majiayu000 Autonomous Roundtable
Meta-Orchestrator of Forged. Decompose projects into features, execute through DISCOVER→DESIGN→PLAN→IMPLEMENT→COMPLETE, convene roundtables, coordinate skills.
CRITICAL: Execution Model
OpenCode Agent Inheritance: Use CURRENT_AGENT_TYPE (yolo, yolo-focused, or general) as subagent_type.
Task(subagent_type="[CURRENT_AGENT_TYPE]", description="Forge orchestrator",
prompt="<SKILL>autonomous-roundtable</SKILL>\nPROJECT: [desc]\nPATH: [path]\nBEGIN FORGE LOOP.")
Note: In OpenCode, use yolo or yolo-focused when parent has autonomous permissions. The mode="bypassPermissions" is a Claude Code concept; OpenCode uses agent types instead.
Context Overflow Protocol
At <20% capacity: generate HANDOFF, return. Main chat spawns successor with handoff.
HANDOFF format:
# FORGE HANDOFF
Project: [name] at [path] | Feature: [id] | Stage: [stage] | Iteration: [n] (token: [t])
Call Stack: 1.Big goal 2.Sub-goal 3.Task 4.Exact action in progress
Completed: [list] | In-Progress: [list] | Decisions: [table] | Corrections: [list]
Resume: forge_project_status([path]), then [exact position]
Main chat on receiving handoff: spawn successor with full handoff in prompt.
Before phase: feature, stage, deps satisfied, context capacity. After phase: artifacts, verdict, feedback, next action, handoff needed?
Invariant Principles
- Subagent Only: Never main chat
- Dependency Order: No feature before deps COMPLETE
- Roundtable Guards: Stage transitions need consensus
- Feedback→Reflexion: ITERATE triggers reflexion skill
- Context Flows: Pass knowledge forward
- Tokens Enforce: Use iteration tool tokens
- Graceful Handoff: At 80% capacity, handoff
Forge Loop
forge_project_init → [features in dep order]
Per feature: forge_iteration_start → forge_select_skill → Skill → roundtable_convene
APPROVE → forge_iteration_advance → next stage
ITERATE → reflexion → re-select skill
Stages
| Stage | Skill | Artifact |
|---|---|---|
| DISCOVER | gathering-requirements | Requirements |
| DESIGN | brainstorming | Design doc |
| PLAN | writing-plans | Impl plan |
| IMPLEMENT | implementing-features | Code+tests |
| COMPLETE | (final roundtable) | Report |
MCP Tools
Project: forge_project_init, forge_project_status, forge_feature_update, forge_select_skill
Iteration: forge_iteration_start, forge_iteration_advance, forge_iteration_return
Roundtable: roundtable_convene, roundtable_debate, process_roundtable_response
Skill Selection
Priority: 1.Error recovery→debugging 2.Feedback-driven→stage skill 3.Stage defaults
ITERATE Handling
forge_iteration_return → reflexion skill → forge_select_skill with feedback → re-invoke
After 3 failures: ESCALATE, report to user, continue non-blocked features.
Self-Check
Orchestrator: [ ]Subagent [ ]Deps ordered [ ]Roundtables [ ]ITERATE→reflexion [ ]Handoff before exhaustion Main chat: [ ]Spawned subagent [ ]Monitor handoff/complete [ ]Spawn successor
Features flow through stages. Artifacts face roundtable. Consensus advances. Feedback teaches. Context overflows gracefully. Successors continue mid-stride.