feat: add CodeAct plugin for code-based agent interaction

[agent-of-mkmeral]

Summary

Implements the CodeAct paradigm as a vended Plugin (P2 from the sandbox design doc). Instead of JSON tool calling, the agent writes Python code that calls tools as async functions.

No sandbox dependency — runs locally in host process (Phase 1 per design doc). Sandbox integration can be layered on later.

What's in this PR

Source (2 new files, ~570 lines)

• src/strands/vended_plugins/codeact/__init__.py — module exports
• src/strands/vended_plugins/codeact/codeact_plugin.py — full implementation

Tests (2 new files, ~740 lines)

• tests/strands/vended_plugins/codeact/__init__.py
• tests/strands/vended_plugins/codeact/test_codeact_plugin.py — 64 tests

How it works

from strands import Agent
from strands.vended_plugins.codeact import CodeActPlugin

agent = Agent(
    tools=[shell, calculator],
    plugins=[CodeActPlugin()],
)

result = agent("Calculate squares of 1-10 and sum them")

The plugin registers two hooks:

1. BeforeInvocationEvent — injects CodeAct system prompt + tool function signatures
2. AfterInvocationEvent — parses code blocks from response, executes locally, feeds results back via event.resume

Key features

• Plugin architecture — uses @hook decorators on a Plugin subclass
• Tool wrappers — generates async wrappers for each agent tool (agent.tool.X())
• Persistent namespace — variables persist across code execution turns
• AST validation — blocks exec(), eval(), compile(), dangerous __dunder__ access
• Import restrictions — configurable allowlist (defaults to safe stdlib modules)
• Max iterations — safety limit (default: 10 rounds)
• Error feedback — errors feed back as observations for self-correction
• final_answer() — clean termination mechanism

Design decisions

Decision	Rationale
Vended plugin (not core)	Follows existing pattern (vended_plugins/skills/, vended_plugins/steering/)
@hook decorators	Uses the Plugin system's auto-discovery, no manual hook registration
event.resume for loop	Native to the hook system — no custom agent loop needed
record_direct_tool_call=False	Tool calls within CodeAct code shouldn't pollute message history
asyncio.new_event_loop()	Creates fresh loop per execution to avoid event loop conflicts
Namespace __ns__ update pattern	Copies locals back to namespace for cross-turn persistence

Test results

• 64 new tests ✅
• 1828 existing tests ✅ (0 regressions)

Related

• Design doc: docs(designs): sandboxes and code execution docs#681
• Sandbox PR: feat: add Sandbox abstraction for agent code execution environments #1968
• Tracking: https://github.com/agent-of-mkmeral/strands-coder-private/issues/3

cc @mkmeral