Verify example catalog against conductor-oss 3.32.0-rc.8; fix broken examples + worker-tool registry lookup

[ling-senpeng13]

Summary

Full verification of the SDK example catalog against a live conductor-oss 3.32.0-rc.8 server, tracked in examples/VERIFICATION_REPORT.md, plus fixes for every reproducible example bug found and one SDK bug.

Documentation only for the server-version issue — no CLI/pipeline version change. The agent examples require conductor-oss ≥ 3.32.0-rc.8 (agent runtime on by default); the CLI's floating latest (3.30.x) has no /api/agent/* endpoints. The READMEs now say so and give conductor server start --version 3.32.0-rc.8. The latest pointer resolves itself at 3.32.0 GA.

Test status (tracked in the report)

Suite	Result
agents (examples/agents/, 146 files)	96 PASS / 46 SKIP (interactive/infra) / 2 ERROR / 2 HUNG
before these fixes, same server	91 PASS / 12 ERROR / 3 HUNG / 40 SKIP
on 3.30.2 (CLI latest)	3 PASS / 103 ERROR — hence the doc change
core (curated 18)	13 PASS / 4 worker-daemons run clean by design / 1 interactive
tests/unit/ai	1744 passed

Example fixes (each verified end-to-end)

• greetings_worker.py — duplicate def greet shadowed the registered worker → spawn-pickling failure on macOS/Windows (broke workers_e2e, helloworld, task_listener_example)
• worker_example.py — hardcoded /Users/viren/ metrics dir → tempfile.gettempdir()
• user_workers.py — examples.* import vs user_example.* module identity crashed spawn children
• kitchen_sink.py — @agent classifiers must take zero args (decorator evaluates them for dynamic instructions)
• 16i / 16j — tools defined inside factory functions aren't importable by spawn workers (SpawnSafetyError) → hoisted to module level
• 94_openai_runner_tools.py — apply function_tool() at Agent(...) construction so the module global isn't rebound to an unpicklable FunctionTool
• llm_chat.py — server requires a non-empty user message (two tasks were system-only), and the example reported success on a FAILED workflow (is_completed() is true for any terminal state); both fixed
• run_all_examples.py — 6 env-dependent examples classified SKIP with reasons (dg skill repo; workflow messages API absent in rc.8)

SDK fix

_try_worker_task in conductor/ai/agents/tool.py: the registry identity check broke once ToolRegistry.register_tool_workers re-registered a task name with a spawn-safe ToolWorkerEntry wrapper — a second get_tool_def() on the same @worker_task tool raised TypeError (broke 14_existing_workers.py). Lookup now walks __wrapped__/fn_direct/fn_ref carriers. Two regression tests added.

Remaining known failures (documented in the report, not fixed here)

• kitchen_sink.py — server-side: translation_swarm sub-workflow definition not registered before /agent/start (HTTP 500)
• 74_cli_error_output.py — flaky LLM-behavior assertion
• 86_coding_agent.py — hangs >420s even solo
• 68_context_condensation.py — passes solo (~346s), exceeds suite timeout under contention

Follow-ups suggested (out of scope)

AgentNotFoundError should name the minimum server version; run_all_examples.py could preflight GET /api/version; spawn-safety layer could learn a closure hop for openai-agents FunctionTool.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 87.50000% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/conductor/ai/agents/tool.py	87.50%	2 Missing ⚠️

Files with missing lines	Coverage Δ
src/conductor/ai/agents/tool.py	91.54% <87.50%> (-0.40%)	⬇️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[ling-senpeng13]

Closing to split this into two draft PRs: (1) verification report + docs only, (2) the example/SDK fixes — so the fixes can be evaluated separately.