agentsclimarketplace

Orchestrate

Skill gabros20/orchestrate-skill/skills/orchestrate

Agent skill for bounded multi-agent strategy selection, dispatch, durable state, review gates, and recovery.

Install
npx -y skills add gabros20/orchestrate-skill --skill orchestrate

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • 26 days oldThe repository was created 26 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Coordinate multi-agent work by selecting and running a bounded orchestration strategy. Use for implementation plans or multi-step tasks needing staged reviews, parallel worktrees, hierarchical fleets, agent teams, workflows, goal loops, advisor/executor splits, adversarial planning, or external coding CLIs. Triggers include "orchestrate this", "run this plan with subagents", "fan out agents", "goal loop", and "agent swarm". Do not use for small coupled tasks one agent can complete efficiently or for choosing a digital-product lifecycle; orchestrate only coordinates agents and execution.

SKILL.md

10.4 KB, ~2.2k tokens by cl100k_base, as published. Nobody here has run it

Orchestrate

Mission and boundary

Coordinate, dispatch, review, and gate multi-agent work through one of nine strategy presets. The invoking agent is the controller: it resolves the execution topology and maintains durable state; subagents perform the implementation or investigation.

orchestrate owns how work is assigned across agents. It does not decide which product-lifecycle domains are required, replace a domain skill, or own product progress. A lifecycle composer such as digital-product may request orchestration, but neither capability depends on the other.

Invocation grammar

Use the explicit form supported by the client: $orchestrate in Codex, /orchestrate in slash-command clients, or the host's equivalent. Documentation uses /orchestrate as shorthand.

<plan-file | task description>
  [strategy=auto|staged|parallel|hierarchical|team|workflow|loop|advisor|adversarial|xcli]
  [review=dual|spec|quality|panel:N|consensus:N|off]
  [engine=claude|codex|grok|cursor|agy|opencode|hermes|kimi|mixed]
  [models=orchestrator:<tier>,worker:<tier>,advisor:<tier>,reviewer:<tier>]
  [isolation=worktree|branch|off]
  [trigger=once|goal:"<stop condition>"|interval:<t>|schedule:"<cron>"]
  [workers=N] [budget=<cycles|agents|tokens>] [alias=<saved-preset>]

Selection priority: explicit strategy= > alias= from config.yaml > auto-triage > ask. Ask only when triage signals genuinely conflict.

Route before acting

  1. Reject orchestration when the coordination cost exceeds the expected benefit.
  2. Resolve the host and its available primitives before choosing a strategy.
  3. Select one strategy and record all resolved dimensions in .orchestrate/run.md.
  4. Read the selected strategy plus only the supporting routes its roles and gates require.
  5. Never preload the entire reference library.

Strategy route

StrategyUse whenReadContribution
staged (default)Mostly independent plan tasks need fresh implementation and ordered dual reviewStagedSequential task cycle, fix-wave rules, and merge gate
parallelTasks share no files and can run concurrently in isolated worktreesParallelPartition, fan-out, integration, and overload behavior
hierarchicalWork is too broad for one context and needs domain sub-orchestratorsHierarchicalDomain decomposition and delegated controller hierarchy
teamWorkers must message, debate, or coordinate across layersTeamShared-context team protocol and host degradation
workflowLarge deterministic fan-out should be held by a script rather than model memoryWorkflowPilot, budget, batch, verify, and aggregation procedure
loopWork repeats until a verifiable goal or scheduled condition is metLoopBounded cycle, stop condition, kill switch, and evolution pass
advisorExpensive reasoning should be separated from cheaper executionAdvisorConsultation cadence and advisor/executor contract
adversarialA high-stakes plan deserves independent challenge before executionAdversarialDebate roles, synthesis, and hardened plan
xcliExternal coding CLIs act as workers, peers, or second opinionsExternal CLIsVerified CLI invocation, isolation, monitoring, and result capture

Strategies compose through dimension overrides: strategy=staged engine=codex, strategy=loop topology=parallel, or strategy=parallel review=panel:3.

