Skip to content

chore: rebase TickTick takeover branch onto main (resolves #29289 conflict) - #29626

Closed
martchellop wants to merge 236 commits into
kirodotdev:feat/connections-ticktickfrom
martchellop:feat/connections-ticktick
Closed

martchellop wants to merge 236 commits into
kirodotdev:feat/connections-ticktickfrom
martchellop:feat/connections-ticktick

Conversation

@martchellop

Copy link
Copy Markdown
Contributor

Resolves the merge conflict in #29289: feat/connections-ticktick rebased onto current main (e71bf3d85e), with _PACKAGE_LINE_BUDGET in test/test_security_gate_liveness.py set to main + 1 (34_313) and both #: notes kept. The tree is otherwise identical to c15ea302c4 rebased (git range-diff shows only the budget hunk and author metadata). Merging this into the takeover branch makes #29289 mergeable; no review needed beyond the budget line. Fork PR #26407 carries the same tree.

Rohan Khanderia and others added 30 commits November 3, 2026 11:33
…te (kirodotdev#29176)

The custom-theme editor threw four hard-coded English messages for
JSON and name validation, and useTheme fell back to an English
"Failed to create theme" when the server gave no reason. All five
reached the error notice in every language. A JSON value such as
null passed the parse step and then surfaced a raw TypeError.

The refusals now read from components.themeEditor catalog keys in
all twelve languages and render as a plain status hint instead of the
error notice, which stays for real save failures. A non-object JSON
value gets the "needs a name" hint, and the create fallback reuses
failed_to_save_theme.

The colour-group headers carry aria-expanded and aria-controls, the
up/down arrow is a hidden lucide chevron instead of a text glyph, and
the picker / JSON / CSS mode buttons carry aria-pressed.
kirodotdev#29133)

Each step of the memory backup restore flow unmounted or disabled the
control that had focus: Restore disables itself once armed, Cancel and a
staged restore unmount the confirm strip, and Confirm is disabled while
the request runs. Keyboard and screen-reader users were left on the
page body after every step. Cancel also left a refused attempt's error
on screen for a restore nobody was asking for any more.

Move focus with the flow: arming focuses the strip's Cancel, Cancel
returns it to the row's Restore, a refusal hands it back to the strip,
and a staged restore moves it to "Cancel staged restore". The strip's
buttons carry its warning as their description. Cancel clears the
refused attempt's error.
kirodotdev#29122)

The polling loop dispatched each update inside a single `for` guarded
only by the loop-level `except Exception`, so a malformed envelope that
made `_build_inbound` raise (a null `message.from`/`message.chat`, a
non-numeric `reply_to_message.from.id`, or a non-list `photo`) aborted
the loop and left every LATER update in the same batch undispatched.
The offset had already advanced on observation, so Telegram never
redelivered them: the batch-mates were lost.

Guard each `_dispatch` on its own (log and continue), and make
`_build_inbound` read `chat`/`from` as `or {}`, coerce the reply
sender id defensively, and ignore a non-list `photo` -- so a malformed
field parses to a valid-shaped inbound (an absent sender reads as
user_id=0, which deny-by-default auth refuses) instead of raising.
…dotdev#29119)

The Schedule page's Next run column built "in <1m", "in Xm",
"in Xh Ym" and "in Xd Yh" from English templates, so every language
showed English. The value goes through fmtRelative, the repo's
locale-aware relative-time helper, in the largest unit that fits
("in 2h", "in 2 Std.", "2小时后"). fmtRelative takes a numeric option,
and the column passes "always" so a run 24-48 hours away reads
"in 1d", never "tomorrow".
…otdev#29116)

* fix(pi): write the gate launcher .cmd in the console code page

cmd.exe reads a batch file in the console code page, which is the
system OEM page by default, not UTF-8, so the pi gate launcher written as UTF-8 under a non-ASCII path (a
profile folder named José) named a misread pi binary or gate extension
and the session could not start pi.

The launcher is now encoded strictly in the OEM code page on Windows
(UTF-8 elsewhere) and written as bytes, the rule pod/windows.py and the
PPTX Maker pdftoppm.cmd launcher (kirodotdev#28972) follow. A path the page cannot
spell refuses the spawn with an AcpError naming the encoding, before any
launcher file is written.

