Skip to content

fix: cherry-pick the crewmate launch list into release/0.11.0 - #25643

Closed
bolichen97 wants to merge 43 commits into
cherry-pick/0.11.0-crew-log-kasfrom
cherry-pick/0.11.0-ray-crewmate
Closed

bolichen97 wants to merge 43 commits into
cherry-pick/0.11.0-crew-log-kasfrom
cherry-pick/0.11.0-ray-crewmate

Conversation

@bolichen97

Copy link
Copy Markdown
Contributor

Stacked on #25642. Cherry-picks the 0.11.0 insider (crewmate launch) list from main into release/0.11.0: 42 of the 44 PRs proposed for it, plus one test-helper commit.

Picked (in merge order, git cherry-pick -x)

Conflicts resolved by hand

Left out

Tests (on this branch's head)

  • Frontend vitest: 3,310 files, 54,644 passed, 0 failed. Electron: 2,865 passed, 0 failed. npm run build ok.
  • flake8, mypy, black, subprocess-encoding, docs-lint and the harness gate pass.
  • Backend full suite: 216,793 passed. 118 failed and 65 errored, and the same tests fail on a clean release/0.11.0 and on main on the same host (the profile resolves to enterprise, and /local/home reads as a sensitive path). Five tests failed only under full-suite load. They pass alone on both this branch and release.

buluoray and others added 30 commits October 29, 2026 01:27
… on the reading (#23912)

When two rows share one identity and either has no start time, recency
cannot order them and the more conservative bucket is kept. The reading
now says so: it reports partial with a note naming each check whose
bucket was chosen, as the spec already described. Rows that agree add
no note, and checks_complete stays true because every row was read.

Closes #21812

(cherry picked from commit 95640ac)
…ons (#23915)

The optimizer prompt told the model to add scope, constraints and
uncertainty notes, and two of its examples added requirements the draft
never named (a launch plan's timeline and risks, exponential backoff).
Rewrites copied that: a precise function request gained validation,
docstrings and examples, a bug report gained the model's own diagnosis,
and an open "explore" request became a fixed list of outputs.

The rules now say a rewrite asks for what the draft asks for and nothing
more, leaves the task to the receiving agent (no answer, diagnosis or
partial solution), and keeps an open request open. What it may change is
limited to hedge removal, one numbered step per compound ask, and the
existing preserve-behaviour clause. The examples follow that contract,
and TestOptimizerPromptContract checks every example rewrite uses only
its draft's words plus the preserve clause.

Closes #21692

(cherry picked from commit 1f0e872)
…22865) (#23920)

When several connections opened the same new memory file at once, some
failed with "database is locked" at PRAGMA journal_mode=WAL. Leaving a
rollback journal for WAL takes an exclusive lock by upgrading a read lock,
and SQLite reports a conflict on that upgrade at once without calling the
busy handler, so busy_timeout did not cover it.

The new kiro_crew.sqlite_wal.enable_wal reads the journal mode first and
returns when the file is already in WAL, which takes no write lock. Otherwise
it retries the switch on a lock conflict with a short doubling wait, until the
store's own busy timeout, and raises anything else at once. The memory store
open now sets busy_timeout first and calls it; the member database, history
index, knowledge store, skill search index and research campaign opens use
it too. The knowledge store waits at most its own 10 s busy window, since
some of its callers still open it on the event loop.

Closes #22865

(cherry picked from commit 03f3b2a)
Two crewmates handed one repository shared its checkout: one committed the
other's edits, or landed a commit on the branch the other had checked out.
The gateway now does only what a model cannot do for itself:

1. Gives each crewmate one worktree per repository, at
   <workspace>/worktrees/<slug>/<repo>-<hash>, on crew/<slug>/<repo>, cut
   from the default branch the first time and reused after that.
2. Points a worker turn's slot.project at it, from the paths the turn names
   or the slot's project, before the provider is claimed. The repository's
   aws.project_profiles binding still applies there: a cwd in a crewmate
   worktree is matched as the same path in the checkout, asked of git each
   time (source_path), so it holds after a restart.
3. Removes those worktrees when the crewmate is deleted, keeping one with
   uncommitted or ignored files; branches and the repository's own checkout
   are never touched.

Git runs through the follow-up card's sandboxed runner (no hooks, no
fsmonitor, checkout filters refused), and the folder is made and unwound
through no-follow descriptors. Topic branches, rebasing onto the default
branch and checking with merge-tree before a push are the worker's own,
asked for in worker.principles.

