feat: AgentMemoryService session cache + AgentMemoryCachePrimer

[sneefyyy]

Summary

This PR is the first in a stack of three that implement user-scoped agent memories. It focuses on the caching layer and service infrastructure with no new user-facing behavior is introduced here.

Changes:

1. AgentMemoryService — session-keyed cache + stampede guard

   • getMemoryPrompt() now checks a per-session cache (Map<sessionId, MemoryPromptResponse>) before making a network request.
   • A second _inflightPrompts: Map<string, Promise<MemoryPromptResponse | undefined>> map deduplicates concurrent callers that arrive before the first response lands. Without this, multiple tools resolving in parallel on turn 0 would each fire a separate fetch.
   • Both maps are cleared when the chat session disposes (onDidDisposeChatSession).

2. AgentMemoryCachePrimer — renamed from AgentMemoryToolRegistrar

   • The old name implied it registered VS Code tools. It doesn't — it warms the memory cache before the first tool call. Renamed to reflect actual behavior:
     • File: agentMemoryToolRegistrar.ts → agentMemoryCachePrimer.ts
     • Interface: IAgentMemoryToolRegistrar → IAgentMemoryCachePrimer
     • Class: AgentMemoryToolRegistrar → AgentMemoryCachePrimer
     • Method: registerMemoryTools() → primeCache()
   • Registered in services.ts alongside AgentMemoryService.

3. Test coverage (agentMemoryService.spec.ts)

   • Cache hit: second call with same session ID returns cached value without a second fetch.
   • Stampede guard: two concurrent calls resolve to the same promise and produce one fetch.
   • Cache eviction: cache is cleared when the session disposes.

What's NOT in this PR

• The store_memory tool and turn-0 fix are in PR2.
• The MemoryContextPrompt migration (CAPI coexistence) is in PR3.

Test plan

• Confirm AgentMemoryService unit tests pass (npm run test:unit)
• Confirm no compilation errors in watch output
• Verify AgentMemoryCachePrimer import resolves correctly in agentIntent.ts (part of PR2/PR3 stack)

[Copilot]

Pull request overview

This PR adds conversation/session-scoped caching infrastructure for Copilot agent memory prompts, plus a small service to proactively warm that cache, as the first step toward user-scoped agent memories.

Changes:

• Add per-session memory prompt caching with in-flight request deduplication in AgentMemoryService, and clear cache entries when a chat session disposes.
• Introduce AgentMemoryCachePrimer service (new file) and register it in the extension’s service container.
• Expand unit tests for the new caching/clearing behavior and for storing repo/user memories via the new memory API wrapper.

Show a summary per file

File	Description
extensions/copilot/src/extension/tools/node/agentMemoryCachePrimer.ts	New primer service that fetches the memory prompt early to warm the cache.
extensions/copilot/src/extension/tools/common/agentMemoryService.ts	Implements session-keyed cache + in-flight dedupe and switches to @github/copilot-agentic-tools/memory for API calls.
extensions/copilot/src/extension/tools/common/test/agentMemoryService.spec.ts	Reworked tests to cover caching, cache clearing, and store operations against mocked memory API functions.
extensions/copilot/src/extension/extension/vscode-node/services.ts	Registers the new IAgentMemoryCachePrimer service.

Copilot's findings

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

[Copilot]

The log message still uses the old "AgentMemoryToolRegistrar" prefix and reports that the cache was primed even when sessionId is undefined (no caching happens in getMemoryPrompt) or when the fetch returns undefined. Update the prefix and only log a "primed" message when a response was successfully cached for a concrete session (otherwise log a more accurate message or lower the level).

Suggested change

	this.logService.info(`[AgentMemoryToolRegistrar] primed memory prompt cache for session ${sessionId || 'default'}, definitionVersion=${response?.storeToolDefinition?.definitionVersion ?? 'none'}`);
	if (!sessionId) {
	this.logService.debug('[AgentMemoryCachePrimer] fetched memory prompt without caching because no sessionId was provided');
	return;
	}
	if (!response) {
	this.logService.debug(`[AgentMemoryCachePrimer] did not prime memory prompt cache for session ${sessionId} because no prompt response was returned`);
	return;
	}
	this.logService.info(`[AgentMemoryCachePrimer] primed memory prompt cache for session ${sessionId}, definitionVersion=${response.storeToolDefinition?.definitionVersion ?? 'none'}`);

[Copilot]

clearCache(sessionId) is invoked on session disposal, but an in-flight _fetchMemoryPrompt() for that same sessionId can still resolve later and repopulate _conversationMemoryCache, defeating eviction and potentially leaking entries past session lifetime. Consider (1) clearing _inflightPrompts for the session and (2) adding a guard (e.g., per-session generation token / disposed-session set) so _fetchMemoryPrompt only caches if the session is still active / generation matches.

[Copilot]

There is trailing whitespace on the blank line after super();. This commonly fails lint/format checks; please remove the whitespace so the blank line is empty.

Suggested change

[Copilot]

The PR description mentions a unit test covering the stampede guard (two concurrent getMemoryPrompt calls should share one fetch), but this spec file doesn’t currently include a concurrency test for _inflightPrompts. Adding that test would ensure the deduplication behavior stays correct.