[raymondk]

Stacked on #232 (deploy-3). Fourth step of the deploy stack, and the first of the work that brings the deploy tool to the hosted server: the call shape both servers will share.

What

icp_deploy_bundle stays one blocking call, but now reports each phase as it completes and stops when the call is cancelled.

Why

A deploy lasts from seconds to minutes (canister creation, installs, a frontend's asset sync) and the call said nothing until it ended. On the hosted server a silent minutes-long call is also what trips a client's per-call timeout. Progress notifications fix both: the client sees each phase, and the MCP SDK clients reset their timeout on every progress notification.

How

• deploy::Watch is handed to deploy_bundle: a stop token and an optional lines sender. collect_log sends each log line the moment it is made, so the outcome's log and the streamed lines are the same lines in the same order (test: the_log_is_streamed_as_it_is_made).
• The run is raced against stop with a biased select!: cancelling drops the run at its next await, which is the next canister call or plugin step. A canister call already in flight completes on the network. The outcome reports a new status, interrupted, with every canister created before the stop in created and canister_ids, like a failure does. The tool returns it as an error result, like a failure.
• An interrupted deploy keeps its working directory (test: an_interrupted_deploy_stops_and_keeps_its_directory): whoever asked for the stop may be a client that went away and will never read the answer, so the directory's mapping file stays as the record, as it does for a run that is cut off.
• The tool forwards each line to the client mid-call: as a progress notification when the request carried a progress token, else as a logging message. MCP logging is deprecated (SEP-2577) and rmcp marks it so; it is used only as the fallback, under #[allow(deprecated)] with the reason. The local server declares the logging capability so a client expects it.
• stop is the request context's cancellation token, which rmcp cancels on the client's notifications/cancelled. The hosted server will add the connection going away as a second source in a later PR.
• Skill, tool description, server instructions and README say the call lasts as long as the deploy, to send a progress token, and what cancelling does.

Not here: the hosted composition, the bundle upload, concurrency across sessions. Those are the next PRs on this stack.

🤖 Generated with Claude Code