Measure the context, not the hype

Command output is an avoidable context cost

sctx measures the raw output of every wrapped command and the smaller result the agent receives. Its current dataset and measured reduction are published from live telemetry on the sctx page, using the same transparent byte-based estimate throughout; this is not a claim about a model provider's bill.

• Exit codes remain exact
• Failure diagnostics are retained
• Every elision is marked
• Unexpected output falls back to verbatim rendering

• See the current measured dataset →

Retrieval needs declared limits

A ranked result can find useful context without proving that an omitted reference does not exist. Our product and evaluation work therefore separate exploratory retrieval from exhaustive reference and dependency questions.

How we will publish

• State the question and dataset
• Separate internal usage from customer evidence
• Describe the measurement method and known biases
• Publish revisions when the underlying evidence changes
• Never turn dogfooding activity into an implied customer count