REVIEW.md

REVIEW — what changed this session, for your eyes

Quick review file. Delete after you've read it.

---

What the numbers look like now

=== CLAUDE (./sigrank) ===
source: ccusage claude
ledger   R 2.65B · C 136.21M · I 2.90M · O 12.32M
SNR       0.810
10x DEV   2.96
velocity  4.25×
leverage  914×
$/1M      $0.715
Υ yield   3,886
class     Closed-Loop Kinetic · holds both axes
rank      #2 of 8

=== CODEX (./sigrank --codex) ===
source: ccusage codex
⚠ estimated via turn-delta (cache_create from daily context growth)
ledger   R 707.30M · C 26.17M · I 58.92M · O 4.01M
SNR       0.064
10x DEV   1.08
velocity  0.07×
leverage  12×
$/1M      $0.561
Υ yield   1
class     Archival Sponge · high reuse, low generation
rank      #5 of 8

---

Problem being solved

ccusage --json combines Claude + Codex + every other agent into one total. When combined, Codex's large inputTokens (58.9M) tanked Υ quadratically. Fix: run them separately — ccusage claude --json and ccusage codex --json.

---

What was wrong with Codex before (and the fix)

Bug 1 — detection miss (already fixed earlier):
is_codex_shape() checked for cached_input_tokens (snake_case) but ccusage codex --json emits cachedInputTokens (camelCase). Codex JSON fell through to parse_ccusage — no split, no cache, raw 58.9M input. Fix: check both cases.

Bug 2 — anchor now uses turn-delta (CODEX.md item 1):
Old: est_fresh = 2 * output (fixed 2:1).
New: with daily granularity present (we have 29 days of data), estimates cache_create from per-day context growth deltas instead.
Fallback: if no daily data, uses Claude's measured I/O ratio (0.236:1) instead of fixed 2:1 — grounded in your real data, not a constant.

---

What the Codex numbers mean

field	value	source
input (I)	58.92M	inputTokens — fresh input directly from Codex JSON
output (O)	4.01M	outputTokens + reasoningOutputTokens
cache_create (C)	26.17M	estimated via turn-delta
cache_read (R)	707.30M	cachedInputTokens — measured directly
cost	real ($0.561/1M)	from costUSD in the JSON

The 707.3M cache reads are real and measured. Codex is reading a LOT of cached context. It just generates very little output relative to input (velocity 0.07×), so Υ is low. That's Codex's architecture, not a bug.

---

Board marker (CODEX.md item 4)

Estimated rows (Codex-anchored) now show a ~ next to the operator name in the leaderboard. Measured rows (real ccusage) show clean.

---

What is NOT right (open questions for you)

1. Codex inputTokens interpretation — the turn-delta treats inputTokens as already-fresh (not combined with reads). If OpenAI actually reports combined fresh+cached in that field, the I value (58.92M) is too high and Υ will be artificially low. Do you know which it is?

2. Υ=1 for Codex — with 58.9M fresh input and only 4M output, Υ is nearly 0. That may be correct (Codex is a heavy reader, light generator) or it may mean inputTokens is the combined figure and needs splitting.

3. CODEX.md items 2 + 3 still open:

   • Item 2: test_metrics.py (pytest for canonical numbers)
   • Item 3: self-cost via OpenAI per-1M pricing table in parse_codex (cost already comes through from costUSD, so this may be done?)

---

Files changed this session

file	what changed
ingest.py	is_codex_shape → camelCase fix; parse_codex → turn-delta + io_ratio param; ingest_meta → io_ratio passthrough
sigrank.py	default changed to ccusage claude --json; --codex path fetches Claude ratio first
app.py	html.escape(name) XSS fix; ~ marker on estimated board rows