Repository navigation
fix: backport the first 74 commits of #34021 to 0.12.1 (rebase-merge limit) - #34032
Merged
Merged
Conversation
…lds (#30053) A Bedrock KB source row on Knowledge > Sources always said "Live", even after the consent grant behind it was revoked or the account drifted, so a knowledge base that every search now refuses looked healthy. The row now asks the same consent query the add form polls, for the row's own (profile, region). It keeps "Live" while that query says granted, and while it is loading or fails. When the query answers that the grant is gone, the chip says "Grant revoked" and opens the add form's consent card for that target. Refs #11113 Co-authored-by: w <w@k.local> (cherry picked from commit 3e97394) Backport note: the 13 locale files keep release/0.12.0's adjacent sync_failed_notice key and add this fix's four keys after it.
kiro-cli reads an agent spec's toolsSettings itself and KAS receives it as permissions rules and gate allowlists, but no mirrored harness gets it. On claude and codex the shell denyByDefault allowlist, the write allowlist, deniedPaths and web_fetch.blocked were dropped with no ruling and no notice. Add Concern.TOOLS_SETTINGS so every mirror has to rule on it: no-channel on claude, codex, opencode, goose and strands (the channel is the spec-shell-allowlist / spec-write-allowlist gate tiers), and withheld on copilot and deepseek. A new chat session on a backend whose mirror rules it so, and where Crew does not fire spec hooks, now gets one notice naming toolsSettings whole. That is the same wording goose, opencode and strands already get from the spec-key read. The check is a positive mirror lookup (spec_tools_settings_applied_nowhere), so kiro-cli, KAS and pi gain no step. The concern is recorded off the backend card with its reason. Enforcement is left to a separate change. Part of #29275 (cherry picked from commit 9210652) Backport note: conflicts were context only. Kept release's spec_keys_without_carrier docstring (toolAliases, dropped knowledge bases) and release's agent-host-contract and agent-spec-fields text around the change, adding only this fix's sentences.
…30156) The governed-scope rows on Settings > Security fill a catalog label like "Block-list · {{rules}}", but the rule count inside it was built in JavaScript as "N rule" or "N rules". Every language therefore showed the English noun and English plural rules, for example "Verbotsliste · 3 rules" in German. Move the count into a pluralized catalog key with the forms each language needs, so the noun and its plural come from the catalog. That was the last whole-word plural ternary in the UI, so the hardcoded plural ceiling drops to the seven sites left, all English prompts written for the agent. (cherry picked from commit 8b57dd2)
claude reads neither toolsSettings.shell.denyByDefault/allowedCommands nor fs_write.fallbackAction "deny"/allowedPaths, and nothing stamped them on its permission requests, so an unattended claude run executed commands and writes the spec forbids. The direct client now reads both allowlists from the mirrored spec its session array uses and stamps them on the session's requests, a native sub-agent's included, for the existing spec-shell-allowlist and spec-write-allowlist gate rows, AUTO_APPROVE included. While either is set, permissions.ask rules for Bash and the file-write tools go into the settings seed and the session/new envelope, so claude asks about calls it would approve itself even beside a project settings.local.json Crew declines. The toolsSettings session-start notice on claude now names only what its gate does not apply, and the claude mirror's ruling says which allowlists reach the gate. Membership is the positive set ACP_BACKENDS_CREW_ENFORCES_SPEC_TOOLS_SETTINGS (claude only); kiro-cli enforces the field itself and gains no step. Closes #30100 (cherry picked from commit 5de3642) Backport note: release does not import worker_array_outruns_seed in client_mcp_array.py, so only the spec_tools_allowlists import was added there; the client_claude_session docstring keeps release's _include_memory entry; turn_spec_hooks keeps release's keeps_mcp_oauth import. The toolsSettings notice hunks apply on top of the #30088 backport above. Doc conflicts were context only.
…#30241) A Bot's worker on claude got the adapter's default MCP discovery, so beside the servers its spec grants it also started every server in the user's own Claude config, the project's .mcp.json and installed plugins, once per worker. With many such servers and several workers the host ran out of memory and the gateway restarted. Send strictMcpConfig for a worker session, as the tool-less and Side Chat sessions already do. Ordinary and member chats are unchanged. Refs #30171 (cherry picked from commit 6c89c09) Backport note: claude-code-provider.md takes only the new "A Bot's worker session" section; the "cold start" section beside it on main is not on release. The client_claude_session docstring keeps release's _include_memory entry and adds _session_key.
…ssion (#29739) A code channel is one session: its binding names it, and Stop and action answers follow the binding while typed messages follow the Slack link. Linking the channel's anchor thread to another dashboard slot moved only the link, splitting the channel between two conversations. The slack-link path now refuses that handoff (409 code_channel_anchor) before any Slack write; the binding's own session may still re-link its anchor, and other threads in the channel link as before. Fixes #29606. (cherry picked from commit 7fe4c86) Backport note: release's _bind_channel has no planned_root (feat #28305), so its own body is wrapped in hold_anchor and keeps release's repo_root line. store.py adds the loaded and unreadable properties over fields release already has. In interactions_options.py the anchor store load runs before the thread read, where main has it since #32336 (already on release), so nothing awaits between the read and the link.
