Claude Code subagent transcripts are never discovered (depth-2 cap) — ~54% of messages invisible on subagent-heavy workflows #207

Context: surfaced while validating the #204 rewrite fix in #200 — the token-exact cross-check there covers main transcripts only; this is the orthogonal coverage gap mentioned in that thread, filed separately as requested.

Summary

The Claude Code analyzer discovers only ~/.claude/projects/<slug>/<session>.jsonl. Claude Code additionally writes per-subagent transcripts under ~/.claude/projects/<slug>/<sessionId>/subagents/**/agent-<agentId>.jsonl (Task tool: Explore / general-purpose / custom agents). Those files sit at depth ≥ 4, so their usage never enters splitrail's totals, uploads, or the leaderboard.

For anyone using subagents heavily, splitrail undercounts by a large factor. On my corpus (~50 days, mixed CLI + subagent workflows), live main transcripts hold 13,704 messages / 18.5M output tokens while live subagent transcripts hold 16,160 messages / 13.7M output tokens — 54% of live messages missing.

Where the cap lives (v3.6.0)

All three discovery paths hard-code depth 2 in src/analyzers/claude_code.rs:

• discover_data_sources() — WalkDir::new(projects_dir).min_depth(2).max_depth(2)
• is_valid_data_path() — relative.components().count() == 2
• get_data_glob_patterns() — "{home}/.claude/projects/*/*.jsonl"

Observed layout (Claude Code 2.1.150–2.1.198, macOS)

~/.claude/projects/<project-slug>/
├── <sessionId>.jsonl                                  # main transcript (scanned)
└── <sessionId>/subagents/
    ├── agent-<agentId>.jsonl                          # NOT scanned
    └── agent-<agentId>.meta.json                      # agentType, e.g. "Explore"

Subagent lines carry the same schema as main-transcript assistant lines (message.id, model, usage, requestId), repeat the parent sessionId, and add agentId + isSidechain: true.

Repro

# any machine that has used the Task tool / subagents:
find ~/.claude/projects -mindepth 3 -name '*.jsonl' | wc -l    # files splitrail never sees
find ~/.claude/projects -mindepth 3 -name '*.jsonl' -exec grep -l '"type":"assistant"' {} + | head

Compare splitrail stats totals before/after copying one subagent file up to depth 2 (in a scratch $HOME): the totals jump by exactly that file's usage.

Numbers from my corpus (message.id-level dedup, append-only log as reference)

Model mix makes the blind spot obvious: on my machine splitrail sees 22 of 5,059 sonnet messages and 0 of 1,566 haiku messages — those models run almost exclusively inside subagents.

Suggested fix

Lift the depth cap and include */*/subagents/**/*.jsonl (or simply walk without max_depth and filter on .jsonl):

• The existing cross-file dedup already keys on local_hash = fast_hash(requestId_messageId), so any duplication between main and subagent records dedups correctly (I verified token-exact agreement between 3.6.0's main-transcript totals and an independent message.id-keyed ingest — the dedup semantics are sound).
• is_valid_data_path and the watcher need the same relaxation, plus the 3.6.0 history store keys by conversation-path hash and should be unaffected (each subagent file becomes its own conversation).
• Heads-up for Splitrail Cloud: totals for subagent-heavy users will jump on upgrade; may be worth a release-note line.

Happy to test a branch against my corpus (same fixture as in #200's validation) and report before/after.