Shared route

ConditionReadContribution
Any non-reference host or uncertain primitiveHostsHost detection, primitive binding, and honest degradation
Every dispatched taskContractsBrief, status, report, findings, and workspace schemas
Any review-enabled runReview gatesOrdered spec/quality gates and panel behavior
Any role or engine selectionModel routingExplicit model tiers, cost posture, and drift verification
More than one writerIsolationWorktree/branch rules and integration ownership
Background, long-running, or external workMonitoringPolling, liveness, timeout, and recovery rules
Every runSafety railsMain-branch, overload, loop, budget, and reward-hacking guards
Resume, compaction, or controller transferHandoffDurable state and clean controller handoff
Every dispatch and returned reportToken economyRole-scoped communication blocks and priming anatomy

Prompt route

RoleReadContribution
Implementation workerImplementerExecutable task brief and dense report contract
Spec reviewerSpec reviewerRequirement-compliance review prompt
Quality reviewerQuality reviewerMaintainability and correctness review prompt
Integration ownerIntegratorCross-worktree merge and conflict-resolution prompt
Domain sub-orchestratorSub-orchestratorDelegated planning, dispatch, and summary contract
AdvisorAdvisor promptBounded consultation prompt
Triage assessorTriage assessorIndependent strategy-assessment prompt
Verification-only workerVerifierMinimal objective verification prompt
Adversarial plannerPlanner debateIndependent proposal and challenge prompt
Loop evolution passEvolvePeriodic pattern extraction and process-improvement prompt

Dimensions

DimensionValuesDefault
topologysolo · staged · parallel · hierarchical · team · workflow · loopselected strategy
planningnone · plan-first · interview · adversarialplan-first
reviewoff · spec · quality · dual · panel:N · consensus:Ndual
engineclaude · codex · grok · cursor · agy · opencode · hermes · kimi · mixedhost-appropriate
modelsadvisor · orchestrator · reasoner · worker · reviewer · peer tier mapmodel routing
isolationnone · worktree · branchworktree for multiple writers
triggeronce · goal · interval · scheduleonce
budgetmax cycles · agents · tokens · open PRsselected strategy

Universal rules

  1. The controller coordinates; subagents work. Do not silently implement a failed task in the controller. Fix the brief, model, context, or worker and re-dispatch deliberately.
  2. Artifacts on disk are the interface. Keep briefs, reports, findings, raw logs, and ledgers in .orchestrate/; do not rely on chat memory.
  3. Pin a model on every dispatch. An omitted model may inherit an unintended expensive default.
  4. Enforce typed gates. When dual review is enabled, spec review precedes quality review.
  5. Ledger before memory. Append progress after every gated unit; on resume, trust the ledger and repository state over recollection.
  6. Keep safety rails active. Never begin on main/master without consent; cap loops and open work; on overload resume or nudge the existing agent rather than spawn a duplicate.
  7. Prime with pointers, work silent, report dense. Preserve required safety and coordination messages while eliminating routine narration and raw output dumps.
  8. Do not orchestrate small coupled work. If one agent can complete the task efficiently within one coherent context, use solo execution and optionally one reviewer.

Core workflow

  1. Inspect the task, plan, repository state, host capabilities, and stop condition.
  2. Resolve strategy, dimensions, roles, models, budget, isolation, review, and degradation.
  3. Initialize .orchestrate/ with workspace; record the resolved run.
  4. Create task briefs with task-brief and validate them with brief-check.
  5. Dispatch only ready work. Monitor without duplicating agents and integrate through the selected strategy's owner.
  6. Package review evidence with review-package, enforce configured gates, and send failures back to the correct worker or owner.
  7. Append durable progress and finish or hand off only when the stop condition is verified.

Use toolbox to inventory available tools once and reuse the recorded result.

Artifact contract

Every run owns .orchestrate/run.md, progress.md, task briefs, worker reports, review findings, and any raw evidence required to reproduce a gate. Record resolved dimensions, exact role/model assignments, budgets, branch/worktree ownership, decisions, failures, and the verified stop condition.

Completion and handoff

Complete only when every in-scope task has a terminal status, required reviews passed, integration is verified, temporary worktrees or processes are accounted for, and the goal or one-shot stop condition is satisfied. On interruption or controller transfer, emit the durable handoff defined in shared handoff; never force the next controller to reconstruct state from chat.

What ships with it: 36 files

136.3 KB alongside SKILL.md, 5 of them executable

agents/

scripts/

Keep looking

Skills are one crate of 326,984. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.