From agent-orchestration
Lay out a parallel-vs-serial dispatch plan for a queue of tasks. Reads the dispatch policy in docs/POLICY.md (slot ceiling, parallel-when-independent, no recursive dispatch) and proposes phases, agent assignments, and tier rollups for cost estimation. Use before launching multi-agent work to validate that the plan respects the slot ceiling and surfaces independence assumptions explicitly.
npx claudepluginhub laran/claude-agent-orchestrationThis skill uses the workspace's default tool permissions.
Walks through a planning exercise for multi-task / multi-agent work.
Mandates invoking relevant skills via tools before any response in coding sessions. Covers access, priorities, and adaptations for Claude Code, Copilot CLI, Gemini CLI.
Share bugs, ideas, or general feedback.
Walks through a planning exercise for multi-task / multi-agent work.
/dispatch-plan with a description of the work, e.g.:
/dispatch-plan
Tasks:
- Issue #1316: implement coach prompt loader (touches PromptLoader.java + V141 migration)
- Issue #1320: rename BackfillMetrics tag (touches BackfillMetricsService + tests)
- Issue #1322: fix earnings test fixtures
- Issue #1325: docs/Backend/Coach.md update
A plan with these sections:
Tasks bucketed into ordered phases. Within a phase, tasks run in parallel; phases run serially. Independence is established by checking:
gh pr diff if a PR draft already exists)For each task, name a recommended agent + tier:
| Task | Agent | Tier | Notes |
|---|---|---|---|
| #1316 implement | feature-dev:feature-dev | sonnet | medium scope, contained |
| #1320 rename | code-simplifier | sonnet | mechanical |
| #1322 fixtures | feature-dev:feature-dev | sonnet | test author |
| #1325 docs | docs-architect | sonnet | doc-only |
Per the dispatch policy: the parent owns reviewer dispatch, not the
implementer. For each phase, the plan names the reviewer agent and the
trigger (typically: after each implementer pushes its branch). Reviewer
agents must be namespaced — feature-dev:code-reviewer or
workflow-core:code-reviewer, never bare code-reviewer.
| Phase | In-flight implementers | Reviewers queued | Within ≤3 ceiling? |
|---|---|---|---|
| 1 | 3 (#1316, #1320, #1322) | 0 | yes |
| 2 | 1 (#1325) | up to 3 from phase 1 | yes |
If agent-cost is installed, the plan can include an estimated cost
range based on the tier rollup × pricing table. Note that this is an
estimate — actual cost is reported by Claude Code's
claude_code.cost.usage metric after the fact.
Items that could invalidate the parallelism assumption:
For 1-2 simple tasks: just dispatch them. The planning overhead isn't worth it.
For tasks with a clear dependency chain (issue B needs issue A's migration first): just serialize.
This skill earns its keep when you've got 3+ tasks, want to maximize parallelism, and need to confirm independence before fanning out.
docs/POLICY.md — full dispatch policy.docs/CONVENTIONS.md — agent naming
and frontmatter.config/tiers.yml — tier mapping.