Mtg Commander

Rationale:

• Context isolation — each agent receives only its defined inputs (deck state, intake params, reference files). No bleed from prior agent reasoning.
• Adversarial independence — challengers MUST NOT share context with the primary they are reviewing. Shared context defeats the entire adversarial architecture.
• Correction loop integrity — self-correction routing depends on clean agent boundaries. Inlined steps corrupt the correction signal.

Anti-pattern (from real session 0876a59e):

"I'll run the agent roles inline to keep things moving..."

This collapsed all four agents into a single context window, destroyed adversarial independence, and produced a deck with 14 undetected color identity violations. The correction loop never fired because there were no agent boundaries to trigger it. NEVER repeat this.

All 8 dispatches (4 primary + 4 challenger) MUST be separate Agent tool invocations. No exceptions. No "just this once." No "it's a simple deck."

Configuration (.mtg-commander.yml)

At pipeline start (after intake confirmation, before the pipeline banner), check the user's working directory for .mtg-commander.yml.

If absent: Use all defaults. Output a one-line note: No .mtg-commander.yml found — using defaults. Create one to customize loop caps, price goals, and escalation.

If present: Validate keys against the schema below. Apply overrides for recognized keys. Use defaults for any missing keys. Warn on invalid or unrecognized keys (list their names) but NEVER fail the pipeline due to config. A parse-failure (invalid YAML) also warns and falls back to all defaults.

Status line (shown after intake, before banner):

• Config loaded: Config loaded from .mtg-commander.yml (version 1)
• Config absent: No .mtg-commander.yml found — using defaults.

Schema (version 1)