(cherry picked from commit 6b21419)
On a phone the Crewmates single bar keeps a way to the navigation while a
crewmate is read, shows the update notice with its text, and wears bare
icons like the Sessions bar. The update pill asks the page before it
leaves, so a New crewmate draft is not dropped.

Closes #19589

(cherry picked from commit a6150d5)
…#23916)

The Discord, Teams and Webex queue pumps had the same loop Telegram
had before #23814. When a channel's turn ends, its drain replays the
queued messages. If another turn took the session lease in between (a
dashboard, cron or auto-nudge turn on the same key, or another
channel's turn under a unified dm_scope), the replay reached the busy
path, which queued the same text again, and the pump dequeued it
straight back. Every lap flipped the receipt to "Now answering" with
no turn started.

Each pump now checks is_busy under its queue lock before it dequeues
or flips anything. When the session is held it leaves the queue in
order and calls messaging.queue_drain.drain_when_free, which runs the
channel's registered drain once the lease is free.

Each client's close now also retires the dispatcher's queue drain
(compare-and-drop), so a retry that outlives a disabled channel finds
no drain instead of answering on it. Discord already had an on_close
hook for its spawn-approval delivery; Teams and Webex gain the same
hook, called first in close().

Closes #23749

(cherry picked from commit 1f3210d)
On macOS the agent.spawn_min_memory_gb floor compared against the loose
reclaimable sum (free + inactive + speculative + purgeable). That sum counts
speculative pages twice (free_count already holds them) and counts dirty
anonymous inactive pages as free, so on a loaded Mac it read about 3.5 GiB
above what a new process can take and the floor admitted starts with less
headroom than the operator set.

check_memory_available now reads platform_compat.host_available_mib on
macOS through the same GiB reader Windows already uses, renamed from
_windows_available_memory_gb to _host_available_gb since it serves both.
The memory-sized parallel cap, the TaskRunner auto value and the memory
posture keep the loose sum; the specs say so and a test pins both readings.

Closes #21574

(cherry picked from commit ed83ad2)
A crewmate's worker now runs on every host whose worker session is
captured live: ACP_BACKENDS_MATE_WORKER_KIT widens from KAS to KAS and
claude. codex and deepseek build their array the same way but have no
live capture of it yet, so a worker there is still refused. kiro-cli
stays out, because it reads --agent off disk from ~/.kiro/agents/, where
the worker spec is not. A worker anywhere else is refused with the member
chat's code (capability_harness_unsupported), worded with the worker's
own remedy. A crewmate whose file turns off single tools is refused only
outside ACP_BACKENDS_MATE_WORKER_TOOL_LIMITS = {kas}: KAS carries the
limits as the worker spec's excludedTools, while claude mounts servers
whole (capability_tool_limits_unsupported). claude builds the worker's
MCP array from its own spec, read by the session through
session_mcp_projection; KAS is unchanged.

The single-tool check counts the crewmate file's own excludedTools as well
(a server or a native tool, as a copied kit keeps them), since claude
projects neither.

The worker's prompt reaches claude in the [AGENT SYSTEM PROMPT] block,
since its mirror projects no spec prompt
(ACP_BACKENDS_MATE_WORKER_PROMPT_IN_CONTEXT). After the prompt is
selected, the block appends the worker's own principles and then the
crewmate's prompt (context_assembly.agent_spec._crewmate_worker_prompt),
the order the on-disk worker base has them; the worker is a default-named
agent, so the selected prompt is the default contract. KAS's wire spec
already carries both, so its block stays the base.

The single-tool refusal is decided again inside the projection, on the
same read of the crewmate's file the worker spec is built from, so an
edit landing after the allocation check still refuses; the allocation
check stays as a fast path. Which prompt channel a worker uses is its
live provider's answer (mate_worker_prompt_in_context), recorded at
allocation, so changing agent.member_acp_backend does not move a running
worker's prompt.

(cherry picked from commit 32362f8)
…3921)

* fix(dashboard): read every body integer through body_int (#22502)

json.loads reads 1e999 and -1e999 as infinite floats, and int() of an
infinite float raises OverflowError. Nine dashboard sites guarded that by
adding OverflowError to their own except tuple, and five more caught only
TypeError/ValueError, so an infinite value raised a 500 there:
history_max_days, restore_window_minutes, the artifact folder order, and an
instance's remote_port/via_remote_port.

Every one of these now reads through dashboard/body_numbers.body_int, which
raises ValueError for an infinite number, and keeps the answer it gave a
malformed value before: the same 400 and code, or the same skipped field or
zero fallback. The skills POST read (limit, capacity) moves too. The
error-handling spec row no longer lists exceptions.

