[AAASM-6204] 🐛 (examples): Declare the dry-run posture the offline Python demos depend on

[Chisanan232]

What changed

Two things, in the order they have to happen.

1. metadata/sdk-versions.yaml advances the python SDK pin from 0.0.1rc6 to 0.0.1rc7 — the current published release. The generator rewrites the 40 files that derive from it (18 manifests, 20 bounded README blocks, the live scenario's Dockerfile and verify-live.yml), and all 18 lockfiles are relocked with a single-package upgrade.
2. All 16 Python examples declare the dry-run posture they actually depend on, in src/main.py and in tests/test_smoke.py.

Step 2 is not optional cleanup. Without it the pin advance turns 18 red PRs into 16 red example lanes.

Why

Every Python example calls init_assembly in sdk-only mode with no enforcement_mode, so it inherits the advertised default. On rc.6 that default silently failed open when registration could not reach a gateway, which is the only reason the offline demos ran at all.

rc.7 closed that hole (AAASM-4130). _local_posture_is_enforce now treats a None posture exactly like enforce for the SDK's own error posture, and only the explicit dry-run postures fail open locally:

def _local_posture_is_enforce(enforcement_mode: str | None) -> bool:
    """Whether the SDK's *local* failure posture must fail closed (AAASM-4130).

    ``enforcement_mode is None`` is the advertised default: it registers the agent
    under the gateway's server-side default, which is live ``enforce``. So for the
    SDK's own error posture ``None`` must behave exactly like ``enforce`` -- an
    unreachable runtime or an unauthoritative query denies rather than silently
    proceeding. Only the explicit dry-run postures (``observe`` / ``disabled``) fail
    open locally.
    """
    return enforcement_mode is None or enforcement_mode == ENFORCE_MODE

So on the current published SDK every demo raises out of registration before reaching its first tool call, while its README promises:

No API key or running gateway is required for the offline demo.

This is not an SDK regression to work around. rc.6 was the version that was wrong — it let an agent proceed ungoverned under the advertised default. The examples were written against that hole and never declared the posture they depend on. Declaring it is the fix.

observe rather than disabled, because it is the documented dry-run posture and it keeps the offline demos symmetric with the live scenario, which already passes an explicit enforce at scenarios/live-core-enforcement/python-agent/agent.py:95.

Why the change lands in src/main.py too, not only the tests

The call in src/main.py sits inside the quickstart region that scripts/extract_snippets.py lifts into the published snippet. Fixing only the tests would turn CI green and leave the snippet a reader copies still broken. snippets/python/*.py are the regenerated output, and the third commit corrects the one README whose production note told the reader to remove arguments one at a time — following it verbatim would now leave a fail-open posture in a production agent.

How to verify

Every claim below was reproduced locally against the published rc.7 wheel, not against an unreleased build. Both rc.6 and rc.7 are on PyPI today.

Per example:

cd python/<name>
uv sync --extra dev --locked --no-build
uv run --locked --no-build pytest tests/ -v

Results on rc.7 with this branch — 16 of 16 suites pass:

agno-tool-policy                          6 passed
autogen-tool-policy                       6 passed
crewai-research-crew                      6 passed
custom-tool-policy                        7 passed
google-adk                                7 passed
haystack-tool-policy                      6 passed
langchain-basic-agent                     6 passed
langchain-research-agent                  8 passed
langgraph                                 5 passed
llamaindex-tool-policy                    6 passed
microsoft-agent-framework-tool-policy      4 passed
openai-agents-sdk                         7 passed
pydantic-ai                               7 passed
semantic-kernel-tool-policy               6 passed
smolagents-tool-policy                    8 passed
strands-agents-tool-policy                6 passed

Anti-vacuity: with the posture line stripped from all 16 test files and nothing else changed, 16 of 16 fail, one failing test each.

agno-tool-policy                          1 failed, 5 passed
autogen-tool-policy                       1 failed, 5 passed
crewai-research-crew                      1 failed, 5 passed
custom-tool-policy                        1 failed, 6 passed
google-adk                                1 failed, 6 passed
haystack-tool-policy                      1 failed, 5 passed
langchain-basic-agent                     1 failed, 5 passed
langchain-research-agent                  1 failed, 7 passed
langgraph                                 1 failed, 4 passed
llamaindex-tool-policy                    1 failed, 5 passed
microsoft-agent-framework-tool-policy      1 failed, 3 passed
openai-agents-sdk                         1 failed, 6 passed
pydantic-ai                               1 failed, 6 passed
semantic-kernel-tool-policy               1 failed, 5 passed
smolagents-tool-policy                    1 failed, 7 passed
strands-agents-tool-policy                1 failed, 5 passed

Reverting the pin to rc.6 on the same machine passes either way, which is what isolates the pin as the only variable.