Process-tree walks can hold all eight subprocess_executor workers past their callers' timeouts. The archived-session scan behind folder deletes, the folder suggestion's directory match and sample titles, the artifact folder list, and the terminal's cwd and title probes and selection redaction all queued on that pool, so they waited behind the walks. Move them to discovery_executor, the pool the terminal's path completion and other dashboard reads already use. Fixes #30666 Co-authored-by: Zezhen Xu <zezhexu@amazon.com> (cherry picked from commit 587e8fb)
…0564) Between a 340px shelf and the width its full row needs, the context readout and model chip kept their whole width and painted over the agent and project chips, so neither chip could be read or tapped on a phone. The right group now shrinks (the model name truncates) and the chip group keeps a floor sized for two icon-only chips. Closes #16849 Backport note: release's right shelf group is 'flex items-center shrink-0' with no gap-2 or presence slot (both from unreleased #29589), so the fix lands there as 'flex items-center min-w-0'. (cherry picked from commit f19babf)
A mate's working state read only its own thread slot (turn, children, queued children), so a mate whose dispatched worker sessions were busy looked idle on the Mates page, the switcher, the Working filter, the nav dock, the command bar and its sidebar folder face. The needs-you signal already rolled worker sessions up through mate_root; working did not. Add busySlotKeys in lib/crewMode: every busy slot's own key plus its mate_root. Every mate status surface now reads it. Fixes #30852 Co-authored-by: CrysisDeu <crysisdeu@users.noreply.github.com> (cherry picked from commit 2dc4573) Backport note: crewMode.ts takes only busySlotKeys, placed after release's mateRootOf; the mate sessions preview helpers beside it on main (mateSessionsHidden, hiddenMateRootOf, liveMateThreads) are from a feature not on release, and so are their tests, so crewMode.test.ts takes only the busySlotKeys tests and import. MembersPageRoster.tsx reads the live mate threads with an inline mode === 'member' filter, release's own keying, in place of main's liveMateThreads helper.
A message waiting in a session's queue is judged against the session's current state after the drain's last wait, so a change the session went through during that wait applies to the message before it starts. Closes #26979 Backport note: release has no Pause all re-check after the parked-work probes (crew_pause, feat #29928, unreleased), so the probe block moves above the opening sweep and the crew_pause lines and their reworded comment are left out. (cherry picked from commit 312a3c5)
…30951) A pure-KAS agent spec that grants the shell or the filesystem through its own permissions block prompted on every call: the KAS projection drops such an allow on purpose (it must not reach KAS as an allow rule, or the call would skip Crew's deny floor), and nothing fed it to the gate's spec-builtin-grant row the way allowedTools and allowedCommands are. The projection now records an unscoped shell/fs_read/fs_write allow as the built-ins it covers, and a shell allow's match globs as command patterns (* read as .*, every other character a literal). The gate answers them after every deny row and governance, and an allow with an exclude, an authored ask/deny, or a ceiling with an opinion keeps the prompt. Backport note: resolved onto release's KAS projection: the crew_answered_commands call keeps release's tools_settings local and 3-arg _allowed_commands_withheld, kas_permissions.py keeps release's crew_answered_paths, the agent-spec-fields.md permissions row takes only the added sentence, and the new test sits beside release's tests. (cherry picked from commit bfbfb3d)
A crew member whose template declares a relative steering glob, such as the shipped file://.kiro/steering/**/*.md, refused to start when that tree held a link to a file or folder outside the project. The implicit steering scan already leaves such a link out, unread, and names it. The declared glob now does the same: the link is left out unread and named in the linked-steering note, in the member snapshot and in the KAS document reader. A link into managed memory or member state, and a link the path screen refuses, still refuse, as does any link under a relative glob without a steering component. (cherry picked from commit 477eebf) Backport note: release/0.12.0 splits the steering test into steering_glob and left_out in _resource_paths (#31131, #31152); the fix keeps those lines and changes only the relative-declaration fallback for outside= to left_out.
A crewmate's reply on the Crewmates page offered no Read aloud, and auto-speak never spoke it. The crewmate bubble passed no onSpeak, and voice playback accepted audio only for chat.activeSlot, which the Crewmates page never moves. Playback now follows storeFocusedSlot(), the resolver attention.ts already used for the focus announcement, moved to lib/ so both share it: the mounted crewmate thread, else the active chat. A crewmate reply's More menu offers the same Read aloud as a Sessions reply through one shared speakReply() action. Opening or leaving a crewmate thread stops speech like a slot switch. The crewmate pane shows its own playback failure notice, without the Ask the agent hand-off because the pane holds an unsent draft, and reveals its replies' action rows while the browser blocks playback. When a turn ends with no streaming row, auto-speak falls back only to a reply of that turn, so a turn that failed before replying reads nothing. Fixes #29689 Co-authored-by: Matt Briones <mdbrio@amazon.com> (cherry picked from commit 2ee14a3) Backport note: release messageActions.ts keeps its own chatSlice import list (no openActivityPanel, TOGGLE_SIDE_TERMINAL_EVENT or usePanelTerminalsPending from feat #25778, which is not on release) and only drops setVoiceAudio and store, which speakReply now owns. docs/feature-map/README.md takes only this PR's Voice reply row and leaves out the Code-channel sessions row, which main has and release does not.
