Execution guarantees (spec)

This document describes what a caller can rely on after invoking a run.

CLI guarantees (normal run)

For python -m bijux_agent.main run … without --dry-run:

• The CLI MUST create a run directory at --out.
• The CLI MUST write result/final_result.json.
• The CLI MUST write trace/run_trace.json.
• The trace MUST validate (or be upgradeable) under the local trace schema.

CLI guarantees (dry-run)

For run --dry-run:

• The CLI MUST validate input discovery and configuration loading.
• The CLI MUST write result/final_result.json.
• The CLI MUST NOT claim model-computed content.
• The CLI MAY skip producing a trace.

API guarantees (v1)

For POST /v1/run:

• The handler MUST validate the request payload with the v1 schema.
• The handler MUST return a structured response with success and context_id.
• On failure, the handler MUST return a structured error object.

Non-guarantees

• No guarantee that outputs are “correct” or “complete”.
• No guarantee of identical outputs across time, even under temperature 0 (providers can change).
• No guarantee of stable performance, throughput, or token usage.