Evaluate: Steering and follow-up message queues for ADF agent interaction

[AlexMikhalev]

Parent Epic

#682 -- Evaluate Pi architectural patterns

Pattern

Pi distinguishes two types of messages sent while the agent is working:

Queue	Trigger	Effect
Steering	Enter while agent runs	Delivered after current tool completes. Remaining tools skipped. Agent responds to interruption.
Follow-up	Alt+Enter while agent runs	Delivered only after agent finishes all work. Triggers another turn.

Both queues are configurable: "one-at-a-time" (waits for response before next) or "all" (delivers all queued at once).

Current State

ADF agent interaction model:

• Agents are subprocesses (claude -p, codex exec) that run to completion
• No mechanism to steer a running agent mid-task
• No follow-up queue -- must wait for agent_end then spawn new process
• The reconciliation loop runs on a 30s tick -- no real-time interaction
• Claude Code's --resume enables multi-turn but requires process restart

Evaluation Questions

1. Is real-time steering valuable for ADF? Current agents run autonomously. Would "stop, do this instead" improve outcomes for long-running tasks?
2. How would this work with subprocess agents? Pi's queues work because the agent loop is in-process. ADF spawns external processes. Options: (a) migrate to Agent SDK (in-process), (b) use stdin pipe to subprocess, (c) signals.
3. Does the Agent SDK support steering? Check if query() accepts mid-stream prompts or if it requires session resume.
4. Follow-up vs reconciliation loop: ADF's 30s reconciliation tick is a crude follow-up mechanism. Would explicit follow-up queues improve this?

Acceptance Criteria

• Map ADF agent interaction points where steering would add value
• Assess Agent SDK capabilities for mid-stream interaction
• Design: how would steering/follow-up integrate with the reconciliation loop?
• Decision: implement (via Agent SDK migration or subprocess protocol), defer to Agent SDK migration, or reject

Dependencies

• Agent SDK evaluation (if steering requires in-process agent loop)
• Phase 2: Session persistence for Claude Code agents in ADF #639 Session persistence for Claude Code agents in ADF

References

• Pi steering/follow-up: https://github.com/badlogic/pi-mono/tree/main/packages/agent (see "Steering and Follow-up" section)
• ADF reconciliation loop: terraphim-ai/crates/terraphim_orchestrator
• Agent SDK sessions: knowledge/claude-agent-sdk-reference.md