#31002) A KAS session refuses an agent spec's whole `permissions` block when it names a `policies` bundle, or is malformed or too large. None of its rules travel, and with a bundle the shell and file built-ins keep prompting even where `allowedTools` grants them, but the user was told nothing: the refusal reached only the gateway log. The session-start notice now names `permissions` whenever the projection refuses the block, read through the same parser the merge uses, and says why a block is refused and what a bundle keeps asking. The refusal itself is unchanged. Part of #30165 Backport note: kas_agents.py keeps release's import lists (shell_denies_carried, and the multi-line mcp_cleanup import with mcp_entry_is_registry_governed and mcp_server_muted_in_any_store) and adds projection_accepts. security.md keeps release's grant-only paragraph and adds the start-notice clause. (cherry picked from commit 329ebc1)
… be dismissed (#31209) Co-authored-by: Joe Guo <zejiangg@amazon.com> (cherry picked from commit e41b9f0) Backport note: only the running-clears half is taken: awaits_person's running flag and its with_worker_facts and api_work_ledger_board call sites, crewNeedsYou in crewWork.ts and milestoneState in UltraCrewPanel.tsx, with their tests. The dismiss half (the dismiss action and _dismiss branch, the _is_outstanding tail-order change, the Skip question button and its three locale keys, the crewBoardAction reportId) is a new feature and is left out, so test_crew_board_needs_you_dismiss.py keeps only the running test.
…#31404) The two confirmations shown after a user trusts one exact command (the chat composer's timed-grant line and the channel card's agent-scoped line) wrapped the command in the locale's quote pair. A command that itself ends in that closing quote showed it twice, and the quotes were the only boundary between command and sentence. Every catalog now carries a <mono/> slot instead of a quoted {{cmd}}, and both callers render through <Trans> with the command as the chip's child, the same chip the trust menu rows use since #20305. The command is never an interpolation value, so it is never parsed as markup. Closes #23893 (cherry picked from commit 4991e9f) Backport note: ChatInput.tsx takes only the approvalNotice guard; the optimizeBusy status line beside it on main is from a feature not on release. en.context.json keeps release's trustDropdown descriptions and takes this fix's two entries.
On the KAS backend the engine keeps a remote MCP server's OAuth tokens in Crew's dedicated grant vault, not in kiro-cli's cache, so the Connections probe found no grant and showed "Sign-in required" for a server that was signed in. The probe now also asks whether that vault lists the engine's kiro.mcp.<hash>.tokens entry for the url. It reads entry names only: the key file is never opened and nothing is decrypted. Either store holding a grant reads "Signed in"; both definitively empty reads "Sign-in required"; anything else, including a url entry with headers (whose engine key cannot be derived from the config), reads "Not verified". With agent.kas_persist_mcp_sign_in off the engine does not use the vault, so the check answers "no grant" without reading it; the fail-closed setting reader moves into the new leaf module and kas_secret_storage re-exports it. docs/architecture/mcp.md describes both stores. The lookup is SEL-audited under a new read id, mcp_discovery.kas_grant_presence. A recorded engine hash pins the key format against a KAS upgrade. Fixes #30531 Not generalizable: the fix mirrors one engine's private key format for one badge; the general lesson (a status check must consult every backend's credential store) is already the harness-parity rule. Co-authored-by: Matt Briones <mdbrio@amazon.com> (cherry picked from commit e9919d2) Backport note: release already carries #31467's once-per-run warning inside kas_secret_storage.persistence_enabled. The reader moves into mcp_grant_kas.py with that warning and its once-flag, as main has them, kas_secret_storage keeps _server_id and _is_tokens_key, and test_saving_silently_off_warns_once_per_run reads the new module's flag and logger as main's does.
…a timed-out close back (#31312) (cherry picked from commit 53eb62e) Backport note: the ChatSidebar.tsx and dashboardSlice.ts import lines keep release/0.12.0's crew-mode imports; main's adjacent useMateSessionsHidden, mateSessionsHidden and selectLiveMateThreads come from #30220, which is not on release.
