Astromesh Leia
Astromesh Leia is a Claude Code plugin. You describe the
agent you need, in English or Spanish, and Leia designs a complete astromesh/v1 manifest with
a real system prompt, shows it to you, and can deploy it, test it and diagnose it on a
Nexus hub. It is named after Leia, the author’s lemon beagle: a scent hound, bred to follow a trail
until it finds what it was sent for — which is the job here too.
What Leia is
Section titled “What Leia is”- Prompts, not a program. Leia is ten slash commands, five subagents, six templates and eight
reference schemas, all Markdown and YAML. Claude Code runs them; the actual work happens
through
curl,kubectlandollamaon your machine. - An agent designer. Its architect subagent knows the agent schema up to astromesh core
0.74.1: per-role models, the real orchestration patterns,
spec.chain,spec.output_schema, every tool type that loads (includingintegration,api,mcpanda2a),prefetch,prompts.context, OKF knowledge bundles, guardrails, permissions,spec.skillsand A2A publishing, and the version each one needs. - A guide to the ecosystem.
schemas/ecosystem.mdlists the current version of every component and when to point at each library:astromesh-a2afor agents in other frameworks,astromesh-okfto check a bundle,astromesh-ragandastromesh-contextfor your own code. - A deploy-and-test loop for a Nexus hub: publish, poll, chat with the agent, run scored test scenarios, and walk a diagnostic checklist when something fails.
It is not a hosted service, a GUI or an agent framework. The agents it builds run on Nexus and the Astromesh runtime.
How it works
Section titled “How it works”flowchart LR you["You<br/>/leia …"] --> interp["leia-interpreter<br/>intent + entities"] interp --> arch["leia-architect<br/>YAML + system prompt"] interp --> op["leia-operator<br/>curl · kubectl"] interp --> test["leia-tester<br/>chat · scenarios"] interp --> doc["leia-doctor<br/>checklist"] arch -- "preview → yes" --> op op --> nexus["Nexus API"] test --> nexus doc --> nexus arch -. reads .-> kb["schemas/ · templates/"]
| Subagent | Model | Does |
|---|---|---|
leia-interpreter | Sonnet | Turns a sentence into an intent (create, deploy, status, test, diagnose…) and entities (business, channel, name, tenant, vertical), then routes |
leia-architect | Opus | Reads the bundled schemas and the closest template, checks ollama list, picks a pattern, and writes the manifest |
leia-operator | Sonnet | Calls the Nexus API with curl and the cluster with kubectl |
leia-tester | Sonnet | Proxies a chat to the agent, or runs five scenarios and scores them |
leia-doctor | Sonnet | Walks a fixed checklist, stops at the first failure, and gives a root cause and the fix |
Compatibility
Section titled “Compatibility”| Leia | Nexus | Astromesh core knowledge |
|---|---|---|
| 0.11.x | 0.26 – 0.28.x | 0.29 – 0.74.x |
| 0.10.x | 0.26 – 0.28.x | 0.29 – 0.74.x |
| 0.9.x | 0.26.x | 0.29 – 0.71.x |
| 0.8.x | 0.26.x | 0.29 – 0.70.x |
| 0.7.x | 0.3.x | 0.29 – 0.70.x |
| 0.6.x | 0.3.x | 0.29 – 0.68.x |
| 0.5.x | 0.3.x | 0.29 – 0.38.x |
| 0.4.x | 0.3.x | 0.29 – 0.36.x |
| 0.2.x | 0.3.x | 0.18 – 0.28.9 |
From 0.8.0 the operator, tester, doctor, bootstrap and logs send the API key in X-API-Key,
and the bundled schemas/nexus-api.md describes Nexus 0.26.11: /run, /run/stream,
invocations, usage, versions, subscriptions and the /admin plane. The tester reads the 0.70
response (blocked/guardrail, propuestas, usage.by_model): an expected block passes, an
unexpected one fails and names the rule. Nexus 0.26.11 does not forward blocked/guardrail
(0.26.12 does), so against 0.26.11 the tester infers the block from the answer.
Leia 0.10 adds the A2A server of Nexus 0.27–0.28.2 to nexus-api.md: card, SendMessage
(blocking or returnImmediately), streaming, GetTask, CancelTask, SubscribeToTask,
ListTasks and the run usage part. Publishing needs Nexus 0.27+, spec.skills core 0.73+ and
type: a2a tools core 0.74+; the rest works against any 0.26.x.
