Skill2team

Design profile-based agent teams from skills, workflows, prompts, tools, or briefs.

Use Skill2Team to convert source material into a compact, accountable, profile-based agent team design and, when requested, a concrete Codex package.

Source material can be a skill zip/folder, SKILL.md, existing agents, generated team package, prompts, tools, scripts, docs, SOPs, manual workflows, or a natural-language brief.

Startup reminder

At startup, tell the user:

Skill2Team 1.9.2 defaults model execution to OpenAI Codex. Running Skill2Team, running the fixed s2t-meta-* agents, and running generated target-team agents should use Codex custom-agent/runtime invocation, not direct model API calls.

If the user explicitly chooses API model service, label it as API-service follow-up or API-run role simulation. Do not describe API runs as Codex custom-agent execution.

Context loading policy

Keep the entry skill lean. Load supporting files only when needed.

Need	Load on demand
Bare startup	references/startup-page.md, references/startup-routing.md
Design	references/design-workflow.md, references/multi-agent-architecture-best-practices.md, references/workflow-aligned-orchestration.md, references/agent-architecture-and-workflow-method.md, references/orchestration-design.md, references/flow-control-and-resume.md, references/design-package-conformance-and-reexecution.md, references/output-contracts.md, references/local-resource-allocation.md
Package	references/package-workflow.md, references/package-to-register-readiness.md, references/agent-registration-and-entrypoints.md, references/runtime-invocation-and-prompt-rewrite.md, references/target-team-execution-guard.md, references/team-usage-guide.md, references/design-package-conformance-and-reexecution.md, references/local-resource-allocation.md, assets/prompt-templates/
Meta-team-first	data/meta_team_contract.json, references/meta-team-execution.md, references/design-package-conformance-and-reexecution.md
ClawHub / Skill Creator / MIT-0	references/skill-creator-packaging.md, references/mit0-openclaw-clawhub-compliance.md, references/clawhub-publish-checklist.md
Deterministic helpers	only the specific script in scripts/ needed for the requested operation

For the detailed policy, read references/context-loading-policy.md only when context-loading or packaging compliance is in scope.

Do not preload all references. Do not copy long reference text into the answer unless the user asks for it.

Routes

Route	Use when
source-to-team	Convert or restructure existing source material into a cleaner target team. Old-team-to-new-team restructuring belongs here.
brief-to-team	Turn a natural-language brief, rough prompt, or SOP summary into a first team design.
guided-to-team	Ask focused questions before design/package when the source or risks are unclear.

Delivery modes

Skill2Team exposes only:

text

design -> package

Delivery	Meaning
design	Inventory, diagnosis, Agent Architecture Map, Workflow Orchestration Map, Control-Flow & Resume Contract, profile-based agent roster, skill allocation, gates, migration plan, design quality gate, and design-continuation prompts.
package	A concrete Codex artifact continuation after design. Generate target-agent TOML, profiles, manifests, usage guide, design intermediate results, design-output.zip, entry-agent startup welcome page, package release gate, and package-end Codex register/start/use prompts.

package is a built-in concrete continuation after design for Codex artifact delivery. Design-continuation prompts may target Codex, API runners, Hermes, OpenClaw, or other frameworks, but package-end prompts must only explain how to use the converted target team in Codex.

validate is not a user-facing delivery. Validation logic is split into the design quality gate and the package release gate. scripts/validate_package.py is a local QA helper only.

Design rules

• Keep Agent Architecture Map separate from Workflow Orchestration Map.
• Use a framework-neutral agent architecture relationship graph by default: profile agents are nodes; delegation, handoffs, routing, state/artifact flow, guardrails, human waits, fan-out/fan-in, review gates, checkpoints, and terminal boundaries are typed edges. Do not describe the default design as tied to any one runtime framework.
• Preserve source workflow control as an explicit migration object. A target team is incomplete if the workflow map only summarizes stages and omits concrete nodes, edges, stage mappings, stage-internal deliverables, required user input nodes, human intervention points, gates, checkpoints, resume rules, or terminal boundaries.
• Stage-internal duties are not optional. If a source stage generates named artifacts, candidate sets, prompt packages, registries, matrices, checkpoints, or closing handoff prompts before the next public step, the target workflow must preserve those deliverables or record a deliberate, reviewed rewrite.
• At the start of source conversion, ask whether to preserve the source workflow's original human-interaction steps, preserve only selected steps, or run fully automated with audit. The default is preserve_source_human_interaction_steps.
• Before finalizing a generated target team, inventory human intervention points and ask the user which ones to preserve, convert to reviewer gates, auto-advance with audit, or remove as redundant. If the user does not explicitly choose automation, preserve all source-mandated user input nodes, human waits, approvals, no-auto-continue boundaries, and terminal boundaries.
• Before entering the migrated workflow's first real stage, collect source-required startup settings, user choices, runtime settings, route/delivery constraints, and unresolved human-intervention retention choices. If a required setting is absent and the source does not name a safe default, ask the user or block.
• For package, generate an entry-agent startup welcome page (entry-agent-startup-welcome.json and docs/entry-agent-startup-welcome.md) adapted from the source skill's startup behavior into target-team entry-agent language. The entry agent must render or summarize it before the first real migrated workflow stage.
• Extract source_derived_runtime_constraints from the source workflow and materialize them in design maps, package contracts, agent profiles, generated runtime instructions, usage guidance, and manifests.
• Verify design/package conformance before package handoff. If a generated team does not conform to the design or package contract, set reexecution_required=true, name reexecute_from, and rerun the earliest responsible Skill2Team phase instead of patching downstream package text.
• Prefer 5-6 top-level agents for nontrivial target teams. Fewer or more requires a strong rationale.
• Do not turn every workflow step, prompt, tool, or skill into an agent.
• Include an entry agent, clear accountability boundaries, independent review where quality risk exists, and explicit state/artifact ownership.
• Stay source-agnostic. Do not hard-code identifiers, prompts, agent ids, package names, or rewrite rules from one specific source.