#31345) On the default "more" widget density the dashboard widget pointer said how to draw a widget but never when, and with the skill trigger off by default nothing else did, so sessions stopped drawing widgets on their own. The pointer now says to draw an answer that reads better drawn as one widget without being asked; "less" still prefers markdown. The agents that get the pointer appended to their contract block were a hand-written list that missed the pipeline and security conductors. It is now the shipped conductor set, crew_mode.CONDUCTOR_TEMPLATES, and every conductor maps the skills the pointer leads to (skill_resources.WIDGET_POINTER_SKILLS), including the crewmate's conductor when its spec mounts the work ledger. The crewmate's worker, which inherits the default prompt and its token, drops the pointer: workers report to their conductor, not to a person. Closes #31294 Backport note: the test_agent_refactor_spec_bytes.py conflict was adjacency only. Release pins different team-lead and worker digests from main's, so those lines keep release's values; this fix's own pin changes (the mate, pipeline and security conductors) are kept as on main and match a DevSpace run on release. (cherry picked from commit e7278b2)
…31126) In a group in mention mode, _is_addressed read contextInfo only from the text, image, video and document submessages. A quote-reply to the agent's message sent as a voice note or audio file, a sticker or a video note carries its contextInfo on audioMessage, stickerMessage or ptvMessage, so it read as not addressed and the agent stayed silent. The submessages read for contextInfo are now typed text plus every media field the media module already reads, so a new media kind cannot be missed again. Mention and quote rules and sender admission are unchanged. Backport note: the media import keeps release's names and adds _MEDIA_FIELDS; KIND_VOICE comes from unreleased #31108 and is left out. (cherry picked from commit e036aa5)
…opping the brief (#31213) Co-authored-by: Joe Guo <zejiangg@amazon.com> (cherry picked from commit 9486b89) Backport note: release/0.12.0's Brief has no lines or checked field and no wake checks (#29679, #30527 are not on release), so _marked tags only the fields release has, the wake-check test and the spec sentence about wake checks are left out, and _redact_whole (removed on main by #30234) stays. The scanners move to warm_task_brief_check.py as on main; the run keeps release's own link row step. The grounding-miss parametrize case keeps release's existing rows and adds this fix's three cases.
) A turn that ended while the person was still typing in a queued card's editor drained the entry as it was: the agent got the unedited words and the save then landed on an entry that was already gone. Opening the editor now places a short server-side hold on the entry (POST .../queue/{queue_id}/editing). While it is held the drain keeps user entries queued, the same way the sub-agent hold does, and system injections still flow. Saving the edit releases the hold after the text is stored, closing the editor without a save releases it, and a 300-second deadline releases it on its own; each release starts the drain on an idle slot. Backport note: the conflicts were context only. The feature-map Sessions row and the ComposerDockView destructuring line carry unreleased wording and props (composerPresence), so they keep release's text and gain only this fix's additions. (cherry picked from commit d4b20d0)
Up-arrow recall put a sent prompt's [ Paste #N ] tokens back in the composer without the blocks behind them, so the resend carried the literal token text instead of the pasted lines. Down-arrow back to the draft lost the draft's own blocks the same way. Each history entry now carries the blocks its row (or its send record) already holds, the browse cursor keeps the draft's blocks, and both composers install them with the recalled text. Closes #29722 (cherry picked from commit 1eb1d21) Backport note: ChatInput.tsx keeps release's fireComposer argument (main's sendOrOptimize is not on release) and adds only onPasteBlocksChange to useComposerKeyDown. messages.ts keeps release's imports (main's toolReturned import is not on release) and adds only type SendAttempt.
The Normal approval mode was described as checking with you before doing anything, and Reads as the mode that runs read-only shell commands on its own. The gate auto-approves look-ups in every mode (file reads and searches, web fetches, read-only shell commands that do not run git), so a read-only command ran with no card in Normal while the picker promised one, and Reads looked like it added what Normal already did. Every description of the two modes now says it: Normal looks things up on its own and asks for everything else, and Reads also searches its memory, past chats, skills and knowledge library on its own. That covers the picker, the Bot change and import cards, the Bot profile, the Mochi settings and tip, the picker's find_ui description and the Bot's set_approval_mode tool text, in all twelve catalogs. en-XA and the UI index are regenerated, and the governance spec's mode list says the same. No gate behaviour changes. Closes #29731 (cherry picked from commit 0e2e8ce) Backport note: de.json takes this fix's profile_mode_normal_desc and profile_mode_reads_desc values in release's file (the conflict was with neighbouring keys only). ui-index.generated.json is regenerated on release with npm run gen:ui; it differs from main's only in input_digest, because the adjacent entries come from feat #31473, which is not on release.
A destructive command the check let through, when a nested shell received a value the outer shell had already filled in, is now refused. Printed text and literal paths that only look like that command keep their verdict. Closes #25385 (cherry picked from commit a62642a) Backport note: release's security package line budget is 34,936, so it is raised by this fix's 24 to 34,960 (main's value and its two earlier raise notes do not apply to release).
A disabled primary button faded its accent fill to 30% opacity, which on a light canvas reads as a pastel accent rather than a control that is off. Btn's primary variant and SendBtn now take the muted surface and label when disabled, and drop their hover glow and sheen; the outline Btn variants keep the fade. Three call sites that painted the accent onto a plain Btn use primary, and the embedded chat's send no longer fades on top of the muted state. Closes #23978 Backport note: release's Switch All Sessions button has no failed-retry state (bulkRetriesFailed is from #31333, not on release), so only its accent classes are swapped for the primary prop on release's line. Backport note: release's Switch All Sessions button has no failed-retry state (bulkRetriesFailed is from #31333, not on release), so only its accent classes are swapped for the primary prop on release's line. (cherry picked from commit 5658952)
A code file opens in the editor, and diff mode there draws the editor's live diff. When the file is over the inline budget (400 lines or about 128 KB) the editor dropped the live diff without any fallback, so the Files panel showed only the line-count badge: no diff and no "Show line-by-line diff" button. The editor now reports when it drops its live diff. With a clean buffer the panel shows its diff view in place of the editor, and that view starts the existing off-thread line-by-line diff on its own, with the existing computing strip and Cancel. Only the visible tab auto-starts, so hidden tabs never queue work on the shared diff worker. Chat file cards and the Turn Diff tab stay click-to-start. The inline budget is unchanged: it guards main-thread work. Builds on #25295 by LundyLife. Closes #25237 Co-authored-by: Blake Lundy <42579642+LundyLife@users.noreply.github.com> (cherry picked from commit 11aed35) Backport note: MarkdownPanel.tsx gets onLiveDiffDegraded and autoLineByLine on release's ContentRenderer and DiffViewBlock lines, without main's richSource prop (feat #30812, not on release). The learn-cron-dashboard.md sentence is inserted into release's Side-panel file tabs paragraph.
