super-oracle AGENTS.md

Dev notes

Maintainer constraints for plugins/super-oracle/skills/super-oracle/scripts/super-oracle.sh. Keep implementation details here, not in the user-facing README.

Layout

super-oracle/
├── .agents/plugins/marketplace.json
├── .claude-plugin/marketplace.json
├── plugins/super-oracle/
│   ├── .codex-plugin/plugin.json
│   ├── .claude-plugin/plugin.json
│   └── skills/super-oracle/
│       ├── SKILL.md
│       ├── reference/
│       │   ├── briefing-template.md
│       │   ├── code-review-principles.md
│       │   └── LICENSE-Matt-Pocock.txt
│       └── scripts/super-oracle.sh
└── scripts/
    ├── auth-test.sh
    └── smoke-test.sh

Pipeline

Keep the bounded council sequence:

1. enabled reviewers run independently and in parallel;
2. when at least two reports are usable, healthy reviewers perform one parallel skeptical refinement using every initial report;
3. one provider synthesizes the best available reports.

Do not add voting, averaging, open-ended internal loops, fixed peer slots, or a research-specific report schema. Peer lists and prompt paths must be derived from the enabled/usable provider set. A failed refinement remains distinct from its initial report; do not copy the initial file into the refined filename.

The driving agent adjudicates the synthesis and owns the outer review/fix cycle. It may reject oracle findings. A degraded council cannot establish convergence.

Backends

Claude and Codex production commands are fixed. OpenCode keeps fixed command shape but accepts -O MODEL and -V VARIANT; Kimi K3/max is the default:

claude --dangerously-skip-permissions -p --effort xhigh --model fable
claude --dangerously-skip-permissions -p --effort xhigh --model opus
codex --dangerously-bypass-approvals-and-sandbox exec --enable fast_mode \
  -m gpt-5.6-sol -c 'model_reasoning_effort="xhigh"' \
  -c 'service_tier="fast"' -
opencode run --auto \
  -m fireworks-ai/accounts/fireworks/routers/kimi-k3-fast \
  --variant max --format json