# Operator Guide Open the operator UI at `http://:8600`. The screen has four areas: | Area | Use it to | | --- | --- | | **Connected apps** | See which store services the agent can reach and how many tools each one offers | | **Operator Chat** | Ask questions and give instructions | | **Autonomy Decisions** | Review events, read the agent's reasoning, and approve or reject proposals | | **Hermes skill learning** | Appears when the agent has enough operator responses to learn from | ## Ask Questions Type a question and select **Send**. Answers stream in as they are generated. Examples: - *How long is the kiosk queue and what is the wait time?* - *Which menu items are unavailable right now?* - *What is the order accuracy rate for the last 30 minutes?* - *Were there any kitchen safety events this shift?* - *Which tables need cleaning?* What the agent can answer depends on which services are connected. Questions unrelated to the store are declined with: *"I can only help with questions about this store's connected services."* The agent remembers the conversation and your stated preferences. Select **New chat** to start a fresh conversation. Preferences remembered by the agent carry over. After 30 minutes of inactivity, the next question automatically starts a new conversation. ## Review Event Decisions When a store event arrives, it appears in **Autonomy Decisions** with one of these states: | State | Meaning | | --- | --- | | Queued / Processing | The agent is evaluating the event | | Pending | The agent proposes an action and is waiting for you | | No action | The agent decided no change is needed and explains why | | Failed | The evaluation did not complete; the event can be sent again | For a pending proposal, read the reason and select **Approve** to apply it or **Reject** to discard it. You can add a reason when rejecting, which helps the agent learn. Proposals expire after five minutes. While an event is shown, questions you ask in chat are linked to it. For example, *"Why did you suggest that?"* or *"Switch the kiosks to express mode instead."* Select **✕** on the event banner if your question is not about that event. ## Try Demo Events Change the simulated weather. Valid conditions are `clear`, `cloudy`, `fog`, `drizzle`, `rain`, `storm`, and `snow`: ```bash make weather CONDITION=rain TEMP=12 make weather CONDITION=clear NO_EVENT=1 # change silently, no event ``` Other events come from the connected store services themselves. Any store system can send events; see the [API Reference](api-reference.md#post-autonomyevents). ## Teach the Agent The agent records how you respond to events: approvals, rejections with reasons, and instructions you give in chat while an event is shown. When a service has 20 responses (`QSR_LEARN_THRESHOLD`), with at least 5 for one event type (`QSR_LEARN_MIN_EVENT_RECORDS`), the learning card shows **Learn now**. The command line needs only the 5 per event type (`--min-records`). 1. Select **Learn now**. Learning takes about a minute. 2. When the card shows **New skill ready for review**, select **Review**. 3. Read the learned rules, then approve or reject them. Approved rules guide the agent's proposals for future events from that service. Every proposal still needs your approval. The same steps are available from the command line: ```bash make learn ARGS=status make learn ARGS=run make learn ARGS=pending make learn ARGS="approve " ``` Learned rules waiting for review are kept in the `learning` volume, so they survive `make restart`, `make up` and image rebuilds. ### Try learning with sample data To see learning work before you have collected real responses, import the synthetic sample history (`learning/simulated/event_op_response.jsonl`, marked `source: "simulated"`). Use a test deployment only: imported records are added to the same journal as real operator responses, so do not import them into a production store. ```bash docker compose cp learning/simulated/event_op_response.jsonl qsr-agent:/tmp/event_op_response.jsonl make learn ARGS="import /tmp/event_op_response.jsonl" make learn ARGS=status make learn ARGS="run --service " make learn ARGS=pending make learn ARGS="approve " ``` Importing is safe to repeat: records already in the journal are skipped. After approval, `learned-/SKILL.md` appears under `learning/skills/` in the `learning` volume, not in the repository's `learning/` folder. Read it with `docker compose exec qsr-agent cat /opt/qsr/learning/skills/learned-/SKILL.md`. To discard a proposal, run `make learn ARGS="reject "`. To learn again from scratch for one service, for example after importing new sample data, delete its learned skill first (test deployments only): ```bash docker compose exec qsr-agent rm -rf /opt/qsr/learning/skills/learned- ``` The sample file can be regenerated with `python scripts/simulate_service_history.py`. Notes: - Learning needs the model server to be running, and takes about a minute. - Outside Docker, run `scripts/learn.py` with `QSR_MCP_PYTHON` pointing to a Python that has `fastmcp` installed (the default is `.venv/mcp/bin/python`). Without it, the learner cannot read tool schemas and drops its rules. - Outside Docker, run `python scripts/learn.py setup` once so the learner profile reads and writes the `learning/skills` folder you intend to use. Use `QSR_LEARNING_DIR` and `QSR_LEARNER_HERMES_HOME` to keep a test run separate from real state. ## Command-Line Chat To chat with the agent in a terminal instead of the browser: ```bash docker compose exec qsr-agent hermes ```