Investigate workflow-artifact provenance for triggered agent runs

[Jordak]

Parent

Related to #108.

Current behavior / design

As of this issue being written, Eval Lab can record local trial/run metadata and workspace provenance, but it does not have a clear project-local stance for triggered agent runs, reusable workflow artifacts, record/replay-derived skills, or cross-surface run correlation.

Why change

Agent work is moving from one-off local prompts toward triggered and reusable workflows: scheduled/background automations, GitHub/Slack triggers, recorded workflows, local/cloud handoff, and persistent run stores. Eval Lab does not need to support every vendor surface, but it should decide which neutral provenance fields matter if future trials include triggered or replayed workflows.

Sources:

• Cursor changelog: Improvements to Cursor Automations
• OpenAI Codex Record & Replay
• Existing runtime-metadata parent: Audit runtime metadata against agent control-surface sources #108

What to build

Investigate a neutral workflow-artifact provenance manifest for triggered or reusable agent runs. The output should be a short design recommendation: either no-op because #108 fields are enough, or a narrow schema/reporting follow-up that records only fields needed for comparability.

Candidate neutral fields:

• trigger origin and triggering actor class;
• manual prompt versus scheduled/event trigger versus recorded/replayed workflow;
• workflow artifact provenance and version;
• run correlation id or replay id when supplied by the harness;
• memory/state mutation expectation;
• stop condition and handoff surface.

Acceptance criteria

• The investigation maps each candidate field to existing Eval Lab metadata where possible.
• Unsupported facts are represented as unknown, not invented from vendor UI assumptions.
• The recommendation distinguishes reusable workflow provenance from ordinary task instructions.
• The recommendation includes a non-goal against adding vendor-specific automation fields before a real Eval Lab task needs them.
• If a gap remains, it produces one small follow-up implementation issue rather than reopening broad run-surface metadata expansion.

Blocked by

None - can start immediately