Fixes kirodotdev#28963

* style(test): black-format the pi gate launcher tests

The Backend Lint black gate flagged test/test_acp_pi_backend.py: two
calls wrapped across lines fit within the line length. Formatting only.

Part of kirodotdev#28963

---------

Co-authored-by: Matt Briones <mdbrio@amazon.com>
…lacing them (kirodotdev#29067)

The upload name sanitizer replaced every character outside \w, . and -
with _. A macOS browser sends names in NFD, so an accented letter
arrives as a base letter plus a combining accent, and \w does not match
the accent: cafe plus U+0301 was stored as cafe_.txt.

Normalize the name to NFC first, so a composable accent is one letter.
A mark with no composed form is still replaced as before, because the
image-path grammar also reads \w and a stored path it cannot match would
keep an uploaded picture from reaching the model.
…ns (kirodotdev#29082)

On the import results step, a failed conflict retry replaced the counts
with zeros, and Import another and Continue stayed clickable while the
retry ran. The last landed import now stays on screen under an error
notice, and both buttons wait for the retry.

Scan skip reasons showed as raw backend codes. They now read in words by
kind (credentials, too large, not supported, excluded, path not allowed,
invalid, other) in every catalog; the code stays on data-reason.
…v#29246)

An app installed, updated or removed outside the tab (the CLI, another
tab) fired no mc:apps-changed there, so the nav rail and link previews
kept the old list until a reload. Re-read the shared ['apps'] list on
focus and visibilitychange, skipping focus that returns from the page's
own iframes and folding one return's paired events into one read.

Closes kirodotdev#28740
…hat text (kirodotdev#27700)

* feat(apps): add contributes.selectionActions on the shared app frame

* fix(apps): show the Apps button when the app list fails, and bind selection mints to the current grant

The selection toolbar's Apps button now hides only while a read of the shared app list is in flight or after a successful read with no approved item. When that read failed, or the dashboard shell's own nav read has not landed, it shows (visibility only); the menu's own fresh read then lists the items or shows the error row with Retry, so a failure is never silent. Items still come only from a successful fresh read. In the dashboard shell the layer no longer fetches the list itself (readApps={false}); the popped-out and embedded chats still do.

Every selection mint and renewal names the grant entry the item was resolved at, and the gateway refuses it unless that is the item's current approved entry. A refused first mint says the item was revoked; a renewal refused with 403 or 404 cancels the live invocation with 'revoked' and removes the frame.

App items get the selected characters as shown, never a link card's hidden address in place of its title words. The feature-map row describes the single Apps button. App.appsMountRead.test.tsx mocks the lazy overlay layers so its nav-read count does not depend on chunk timing. The pseudolocale is regenerated.

A held selection-action grant entry on the App Detail page is now shown in words (the item title, its text limit and its buttons) instead of the raw entry string, in the list and in the approval confirmation.
The toolbar re-measures its docked position when the app row appears while it is open, so the Apps row never hangs over the composer.
…irodotdev#29035)

_rewrite_update_frontmatter stripped every frontmatter line before
reading its key and dropped blank lines, so an indented line inside a
block-scalar description that began with name:, version:, pinned: and
the like was removed, and paragraphs were joined. Its LF-only fence
also missed a BOM or CRLF candidate, which then kept its whole old
frontmatter inside the new body.

Filter only column-0 keys, keep blank lines inside the block, and
accept a BOM and CRLF fence.
…irodotdev#29020)

json.loads raises RecursionError, not ValueError, on a line nested past
the interpreter limit. The session digest scanners promise never to
raise, and the history search corpus skips unparseable lines, but all
four caught ValueError only. One such line in a transcript or a kiro-cli
event log made the session-storage detail route answer 500 and the
history search raise. They now catch both and skip the line.
…irodotdev#28961)

The shared ingest took everything after the last dot as the temp-file
extension, so a dotless name ("README", "voice") became the whole
extension, and a long one made the temp path too long to create, which
reported "download failed". WeCom named an unnamed video or file
".video" / ".file".

attachment_suffix picks the name extension, then the one the mimetype
implies (audio/ogg pinned to .ogg), then .bin; safe_suffix caps any
hint at 16 characters. The shared ingest and WeCom use it.
)

* fix(apps): refuse app names longer than 128 characters

An app name of 251 or more characters passed app_name_error, and then
install failed creating the <name>.lock record lock with an OSError, so
POST /api/apps/install answered 500. The name is part of several file
names the gateway builds, the longest being
.<name>-deps-doomed.kirocrew-deps-staging-<pid>-<hex> (name + 56), and
common filesystems cap one component at 255.

app_name_error now refuses names over APP_NAME_MAX_LENGTH (128) with a
message that names the limit, so every admission door refuses them and
the install route answers 400. 128 leaves room for every derived
component with headroom; raising the cap later costs nothing.

