Orchestration
Multi-agent orchestration. Lane contracts, worker-model doctrine, fault tolerance, worktree isolation. Triggers: orchestrate, coordinate, agents, parallel, spawn, contract, tier 2, tier 3.
hashgraph-online
Updated Orchestration by hashgraph-online · 7e009ff
npx skillmds@latest add hashgraph-online-awesome-codex-plugins/orchestration File contents
---name: orchestrationdescription: Multi-agent orchestration. Lane contracts, worker-model doctrine, fault tolerance, worktree isolation. Triggers: orchestrate, coordinate, agents, parallel, spawn, contract, tier 2, tier 3.---<skill id="orchestration"><purpose>Orchestration is coordination, not implementation. You define contracts, agentsexecute, AgentDB is the bus. Never assume completion without reading the file.Reference on demand: skills/orchestration/reference/orchestration-research.md.</purpose><lane_contract>Every spawned lane gets ALL of these fields; a missing field is where the lane fails:1. **Deliverable**: the observable artifact, named exactly (file path, PR, report).2. **Read-first list**: the files/docs the lane must read before acting.3. **Files table**: exhaustive list of files it may touch (`constraints.files`). No two concurrent lanes may overlap. Contract JSON: `{"goal":"X","constraints":{"files":["a.sh","b.md"]},"tier":2}`4. **Known traps, restated**: gotchas relevant to this lane, inlined, not linked.5. **Verification loop with exact commands**: the literal commands the lane runs to prove its own work (test invocation, grep, curl), plus expected output.6. **Forbidden list**: what the lane must NOT do (push, touch _meta/, add deps, ...).7. **Raw-data return format**: counts, file lists, command output. Never narrative alone; a lane that returns only prose has returned nothing checkable.</lane_contract><output_integrity>Structured long-report lanes degrade silently to placeholders under load. Validateevery lane return mechanically before using it: minimum-length check on requiredsections, placeholder detection ("TBD", "...", repeated boilerplate), counts matchthe claimed work. Reject-and-retry a degraded return; never synthesize over it.Pass large input pools to lanes by FILE PATH, never as an inline slice (silenttruncation reads as full coverage). Each lane keeps a per-lane journal/checkpointso a degraded final message is not the only record of what it did.</output_integrity><single_coordinator>One coordinator per repo at a time. Before coordinating, check for a live secondsession on the same working directory (stale sessions can survive as backgrounddaemons and produce split-brain: two coordinators mutating one repo in parallel).When killing a stuck session, kill its whole process pool, not just the visible pid.</single_coordinator><worker_model_doctrine>Select model and effort from the lane's task shape and measured evidence, never fromrole prestige. Mechanical total-spec execution with deterministic checks can use thelowest setting proven adequate. Ambiguous judgment or a measured miss justifies moreeffort; protected work requires a fresh independent verifier. The lane contract recordsthe requested model and effort, whether inheritance is intentional, and the escalationcondition. A prompt that says "use your judgment" is evidence of ambiguity, not a licenseto choose a prestige model by role. Receipts keep `requested_model` and `requested_effort`separate from `observed_model` and `observed_effort`. Unsupported or unexposed values are`unavailable`, never inferred. Protected receipts require distinct `builder_identity` and`verifier_identity`; the builder never grades its own protected work. The coordinatorreproduces acceptance evidence.</worker_model_doctrine><fault_tolerance>1. RETRY transient failures with backoff, max 3. 2. CLASSIFY the failure before choosingrecovery. 3. Never silently substitute model, effort, or provider. Use an alternative onlywhen the contract names a pre-authorized fallback; record the original request, the failure,and the observed fallback identity separately. Otherwise stop and re-contract. 4. CHECKPOINTstate to AgentDB at every boundary so a respawn resumes instead of restarting.</fault_tolerance><worktree_safety>Parallel lanes (tier 2+) run in isolated git worktrees (`isolation: "worktree"`),never the main worktree; failed work is discarded by deleting the worktree. Pre-spawn:working tree clean or stashed; each lane's `constraints.files` disjoint from allactive lanes. Post-agent validation: read the lane's checkpoint, then`git diff --name-only {base}..{lane_branch}`; every changed file MUST appear in`constraints.files`, and an out-of-scope file means reject, do not merge, re-contract.Tier 1 skips worktrees (unnecessary overhead).</worktree_safety><knowledge_injection>Inject context BEFORE spawn, never let lanes discover it at runtime: build the slicewith `agentdb inject-context <agent_type>` and inline it in the prompt (surgeon getsgotchas + patterns + contract; adversary/reviewer get past failures + recent errors;researcher gets domain learnings). The orchestrator owns injection. Every agentboundary is lossy compression: structured briefing in, structured checkpoint out;never rely on conversation history across agents.</knowledge_injection><anti_patterns>Holding context in memory instead of AgentDB · assuming a lane finished withoutreading the deliverable file (receipts describe intent; files describe reality) ·parallel lanes touching shared files (N-way merge conflicts) · serial execution whenparallel is genuinely safe · retrying without new information from the failure ·autonomous loops without a budget cap (`max_budget_usd` on the contract) ·accepting a lane return without the output-integrity check (placeholder degradationis silent) · two coordinators on one repo.</anti_patterns></skill>
hashgraph-online/awesome-codex-plugins/tree/main/plugins/ariaxhan/kernel-claude/skills/orchestration commit 7e009ff268
Frequently asked questions
Run npx skillmds@latest add hashgraph-online-awesome-codex-plugins/orchestration in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.
Multi-agent orchestration. Lane contracts, worker-model doctrine, fault tolerance, worktree isolation. Triggers: orchestrate, coordinate, agents, parallel, spawn, contract, tier 2, tier 3. It is listed under AI & ML on SkillMD.
This skill has not completed SkillMD's automated safety review yet. Capability flags: makes network calls. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.
This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.
Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.
hashgraph-online (@hashgraph-online-awesome-codex-plugins) published this skill. Their other Agent Skills are listed on their SkillMD profile.