La interfaz está localizada; las descripciones y documentos de las Skills se traducen automáticamente con IA a su idioma. El código, los comandos, rutas y términos técnicos se mantienen en el original, y prevalece el documento del autor.
Explorar Skills
Desarrollo · SKILL.md

python-sdk

python-sdk

Guide work on the Strands Python SDK: agent loop, tools, models, MCP, memory, sessions, hooks, interventions, plugins, sandbox, telemetry, multi-agent graphs, snapshots, checkpoints, and Python tests.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentación de la Skill

GitHub

name: python-sdk description: "Guide work on the Strands Python SDK: agent loop, tools, models, MCP, memory, sessions, hooks, interventions, plugins, sandbox, telemetry, multi-agent graphs, snapshots, checkpoints, and Python tests." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Python SDK

Use this sub-skill for Strands Python SDK work in the strands-agents package and its Python source tree.

Start here

  1. Read overview.md to confirm scope, package identity, and exclusions.
  2. Read api-reference.md for verified signatures, exports, extras, and provider surfaces.
  3. Read workflows.md for the smallest safe path for the task.
  4. Read testing-and-maintenance.md before changing code or tests.
  5. Read troubleshooting.md when the task is about recovery or diagnosis.
  6. Run python-core-check.sh for import and signature sanity; add --pytest only in a Strands Agents checkout with test dependencies installed.

Use this sub-skill for

  • Agent, invocation, event loop, concurrency, limits, structured output, checkpoints, snapshots, and direct tool calls.
  • @tool, tool schemas, ToolContext, tool loading, tool providers, executors, and MCP client integration.
  • Model providers, model config validation, streaming, token counting, context-window behavior, and provider error translation.
  • Conversation managers, context manager strategies, memory manager, sessions, storage, hooks, interventions, plugins, sandbox, and telemetry.
  • Multi-agent graph, swarm, A2A, shared result/state objects, and multi-agent hook/plugin behavior.
  • Python unit tests, focused integration-test selection, package extras, and Python-specific coding conventions.

Route elsewhere

  • TypeScript SDK implementation or Node/browser package checks: use typescript-sdk.
  • Docs-site authoring, snippets, sourceLinks, generated API docs, or navigation: use docs-site.
  • The docs-search MCP server package and its search_docs / fetch_doc tools: use mcp-server.
  • AWS test-infra or live provider credentials: treat as optional and credential-bound unless explicitly requested.

Operating rules

  • Prefer verified signatures and package metadata from the bundled references over memory.
  • Public package surfaces should be declared through __all__; heavy optional providers should remain lazy-loaded.
  • Use precise types, X | None optionals, coded # type: ignore[code], Google-style docstrings, and specific exceptions.
  • Provider boundaries translate vendor context-window and throttling errors to typed SDK exceptions and chain the original cause.
  • Hook events use shared names and before/after ordering; compare sibling TypeScript behavior before changing cross-SDK event names.
  • Keep provider credentials, network services, AWS resources, and optional backend extras explicit in the verification handoff.

Fast path

  1. Identify the owning subsystem and read the closest reference section.
  2. Inspect the target file and smallest behavior test for line-level confirmation; do not widen to unrelated provider or integration surfaces.
  3. Run the import/signature smoke helper first, then the smallest relevant pytest slice.
  4. If a task changes public API, check docs and TypeScript parity before final handoff.
  5. Record skipped provider, credential, network, or AWS-backed checks instead of implying broad integration coverage.
Skills guardadas