Menu
Atulya Cortex (TUI/WhatsApp), Hand tools, internet stack (SearXNG/Firecrawl), config, and what-if troubleshooting. Use for cortex wiring, web_search/web_extract, deliberation limits, or docker internet-search compose.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Baseado na classificação ocupacional SOC
Disciplined development workflow for AI coding agents working in the Atulya monorepo — covers context gathering, branch hygiene, edit/lint/test cycle, API surface changes that require client regen, parallel vs sequential tool use, and the report-back template. Use at the start of any non-trivial Atulya coding task, or when the user asks how to "implement X", "port Y", "change the API", or "ship a feature" in this repo.
Run the Atulya repo through end-to-end system testing — toolchain check, lint, Python pytest, Rust cargo check, TypeScript build, control-plane typecheck, OpenAPI + client regen as a contract test, cross-client field smoke test, and legacy-leak audit. Use when validating a change before commit/PR, after pulling main, before a release, or when the user asks to "run all tests", "verify everything", "system sanity check", or "end-to-end test".
Design a meaningful git commit for the Atulya repo end-to-end — what commands to run to inspect the change, how to decide if it's one commit or several, how to write a message in this repo's `[type] ~ subject` convention with a structured body grouped by surface, and the exact `git add` / heredoc `git commit` invocations to use. Use whenever the user says "commit this", "git commit", "draft a commit message", "what should the commit be", or you've just finished a feature/fix and want to propose a commit before they ask.
Complete Atulya documentation for AI agents. Use this to learn about Atulya architecture, APIs, configuration, and best practices.
Review changed code against project standards. Checks for missing tests, dead code, type safety, lint issues, and coding conventions. Run after completing any implementation work.
Execute clean Atulya releases with preflight validation, version bumping, SDK regeneration, docs deployment verification, and recovery from failed release/doc workflows. Use when shipping patch/minor releases, tagging versions, or debugging why release/docs did not publish.
| name | atulya-cortex |
| description | Atulya Cortex (TUI/WhatsApp), Hand tools, internet stack (SearXNG/Firecrawl), config, and what-if troubleshooting. Use for cortex wiring, web_search/web_extract, deliberation limits, or docker internet-search compose. |
From repo root (compose file path relative to atulya/):
docker ps --filter name=atulya-searxng --filter name=atulya-firecrawl-api
curl -sS -o /dev/null -w "searxng %{http_code}\n" http://127.0.0.1:18080/
curl -sS -o /dev/null -w "firecrawl %{http_code}\n" http://127.0.0.1:3002/v0/health/liveness
Expect 200 on both when the internet-search compose is up.
uvRepo root (atulya/ workspace):
cd atulya-cortex
uv sync
uv run atulya-cortex
# same: uv run atulya-cortex chat
From monorepo root without cd:
uv run --directory atulya-cortex atulya-cortex
Dumb terminal / pipes: uv run atulya-cortex chat --simple.
tui)Edit cortex home config.toml (default ~/.atulya/cortex/ unless ATULYA_CORTEX_HOME / --home):
[tools]
enabled = true
allowed_channels = ["tui"] # add "whatsapp" only if you trust that surface
max_actions = 4 # search + extract + headroom for answer
internet_search_enabled = true
internet_extract_enabled = true
# Optional overrides (else env defaults):
# internet_searxng_base_url = "http://127.0.0.1:18080"
# internet_firecrawl_base_url = "http://127.0.0.1:3002"
# internet_firecrawl_api_key = "11111111-1111-4111-8111-111111111111"
Env (same names as atulya-api): ATULYA_API_CORTEX_SEARXNG_BASE_URL, ATULYA_API_CORTEX_FIRECRAWL_API_URL, ATULYA_API_CORTEX_FIRECRAWL_API_KEY. Put secrets in <home>/.env, not committed TOML.
Bring up stack: docker compose -f docker/docker-compose/internet-search/docker-compose.yaml up -d (from atulya/).
One tag per turn, JSON body. Examples:
<tool name="web_search">{"query": "IANA reserved IPv4 documentation prefix"}</tool>
Then (after digest):
<tool name="web_extract">{"url": "https://www.iana.org/help/example-domains"}</tool>
Prefer web_search → web_extract; web_fetch only for tiny static HTML (raw markup burns context).
| Symptom | Check |
|---|---|
Tool unknown / no <tool_result> | [tools].enabled, max_actions, model actually emitted valid JSON inside the tag |
web_search errors | SearXNG up, JSON format allowed in searxng/settings.yml in compose, URL env |
web_extract errors | Firecrawl liveness 200, Bearer UUID when self-hosted |
| No tools in prompt | allowed_channels must include channel root (tui, whatsapp, …) |
atulya_api import warning | atulya-cortex needs atulya-api installed (workspace uv sync) |
Small models often bash curl or web_fetch Google SERP HTML — useless. With internet_search_enabled, web_search is listed first in the tool catalogue, and web_fetch is rejected for common search-engine URLs with an error that points at web_search. Still set internet_search_enabled=true and run SearXNG.
web_search: digest cap ~1100 chars; tune internet_search_max_hits, internet_search_snippet_max_chars.web_extract: default markdown cap 1200 chars; per-call max_chars in tool args or internet_extract_max_chars in config.atulya-cortex/motors/fine_motor_skills.pyToolSpecs: atulya-cortex/cortex/_runtime.pyatulya-api/atulya_api/cortex/internet_connectors.pyatulya/docker/docker-compose/internet-search/