Compute Token Stats (Rust Ultra)

[Created by Codex: 019beb87-b904-7a81-a7c9-5b5d5575d2f7 2026-01-23] [Edited by Claude: 9d1228e9-aabe-4ab3-9b9c-c0c1d1f64ebc 2026-01-27]

⚠️ CRITICAL: Agent Instructions

If the user asked you to load this skill, you MUST use it for token cost computation.

1. ALWAYS use the Rust binaries - they are ultra-fast and purpose-built for this task
2. NEVER write custom Python to parse tokens - the Rust tool already does this
3. If you think you need Python, STOP and ask the user first - 99% of the time the answer is NO
4. These binaries are PIPE-ONLY - they read from stdin, not files directly

⚠️ NEVER RUN STANDALONE - WILL HANG!

# WRONG - will hang waiting for stdin forever!
/tmp/.codex/skills/compute-token-stats-rust-ultra/rust/target/release/token_costs

# CORRECT - always pipe data to it
cat file.jsonl | /tmp/.codex/skills/compute-token-stats-rust-ultra/rust/target/release/token_costs

Filtering by SID

To compute tokens for specific sessions, filter first then pipe:

# Filter by single SID
grep '"sid":"abc123"' ~/centralized-logs/claude/sse_lines.jsonl | \
  /tmp/.codex/skills/compute-token-stats-rust-ultra/rust/target/release/claude_token_costs

# Filter by multiple SIDs (use grep -F with file)
grep -F -f /tmp/sids.txt ~/centralized-logs/codex/sse_lines.jsonl | \
  /tmp/.codex/skills/compute-token-stats-rust-ultra/rust/target/release/token_costs

Compute token totals + estimated costs from massive JSONL SSE logs:

• Codex: ~/centralized-logs/codex/sse_lines.jsonl
• Claude: ~/centralized-logs/claude/sse_lines.jsonl

This skill includes:

• Rust Ultra (streaming, byte-scan, no JSON DOM): rust/ → tokencosts + claudetoken_costs
• Python anchor (ground truth): python/runcomputetokencostsanchor.py → runs ~/swe/token-computation/computetokencosts.py
• Python fast (byte-scan): python/compute{codex,claude}tokencostsfast_pipe.py

Build (Rust Ultra)

cd /tmp/.claude/skills/compute-token-stats-rust-ultra/rust && \
  cargo build --release --bins

Run (Rust Ultra)

Codex:

cat ~/centralized-logs/codex/sse_lines.jsonl | \
  /tmp/.claude/skills/compute-token-stats-rust-ultra/rust/target/release/token_costs

Claude:

cat ~/centralized-logs/claude/sse_lines.jsonl | \
  /tmp/.claude/skills/compute-token-stats-rust-ultra/rust/target/release/claude_token_costs

Run (Python anchor / ground truth)

Generate a pricing snapshot (recommended; makes runs reproducible):

python /Users/sotola/swe/token-computation/pull_pricing.py \
  --providers openai,anthropic --timeout 20 --snapshot-dir ~/centralized-logs/pull-pricing/snapshots

Then run:

python /tmp/.claude/skills/compute-token-stats-rust-ultra/python/run_compute_token_costs_anchor.py \
  --codex ~/centralized-logs/codex/sse_lines.jsonl \
  --claude ~/centralized-logs/claude/sse_lines.jsonl \
  --pricing-snapshot ~/centralized-logs/pull-pricing/snapshots/<pick-one>.json

Quick benchmark + cost diff (%)

This compares Rust Ultra (streaming stdin) vs Python anchor on the same datasets and prints:

• seconds
• lines/sec (Python uses file line count / wall time)
• cost diff (%)

python /tmp/.claude/skills/compute-token-stats-rust-ultra/python/bench_rust_vs_anchor.py

Notes on “unknown” Codex sessions

When turn.session_configured is missing for some sid (log truncation / invalid state machine), per-sid model attribution is unknowable.

• The anchor (computetokencosts.py) prices Codex totals with a single detected codex.model.
• Rust Ultra keeps per-model attribution when possible; for (unknown) it prices using the detected pricing model to stay consistent with the anchor.