Connor Peet

wip: connection unification

agentHost: first take of session settings and enhanced tool call approval I started this branch to just make 'Allow in this Session' work, and rabbit holed a bit down into making sessions settings visible+editable. - Sessions settings are available via a context menu action on the session. - This opens an editor backed by a new filesystem provider that maps to the agent host's settings. - Tool approval options are available as agent host-provided `options` on the tool call. Stil todo: - Register a JSON schema for agent sessions for intellisense - Make options appear in a dropdown like normal tool call options - Currently for new sessions we do an untitled->titled swap which closes settings if they're open. A very edge case but will fix if it's easy. - I think we want a more formal way to say what settings should be editable prior to a session being started (worktree etc.) and what settings are just there if the user wants them. - Currently `permissions` are just tool names, but I want to do more claude code-esque wildcard matching.

agentHost: enable @ file completions in new and existing sessions

agentHost: implement shell integration for UI side of terminals This _almost_ works, but the first terminal in a session (before OSC sequences are detected) presents malformed output to the model. Otherwise, it does the thing.

agentHost: support checkpointing and forking This enables restore/undo checkpoint for agent host sessions. It also does some initial work for forking although this is not yet fully implemented. The copilot SDK does not actually support these yet, so to do these we shut down the session, rewrite copilot SDK's disk storage, and start it back up again. It actually works. We'll need to make sure it works when we upgrade the SDK, but I don't expect it to break terribly, as the Copilot SDK folks must already be backwards-compatible to arbitrary old SDK versions that exist on the user's device, and we'd essentially just be an 'old SDK' with some dependency on internals. Of course that should all be swapped out when they eventually add proper support for it. I just flagged the specific scheme of agent host sessions thus far while developing, but will clean up prior to merging.

incremental

agentHost: actually really track (and restore) file edits Historically we've tracked edits in the ChatEditingSession, which is owned by the editor and _very_ internal and _very_ complex, for Reasons. In the agent host world, the agent host now owns edits. This is a minimal implementation of an IChatEditingSession that is used for the agent host. It does not have keep/undo (the writing has been on the wall for that for a while) which removes a large chunk of complexity. Nevertheless, it can deliver diffs, undo/redo and restore state. Diffs still happen client-side, but this could be optimized in the future. Closes #305332

agentHost: finish implementing host-level settings

agentHost: integrate terminals with tool call We need proper command detection to make this work flawlessly, but this does the initial hookup replacing the copilot SDK's terminal tools with our tools which are backed by real PTY's.

finally works v.v

agentHost: implement plan and native autopilot modes

agenthost: customization syncing data flow

agentHost: add session-specific metadata Adds a SQLite DB for session-specific metadata. Stores edits in there. It can _almost_ restore edits, but I still need to make undoStops be similarly persisted. That is a project for later this evening.

agentHost: support askQuestions/elicitation and InProgress Adopts new AHP additions to support this stuff. askQuestions work and state is updated in the session list.

cli: implement client AHP control functionality

agentHost: support running tasks in remote workspaces Adds a pluggable task-runner registry (ISessionTaskRunnerRegistry) so the agents window can dispatch session tasks through different runtimes: - WorkbenchSessionTaskRunner (priority 0) preserves today's ITaskService path for local file workspaces. - AgentHostSessionTaskRunner (priority 100) opens an agent-host terminal via IAgentHostTerminalService and sendText's the resolved command, so tasks now work in remote agent-host sessions where there's no extension host. Also adds a WorktreeCreatedTaskDispatcher contribution that runs `runOptions.runOn === 'worktreeCreated'` tasks once per session; the cloud-agent provider opts out via a new ISessionCapabilities.runsWorktreeCreatedTasks flag because its server-side host already runs those during provisioning. resolveTaskCommand handles command/args (with POSIX shell quoting honoring strong/weak/escape CommandString metadata), OS overrides, npm scripts, and dependsOn chains (sequence -> '&&', parallel -> backgrounded subshells with trailing wait). Fixes #312052 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Soft reconnect for remote agent host protocol client Persist the protocol client across transport drops. On WebSocket close or watchdog timeout, install a fresh transport via an optional constructor-supplied factory and issue a `reconnect` RPC with the existing clientId, last-seen server seq, and active subscription URIs; server replies with a replay buffer or fresh snapshots. Outgoing requests and buffered notifications gate on a DeferredPromise until the handshake completes, then drain (with dedup of optimistic actions the server already echoed). State is modeled as a discriminated union so reconnect-only fields can't leak into other states. Service-layer surfaces Reconnecting as `connecting` so the UI doesn't flicker.

wip: connection unification renderer side

agentHost: ui side of edits

agentHost: support for client-provided tools Does a half-way cleanup of state related to the active client and supports client-provided tools -- initially the runTests tool, task tools, and problems tools.