Fixed S2T meta-team

The internal Skill2Team service team is fixed by data/meta_team_contract.json and uses the same profile/framework-neutral convention as target-team packages.

Registered agent id	Role
s2t-meta-entry	Intake, routing, delivery selection, state tracking, synthesis
s2t-meta-source-mapper	Source inventory and original workflow extraction
s2t-meta-architecture-designer	Agent Architecture Map and role boundaries
s2t-meta-workflow-orchestrator	Workflow Orchestration Map, handoffs, checkpoint/resume
s2t-meta-runtime-adapter	Codex package/profile artifacts and prompt templates
s2t-meta-evaluation-reviewer	Independent design/package conformance, runtime-instruction, and package-quality audit

When Execution path: meta-team-first is requested with Target runtime: codex, first generate/reuse/register the fixed six-agent s2t-meta Codex meta-team. Then check the active Codex session separately: newly written .codex/config.toml entries may require a new thread or workspace reload before spawn_agent can invoke those custom agent types. If registered s2t-meta-* activation is unavailable but the active session can still spawn real independent Codex subagents, you may continue only as current_run_fanout_status=real_session_subagents, record that registered s2t-meta-* smoke tests remain pending, and route fixed-role work orders through real subagents. Fallback role-play is not allowed in Codex meta-team-first. If neither registered activation nor real session subagents can be confirmed, stop with meta-team-first blocked, explain the concrete reason, and give recovery steps; do not continue under the meta-team-first label.

Generated meta-team packages and meta-team-first target packages require independent audit by s2t-meta-evaluation-reviewer or an equivalent real independent reviewer subagent. The reviewer must inspect source extraction, design/package conformance, runtime instruction materialization, local-resource allocation, human-intervention choices, and package integrity. If the audit blocks, rerun the producing meta role from the earliest failed phase before package handoff.

Required outputs

For design, include the design result plus continuation prompts for package, Codex/OpenAI, framework-neutral API-service runners, Hermes profile conversion, OpenClaw profile conversion, and other explicitly selected frameworks.

For package, include generated target-team agents and functions, profiles, manifests, design intermediate results, design-output.zip, design-output-manifest.json, entry-agent-startup-welcome.json, docs/entry-agent-startup-welcome.md, local-resource-allocation.map.json, source-resource-manifest.json, workflow preservation gate, design-package-conformance.contract.json, runtime-instruction-conformance.json, meta-team-audit.contract.json, package release gate, target-team execution guard, and package-end Codex registration/start/use prompt templates. The local resource allocation map and source-resource manifest are required for source packages with bundled local files.

Every design reply must end with paste-ready follow-up prompts for legal next actions or further analysis, including package generation, further design/resource-gate analysis, Codex/OpenAI registration guidance, API-runner construction, API-run role simulation, Hermes profile conversion, and OpenClaw profile conversion prompts. Every package reply must end only with Codex package-use prompts: artifact inspection, package release/resource-gate analysis, Codex registration/use after smoke tests, and current-session target-team fan-out when registered target-agent types are unavailable. These prompts are not additional Skill2Team delivery modes.

Direct-skill and meta-team-first must produce target-team packages with the same artifact contract, design archive, conformance gates, Codex package-end prompts, and target-team execution guard. The only allowed difference is the recorded Skill2Team run fan-out status and meta-team activation evidence.

Generated target-team execution in Codex must not degrade into sequential or single-agent simulation. A registered target-team entry agent should hand off to real registered specialist agents. If newly registered target-agent types are unavailable in the active thread but real independent current-session Codex subagents can run, the target run may continue only with target_run_fanout_status=real_session_target_subagents, registered_target_team_smoke_status=pending_hot_reload_or_new_thread, and explicit entry/specialist/reviewer boundaries. If neither registered handoff nor real current-session target subagents are available, it must stop with target-team execution blocked, a concrete reason, and recovery steps.

Use the user's language unless they request otherwise. Keep file names, ids, and placeholders unchanged.

Helper scripts

Use deterministic helpers only when useful for the requested task:

bash

python scripts/ensure_codex_meta_team.py --codex-root <CODEX_PROJECT_ROOT> --register
python scripts/generate_deployment_package.py examples/sample_team_plan.json /tmp/s2t_deploy --target codex --delivery package
python scripts/validate_package.py .

Helper scripts do not create additional Skill2Team delivery modes.