SubagentStop agent_type: null — hookSSEBridge registry lookup fails for nested agents

[Number531]

Problem

ALL SubagentStop hook events arrive with agent_type: null, phase: null, stage: null — both locally and on GCE. This occurs because:

1. The SDK does not reliably populate agent_type in SubagentStop input (despite Stage-1 deep-plan: repoint paths to worktree + correct START-HERE launch #226 claiming this was fixed in v0.2.69+)
2. The agent_id on SubagentStop events often does not match the agent_id from the corresponding SubagentStart — making registry lookup in hookSSEBridge fail
3. Nested subagent stops (from Agent tool calls inside subagents) bubble up to the parent's hooks, but their corresponding SubagentStart never fires (Fix billable-canary cache-source false-fail (srcOK accepts 'cache') #225), so they were never registered

Impact

• Agent observability timeline shows "unknown" agent type for all stop events
• Phase/stage/wave classification unavailable on stop
• Gate check status, report verification, and tool usage metrics are captured but cannot be attributed to a specific agent type
• The agentTypeMap correlation (for agent_progress) partially compensates but only covers agents that had a SubagentStart

Current behavior (hookSSEBridge.js)

SubagentStart fires → agentRegistry.set(agent_id, { agent_type, classification })
SubagentStop fires  → agentRegistry.get(agent_id) → returns undefined (different ID)
                    → falls back to null for all classification fields

Data from production

• Local: 9 SubagentStart, 43 SubagentStop — 34 orphaned stops with no matching start
• GCE: 6 SubagentStart, 28 SubagentStop — 22 orphaned stops with no matching start
• 100% of SubagentStop events have agent_type: null in both environments

Potential mitigation

Use temporal correlation: when a SubagentStop arrives with unknown agent_id, match it to the most recently started agent that hasn't stopped yet. This is imperfect with concurrent agents but better than null.

Alternatively, track the tool_use_id from SubagentStart and correlate with the Agent tool_result that arrives around the same time as SubagentStop.

Related upstream issues

• anthropics/claude-agent-sdk-typescript#226 — agent_type in SubagentStop (CLOSED, supposedly fixed)
• anthropics/claude-agent-sdk-typescript#225 — No hook propagation to subagents (OPEN)
• SubagentStart/SubagentStop hooks not firing on GCE — root cause analysis + debug test #53 — Parent tracking issue for hook firing on GCE

[Number531]

Comprehensive reference doc: https://github.com/Number531/Legal-API/blob/main/super-legal-mcp-refactored/docs/docker-hooks-issue.md