The messaging, files, prompts and chat-tags facades name body_int, because
their owner modules run against the facade's namespace.

Closes #22502

* fix(dashboard): keep the skills offset and promote min_sim guarded

The skills POST read still converted offset with bare int() after its
except clause stopped catching OverflowError, so offset=1e999 became a
500. It now reads through body_int like limit and capacity.

The memory promote clause also guards float(min_sim), and float() of a
JSON integer beyond float range raises OverflowError, which body_int does
not cover. That clause keeps OverflowError.

Also drops black reflows that rode along in four baselined files.

(cherry picked from commit dad46e9)
…owser (#23922)

In the desktop app the only pop-out failure message told the user to allow
pop-ups in their browser, which does not exist there.

A session shown in a connected crew's pane runs on the crew's tunnel port.
Its pop-out URL is therefore never the window's own origin, so the desktop
window-open handler sends it, without its token, to the system browser and
returns no window. The pane now learns from the host model that its hub is
the desktop app and refuses before calling window.open, with a message that
names the real cause. A refusal in the desktop app's own window gets a
desktop message instead of the browser advice. Browsers are unchanged.

Refs #23535

(cherry picked from commit 99c2474)
bind_cron_memory stamps member_id on every job a crewmate creates,
script and command jobs included, but wakesCrew dropped script and
command jobs before reading member_id. A crewmate's own poller or
cleanup script therefore never appeared on its Schedules list. Read
member_id first and keep the script/command exclusion for the
name-based fallbacks only.

(cherry picked from commit d2d8d14)
#23925)

Every App Store listing re-ran the app.json fetch, often a shallow clone
bounded only by the 60 s clone timeout, for every source whose fetch had
failed before, because _resolve_manifest cached only a successful fetch.

_resolve_manifest now writes an empty .miss.json marker on a failed fetch,
keyed through _manifest_cache_path on the same source coordinates as the
success cache, and skips the fetch while the marker is under an hour old.
Only the listing reads it: the install path calls _fetch_app_manifest
directly and always fetches. refresh_registries expires each row's marker
with its manifest file, so an explicit refresh always retries. No
concurrency cap is added.

Closes #21964

Co-authored-by: Kyle Seaman <2904635+kyleseaman@users.noreply.github.com>
(cherry picked from commit 455e8ec)
…18259)

On S0 low-power-idle (Modern Standby) laptops the SetThreadExecutionState
hold is registered yet Windows still enters standby once the display goes
off after lock. Say so in the Prevent sleep hint (12 locales, en-XA), the
dashboard.prevent_sleep help text, the config baseline and both generated
settings registries, and in the power.py docstring.

Tighten two i18n style gates: the hi formal-imperative boundary now
treats danda as punctuation, and the pt-PT screen-noun pattern uses
Unicode-letter lookarounds and also matches the plural, with fixture
assertions for both forms.

Refs #14704

(cherry picked from commit 0480058)
…24492)

A conductor context-cap trim no longer writes a transcript row: the person cannot act on it, and the damping guard reads its in-memory trim record, not the transcript. The model still learns of the trim from the seed's note, and the replay still skips a context-cap notice row a transcript already holds.

(cherry picked from commit 1411fcc)
…s after it (#23904) (#24632)

Picking Rename on the Schedule page's empty-folder chip opened an editor
that closed itself: Radix restored focus to the folder-actions trigger
after the input had autofocused, the blur committed the unchanged name
and unmounted the input. When the editor did close on Enter or Escape,
focus fell to document.body.

Skip the Radix close-focus restore when Rename was picked (Delete keeps
it), and hand focus back to the folder-actions trigger on Enter and
Escape. A pointer blur still commits and moves focus nowhere.

Closes #23904

Co-authored-by: Yusheng Xu <rayrayxu@amazon.com>
(cherry picked from commit 9f51b91)
…4674)

* feat(website): swipe to dismiss the notification banner on mobile

On the mobile layout a horizontal swipe on the live notification card
dismisses it through the same `removeNotes` the close control calls, so the
note waits unread in the bell either way. A short swipe snaps back, a
vertical drag is left to the browser (touch-action: pan-y, a scroll-slop
decline before the horizontal lock, pointercancel snaps back), the release
click is swallowed, reduced motion drops the snap-back transition, and the
close button stays visible. Desktop is unchanged.

Closes #24667

* fix(website): time-cap the banner swipe's click swallow

A release past the browser's tap slop may fire no click, and a pointercancel
never does, so a swallow held as a bare flag waited for the next activation
with no pointerdown of its own (a keyboard Enter on the card) and ate it. The
swallow is now a 350 ms deadline, the drawer swipe's own ceiling, armed only
by a release; two tests pin that an Enter after a cancelled swipe, and one
350 ms after a snapped-back swipe, still opens the note.