Fixes kirodotdev#27872

* fix(apps): re-export APP_NAME_MAX_LENGTH from the manifest facade

kiro_crew.apps.manifest binds every name its owner modules define, and
test_manifest_facade checks that. Bind the new cap there too.

Part of kirodotdev#27872

---------

Co-authored-by: Matt Briones <mdbrio@amazon.com>
…irodotdev#28996)

On the Workflows Runs view the Cancel and Save outcomes were held once
for the whole view. A failed Cancel showed its error in the detail pane
of whichever run was open and kept showing it after another run was
opened; with two cancels in flight only the latest row read Cancelling,
and an earlier failure was lost. A save that finished after the reader
opened another run labelled that run, a second save replaced the first
run's /workflow name, and a save that failed after its dialog was closed
said nothing.

Keep these per run id. Every in-flight cancel keeps its row busy; a
failed cancel opens its own run and shows the error there. Each run
keeps the /workflow name its own save produced, and a save that failed
after its dialog closed shows a notice on its run.
…otdev#28894)

The cron guide said the zombie reaper uses a 30 min deadline. The
reaper reads each job's timeout_secs, capped at 24 hours and never below
30 minutes, plus the queue, gate and vet allowances for script and
command jobs.

Fixes kirodotdev#28890
kirodotdev#28897)

save_progress maps each task status to an icon, and the map had no
entry for TaskStatus.CANCELLED. Cancelling a run marks its unfinished
tasks cancelled, so the write raised KeyError inside the writer's
broad except and was only logged at debug: TASK_PROGRESS.md kept the
last running snapshot, and every later save of a run that still held
a cancelled task was dropped the same way. Cancelled tasks get their
own icon, and a test pins every TaskStatus member.
…irodotdev#28948)

The download Range parser read positions with int(), which also takes
"1_0", "+5" and non-ASCII digits, and it answered a malformed header
(reversed range, multi-range, unknown unit) with 416. RFC 9110 says a
malformed Range is ignored: the server sends 200 with the whole body.
The parser now takes ASCII digits only and tells the two cases apart:
a well-formed range past the end is still 416, a malformed one is
ignored. Every valid range is answered exactly as before.
…rodotdev#28881)

The Commands section sends readers to "Who can reach it", a heading the
WeCom guide does not have. The group-refusal rule it means lives under
"## Access control", so the note links there.
…kirodotdev#28880)

head_content split the header on "filename=" and stripped quotes, so a
later parameter leaked into the name ('report.pdf"; size=123') and the
RFC 5987 form filename*=UTF-8''... read as empty, leaving the attachment
named "attachment". The value is parsed as header parameters with the
stdlib email parser: a quoted name ends at its quote, filename* is
decoded in its charset, and it wins over a plain filename as RFC 6266
asks.
… unattended turns (kirodotdev#27745)

A cron turn now hands the PreToolUse gate its own session key, so a
governance profile bound to that session's surface governs it the way it
governs a chat turn.

When the governance tier's own evaluation raises, a turn nobody is
watching (cron, heartbeat, a monitor wake) now refuses the call instead
of allowing it. An attended turn keeps "no governance opinion", so a
transient fault never refuses an ordinary user action. Both paths log
the error.

Co-authored-by: w <w@k.local>
Luca Chang and others added 19 commits November 4, 2026 10:24
kirodotdev#29469)

The composer box unmounts for the approval ghost and for a collapse,
and mounts a new textarea when it returns. The autosize effects read
inputRef only when a dep changes, and a remount changes none, so the
new textarea was not measured and the width observer stayed on the
detached one. Add the box's mount state as a dep of both effects.
…irodotdev#29524)

Tests in one file whose bodies were the same AST become one parametrized test.
Every old case stays as a param whose id carries the old test name. Three pure
copies are deleted; each names a twin that still runs the same check.