A state import wrote its uploaded archive (up to 2 GiB) to the system temp dir, which a tmpfs /tmp holds in RAM. The spool now lives in a per-request directory of <data home>/staging, like the extraction. The data home may be NFS, so no call on it runs on the event loop: a new part_stream.stream_part_off_loop runs the create, every write, the close and the discard on one worker thread in submission order. The staging leaf is agent-writable, so the spool directory and its file are created and removed through pinned descriptors and a name swapped for a link is never followed. The other two upload handlers are unchanged. Builds on #28493 by @dodgeblaster. Closes #22565 (cherry picked from commit 10e0af5) Backport note: release's apply_import_zip comment is the longer one docs #30048 rewrote on main; it keeps release's wording and only replaces the sentence that put the upload spool in the system temp dir.
…32299) A folder Default agent saved as a crewmate's label, which the earlier picker led people to do, binds a same-named template, so new chats in that folder lose the crewmate's model, workspace and memory with no sign anything is wrong. Once per data home, after the dashboard is up, such a value is rewritten to the crewmate's config key. A label two crewmates share is left as saved, and a template picked under that name afterwards stays. Closes #22713 (cherry picked from commit ddc8f9d) Backport note: chat_folders_records.py takes only FOLDER_MEMBER_LABELS_MARKER beside it; the default_prompt cap and clean_folder_default_prompt next to it are feat #30802, not on release.
The spawn gate wrote approval/decided with no `by` after every normal return of the spawn approval callback, which reads as a person's answer. The callback often answers alone: an auto-approve source, --approval, the YOLO override, slot trust, no UI, or a Slack prompt whose window closed. The spawn callback now returns a HostSpawnDecision(approved, cause) when the gateway answered with nobody asked, learned through a new optional on_host_decision hook on _interactive_approval that only the spawn gate passes. A bare bool still means a person answered, so tool approvals and every approve/reject outcome are unchanged. The gate records `by: host`, with `cause: approval_timeout` for an expired Slack prompt. Closes #16574 (cherry picked from commit 7cb813b) Backport note: turn_approvals.py, its composition-contract signature and the slack-gateway.md list add on_host_decision without the requester parameter and its card line (feat #31393, not on release), so the doc entry is numbered 6 on release.
#32327) A remote crew's tab is a cross-origin iframe, so its toast check read the "a background chat finished" setting from its own origin's storage, which the hub's Settings toggle never writes. Every pane saw the default (off) and never relayed a finished turn to the hub's OS banner; only the chime, which has no such setting, was heard. The hub now sends its setting in the host model it posts to each pane, and re-sends the model when the toggle flips. A pane follows that value unless its own storage holds an explicit choice; a hub that sends none leaves the pane on its own storage. Part of #20191 (cherry picked from commit 9670914) Backport note: NotificationsPanel.tsx takes only the chatCompleteNotify import swap; the AppSessionMuteSection import beside it on main is from a feature not on release.
Boot restores slots and publishes their keys to the dashboard-surfaced set before any browser connects, and they stay published after the last one closes. The memory.recall gate read only that set, so a recall in a restored tab could send recalled memory to the decision provider with nobody there to read the receipt. The gate now also needs an open owner dashboard socket, read in the same predicate at the early refusal and at the egress. Part of #12532 (cherry picked from commit c2d3166) Backport note: the recall call keeps release's arguments (no conversation_id, feat #30968) and gains the state argument. The memory_recall.py docstring and the tab-closed test's docstring keep release's wording with the open owner dashboard added.
…1882) A tunnel or proxy answers with its own HTML page while the gateway restarts, and since #9484 that page reads as a bare "HTTP 502", which looks like a lasting failure. When such a page arrives with 502 or 503, the dashboard and app clients now say the server is temporarily unavailable, in every shipped language. Like the "HTTP 502" it replaces, the text names only the cause and has no end punctuation, so a message that wraps it keeps its own period and retry advice and says each once. A 504 keeps "HTTP 504", since the request may have run. A JSON or plain-text body keeps its own message, and a caller with a sentence of its own keeps it. Closes #9582 (cherry picked from commit b3be506) Backport note: RestartButton.cov80.test.tsx keeps release's imports and window.confirm stub (main's confirmAction dialog helper and within import come from a later change not on release), so the new 502 case drops its confirmAction step.
