建立聯邦式 Copilot：第 4 部分，讓 Copilot Studio 接受正確的 Token

The most time-consuming part of the integration was not generating another JWT. It was proving that the particular credential sent to a Microsoft endpoint belonged at that endpoint.

We had already verified real PF exchanges and policy. Copilot Studio still returned HTTP 400. Later, sign-in worked while a connector returned HTTP 401. Each result described a different boundary.

Step 1: Establish the Microsoft environment

The project began without a usable Copilot Studio environment. We prepared Power Platform tooling and an environment plan, then worked through the actual tenant's capacity, maker access, licensing and publishing requirements.

An Azure subscription alone did not prove Dataverse capacity or a Copilot maker entitlement. The observed errors included a self-service signup requirement and User license not found while creating or saving an agent.

The practical sequence was to verify the signed-in tenant and account, inspect environment capacity, configure the approved billing or licensing arrangement, grant the maker access allowed by tenant policy, and publish a dedicated test agent.

We created a small budget with notifications. That budget was an alert mechanism, not a spending cutoff. Product licensing and trial conditions change, so reproduce these checks against current Power Platform environment requirements, rather than treating a historical portal screenshot as a license guide.

The maker account is an internal administrative identity. None of this creates Entra accounts for the external PF users.

Step 2: Put only necessary PF endpoints on public HTTPS

Microsoft's callback and the user's browser need reachable provider endpoints. We used an existing locally managed Cloudflare Tunnel on the Windows host.

The public names in this article use examples:

DNS records, edge certificates and tunnel ingress are separate objects. Creating a DNS route did not populate a locally managed connector's configuration. A wildcard certificate did not authorize wildcard application routing.

We kept deny-all fallbacks, preserved origin TLS verification and checked the connector process before restarting it. PF administration ports were not added to public ingress. The host and connector version remained explicit deployment dependencies.

Step 3: Configure Generic OAuth and publish

The agent uses manual authentication with Generic OAuth 2. Its callback for the original browser integration was:

https://token.botframework.com/.auth/web/redirect

The authorization template used fixed server configuration, a state value and the provider scope. An illustrative PF-specific query is:

?client_id={ClientId}&response_type=code&redirect_uri={RedirectUrl}&scope={Scopes}&state={State}&pfidpadapterid=consentHTMLForm&resource=urn%3Acopilot%3Alogin

The token endpoint was the PF A /as/token.oauth2 endpoint. The authorization-code body used the exact placeholders expected by the editor:

code={Code}&grant_type=authorization_code&redirect_uri={RedirectUrl}&client_id={ClientId}&client_secret={ClientSecret}

These are examples of the original dedicated Generic OAuth client configuration. They are not instructions to paste a real secret into an article or to replace the later mobile client's signed authentication with a shared secret.

The configuration review also covered the less visible fields:

The original shared Generic OAuth client was authorization-code-only. Filling in a refresh template does not enable its refresh grant or prove renewal. The new mobile client uses a separate signed-authentication refresh path and still needs live acceptance. We did not copy the mobile offline_access scope into the original agent's scope field by assumption.

We checked saved values after editing. One observed problem was a malformed saved token URL. A field that looked right before saving was not evidence of the stored configuration.

Save and publish are separate operations; Microsoft documents publication as necessary for authentication changes to take effect. We also used a fixed Boolean topic reporting User.IsLoggedIn; a greeting or a model-generated claim of federation was not a verification result.

The Native App connection string shown by the Microsoft 365 Agents SDK view was not automatically the Direct Line endpoint expected by our adapter. We treated the transport contracts separately instead of substituting URLs with similar names.

Step 4: Separate four credential types

One confirmed bug crossed these boundaries. An earlier runner posted a PF B delegated tool token to the Generic OAuth secure URL. That token passed its own local checks, but the agent was configured for the PF A provider user token.

The repair introduced a separate provider verifier and an authoritative binding from that verified provider identity to the already verified broker human. It did not invent missing actor or grant claims in the user token.

