Model-visible payload
Character and local token estimates compare the text shown to the model before and after a policy activates.
Codex token efficiency
Reduce repeated context without losing the evidence.
Capsule is designed for the part of Codex usage we can observe: tool results, file reads, project exploration, web payloads, and repeated context. It keeps a compact view in the conversation and an exact local capsule for recovery.
What is measured
Exact recovery
Content hashes and capsule IDs verify that a compact view can recover the original local result.
Monotonic safety
If a compact representation would be larger or less safe, Capsule passes through or falls back to the existing projector.
What is not claimed
Capsule cannot see hidden provider reasoning, provider-side caching, subscription limits, or billing counters. A local reduction is not automatically a billing reduction. Every benchmark therefore states its workload, control arm, recovery invariant, and exclusions.
Where the savings come from
| Workload | Technique | Safety rule |
|---|---|---|
| Terminal output | Keep failures, critical lines, and novel evidence; omit repeated boilerplate. | Failed or explicit full output passes through. |
| Files and projects | Return task-conditioned ranges, symbols, and dependency impact first. | Exact expansion remains available. |
| Web and tools | Keep structured fields, URLs, and references in a bounded view. | Real URLs and recovery IDs are required. |
| Repeated reads | Replay unchanged content from a verified hash. | Any content change invalidates replay. |
Run your own measurement
npm run verify
npm run benchmark:get-content
npm run benchmark:get-content:historySee the full methodology in BENCHMARK.md and the implementation in the MCP source.