remotion_service/docs/superpowers/specs/2026-04-05-codex-team-policy-fixes-design.md

Codex Team Policy Fixes

Date: 2026-04-05

Scope

• Fix the .codex/memories path convention so the shared rule and per-agent instructions use the same agent IDs.
• Tighten the team-first wording so non-trivial repo work consults the team by default.
• Remove role skill assignments that depend on unavailable review infrastructure.

Approved Approach

Use a minimal patch:

• Standardize memory paths on the actual Codex agent IDs, which use underscores.
• Change the consultation policy from "before deep analysis" to "before any non-trivial repo task", while keeping a narrow exception for purely mechanical actions and explicit user opt-outs.
• Replace non-executable requesting-code-review entries with executable installed skills only.

Intended Changes

Memory paths

• Update .codex/agent-team.md to state that memories live under .codex/memories/<agent_id>/.
• Update every .codex/agents/*.toml file to reference underscore-based memory directories matching the agent names.
• Update .codex/memories/README.md examples to use <agent_id> wording.

Team-first policy

• Update AGENTS.md and .codex/agent-team.md to require team consultation before any non-trivial repo task.
• Keep a narrow local-only exception for purely mechanical actions that cannot materially change behavior, architecture, or risk.

Skill map

• Remove requesting-code-review from roles because the required superpowers:code-reviewer subagent is not available in this workspace.
• Keep the map limited to executable skills already installed in the current environment.

Success Criteria

• Shared policy and per-agent instructions point to the same memory paths.
• The root guidance no longer leaves "deep analysis" as the main threshold for consulting the team.
• The skill map contains only practically usable role assignments for this environment.