…#32393) Backport note: queue_drain.py imports only dashboard_slot_key; the adjacent queue_edit_hold import belongs to unreleased #30549. The new channel-thread test resets only release's _dashboard_surfaced registry; session_surface._dashboard_tabs is from unreleased #31139. Picked after apply clean. (cherry picked from commit c4d1884)
…32477) A Connect whose mint found the stored grant does not work ended with reason mint_grant_unproven, but the card still folded "grant files on disk" plus the probe's needs_auth into Connected, beside the generic "failed before an approval address was ready" text. The page now remembers that provider for the visit and treats its grant files as unconfirmed until a later mint reports granted, so the card shows Not verified with its Connect button, and the failure line names the real cause in every shipped language. Closes #29764 Backport note: release pushes a mint failure as a one-line object without the reasonless-expiry fallback (#31821, unreleased), so the fence field is added to release's slug, reason and endpoint. The connect tests keep only this fix's two new cases; the neighbouring timeout, remount and stale-feed cases belong to unreleased PRs. Locale conflicts were adjacency only. (cherry picked from commit 7f11b5c)
) The crew window's mismatch notice printed "?" for a version it could not read, a placeholder "0" from a peer as if it were a release, and told the user to "update one" without saying which side was behind. Every reader of a crew's /api/version now treats an all-zero release as no version. An unknown version always carries its reason, the capabilities reply reports which side is newer, and the crew window and New chat on crew word each case. Closes #25246 Backport note: en.context.json adds only the version_unknown line (the neighbouring elicitationCard.dismiss key belongs to unreleased #23973), and the instances.md rows take only this fix's sentences on release's row text. (cherry picked from commit 3ea0a5f)
A reload, or opening a chat by its URL, drops the in-memory tool log, so the running shell's row fell to the history branch, which always drew a finished, non-shell row: no spinner and no "Running · Ns" line for the rest of the command. The history branch now reads the row itself: the persisted kind marks a shell, and the row is still running while its turn runs, the server has not marked it done, and nothing but its own tool run follows it. Every final tool result sets the row's done flag, as the server does, so the line goes away when the command ends, and a row whose log entry was trimmed still reads finished. Closes #29760 (cherry picked from commit 1b57675) Backport note: release/0.12.0 has no noteToolReturn helper (#29356 is not on release), so activity.ts adds only markToolRowsDone and calls it in sseToolResult's own `final !== false` block, where main's helper calls it.
…not 0% (#32769) (#32831) A usage_update moves the provider's context counts mid-turn but emits no event, and the only context_usage frame was the turn-end one, so a long tool-calling turn left the meter on its old reading. The stream loop now broadcasts a measured, changed reading at most every 2 s; the turn-end frame stays authoritative. The composer no longer collapses an absent reading to 0%: a slot with no reading renders a dashed track and an em dash with a 'No context reading yet' tooltip (new catalog key in all 12 catalogs + en-XA). Co-authored-by: Zezhen Xu <zezhexu@amazon.com> Backport note: the stream loop on release has no turn_timing mark or owed agent cancel, and _ChatTurn has no _drained_subagent_notices, so only the _broadcast_live_context call and the two _live_context_* attributes are added. ChatInput and ContextShelf gain contextUnknown without the unreleased cacheSlot prop. Cutoff rebase onto release/0.12.0 34898f9, which already carries #31305: ContextUsageControl takes contextUnknown without ctxWrapRef, the chip keeps the #31305 press handler with the contextUnknown tooltip, and the two unknown-reading tests follow the #31305 portal test. The three files match the pre-cutoff staging tip. (cherry picked from commit e03eb09)
…2369) Reopening a cron chat from History built an unbound tab named by the transcript stem: it showed cron_<id>.jsonl but saved to the dashboard:cron_<id> fallback, so the delete check refused every save and new turns were never kept. The resume now opens the job's own tab (cron_tab_name) bound to the line's own cron: link, decided before its bookkeeping keys on the name, so the next run reuses that tab and the saves stay on the cron transcript. Part of #23132 (cherry picked from commit 340b151) Backport note: cron_tab_name comes from feat #31139 (new chat per run), which is not on release, so resume.py names the tab inline with release's rule: every cron:<id>[:<run>] key shows in cron-<id>, as release's dashboard_slot_key maps it. The cron spec says cron-{job_id} for the same reason. The comment edit in chat_utils_slots.py and the test_cron_new_chat_per_run.py hunk are dropped, since both belong to #31139.
…ost (#32488) On a host whose environment refuses every name grant (an exported shell function, BASH_ENV, a relative PATH entry, a PowerShell profile), a per-command or per-program trust grant on a channel approval card is recorded but never auto-approves the next run. The card still said the command "is auto-approved ... until Kiro Crew restarts", and the card that asked again gave no reason. The approval card now carries a name_grants_refused fact, and the channel page confirms those tiers with "Approved this run - this host's shell setup means @agent will still ask each time". A card posted after a matched grant was declined ends with the refusal's constant reason. The refusal itself is unchanged. Closes #29787 Backport note: the i18n catalogs on release still carry the quoted {{cmd}} wording of trusted_channel_command (the <mono/> rewording is gains the new trusted_channel_still_asks key, and en.context.json gains only its context entry. (cherry picked from commit 3da2d94)
The star on an artifact card was disabled while its pin save ran. A disabled button leaves the tab order, so a keyboard user who pressed it lost focus to the page. Mark it aria-disabled and aria-busy instead, and claim the press in a ref so a second press before the save shows as pending is refused. (cherry picked from commit 7d0d0ad) Backport note: release ArtifactsPageGallery.tsx imports only useMemo from react (useId and the publication badge beside it come from feat #32337, which is not on release), so the import adds useEffect and useRef to it.
