Optimized usage

How to run Metrecept at the standard it was built for. Short version: let the cache do the repetitive work, fetch with intent, and read your own meters.

Prerequisites: a connected host (Integrations) and the command catalog for parameter detail.

The pipe replays identical prompts from Redis instead of re-billing the provider. To benefit:

• Keep repeated prompts byte-identical. Templates, lint explanations, doc lookups — same wording, same model, same replay.
• Don't inject noise. Timestamps, request ids, or "please" variations in otherwise-identical prompts turn cache hits into misses.
• Check it's working. ohm_savings shows estimated provider $ avoided, pipe rent, and ROI; ohm_usage shows the hit ratio.

Agent loops inflate context every turn. Exact-match still wins on mechanical repeats:

• Stable system / tool preambles; strip clocks and UUIDs from cacheable cores
• Retries after tool errors should HIT — if they always MISS, something in the payload is mutating
• Use cache_control: "no_store" when you need fresh sampling
• Host “context summarised” reduces window bloat; Metrecept removes duplicate full-price journeys when prompts repeat

Two ways to bring the public web into context — pick deliberately:

• ohm_fetch_web when the page content is the deliverable: scraping docs, reading a changelog, comparing listings. format="json" gives structured title/text/meta/JSON-LD; query focuses the summary.
• ohm_chat with fetch_urls when you want the model to reason over pages in one round trip: "read these three pages and answer X."

Rules of the road: public pages only, and every fetch declares a purpose. If a fetch is refused, ohm_policy tells you which purposes your seat allows — don't guess.

BYOK ("bring your own keys") means your provider key — OpenAI, Anthropic — rides each request in the X-Ohm-Upstream-Key header and is not stored. Set OHM_UPSTREAM_KEY once in your host config and forget it; only reach for ohm_chat's upstream_api_key param when a single call needs a different key. The mock model needs no key at all — use it to smoke-test a new connection.

Once a week (or when a bill surprises you):

1. ohm_usage — hit ratio, fetch counts, estimated pipe rent.
2. ohm_savings — provider avoided vs pipe rent + roi_ratio (estimate_only).
3. ohm_providers — upstream health, if calls have been flaky.

You pay pipe rent (routing, caching, compliant fetch), not token wholesale — so the lever you control is the hit ratio and the fetch count, and both are visible on demand.

1. ohm_usage returns? The pipe and your key are fine.
2. mock works but gpt/claude fails? BYOK key missing or wrong — see step 4 above.
3. Fetch refused? ohm_policy for the allowed purposes.
4. Still stuck? ohm_providers for upstream/failover status, then Architecture.