Web evidence without repeated payloads

Keep web search useful without carrying the whole page every turn.

Capsule can expose a bounded web result while preserving the complete response, URLs, and reference identifiers in a local exact-recovery lease. The feature is opt-in because a compact view must never silently discard evidence.

First call and repeat call

First call

Run the web request through fetch. Capsule stores the original response, fingerprints it, and returns structured evidence only when the projected view is smaller and safe.

Repeat call

For an unchanged response, a content-addressed lease can avoid carrying the same page text again. Recover exact text, URL, or ref_id only when the task needs it.

Measure the real workload

$env:CAPSULE_WEB_ZERO_COPY="1"; npm run benchmark:web:zero-copy

Compare the baseline response and the Capsule lease for the same query, URL set, and output requirement. Record visible bytes or token estimates, recovery success, latency, and fallback cases. A small or explicit full request should pass through when the lease would add overhead or hide detail.

What remains exact

The original response is written before a compact view is emitted.
Recovery is addressed by a real content hash and verified locally.
URLs and reference identifiers remain available for follow-up work.
Failed, unsafe, ambiguous, and token-negative candidates use the conservative path.

A web-search percentage is one model-visible workload measurement. It is not a promise about hidden provider reasoning, provider caching, billing, subscription quotas, or every search result.

Read the implementation in mcp/unified.cjs, compare related MCP context workflows, and ask reproduction questions in the measurement discussion.