Repository navigation
Let chat orchestration propose, create, run and hand off saved workflows #1543
Description
Summary
Let V2 chat orchestration decide when a saved workflow is the better way to handle a request, then help the user set it up, run it, and get the results back in chat. This umbrella tracks the phased plan in docs/explanation/features/CHAT_ORCHESTRATION_WORKFLOWS_ROADMAP.md.
User Value
Chat orchestration handles one-time requests well. It can't repeat work on a schedule, react to new File Sync documents, or run work bigger than a plan allows (8 steps by default, 180 s per step, 900 s per run by default and 7,200 s at most, no loops). Durable workflows already do all of that, but today the user has to know to build one by hand.
With this work:
- "Every Monday read my email and tell me what to do this week" answers for this week now, and offers a card that creates the recurring workflow.
- "Review new documents from File Sync" proposes an event-driven workflow.
- "Run my weekly digest now" runs an existing workflow and posts the answer back into the chat when it finishes, even hours later with the browser closed.
- A 200-document review that's too big for a plan is handed off to a one-time durable workflow that posts a summary back.
Proposed Behavior
One PR per item. Each item is planned on its own when it starts (roadmap §10).
| Item | Scope | Depends on |
|---|---|---|
| Phase 1 (#1544) | Calendar schedules: day, time and IANA timezone | — |
| Phase 2 (#1545) | Workflow draft service: dry-run build, blueprint builder, provenance | 1 |
| Phase 3 (#1548) | AI workflow assistant in the V2 editor: highlighted AI changes, per-turn undo, save confirmation | 2, A1, A2 |
| Phase 4 (#1547) | Orchestration proposes workflows: workflow_propose plus an Approve / Deny / Edit card |
2 |
| Phase 5 (#1551) | Orchestration runs existing durable workflows: workflow_run, start-and-link |
4 |
| Phase 6 (#1546) | Results back in chat: 6a results reader and Follow up; 6b post-back delivery, run card and chat-list indicator; 6c in-plan wait (later) | 6a: 4 · 6b: 5, 6a, N1 |
| Phase 7 (#1549) | Hand-off of big one-time jobs | 4, 6b |
| Phase 8 (#1550) | Follow-ons: group workflows, #1347 parity, plan-replay task | 4+ |
| A1 (#1552) | Shared AI-assist thread: your message appears immediately, Cancel and Retry, one component for all four assist editors | — |
| A2 (#1556) | # document references in AI-assist inputs (plan editor now, workflow assistant in Phase 3) |
A1 |
| N1 (#1554) | V2 notification bell and panel; the Desktop notifications preference works in V2 | — |
| N2 (#1553) | Animated workflow alerts in V2 | N1 |
| P (#1555) | Document provenance for documents that workflows and chats create | — |
Key decisions (roadmap §2):
- Personal workflows only in v1. Group workflows are Phase 8.
- Creating a workflow from chat always needs a manual decision on a card, even when plan approval is Auto. Plans that run a saved workflow or hand off work always need manual plan approval.
- The plan never waits for a workflow run. It ends once the run starts, and the server posts the results into the chat later. Results are never posted into shared or collaborative chats.
- The chat capability sits behind a new admin setting that's off by default.
- Planner and assistant output is untrusted. Only server-side deterministic builders create workflow definitions, and only request-thread routes with fresh authorization save them.
Acceptance Criteria
- Every child issue is closed or explicitly deferred.
- The roadmap doc's status table matches what shipped.
- In V2 chat, "every Monday read my email…" returns this week's answer plus a proposal card, and Approve creates exactly one workflow that runs Monday at the chosen local time and notifies.
- "Run my weekly digest now" asks for manual approval, starts exactly one run, and posts exactly one answer back into the private chat when the run finishes.
- A one-time job beyond plan limits is handed off to a durable workflow and posts one summary back.
Notes
- Roadmap sections: §6 phase detail, §8 gotchas, §9 open decisions with recommendations, §11 validation.
- Related issues:
- Update the SimpleChat action for task-based workflows and reorganize its capability list #1347 SimpleChat action for task-based workflows (Phase 8 builds on the draft service)
- Split Microsoft 365 actions with sharing approvals, live file RAG, and delegated workflow execution #1493 Microsoft 365 split and delegated workflow execution (Run-as and M365 sources)
- Design generic monitor trigger model for workflows #949 Generic monitor trigger model (stays a separate design)
- Generalize workflows into repeatable AI automations without required documents #1082 Generalize workflows into repeatable AI automations
- Complete the orchestration harness with Gather, Reason, Render and explicit file creation #1509 Orchestration harness: Gather, Reason, Render
- Introduce turn-level orchestration across chat capabilities #1021 Turn-level orchestration across chat capabilities
- Every row above is a sub-issue of this issue, with blocked-by links that match the "Depends on" column.
Metadata
Metadata
Assignees
Type
Projects
- Status
Activity
31 remaining items
Update roadmap status: Phase 7's settled 7a design and two follow-ups
Align the roadmap with container-only result access and self-authored…
Set the run card harness's first stream number through the constructor
Re-prove a hand-off report's lineage before reading it
Document the hand-off report's lineage re-proof
Refresh the Phase 4 schema digest pin after microsoft#1641
Show workflow hand-offs under a chat answer in V2 (Phase 7b)
Document the V2 workflow hand-off card (Phase 7b)
Bump version to 0.261.256 for the V2 workflow hand-off card (Phase 7b)
Update roadmap status: 7b done; the capability is complete
paullizer commentedon Oct 8, 2026
Status update: 2026-10-08
Everything below landed in
paullizer-react-v2-ui. With Phases 6 and 7 merged, chat can propose, create, run and hand off workflows, and their results come back into the chat. The chat features are behind four admin settings, all off by default: Propose Workflows From Chat, Run Workflows From Chat, Hand Off Large Work From Chat and Use Workflow Results In Chat.Merged into
paullizer-react-v2-uiworkflow_resultsplanner capability#references in assist inputsAlso merged: #1597 cleared the 19
py/log-injectionalerts (0.261.216), #1598 fixed test-harness drift, and the status PRs #1595, #1601, #1627, #1633, #1648 and #1707 kept the roadmap up to date.Optional, not started
Acceptance criteria
Verified from tests and review, not on a live tenant.
Child issues
These PRs target
paullizer-react-v2-ui, not the default branch, so theirCloseskeywords didn't close anything. Everything these issues track has merged, so they're ready to close: #1544, #1545, #1546, #1547, #1548, #1549, #1551, #1552, #1553, #1554, #1555 and #1556. #1546 defers 6c, and #1550 (Phase 8) is optional and not started.Once the child issues are closed and #1550 is deferred or moved out, this issue can close too.
Give replay workflows a default 'Run failed' alert
Bump version to 0.261.305 for plan replay
Document workflow plan replay
Add the wait-for-quick-workflows setting and its wait cap