agent-orchestrator

// Coordinate multiple OpenClaw agents as one governed system with registry discovery, hierarchy-aware planning, task delegation, and permission boundaries. Use when Codex needs to route work across agents under ~/.openclaw/agents, design or enforce manager/staff hierarchies, generate delegation workflows, or replace skill-local agent bridges with a reusable orchestrator.

name	agent-orchestrator
description	Coordinate multiple OpenClaw agents as one governed system with registry discovery, hierarchy-aware planning, task delegation, and permission boundaries. Use when Codex needs to route work across agents under ~/.openclaw/agents, design or enforce manager/staff hierarchies, generate delegation workflows, or replace skill-local agent bridges with a reusable orchestrator.

Agent Orchestrator

Build or operate a reusable orchestration layer above OpenClaw agents. Prefer this skill over ad-hoc bridges when the system must discover agents from ~/.openclaw/agents, enforce delegation boundaries, plan multi-step workflows, and keep transport/runtime concerns separate from business skills.

Workflow

Discover runtime agents from OPENCLAW_HOME or the default ~/.openclaw.
Load manifest overrides from the orchestrator manifests/ directory or an external manifest directory.
Merge runtime metadata and manifest policy into one registry.
Create a plan:
- direct: send to one agent
- auto: select the best allowed child
- hierarchy: walk the org tree for review/delegation chains
Execute through the transport layer using each agent's real session key.
Keep wrappers thin. Feature skills should call the orchestrator, not reimplement routing.
When invoking the CLI from an agent lane on Windows, prefer --message-file with a UTF-8 brief file instead of passing raw Vietnamese text directly in shell arguments.

Fixed Department Flow (Facebook Campaign)

When user commands truong_phong directly, hierarchy mode should behave like this:

If user explicitly asks for a detailed plan first, truong_phong returns a proposal for approval.
If user asks to "trien khai bai viet", "viet bai", or "soan bai" for a product, truong_phong must execute immediately instead of asking whether to draft or publish.
The default scope for the case above is content-only so user can review the draft first.
pho_phong triggers mandatory product research via search_product_text.
pho_phong assigns nv_content to write content.
pho_phong reviews content.
Only if the brief explicitly asks for image/video/media does pho_phong open nv_media, review media, and compile the package.
nv_media uses product reference images from artifacts/references/search_product_text/<product-slug>/ when calling gemini_generate_image and generate_video.
truong_phong performs the real final_review.
Publishing with facebook_publish_post happens only after explicit user confirmation to publish.

Detailed-plan approval gate is applied only when user explicitly requests planning before execution.

CLI Notes

--product-keyword: keyword passed to skills/search_product_text/action.js.
--target-site: target domain for product research (default uptek.vn).
--artifacts-dir: custom output folder for simulation artifacts.
--no-simulation-artifacts: disable artifact writing.
--strict-review-gates: enforce strict blocking review behavior.

Simulation output default folder:

artifacts/campaigns/agent-orchestrator-simulations/<run-id>/
Includes plan/result, per-step snapshots, final content, image/video prompts, and copied product original images.
Includes publish result artifacts when compile/publish runs successfully.

Rules

Treat agents/* as the runtime source of truth for which agents exist.
Treat manifests as the source of truth for hierarchy, permissions, and capabilities.
Do not hardcode session keys inside business skills.
Keep transport generic so the system can later move from chat-completions relay to session tools or other runtimes.
Prefer explicit reportsTo, canDelegateTo, and capabilities over keyword-only routing.

Resources

Scripts: use scripts/orchestrator.js as the CLI entrypoint.
References: read references/architecture.md, references/manifest-schema.md, and references/openclaw-config-template.json5 when editing the hierarchy model or gateway policy.
Manifests: keep default agent hierarchy in manifests/*.json; allow external overrides through env/config.

Agent Orchestrator

Workflow

Discover runtime agents from OPENCLAW_HOME or the default ~/.openclaw.

Load manifest overrides from the orchestrator manifests/ directory or an external manifest directory.

Merge runtime metadata and manifest policy into one registry.

Create a plan:

direct: send to one agent
auto: select the best allowed child
hierarchy: walk the org tree for review/delegation chains

Execute through the transport layer using each agent's real session key.

Keep wrappers thin. Feature skills should call the orchestrator, not reimplement routing.

When invoking the CLI from an agent lane on Windows, prefer --message-file with a UTF-8 brief file instead of passing raw Vietnamese text directly in shell arguments.

Fixed Department Flow (Facebook Campaign)

When user commands truong_phong directly, hierarchy mode should behave like this:

If user explicitly asks for a detailed plan first, truong_phong returns a proposal for approval.

If user asks to "trien khai bai viet", "viet bai", or "soan bai" for a product, truong_phong must execute immediately instead of asking whether to draft or publish.

The default scope for the case above is content-only so user can review the draft first.

pho_phong triggers mandatory product research via search_product_text.

pho_phong assigns nv_content to write content.

pho_phong reviews content.

Only if the brief explicitly asks for image/video/media does pho_phong open nv_media, review media, and compile the package.

nv_media uses product reference images from artifacts/references/search_product_text/<product-slug>/ when calling gemini_generate_image and generate_video.

truong_phong performs the real final_review.

Publishing with facebook_publish_post happens only after explicit user confirmation to publish.

Detailed-plan approval gate is applied only when user explicitly requests planning before execution.

CLI Notes

--product-keyword: keyword passed to skills/search_product_text/action.js.

--target-site: target domain for product research (default uptek.vn).

--artifacts-dir: custom output folder for simulation artifacts.

--no-simulation-artifacts: disable artifact writing.

--strict-review-gates: enforce strict blocking review behavior.

Simulation output default folder:

artifacts/campaigns/agent-orchestrator-simulations/<run-id>/

Includes plan/result, per-step snapshots, final content, image/video prompts, and copied product original images.

Includes publish result artifacts when compile/publish runs successfully.

Rules

Treat agents/* as the runtime source of truth for which agents exist.

Treat manifests as the source of truth for hierarchy, permissions, and capabilities.

Do not hardcode session keys inside business skills.

Keep transport generic so the system can later move from chat-completions relay to session tools or other runtimes.

Prefer explicit reportsTo, canDelegateTo, and capabilities over keyword-only routing.

Resources

Scripts: use scripts/orchestrator.js as the CLI entrypoint.

References: read references/architecture.md, references/manifest-schema.md, and references/openclaw-config-template.json5 when editing the hierarchy model or gateway policy.

Manifests: keep default agent hierarchy in manifests/*.json; allow external overrides through env/config.