Alexandru Dima

agentHost: await real signals instead of polling timers in the remaining agentService tests (#343603) Replace the remaining 36 timer-polling loops, and seven unreported polling loops, with the signals the tests wait for: waitForCondition now re-checks after a given change event (state manager envelopes, agent and database events), stubs complete deferreds, persistence waits use the service's idle promises or write events of the test databases, and the wall-clock spin uses a faked clock. agentService.test.ts leaves the rule's exclude list. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

agentHost: await real signals instead of polling in the local model bridge tests (#343737) Remove settle() and waitFor() from fakeLlamaServer.ts, which yielded to setImmediate a fixed number of times or until a predicate held. Status and supervisor-state waits use waitForState; the fake downloader and runtime complete deferreds; health-poll loops advance the fake clock only when a poll is pending; the dispose test awaits the supervisor's record and lock work; and negative checks await the provider's in-flight load or swap, or assert that no fake-clock timer ran. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

agentHost: await real signals instead of polling timers in unit tests (#342840) agentHost: await the real signals instead of polling timers in unit tests Replace the timer-polling loops in 10 agent host unit test files with awaits of the work each test waits for: transport sends and factory calls, connection state events, fake PTY writes, residency evictions, session state store idles, SDK setup publications, local stream reads, title controller fake calls and writes, and the worktree folder lock. Remove these files from the exclude list of local/code-no-timeout-polling-loop. agentHostDebugLogs.test.ts stays excluded: its lease-expiry test waits for a real fs.rm that the collector exposes no completion signal for. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

agentHost: await real signals instead of polling timers in copilotAgent tests (#343092) Replace the 46 timer-polling loops reported by local/code-no-timeout-polling-loop (and three unreported polling loops) with awaits of the work each test waited for: calls recorded by the fake Copilot client, the agent's proxy refresh, stub signals, the chat discovery pass and customization publications. Model-refresh retry tests run in virtual time. Remove the file from the rule's exclude list. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

processes: resolve killTree when the process exits while its tree is killed (#342451) * processes: resolve killTree when the process exits while its tree is killed taskkill /T and terminateProcess.sh kill the children before the process itself. When the process exits on its own once its children are gone, such as the cmd.exe that runs an SSH ProxyCommand, the command fails to kill it and exits with an error although the whole tree is gone. killTree rejected in that case, which made `SSH ProxyCommand > uses forceful process-tree termination only on Windows` flaky on loaded Windows machines and makes callers such as SSHProxyCommand log a spurious failure. killTree now resolves when the command fails but the process is no longer running (`process.kill(pid, 0)` reports ESRCH). The command is spawned through a new optional parameter so that a test can stand in for taskkill. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * processes: require diagnostics for ignored killTree failures Record suppressed termination-command failures through a required callback, and wire production callers to their diagnostic logs without adding console output in the base layer. Cover logging, preserved failures, and callback forwarding with focused tests. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Co-authored-by: Christof Marti <chrmarti@microsoft.com>

sessions: fix stuck Cloud tasks in shared sandbox environments (#342189) * sessions: preserve Cloud task routing in shared sandbox environments Keep task ownership per session when several Cloud tasks share a sandbox environment, so every row retains its archive, rename, and recorded-history routes. Preserve sibling tasks during discovery and deletion, and persist read-state changes for disconnected recorded conversations. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * sessions: fix shared sandbox lifecycle regressions Remove vanished task ownership independently of live session membership, reconnect through the surviving task, and allow live-only content without a recorded task owner. Cover connected editor discovery, first-task removal before reconnect, host-action fallback, and live-only history/cache behavior. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

editor: add undo stop between typed char and multi-character auto-close (#339991) * editor: add undo stop between typed char and multi-character auto-close When typing the last character of a multi-character auto-closing pair opening (e.g. `begin`), the typed character and the auto-closed text are now inserted as two edits with an undo stop in between, so that undo first removes only the auto-closed text. Single-character pairs keep the existing behavior. Fixes #338693 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * editor: guard multi-character auto-close follow-up against reentrant edits Skip inserting the auto-closed text when a content change listener reentrantly changed the model so that a cursor is no longer right after the typed character. Also compare the opening pair length against the typed character instead of 1, so single supplementary Unicode characters keep the previous behavior. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * editor: reject stale auto-close follow-up selections Skip the multi-character auto-close follow-up when a reentrant content listener changes the model version, preventing the closing text from being inserted at a stale cursor position. Add a regression test for a listener that inserts text before the typed character. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