The metadata drift gate, all five steps, run locally from the repo root:

python scripts/generate_example_metadata.py   -> No changes -- every example is already in sync with the SoT.
python scripts/extract_snippets.py            -> Extracted 26 snippet(s) across 3 SDK(s). No changes.
git diff --exit-code                          -> clean
python scripts/generate_example_metadata.py --check
                                              -> --check: every audited surface matches the SoT.
python -m unittest scripts.test_generate_example_metadata
                                              -> Ran 39 tests, OK

uv lock --check is clean on all 18 locked directories. Each relock moved exactly one package and preserved revision = 3.

Three examples (google-adk, semantic-kernel-tool-policy, strands-agents-tool-policy) cannot be installed locally with the exact CI flag combination, because watchdog publishes no macOS-arm64 wheel for this interpreter and pybars4 ships no wheel at all, so --no-build refuses them. That is pre-existing and unrelated: it reproduces identically on unmodified main at rc.6. Their numbers above come from the same command without --no-build, so the same test code ran on rc.7 and only dependency acquisition differed. CI runs ubuntu, where manylinux wheels exist. The command the READMEs actually give users is uv sync --extra dev, which has no such restriction.

What this means for the 18 open Dependabot PRs

They were recorded as blocked on a future release. That was wrong. Their two failures are both release-independent:

• metadata drift — the generator regenerates pyproject.toml from the SoT and reverts Dependabot's rc.7 back to rc.6.
• the per-example lane — uv sync --locked refuses because Dependabot bumped the manifest without relocking.

This PR resolves both by satisfaction rather than by closure. A separate ticket covers the structural half: Dependabot's pip ecosystem edits a generated pin, and advancing it correctly also rewrites a Dockerfile and a workflow that pip will never touch, so those PRs cannot satisfy the drift gate by construction.

Fourth commit: the SonarCloud duplication gate (AAASM-6206)

The quality gate initially failed on one condition, new_duplicated_lines_density 4.1% against a 3% threshold, localised to 3 new duplicated lines in each half of one pair:

dup=3  new=3  snippets/python/llamaindex-tool-policy.py
dup=3  new=3  python/llamaindex-tool-policy/src/main.py

The reported block is 31 lines — the whole extracted quickstart region, not the 3 added lines. snippets/python/llamaindex-tool-policy.py is a verbatim extraction of that region by scripts/extract_snippets.py, drift-gated by example-metadata-check.yml, so the two files are identical by construction. I verified that directly: all 16 Python snippet/source pairs are byte-identical, modulo dedent, on main as well as on this branch.

main reports duplicated_lines 0 and duplicated_blocks 0 for that same file, with an empty duplications list, so nothing about the content became duplicated here — only the detection changed. .sonarcloud.properties already excludes three sibling pairs (agno, crewai, haystack) from copy-paste detection for exactly this reason; its comment records that the list was "taken from the duplication report as of 2026-07-23", and llamaindex was below the detector's minimum block size on that day. One added argument plus its comment pushed it over.

So the fourth commit adds both halves of that pair to the existing list and replaces the dated-snapshot comment with the invariant. Both files stay fully analysed by every other rule, and every hand-written source file stays under full duplication analysis — nothing is removed from analysis scope, and no paid SonarCloud capacity is involved. The gate now passes all five conditions, with the PR's total duplication back to 101 lines / 7 blocks, exactly matching main.

AAASM-6206 records the structural half: the exclusion list is wildcard-free by platform restriction, so it is a hand-maintained snapshot with no gate, and the next region to cross the boundary will fail the same way.

Coverage caveat: the two rc.7 pins this PR's checks do not exercise

Two of the 58 files whose pin this advances are not reachable from any check on this PR:

• scenarios/live-core-enforcement/python-agent/Dockerfile — a RUN pip install line.
• .github/workflows/verify-live.yml — a run: string.

verify-live.yml is schedule (daily 06:00 UTC) plus workflow_dispatch only; it has no pull_request trigger. And verify-scenarios.yml's live: live-core-enforcement (real gateway) job is gated if: github.event_name == 'workflow_dispatch', which is the one skipped check on this PR. Its sibling validate: live-core-enforcement (offline, no gateway) does pass, but it only installs PyYAML to validate policy documents and never runs the agent, so it carries no posture exposure either.

Those two pin edits are therefore verified by inspection only here, and will first be executed by the next scheduled live run or a manual dispatch. Calling this PR fully end-to-end validated would overstate it.

Not done here

No release is created, and none is needed — rc.7 is already published.

Refs: AAASM-6204

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

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

See analysis details on SonarQube Cloud