fix: zero-pad percent-encoding for code points below 0x10 in UriTemplate
pctEncode emitted '%' + chr.toString(16) without padding, so a code point < 0x10 produced a single-digit escape (e.g. tab -> %9, newline -> %0A becomes %A, CR -> %D) instead of the RFC 3986/6570 two-digit form. The result is a malformed URI that decodeURIComponent rejects with "URI malformed". UriTemplate is used to resolve MCP and agent resource URIs, where a variable value containing a control character would yield a broken URL.
Zero-pad to two hex digits with padStart(2, '0'); this is a no-op for code points >= 0x10, so existing escapes (%20, %21, ...) are unchanged. Adds a regression test.
* sessions requiring input
* Dismiss blocked-sessions dropdown + context-specific requires-input messages
- Close the dropdown on Escape or an outside click (contextView onDOMEvent)
- Show a specific titlebar message when all blocked sessions share one reason (terminal approval, failing CI, question, unresolved comments); fall back to the generic 'N sessions require input' for a mix
- Animate the first requires-input appearance (preserve the blink class across redundant re-renders)
Removes the standalone "Subagents" dropdown above the chat input and
instead surfaces a chat's subagents as a separate group at the bottom of
the Conversations menu, scoped to the currently-active chat.
- Delete SessionAgentsControl and its CSS; drop it from ChatView.
- SessionConversationsMenuContribution now registers the active chat's
subagents in a `2_subagents` group below the session's chats.
- Add SessionActiveChatHasSubagentsContext so the Conversations menu
surfaces even when the parent is the only committed chat.
- Show the chat tab strip as soon as the session has any subagent, so
the Conversations menu moves to the tab bar.
- Add a regression test asserting a spawned subagent chat is resolvable
from the inline pill resource via parseChatUri (the Open-Subagent
contract), guarding against the pill/catalog identity desyncing.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* theme: add foreground color for problems warning icon in 2026 Light theme
css: remove opacity adjustments for warning and info labels in chat widget
* theme: update foreground color for problems warning icon in 2026 Light theme
Co-authored-by: Copilot <copilot@github.com>
---------
Co-authored-by: mrleemurray <mrleemurray@users.noreply.github.com>
Co-authored-by: Copilot <copilot@github.com>
The chat list rows span the full viewport width while the item container is centered with a max-width, leaving gutters on either side. The base list stylesheet sets cursor: pointer on rows, so the gutters showed a pointer cursor even though the item container uses cursor: default. Force default cursor on chat list rows so the gutters match.
Fixes#324173
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Some deps ship their native binary in a per-platform package declared as an
*optional* dependency of a small base package (e.g. `@openai/codex` ->
`@openai/codex-<platform>-<arch>`, `@anthropic-ai/claude-agent-sdk` ->
`@anthropic-ai/claude-agent-sdk-<platform>-<arch>`). npm does not fail when an
optional dependency can't be installed, so a transient hiccup can leave the base
package present but the per-platform package missing — and that broken tree then
gets frozen into the shared node_modules cache and served to every consumer
(see #323881).
Add build/azure-pipelines/common/checkNativeOptionalDeps.ts and run it after
`npm ci` in the linux, macOS and windows cache-build jobs (before the cache is
saved). For each configured base package that is installed, it asserts the
matching `<base>-<platform>-<arch>` package exists, and fails the job otherwise
so a poisoned cache is never saved.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Remove the system-wide keybindings first-run dialog
The first-run notice could appear in multiple windows at once on startup, and
the single-window election meant to prevent that was racy across concurrently
syncing renderers. Rather than harden the election, remove the dialog entirely:
the feature is always on and users who opt a keybinding into "systemWide" know
what they are doing.
This deletes the whole notice path, which existed only to serve the dialog:
- renderer: drop the first-run notification, the acknowledged-storage gate, and
the IDialogService/IStorageService deps; sync() now just warns about ignored
when-clauses and pushes to the main process.
- main process: drop the elected-window bookkeeping so INativeSystemWideKeybindingResult
only reports failures.
- update the systemWide keybindings.json schema description and drop the
now-obsolete election unit tests.