agentHost: add experimental Computer Use support to local Copilot sessions (#342312) * agentHost: add experimental Computer Use support to local Copilot sessions Offer GitHub's Computer Use plugin to local agent-host Copilot sessions on macOS and Windows behind the experimental `chat.agentHost.computerUse.enabled` setting. - Download the pinned, GitHub-signed `github/computer-use-plugin` release for the platform on first use, verify its SHA-256, and register it with the Copilot client via `builtinPluginDirectories`. The plugin can't be bundled: product signing re-signs bundled binaries, and the macOS helper only accepts GitHub's signature. - The setting restarts the client when changed, applies only to the local desktop agent host, and enables the plugin when turned on while running. - Point the built-in computer-use MCP server at the setting in the customizations UI. - Honor the runtime's `permissionRecommendation: "approve"` for MCP tools so trusted built-in servers don't prompt twice; managed approval and a consulted assisted-approval judge still outrank it. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * agentHost: don't block startup on the Computer Use download or auto-enable it Address review feedback: - Never block Copilot client startup on the plugin download. On a cache miss the client starts without Computer Use while the plugin downloads in the background (bounded by a deadline that also aborts a stalled response body, and cancelled when the setting is turned off or the agent shuts down); the client restarts once the plugin is installed. - Stop enabling the plugin when the setting turns on. The host can't tell a user toggle from the initial configuration sync, so doing so could override a saved `/computer off`. Like the GitHub Copilot app, the setting now only offers Computer Use; enabling it stays an explicit `/computer on`. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * agentHost: don't publish cancelled Computer Use installs; scope the setting to the local host Address review feedback: - A cancelled extraction could still resolve (or fail with an unrelated read error) and then be published with a `.complete` sentinel that later starts trust. Treat cancellation during extraction as a cancellation error and check again before publishing; extract-and-publish is now a separate, tested function. - Mirror `chat.agentHost.computerUse.enabled` only to the local agent host, and ignore it on other hosts, so toggling it no longer restarts remote Copilot clients that can't use the plugin. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * agentHost: test Computer Use installing while the client starts Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * agentHost: start the client again when Computer Use installs during startup Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * agentHost: stage each Computer Use install in its own folder Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * agentHost: never ship a Computer Use plugin older than the Copilot runtime's Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * agentHost: drop the Computer Use pin check against the Copilot SDK The check compared the pin with the plugin bundled in @github/copilot-sdk, which does not protect the shared helper app. The comment now names what does: VS Code hands the MCP server an unmodified, GitHub-signed plugin. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * agentHost: update the Computer Use plugin to 0.1.95 The 0.1.95 helper trusts a copilot-runtime signed by VS Code. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Co-authored-by: Giuseppe Cianci <gcianci@microsoft.com>

chat: constrain symbol tools to active workspace (#333144) * chat: constrain symbol tools to active workspace Normalize symbol tool resources with provider-aware URI semantics and keep target resources, returned references, preview searches, and rename edits within the active workspace boundary. Add focused coverage for workspace, working directory, remote URI, and edit-result cases. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * chat: address symbol tool review feedback Localize the new error messages, clarify file path resolution, and reject rename edit types that cannot be represented in the chat edit stream. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

agentHost: await changeset signals instead of polling timers in changeset service tests (#342992) Replace the timer-polling loops in agentHostChangesetService.test.ts with awaits of the work each test waits for. Changeset status, summary counts and chat summaries are re-checked after each state manager envelope instead of on a timer; stub call counts and telemetry use DeferredPromises and an event on the capturing telemetry service. The two repository fan-out tests gate every diff and release one at a time once every started repository has parked, so the concurrency-limit check no longer guesses a delay. Remove the file from the exclude list of local/code-no-timeout-polling-loop. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>