Improve shell command output handling to avoid context overflow

When invoking shell commands (e.g. build), the tool often produces a very long output, which can easily blow the LLM’s context window during evaluation or downstream reasoning.

Proposal:
Let the LLM decide how much of the output to read:
Either full output (if needed for context),
Or just the last N lines (e.g., tail 20),
Or allow tool to summarize the output (e.g., by heuristics or patterns).

Start at the shell-command invocation path used for builds and inspect how command output is collected and passed into downstream reasoning. Compare the proposed full-output, last-N-lines, and heuristic-summary behaviors, then define which option is selected and how completion is verified without context overflow.

Scritto dal modello di indicizzazione a partire dal testo della issue.