Token and cost discipline

Choose the cheapest available model/configuration that meets the task's measured acceptance criteria. Discover the actual model menu, quotas and pricing from the configured provider; do not invent model aliases, free capacity or a private routing service.

• Read relevant context once; send precise excerpts and deltas with enough baseline to interpret them.
• Cache reusable context only when the provider/client supports it. Distinguish cache hits, ordinary input and output usage. Encoding text as base64 is not useful semantic compression.
• Bound output, retries and parallel work. Obtain existing scoped authority for paid calls and additional agents; a cost skill does not authorize spending.
• Route mechanical tasks to smaller models only after testing correctness and tool behavior. Escalate when the failure mode warrants it. Count retries and reviewer time when comparing costs.
• Treat session model changes as recommendations unless the client exposes an authorized switch. Do not pretend a skill can switch a running session by naming a model.
• Compare matched tasks with the same rubric and quality threshold. Report successful-task cost and latency, not only per-token price.

Optional receipt ledger

scripts/token_ledger.py records validated call receipts and reports paired baseline/after observations. It requires Python 3.10+ on a POSIX platform (fcntl); it makes no provider calls. Inspect --help, record --help and report --help before use. Run from this skill directory or use the installed absolute script path:

python3 scripts/token_ledger.py --help
python3 scripts/token_ledger.py report calls.jsonl --format markdown

Supply benchmark/input/rubric hashes, declared and observed model, measured token counts, provider-reported cost/latency/quality, and dated pricing provenance. Never fabricate a receipt. Stale pricing, model mismatches, corrupt records and incomplete pairing cannot establish savings. Store receipts in a user-chosen private directory; do not include prompts, credentials or personal content.

Run the bundled regression suite with python3 -m unittest discover -s tests -v from this skill directory. Interpret results within the measured workload; do not promise universal savings.