[Sub-Task] Extend Feedback type with messageId for ACP `session/inject` (RFD #1261) support

[Leoyzen]

Context

Parent: #135 (AgentWolf v1 Foundation Architecture — Layer 4 RunLoop)

ACP RFD #1261 proposes session/inject with mode: "queue" | "steer", agent-owned messageId, user_message echo notification, and mandatory session/revoke_inject. This is the upstream protocol standardization of the steer/queue pattern AgentPool already implements internally via CommChannel + Feedback.

Related Bug Fix (Completed)

The steer() bug (PR #144 comment) was fixed in this session: steer() now checks enqueue("asap") before the CommChannel feedback path, ensuring ProtocolChannel sessions (ACP/OpenCode/AG-UI) get mid-turn injection. TDD tests added in tests/lifecycle/test_run_loop.py.

What's Missing

AgentPool's Feedback dataclass (src/agentpool/lifecycle/types.py L61-72) currently has only:

@dataclass
class Feedback:
    content: str
    is_steer: bool

No messageId, no revoke, no replace, no steer_in_stream capability. The RFD requires all of these for session/inject compliance.

Tasks

P1: Extend Feedback type

• Add message_id: str | None = None field to Feedback dataclass
• Add content_blocks: list[ContentBlock] | None = None for structured content (ACP ContentBlock[] shape)
• Generate message_id in ProtocolChannel.deliver_feedback() when not provided (UUID-based)
• Update steer() and followup() to accept and propagate message_id

P1: Define cancel() interaction with pending steer/followup

• RFD #1261 specifies: pending injects survive session/cancel, deliver at next idle
• Audit current RunHandle.cancel() — does it clear CommChannel feedback queue?
• Add test: cancel() during pending steer → steer survives and delivers next turn
• Document the interaction in steer() docstring

P2: Implement session/revoke_inject semantics

• Add RevokedFeedback tracking (tombstone by message_id)
• ProtocolChannel.recv() skips revoked feedback
• Add RunHandle.revoke_steer(message_id: str) -> bool method
• Add test: revoke before delivery → no user_message emitted
• Add test: revoke after delivery → returns already_delivered error

P2: Add steer_in_stream capability declaration

• Define capability: ["interrupt"] (truncate stream) vs ["finish"] (let stream complete)
• Expose via agentCapabilities in ACP initialize response
• Native agents (PydanticAI) → ["interrupt"] (PendingMessageDrainCapability supports this)
• ACP agents → ["finish"] (must wait for tool boundary, can't interrupt mid-stream)

P3: Evaluate session/replace_inject (opt-in)

• Add pending.replace: bool capability flag
• ProtocolChannel.replace_feedback(message_id, new_content) -> bool
• Only if underlying runtime supports it (content not yet serialized into prompt envelope)

P3: ACP protocol mapping

• Implement _session/inject extension method on ACP v1 (underscore prefix, per ACP extensibility spec)
• Map RunState.RUNNING → ACP v2 state_change(state="running")
• Map RunState.IDLE → ACP v2 state_change(state="idle")
• Evaluate upgrade to v2 prompt lifecycle (non-blocking session/prompt)

RFD #1261 Key Design Points

Aspect	RFD Design	AgentPool Current
Method	session/inject with mode: "queue" | "steer"	steer() + followup() (separate methods)
Response	Immediate messageId (agent-owned)	Fire-and-forget (no ID)
Delivery signal	user_message notification with messageId	CommChannel feedback queue drain
Revoke	session/revoke_inject (mandatory)	Not implemented
Replace	session/replace_inject (opt-in via pending.replace)	Not implemented
steer_in_stream	["interrupt"] / ["finish"] capability	Not declared
cancel interaction	Pending injects survive cancel	Undefined (independent paths)
Version	v2-only (depends on v2 prompt lifecycle)	v1 (v2 not yet implemented)

References

• RFD: docs(rfd): mid-turn input via session/inject (queue and steer) agentclientprotocol/agent-client-protocol#1261
• Discussion: https://github.com/orgs/agentclientprotocol/discussions/1220
• Fork implementation: yoke233/claude-agent-acp session/steer notification
• Production feedback: Raft team reports steer absence is ACP adoption blocker (comment)
• PR Refactor/agentwolf v1 #144 steer bug fix: Refactor/agentwolf v1 #144 (comment)
• Parent issue: [RFC] AgentWolf v1 Foundation Architecture — Six-Layer Design & Migration Plan #135

[Leoyzen]

Linked PR #157 which adds the OpenSpec change v2-message-id-infrastructure and RFC-0054 covering the full design for message ID unification, Feedback extension with revoke/replace, and cross-protocol (ACP + OpenCode) considerations.

The PR has been reviewed by Oracle (2 rounds, APPROVED) and Momus (PASS). Ready for implementation via /opsx:apply.