(cherry picked from commit e4fad39)
* fix(mcp-gateway): never pool a credential vendor

creds-agent and local-creds-agent-mcp join UNPOOLABLE_SERVERS, so the
rewriter leaves them unwrapped and the dashboard reports them as
can_stub: false. A single toggle, a bulk "Enable all" and the batch stub
endpoint therefore all refuse them.

Each chat's own copy of a credential vendor runs inside that chat's
sandbox, which hides the credential directories and the ~/.aws profile
files (sandbox_tiers). A shared backend is spawned by gatewayd outside
every sandbox. Sharing one would hand every chat a vendor that can read
what its own sandbox withholds.

Comments that described the set as empty now say it lists only the
credential vendors.

* fix(mcp-gateway): refuse a credential vendor on every pooled path

The rewriter applies UNPOOLABLE_SERVERS while it writes an overlay, but an
overlay, env sidecar or approval written before the vendors were listed (or
kept by a failed regeneration) could still pool one.

- _FINGERPRINT_SCHEMA 10 -> 11. The denylist is not a fingerprint input, so
  a kept overlay from before it (same version, unchanged inputs) would keep
  the vendor's stub; the bump rebuilds every kept overlay.
- Every reader asks the same rewriter_spec.is_unpoolable (name or alias):
  pooled_session_servers and injection_server_names skip a vendor's stub, so
  a session resolves it from its agent spec, inside its own sandbox;
  filter_target_env drops a vendor's KIROCREW_MCP_TARGET_* entry before it
  becomes gatewayd's env, without recording a refusal; and
  LaunchApprovals.admits_launch, the check gatewayd runs before it spawns a
  backend, refuses a vendor whatever its recorded approval.

Tests: a schema 10 overlay that pooled creds-agent is rebuilt unwrapped; a
stale overlay that still holds a vendor stub injects nothing for it; a
recorded approval admits an ordinary server but not a vendor, in the target
filter or admits_launch. Each fails without its fix.

(cherry picked from commit 2464225)
…earch (#24297)

KAS defers every MCP tool not on toolSearch.neverDefer and never reads
toolSearch.enabled, so a spec without the tool_search loader still had all
of its MCP tools deferred. For such a spec the handshake now lists every MCP
tool it can call as <server>::<tool>, expanding whole-server grants from the
tool names Crew already holds. A spec that grants the loader is unchanged.

(cherry picked from commit ddc1fea)
… tear (#24678)

A working icon-tier mate animated inside its own SVG, drawn through an
<img>. The browser re-rasters such an image every frame, and iOS Safari
re-rastered it in pieces: in the Mates row the head, the tail and one eye
drew out of step, so the face looked torn. It shows right after a mate is
created because its greeting turn runs at once; an idle mate draws the
still frame.

The icon still moves with the ghost's own full body motion, but the motion
is now page CSS (crewIconMotion.css, fed by iconMotionVars) on four
wrappers around one still image of the silhouette and eyes
(poseBodyDataUri), over a tile the avatar paints itself. Offsets and
origins are percentages of the box, so every avatar size moves alike.
Reduced motion keeps the still frame.

(cherry picked from commit e062153)
A member execution record dispatched its stored template_id verbatim, so a
thread recorded before the conductor kept running kirocrew-conductor. The
chat and is_conductor_execution now share mate_worker.member_record_template:
when the member's current bound kiro_agent owns a crewmate file the record is
the conductor, whatever template it kept; any other member keeps its recorded
template. Template records stay exact and no record is rewritten.

On the event loop a crewmate file the snapshot did not know yet answered
not mate-owned, so a just-created crewmate chatted on its own file. The
on-loop answer is now two file stats (mates/agents/<name>.json present,
agents/<name>.json absent), never a rescan.

(cherry picked from commit 7582829)
* fix(memory): give a damaged memory.db its own remedy instead of restore advice

A damaged global memory.db ("database disk image is malformed", "file is not
a database") failed the default store with "Inspect or cancel this store's
pending restore", which names a restore that does not exist. The gateway
itself keeps running; only the default store is unavailable.

The failure now carries a remedy that names the file, says it was left
untouched, and tells the user to copy it (and any -wal/-shm) somewhere safe
and rename it before restarting. The file is never moved or deleted by
startup: unlike knowledge.db (#18435), it holds the store's only copy of its
memories. Any other failure keeps the restore advice.

