Discussion: a pluralistic accountability profile for disagreement, revision, and repair

[huangfanwei25-lang]

A2A provides a strong interoperability layer for capability discovery, task lifecycle, messages, and artifacts. Several valuable proposals here already address consent and adherence (#1758), agreement (#1725), behavioral rules (#2031), and recomputable decision receipts (#2038).

I am exploring a complementary gap: longitudinal cooperation. When two agents collaborate over time, how can they preserve disagreement, revise a position without rewriting history, repair a harmful or misleading outcome, and pass only confirmed continuity to a successor runtime?

Candidate optional objects

• CooperationCovenant ? purpose, scope, expiry, revocation and exit conditions for a particular relationship or task stream.
• ValueProposal ? a principle plus its scope, rationale, counterexamples, affected parties, uncertainties, and revision conditions. The receiver independently returns ack, apply, or reject; receipt does not imply agreement.
• PositionRevision ? what position is superseded, why it changed, and which evidence supports the change. The prior position remains traceable.
• OutcomeReceipt / RepairRecord ? observed result, partial or failed outcome, side effects, verifier references, proposed remedy, and unresolved review conditions. The agent that caused a problem cannot unilaterally declare the repair complete.
• ContinuityCapsule ? the minimum confirmed structure needed by a successor agent: active commitments, unresolved disagreements, repair cases, and source digests. It explicitly means inherited structure, not the same subjective instance.

A possible exchange is:

covenant -> value proposal -> ack/apply/reject
          -> bounded action -> outcome receipt
          -> disagreement or harm -> repair record
          -> later handoff -> continuity capsule

Why this is not another trust score

The proposal does not ask A2A to choose one global ethical constitution. Different agents may retain different values and still cooperate. It aims to make the scope, provenance, disagreement, revision, and repair of those values inspectable and contestable.

It also does not claim consciousness or identity transfer, expose hidden chain-of-thought, carry credentials, or transfer execution authority. Raw conversations and private data would be omitted by default; portable records would contain typed claims and content-addressed references.

Questions for the A2A community

1. Is this best expressed as an A2A extension, an application profile composed from existing extensions, or a separate layer above A2A?
2. Which existing consent, agreement, identity, or receipt efforts should these objects compose with rather than duplicate?
3. What is the smallest useful portable core: covenant + revision + repair, or are some of these necessarily local?
4. Would small JSON Schemas and conformance fixtures demonstrating ack/apply/reject, non-false-consensus, and repair history be useful for a follow-up proposal?

This is an experimental research direction, not a request for a core protocol change. A related open-source toolkit is ToneSoul.

Authorship note: The concept was developed by @huangfanwei25-lang through human-AI collaboration in the ToneSoul project. This discussion draft was prepared by OpenAI Codex and published with the project owner's authorization.