Decision Session: Structured multi-agent decision reconciliation tool #85
Description
Overview
A focused tool that walks through a set of decisions requiring reconciliation — specs, architecture, process disambiguation — via a structured, step-through session shared between a human and one or more CLI-accessible AI agents.
The core value is providing a heads-up display for decision-making that contextualizes each decision within the broader codebase, plan, or process while enabling collaborative refinement across multiple AI agents (Claude, Gemini, Codex).
Decision Context Display
For each decision in a session, the UI should present:
- Core context — The real, grounded context for this specific decision
- Decision statement — A very clear articulation of what is being decided
- Downstream impacts — What this decision affects or unlocks
- Upstream influences — Prior decisions that constrain or inform this one
- Blockers / dependencies — What must be resolved first
- Decision tree position — Where this fits in the broader structure:
- Timing: Must this be decided now, or can it be deferred?
- Scope: How does it relate to the rest of the codebase/plan/process?
Session Model
Single shared session that steps through a decision set:
- Session is shared between the human operator and one or more CLI AI agents
- Leverages different agents' strengths at different phases
- All agents operate within one shared thread / "mini sandbox"
- Shared state is either in-memory or persisted to a file
- Functions as a collaborative decision draft / scratchpad space
Key properties:
- Agents can be swapped in/out during a session
- State persists across agent invocations
- Human can intervene, edit, or redirect at any point
Artifacts & Outputs
The session can produce various artifacts:
- Architecture diagrams and visualizations
- Specifications
- Plans or roadmaps
- Any format appropriate for the decision set being reconciled
The primary goal is to walk through, contextualize, and reconcile a set of decisions, then produce a final artifact capturing those decisions.
Open Questions
- What's the minimal viable interface? (TUI, web UI, file-based?)
- How do we represent decision graphs / trees?
- What's the agent handoff protocol?
- How much structure vs. free-form in the decision model?
This is an epic-level issue describing the concept and intended behavior. Sub-issues for implementation will be created separately.