* fix(memory): name memory.db only when its own open failed

A damage error from the FTS index open or rebuild named memory.db in the remedy. Only the vector store's open earns the keep-the-file advice now; any other failure keeps the restore advice.

(cherry picked from commit efe7624)
…#24563)

On Windows, Builder Toolbox installs tools under %LOCALAPPDATA%\Toolbox\bin,
not ~/.toolbox/bin, and leaves adding that dir to PATH to the user. A gateway
whose inherited PATH lacks it reported a Toolbox-installed Claude Code as not
installed in Settings > Agent Harness. augmented_path now appends that dir on
Windows (the pinned home's AppData\Local when a home is pinned).

(cherry picked from commit 686d39a)
…1776)

* fix(discord): re-admit bot-opened threads after a gateway restart

* fix(discord): close two gaps in restart re-admission of bot threads

- reconfigure now remembers ids an operator removes from allowed_thread_ids,
  and readmit_own_thread refuses them before and after the Discord lookup,
  so a revoked bot-opened thread does not come back on the next message.
- thread_origin opts out of the REST ladder's mid-send roster re-read
  (recheck_destination=False). Its thread is not yet on any roster, so the
  re-read refused the admission lookup whenever a rate-limit hold was served.
  readmit_own_thread re-reads the rosters itself after the answer; every
  other route, GETs included, is still re-judged after a wait.

---------

Co-authored-by: EC2 Default User <ec2-user@ip-172-31-10-71.ec2.internal>
(cherry picked from commit 072a37b)
buluoray and others added 10 commits October 29, 2026 01:29
…r's language (#24966)

A crewmate's first welcome and its goal check-ins are hidden gateway turns
with no user text to follow. Their preambles said to write in the
[UI LANGUAGE] when shown, else English, but that block scopes itself to
tool-call purpose text, so these turns came back in English even with
dashboard.language set to zh-CN.

mate_welcome.with_language now swaps that one sentence for one naming the
tag: the configured dashboard.language, else (welcome only) the opening
tab's X-UI-Lang, sent by api.memberGreet, validated and never persisted.
Check-ins follow the user's language in the thread, falling back to the
tag. English and unusable tags swap nothing, so English prompts stay
byte-identical.

(cherry picked from commit d5a76e6)
…#24954)

* feat(crew-mode): move every used crewmate to its conductor on upgrade

A crewmate whose worker would lose a grant now moves and its one upgrade
notice names the loss; a grant the worker would gain is stripped from the
file and offered only through the capability card; an unreadable agent
moves onto the default template's file with the same notice. A change that
would loosen a limit (gain, relaunched server, hook or permissions /
toolsSettings restriction) still keeps the crewmate where it is. The
ledger records how each crewmate moved under reasons.

* fix(crew-mode): never loosen a parseable agent's limits on the default-template move

An agent whose spec parses but whose prompt does not resolve has its limits
read raw. The default template's file then writes its switched-off servers
as disabled, adds its excluded and disabled tools to excludedTools, and
drops any grant it never had; a hook or permissions / toolsSettings
restriction it cannot carry keeps the crewmate where it is. Drops dated
history from comments and links the PR from the decision entry.

(cherry picked from commit 76e89ae)
…ols (#24910)

Under the dashboard's Reads approval mode (trust_reads) every MCP call
still reached the card, Kiro Crew's own reads included, because an
MCP-served tool carries no host-trusted read-only marker. The read-only
tier now accepts one more proof, only when the surface passes
trust_reads=True: the call's trusted (server, tool) pair is in the closed
_FIRST_PARTY_READ_ONLY_MCP_TOOLS set (@kirocrew-core/memory_recall,
search_chat_history, get_chat_session, skill_search,
local_knowledge_search). The match is on the identity the transport
stamped, never the title or payload, so the same name on another server,
an untrusted identity and every other kirocrew-core tool still prompt.
Deny tiers, governance and the operator confirm list keep running first.

Closes #24899

(cherry picked from commit d8f2038)
…hat (#24987)

* fix(members): list crewmates the person created, even with an empty chat

A crewmate made by approving Mate's crewmate.create card opens no chat, so
its thread stayed empty and the roster's empty-thread rule hid it. Rows the
person created (dashboard_created: source kirocrew and a member id) are now
listed unasked; product-made rows keep the old rule. The reserved default
member is never dashboard_created and is listed only once its thread holds a
message the person sent (has_user_dm_message), so a fresh install shows only
Mate.

