[bug]Preserve Claude subagent token totals during condensation

[stale2000]

Entire Logs: https://entire.io/gh/stale2000/cli/session/019eaa64-5270-74c0-9f8a-e4df88c7014e
SUMMARY:
After finding and making the fix, I noticed the log saying that the "bug" might be intentional? Or it is unknown why it is written this way?

I did an analysis as for why it might have been written this way and got the following response that implies that its indeed a bug, and the previous implementation didn't know how to solve it. But this PR does fix the issue.

Explanation/response to the TODO statement:
"
So the nuanced answer is:

For live transcript condensation: bug.
For shadow-only fallback with no live path: empty dir may be unavoidable.
The TODO was probably someone noticing the bug-prone gap during cleanup but not resolving the live-vs-shadow distinction.
Our fix handles exactly that: derive the dir when a live transcript path exists, otherwise helper returns "".
"

Summary

This fixes a metadata loss bug in checkpoint condensation for Claude Code sessions that spawn Task subagents.

Before this change, live session handling could see subagent token usage, but the later condensation step recalculated token usage without passing the session's subagent transcript directory. As a result, SubagentTokens were silently dropped from the committed checkpoint metadata written to entire/checkpoints/v1.

Issue

Claude Code stores Task subagent transcripts separately from the main transcript:

<transcriptDir>/<sessionID>/subagents/agent-<agentID>.jsonl

The lifecycle path already knows that convention and passes the subagent directory into token/file extraction. Condensation did not. Both condensation paths called:

agent.CalculateTokenUsage(ctx, ag, data.Transcript, checkpointTranscriptStart, "")

For subagent-aware agents, the empty directory intentionally disables subagent transcript reads. That means the main transcript token totals survived, but spawned subagent token totals disappeared when checkpoint metadata was committed.

How to reproduce

One reproducible shape is:

1. Enable Entire in a repo and start a Claude Code session.

2. Have Claude invoke the Task tool so the main transcript contains a Task tool_use and a corresponding tool_result with an agentId, for example agentId: sub1.

3. Ensure the subagent transcript exists at:

   <main-transcript-dir>/<sessionID>/subagents/agent-sub1.jsonl
   

   and contains assistant usage metadata.

4. Let the normal stop hook run. At this point live session state can include subagent token usage because the lifecycle path passes the subagent directory.

5. Make a user commit so the post-commit hook condenses the session into committed checkpoint metadata.

6. Inspect the committed checkpoint metadata from entire/checkpoints/v1.

Before this fix, the condensed checkpoint metadata had main token usage but no subagent_tokens field, even though the subagent transcript was present and readable.

The added regression test builds that shape directly: a main Claude transcript with a Task result referencing sub1, a matching agent-sub1.jsonl subagent transcript, and a condensation call that verifies SubagentTokens survive in committed metadata.

Impact

This affects checkpoint metadata for Claude Code sessions that use Task subagents.

Observed effects:

• entire/checkpoints/v1 underreports total work done by spawned subagents.
• Later checkpoint consumers such as activity, explain/status views, or metadata sync consumers cannot recover those subagent token totals from the committed metadata.
• The bug is silent: checkpoint writes succeed, but the committed metadata is incomplete.

The code path is shared through the SubagentAwareExtractor token interface, so passing the directory also keeps condensation aligned with other subagent-aware implementations that use the same transcript layout convention.

Fix

• Add subagentsDirForTranscript(transcriptPath, sessionID) in the strategy package.
• Pass the derived subagent directory into agent.CalculateTokenUsage from both condensation paths:
  • shadow-branch extraction with live-transcript preference
  • direct live-transcript extraction
• Add TestCondenseSession_ClaudeSubagentTokenUsage to lock the end-to-end condensation behavior.

Verification

go test ./cmd/entire/cli/strategy -run 'TestCondenseSession' -count=1
go test ./cmd/entire/cli/agent/claudecode -run 'TestCalculateTotalTokenUsage|TestExtractAllModifiedFiles' -count=1
mise run lint

[stale2000]

Answer:

"So the nuanced answer is:

For live transcript condensation: bug.

For shadow-only fallback with no live path: empty dir may be unavoidable.

The TODO was probably someone noticing the bug-prone gap during cleanup but not resolving the live-vs-shadow distinction.
Our fix handles exactly that: derive the dir when a live transcript path exists, otherwise helper returns "".

"