| # Inference contract |
|
|
| This describes supported structures, not a guarantee that English intent will be |
| recognized correctly. Inputs outside these structures may be falsely accepted. |
|
|
| ## Input |
|
|
| Each JSONL record is an object with: |
|
|
| - `instruction`: an English string explicitly naming catalog tools and arguments. |
| - `tools`: 1–8 catalog entries, each containing exactly `name`, `arguments`, |
| `required`, and `returns`. |
|
|
| A catalog entry looks like: |
|
|
| ```json |
| { |
| "name": "list_users", |
| "arguments": { "team": "string" }, |
| "required": ["team"], |
| "returns": { "kind": "array", "properties": ["id", "name"] } |
| } |
| ``` |
|
|
| Tool names must be JavaScript identifiers. Argument schemas contain at most four |
| properties of type `string`, `number` or `boolean`; nested objects, arrays and |
| null arguments are unsupported. Required argument names must exist in the schema. |
| Return kinds are `text`, `object` or `array`, with declared property names where |
| projection is needed. Quoted values use JSON string syntax and remain opaque. |
| Numbers must be finite and compatible with exact JavaScript numeric handling. |
|
|
| The frontend diagnoses inputs over 128 tokens rather than truncating them. |
| The model does not discover tools or interpret free-form tool descriptions. |
|
|
| ## Plan |
|
|
| A successful response contains a `plan`, for example: |
|
|
| ```json |
| { |
| "mode": "single", |
| "calls": [{ "tool": "list_users", "args": { "team": "blue" } }], |
| "output": { "kind": "last" } |
| } |
| ``` |
|
|
| Supported schedules are one call, or 2–3 independent calls sequentially or in |
| parallel. Repeated tools are allowed. Sequential calls do not pass earlier |
| results into later arguments. Dependencies, mixed serial/parallel groups, |
| conditions, retries, loops and arbitrary code generation are unsupported. |
|
|
| `output.kind` is `all` or `last`. For one array-returning call only, optional |
| `fields` projects 1–3 declared properties and optional `limit` takes the first |
| 0–100 rows. Projection precedes the limit. Filtering, sorting and aggregation |
| are unsupported. |
|
|
| Rejected requests return a `diagnostic` string instead of a plan. Diagnostics are |
| not evidence of reliable semantic rejection. Valid plans can be semantically |
| wrong even when schema checks pass. |
|
|
| ## Compilation, not execution |
|
|
| The CLI's `--js` option also returns a `js` string: an async-function body intended |
| for a host exposing `tools[name](args)`, top-level `await` and `return`. Nothing in |
| the CLI invokes these tools. |
|
|
| Generated code awaits every started call. Parallel mode uses `Promise.allSettled`; |
| serial mode catches each call's failure and continues in mention order. Results |
| are `{ok: true, value: ...}` or `{ok: false, error: ...}` records. `all` returns |
| records in call order; `last` returns the final call's record, not the last call to |
| finish. Single-array shaping reports invalid result shapes explicitly. |
|
|
| The compiler escapes names and values mechanically and validates structural |
| constraints. These properties do not grant permission, establish intended |
| semantics, or provide execution isolation. Real tool use needs independently |
| implemented authorization, intent review and a suitable execution environment. |
|
|
