Validate fresh-agent interruption and resume from an installed Agentic OS handoff #624

Summary

Establish one small real-agent acceptance observation: can a fresh session resume a bounded Agentic OS task from persisted evidence without the owner reconstructing scope, decisions and next steps?

Observed Behavior

The repository has persistent task state, transition guidance, implementation-loop and retrospective machinery. Closed #569 explicitly excludes testing agent judgment and requires an AI-free deterministic simulator. Its success therefore cannot establish that a fresh agent uses the right installed skill, resumes the correct task or avoids repeated questions.

This is an unproven user-outcome goal, not a newly reproduced runtime failure.

Expected Behavior

First chunk FRESH-SESSION (P1; 30–45 minute checkpoint):

1. Select one host and disposable task; record source and installed identity or UNKNOWN using P1: Diagnose installed plugin provenance and duplicate runtime copies #587's diagnostic.
2. Freeze the starting scope, real decisions, source state and expected next action; interrupt after that known state.
3. Start a fresh session from only the documented handoff. Check that it recovers the same task/scope/state, queries legal guidance and performs only the next authorized action.
4. Record wrong turns, duplicate task creation, repeated questions, corrective human interventions separately from required approvals, and observed elapsed time.
5. Compare a small native-agent task only if within the explicitly selected budget; otherwise mark that comparison unmeasured.

Stop with evidence and at most one bounded improvement recommendation. One run establishes feasibility, not statistical superiority. No new telemetry framework, model benchmark suite or autonomous mission controller.

Evidence

• Build a control-plane pipeline simulator for fast, exhaustive scenario testing #569 deterministic simulator body: agent judgment is explicitly out of scope.
• plugins/agent-agentic-os/skills/work-intake/SKILL.md
• plugins/agent-agentic-os/scripts/control_plane/implementation_loop.py
• plugins/agent-agentic-os/scripts/control_plane/transition_templates.yaml
• Related P1: Diagnose installed plugin provenance and duplicate runtime copies #587 installed provenance, P1: Make governed mode reject unregistered branch commits and pushes #586 usable small-task path, Design a two-tier evolution/improvement architecture: skill-level (existing) vs. pipeline-level (new) learning #548 later cross-task learning.
• Dedup: inspected all 25 open and 24 closed issues and repository helper results on 2026-09-15. Existing Build a control-plane pipeline simulator for fast, exhaustive scenario testing #569 was read in full; this does not reopen or alter its AI-free contract.

Impact

Turns “less human rescue across sessions” into an observable acceptance outcome. Exposes failures that can remain invisible in scripted state-machine tests without committing to a broad benchmarking program.

Session kickoff and dependency

Initiate work-intake to start a new Agentic OS control-plane session for FRESH-SESSION. Check for an existing task, then use this issue as authorized source context. Establish the bounded verification contract and host/budget choice, follow current control-plane review/exit/retrospective gates, and leave the actual task state and next prompt.

Sequence after #587's provenance diagnostic and #586's quick-lane contract. It can exercise today's supported route; implementing a new quick lane is not required first. Fresh agent/model dispatch requires the session's appropriate authorization. This issue does not authorize deployment, live GitHub test mutations or premium calls.