docs(conformance): record Microsoft Agent Framework cross-validation report (33/47)

[MohammadHaroonAbuomar]

What

Adds a registry entry and reproducible claim artifacts (conformance/claims/maf/: per-part REPORT.md, harness.py, test_conformance.py) for the AGENT-HOOKS-0.1 middleware merged upstream as microsoft/agent-framework#7515.

• Adapter: agent-framework-core 1.13.0, upstream main @ 4b1afd90520310547cb0e9cdc70f644d80161e82
• SDK/CTK: agent-hooks-sdk[ctk] 0.1.0a4 (PyPI), Python 3.12, pytest plugin runner
• Corpus: the 47 wheel-vendored vectors, byte-identical to this repo's conformance/vectors/ at the PR base

Declared surface

All eight interception points; model_calls, tool_calls, int64_json, bigint_json; all four §7.2 profiles and knobs (factory composition= passthrough, default sequential/first_deny, on_approval: stop); jcs-sha256 (+ null, vector-scoped); enforce and evaluate_only; buffered_output: true (streaming fully buffered and gated); hosted (service-side) tools disclosed as uninterceptable at the tool seam (surfaced in the post_model_call projection).

Result — recorded as partial cross-validation, not a §13.1 conformance claim

33 passed, 14 failed, 0 skipped. Every failure is root-caused in the report:

• F1 (13 vectors, sole failing assertion run_outcome blocked-vs-completed): MAF deliberately halts the run on a host_error:* deny at the tool seam (documented + unit-tested upstream). §6.2's "unless the host's own semantics terminate the turn" permits this posture; the CTK's single-valued run_outcome cannot express it. CTK expressibility gap — issue material for this repo. Secondary MAF-side question: the halt keys on the host_error: prefix, so it also fires for composition-produced denies (transform_conflict, composition_disagreement).
• F2 (AH-CTK-100): the vector pins the reference host's blocked-tool transcript (messages[1].role == "tool", content string blocked: <reason>); MAF keeps the protocol-required assistant function-call message and uses a structured error payload. §6.2 prescribes no transcript shape. CTK over-prescription — issue material for this repo.
• F3 (observation, no vector failure): constructor-registered tools do not surface in agent_startup.tools_registered (Agent has no .tools attribute; they live in default_options). MAF-side fix material.

Production-path confirmation (per CLAIMS.md)

The harness drives a real Agent through create_agent_hooks_middleware (per-run session; create_agent_hooks_middleware_from_emitter only for the zero-interceptor and redact_for_approval vectors, with harness-owned session brackets in the documented host role). Only the chat client and tools are mocked; every emission, write-back, and block is the framework's. One disclosed host policy: an inner (inside-the-bundle) presentation middleware makes the caller-visible output the final assistant message before the output point guards it.

Reproduce:

pip install "agent-framework-core[agent-hooks] @ git+https://github.com/microsoft/agent-framework.git@4b1afd90520310547cb0e9cdc70f644d80161e82#subdirectory=python/packages/core" "agent-hooks-sdk[ctk]==0.1.0a4" pytest
cd conformance/claims/maf && pytest test_conformance.py --agent-hooks-harness=harness:AgentFrameworkHarness