Aicrew Quick

aicrew /quick (Codex)

Use for scoped tasks that don't need the full /dev ceremony. Scout the relevant code first, then act with surgical, goal-driven changes.

Token foundation (mandatory)

/dev, /fix, and /quick all share the same 11-capability token-saving stack — only pipeline depth differs. Full reference: ~/Agents/docs/token-foundation.md. Stack: graph-first (codebase-memory-mcp), speculative Scout → verify (SCOUT schema, two-model routing), Karpathy guardrails, layered guardrails (guardrails-taxonomy.md), context-economy read policy, security-guard.py hooks, .ai/state checkpoints, /compact between phases, /handoff on tool switch, optional context-mode + token-optimizer-mcp, caveman default output. For /quick, Scout IS Phase 1 (built-in, always first); Act is Phase 2.

Default output

Caveman/lean style by default. See ~/Agents/agents/caveman.md and ~/Agents/agents/context-economy.md. /normal or /lean off restores verbose.

Source of truth

• ~/Agents/commands/quick.md
• ~/Agents/agents/karpathy-guardrails.md
• Project overrides in .ai/skills/ and repo AGENTS.md (if present)

Token foundation (mandatory — both phases)

• Graph-first: codebase-memory-mcp (search_graph → trace_path → get_code_snippet) before any Grep/Read
• Speculative context: Scout IS Phase 1 (cheap model); emit SCOUT: schema; verify before Act
• Layered guardrails: security-guard.py (input) → scope lock (SCOUT Goal:) → karpathy-guardrails (Act) → context budget
• Context economy: always on; state saved at Checkpoint B
• Two-model routing: Scout on haiku/mini; Act on sonnet/opus
• See: ~/Agents/docs/token-foundation.md

Workflow summary

1. Scout — graph-first discovery before touching anything:
   • Graph MCP (codebase-memory-mcp): list_projects → search_graph → trace_path → get_code_snippet
   • Fallback: git diff --name-only, targeted grep, slice-reads
   • Emit and verify SCOUT: schema (Goal / Status / Constraints verbatim / Files / Call chain / Next action / Tests / Risks)
2. Act — implement with Karpathy guardrails:
   • Think Before Coding: Scout verified before any edit
   • Simplicity First: fewest files, smallest change
   • Surgical Changes: diff-sized edits only, no unrelated reformats
   • Goal-Driven Execution: every edit traces to the Goal line; stop when done
3. Write/update .ai/state/AI_STATE.<tool>.<session>.md at Checkpoint B and after Act

When to use /quick vs /fix vs /dev