Super Astra
English | 中文
Route ordinary tasks through relevant official Astra guidance. The current agent model selects modules before working; users do not need to complain first. No embedding service or extra model API is required.
ordinary prompt -> hook or persistent rule -> current model selects
-> reads selected prompts/guards -> completes the taskThe prompts target Astra, the model; Codex, Hermes and OpenClaw are hosts. Check the actual selected model before calling this an Astra setup. Claude Code/Cursor adapters provide reusable rule formats, not verified Astra availability or equivalent effects on other models.
Install and initialize
In a desktop agent, ask:
Install LabKit's super-astra for my current client and project. Complete initialization, preserve existing instructions, and explain what remains to verify in this client.
The agent can run the installer for you. It checks the actual client and recommends Hook for supported Codex runtimes; rules remain available when selected. Missing hook trust is a pending step, not a reason to switch automatically. Copying files alone cannot execute initialization, so a downloader should follow SKILL.md.
First-use guidance lists the five module switches and current hook registration facts, then asks which modules to turn off, whether to check existing Skills and instruction conflicts, and whether to adjust routing. All five modules, including delegation, start enabled in a new install; reopening setup preserves existing preferences. Enabled modules are selected only when useful, and delegation does not enable a host's multi-agent feature. The audit runs only if chosen. Initialization conversation.
| Client | Implemented entrypoint |
|---|---|
| Local Codex desktop / CLI / IDE | Codex Hook with trust review; optional scoped AGENTS.md block |
| Hermes project | Existing project context file, or AGENTS.md; guide |
| OpenClaw Agent workspace | Workspace AGENTS.md and skills/; guide |
| Claude Code (additional format adapter) | Scoped CLAUDE.md block; model applicability separate |
| Cursor (additional format adapter) | Always Apply .mdc rule; model applicability separate |
| Other local file-reading agents | Explicit host-loaded rule file |
| UI-only settings / chat / cloud | Short rule export or portable manual prompt pack |
Client-specific guidance, limitations and actual-client verification.
From a LabKit checkout, preview and then apply (Python 3.10+, scripts tested on macOS/Linux):
python3 skills/super-astra/scripts/install.py --host codex --surface desktop --project /path/to/project
python3 skills/super-astra/scripts/install.py --host codex --surface desktop --project /path/to/project --applyUse --host claude-code or --host cursor for those clients. --user selects user scope instead of --project PATH; Cursor user scope exports manually instead of editing the app settings database. Unknown hosts fall back to clearly labeled manual mode. Codex on macOS/Linux defaults to Hook registration and its separate trust review; --mode rules selects the persistent-rule fallback. --skill-only means manual mode.
Select by task intent
An implementation can use initiative and testing. Independent deliverables may benefit from delegation. Writing may need prose guidance. A simple factual question can use no modules. These are semantic decisions made by the current host model, not hard-coded keyword classifications.
Rules ask the model to read current config and module descriptions each task; hooks supply a short catalog on each valid submission for an allowed model. Selected module bodies are read as needed. Rules rely on instruction adherence, hooks provide an event entrypoint, and neither guarantees correct model decisions. Manual packs include all enabled bodies and require re-export after edits.
| Module | Intended use |
|---|---|
initiative |
Complete authorized multi-step work |
instruction-following |
Resolve relevant instruction conflicts |
writing-style |
Clear writing, explanations, and summaries |
delegation |
Useful, permitted independent subtasks |
testing |
Proportional meaningful verification |
All five are available by default in new installs; availability does not mean selection or spawning. The delegation guard preserves user restrictions, tool permissions, and coordination tradeoffs.
Modular configuration
Each module has module.json (including semantic when), prompt.md (official blocks), and guard.md (separate LabKit conditions). Toggle modules in config.json. Add/remove modules without changing the engine. routing: keyword provides optional regex matching; semantic routing is the default.
Included tools
install.py: complete-folder installation followed by initialization.initialize.py: rules/hook setup, scoped removal, manual exports, and--onboardingread-only module/hook facts.onboarding.py: separate inventory component for the initialization conversation; no Skill scan or hook execution.adapters/hosts.json: independently editable host profiles.astra.py router: preview the actual semantic entrypoint.astra.py compose: output selected guidance separately from the original prompt.audit.py: read-only scan of agreed configuration paths, offered during onboarding and run only if chosen.
Setup, configuration, and removal.
Validation and provenance
Evaluation documents automated coverage and the checks needed for native host delivery and module selection. Codex hooks filter gpt-6-astra by default. Rules can deliver the guidance to other models, which must judge its applicability; this does not switch models or add tools. File adapter tests pass, but native desktop loading has not been verified across these clients.
The five prompt blocks are from OpenAI's Astra guide, checked 2026-09-05. Routing and onboarding are LabKit additions. See sources and the PolyForm Noncommercial license.