fix(outlook): find scheduled sends outside Drafts

[stevenobiajulu]

Fixes #204.

The bug, confirmed live

Read-only Graph check against a real mailbox — the Drafts-scoped query the provider used found 1 pending send; a mailbox-wide scan found 4:

DEFERRED "OpenAgreements lab update"                            2026-09-08T13:00:00Z  folder=Drafts
DEFERRED "RE: Accord Project GSoC Agentic Workflow project"     2026-09-08T15:00:00Z  folder=Deleted Items
DEFERRED "RE: Rubicon legal status monitor - next steps"        2026-09-09T15:00:00Z  folder=Deleted Items
DEFERRED "Re: Quick Intro Request …"                            2025-09-16T13:00:00Z  folder=Deleted Items

Outlook's own Schedule send parks the deferred message outside Drafts, so folder membership says nothing about whether a send is pending. Note the last row: a past deferred time, i.e. residue Outlook left behind after delivery.

The fix

listScheduledSends now scans /messages mailbox-wide, narrowed to $filter=isDraft eq true, and matches the deferred property client-side.

Server-side filtering on the property is not available — all three shapes were probed against live Graph and rejected:

filter	result
svep/any(ep: ep/id eq 'SystemTime 0x3FEF')	400 ErrorInvalidUrlQueryFilter
… and ep/value gt '2026-09-07T00:00:00Z'	400 ErrorInvalidUrlQueryFilter
… and casted(ep/value, Edm.DateTimeOffset) gt …	400 BadRequest — unknown function casted

So isDraft eq true is what keeps the scan bounded; a pending deferred send is always still a draft, and without it this would page the entire mailbox.

Past-dated matches are now dropped. The Drafts-scoped scan rarely saw that residue; a mailbox-wide one sees all of it, and reporting a year-old delivered message as "pending" would be worse than the bug being fixed.

Page size 500, measured on the 1,200-draft mailbox above: 3 pages / ~7.5s, vs 13 pages / ~9.2s at 100 and 2 pages / ~8.3s at 999 — Graph's per-item cost dominates, so bigger pages stop paying off.

Verification

• npm run test:run — all workspaces green (511 + 236 + 116 + 219 + 31); npm run lint clean; openspec validate add-scheduled-send --strict passes.
• New regression test covers an Outlook-UI-scheduled message whose parent is Deleted Items, alongside the MCP-created variant, plus the past-dated exclusion. Existing scenarios keep their fixtures against a pinned clock.
• Live smoke through the built MCP server against the real mailbox: list_scheduled_sends returns the 3 pending sends (was 1), residue excluded, stable across repeat calls.

Spec delta in openspec/changes/add-scheduled-send/ updated — it previously said listing happens "only from Drafts".

https://claude.ai/code/session_01A7i4dE7jfKMCEM5BRoTwtw

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ All tests successful. No failed tests found.

📢 Thoughts on this report? Let us know!

[stevenobiajulu]

✅ Post-merge smoke passed

Merged: b83046dae5b00ca280ef605c5575e9e11974b5b8
Built from: main @ b83046d
Smoke: clean build → lint → full suite → live list_scheduled_sends through the built MCP server

Steps

• ✅ build (clean — npm run clean --workspaces then npm run build, no stale dist from the feature branch)
• ✅ lint (all workspaces)
• ✅ tests (511 + 236 + 116 + 219 + 31 passed) — see flake note below
• ✅ feature path: list_scheduled_sends over stdio against the real Microsoft mailbox

Feature-path result (live mailbox, read-only)

list_scheduled_sends: 3 pending in 13391ms
  FUTURE 2026-09-08T13:00:00.000Z  to=1 recipient(s)  subject="OpenAgreements lab update"
  FUTURE 2026-09-08T15:00:00.000Z  to=1 recipient(s)  subject="RE: Accord Project GSoC Agentic Workflow"
  FUTURE 2026-09-09T15:00:00.000Z  to=1 recipient(s)  subject="RE: Rubicon legal status monitor - next "
gmail arm: error.code=NOT_SUPPORTED
SMOKE PASS

Pre-fix this returned 1 (the Drafts-resident one). The two RE: rows are the Outlook-UI-scheduled messages living in Deleted Items — exactly the miss #204 reported. Every returned scheduledSendAt asserted in the future, so the past-dated delivered residue stays excluded. Gmail still fails closed with NOT_SUPPORTED.

Flake, unrelated to this PR: the first full run had email-core > security/WatchedAllowlist > Scenario: Resets config on file deletion fail on waitFor timed out (2009ms). It's an fs-watch timing test in a package this PR does not touch; it passed 3/3 in isolation and green on the full-suite retry, and CI was green. Flagging rather than swallowing it — the watcher test looks timing-sensitive under load and may be worth a separate issue.

Not live yet for agents: this repo publishes on v*.*.* tags, not on merge. The globally-installed server that MCP clients actually run is email-agent-mcp@0.3.0 at /opt/homebrew/bin/email-agent-mcp, so the fix reaches agent sessions only after a release + reinstall.

Prod: n/a — repo does not deploy on merge (release workflow is tag-triggered).
Log: /tmp/automerge-smoke-205-20260907-014434.log (local).

Cleanup

• ✅ deleted remote branch fix/204-scheduled-sends-mailbox-wide
• ✅ removed worktree /Users/stevenobiajulu/Projects/email-agent-mcp-wt-204
• ✅ deleted local branch + pruned refs

https://claude.ai/code/session_01A7i4dE7jfKMCEM5BRoTwtw