Capability: demo

One harness instance on Temporal Cloud with gpt-6.1-sol, two MCP servers, a skill and both renderers on one task; its e2e tests prove the flow and crash recovery.

What it is

The definition of done of issue-3, kept runnable: examples/demo is an Agent Plugin (the orders and policy stdio MCP servers, the support-agent skill, a prompt extension) with a config.toml, started by python -m examples.demo. The support agent example is the walkthrough.

Current behaviour

• python -m examples.demo [CONFIG] SHALL start the A2A server with an in-process worker on 127.0.0.1:8080 through tiny_harness.service.serve, serve the built web renderer under /ui and print how to start the TUI; both surfaces reach the same instance. CONFIG defaults to config.toml (Temporal Cloud); config.embedded.toml runs the same demo on an embedded Temporal with no Temporal account.
• WHEN the complaint about order #48213 is sent THEN the agent SHALL load the skill, create a plan, call get_order and policy.lookup, emit an A2UI card with the resolution options and a Confirm button, ask the reporter and wait in INPUT_REQUIRED; on the reply or the card's action it SHALL act with ship_replacement or refund and complete with a two-sentence summary.
• The orders server SHALL append every call to $PLUGIN_DATA/orders-ledger.jsonl and, while $PLUGIN_DATA/slow-get-order exists, hold get_order open, so a test can kill the worker mid-activity.
• tests/e2e/test_demo.py SHALL drive the conversation through the A2A client (pressing the card's button for the first question) and assert the states, the artifact, the ledger (a non-idempotent tool at most once), one chat span per invoke_llm run, cached tokens from turn two and no secret in the trace or log.
• tests/e2e/test_crash_recovery.py SHALL run the server and the worker as separate processes, kill the worker with SIGKILL during get_order and again while the task waits for input, restart it, and assert from Temporal's history that every invoke_llm was scheduled once and completed once and that the ledger shows ship_replacement at most once.
• tests/e2e/test_demo_embedded.py SHALL run the demo in embedded mode as one serve process with no TEMPORAL_API_KEY in its environment and assert the task completes, the embedded server's start and warning are logged, and no secret is in the trace or log.
• tests/e2e/test_demo_ollama.py SHALL run the demo with config.ollama.toml's settings — embedded Temporal, a loopback Ollama over Chat Completions — with neither OPENAI_API_KEY nor TEMPORAL_API_KEY in its environment, and assert the task completes, the model endpoint line names the loopback endpoint, and no secret is in the trace or log. It skips with the reason when no Ollama answers on 127.0.0.1:11434 or the model (TINY_HARNESS_OLLAMA_MODEL, default qwen3:8b) is not pulled.
• The e2e tests SHALL skip with the missing variable's name when a variable they need is absent (OPENAI_API_KEY and TEMPORAL_API_KEY; OPENAI_API_KEY alone for the embedded demo); each run uses its own task queue, port and state directory, and writes redacted evidence to TINY_HARNESS_E2E_EVIDENCE when set.

Design

design.md § Testing strategy, testing-plan.md rows T4 and T12.

History

Work item	What changed	Links
issue-3	The demo plugin, configuration and e2e tests (Layer 9)	spec, issue #3
issue-17	config.embedded.toml, the optional config path, the embedded-mode e2e	spec, PR #18
issue-19	config.ollama.toml, the offline Ollama e2e	spec, PR #22