Skip to content

Systemorph/MeshWeaver / .claude/skills/teams

Microsoft Teams access from MeshWeaver — what the shipped Teams integration actually is (a bot people message; it reads and sends NOTHING on a user's behalf), and what reading or sending a user's teams, channels and chats needs — the delegated Graph scopes on the Executive Assistant consent link, the tools on the EA plugin, the consent rules, and the TENANT wall that makes "can you see the client's team?" a different question from "can you see my teams?". Use when someone asks whether an agent can see, read, summarise or post to Teams, when adding Teams scopes or tools, or when a Teams-related answer from an agent has to be checked against what is wired.

/teams — reading and sending Teams on a user's behalf

Measured 2026-09-09 by asking the Executive Assistant on the control instance "can you see my Teams channels, in particular the enterprise client's team?": it answered no, and it was right — then. Since Mail 1.5/1.6 (read, 2026-09-09/10) and Mail 1.7 (post, 2026-09-10) the EA reads a user's own-tenant teams, channels and chats, and posts to them where the deployment sets Teams:AgentSend=Send. This skill records what exists, what the ask needs, and the one wall that no scope can climb.

What exists: a bot, not a reader

MeshWeaver.Teams (MeshWeaver.Plugins, src/MeshWeaver.Teams/) is the Bot Framework channel documented in Doc/AI/TeamsBot: a person messages the bot in Teams, TeamsInboundProcessor finds-or-creates one agent thread per conversation, TeamsReplySender posts the agent's reply back. Its outbound credential is an app-only connector token for api.botframework.com — it can answer in a conversation the bot is part of and nothing else. On the control instance it is inert (Teams:Enabled=false, zero TeamsConversation nodes).

The Executive Assistant (src/MeshWeaver.Mail.MicrosoftGraph/ExecutiveAssistantPlugin.cs) has mail, calendar and Teams tools. Its credential is the delegated one minted by the EA consent link, and that credential is the seam everything below hangs on.

The seam: the EA consent link IS a delegated Graph token

{BaseUrl}/auth/ea/connect sends the user through Microsoft's authorize endpoint with the scope string in EaGraphAuth.Scopes — this repo, memex/Memex.Portal.Shared/Authentication/EaGraphAuth.cs:

Mail.ReadWrite Mail.Send Calendars.ReadWrite
Team.ReadBasic.All Channel.ReadBasic.All ChannelMessage.Read.All Chat.Read
ChannelMessage.Send ChatMessage.Send offline_access

🚨 Widening that string invalidates every stored grant. Entra refuses to redeem a refresh token for scopes the user never consented to (400 invalid_grant), and the consent controller's fast path bounces a user whose credential node exists straight back — so on 2026-09-10 the read scopes landed, the reconnect link "just brought me in", and every Teams call kept failing. EaGraphAuth.Classify now compares the credential's stored Scopes with the constant and reports a mismatch as NotConnected, which hands the user the link AND makes /auth/ea/connect run the dialog. Do not add a scope without that classification in place; ?force=true on the connect link is the manual escape hatch for a portal that predates it.

Add Teams scopes to that string and every user reconnects once; from then on the same GraphServiceClient the plugin already builds reaches /me/joinedTeams, /teams/{id}/channels, /teams/{id}/channels/{id}/messages, /me/chats, /chats/{id}/messages — as the user. Nothing app-only, no new secret, no new token store. That is the whole point of routing it through the EA.

Read

Scope (delegated)Lets the EAAdmin consent
Team.ReadBasic.Alllist the user's teamsuser-consentable
Channel.ReadBasic.Alllist a team's channelsuser-consentable
ChannelMessage.Read.Allread channel messagesadmin consent required
Chat.Readlist and read the user's chatsuser-consentable

Send

Scope (delegated)Lets the EAAdmin consent
ChannelMessage.Sendpost to a channel as the useruser-consentable
ChatMessage.Sendpost in a chat as the useruser-consentable

Re-verify the consent column against the Graph permissions reference before adding a scope — a tenant's consent policy can make any of them admin-only. The Entra sign-in app registration must list the same delegated permissions, or the authorize call fails before the user sees a screen.

🚨 Teams has no draft state. Mail is DraftOnly by design (Email:AgentSend, the human presses Send); a Teams post is immediate and irreversible. The gate is the same shape: Teams:AgentSend defaults to Off, in which PostChannelMessage / ReplyToChannelMessage / SendChatMessage are never handed to the model (and refuse by name when reached directly); Send hands them over, with no per-message confirmation — the boundary Email:AgentSend=Send has too. The scopes are consented regardless of the mode, so flipping the mode later costs nobody a reconnect. The public instance runs Teams__AgentSend: "Send" (its deployment values); the control instance does not.

The three pieces (built 2026-09-09/10)

  1. This repo: the scopes in EaGraphAuth.Scopes plus the stale-grant classification above; the delegated permissions on the app registration (Roland's Entra tenant for the Systemorph portals) are optional under dynamic consent except the admin-restricted ChannelMessage.Read.All.
  2. Plugins repo: the tools on ExecutiveAssistantPlugin — read: ListTeams, ListChannels, ReadChannelMessages, ListChats, ReadChat (every mode); post: PostChannelMessage, ReplyToChannelMessage, SendChatMessage (Teams:AgentSend=Send only). 🚨 A tool exists for the model only when CreateTools() LISTS it — Mail 1.5 shipped the read methods without listing them and the EA answered "not in my toolset" (Plugins#1611).
  3. Plugins repo: src/MeshWeaver.AI/Data/Agent/ExecutiveAssistant.md names the surface, or the agent keeps answering "I have no Teams tool" from its instructions.

🚨 The wall: a client's team lives in the client's tenant

A delegated token is minted by the user's home tenant. A team the user reaches as a guest (B2B) lives in the resource tenant, and Graph does not cross that line on a home-tenant token: /me/joinedTeams does not list guest teams, and the channel endpoints answer 403/404 for them — measured and documented by others (Q&A, Tech Community). So the scopes above make the EA see Systemorph's teams. A client's team (e.g. Globex/Team) needs one of:

  • A token for the client's tenant. Run the same consent flow against login.microsoftonline.com/<tenant-id> (the app registration must be multi-tenant), store the resulting refresh token per (user, tenant), and route calls for that team through it. ChannelMessage.Read.All then needs the client's admin to consent to the Memex app in their tenant — a conversation with the client's IT owner and their security lead, not a setting on our side.
  • The bot, installed in their team. The existing MeshWeaver.Teams bot with resource-specific consent (ChannelMessage.Read.Group) in its manifest, added to the client's team by a team owner. Messages are then PUSHED to /api/teams/messages as they happen; nothing is enumerated. Needs Teams:Enabled on the receiving portal and the client's app-upload policy to allow it.
  • Their export. Ask for a channel export; ingest it as content. No integration at all.

Do not promise "I'll read the client's channel" until one of these exists; the honest answer today is the Graph-visible substitute — the mailbox — which the EA already gives.

Checking what an agent says about Teams

The EA's answer is only as good as its tool list. Before trusting a "no Teams access" or a "connected" claim: get @Agent/ExecutiveAssistant/data/ for the plugins, then read ExecutiveAssistantPlugin.cs for the tools actually exposed, then EaGraphAuth.Scopes for what the token can carry. All three moved independently in the past; the doc Doc/AI/TeamsBot describes the bot only.

reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 466kUpdated Claude CodeCodexUnverifiedInstall
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 466kUpdated Claude CodeCodexUnverifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 319kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 319kUpdated Claude CodeCodexVerifiedInstall