agentleFS
MaicololiverasMaicololiveras/praxisgenai-agent-orchestrator/skills/agent-team-orchestrator/SKILL.md

agent-team-orchestrator

Skill1 starsChanged 8 months agoMITView on GitHub

Coordinate multi-agent work in OpenCode using skills, Engram memory, and host-native delegation primitives instead of file-based mailboxes or lock files. Trigger: When work spans multiple areas, requires delegation, or the assistant must act as a director rather than a solo executor.

1 other project uses a version of this.

Adopted by

Most recently taken up by

  • Other versions1 project uses a version of this.

    Word for word · 0

    None found.

    Edited · 0

    None found.

    Wrote their own version · 1

    1. InnovationMachineTeam/skills★ 1same name

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

  • In this repository10 other files it gives its agents.

The file

Show its rules and text

We split this skill into the rules it gives an agent

20 rules. About 0 are common, 8 are distinctive, and 12 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈8

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 12

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Do NOT force heavy orchestration for trivial factual answers.

    line 21
  2. Very specific:

    Do NOT use tasks.json, mailbox folders, .lock files, or claim files as the primary coordination layer

    line 25
  3. Very specific:

    Use host-native primitives: subagents, matching skills, slash commands, Engram, and TodoWrite when useful

    line 26
  4. Distinctive:

    Do NOT parallelize overlapping paths

    line 28
  5. Very specific:

    Every delegated worker MUST return valid JSON conforming to schemas/worker-output.schema.json.

    line 73
  6. Distinctive:

    After reading sub-agent output, the orchestrator MUST reject:

    line 121
  7. Distinctive:

    If rejected, log the error and optionally retry the sub-agent.

    line 128
  8. Very specific:

    Use enter_plan_mode for structured planning before implementation

    line 148
  9. Very specific:

    Windows (PowerShell): Use Start-Job + Wait-Job with codex.cmd

    line 159
  10. Very specific:

    Linux/Mac (bash): Use & + wait with codex

    line 160
  11. Very specific:

    Always use --ephemeral (no session persistence for sub-agents)

    line 195
  12. Very specific:

    Always use --full-auto (no approval prompts)

    line 196
  13. Very specific:

    Always use -o <file> (capture output to file)

    line 197
  14. Very specific:

    Windows: Use codex.cmd (not codex.ps1), $env:TEMP for temp dir, Start-Job/Wait-Job for parallelism

    line 198
  15. Very specific:

    Linux/Mac: Use -C "$(pwd)" (same working directory), /tmp/ for temp dir, &/wait for parallelism

    line 199
  16. Very specific:

    Use praxisgenai-sub{N}.json naming convention (JSON output, not markdown)

    line 200
  17. Distinctive:

    If a sub-agent fails, read its output for error info

    line 202
  18. Distinctive:

    If you are running in an editor without native sub-agents:

    line 209
  19. Distinctive:

    ALWAYS save results to Engram after each phase

    line 212
  20. Distinctive:

    ALWAYS load prior artifacts from Engram before each phase

    line 213

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

When to Use

  • Requests that touch multiple files, modules, or architectural areas
  • Work that benefits from exploration before implementation
  • Multi-step execution that risks context bloat
  • Any SDD workflow

Do NOT force heavy orchestration for trivial factual answers.

Hard Rules

  • Do NOT use tasks.json, mailbox folders, .lock files, or claim files as the primary coordination layer
  • Use host-native primitives: subagents, matching skills, slash commands, Engram, and TodoWrite when useful
  • Assign ownership by area, path, capability, or phase
  • Do NOT parallelize overlapping paths
  • Save meaningful discoveries, bug fixes, decisions, and conventions with mem_save

Coordination Stack

  • Task when you need bounded delegated work
  • TodoWrite when tracking adds clarity
  • engram for persistent state and recovery
  • SDD skills for substantial changes
  • explore for investigation and general for implementation when those subagents are available

Scaling Rules

  • Small task: one bounded delegation or one matching skill
  • Medium task: explore -> implement -> verify
  • Large task: sdd-explore -> sdd-propose -> sdd-spec/design -> sdd-tasks -> sdd-apply -> sdd-verify -> sdd-archive

Ownership Model

Good boundaries:

  • src/auth/**
  • docs/api/**
  • verify regression coverage
  • design state persistence

Bad boundaries:

  • "whatever is free"
  • "take backend maybe"
  • lock-file claims

Memory Rules

Save important outcomes with mem_save using:

  • title
  • type
  • project
  • content with What, Why, Where, and Learned

For evolving topics, reuse topic_key.

Sub-Agent Output Contract

Every delegated worker MUST return valid JSON conforming to schemas/worker-output.schema.json.

Schema reference: schemas/worker-output.schema.json (repo) or ~/.codex/schemas/worker-output.schema.json (installed).

Enforcement by editor:

  • Codex: Uses codex exec --output-schema <path> to enforce the schema at runtime.
  • OpenCode and Gemini: Follow the same contract via prompt instruction (no native schema enforcement).

Status Values

StatusMeaning
successTask completed fully. verification.performed MUST be non-empty.
partialTask partially done. Some work completed, but not everything.
blockedCannot proceed. Explain why in risks.

Example Output

{
  "contract_version": "1.0",
  "worker_id": "sub1",
  "task_id": "explore-auth",
  "status": "success",
  "executive_summary": "Traced the full authentication flow...",
  "ownership": {
    "area": "src/auth/**",
    "touched_paths": ["src/auth/login.ts", "src/auth/session.ts"],
    "untouched_paths": ["src/payments/**"]
  },
  "artifacts": { "created": [], "updated": [], "references": ["docs/auth-flow.md"] },
  "findings": ["Sessions persist in localStorage — vulnerable to XSS"],
  "decisions": ["Focus on token lifecycle before session storage"],
  "changes": [],
  "verification": {
    "performed": ["Code reading", "Flow tracing"],
    "not_performed": ["Unit tests", "E2E flow"]
  },
  "risks": ["The interceptor is a global singleton — changes affect ALL API calls"],
  "next_recommended": ["Implement refresh token rotation"],
  "memory_saved": []
}

See schemas/worker-output.example.json for a complete example.

