Everything a model writes costs output tokens and is bounded by its output limit, including the arguments of a tool call. An agent asked to turn a large query result into a file had to retype the rows, and a long result was cut off. This release changes how results travel between tools, and how tools are found.
Results by reference
- Captured by size A tool result of 4,000 characters or more, configurable, is stored in the conversation's scratchpad, and the model receives it followed by a line naming its key. Errors and results that end the turn are not captured.
- Exported by key A chat tool writes a stored result to a conversation file by its key: as it is, or, for rows in JSON, as CSV. Other JSON is refused as CSV rather than flattened by a guess.
- Notes on stored files A tool that writes a file can add what it knows about it, such as rows written or a limit reached, after the file's identity.
- Delegation A sub-agent's summary now asks for the key of a stored result instead of its rows.
Calls that do not need to run
- Repeats answered A tool can declare that an identical call is answered from what the model already holds. Loading a skill already served returns "already loaded, follow it" instead of an error; one agent had been reloading the same skill two to six times a run. The repeat is still charged against the run's budget.
- Unknown arguments named An argument a tool does not take is named back to the model in the same result, with the arguments it does take, instead of being dropped silently.
Tool search, ranked
- BM25 Deferred tools are ranked by BM25 over the name's parts, weighted three times, and the description's first paragraph, with English stemming; a tool must score at least half the best match, and at most five load.
- Names first An exact name wins, a name without its namespace resolves, and a name that is not offered loads nothing.
- Open-world tools by name only Web search and HTTP request load only when named, never on keywords. Production logs had shown them loading on unrelated queries through substring matches.
- Replayed from the log The tests replay 174 logged queries against the tool set of their session.
The model's context is spent on reasoning about data rather than copying it, and the tools it loads are the ones it asked for.