Part of kirodotdev#29521

Co-authored-by: zejiangg <zejiangg@amazon.com>
…plies (kirodotdev#28655)

`parse_ics` in the Meetings calendar provider appended in-window events in
feed order and stopped reading at `MAX_CALENDAR_EVENTS` (500), and only
then sorted by start. RFC 5545 does not require VEVENTs in time order, so
a feed with 500 later events ahead of a nearer one returned the later 500
and dropped the nearer meeting from the dashboard and the poller.

The parser now reads the whole size-capped document and keeps the nearest
500 events in a bounded heap keyed on start, with feed order kept on a
tie, then returns them sorted by start. `MAX_ICS_BYTES` (4 MiB) admits
about 85,000 minimal VEVENTs, so the heap keeps memory at the cap instead
of holding every event. An event no nearer than the latest kept one skips
`_finalize_event`, so a chronological feed does not pay for redaction and
hashing of events that cannot be kept.

Closes #ISSUE
…kirodotdev#29549)

Python's _handle_fromlist reads __path__ on a module before iterating
a from-import's name list. For a MirroredModule facade that is not a
package, __path__ is absent from __dict__, so _handle_fromlist reaches
MirroredModule.__getattr__('__path__'): one extra facade call per
from-import statement.

Two production sites fire per slot on the test thread:

1. slots_frame._slot_links (slots_frame.py:609): from chat_utils import
   effective_session_key, mirror_is_paused, slack_mirror_is_paused.

2. slot_projection.SlotProjection.to_dict (slot_projection.py:264):
   from chat_utils import chunk_generation.

Both are deferred inside function bodies to avoid a circular import at
module load time (chat_utils -> dashboard.state -> slots_frame /
slot_projection). The constraint applies only at load time; at call time
the module is already in sys.modules. Switching from 'from X import Y'
to 'import X as _cu; _cu.Y' skips _handle_fromlist entirely. Python's
'import X' resolves sys.modules['X'] without calling _handle_fromlist.

A third source of nondeterministic __path__ reads: the maintenance-pool
thread. _request_lineage_seed() submits _seed to a thread pool when the
session-tree projection is not yet seeded. In tests the seed attempt
fails fast (no crew log on disk) and resets the in-flight flag, so the
next flush submits another seed. The facade_reads fixture patches
MirroredModule.__getattr__ at class level, making it active on every
thread. The background thread fires __getattr__('projection') (or
__getattr__('__path__')) while the measurement counter is live -- reads
that belong to background work, not the synchronous flush cost the
budget tracks. On Windows the thread races the measurement window
near-deterministically (PR kirodotdev#29454, 2/2 failed attempts).

Fix: thread-filter the counter in facade_reads. Capture
threading.get_ident() when the fixture installs the patch; only reads
from that thread are counted. This closes the class for every budget
test in the file. The two production import conversions stand: they are
real hot-path improvements (2 fewer facade calls per slot) and the
ceiling correctly reflects them at 4.

Result: per-slot facade read count drops from 6 to 4 (two __path__
probes removed); background-thread reads excluded from measurement;
test ceiling and comment updated.

Co-authored-by: CrysisDeu <crysisdeu@users.noreply.github.com>
…h ShipIt (kirodotdev#29449) (kirodotdev#29551)

Electron's native quitAndInstall emits before-quit-for-update, closes every
window, and only once the last one closed asks Squirrel to launch ShipIt and
quit the app. The main window and the Mochi panel learned about a quit from
before-quit, which comes after the closes, so the installer's close hid them
to the tray, the window list never emptied, and the 5s failsafe armed on
before-quit-for-update force-exited the app before ShipIt had the request.

- main.js and the Mochi panel set their quitting flag on
  before-quit-for-update (close-action.js onQuitForUpdate).
- On darwin the exit failsafe now arms on Squirrel's own quit; the 5s timer
  destroys windows still open instead of exiting.

Closes kirodotdev#29449
…ent (kirodotdev#29491)

The stall case in test_an_ineligible_crew_is_skipped used a real
200 ms wall-clock probe timeout (PROBE_TIMEOUT_S=0.2).  On a loaded
CodeBuild shard the event loop can be starved for longer, so crew
'b' also hits the timeout and placement reports no room instead of
placing on 'b' -- the over2/over3 variants.

