Savings report

Savings report

Prove (or disprove) that the policy is paying off. Be honest — if the policy saved nothing, say so.

Step 1 — Collect usage

Claude Code writes session transcripts as JSONL under ~/.claude/projects/<encoded-project-path>/. Computing actual dollar figures requires read access to that directory — which is outside the project root, so a sandboxed or restricted session cannot reach it. If you cannot read ~/.claude/projects/, skip to "Degraded mode" below; do not invent token counts.

Try the community tool first (it reads the same transcripts):

npx -y ccusage@latest daily --json    # bare `--json` without a subcommand prints nothing

and filter to this project. If ccusage is unavailable, parse the JSONL directly: sum usage blocks (input_tokens, output_tokens, cache_read_input_tokens, cache_creation_input_tokens) grouped by message.model.

Degraded mode (transcripts unreadable and ccusage absent): you can still produce a useful report from .calibrate/route-log.ndjson (tier decisions) and the policy file (per-phase tiers). Report theoretical per-phase savings vs all-top-tier, clearly labelled as theoretical, and tell the user to either install ccusage or re-run this skill from the main interactive session (which has transcript access) to get actual dollars. Never present theoretical numbers as measured.

Step 2 — Compute the comparison

• Actual: tokens × current per-token price for each model actually used. Look up current Claude pricing rather than assuming remembered prices.
• Baseline: the same token volumes priced as if every request ran on the project's top-tier model (the policy's plan model, or Opus if unset).
• Delta: baseline − actual, in dollars and as a percentage.

Step 3 — Quality check

Savings are meaningless if quality dropped. Check .calibrate/route-log.ndjson (if present) for escalation events, and scan recent sessions for signals of cheap-tier failure: immediate retries, user corrections right after a routed task, test failures following a coding task. Report the count.

Step 4 — Report

Output:

1. Spend by model for the period (table).
2. Actual vs all-top-tier baseline, with the delta.
3. Escalations and quality incidents observed.
4. One recommendation: keep the policy, or re-run /calibrate:project because the data suggests a phase is mis-tiered (e.g. frequent escalations in code → raise its default).

State clearly that dollar figures for subscription (Max/Pro) users represent quota value, not cash.