fix(rehydrate): wire agent MCP servers into restarted/recovered sessions

[ishaan-berri]

What

When a session is restarted (or auto-recovered after its sandbox dies), rehydrateSession brought the harness back up without the agent's external MCP servers — so a restarted session silently lost access to e.g. linear. Only the create path (finishBringUp) wired them (fixed earlier in #290).

Why

rehydrate.ts called harnessCreateSession({ sandbox_url, title, files }) — no mcp_servers. Both the K8s path and the brain-inline path had the gap.

Fix

• Extract resolveAgentMcpServers into src/server/agent-mcp.ts so the create path and the rehydrate path share one implementation (no drift).
• Forward mcp_servers (+ agent_id, platform_session_id) on both rehydrate paths, mirroring finishBringUp.
• Refactor the session route to use the shared helper.

Effect

A restarted/recovered session now keeps its attached MCPs. Concretely: restarting a session on an agent configured with linear will list linear again, instead of dropping back to only the built-in lap-* servers.

Test

tsc --noEmit clean. Companion to #290 (create path) — this closes the same gap on the restart path.

[greptile-apps]

Greptile Summary

This PR fixes a gap where restarted or auto-recovered sessions lost their agent-attached MCP servers (e.g. linear) by wiring mcp_servers into both rehydrate paths. It also extracts the shared resolveAgentMcpServers / agentMcpServerIds helpers into a new src/server/agent-mcp.ts module to eliminate drift between the create and restart paths.

• src/server/agent-mcp.ts (new): shared module with resolveAgentMcpServers and agentMcpServerIds, extracted verbatim from the session route with no logic changes.
• src/server/rehydrate.ts: both the brain-inline and K8s rehydrate paths now call resolveAgentMcpServers and forward mcp_servers, agent_id, and platform_session_id to harnessCreateSession, mirroring finishBringUp.
• route.ts: removes the local duplicate of resolveAgentMcpServers and replaces the inline ID-extraction with the shared agentMcpServerIds helper.

Confidence Score: 4/5

Safe to merge — the fix is correct and the refactor is a clean extraction with no logic changes.

Both rehydrate paths now correctly forward MCP servers. The only gap is that the warning string from resolveAgentMcpServers is silently dropped on rehydrate, so a LiteLLM outage at restart time leaves no log trail — a minor observability blind spot compared to the create path, but not a correctness issue.

src/server/rehydrate.ts — both resolveAgentMcpServers call sites discard the warning return value.

Important Files Changed

Filename	Overview
src/server/agent-mcp.ts	New shared module extracting resolveAgentMcpServers and agentMcpServerIds — clean extraction of the pre-existing implementation with no logic changes
src/server/rehydrate.ts	Both rehydrate paths (brain-inline and K8s) now forward mcp_servers, agent_id, and platform_session_id; the warning from resolveAgentMcpServers is silently discarded on both paths unlike in finishBringUp
src/app/api/v1/managed_agents/agents/[agent_id]/session/route.ts	Removes the local resolveAgentMcpServers duplicate and replaces raw ID extraction with agentMcpServerIds — straightforward call-site cleanup with no behaviour change

_{Reviews (1): Last reviewed commit: "refactor(sessions): use shared resolveAg..." | Re-trigger Greptile}

[greptile-apps]

The warning returned by resolveAgentMcpServers is silently discarded on both rehydrate paths (brain-inline and K8s), unlike finishBringUp which logs it. When LiteLLM is unreachable during a restart or auto-recovery, the session will silently come up with no MCPs and there will be no trace in the logs — making this failure mode invisible in production.

Suggested change

	const { specs: inlineMcpServers } = await resolveAgentMcpServers(
	agentMcpServerIds(agent),
	);
	const { specs: inlineMcpServers, warning: inlineMcpWarning } = await resolveAgentMcpServers(
	agentMcpServerIds(agent),
	);
	if (inlineMcpWarning) console.warn(`rehydrateSession (inline) session_id=${session_id}: ${inlineMcpWarning}`);

[greptile-apps]

Same missing warning log on the K8s/local-dev rehydrate path. If LiteLLM is down at restart time, MCPs drop silently with no log entry.

Suggested change

	const { specs: mcpServers } = await resolveAgentMcpServers(
	agentMcpServerIds(agent),
	);
	const { specs: mcpServers, warning: mcpWarning } = await resolveAgentMcpServers(
	agentMcpServerIds(agent),
	);
	if (mcpWarning) console.warn(`rehydrateSession session_id=${session_id}: ${mcpWarning}`);