fix: discover subagent files by directory scan instead of JSONL references

[neoeny152]

Summary

• Bug: getSubagentTranscriptPaths() relies on getReferencedSubagentIds() finding agentId fields in the main session JSONL. However, the main JSONL contains zero entries with agentId — all agent IDs exist only in the separate subagent files under {sessionDir}/subagents/. This causes getReferencedSubagentIds() to return an empty set, making getSubagentTranscriptPaths() short-circuit and return [], so no subagent files are ever read.
• Impact: Both speed metrics (tok/s) and token counts miss all subagent tokens. In sessions with heavy agent usage (16+ parallel agents via GSD, team agents, etc.), 90%+ of tokens are invisible — e.g. showing ~40 tok/s when actual throughput is ~3,333 tok/s.
• Fix: Scan subagents/ directories for all agent-*.jsonl files directly instead of requiring pre-discovered IDs. Also add subagent token aggregation to getTokenMetrics().

Changes

1. getSubagentTranscriptPaths() — Remove referencedAgentIds parameter and the early-return gate. Scan candidate directories for all agent-*.jsonl files directly. This is safe because the subagents/ directory only contains files for the current session.

2. getTokenMetrics() — After reading the main JSONL, also discover and read all subagent files, summing their token usage (input, output, cached).

3. getSpeedMetricsCollection() — Remove the broken getReferencedSubagentIds(mainLines) call and use the simplified getSubagentTranscriptPaths(transcriptPath) directly.

Verified

Tested against real Claude Code sessions on Windows. Before patch: main JSONL had 0 agent IDs, 0 subagent files discovered. After patch: all subagent files found, ~48% more tokens counted in a 2-agent session. For 16+ agent sessions the difference is ~90%+.

Test plan

• Verify speed widgets show combined main+agent tok/s during sessions with subagents
• Verify token count widgets include subagent tokens
• Verify backward compatibility when no agents are running (subagents/ dir doesn't exist)
• Verify no double-counting (main JSONL has isSidechain: false only, subagent files have the agent data)

🤖 Generated with Claude Code

[sirmalloc]

@neoeny152 I'm not sure this is necessary. After reviewing my session transcript JSONL files, I've found a 1:1 match of unique agentIds in the JSONL files and the subagent transcripts. Can you expand on your assertion that the main JSONL has no agentId fields in it, and provide a recent example with the latest Claude Code version?