feat: total session tokens including sub-agents

[elhoim]

Summary

Adds subagent-inclusive session token accounting. The existing token widgets
count only the main agent — tokens spent by Task-tool sub-agents are invisible.
This wires sub-agent usage into token counting, reusing the same sub-agent
transcript discovery that already backs the speed-metrics widgets.

Two surfaces (opt-in, zero change for existing setups):

• (s)ubagents toggle on the existing Tokens Input / Output / Cached / Total
  widgets. When on, that widget's number includes sub-agents, shows a Σ marker,
  and a [+sub] modifier in the editor.
• New Session Total Tokens widget (tokens-session-total) — session grand
  total including sub-agents, with an optional (b)reakdown of input/output/cache:
  Σ Total: 152k (in 90k/out 40k/cache 22k).

How it works

• getTokenMetrics(path, { includeSubagents }) sums referenced
  subagents/agent-*.jsonl files via the existing
  getReferencedSubagentIds → getSubagentTranscriptPaths discovery.
• Default behavior is byte-identical — the option defaults off; existing
  widgets and the stdin context_window path are unchanged.
• Double-count guard: when separate sub-agent files exist, inline
  isSidechain rows are dropped from the main pass (they're represented by the
  files); when no files exist (older transcript format), inline sidechain rows are
  still counted. contextLength remains a main-chain-only concept.
• The subagent-inclusive figure is computed once per render, on demand — only
  when a session-total widget or a subagents-enabled token widget is configured
  (RenderContext.sessionTokenMetrics). No extra file I/O otherwise.

Testing

• New unit tests: data layer (incl. double-count guard, multi/unreferenced
  sub-agents, back-compat lock), the metadata helper, the four widget toggles, and
  the new widget. 49 tests pass including the pre-existing TokensWidgets
  suite (no regression).
• tsc --noEmit and eslint clean across the project; bun run build succeeds;
  example-payload smoke render verified.

🤖 Generated with Claude Code

[astral303]

Merge ed087b6 is bad, compilation errors starting in src/utils/jsonl-metrics.ts:256

[elhoim]

Fixed it in 0f89739.

[elhoim]

Rebuilt against current main — conflicts resolved and CI green.

The conflicts weren't textual: main replaced the array-based getTokenMetrics() this branch was written on with a single streaming pass (scanTranscript), and the token widgets picked up number formats and hideable zero states. So the feature is re-ported, not merged around:

• jsonl-metrics.ts — new includeSubagentTokens option puts sessionTokenMetrics on the analysis. Each sub-agent transcript is read once and feeds the token and speed collectors together, so the "computed once per render" promise still holds. Inline isSidechain usage now accumulates into its own tally and is subtracted from the session totals when separate sub-agent files exist — the old skipMainSidechain guard, in collector form. contextLength stays main-chain only.
• ccstatusline.ts — the subagent-inclusive pass is requested only when a widget actually asks for it, so nothing extra is read for existing configs.
• Token widgets — the (s)ubagents toggle, Σ marker and [+sub] modifier now sit on top of resolveNumberFormat and the zero-hideable state. With subagents on, the main-only context_window fallback is skipped.
• Housekeeping — token-subagents.ts and SessionTotalTokens.ts use the shared metadata helpers main introduced, and Session Total Tokens gained number-format support plus the zero-hideable state for parity with the rest of the Tokens category.

Default behavior is unchanged: without the opt-in, no sub-agent file is opened and every existing token number is identical.

Verification, beyond CI:

• The three subagent token tests were re-expressed against the new API (main metrics untouched, inline-sidechain guard, referenced-files-only) and pass alongside the existing widget tests.
• On a real session with 23 sub-agent transcripts, main metrics are unchanged (in=334 out=318378 cached=27552149) and the session totals (in=810 out=519983 cached=47512932) match an independent tally of that transcript plus its referenced agent files, to the token.
• Locally: bun run lint clean, bun test 2376 pass / 2 fail — both pre-existing flaky TUI/subprocess cases in my environment (fetchUsageData subprocess timeout, UsageTimezoneEditor), neither in metrics; CI is green on all three jobs.

One note for scheduling: #591 fixes a double-count in the same collector (one API response is logged once per content block, so every transcript-derived token number currently reads ~2x). The two are independent, but whichever lands second needs a trivial re-merge of jsonl-metrics.ts, and these session totals inherit that fix once it's in. Happy to rebase this on top of it in either order.

🤖 Generated with Claude Code