AgentFlow OS Compatibility Entry

This legacy Skill ID delegates to MakeCrew. Use it when a user wants AI to clarify an underspecified task, discover the best available Skill or method, route work to one or more Agents, preserve project memory, enforce verification, resume interrupted work, or improve from feedback. Search aliases: MakeCrew, AI Company OS, AI work operating system, multi-agent orchestration, agent routing, and Skill discovery.

The compatibility entry also uses MakeCrew's local method catalog after a task is clear. It can match source-first research, theory/case/counterexample research, transferable benchmark selection, content asset maps, two-layer pre-publication review, knowledge-base navigation/incremental auditing, design checks, and Skill trigger contracts. Each card exposes a stable ID, use signals, boundaries, deliverables, acceptance evidence, cost, and source level. Cards are ranked by task signals with explicit match reasons; the display window is bounded while the adapter reports the total external result count. See docs/method-catalog.md; method cards are optional planning aids and do not install Skills, create employees, or scan all files. Use makecrew method-audit before a catalog release; malformed host entries are skipped and counted.

Default path

Keep one clear task in the current conversation. Ask only the missing questions that change the deliverable (0-N total, 1-3 per round), then inspect installed Skill metadata and local methods. Use local matches directly. Search external candidates only for unresolved capability gaps or a requested fresh comparison; show purpose, source, license, and trade-offs before the user chooses whether to install or use one.

Post-install welcome

After installation or first activation, immediately tell the user in plain language that MakeCrew is ready and explain the three ways to use it: lightweight current-conversation mode, approved basic-team mode, and on-demand employee mode. Mention that existing conversations and project memory are kept, that a single clear task stays in the current conversation, and that missing employees are proposed with reasons before creation. Show this welcome once per workspace and do not send recurring daily notices.

Employee creation gate

A fresh install loads role templates but creates no employee conversations. Existing ordinary conversations, employees, project memory, and configuration remain untouched. When no suitable employee exists, return an employee_proposals list with:

• reason and responsibilities;
• required Skills and tools;
• memory scope;
• estimated Token/time cost;
• workspace and conversation impact.

Set status to awaiting_employee_approval. Create or bind an employee conversation only after explicit user approval. CEO-001, PM-001, and QA-001 are optional templates; a single task does not require a CEO.

Multi-task path

Use CEO/PM scheduling only when the user submits multiple tasks, asks for parallel work, or needs cross-project resource decisions. Reuse an approved employee thread for the same project; propose missing roles first, then dispatch independent tasks with explicit dependencies, budgets, and one QA summary.

On Codex, use the bundled CodexAdapter to map the parent Agent to the supervisor and native child Agents to employees. Host callbacks create/reuse threads, send compact packets, and return structured results; missing callbacks remain explicitly queued.

The CEO-001 role may rotate across host threads. At 85% context usage prepare a compact handoff; at 95% complete it before complex work. The phrase 给 CEO 换一个新对话 triggers the same flow. Use CeoContinuityManager and CodexAdapter.rollover_ceo() so the candidate verifies active project, task, decision, high-risk IDs, and the handoff digest before the active pointer moves. Never carry the full transcript into the new CEO thread.

P0 quality and recovery

The compatibility entry also exposes three optional contracts from MakeCrew: skill-audit validates Skill metadata and progressive-disclosure layout; review-and-critique supplies milestone and pre-release independent review; checkpoint-recovery persists compact, idempotent node state and bounded retry decisions. Add them only when the task needs those guarantees, keeping ordinary single-task work on the shortest path.

Verification and learning

Return the route, employee IDs, selected Skills/tools, project context, acceptance gates, budget, evidence, blockers, and next action. Do not claim a host-side action without execution evidence. Trigger self-evolution only after negative feedback, failed verification, rework, repeated issues, or an explicit retrospective; replay a candidate against a baseline before adoption.

Compatibility

The legacy Skill path skills/makecrew/SKILL.md, Skill ID makecrew, command makecrew, Python module ai_company_os, and workspace directory .makecrew remain supported.