agent-orchestration

claude-agent-orchestration

Canonical agent roster, model-tier mapping, and dispatch policy for Claude Code sub-agent workflows.

What this plugin is

The intent layer of the laran/claude-methodology family — agent-telemetry and agent-cost observe; this one decides.

Ships:

Path	Purpose
config/tiers.yml	Canonical subagent_type → haiku|sonnet|opus mapping
docs/POLICY.md	Dispatch rules (slot ceiling, parallel-when-independent, no recursive dispatch, reviewer namespacing, project-state verification)
docs/CONVENTIONS.md	Agent frontmatter format, project-level overrides, dispatch description hygiene
bin/agent-roster.sh	Resolves and prints the active mapping
skills/agent-roster	/agent-roster [name] [--format json]
skills/dispatch-plan	/dispatch-plan — lay out parallel-vs-serial dispatch for a queue of tasks
agents/orchestrator.md	Sonnet-tier meta-agent for sustained planning over many tasks

Install

/plugin marketplace add laran/claude-methodology
/plugin install agent-orchestration@claude-methodology

No backend or hooks required. This plugin is documentation + configuration + a planning skill.

Tier mapping resolution

When agent-telemetry or agent-cost reads tier info, the lookup chain is:

1. $AGENT_TIERS_FILE env var
2. ${CLAUDE_PROJECT_DIR}/.claude/agent-tiers.yml (project override)
3. ${HOME}/.claude/agent-tiers.yml (user override)
4. agent-orchestration/config/tiers.yml (canonical, this plugin)
5. plugin-bundled fallback (agent-telemetry's / agent-cost's own copy)

Editing the canonical file requires a PR against this repo; per-project or per-user overrides are file drops at the paths above.

Dispatch policy summary

See docs/POLICY.md for full rationale. Quick reference:

1. Slot ceiling — ≤3 in-flight implementers (read-only research can fan out to ~6).
2. Parallel-when-independent. Default to fanout; verify independence via gh pr diff and migration version coordination.
3. No recursive dispatch — parent owns reviewer. Implementers implement + push + STOP; the parent dispatches the reviewer.
4. Reviewer agent must be namespaced (feature-dev:code-reviewer, workflow-core:code-reviewer).
5. Comment + label, not GitHub formal review. Single-developer workflows hit GitHub's self-approval block.
6. Verify project-state claims before asserting them. 1-5 second gh / git / grep check beats memory recall.
7. Verify runtime claims via ps / lsof / stat. The Claude Code harness spawns helper processes you don't track.
8. Save validated approaches as feedback memory, not just corrections. The quieter "yes, exactly" signal matters too.

Relationship to the other two plugins

                        ┌─────────────────────────────┐
                        │  agent-orchestration        │
                        │  (intent — this plugin)     │
                        │                             │
                        │  • canonical tiers.yml      │
                        │  • dispatch policy          │
                        │  • /dispatch-plan           │
                        │  • orchestrator agent       │
                        └─────────────┬───────────────┘
                                      │
                                      │ supplies tier mapping to
                                      ▼
       ┌──────────────────────────┐         ┌──────────────────────────┐
       │  agent-telemetry         │         │  agent-cost              │
       │  (observation)           │         │  (analytics)             │
       │                          │         │                          │
       │  • SessionStart enrich   │         │  • cost-report by tier   │
       │  • outcome reconcile     │         │  • budget alerts         │
       │  • observe-trace         │         │  • cost-analyst agent    │
       └──────────────────────────┘         └──────────────────────────┘

The three plugins are independently installable. agent-orchestration adds the most value when paired with the other two — its tier mapping becomes the lens through which observation rolls up.

License

MIT.