User reference / Run an experiment
4. Execute the case
Understand the stages, selection rules and records produced by a collaboration.
Run from the repository root with your keys configured:
slow-thinker examples/software_incident_management_process.json \
--base config/slow-thinker.study.json
The equivalent module entry point is python -m slow_thinker. Every invocation registers a new task UUID. Re-running the same task does not resume an interrupted run; it creates a new experiment with the same prompt identity and potentially the same configuration identity.
| Stage | What happens |
|---|---|
| RegistrationReviewed E 0.1.5 | Save the prompt and effective configuration, then create the task record with component versions and timestamps. |
| Initial proposalsReviewed E 0.1.5 | Launch one call per proposer concurrently. Validate each answer, assign step UUIDs and persist valid proposals as round 0. |
| RefinementReviewed E 0.1.5 | Wait for the previous round, then launch the next round concurrently with its valid proposals as context. Repeat the configured number of times. |
| VotingReviewed E 0.1.5 | Give the final proposals to the configured voters in parallel. Persist only votes with a valid selected proposal ID. |
| SelectionReviewed E 0.1.5 | Select the most-voted proposal, save its steps as a plan and update task status. No extra synthesis call is made. |
| CompletionReviewed E 0.1.5 | Mark the task completed if a plan was selected. Generate HTML automatically if enabled and available. |
Progress lines show the elapsed process time, agent, provider/model, step count, reported token usage and call duration. Full prompts and responses also appear in console output. Save console output if you need to investigate attempts that did not become valid trace records.
| Situation | Recorded behavior |
|---|---|
| One proposal call fails or returns an invalid proposalReviewed E 0.1.5 | Its error is printed and the proposal is skipped. Other valid proposals continue. A later round still attempts all configured proposer slots. |
| A whole round produces no valid proposalsReviewed E 0.1.5 | The graph routes to the error handler and the task is marked failed. |
| A vote is invalid or its call failsReviewed E 0.1.5 | It is omitted from the saved votes. Actual voter counts in reports may be lower than configured. |
| No valid votesReviewed E 0.1.5 | The first final proposal is selected as fallback. A selected plan therefore does not by itself demonstrate voter support. |
| Tie for the highest vote countReviewed E 0.1.5 | The first tied proposal encountered while counting valid votes wins; valid votes are processed in configured voter order. There is no tie-breaking model call. |
| Keyboard interruptionReviewed E 0.1.5 | The CLI marks the task interrupted when handled. Records already persisted remain available. |
| Unhandled execution exceptionReviewed E 0.1.5 | The outer handler attempts to mark the task failed and stores the error summary. Storage failures may prevent complete persistence. |
Transient call errors are retried up to three attempts by the wrapper, with one- and two-second waits. The provider library may have its own retries. The wrapper does not retry every bad request, and does not run a separate semantic repair dialogue for an invalid answer. See execution internals.
With JSON storage, the new UUID directory contains task, proposals, votes and plan records as they become available. Shared prompt and configuration records live in sibling directories. Missing plan.json means no selected plan was persisted. The report library is a static snapshot, so regenerate it to see later trace changes.
“Completed” means a plan was selected and the collaboration finished. The plan's actions are not executed by Slow Thinker, and their success has not been established.
Continue with optional analysis.
Source files used for this reference
Reviewed at 7361ea0b71a5.