Every version, and what it changed for you.

Asking whether a file contains a string, without buying the file.

• Newcontains — your agent can ask whether literal strings appear in a set of files and get back paths, occurrence counts and line numbers, never file contents. This was the question that had been costing whole files: a phrase inside a comment has no declaration and cannot be guessed from a name, so the agent bought all twelve files to settle it. On 0.18.0 the answer costs 198 tokens where reading those twelve files costs 40,484.
• ImprovedA file whose compact form would come out larger than the file itself is served as the file. Small and highly repetitive files — a compose file whose every line is already one key and one value — could project to more than their own source, and nothing was checking.
• FixedA shortened answer said it had served everything. Batch reads counted what had been rendered rather than what was sent, so a response carrying nine files announced twelve — and an agent that cannot reconcile those two numbers concludes the payload is broken and goes looking for a way around it.
• FixedA batch that served every file but shortened two of their symbol lists now names those two files, and says plainly when there is nothing left to ask for. It used to report no truncation at all, so the agent re-read fifty-seven files to find the two.
• FixedAsking for one file on its own now returns all of it. A per-file ceiling was fixed regardless of the budget you asked for, so the advice printed beside a shortened list bought a turn and returned the identical list.
• FixedResponses say how much of the exact-source budget is left, so your agent can size its next request instead of bisecting toward the limit. One measured run spent five turns doing exactly that.