#32469) kiro-cli answers the loopback callback before it exchanges the code, then reports a rejected exchange only as a server_init_failure frame and closes the listener. Nothing read that frame, so a Connections card said "Checking the connection..." until the 600 s mint TTL ran out and then kept the stale "Return address delivered" line. The cold mint watcher now reads its held process's frames on each grant poll, and the warm reaper reads each adopted row's session; a failure for the row's server with no grant on disk ends the attempt as failed with reason mint_sign_in_rejected, which the card explains. A return address relayed for the card's minted approval link holds the tab's wait when it had none, so after a reload the verdict still replaces the delivered line; a chat banner's sign-in keeps reporting its own, even beside a waiting mint. A failed attempt re-reads the status feed at once, so the card leaves its waiting state instead of holding it until the next 30 s poll. (cherry picked from commit cc2b931) Backport note: the backend (mint.py, warm.py, shared_rows.py), the mint_sign_in_rejected card message, the failure-time status re-read and the relayed-address wait are taken. Release has no unprovenGrants fence (it comes from #32477, which is in another lane of this hotfix), so the fence hunks, the test that pins the fenced badge and the connections.md sentence about the fence are left out. When #32477 is applied, mint_sign_in_rejected joins its unproven filter as on main.
…ers (#32600) KAS starts a stdio MCP server with a fixed minimal environment plus the entry's own env, so the KRB5CCNAME and SSH_AUTH_SOCK the gateway arms for the KAS spawn never reached a server KAS launched. On an AL2023 host the sandboxed server fell back to an empty per-namespace keyring and reported no Kerberos ticket, and an unpooled builder-mcp or slack-mcp could not use the forwarded ssh-agent for MCS Session Forwarding. A kiro-cli MCP child inherits both. kas_env_launch.inherited_auth_env reads the two pointers from the gateway's live environment under the rules the KAS spawn applies: KRB5CCNAME as env.resolve_krb5_ccname settles it, and SSH_AUTH_SOCK only when sandbox._forward_ssh_auth_sock keeps it for the KAS spawn (the configured tier, no credential mask). relay_withheld_env writes them into the owner-only launch snapshot under the server's own env, so they never cross the wire and a key the spec sets or the wire keeps wins. A server with nothing else withheld whose snapshot cannot be written is sent as projected, so it still starts. Managed, pooled and Connection servers are unchanged. A test floor pins the relay empty so other projection tests do not depend on the runner's ssh-agent or Kerberos cache. The acp-client, security, agent-host-contract and agent-spec-fields docs describe the relay. Backport note: the hunk for acp/harness/kas_cloud.py is dropped; that harness (Kiro Cloud Sessions, #25347) is not on release. The new test pins release's ccache seam: env.usable_file_ccache is from unreleased #26373, so the gateway fixture pins the uid and env._reject_reason instead. (cherry picked from commit 178718e)
) A task-review or workflow tab writes its transcript under a linked session key that its slot name cannot reach. Once the tab closed, the Telemetry card looked its title up by slot name, found nothing, and showed the row unnamed. The dashboard's usage rows now record the transcript key the tab writes to, and the card resolves a closed tab's title through it before falling back to the slot name. The key stays server-side. No new store is added: every row the card lists was written while its tab was open. Part of #16077 Backport note: release builds the usage row as a returned dict literal and has no delivered field (#32035, unreleased), so the row becomes a local before return and only the transcript kwarg, docstring paragraph and metrics spec row are added. (cherry picked from commit d7e9257)
A channel sender's access is decided once more right before their message reaches the agent. Closes #31135 (cherry picked from commit 88e8981) Backport note: release writes request frames without the on_written=self._record_sent hook (an unreleased feature) and has no _sent_at stamp after the prompt write, so both stay out; only the before_write and prompt_gate lines were taken in client_transport.py, runtime_wire.py and session_handle_turn.py. providers/acp_turns.py also imports Callable for the new prompt_gate parameter, since release's module does not import it (main's comes from a change not on release).
This was referenced Nov 9, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem / Motivation
Goal: Ship the first half of the remaining 141 user-facing fixes of the 0.12.1 hotfix list to 0.12.1 users in one batch (slices hf4 to hf10 of the list, after #33974 carried hf1 to hf3).
These fixes are on main and not on
release/0.12.0. Each one is a bug that 0.12 users can hit today. 31 cherry-pick onto release with no conflict or apply as is once the earlier picks are in; 110 needed a hand-port onto release's own code, listed below.Why it matters
0.12 users keep hitting these bugs until a patch release carries the fixes. This batch covers chat and sessions, the gateway and service, KAS and agent specs, bots and workers, schedules, MCP and connections, channels, Windows and desktop, and accessibility, plus the 26 fixes merged to main before the 05:00 PST cutoff.