Global registration and trigger routing are unchanged.
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
* Add a specification doc for system-wide keybindings
Co-locate an architecture/spec document next to the renderer contribution so
future work has the full feature context (user contract, renderer<->main-process
flow, the systemWide flag plumbing, design rationale, tests, and gotchas).
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
* Bump @anthropic-ai/claude-agent-sdk to 0.3.198
Bumps the pinned Claude agent SDK from 0.3.187 to 0.3.198 following
build/agent-sdk/README.md: updates the exact version in the build
agent-sdk pin and its lockfile, plus the matching root devDependency
types pin and root lockfile so shipped types and the build-time pin
stay in lockstep.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Implement Query.reinitialize in test fakes
0.3.198 adds reinitialize() to the SDK's Query interface. Add the
matching stub to the three test doubles that implement Query
(FakeQuery, RoundTripQuery, ImmediatelyDoneQuery), following each
file's existing "not modeled" pattern.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Refine how VS Code-configured MCP servers reach the agent host:
- resolveCustomizationRefs now excludes workspace-discovered MCP servers
by default, since the agent host discovers workspace `.mcp.json` itself
(via enableConfigDiscovery). The exception is `.vscode/mcp.json`, which
the SDK does not discover: those servers are synced explicitly, but only
when their config resolves without user interaction. Resolution uses
IConfigurationResolverService.resolveAsync + ConfigurationResolverExpression
and skips servers that reference interactive variables (${input:…},
${command:…}) or fail to resolve.
- Enable `enableConfigDiscovery` when creating/resuming Copilot agent-host
sessions so the host discovers workspace `.mcp.json` on its own.
- Move MCP_CONFIGURATION_COLLECTION_ID_PREFIX and add
McpCollectionDefinition.isWorkspaceDiscovered / isVscodeMcpJson helpers to
mcpTypes so the collection-classification logic lives in one place.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Support local folder paths in Install Plugin from Source
"Install Plugin from Source" now accepts local folder paths (file:// URIs,
absolute paths, and ~-prefixed paths) in addition to GitHub shorthand and
git URLs. Local folders are inspected to route them to the right setting: a
marketplace (has marketplace.json) is registered under chat.plugins.marketplaces,
while a standalone plugin (has a single-plugin manifest) is registered under
chat.pluginLocations.
Also consolidates installPluginFromSource and installPluginFromValidatedSource
into a single result-returning installPluginFromSource; callers now surface
source-level error messages themselves.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address review: verify trust before persisting local marketplace
Declining the trust prompt for a local folder that is a marketplace no longer
persists it under chat.plugins.marketplaces; the trust check now runs before
config is written, mirroring the git path. Also fixes a chat.plugins.marketplaces
doc typo in AGENTS_PLUGINS.md.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* chat: show "Starting MCP servers…" hint for slow agent-host startup
When a new agent-host turn begins and one or more MCP servers stay in the
`starting` state past a short grace period (5s) without any content arriving
from the host, surface a lightweight progress hint listing the servers still
starting. The hint is observable-backed: it updates as servers finish and hides
itself once every server has started, content starts being received, or the
turn ends — whichever happens first. There is no interactive affordance (no
"Skip").
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
Verified end-to-end on an rt=1 enterprise account: rt/tid resolved from the minted
CAPI Copilot token, endpoint routed to the enterprise host, request.options.tools
emitted to the enhanced iKey and accepted (HTTP 200). Reverts the emit log to trace
and drops the rt-debug/POST-accepted diagnostics.
The restricted-telemetry gate read rt/tid from copilotAgent's _githubToken, which
is the GitHub OAuth token (no rt/tid claims) — so rt was always false and enhanced
telemetry never sent for anyone. Source them from the minted CAPI Copilot session
token instead (copilotApiService.resolveRestrictedTelemetryContext), mirroring how
the Copilot extension reads token.getTokenValue('rt'|'tid'). Resolved async on auth
with a synchronous disabled default; also fixes copilot_trackingId (tid) never being
set. Keeps a temporary rt-debug log for verification (to be removed before merge).
