Parse narrowly
Reject shell composition, variables, wildcards, non-UTF-8 encodings, ranges, waits, and unknown flags.
Measured file-read optimization
Capsule recognizes only a narrow, safe form of PowerShell Get-Content or gc: one literal UTF-8 file, no pipeline, wildcard, variable, wait, range, or unknown switch. Everything else keeps the normal shell semantics.
| Arm | Model-visible output | Reduction |
|---|---|---|
| Generic projector | 315,357 local token estimates | Control |
| Native path + verified fallback | 42,120 local token estimates | 86.64% |
| Unchanged-file replay | 2,814 tokens across 201 calls | 94.21% vs control subset |
The bounded replay fixture retained exact recovery for every replayable case and recorded no regressions. Counts and payloads are intentionally summarized rather than publishing local replay-set details. These are local model-visible proxies, not provider billing or hidden reasoning telemetry.
Reject shell composition, variables, wildcards, non-UTF-8 encodings, ranges, waits, and unknown flags.
Replay requires the same request and an unchanged SHA-256 source hash.
If the generic projector is shorter, the native path delegates back instead of increasing model-visible context.
npm run benchmark:get-content
npm run benchmark:get-content:history
Read the implementation in mcp/get-content.cjs and the full A/B report in BENCHMARK.md.