Fix: split the stall variant out of the parametrize list.  The eight
data-only ineligibility cases (disconnected state, old version, bad
headroom fields) need no timeout at all -- their probes return
immediately with ineligible data.  Remove the PROBE_TIMEOUT_S patch
from that test.

Add test_a_stalling_crew_probe_is_skipped, which stubs _candidate to
return None for crew 'a' (the same value production returns on a real
timeout), so the test is driven entirely by data and never races the
wall clock.  Replace asyncio.sleep(60) in the Crews stub with an Event
that never fires, keeping the stall path clean for future use.

Co-authored-by: CrysisDeu <crysisdeu@users.noreply.github.com>
…elative cap (kirodotdev#21098)

* fix(schedule): give the job detail dialog's run surfaces a viewport-relative cap (kirodotdev#21061)

The job detail dialog's Last Error notice is a direct `overflow-y-auto` flex
child of the column-flex DialogBody, which gives it an automatic `min-height: 0`.
The column squeezes it to its content's min-content height (~16px, one line)
regardless of its `max-height`, so a multi-frame traceback -- the common failed-
cron case -- collapses to one clipped line and the line naming the failure is the
half the reader loses.

The fix is `shrink-0` (the box takes its content height instead of collapsing)
plus a viewport-relative `max-h-[min(45vh,420px)]` cap (bounded by the dialog's
own `max-h-[86vh]`, still internally scrollable). The sibling Last Output `<pre>`
gets the same cap and `shrink-0`, plus a stable `data-testid` so the layout is
asserted directly; the repo's PNG fixtures are replaced by a scripted layout
verification under website/scripts.

* fix(schedule): drop the redundant Last Error wrapper; cap only Last Output

Per the First Principles review: main already caps and floors the Last
Error notice (kirodotdev#21113). Keep only the Last Output cap, drop the block
wrapper, its structural test, the harness's hostile mode and the
test:layout npm script.

---------

Co-authored-by: EC2 Default User <ec2-user@ip-172-31-10-71.ec2.internal>
…irodotdev#29504)

A Slack thread's plain auto-nudge wake streams with stream_and_collect,
which never consumed session-directive markers. autonudge_stop,
monitor_update and nothing_to_do called from inside the wake turn
reported "Stop REQUESTED" while the effect was dropped, so the loop
kept firing.

Move TurnDriver's per-turn directive bookkeeping (forgery gate,
sub-agent isolation, single-consume) into messaging.directive_tracker
unchanged, let stream_and_collect feed it through a new on_event hook,
and wire it into the Slack wake. Both Slack wake branches now pass the
loop id to build_directive_consumer, so the applier treats the turn as
the loop's own wake, as the dashboard wake already does: a quiet end
is honoured and a stop is checked against the loop that woke it.
…iters (kirodotdev#29267)

* fix(windows): retry the final rename in the remaining tmp + rename writers

Twenty-two writers still published their temp file with a bare os.replace
(or Path.replace). On Windows a reader holding the destination open fails that
rename with PermissionError, and the write was lost on the first fault. They
now rename through atomic_write.replace_with_retry, which is a plain rename on
POSIX and on the event loop.

The Mochi queue and the Weixin state writers also remove their temp when the
publish fails, so a hold that outlasts the retry no longer leaves a leftover.
The sage store binds the retry in a guarded import so it stays importable
without kiro_crew. The platform-compat spec gains the write-side row.

Closes kirodotdev#18460

* fix(windows): re-check the destination before every retried rename

The descriptor-pinned source write and the transcript repair check the
destination is still the file they read just before renaming. Through the plain
retry that check ran once, so a save landing while a scanner held the temp
would be overwritten by the next attempt. atomic_write gains
replace_with_retry_checked, which runs the check before every attempt, and both
writers use it.
…t pass (kirodotdev#25599)

An Outlook on the web mail or calendar item id is the base64 of an
Exchange id with its +, / and = percent-escaped, and a Slack client
link joins its two ids and its message stamp into one 40+ char run, so
pass 3 masked both as a bare AWS secret key.