* fix(members): keep a person-messaged default listed and share the person-row rule

(cherry picked from commit 163652b)
)

A no-LLM script cron had no way to put its result on the crewmate's
Dashboard page: it holds no kirocrew-panel server, nothing exposed an
append to the crewmate's crew log, and the fold catalog was closed, so a
poll every minute needed a model turn just to copy a number into a field.

ScriptContext.crew_log_append(type, data) posts the record to the new
strict-internal route POST /api/agent-panel/dashboard/record with the
cron's existing credential. The gateway resolves the writer from the
attested cron scope (the scheduler's own record of the job and its bound
member), appends ONE registered dashboard/custom_value entry to the
crewmate's own DM session log through crew_log_emit, the path
dashboard_write already takes, and stamps writer (cron:<job id>), member
and crew_key itself: the body is read for type and data only, so a script
can neither choose who it writes as nor whose log it writes. A cron bound
to no crewmate, a non-cron caller and an app-token caller are refused;
the agent.crew_panel switch closes the door with the rest of the surface.

Caps refuse and never truncate: one plain type name of at most 64
characters, data at most 8 KiB serialized and 8 levels deep, the whole
line under MAX_ENTRY_BYTES, 30 records a minute per job, 64 types
retained per slot. The new slot fold `custom` keeps the latest record per
type as latest.<type>.{data, at, seq, writer, member, crew_key} and
re-applies every cap on read, so a page binds
custom.latest.<type>.data.<key> and redraws on push. dashboard_types
lists it with latest as an open-keyed map; the fold catalogue, the
dashboard skill, the crew-log reference and the specs name it. Context
assembly reads no crew-log fold, so a record reaches an agent only as the
one-line fold summary dashboard_fields already gives fold-sourced fields.

No app-facing grant is added: an installed app's write path stays a
separate decision, and the writer resolution is one function so it could
be added there without touching the append or the fold.

Closes #24669

(cherry picked from commit 959b35a)
…ns (#25106)

On the KAS backend a spec's mcpServers.<name>.autoApprove was dropped
outright, so an owner-written list that kiro-cli honours still raised an
approval card for every call. KAS has no slot for the key; its only
auto-approval input is the permissions block derived from allowedTools.

The projection now runs the spec's servers through the same
strip_ungoverned_auto_approve predicate the kiro-cli spec writers use and
adds each kept verb as an @server/tool entry to that derivation input. It
then meets the per-entry governance ceiling and the PreToolUse-hook
withholding like any other grant. The key itself still never reaches the
wire, and a verb that is not one literal tool name keeps prompting.

(cherry picked from commit 073c83a)
…#24591)

On WebKit (no scroll anchoring) a measure-farm reprice that shrinks
the transcript by more than the reader's distance from the bottom
clamps scrollTop to the new ceiling. The height-sync consumer read
that clamp as the reader moving, dropped the correction, and left the
reader flush at the bottom: the iPhone idle snap in a fully loaded
chat. The consumer now also corrects when the captured scrollTop is
past the new ceiling and the live one sits on it, but only when
scrollTop had not moved by the consuming commit's render, so a reader
who scrolled to the bottom after the capture stays there. The
owner-swap re-base reuses the same ceiling check.

(cherry picked from commit 9a73292)
…al ones (#20487)

Allow for this host did nothing when one reply held two different
addresses on that host (two dashboard links that differ only in their
query): restore_allowed_links restored a host only when every record
for it kept the same address, because a placeholder names only its host.

Pair a host's placeholders with its records by order instead. Records
are written in first-appearance order and a repeated address yields one
record, so exactly one placeholder per record means each address stood
in the text once, in that order. Every existing refusal stays: an
unallowed host, a withheld address, a rule the allow does not relax, or
any count mismatch (an echoed placeholder, a repeat) keeps all of that
host's placeholders. A new read-back check also refuses a host when an
address put back does not read as one whole URL match where it lands
(a placeholder that stood for only the prefix of a longer link) --
which the one-address path could get wrong before.

_redact_segment now keeps records it takes from the raw copy in the raw
copy's order, so the pairing holds when the streamed text and the raw
copy disagree. The stored placeholder format is unchanged.

Rows saved before this change can hold a host's records out of reply
order (the raw copy's records were appended after the text's own), so
pairing by order would show two addresses in each other's places. Several
different addresses on one host now pair only on a row marked
blocked_links_ordered, which _flush_segment and the interrupted-reply path
set when _redact_segment vouches for the order, and which moves with its
records into and out of regenerate variants. An unmarked row keeps the
old rule: restored only when every record for the host keeps one address.

The read-back now also requires the URL match to START where each address
lands, so an address whose placeholder sat inside another link (a second
address carrying the first in its query) is never put back.