Leia 0.7.x and earlier target Nexus 0.3.x; against 0.26.x design with them and publish the YAML with Cortex or the Nexus API:
curl -X POST "$NEXUS_URL/api/v1/agents" \ -H "X-API-Key: $NEXUS_API_KEY" -H "Content-Type: application/yaml" \ --data-binary @agents/my-agent.yamlWhat each release taught it
Section titled “What each release taught it”Leia’s value is what the architect knows, so each release is a knowledge sync with a core version.
| Version | Synced to | Added |
|---|---|---|
| 0.11.2 | Nexus 0.28.2, core 0.74.1 | Herald 0.14.0 in schemas/ecosystem.md: a binding with an agent can also send attachments to media_callback_url; a new row in the request table. |
| 0.11.1 | Nexus 0.28.2, core 0.74.1 | Herald 0.13.0 in schemas/ecosystem.md: new conversations per day, hour and channel. |
| 0.11.0 | Nexus 0.28.2, core 0.74.1 | A request-to-component table at the top of schemas/ecosystem.md (Spanish and English); Cortex 0.29, Forge 0.29.1, Glyph, Orbit vs. Prisma and Nebula/Centinela with their status; Spanish verbs mapped to every intent; the context eval’s level 2 marked as not run. |
| 0.10.0 | Nexus 0.28.2, core 0.74.1 | type: a2a tools (card_url, auth by env var name, policy) and the warning that the Nexus pool has no tenant env vars; spec.skills, spec.a2a.publicar and origen (an agent’s turn never confirms a write); sub-agent usage counted in the parent’s cost_limit and billed; the Nexus A2A server in nexus-api.md; an A2A check in the operator, doctor and tester; schemas/ecosystem.md. |
| 0.9.0 | Nexus 0.26.13, core 0.71.0 | Schemas and the architect at core 0.71: permissions (fnmatch over registered tool names, denied wins), semantic/episodic memory scoped to the session, real swarm handoff with max_handoffs, enforced timeout_seconds (120 s recommended); /leia logs marks invocations blocked by a guardrail. |
| 0.8.0 | Nexus 0.26.11, core 0.70.0 | Operations use X-API-Key (Bearer gave 401); nexus-api.md rewritten for 0.26.11; the tester understands blocked/guardrail, propuestas and by_model; kind: Eval, OKF bundle and RAGPipeline schemas; /leia deploy --node for standalone nodes. |
| 0.7.0 | core 0.70.0 | Guardrails are enforced in core 0.70, so the architect emits them again: the real rule types per side, message, and the cautions (no input PII redaction when tools need IDs or phones, cost_limit ≈ 4000 for the whole run, US-format PII patterns). Templates carry input max_length and cost_limit, and output PII redaction where it fits. |
| 0.6.0 | core 0.68.0 | The whole manifest again: integration, api and mcp tools with confirm/mode, prefetch, prompts.context and context_window, max_result_tokens, pipeline stages, Glyph, and spec.knowledge with OKF bundles. It also says what 0.68 accepts but never applies: spec.guardrails, permissions.allowed_actions, semantic and episodic memory, a swarm handoff, orchestration.timeout_seconds |
| 0.5.0 | core 0.38.1 | spec.chain and spec.output_schema: when to offer them, that every agent in on_complete must be deployed first, and that an older runtime ignores both silently. The doctor checks chain declarations |
| 0.4.0 | core 0.36.0 | Only builtin, agent and client tools load from YAML; which model keys each source reads; the streaming contract and usage.by_model; RAG pipeline and workflow endpoints |
| 0.3.0 | core 0.29.0 | Per-role models (default + roles), litellm, the roles each pattern asks for |
| 0.2.1 | core 0.28.9 | Moonshot / Kimi through openai_compat |
| 0.2.0 | core 0.28.5, Nexus 0.3.0 | Fixed templates that the runtime silently ignored (flat memory, limit instead of max_chars) and invalid patterns; dual auth |
Next steps
Section titled “Next steps”- Quick Start — install the plugin and design your first agent
- Command Reference — every command, step by step
- Templates & Agents — the templates, the subagents and the bundled schemas