fix(space): attach space-agent-tools to post-approval merger session [#852]

[lsm]

Post-approval spawns (e.g. the built-in merger) built init.mcpServers with only node-agent + agent-memory. A post-approval session carries no NodeExecution row and its id has no :exec: segment, so resolveSpaceMcpSessionPolicy classifies it as ad_hoc_member — and QueryRunner.ensureMemberSpaceMcpInvariant therefore requires space-agent-tools on it. The missing server made the merger's first turn throw the MCP-invariant error, which approvePendingCompletion misrecorded as "Interrupted by user". #850/#2374 fixed a separate reuse-path issue and never touched this omission, which is why the same failure recurred after merging it.

This attaches space-agent-tools in spawnPostApprovalSubSession via the same builder attachSpaceToolsToMemberSession uses (extracted to SpaceRuntimeService.buildMemberSpaceToolsMcpServer — no hand-rolled server). Because init.mcpServers is merged into the session's runtime MCP map inside createSubSession before startStreamingQuery, the server is present when runQuery runs the invariant at first turn — race-free. It also wires onMissingMemberSpaceMcpServers → reattachMemberSpaceTools so a future regression self-heals instead of throwing. Normal workflow-worker node-agent sessions are untouched (they remain workflow_worker, require only node-agent, and never hit this spawn path).

Recovery for already-parked tasks (#785, #816, #848, …): their dispatch already failed, so they still need a manual Mark done or a send-back → re-approve to exercise the now-working path.

[lsm]

🤖 Review by glm-5.1 (GLM)

Model: glm-5.1 | Client: HyperNeo | Provider: GLM (Zhipu AI)

Recommendation: APPROVE — zero findings. (Posting as COMMENT because the workflow bot and this review share a GitHub author; GitHub rejects self-APPROVE.)

Verified the root cause and the fix against the actual code paths (not just the diff), with three independent traces plus empirical test runs.

Root cause confirmed. A post-approval session id (space:S:task:T:post-approval:agent) has no NodeExecution row and no :exec: segment, so resolveSpaceMcpSessionPolicy classifies it ad_hoc_member → requiredServers = ['space-agent-tools'], which ensureMemberSpaceMcpInvariant enforces. The old spawnPostApprovalSubSession attached only node-agent + agent-memory, so the first turn threw — misrecorded as "Interrupted by user" by approvePendingCompletion. Distinct from the reuse-path bug #850/#2374 fixed.

Fix is correct and race-free. space-agent-tools is added to init.mcpServers via the shared buildMemberSpaceToolsMcpServer (extracted from attachSpaceToolsToMemberSession; extraction is faithful — only mySessionId: session.id → sessionId). Traced the ordering: createSubSession merges init.mcpServers into the runtime MCP map synchronously at task-agent-manager.ts:1460, before startStreamingQuery at :1523; the invariant at query-runner.ts:1552 reads that same live map. Server is present at first turn.

Self-heal genuinely wired. onMissingMemberSpaceMcpServers → reattachMemberSpaceTools. The callback exists on AgentSession, is invoked+awaited by the invariant before throwing, and reattachMemberSpaceTools re-resolves policy (still ad_hoc_member for post-approval) and re-attaches. The long-horizon guard in attachSpaceToolsToMemberSession cannot skip post-approval (different id format).

Normal workers untouched. Worker spawns keep their :exec: id → workflow_worker → require only node-agent; post-approval path is disjoint, requiredWorkflowSubSessionMcpServers unaffected. No double-attach/overwrite risk (mergeRuntimeMcpServers is per-key; nothing else attaches space-agent-tools here).

Empirical:

• New policy test + CREATE-branch regression test pass (15/15).
• Reverted just the source fix and re-ran the regression test → it fails (space-agent-tools undefined in the merged map), confirming it pins the real bug.
• lint + typecheck + knip + session-guards clean; CI fully green (all daemon unit + online shards).

One unrelated note (not a finding against this PR): the local bun run check fails its trailing check:test-quality step on provider-registry.test.ts:748 — pre-existing on origin/dev (last touched by #2264), in a file this PR doesn't touch, and not part of the PR CI gate. Worth a separate cleanup.

At this space's autonomy level the merge still needs human sign-off to finalize.