Inside the decision agent
This page is for people working on the demo's decision agent itself: its classes, the stream the page draws it from, and the tests that pin it. What the agent does and what a run looks like are in The decision agent.
The runtime's tools.deciding.DecisionThinker is the loop; this package is the workload it
runs, the artifacts and tools of that workload, and the stream a page draws it from.
What is here
| Class | Role |
|---|---|
| PriceChangeFinder | The agent: a DecisionThinker<Folder, Statement> whose objective is to find the statements in a folder's documents that decide a change of a price, over a palette of three DecisionTools that call no model, SplitStatementsTool being the key tool (the one that produces the statements the answer is a list of). The palette is legal because it compiled. capabilities() is the objective and the palette for a page. |
| Folder, FileEntry, Document, Statement | The artifacts, each with a @TypeAlias: the folder a run starts from; a file as the listing saw it (name, kind by its bytes, size, first words); a file read as text; one statement of a document with its source and place. |
ListFolderTool (list_folder) | Folder to a list of FileEntry: every regular file, by name, sniffed the extract step's way (Sniff), a text file's first words decoded; a path that is not a folder refused with the reason (InvalidInputException). |
ReadDocumentTool (read_document) | FileEntry to Document: a text file decoded (refused as not text past 5 percent undecodable characters), an Office document with its tags stripped (DeterministicExtractTool.officeText); a PDF, an image, a binary or an empty file refused with the reason (InvalidInputException), which the thinker feeds back to the model as a fact of the run. |
SplitStatementsTool (split_statements) | Document to a list of Statement: a paragraph's sentences, each bullet or numbered item; headings dropped, fragments under three words dropped. |
| Digests | What the model reads of each artifact type, registered as the types' text formatters: a decision thinker's state is one digest line per artifact and nothing else of it, so these lines are what the model decides on. A file reads as name (kind, size): first words, a document as name (N characters): first words, a statement as source #n: text. |
| DecisionTrace | The run streamed as newline-delimited JSON to any sink: every event of the thinker and of every job under it, a decision's completion carrying the whole exchange (the state shown, the questions, every distribution answered, the call's account), a tool's completion carrying what it produced (ref and digest, a list's iterands the same way), and a last line with the statements selected, the thinker's recorded turns and the totals, or the failure. Ends the stream from the run's own terminal event and tells the host; awaitEnd and fail are for the handle's holder, as with the agent's trace. |
| DecideCapabilities | The palette as a page shows it (PriceChangeFinder.capabilities(): the objective, the input and answer aliases, and per tool its name, description, and the aliases it takes and produces, list for a list). A run takes no request: the host hands the agent the shipped corpus. |
The tests
PriceChangeFinderTest drives the whole run through the dispatcher against the runtime's
fake decision provider on a scripted policy that opens the two meeting documents, splits them
and selects the two statements that decide a price (the 20 percent rise from 1 June 2026, the
Kestrel 1 prices effective 1 April 2026), and reads the stream back; it also pins that a sink
which loses its reader stops the lines and not the run, and what capabilities() reports.
DecideToolsTest pins the three tools and the digests on the corpus.