[AAASM-3536] ✨ (adapters): Add LlamaIndex framework adapter

[Chisanan232]

Description

Adds full LlamaIndex framework support to the Python SDK. A new
agent_assembly.adapters.llamaindex adapter installs SDK-layer pre-execution
allow/deny and audit hooks by patching the concrete LlamaIndex tool-execution
methods: FunctionTool.call (sync) and FunctionTool.acall (async). The
modern agent stack (FunctionAgent / ReActAgent via AgentWorkflow) awaits
tool.acall(...); the legacy/sync path calls tool.call(...). Both are hooked
so a denied tool's underlying function never executes. On a deny verdict the
adapter returns a ToolOutput(is_error=True) so the agent loop receives a
well-formed result instead of crashing. The adapter is registered as a builtin
in the registry and advertises get_supported_versions() == [">=0.10.0"].

Decision normalization (fail-closed under enforce, AAASM-3107) is reused from
the shared CrewAI patch rather than re-implemented.

Type of Change

• ✨ New feature

Breaking Changes

• No

Related Issues

• Related JIRA ticket: AAASM-3536 (parent Story AAASM-3535)

Testing

• Unit tests added/updated

The tests prove genuine governance, not a no-op (the AAASM-3528 lesson):

• Deny path asserts the tool function body never runs — verified against both a
  fake module and a real FunctionTool (via importorskip), sync and async.
• Allow path asserts the body runs and the result is recorded.
• A negative-control test demonstrates the deny assertions fail if the patch
  is reverted to a transparent pass-through.
• Manually confirmed: neutering the deny branch to a no-op makes both deny tests
  FAIL (tool body executes); restoring the patch makes them pass.
• llama-index-core added to the test dependency group so the real-package
  governance tests run in CI rather than skipping.
• Full suite: 593 passed, 13 skipped (environmental); mypy clean on the adapter;
  ruff + pre-commit clean on changed files.

Checklist

• Code follows project style guidelines
• Self-review completed
• Documentation updated (README + docs/compatibility/frameworks.md)
• All tests passing

Refs AAASM-3536

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 89.52381% with 22 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
agent_assembly/adapters/llamaindex/patch.py	87.91%	22 Missing ⚠️

📢 Thoughts on this report? Let us know!

[Chisanan232]

Claude Code — review result

✅ Real CI green (SonarCloud not blocking).

• Adapter: hooks the concrete llama_index.core.tools.FunctionTool.call (sync) + .acall (async) — BaseTool.call/acall are abstract, so the patch targets FunctionTool. Correctly handles LlamaIndex's @dispatcher.span instrumentation by invoking the original via the descriptor protocol and capturing it from the class __dict__. Registered, get_supported_versions() = >=0.10.0.
• Governance proven: deny verified against both a fake module and a real FunctionTool (sync + async); negative control confirmed (manually neutering the deny branch makes both deny tests fail). ✔
• Shared files: trivial keep-all overlap with siblings (distinct registry key llama_index.core).

Ready to merge. Merge before examples #152 + a python-sdk release. Full cross-repo review on AAASM-3535.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
89.5% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud