Task Human Interactions

​

Authentication and enablement

​

Inputs

​

Output

{
  "hitl_triggered": true,
  "hitl_completed": true,
  "interactions_total": 1,
  "interactions": [
    {
      "kind": "review",
      "channel": "in_app",
      "tool_name": "internal_review",
      "stage_name": "AP Review",
      "await_id": "…",
      "step_id": "…",
      "status": "completed",
      "created_at": "…",
      "completed_at": "…",
      "wait_ms": 542000,
      "edits": {
        "fields_shown": 14,
        "fields_edited": 2,
        "fields": [
          { "key": "total", "label": "Total", "kind": "header", "before": "1,200.00", "after": "1,250.00" }
        ]
      }
    }
  ]
}

• hitl_triggered: a human was asked — a review form, approval, ask_user question, or email wait exists, even if still pending or later cancelled (“skip & send now”).
• hitl_completed: a human actually responded (a review submitted, an approval answered, a question replied to).
• One interaction per human touchpoint. kind is review, approval, question, or email_wait; channel is in_app, slack, teams, or email. A multi-channel approval (e.g. Teams + in-app) is a single interaction reflecting the channel that answered.
• status: pending, completed, cancelled, expired, or (for question steps) processing/failed.
• edits appears only on completed reviews where the reviewer changed something: every edited field with its before/after value. Table columns report cells_changed counts instead of per-cell values. Values longer than 200 characters are truncated and truncated is set.

​

Limits and side effects

• No side effects; single read of the current task’s state.
• Reports the task’s own interactions only — human involvement inside child tasks spawned by this task is not included.
• Field diffs cover review-form-family tools only. Edits to other tool outputs and the specifics of match-config corrections (old vs. new matched catalog entry) are not broken out.
• At most 100 interactions are listed — the newest 100, with interactions_total giving the full count and interactions_truncated set when the list was cut. The booleans always reflect the full set.
• Field diffs are attached to at most the 20 most recent completed reviews; at most 50 edited fields per review.

​

Expected errors

• no task id in context — called outside a running task.
• task human interactions are unavailable: repositories not configured — misconfigured deployment.