(cherry picked from commit b161144)
…25151)

A notification banner card on the mobile layout now also goes with an
upward swipe, like an iOS banner, beside the sideways swipe. The wrapper
locks whichever axis travels the slop first and becomes touch-action:
none so the browser no longer cancels the upward drag as a scroll.

Closes #25143

Co-authored-by: Yusheng Xu <rayrayxu@amazon.com>
(cherry picked from commit e68a716)
#24910's backported tests call _card_outcomes, which main gained in
#24650 (a test-only PR not on release/0.11.0). Add the same helper so
the file passes flake8 and its tests run on the release branch.
@bolichen97
bolichen97 requested review from a team and KiroCrewCaptain as code owners October 29, 2026 04:36
@bolichen97
bolichen97 requested review from mbriones98 and removed request for a team October 29, 2026 04:36
@github-actions

Copy link
Copy Markdown
Contributor

Sensitive Change Review

Sensitive files changed:

  • src/kiro_crew/hook_runtime/gate_tiers.py

  • src/kiro_crew/hook_runtime/tool_identity.py

  • src/kiro_crew/hooks.py

  • src/kiro_crew/security/exfil.py

  • src/kiro_crew/security_posture_sinks_delivery.py

  • No reasoning template for 9df4d01 yet.

How to clear this check:

  1. The PR author posts the template below as a PR comment, with
    every field filled (N/A needs a reason). Keep the
    Head commit line: it must name the current head SHA.
  2. The check turns green within a minute, with no push.
  3. A new push resets this check: post it again for the new head.

Only the author's comment counts this way. A reviewer can instead
paste the filled template into an "Approve" review on this head.
Merging still needs a reviewer's approval.

## Sensitive change review
- Head commit: 9df4d012cced4b590b12887607305bf9e3cc7804
- What rule changed:
- Before -> after (what was allowed/blocked, now):
- Worst case if wrong (breaks commands / leaks secret / locks user out):
- Evidence checked (which sandbox result proves it):
- No regression:
  - Tools still work (which tools/commands were run, result):
  - Backward compatible (old config / old data / old callers still work, how checked):
  - Existing tests (which suites passed on this head):
- How to undo:

@github-actions github-actions Bot added readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Oct 29, 2026
@github-actions

Copy link
Copy Markdown
Contributor

First Principles Review (Opus 5.5) — ✅ PASS

Premise-level review of 9df4d012cced4b590b12887607305bf9e3cc7804 — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

First-Principles-Verdict: PASS

This backport opens no new decisions. Each of the 42 picks has already merged on main through its own PR. The two default-changing picks are covered by the accepted rfc-crewmates-launch, and both are recorded under docs/decisions/.

Notes

What this change ships

Inventory (10 items) — 9 justified

Intent: ship the 0.11.0 crewmate-launch list on release/0.11.0 by cherry-picking PRs that already merged on main. This is a FIX (a backport).

  1. Each crewmate's worker runs in its own git worktree, on claude as well as KAS — justified
  2. On upgrade, every crewmate in use moves to its conductor, with a one-time notice of what it lost — justified
  3. A crewmate's script cron feeds its Dashboard page, and its own schedules are listed — inherited ("fix" framing on a feat)
  4. Crewmates the person created show up in the members list even with an empty chat — justified
  5. Reads mode auto-approves Kiro Crew's own read-only MCP tools, and KAS honours a server's autoApprove — justified
  6. Discord, Teams and Webex stop re-announcing while held, and Discord re-admits bot-opened threads after a restart — justified
  7. On mobile, the notification banner dismisses with an upward swipe — justified
  8. Error and status wording: gateway running, a damaged memory.db, an unreadable mcp.json, pop-out refusals, Modern Standby — justified
  9. Fixes for spawn, the scratch environment, the WAL retry, macOS memory and the Windows bin path — justified
  10. Pins re-counted for release, plus the _card_outcomes test helper — justified

[FIRST-PRINCIPLES-REVIEWED] 9df4d01

Description read

[DESCRIPTION-READ] b5d0fd61008e3640ce91d0cba3ed58ca7da9d002fd7d0b26b17d1dd11f670ea7

That sha256 names the PR description this verdict read. Recompute it with .github/scripts/pr-description-capture.sh; if it differs, the description changed after this verdict was formed and any finding drawn from the description is unproven.

