Survive a restart, wait for a person

1. 01Write the four steps as a graph and run one small claim end to end

   uv run --with langgraph --with langgraph-checkpoint-postgres --with "psycopg[binary]" --with openai claim_flow.py start --thread t0 --amount 20

   You pass when: The claim is paid in one run, and a payments table has one row for t0.

   Check step 1 claim_flow.py

2. 02Stop and wait for approval when the amount is over 100

   uv run --with langgraph --with langgraph-checkpoint-postgres --with "psycopg[binary]" --with openai claim_flow.py start --thread t1 --amount 250

   You pass when: The process prints that it is waiting for approval and exits. Nothing is paid.

   Check step 2 claim_flow.py

3. 03Approve it from a new process

   The second run is the real test. Any step that moves money must be safe to repeat.

   uv run --with langgraph --with langgraph-checkpoint-postgres --with "psycopg[binary]" --with openai claim_flow.py resume --thread t1 --approve

   You pass when: It carries on from the saved step and pays. Run the same command again: there is still exactly one payment row for t1.

   Check step 3 claim_flow.py

4. 04See the same contract in a ready-made stack

   Optional. Flisk's agent resource keeps sessions in a checkpoint store in the same way your flow keeps threads in Postgres.

   flisk project agent-stack && flisk up

   You pass when: A local stack starts with an agent, tools and a database. flisk status --json shows the resources; flisk agent sessions <resource> lists stored sessions, and they are still listed after flisk restart <resource>.

   Check step 4

The check runs on the computer, not on your word: a step passes when its command exits with 0 there.