Step 5: Use the fresh OAuth card exactly

sequenceDiagram
    participant B as Trusted backend client
    participant DL as Direct Line
    participant C as Copilot Studio
    participant PF as PF A provider
    participant SP as Secure token-post endpoint
    B->>PF: Obtain and verify provider user access token
    B->>DL: Create a new conversation
    DL->>C: Begin published authentication flow
    C-->>B: OAuth card with tokenPostResource.sasUrl
    B->>B: Validate card, destination and current authority
    B->>SP: POST exact card URL with token JSON
    SP-->>B: Sign-in result
    B->>DL: Continue only after verified sign-in

The backend extracts content.tokenPostResource.sasUrl from an attachment with content type application/vnd.microsoft.card.oauth. It uses that exact URL from the new conversation and sends the configured provider token in the documented JSON property. This is the flow described in Microsoft's Generic OAuth SSO documentation.

A deliberately non-runnable request shape is:

POST <fresh-card-secure-url>
Content-Type: application/json

{"token":"<verified-PF-A-provider-user-access-token>"}

We do not reconstruct the URL, reuse a previous conversation's card or add the Direct Line Authorization header to this POST. Destination rules are pinned; an activity cannot instruct the server to send a credential to an arbitrary host.

Step 6: Diagnose HTTP 400 without guessing

We observed responses classified as expired, including the literal response Token is expired. Some posted PF tokens were locally within nbf and exp, and the card-to-POST delay was short.

That evidence does not identify which Microsoft-side object was considered expired. The secure URL, OAuth state, connection binding and token validation are distinct possibilities. We corrected the proven token-type mismatch without declaring that it explained every historical 400.

The investigation recorded token provenance, verified timing comparisons, fresh-card use, method, content type, property names and sanitized request IDs. When exact captures were explicitly requested, they were retained privately in Vault, not pasted into support messages or committed as fixtures.

A later approved test with a 60-minute PF A provider token returned HTTP 200 and the published signed-in diagnostic succeeded. We extended that dedicated provider lifetime, not the broker and tool authorization lifetimes. The result demonstrates a working lab configuration; it does not prove a universal Microsoft minimum lifetime or explain the earlier failures.

Step 7: Prove the connector's headers before using a real token

Sign-in success still did not prove delegated tool execution. The custom connector initially returned ConnectorAuthorizationError with HTTP 401.

We inspected fixed origin diagnostics containing only header-presence and outcome categories. A synthetic unbound provider value could safely distinguish missing transport from a later authorization denial.

An explicit forwarding policy did not repair the missing header in our observed connector path. Changing the managed parameter to PfUserAccessToken, updating the system-variable mapping and selecting that exact transport at the relay did.

This was a measured workaround, not a claim that all hyphenated headers fail in Power Platform. The managed profile rejects mixed or alternate header formats; it does not merge them or weaken token validation.

Step 8: Bind the provider token to delegated execution

flowchart TD
    U[Verified provider token] --> B[Private exact-token binding]
    A[Verified agent and broker authority] --> B
    C[Separately authenticated connector caller] --> R[Relay]
    B --> R
    R --> P[Current grant and policy checks]
    P --> T[Private delegated tool]
    T --> V[Check result and authority before release]

The provider token alone still grants no tool permission. The relay requires a separate caller credential and a short-lived, one-use exact-token binding created from independently verified human and agent evidence. The private tool checks purpose-specific PF proof and current policy.

The eventual live interactive diagnostic succeeded through the published connector, the real PF chain and the private tool. Independent audit confirmed human, actor, organisation, grant, mode, resource and audience continuity, exactly one successful origin operation and owned cleanup.

The credential-trace limitation remains

The PoC passed System.User.AccessToken through a tool input. We could not establish credential protection across Microsoft tool execution traces. The owner's explicit synthetic-data PoC exception allowed the experiment, but success did not resolve that protection gap.

Review Microsoft's sensitive-data limitations for the relevant action and channel. We do not treat test-pane redaction or the absence of an observed leak as a guarantee. A production user-authenticated connector would change the trust model and requires its own design review.

Part 5 adds document access through a selected-site Entra application while keeping external human and agent authorization separate.