@github-actions

Copy link
Copy Markdown
Contributor

Security Scope Review -- 🔴 could not complete (fails closed)

No scope verdict for 9df4d012cced4b590b12887607305bf9e3cc7804 (the candidate stage failed). This lane fails CLOSED: an unmeasured tightening must not read as "nothing newly refused". See the job logs, then re-run -- or record /ai-review override scope 9df4d012cced4b590b12887607305bf9e3cc7804 <class>: <reason> if a human has judged the scope by hand.

@github-actions

Copy link
Copy Markdown
Contributor

GPT 6.1 Review — 🔴 changes requested (blocking)

GPT 6.1 found at least one blocking issue that must be resolved before merging 9df4d012cced4b590b12887607305bf9e3cc7804. 1 of 2 blocking finding(s) are security-class and were withheld from adjudication, so the blocking verdict stands.

This comment is updated in place on each push.

BLOCKING -- docs/decisions/2026-10-27-a-crewmate-the-person-created-is-listed-with-an-empty-chat.md:19 -- Link the maintainer’s recorded decision

  • The owner's restatement on the pull request that adds this entry.
    Trigger: The PR indexes this new binding decision, but its Evidence section provides no link to a human decision or maintainer restatement.
    Outcome: Subsequent review and issue triage treat the behavior as settled without checkable maintainer authorization; the entry becomes immutable after merging.
    Anchor: recorded-decisions-are-not-regressions
    Fix: Cite a direct permalink to the on-record restatement by a maintainer listed in MAINTAINERS.md.

BLOCKING -- website/src/pages/overview/McpShareRefusedNotice.tsx:83 -- Render failed sharing through ErrorNotice

role="status"
Trigger: A shared server’s command changes; LaunchApprovals.admit() rejects its launch, /api/mcp-gateway/status returns share_refused, and McpTab renders this handwritten box.
Outcome: The failed sharing operation bypasses the required error surface and offers no agent hand-off, even when the page has no staged edits.
Anchor: errors-use-error-notice
Fix: Render the refusal through ErrorNotice, retaining the MCP Management link and the existing staged-edit protection for hand-offs.

[BLOCK-MERGE] 9df4d01
[GPT-REVIEWED] 9df4d01

Adjudication (Opus 5.5) — is blocking on each finding proportionate?

F2: I'm upholding it. It cites errors-use-error-notice, which is marked blocking: true in the base-branch snapshot (.review-base-rules/website-AUTOSDE.yaml:622-623), so I can't downgrade it. The refusal comes from share_refused (src/kiro_crew/dashboard/handlers/mcp_gateway_routes.py:84), meaning a launch that was rejected. That counts as a failure under the rule. It is shown in a hand-built role="status" box at website/src/pages/overview/McpShareRefusedNotice.tsx:83, not through ErrorNotice. The fix is small: the same file already uses ErrorNotice with askAgent={!stagedChanges} at lines 70-76.

F1 (fenced): UPHOLD-FENCED. docs/decisions/README.md:38-43 says an entry must link the on-record artifact, and that a decision is only a proposal until a maintainer restates it. The Evidence section at line 19 only mentions "the owner's restatement on the pull request" and gives no link. A missing link is not an extreme condition, so there's no rarity argument for flagging it.

[ADJUDICATION] 9df4d012cced4b590b12887607305bf9e3cc7804 total=1 uphold=1 downgrade=0
UPHOLD F2 website/src/pages/overview/McpShareRefusedNotice.tsx:83 reason=autosde-blocking-rule
[GPT-ADJUDICATED] 9df4d012cced4b590b12887607305bf9e3cc7804

[ADJUDICATION-FENCED] 9df4d012cced4b590b12887607305bf9e3cc7804 fenced=1 flagged=0
UPHOLD-FENCED F1 docs/decisions/2026-10-27-a-crewmate-the-person-created-is-listed-with-an-empty-chat.md:19 -- The Evidence section describes a restatement but links nothing, which docs/decisions/README.md:38-43 requires, and leaving out a link is an ordinary authoring gap, not an extreme condition.
[GPT-ADJUDICATED-FENCED] 9df4d012cced4b590b12887607305bf9e3cc7804

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 9df4d012cced4b590b12887607305bf9e3cc7804 false-positive: <reason, 20 to 500 characters>

@bolichen97

Copy link
Copy Markdown
Contributor Author

Superseded by the split backport #25663 → #25664 → #25665 → #25666 → #25667 (1–2 commits each, every diff under 1 MB). The same picks, plus the fixes for the GPT 6.1 findings (#25656, #25660, and the #24987 decision evidence links).

@bolichen97 bolichen97 closed this Oct 29, 2026
@github-actions github-actions Bot removed the readiness: action required A blocking check or review needs attention label Oct 29, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants