Files
manual_slop/conductor/tracks/opencode_config_overhaul_20260310/spec.md

2.4 KiB

Track Specification: OpenCode Configuration Overhaul

Overview

Fix critical gaps in OpenCode agent configuration that cause MMA workflow failures. Remove step limits that prematurely terminate complex tracks, disable automatic context compaction that loses critical session state, raise temperature for better problem-solving, and expand thin command wrappers into full protocol documentation.

Current State Audit (as of HEAD)

Already Implemented (DO NOT re-implement)

  • OpenCode MCP integration working (opencode.json:17-25)
  • Agent persona files exist for all 4 MMA tiers (.opencode/agents/tier*.md)
  • Conductor commands exist (.opencode/commands/conductor-*.md)
  • MMA tier commands exist but are thin wrappers (.opencode/commands/mma-tier*.md)

Gaps to Fill (This Track's Scope)

  1. Step Limits: All agents have restrictive steps limits:

    • tier1: 50, tier2: 100, tier3: 20, tier4: 5
    • These terminate complex track implementations prematurely
  2. Auto-Compaction: opencode.json has compaction.auto: true which loses session context without user control

  3. Temperature Too Low:

    • tier2: 0.2, tier3: 0.1, tier4: 0.0
    • Reduces creative problem-solving for complex tracks
  4. Thin Command Wrappers: mma-tier*.md commands are 3-4 lines, lacking:

    • Pre-delegation checkpoint protocol
    • TDD phase confirmation requirements
    • Blocking protocol
    • Context management guidance

Goals

  • Remove all step limits from agent configurations
  • Disable automatic compaction, enforce manual-only via /compact
  • Raise temperatures to 0.2-0.5 range for better reasoning
  • Expand MMA tier commands with full protocol documentation

Functional Requirements

  • All 6 agent files updated with removed steps and adjusted temperature
  • opencode.json updated with compaction.auto: false, prune: false
  • All 4 MMA tier commands expanded with context, protocols, and patterns

Non-Functional Requirements

  • No functional changes to MCP tool usage or permissions
  • Maintain backward compatibility with existing workflow

Architecture Reference

  • docs/guide_mma.md - 4-tier architecture, worker lifecycle, context amnesia
  • docs/guide_meta_boundary.md - Application vs Meta-Tooling distinction

Out of Scope

  • Model tiering (using different models per tier)
  • Changes to Gemini CLI configuration
  • Changes to conductor workflow itself