[d6veteran]

What it adds

• scripts/cost.py (standard library): one session's cost as " cost summary", a row per lane (main session, each subagent by type, and background: what Claude Code's own cost record counts beyond the transcripts) and a column per kind of token (input, output, cache write, cache read), each cell in tokens and dollars.
• Prices come from Anthropic's pricing page (platform.claude.com/docs/en/about-claude/pricing.md), fetched at most once a day into ~/.cache/ck/pricing.md. A model the page does not list is priced as ?, never $0. Each request is priced on its own (5-minute and 1-hour cache writes at their own rates, Haiku 5.5's higher price over 100K tokens).
• The total is checked against Claude Code's own figure; a line says when its record is older than the last turns, or when the page and Claude Code disagree on a price (today: Sonnet 5.5 cache reads, $0.10 on the page, $0.20 in Claude Code).
• /ck:cost [history | <session id>], typed only.
• Hooks, no model involved: a Stop hook prints the cost so far after each answer once the session has a pull request; a PostToolUse hook posts the table to the pull request after gh pr merge and logs a line to ~/.claude/ck-cost.jsonl, which /ck:cost history lists.
• README row, ROADMAP release row and revision entry, version 0.1.19.

Tested

• bash tests/run.sh: 485 passed, 0 failed. Section 22 uses a made-up fixture session and a stub gh.
• Run on 578 real sessions with no failure; on seven with a Claude Code record, totals within a few cents of Claude Code's figure or the gap explained (d20mob "[P0] d20Mob coding": $356.01 against $356.68).

🤖 Generated with Claude Code