chat: gate agent host input completions on trigger characters

[connor4312]

chat: gate agent host input completions on trigger characters

Agent host completion providers were re-invoking provideChatInputCompletions
on every keystroke, producing one RPC round-trip per character typed because
Monaco still consults the provider for filtering / incomplete-result refresh
even when the user isn't editing a trigger-led token.

• Adds isAtTriggerCharacterToken helper in chatInputCompletionUtils.ts
  that returns true only when the cursor sits inside a non-empty token whose
  first char is one of the host-announced trigger characters.
• Wires the gate into both agent host completion providers (workbench chat
  widget + sessions new-chat input) so the host RPC is only dispatched when
  the user is actively editing a @/#/... led token.
• Extracts a small AgentHostInputCompletionsBase template-method base
  class that owns the shared plumbing (trigger-char gate, RPC dispatch,
  range computation, Monaco provider registration). Concrete subclasses
  now only supply _resolveContext (how to find the active session
  resource for a model) and _buildItem (how to build the Monaco
  completion item + accept command), keeping the registration lifecycle
  policy in each subclass.

(Commit message generated by Copilot)

[Copilot]

Pull request overview

This PR reduces unnecessary agent-host completion RPC traffic by gating provideChatInputCompletions calls so they only run when the cursor is within a trigger-led token (e.g. @...). It also factors out shared Monaco-provider plumbing into a reusable base class used by both the workbench chat widget and the sessions (Agents window) chat input.

Changes:

• Added isAtTriggerCharacterToken helper to detect when the cursor is in a trigger-led token.
• Introduced AgentHostInputCompletionsBase to share provider registration, trigger gating, RPC dispatch, and range computation.
• Updated both agent-host completion handlers (workbench chat input + sessions new-chat input) to use the shared base.

Show a summary per file

File	Description
src/vs/workbench/contrib/chat/browser/widget/input/editor/chatInputCompletionUtils.ts	Adds trigger-token detection helper used to gate RPC-backed completions.
src/vs/workbench/contrib/chat/browser/widget/input/editor/agentHostInputCompletionsBase.ts	New shared base class for agent-host input completion providers.
src/vs/workbench/contrib/chat/browser/widget/input/editor/agentHostInputCompletions.ts	Refactors workbench agent-host completion provider to use the shared base + gating.
src/vs/sessions/contrib/chat/browser/agentHostInputCompletions.ts	Refactors sessions new-chat agent-host completion provider to use the shared base + gating.

Copilot's findings

Comments suppressed due to low confidence (1)

src/vs/workbench/contrib/chat/browser/widget/input/editor/chatInputCompletionUtils.ts:79

• isAtTriggerCharacterToken misidentifies the token when the cursor is after trailing whitespace. Example: line "@foo " with the cursor after the space yields token="@foo " and returns true, so the host provider can still be consulted even though the user is no longer editing the trigger-led token. Consider extracting the token as the last run of non-whitespace before the cursor (or explicitly return false when the character immediately before the cursor is whitespace).

	const line = model.getLineContent(position.lineNumber);
	const beforeCursor = line.slice(0, position.column - 1);
	// The current token is everything from the last whitespace char (or
	// start-of-line) up to the cursor.
	const wsIdx = beforeCursor.search(/\s\S*$/);
	const token = wsIdx >= 0 ? beforeCursor.slice(wsIdx + 1) : beforeCursor;
	if (token.length === 0) {
		return false;
	}
	return triggerCharacters.includes(token[0]);

• Files reviewed: 4/4 changed files
• Comments generated: 2