Cost Diff

PR-level cost regression detection. Where cost-counterfactual compares to HYPOTHETICAL baselines (always-haiku/sonnet/opus) and cost-burn compares latest bucket to PRIOR MEAN, cost-diff compares two SPECIFIC known-good snapshots.

Question	Skill
"What would we have spent at always-X?"	cost-counterfactual
"Is daily burn accelerating vs prior mean?"	cost-burn
"Did THIS PR add spend vs main?"	cost-diff ← this

Algorithm

Implementation: scripts/diff.mjs. Consumes the stable JSON contract from cost summary --format json.

1. Load --baseline and --current JSON snapshots.
2. Sanity check: both must have total_cost_usd + sessionCount (cost-summary shape).
3. Per-key delta: byTier (haiku/sonnet/opus) and byModel (each model).
4. Each entry tagged added / removed / changed based on baseline / current zero-ness.
5. Sort table by |delta| descending so the biggest movers are at the top.
6. --alert-on-pct N: exit 1 when total_pct > N.
7. --alert-on-usd N: exit 1 when total_delta_usd > N. Both can be set; first to trigger wins.

PR-gate workflow

# Capture baseline (e.g. on main, via the cost-tracker-smoke CI workflow)
cost summary --format json > baseline.json

# On the PR branch, capture current state
cost summary --format json > current.json

# Compare; fail the PR if total spend grew >10% OR >$5
cost diff --baseline baseline.json --current current.json \
          --alert-on-pct 10 --alert-on-usd 5.00

The combination of both flags catches:

• Percent-only fires: a small absolute change but a meaningful shift (e.g. doubling from $0.10 to $0.20 hits +100% but only +$0.10).
• USD-only fires: a large absolute change with a small percent (e.g. growing from $100 to $110 is only +10% but +$10).

Either signal can fail the PR independently — they're OR'd.

--alert-on-class-pct (iter 86)

The two USD-level thresholds above miss a regression class: when ONE token type grows disproportionately even though total spend grows modestly. Example: a PR introduces a verbose context-cache pattern, total spend grows only 10% (under --alert-on-pct 50), but cache_write tokens grow 900%. The iter-82 driver hides inside the USD signal.

--alert-on-class-pct cache_write:50 exits 1 when cache_write tokens grow more than 50% baseline → current. Multiple classes can be checked in one flag (comma-separated):

cost diff --baseline baseline.json --current current.json \
          --alert-on-class-pct cache_write:50,output:25

First class to breach wins. Valid classes: input | output | cache_write | cache_read.

Recommended PR-gate triad:

cost diff --baseline ... --current ... \
          --alert-on-pct 25 \
          --alert-on-usd 5.00 \
          --alert-on-class-pct cache_write:100

Three orthogonal signals — pct (total grew), usd (large absolute jump), class-pct (composition shifted). Each catches what the others miss; AND-of-OR semantics means any one firing fails the PR.

Smoke transcript (synthetic baseline + current)

| Total spend       | $1.000000 | $1.500000 | +$0.500000 | 50.00% |
| Sessions          | 10        | 13        | +3         | 30.00% |

## By tier
| opus   | $0      | $0.60   | +$0.600000 | new      | added   |
| sonnet | $0.70   | $0.50   | -$0.200000 | -28.57%  | changed |
| haiku  | $0.30   | $0.40   | +$0.100000 | 33.33%   | changed |

Notice the table is sorted by absolute delta, not alphabetically — the biggest mover (opus newly added) bubbles to the top. Operators reading top-down see "what mattered" first.

Exit codes

Exit	Meaning
0	No alert, OR no thresholds set
1	--alert-on-pct or --alert-on-usd threshold exceeded
2	Config error (missing files, invalid JSON, malformed snapshot)

Status column

Status	Meaning
added	This tier/model was $0 in baseline, >$0 in current
removed	This tier/model was >$0 in baseline, $0 in current
changed	Both baseline and current >$0; delta is the difference

Entries with baseline === 0 && current === 0 are dropped (nothing to report).

Composition with cost-summary

cost-diff is the SECOND HALF of a contract that cost-summary started: the stable JSON shape from cost summary --format json. Both pieces have been frozen — adding fields to summary is fine; renaming or removing isn't.

If you're consuming snapshots elsewhere (dashboards, alerting), the same shape works — cost-diff is just one consumer.