Microsoft 365 Copilot desktop CDP

Outcome

Use the existing desktop session and exact requested mode/model to obtain the requested conversation state or a completed new assistant reply. The next consumer receives JSON containing the actual rendered result and its turn identity. A Send click, an empty composer, an old reply, or stable text is not completion evidence.

Chat and Code have separate model/completion contracts. Code needs its native host, not a generic browser facade. Read the UI contract when modes, models, selectors, or completion evidence differ.

Authority

• Inspect/read only what the user requested. Sending a prompt requires explicit authorization for its exact purpose; app access alone is not send authority.
• Keep existing drafts. The adapter rejects composing or changing modes with a nonempty draft. A pre-existing menu must be closed manually before model discovery/selection.
• Code can modify files, run commands, install packages, and connect to services. A prompt authorization does not authorize those actions, project creation, onboarding acceptance, tool approvals, or persistent permission changes.
• After any error from a Send attempt, treat the result as uncertain. Read the conversation and ask the user to reconcile it; never automatically resend.
• Use visible UI roles/test IDs. Keep normal sandboxing, organization policy, authentication, and connection/permission prompts intact. Reuse the existing signed-in profile; do not create independent profiles or copy authentication material. Do not edit profile settings unless the user specifically authorizes the documented setting and restart procedure in setup.md.
• Use a dedicated native Copilot window only when its ownership and independence are established. Native Chat/Code dual-window operation was not established on the observed build; the current session runner is sequential and holds a browser-wide writer lock. Coordinate exclusive use of the selected window. Do not substitute generic CDP-created windows or automate alongside human input.
• Store request files, output, HTML, screenshots, traces, and other private artifacts only under .tmp-artifacts/<bundle>/. Do not put personal conversation data in source files or tracked documentation.

Prerequisites and commands

Use Windows and Bun >=1.3 (verified with 1.4.4). Run bun install --frozen-lockfile in skills/m365-copilot-cdp/scripts/; with only Node, substitute npx --yes bun. No global Playwright install or browser download is needed.

The independent app must already be running with a loopback-only CDP endpoint and a working signed-in native session. Prefer --profile-root pointing to the existing profile: it reads only DevToolsActivePort to find the current dynamic browser WebSocket endpoint. An explicit local --endpoint is also supported. The CLI attaches existing targets and disconnects without closing the app; it never creates pages or navigates a target to another service.

For startup, consent, or unavailable-runtime issues, read desktop setup and recovery. Do not use a PWA, isolated profile, authentication copy, or policy bypass as a substitute.

From the repository root, prefer one persistent connection:

bun skills/m365-copilot-cdp/scripts/copilot.ts --help
$profile = Join-Path $env:LOCALAPPDATA 'Microsoft/Copilot/User Data'
bun skills/m365-copilot-cdp/scripts/copilot.ts --profile-root "$profile" --session

If the native app asks Allow remote debugging?, explain that this connection will control the signed-in Copilot window and have the user choose Allow only for this expected connection. Then keep the process alive. Its initial JSON contains ready, current targets, and JSONL protocol metadata; it is not an app operation result. Send one JSON object per stdin line:

{"operation":"select-target","target":"EXACT_CDP_TARGET_ID_FROM_READY_OUTPUT"}
{"operation":"inspect"}
{"operation":"models","mode":"code"}
{"operation":"read","mode":"code"}
{"operation":"disconnect"}

{"operation":"targets"} refreshes discovery within the same connection. Every app operation requires an explicitly selected exact ID; invalid selection clears the old one. The working webview target is on cr.m365copilotapp.svc.cloud.microsoft, not the edge://olympia.top-chrome/ shell.

Reusing this connection avoids approval for every command. There is no verified forever grant: disconnecting or restarting may require fresh consent. The observed enableCdpWl flag was false during a successful native Code interaction; it is not required by this workflow and its full semantics are undocumented. Do not toggle it or suppress consent prompts.

For a one-shot request, use --profile-root "$profile" --target EXACT_ID --request .tmp-artifacts/copilot-run/request.json; --request - reads JSON from stdin. That separate connection may ask for approval again, so do not make a separate --targets connection the default discovery workflow.

Results use { "ok": true, "result": ... }; session errors use { "ok": false, "error": ... } on stdout without disconnecting. One-shot/startup failures use stderr and nonzero exit.

Requests

Read the request and result contract before constructing requests. Unknown fields, unsupported operations/modes, blank text, and inexact model labels fail before composing.

1. {"operation":"inspect"} identifies the current surface, selected model, draft presence, and any Code prerequisite without exposing draft text.
2. {"operation":"models","mode":"chat"} lists exact selectable leaf labels from the live model menus. Use "code" for the separate Code model menu. Reuse an exact label; never substitute a nearby model or silently fall back.
3. {"operation":"read","mode":"chat"} reads the active conversation without switching it. Use "code" for the active Code session. Incomplete replies are marked incomplete; pending permissions are reported without approval.
4. {"operation":"send","mode":"chat","model":"EXACT LABEL","text":"AUTHORIZED PROMPT"} selects and rechecks the exact model and mode, sends once, and waits for a new prompt-linked assistant turn with mode-specific completion evidence. Use "code" only for an authorized actual Code interaction. Omit model only when retaining the verified active model is intended.

For Code, inspect the actual Code surface. If onboarding or models are unavailable, stop with the explicit prerequisite. Do not start a project, approve execution, or switch to Chat to satisfy the request.

Completion and handoff

Report mode/model, rendered turn ID, reply, and completion signals. Code IDs are UI identities, not execution IDs: native history replay can duplicate display records. Do not infer execution counts, deduplicate text, or resend from them. Report uncertain sends and unavailable UI explicitly; a read is not a new send.

Basis

Observed English builds, live acceptance, and the native replay limitation are recorded in UI evidence.