agentHost: release idle sessions from memory (SDK session + state)

[roblourens]

Fixes #324856

What & why

Re-enables idle-session eviction in the agent host and fixes what the original implementation got wrong. When an idle session loses its last subscriber, we now release all of its in-memory footprint — both the cached AHP state and the provider's SDK session/connection — while preserving all durable data (SDK session log, session database, worktree) so the session resumes transparently on the next access.

Idle eviction was previously disabled (#323824) because the old _maybeEvictIdleSession only dropped the lightweight AHP state cache via _stateManager.removeSession(...) and never released the heavy provider resources (the cached session entry, the live SDK session, active clients, MCP subscriptions). So it freed the wrong thing and dropped observable state out from under clients that still needed it. This change inverts that: it releases the expensive SDK resources, and the state-cache drop is safe because restore-from-disk is lossless.

How

• New non-destructive IAgent.releaseSession(session) — distinct from the destructive disposeSession. It disconnects/tears down the session's in-memory SDK resources but never deletes anything on disk.
• _maybeEvictIdleSession is re-enabled: after confirming no subscribers remain on the root or any subagent descendant, no restore is in flight, and there is no active turn, it drops the cached state for the root + its subagent entries and calls provider.releaseSession(root).
• Implemented for all three providers:
  • Copilot: extracted the in-memory teardown from _destroyAndDisposeSession into _releaseSessionResources (SDK session.disconnect() + entry/active-client/MCP-subscription disposal, no worktree removal, no deleteSession). The destructive disposeSession keeps calling it then reaps the worktree.
  • Claude: releases the session entry (disposing the live SDK subprocess and its MCP-server children via the pipeline), keeping the on-disk session for resume.
  • Codex: thread/unsubscribe on the shared codex process (keeping the resumable on-disk rollout), reusing the existing disposeSession teardown.
• All three guard against provisional (never-persisted) sessions and in-flight turns; a new hasActiveTurn signal backs the Copilot/Claude defensive re-check.

The destructive difference (deleting durable data) stays in the orchestrator, which only invokes it on dispose — so provider releaseSession and disposeSession share the same in-memory teardown.

Notes for reviewers

• Lossless restore is the crux. The prior disable was caused by dropping state clients still observed; here the state-cache drop is paired with a verified restore-from-disk round-trip (unit + integration tests) and guarded against in-flight restores/active turns.
• MCP servers / subprocesses: MCP servers are SDK-owned (per-session for Copilot/Claude inside the SDK-owned CLI runtime; per-thread on the shared process for Codex). Release routes through the same runtime teardown (session.destroy -> session.dispose() -> mcpHost.stopServers() for Copilot; pipeline dispose -> CLI subprocess teardown for Claude; thread/unsubscribe for Codex) that disposeSession already used, so no new subprocess-leak risk is introduced.
• Retention (keep-warm) is intentionally out of scope for this PR. A grace-timeout and/or last-N LRU to keep recently-used sessions warm can build on the same seams (the existing _pendingSessionGc timer pattern and AgentHostChangesetStateCache LRU) in a follow-up.

Tests

• Re-enabled the four disabled eviction tests in agentService.test.ts and added coverage that releaseSession fires, that dispose is not called on eviction, and that a subsequent subscribe restores turns losslessly.
• Added an end-to-end unsubscribe -> eviction -> re-subscribe round-trip to sessionLifecycle.integrationTest.ts over the real protocol server (mock provider).
• Full agent-host node suite green (2536 passing, incl. Claude 181 / Codex 81); typecheck-client and hygiene clean.

_{(Written by Copilot)}

[Copilot]

Pull request overview

Re-enables idle-session eviction in the agent host and corrects the original implementation. When an idle session loses its last subscriber, the orchestrator now drops the cached AHP state and asks the provider to release its heavy in-memory SDK resources, while preserving all durable data (SDK session log, session database, worktree) so the session resumes transparently on next access. This inverts the prior (buggy) behavior that dropped only the lightweight state cache that clients still observed.

Changes:

• Adds a non-destructive, optional IAgent.releaseSession(session) distinct from destructive disposeSession, implemented for Copilot, Claude, and Codex by sharing each provider's existing in-memory teardown.
• Re-enables _maybeEvictIdleSession with a restore-in-flight guard and a provider releaseSession call in lockstep with the state-cache drop; adds hasActiveTurn defensive re-check signals.
• Re-enables/updates the four eviction unit tests, adds lossless-restore coverage, and adds an end-to-end unsubscribe → eviction → re-subscribe integration test.

Show a summary per file

File	Description
src/vs/platform/agentHost/common/agentService.ts	Adds optional releaseSession? to the IAgent interface with documented non-destructive semantics.
src/vs/platform/agentHost/node/agentService.ts	Re-enables _maybeEvictIdleSession; adds restore-in-flight guard and fire-and-forget provider.releaseSession.
src/vs/platform/agentHost/node/copilot/copilotAgent.ts	Adds releaseSession; extracts _releaseSessionResources from _destroyAndDisposeSession, keeping worktree reaping only on dispose.
src/vs/platform/agentHost/node/copilot/copilotAgentSession.ts	Adds hasActiveTurn getter for the active-turn guard.
src/vs/platform/agentHost/node/claude/claudeAgent.ts	Adds releaseSession reusing _teardownEntry with provisional/active-turn guards.
src/vs/platform/agentHost/node/claude/claudeAgentSession.ts	Adds hasActiveTurn getter delegating to the pipeline.
src/vs/platform/agentHost/node/claude/claudeSdkPipeline.ts	Adds hasActiveTurn getter based on queue emptiness.
src/vs/platform/agentHost/node/codex/codexAgent.ts	Adds releaseSession; extracts _teardownSessionInMemory shared with disposeSession.
src/vs/platform/agentHost/test/node/mockAgent.ts	Adds MockAgent.releaseSession recording calls non-destructively.
src/vs/platform/agentHost/test/node/agentService.test.ts	Re-enables/updates eviction tests and adds lossless re-subscribe restore coverage.
src/vs/platform/agentHost/test/node/protocol/sessionLifecycle.integrationTest.ts	Adds an end-to-end release-and-restore round-trip over the real protocol server.

Review details

• Files reviewed: 11/11 changed files
• Comments generated: 0
• Review effort level: Medium