phased-delegation

name	phased-delegation
description	Use when a task is multi-step, risky, or unclear enough that you should search for existing skills first, delegate work in bounded phases, and verify results before reporting completion.

Phased Delegation

OpenSpace is connected as an MCP server. Whether the host uses stdio, sse, or streamable-http, you have the same 4 tools available: execute_task, search_skills, fix_skill, upload_skill.

When to use

• The task is non-trivial — it spans discovery, implementation, validation, or multiple systems
• Risk is meaningful — wrong output would waste time or cause regressions
• You suspect reuse exists — there may already be a local or cloud skill worth following
• One-shot delegation feels too vague — the task needs phases, not a single umbrella prompt

Workflow

1. Frame the task before delegating

Write down:

• the concrete goal
• important constraints
• what would count as success
• what needs validation before you call it done

If the request is still ambiguous, clarify it first instead of passing that ambiguity downstream.

2. Search first

Use search_skills before reaching for execute_task.

search_skills(query="docker health checks with restart policy", source="all")

If the missing piece is current third-party library, SDK, or API behavior rather than workflow reuse, pause before execute_task and check a host-side documentation surface such as Context7. If the missing piece is broader live web evidence, recent announcements, or page content, pause and use a host-side web research surface such as Exa. Bring the confirmed details back into the next phase instead of delegating documentation ambiguity.

Check whether the result suggests:

• Follow it yourself — the skill is clear and you already have the needed capability
• Delegate a bounded phase — OpenSpace should perform one concrete slice
• Delegate the whole task — only when the task is already well-scoped and success criteria are clear

Do not skip this step just because delegation is available.

3. Delegate in bounded phases when the task is risky

For large or uncertain tasks, prefer separate calls such as:

• discovery or planning
• implementation
• verification

Example:

execute_task(
  task="Inspect the repo's existing Docker health-check patterns and propose the smallest change set to add health checks for service X. Do not edit files.",
  search_scope="all",
  max_iterations=12
)

Then follow with a narrower execution call only after you understand the result.

Use a single end-to-end execute_task only when the task is already well bounded.

4. Verify before you report success

After every execute_task result, check whether OpenSpace returned concrete proof:

• changed files or artifacts
• commands run
• validations performed
• unresolved risks or failures

If the result is missing proof, do one of these before reporting completion:

• ask OpenSpace for a focused verification pass
• run your own validation with the host's tools
• tell the user the work is incomplete

Do not translate "task completed" into "done" unless validation exists.

5. Repair or upload deliberately

Use fix_skill when a specific skill is broken and the failure is local enough to describe precisely.

fix_skill(
  skill_dir="/path/to/skills/example-skill",
  direction="The endpoint now requires bearer auth and returns a different status field. Update the skill instructions and examples."
)

If OpenSpace evolves or fixes a skill, decide whether to upload it only after checking whether the improvement is reusable or project-specific.

Decision Guide

Need help from OpenSpace?
├── Existing skill is clear and I can follow it myself
│   → Follow the skill directly
├── Task is large or risky
│   → Search first, then delegate in phases
├── Task is well-scoped and success criteria are clear
│   → Delegate with one execute_task call
└── Skill or workflow is broken
    → Use fix_skill with a precise direction

Notes

• search_skills is the default first step for complex work.
• Recommended order for external-library or live-web work: local repo evidence, search_skills, Context7 if current library or API docs are missing, Exa if broader live web evidence is missing, then execute_task.
• execute_task is strongest when you pass a bounded objective with clear validation expectations.
• Tell the user what OpenSpace actually did, what it proved, and what still needs attention.