Cost attribution

When a PR changes cost, the question isn't just "did it?", it's "why, and by how much per user-facing session?" Cost attribution decomposes the per-session cost delta into three independent sources.

The decomposition

total_delta = model_swap + token_movement + mix_residual

• model_swap: how much of the delta is the candidate's price-per-token vs the baseline's, holding token counts constant at candidate levels
• token_movement: how much is the token-count change, holding price constant at baseline
• mix_residual: non-additive interaction (simultaneous model swap + token change)

When |residual| > 10% of |total_delta| the decomposition is flagged as less trustworthy, the two-factor story is incomplete.

Output

cost attribution (per session):
  session #0: $0.0870 → $0.0174 (Δ $-0.0696, -80.0%)
    model swap claude-opus-4-7→claude-sonnet-4-6: $-0.0696 (+100%)
    token movement:                                $+0.0000 (-0%)
  total: $0.0870 → $0.0174 (Δ $-0.0696)

In the markdown / PR comment

| session | baseline | candidate | Δ | model swap | token move | mix |
|--------:|---------:|----------:|--:|-----------:|-----------:|----:|
| #0      | `$0.0870` | `$0.0174` | `$-0.0696` | `$-0.0696` (+100%) | `$+0.0000` (-0%) | `$+0.0000` (-) |

When it surfaces

Attribution always runs on shadow diff, but only prints when the total cost delta is non-zero. Trace pairs with no pricing data (or --pricing unset) won't produce attribution output.

Pricing table

shadow diff baseline.agentlog candidate.agentlog --pricing pricing.json

With pricing.json:

{
  "claude-opus-4-7":   {"input": 1.5e-5, "output": 7.5e-5},
  "claude-sonnet-4-6": {"input": 3e-6,   "output": 1.5e-5, "cached_input": 3e-7}
}

Rich-dict pricing supports input / output / cached_input / cached_write_5m / cached_write_1h / reasoning / batch_discount. Legacy tuple [input, output] also accepted.

Unknown models contribute $0 (no crash).