feat(models): agent x model cross-breakdown (--by-agent)

[iamtoruk]

Problem

The dashboard shows Skills and Agents and By Model as separate panels, so orchestrator setups with agents pinned to different models cannot see which agent drives which model's spend. Closes #708.

Fix

codeburn models --by-agent: one row per (provider, model, agent), mirroring --by-task exactly across the flag, aggregation, and all four output formats (table, markdown, csv, json). The agent dimension comes from each subagent session's own agentType joined with its calls, which is the authoritative pairing; spawn markers on main sessions are deliberately not used. Ordinary main sessions and non-Claude providers bucket under main. --by-task and --by-agent are mutually exclusive with a clear error. The renderers gained a generic third-column concept rather than a copy of the task path.

Real output from a machine with a heavy orchestrator day (the exact question from the issue, answered):

Claude  Opus 4.8   main             691.7M  446.51
                   general-purpose   61.3M   50.20
Claude  Fable 5    main             203.2M  248.13
                   general-purpose   24.7M   55.21
Claude  Haiku 4.5  general-purpose   68.8K    0.03

The issue's secondary ask, exposing subagentType on per-record exports, is left for a follow-up to keep this diff on one surface.

Verification

• 8 new tests: the model x agent cross (one agent on two models, two agents on one model, main, non-Claude), group ordering, all four renderers, and the mutual-exclusion error. Mutation-checked twice: forcing every bucket to main fails exactly the two aggregation tests; mislabeling the third column fails exactly the two renderer pins.
• tsc clean; full suite green apart from the pre-existing app/electron/cli.test.ts fails on a fresh clone: dev-resolution tests require a built dist/cli.js #681 electron failures and the playwright-dependent sync e2e on this checkout.
• aggregateModels callers without options (usage-aggregator) are behaviorally unchanged.

[ozymandiashh]

Reviewed against a parallel implementation of the same issue that went through a three-model adversarial panel today, so these findings are battle-tested rather than speculative. The core design matches what we converged on independently: session-level attribution from the subagent transcript's own agentType, spawn markers deliberately unused (crossing per-call subagentTypes would bill the parent's Task-invocation tokens to the child agent). The generic third-column renderer is a nicer shape than mirroring the task path. Three findings:

1. The bare main sentinel collides with a real agent named "main". agentType is free-form text from the meta file, not an enum: a user-defined subagent literally named main merges silently into the main-session bucket and the attribution the feature exists for is corrupted for exactly that agent. A non-slug sentinel (we used (main) with parentheses) cannot collide; worth a test pinning that a subagent named main stays distinct.

2. The inherited --min-cost 0.01 default quietly hides sub-cent agent buckets. At task granularity that default is invisible; at agent granularity sub-cent buckets are the norm (scouts, one-shot helpers), so agents disappear from the audit without a trace. One README sentence recommending --min-cost 0 for a complete agent inventory saves the support question.

3. For the deferred export follow-up: when subagentType/model land on sessions.csv, append them AFTER all legacy columns — inserting mid-row shifts positions for by-index CSV consumers. We have the export half already built and panel-reviewed (records.csv with one row per call, sessions.csv with appended fields, formula-injection-safe); happy to open it as the follow-up PR so it composes with this one.