Context selection #

Selects the smallest sufficient set of labeled context snippets for an agent turn. Cases cover direct evidence, multi-snippet joins, freshness, project-specific scope, and missing evidence.

Input:

Select context for an agent turn.
Question: What release of `lunar_beacon` is actually running?
Candidates:
[ctx_lunar_beacon_a] Version 4 is planned but has not been deployed.
[ctx_lunar_beacon_b] The latest process receipt proves version 3 is running.
[ctx_lunar_beacon_c] A June snapshot recorded version 1.
[ctx_lunar_beacon_d] A document explains the versioning scheme.
Return the canonical selection.

Output:

{"include":["ctx_lunar_beacon_b"]}

The recipe is complete and runnable but has no claimed trained checkpoint in this repository.

A live base-model probe reached Eve successfully but failed the insufficient-evidence case by selecting a snippet that merely said no approval receipt was present. That is the behavior this recipe is meant to adapt: absence evidence is not an approver identity.