[sambitcreate]

Bots can watch connected Mail or Calendar, track accepted tasks in Open work, use a chosen model for each routine, and ask approved Bot peers for help. Accepted work and owed results survive restart while each recipient retains its own rules, access and spending limits.

• Add durable watch transitions, crash-safe hand-off and polling while Aiden runs and the owning host is awake. Desktop, iOS and Android can discover, preview, create, edit, pause/resume, run and remove watches. Composio Calendar works on Mac/Linux Electron hosts; iMCP Mail requires a Mac.
• Track explicit tasks with evidence, edit/dismiss/Undo and request-scoped Stop.
• Freeze routine model/thinking/fallback choices per accepted run. An unavailable explicit model stops with Needs you; fallback runs only when configured. Retain provider-cost obligations without repeating uncertain requests.
• Add named pair approval/revocation, protected delegation, linked exchanges and exact-request controls on all three clients. Owed final delivery survives pruning, quiet hours, channel failure and Bot replacement.
• Present Composio as one tool with one service allowance. Existing partial grants require an explicit service choice. Watch reads construct one curated operation with an immutable account/calendar binding; current source authority and connector schemas fence reads and durable publication.

The owner-authorized live Composio check passed exact fabricated-account refusal, discovered seven readable calendars, read real events and completed a matching GET of an actually returned timed occurrence. The reviewed Calendar adapter is now installed lazily at Electron startup without opening connections. Diagnostic files never authorize a source or warm the watch connection. Read compatibility validates a bounded copy; original replies still feed identity, occurrence, pagination and incomplete-result guards. The signed foreground check remains separate from later private feature formats.

Validation

Every phase passed the requested GPT-6.1 Sol high fix/review loops. Final integration review independently passed 252 focused cases and four Cua journeys; the final messaging core review also passed all three original failure reproductions and 50 focused cases.

Surface	Local result
Host	Runtime 319; proactive 505; Open work 51; routine models 120; peers 56; Telegram 281; spending 33; usage limits 31 — passed
Desktop	All four feature journeys passed; final watch management, Calendar preview/edit/replay, onboarding tour and Composio Settings journeys passed 4/4 with retries disabled
Android	1,261 JVM tests, instrumentation compilation and lint passed; Calendar controller flow 15/15
iOS	Adjacent run 395 passed / 4 skipped; peer conflict run 42 passed / 1 existing Simulator skip; Calendar flow 29/29; chat fixture class 365/365
CLI	Final build/types and bundle/runtime/parity/inbox 119/119 passed; earlier full run 182 passed / 3 existing skips
Contracts	Remote 57, closed DTO/IPC/OpenAPI, older-client projections and incremental SSE checked

Actual whole-project ES2020 types, E2E types, scoped lint and CI policy/collection checks pass. React Doctor reports no issues. The authenticated foreground snapshot passed its full build and 423 cases. The CLI does not advertise the optional new management surfaces.

CI findings and remaining release acceptance

Latest published head is f7125cec9. Its CI found two fixture-timing failures: the native Bot browser masking check passed its existing retry and failed the flaky gate; the iOS opening-roster test missed its hold expectation. Both are being reproduced and corrected without production timeout/retry changes. Other completed surfaces are green; Linux x64 is still running. This remains a draft pending corrected-head CI and the documented release acceptance.

• The earlier iOS opening-roster failure exposed fixture epoch and callback-delivery races. Deterministic reproductions now pass, independent XCTest is 6/6 and the full chat class is 365/365. The exact hosted interleaving was not recorded.
• Electron deadline failures led to four measured single-worker shards, preserving the 20-minute limit, retries and flaky-test gate. All four passed on b278f4e4b in roughly 12.5–13.5 minutes. Extra runner/setup cost is documented.
• The Linux Usage data persistence flake passed its existing retry and correctly failed the flaky-test gate. Its poll now treats only the missing initial settings file as pending; persisted false and relaunch checks remain. A deterministic filesystem RED and independent Cua 2/2 GREEN establish the fix.
• Local full-suite validation exposed a fake-clock/durable-timer sequencing race. A delayed mutation reproduced the exact missing-event failure. The test now synchronizes existing dispatcher idle work and checks exact ordering, the full five-minute gap and durable receipts. Independent events 26/26 and full proactive 505/505 pass. No production timeout, retry or scheduling behavior changed.

Earlier CI inventory, fixture and ES2020 compiler failures were corrected locally and independently reviewed. No failed hosted job was rerun unchanged.

Packaged/physical-device lifecycle and live iMCP Mail acceptance remain release gates. Three earlier local iOS journal tests observed unexpectedly missing pending attempts; later passes did not establish the cause: testPublishedLostResponseRestartSameBodyReplayWithoutNewApproval, testStopAcceptedRequestedKeepsOriginalBodyUntilMatchingTerminalAcknowledgement, and testStopLostReplyRestartSameRequestAndBlockedReceiptDoesNotClaimCancelled. Their evidence and Simulator limits remain in iOS contract checks.

See the implementation and acceptance plan and authenticated Composio proof.