Orchestrator Validation Rules

After reading sub-agent output, the orchestrator MUST reject:

  • Invalid JSON (parse error)
  • Missing required fields (any field from the schema's required list)
  • Invented status (status not in ["success", "partial", "blocked"])
  • Paths outside ownership (paths in touched_paths outside declared ownership.area)
  • Success without evidence (status "success" with empty verification.performed — no evidence = no trust)

If rejected, log the error and optionally retry the sub-agent.

Delegation Model per Editor

The orchestrator adapts its behavior based on the runtime's actual capabilities:

Claude Code (native — no orchestrator needed)

  • Uses built-in Agent tool with typed sub-agents (Explore, Plan, general)
  • Full context isolation per sub-agent
  • This skill is NOT needed in Claude Code

OpenCode (full sub-agent support)

  • Uses subtask: true in command definitions to spawn isolated agent contexts
  • Each SDD phase runs as a dedicated agent with its own system prompt
  • The orchestrator delegates via slash commands: /sdd-explore, /sdd-apply, etc.
  • Results flow back through the conversation

Gemini CLI (sub-agent via tools)

  • Uses SubagentTool — sub-agents are exposed as callable tools
  • Skills can be loaded dynamically with activate_skill
  • Use enter_plan_mode for structured planning before implementation
  • The orchestrator invokes skills directly, each running in isolated context

Codex (Simulated Sub-Agents via codex exec)

  • Codex has NO native sub-agents, but can SIMULATE them using codex exec in background
  • The main Codex spawns up to 4 parallel codex exec processes
  • Each sub-process writes its result to a temp file via -o flag
  • Main Codex reads all output files and synthesizes

How to delegate in Codex:

Windows (PowerShell): Use Start-Job + Wait-Job with codex.cmd Linux/Mac (bash): Use & + wait with codex Important: codex.ps1 may be blocked by execution policy. Always use codex.cmd on Windows.

# Windows (PowerShell) — Launch up to 4 sub-agents in parallel
$schema = Join-Path $PSScriptRoot 'schemas' 'worker-output.schema.json'
# Or use the installed path:
# $schema = "$HOME\.codex\schemas\worker-output.schema.json"

$jobs = @()
$jobs += Start-Job { & 'codex.cmd' exec --full-auto --ephemeral --output-schema $using:schema -o "$env:TEMP\praxisgenai-sub1.json" 'PROMPT FOR TASK 1' }
$jobs += Start-Job { & 'codex.cmd' exec --full-auto --ephemeral --output-schema $using:schema -o "$env:TEMP\praxisgenai-sub2.json" 'PROMPT FOR TASK 2' }
$jobs | Wait-Job
$result1 = Get-Content "$env:TEMP\praxisgenai-sub1.json" | ConvertFrom-Json
$result2 = Get-Content "$env:TEMP\praxisgenai-sub2.json" | ConvertFrom-Json
if ($result1.status -notin @('success','partial','blocked')) { Write-Error "Invalid status from sub1" }
if ($result2.status -notin @('success','partial','blocked')) { Write-Error "Invalid status from sub2" }
Remove-Item "$env:TEMP\praxisgenai-sub*.json" -ErrorAction SilentlyContinue
# Linux/Mac (bash) — Launch up to 4 sub-agents in parallel
SCHEMA="$HOME/.codex/schemas/worker-output.schema.json"

codex exec --full-auto --ephemeral --output-schema "$SCHEMA" -C "$(pwd)" -o /tmp/praxisgenai-sub1.json "PROMPT FOR TASK 1" &
codex exec --full-auto --ephemeral --output-schema "$SCHEMA" -C "$(pwd)" -o /tmp/praxisgenai-sub2.json "PROMPT FOR TASK 2" &
wait
# Validate with jq
echo "$(cat /tmp/praxisgenai-sub1.json)" | jq -e '.status' > /dev/null || echo "INVALID sub1"
echo "$(cat /tmp/praxisgenai-sub2.json)" | jq -e '.status' > /dev/null || echo "INVALID sub2"
rm /tmp/praxisgenai-sub*.json

Rules:

  • Maximum 4 parallel sub-agents (to avoid saturating the machine)
  • Always use --ephemeral (no session persistence for sub-agents)
  • Always use --full-auto (no approval prompts)
  • Always use -o <file> (capture output to file)
  • Windows: Use codex.cmd (not codex.ps1), $env:TEMP for temp dir, Start-Job/Wait-Job for parallelism
  • Linux/Mac: Use -C "$(pwd)" (same working directory), /tmp/ for temp dir, &/wait for parallelism
  • Use praxisgenai-sub{N}.json naming convention (JSON output, not markdown)
  • Clean up temp files after reading
  • If a sub-agent fails, read its output for error info

For sequential phases (when parallelism doesn't help):

  • Still use codex exec but run one at a time
  • Save results to Engram between phases for cross-session continuity

Detection Rule

If you are running in an editor without native sub-agents:

  • In Codex: use codex exec to simulate sub-agents (see Codex section above)
  • In other editors without sub-agents: execute phases inline sequentially
  • ALWAYS save results to Engram after each phase
  • ALWAYS load prior artifacts from Engram before each phase

Commands

/orch-help
/orch-status
/orch-doctor
/orch-init
/skill-registry
/sdd-init
/sdd-new change-name
/sdd-apply change-name

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.