Not a goal
git cherry-pick -xof the main merge commit; a hand-port keeps release's own code, applies only that fix's change, and says what it left out in aBackport note:line of its commit message.What changed (motivation → approach → change)
GitHub cannot rebase-merge a PR of more than 100 commits, so the 147 commits of #34021 land in two PRs: this one carries its first 74 commits unchanged, and #34021 keeps the other 73. The table below lists the whole set.
The 0.12.1 hotfix list was re-reviewed PR by PR against
v0.12.0: a bug in released code, not on release yet, and not an app PR. This batch applies slices hf4 to hf10 of that list in main merge order, onecherry-pick -xof each main merge commit, onrelease/0.12.0at3e8ff05a0a, after #33974 merged. The whole list was assembled and tested as one staging branch first (see Tests); this branch's tree is byte-identical to that tested staging tree.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-ported picks. Each conflict keeps release's side and applies only the fix's own change; the full reasoning is in each commit's
Backport note:.else Nonetoelse left_out.export { selectPersonSlots }, added the CloseHold type import, removed the moved interface.final !== falseblock.Release-only commits. Written for release, not picked, each needed by a pick in this batch.
6cddb5f5a70536efda4ca0c30f6d93c24075720d9641204da593cee51946File line cap. This batch grows acp/kas_agents.py 2,008 to 2,013; env.py 2,029 to 2,052; channel.py 1,994 to 2,013 (crosses the cap), dashboard/chat_api/resume.py 2,008 to 2,028, acp/kas_agents.py 2,013 to 2,026; env.py 2,054 to 2,078 (acp/kas_agents.py back to 2,007); channel.py 2,013 to 2,016 (#33766). Release is already over 2,000 lines on
acp/kas_agents.py,dashboard/chat_api/resume.pyandenv.pywithout a baseline row, and no gate on a release PR runscheck_file_line_cap.py. Splitting a file here would be a refactor main never shipped in this form, so the batch ships as is.Dropped: none in this batch. The list's exclusions (app PRs, fixes that need an unreleased feature, regressions the release owner closed) are in the hotfix list thread.
Backwards compatibility
Compatible: each commit is the main commit unchanged, or only that fix's change applied to release's own code. Each main PR states its own compatibility.
Removes nothing: no field, API, key or config is removed beyond what each main PR already shipped and documented.
Tests
The branch tree equals the tested staging tree
a331205756(tree0d242f9d1d), built on the same release base plus #33974's commits.test/*ratchet*.py, andvitest runon the website test files they touch. All green; the one failure in the hf6 to hf8 slice is order-dependent and fails the same way on plainrelease/0.12.0.release/0.12.0or comes from the host), and the website suite.flake8,isort,mypy --platform linux,tsc -p tsconfig.app.json,gen:ui --checkand the black and comment-history gates: clean on the staging tip.Proof: not provable -- other: a cherry-pick batch of already-merged main fixes, each proven in its own main PR
Manual verification
N/A: each fix was verified in its own main PR, and this batch carries them unchanged except for the hand-ports above. The staging build that carries every batch was installed and run on Amazon Linux 2, Amazon Linux 2023 and macOS (install, gateway boot, dashboard, chat, MCP gateway, a scheduled run, restart, stop) with an isolated data home, and the collected bug reports were re-run against it with no regression.
Screenshots / video
Screenshots come from each pick's own main PR; a hand-port changes how a fix sits on release's code, not what it draws.
Related Issues
0.12.1 hotfix, slices hf4 to hf10 (after #33974): #30053, #30088, #30156, #30108, #30064, #30241, #29739, #30776, #30564, #30896, #30784, #30888, #30951, #30722, #30039, #31212, #31002, #31209, #31404, #30766, #31299, #31312, #31345, #31126, #31213, #30549, #31488, #31514, #30887, #31804, #21369, #26497, #26373, #31773, #32116, #31381, #31836, #31792, #31958, #32103, #32195, #32154, #32411, #31802, #32297, #32460, #32326, #32128, #32316, #32487, #30230, #32500, #32420, #32360, #32299, #31954, #32327, #32076, #31882, #32528, #32393, #32477, #32464, #32473, #32831, #32369, #32488, #32889, #32469, #32600, #32907, #32239, #32399, #32459, #32817, #32949, #32334, #32993, #32937, #33010, #31757, #33033, #33091, #32281, #33054, #32602, #33173, #33187, #33016, #33275, #11722, #5933, #33073, #33325, #33125, #33038, #30188, #33416, #33387, #33496, #31172, #32115, #33573, #33551, #33541, #33299, #33380, #33616, #33594, #33602, #33445, #33189, #33617, #33647, #33509, #33453, #33629, #32108, #33653, #33672, #33747, #33752, #33735, #33580, #33499, #33710, #33504, #33763, #30545, #33690, #33714, #33766, #33709, #33651, #33740, #29625, #33706, #33773, #33833, #33784, #30901.
Sibling sweep
Siblings: none -- a backport batch; each pick's own PR carried its sweep
Pattern harvest
Not generalizable: a release backport batch, not a new defect fix.
Checklist
cherry-pick -xof a main commit with a Conventional Commits title, or a release-only commit listed aboveContribution License Agreement
Contributions are accepted under the project's Apache-2.0 license; see the Licensing section of CONTRIBUTING.md.