They are now exempt by structure, with no pattern: the link is cut out
of the text by fixed character sets, split by urllib.parse.urlsplit,
its host compared exactly against a fixed set, its path segments
compared as exact tokens, and its ids checked character by character.
An Outlook id must also decode to one whole Exchange id, and no piece
of the link may be a whole key, so a key as an id, glued to one,
escaped, between escapes, after the route or in a crafted header is
still redacted. Every other link and pass is judged as before.

Co-authored-by: Zezhen Xu <zezhexu@amazon.com>
…irodotdev#29578)

The Discord and Telegram client/renderer contracts and the Discord/Webex
transport-dispatch contracts repeated the same check bodies once per
file. test_facade_shared_checks.py now holds each body once and also
reads opt-ins from test_*composition*.py files, so every old case runs
there with one param id per file.

Part of kirodotdev#29521

Co-authored-by: zejiangg <zejiangg@amazon.com>
…el tests (kirodotdev#22815)

_eager_spawn returns early when the module-global _eager_spawn_sem is already
held, so this class silently depended on no sibling pre-warm being in flight in
the same xdist worker.

Co-authored-by: m <m@e>
…kirodotdev#29555)

Each semantic memory value in the Vector Memory card now renders as a
button named "Edit value of “<key>”", with the value text as its
description. Enter or Space on it opens the inline editor, through the
same cell click handler a mouse click uses, so mouse behaviour is
unchanged. Enter, Escape and the editor's Save and Cancel buttons hand
focus back to that button after the editor unmounts, unless the user has
already moved focus elsewhere (the MeetingTitle requestAnimationFrame
pattern). The icon-only Save and Cancel buttons and the editor input get
accessible names. Strings are translated in all 12 locales.

Closes kirodotdev#26667
Methods copied between sibling test files (the built-in dashboard
templates, the GPT and fork-verdict review workflows, the eval-subagent
and task-runner deny notices, the member dispatch and panel mounts)
move into mixin classes in helper modules. Each class that held a copy
takes the mixin as a base, so every node id, class setup and fixture
stays the same.

Part of kirodotdev#29521

Co-authored-by: zejiangg <zejiangg@amazon.com>
Add TickTick's hosted MCP server (https://mcp.ticktick.com) to the
Connections registry as a tier-2 project-management entry with
launch_gate_passed=false, so the Connect grid does not change until the
manual launch-gate check passes. Sign-in is OAuth with dynamic client
registration and PKCE, so no operator app and no per-host install.

The entry requests tasks:read plus tasks:write, the whole scopes_supported
list the resource metadata publishes; TickTick offers no read-only MCP
scope and the card copy says so and names the server's delete tools.
TickTick's consent URL (ticktick.com/oauth/authorize) joins the OAuth
banner allowlist and the URL corpus.

TickTick ships bare search and fetch tools like Atlassian, Neon and
Todoist, and a list_projects like Linear and Vercel, so it declares
ticktick_search, ticktick_fetch and ticktick_list_projects and joins the
bare search/fetch group in the alias decision table. The user doc counts
30 providers and eight rename declarers; the spec roster gains a row and
ProviderLogo.tsx maps logos/ticktick.svg (the CC0 simple-icons mark).

Closes kirodotdev#26388
@martchellop
martchellop requested review from a team and KiroCrewCaptain as code owners November 4, 2026 14:27
@martchellop
martchellop requested review from iamwhatever and removed request for a team November 4, 2026 14:27
@github-actions github-actions Bot added fork Pull request from a fork (external contributor) readiness: action required A blocking check or review needs attention labels Nov 4, 2026
@github-actions github-actions Bot added the merge conflict Branch has merge conflicts with its base — author must resolve before merge label Nov 4, 2026
@martchellop

ghost commented Nov 4, 2026

Copy link
Copy Markdown
Contributor Author

Superseded by #29627; this one conflicts because of the rebase (single commit on new main vs c15ea30 both edit the budget line); see the new PR.

@martchellop martchellop closed this Nov 4, 2026
@github-actions github-actions Bot removed the readiness: action required A blocking check or review needs attention label Nov 4, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

fork Pull request from a fork (external contributor) merge conflict Branch has merge conflicts with its base — author must resolve before merge

Projects

None yet

Development

Successfully merging this pull request may close these issues.