Add a Microsoft 365 mail provider (Microsoft Graph)

[cvmilocrz]

Why

Exchange Online no longer accepts passwords over IMAP, and SMTP AUTH is off by default in most tenants. So the smtp provider can't read a Microsoft 365 inbox. Without reading replies, Harvey can't stop sequences, honor opt-outs or catch bounces, and it keeps emailing people who already answered.

What

A new native mail provider, channels.email.provider: microsoft, built on Microsoft Graph. It follows the same pattern as the Gmail one: raw REST with httpx, no SDK.

• Sign-in: harvey microsoft auth signs in once with OAuth authorization code + PKCE. It's a public client, so there's no client secret and no stored password. The rotating refresh token is saved to data/m365_token.json (mode 600 on macOS/Linux; Windows relies on folder ACLs). harvey microsoft test checks the connection.
• Sending: drafts, then sends, so the RFC Message-ID is known for bounce and reply matching.
• Replies: threaded with createReply on the prospect's own message, so In-Reply-To/References are right. If that message was deleted, the reply goes out as a new email.
• Reading: reads the last 7 days of the inbox without marking anything read or moving it. Dedup stays with the handler.
• Right mailbox only: refuses to read or send as any mailbox but persona.email. This catches signing in as the admin who approved the app, which is easy to do when a tenant requires admin consent.
• Failures hold sending: get_replies raises GraphError when the inbox can't be read, instead of returning [], so the heartbeat holds sending rather than treating an outage as "no replies".
• Loopback callback: listens on both 127.0.0.1 and ::1. localhost resolves to both, and a browser that tries ::1 first never reaches an IPv4-only listener (this hit the first real sign-in on macOS).

Setup, the admin-consent path, behaviour and troubleshooting are in docs/microsoft365.md.

Files

File	Change
harvey/integrations/microsoft.py	new: MicrosoftProvider + run_auth_flow
harvey/integrations/mail_provider.py	registers microsoft
harvey/config.py, .env.example	M365_CLIENT_ID, M365_TENANT_ID
harvey/cli.py	harvey microsoft auth|test
harvey/dashboard.py	setup checklist knows the provider
docs/microsoft365.md, README.md, CHANGELOG.md	docs

Testing

• tests/test_microsoft.py: 12 tests with Graph and the login endpoint faked through httpx.MockTransport, no network. They cover:
  • token rotation
  • draft-then-send and threaded replies
  • fallback when the original message is gone
  • draft cleanup on failure
  • inbox parsing (bounces, own mail skipped, nothing marked read)
  • raising on an unreadable inbox, including end to end through the handler
  • the wrong-account guard
  • PKCE
• Full suite: 439 passed. One failure, test_brain_provider.py::test_existing_usage_rows_become_claude, fails the same way on main without this change.
• Verified live against a real Microsoft 365 tenant: sign-in (including the admin-consent path), the wrong-account guard, and reading the inbox. No mail was sent during that check.

🤖 Generated with Claude Code

Maintainer follow-up and validation

The original Microsoft 365 implementation and commit by Camilo Cruz (@cvmilocrz) are preserved. Maintainer fixes are separate commits.

• Read every page of the seven-day inbox window, so replies and unsubscribe requests beyond the first 50 messages are processed.
• Raise GraphError on malformed or incomplete responses, including failures on later pages, so sending is held instead of treating the inbox as empty or partially read.
• Validate pagination URLs before attaching credentials; reject off-origin links, unexpected paths, loops, and over-limit pagination. Preserve empty 202/204 send/delete responses.
• Bound reads to 40 pages (about 2,000 messages at the default page size); exceeding this holds sending. See docs/microsoft365.md for limitations.
• Integrated current main, including PRs Optional pre-send compliance check with Precleared (off by default) #8 and Support newer GPT models with isolated text-only execution #15.

Local validation: 65 Microsoft-provider tests passed; the full integrated suite passed with 531 passed and 22 skipped. Regression tests were also run against the original implementation and reproduced failures. Coverage includes later-page unsubscribe handling, malformed pages, failed partial reads, unsafe pagination links, loops, and successful empty send/delete responses. No real emails or live Microsoft Graph calls were made. Live tenant OAuth was not independently retested by the maintainer. The first CI run passed all macOS/Linux jobs and exposed a POSIX-only file-mode assertion on Windows. The assertion is now platform-aware while token rotation/save coverage remains enabled everywhere; documentation explains Windows ACL requirements. All 65 focused tests passed again after this correction. Final GitHub CI results are tracked in the PR checks.