fix: persist agent name on sub-session nodes across a delivery-order race (v6.7.4)

[Diego Colombo (colombod)]

What this fixes

A spawned sub-session's agent name was silently lost whenever the child's event landed before the parent's — leaving :Session.agent empty and undercounting every WHERE s.agent = ... query on the shared server. This restores correct per-agent attribution. Fixes microsoft-amplifier/amplifier-support#484. Bumps server version 6.7.3 → 6.7.4.

Root cause (verified against shipped source)

The agent name for a spawned sub-session arrives only on the parent's delegate:agent_spawned event (top-level agent); the child's own session:start carries none. Both flow through ensure_session_node (pipeline step 2). Its Tier-2 "node already exists" branch upserted only {labels, status, session_id}, so when the child's session:start created the node first, the parent's later agent value was dropped. (upsert_node uses key-present merge SET n += props, so it was never the culprit — the branch simply never included agent.)

The fix (surgical — one branch)

A populate-if-missing rule for agent, alongside the working_dir one #79 introduced:

if data.get("agent") and not existing.get("agent"):
    stub_data["agent"] = data["agent"]

Writes only when the event supplies agent and the node still lacks it, so an already-set value is never clobbered. No SessionHandler/bundle changes (out of scope).

Verification — RED before, GREEN after (real, no mocks at the gate)

• Unit (tests/test_services.py::TestEnsureSessionNodeAgentField) — two writers sharing one GraphState reproduce the child-first ordering: RED without the fix, GREEN with it. Full file: 60 passed.
• E2E, isolated Neo4j (tests/neo4j/test_agent_field_ordering_race.py) — a throwaway neo4j:5.26.22-community container drives the real SessionHandler + DelegationHandler across two stores sharing one Neo4j, child session:start before parent delegate:agent_spawned, and asserts :Session.agent persisted. RED without the fix (agent=None), GREEN with it (~24s).

SERVER_VERSION verified reporting 6.7.4 from package metadata.

Stack

#78 harden-durable-ingestion
  └─ #81 storage-api-on-78          (backend-neutral storage)
       └─ #79 stack/pr70-on-pr78    (operability & data-integrity, v6.7.3)   ← this PR's base
            └─ this PR              (agent-attribution fix, v6.7.4)

Base is #79's head, not main — merge after #79. The fix slots directly into the stub_data populate-if-missing block #79 introduced, so there is no conflict with #79 (an independent merge onto main would conflict on that same hunk).

🤖 Generated with Amplifier

Co-Authored-By: Amplifier 240397093+microsoft-amplifier@users.noreply.github.com

[Diego Colombo (colombod)]

The agent-field session-node fix folds into PR2 (session-node data quality) of the #79 split; this PR will be closed when PR2 lands. See #85 (PR1: ingest correctness, now open) for the first layer of the split.