, following the desktop setting
The host now sends nativeChatCiResultLines resolved in settings.get, the
ready snapshot and a nativeChatCiResultLinesChanged event, from one
publisher that tracks each pushed chat preference separately. The phone
keeps one store per desktop connection and preference, reads the CI value
only from the paired desktop, and hides the lines until that desktop
reports one. CI rows stay in the projection and are filtered after it, so
a flip rebuilds nothing. The phone draws the line with links to the pull
request and commit.
* Let the phone CI line wrap and select like other notices
Long repository names at large font scale no longer push the commit link
past an ellipsis: the line wraps, centered, with the icon on its first
line. Its text follows the system-notice selection policy, so iOS selects
natively and Android keeps the long-press actions sheet. The CI strings
are translated in every mobile locale, the filter copies only once it
meets a CI row, and the tests drop a token assertion and a duplicate case.
* Draw CI results as one-line cards on desktop and phone
The CI result line becomes a one-line card in the same frame as the
opened pull request card, with the Hide button inline at its right end
on desktop, and in the phone's chat card frame without one. Only the
lead and pull request name truncate, so the commit link and the Hide
button stay reachable on narrow widths. The fixture uses real
percent-encoded item ids so its links render.
* Make the CI card's whole lead open its pull request
A translated lead at large font scale can push the ellipsis before
repo#N, which left no visible way to open the pull request. The
truncating segment is now the pull request's target on desktop and
phone, named for it, while "at <sha>" stays the commit's own target.
The ja, zh and zh-Hant leads use a plain colon so the space after it
isn't doubled.
* Read the CI result lines setting only for structured chats
CI result lines come only from structured sessions' journals, so a
terminal chat no longer sends the setting read. The prompt-controller
fixture stubs the setting hook: its tests count only the chat's prompt
writes, and the structured chat's one read per desktop connection is
expected.