[cris005]

Part 4 of 6. This is where the ledger starts counting. The sidebar rows come in #33.

The savings setting

• Where it's set: default on. In the picker (Session and Global), as /cache-manager savings on|off, and as a plugin option. Turning it off hides the rows, but the ledger keeps counting.
• modelsDevPrices option: new, default off. It turns on the models.dev price fallback.

What counts as a save

An idle stretch counts as one avoided cache rewrite when all three hold:

• at least one ping in the stretch was a hit;
• the next real primary or compaction request is sent more than 30 s after the cache would have expired without pings;
• that response is ok and reads at least 90% of the prior context from cache.

Every ambiguous case counts less:

• hybrid: only the part past the long-lived prefix counts. The prefix is the largest 1h write on Claude, or the first primary context on GPT-5.6+. If the prefix size is unknown, no save is counted.
• Return turn: a failed or overtaken return turn drops the save.
• Restart: a restart mid-stretch never produces a save.
• Ping charges: a ping is charged at the TTL it was sent on. A ping answered after its stretch ended still counts against that stretch.

Wiring

• Engine: reports to the ledger fire-and-forget. Snapshots gain showSavings and savings. The readout loads only for managed, non-child sessions with savings shown.
• Middleware: passes split, model and reportedModel to the engine.
• Server:
  • Builds the PriceBook, models.dev fallback and ledger.
  • Prefetches prices with void prices.prefetch(), so nothing on the request path waits for a price.
  • Ledger changes refresh the snapshot.
  • One startup pass sweeps the session/ and ledger/ keys of deleted sessions.
• Capture sample: the sample snapshot gains the two new fields.

Testing

• Engine tests: every rescue edge case.
• Server tests:
  • restart mid-stretch;
  • a catalog that never loads;
  • models.dev on and off;
  • the sweep.

Before release

• Confirm at runtime that a cost declared on a custom provider's model in opencode.jsonc comes back from ctx.model.list().

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
1.1% Duplication on New Code

See analysis details on SonarQube Cloud