Example: document review workflow

This example shows how to use Bijux Agent as an auditable “review runner” over a directory of documents.

Goal

Given a folder of files, produce per-file structured results and an overall run verdict you can audit.

Recommended setup

1. Put your documents in a single directory (non-recursive processing today).
2. Set a review-oriented task goal in your config, for example:

task_goal: "review this document for correctness, missing assumptions, and actionable revisions"

Run

python -m bijux_agent.main run path/to/review_corpus --out artifacts/review_run --config config/config.yml

How to interpret the outputs

• artifacts/review_run/result/final_result.json is the first file to read.
• artifacts/review_run/trace/run_trace.json is what you use to validate:
• which model metadata was used,
• what the replayability classification is,
• what fingerprints identify the run inputs and config.

Practical advice

• Keep the task goal narrow. “Review everything” yields unstable scope and weak auditability.
• Treat the run directory as immutable once produced; regeneration should go to a new --out directory.

See also:

• docs/user/usage.md (CLI mechanics)
• docs/spec/failure_model.md (how failures are represented)