[Feature]: Add a structured expert handoff and acceptance contract

[GTC2333]

Problem or motivation

Delegated work currently returns through a free-form complete_task(reply=...) string. Personas ask experts to include conclusions and artifact paths, but the Principal cannot reliably distinguish:

• verified results from interpretation
• complete work from partial work
• evidence from unsupported claims
• executed checks from recommended checks
• an intentionally absent artifact from a forgotten artifact
• skill grounding from generic model knowledge

This makes acceptance depend heavily on the Principal rereading arbitrary prose and weakens automated evaluation.

There are also coordination ambiguities:

• The high-impact action gate tells an expert to notify the task creator and wait, but does not specify whether to complete the current task, dispatch a new task, or use another protocol.
• Principal is described as the single user-facing orchestrator, while Writer can call ask_user for general writing clarification.
• Multi-hop delegation requires agents to stop after dispatch, which can create unnecessary sequential model turns.

Proposed solution

Define a versioned structured completion envelope while preserving a human-readable summary. Suggested fields:

{
  "status": "complete | partial | blocked",
  "summary": "...",
  "claims": [
    {
      "text": "...",
      "evidence": ["path, citation, or observed command output"],
      "confidence": "high | medium | low"
    }
  ],
  "skill_usage": [
    {
      "name": "...",
      "rules_applied": ["..."]
    }
  ],
  "verification": {
    "performed": ["..."],
    "not_performed": ["..."]
  },
  "artifacts": ["..."],
  "uncertainties": ["..."],
  "requested_action": null
}

A high-impact authorization request should use an explicit requested_action shape and route deterministically to Principal without falsely completing the scientific task.

Clarify which roles may contact the user and under which conditions.

Acceptance criteria

• complete_task accepts or produces a versioned structured result.
• Plain-text compatibility is retained during migration.
• Principal receives explicit status, evidence, verification, uncertainties, and artifacts.
• Partial and blocked work cannot be mistaken for complete work.
• Authorization escalation has one documented tool path and cannot deadlock a task.
• The user-contact policy is consistent across Principal, Writer, and tool allowlists.
• Tests cover Principal-to-expert and expert-to-expert delegation, partial completion, authorization escalation, and malformed envelopes.

Area

Agents / orchestration

Alternatives considered

Prompt-only formatting guidance is simpler but cannot be validated reliably and will drift across models.