juried 0.1.0__tar.gz

juried-0.1.0/.env.example

@@ -0,0 +1,61 @@
+# Environment variables read by juried.
+#
+# juried never reads this file itself: it only reads the process environment. Copy it
+# to .env, fill in the values you need, then load it into your shell before running:
+#
+#   set -a; source .env; set +a
+#
+# or let direnv, your CI secrets store or your shell profile export them.
+
+# ---------------------------------------------------------------------------
+# Judge and generation providers. Only the key for the provider named in
+# [judge] (or [generate]) of juried.toml is needed. The stub provider needs none.
+# ---------------------------------------------------------------------------
+ANTHROPIC_API_KEY=
+OPENAI_API_KEY=
+
+# ---------------------------------------------------------------------------
+# Target credentials. Header values in [target] of juried.toml may reference
+# any variable as ${NAME}; a missing variable stops the run with an error.
+# Name these to match your own config, for example:
+#   headers = { Authorization = "Bearer ${STAGING_TOKEN}" }
+# ---------------------------------------------------------------------------
+STAGING_TOKEN=
+
+# ---------------------------------------------------------------------------
+# Overrides for juried.toml. Any scalar key can be set as JURIED_<SECTION>_<KEY>
+# and takes precedence over the file. Sections: target, criteria, run, judge,
+# generate. Nested values such as target.headers and target.body cannot be
+# overridden this way. Leave unset to use the values in juried.toml.
+# ---------------------------------------------------------------------------
+
+# [target]
+#JURIED_TARGET_URL=https://staging.example.com/api/chat
+#JURIED_TARGET_METHOD=POST
+#JURIED_TARGET_RESPONSE_PATH=reply
+#JURIED_TARGET_TIMEOUT_SECONDS=30
+#JURIED_TARGET_RETRIES=3
+
+# [criteria]
+#JURIED_CRITERIA_FILE=acceptance.md
+#JURIED_CRITERIA_SCENARIOS_DIR=scenarios
+
+# [run]
+#JURIED_RUN_RUNS=10
+#JURIED_RUN_THRESHOLD=0.7
+#JURIED_RUN_CONCURRENCY=4
+#JURIED_RUN_CACHE_DIR=.juried
+#JURIED_RUN_REPORT_DIR=reports
+
+# [judge]  provider is anthropic, openai or stub
+#JURIED_JUDGE_PROVIDER=anthropic
+#JURIED_JUDGE_MODEL=claude-sonnet-5
+#JURIED_JUDGE_TEMPERATURE=0.0
+#JURIED_JUDGE_MAX_TOKENS=2048
+#JURIED_JUDGE_BASE_URL=
+
+# [generate]  provider and model default to the [judge] values
+#JURIED_GENERATE_PROVIDER=
+#JURIED_GENERATE_MODEL=
+#JURIED_GENERATE_SCENARIOS_PER_CRITERION=4
+#JURIED_GENERATE_MAX_TOKENS=4096

juried-0.1.0/.gitignore

@@ -0,0 +1,13 @@
+.juried/
+.venv/
+reports/
+__pycache__/
+*.pyc
+*.egg-info/
+dist/
+.mypy_cache/
+.ruff_cache/
+.pytest_cache/
+examples/faq-bot/scenarios/generated/
+out.xml
+.env

juried-0.1.0/CONTRIBUTING.md

@@ -0,0 +1,37 @@
+# Contributing
+
+## Setup
+
+```
+uv venv --python 3.12 .venv && source .venv/bin/activate
+uv pip install -e ".[dev]"
+make check
+```
+
+`make check` runs ruff, mypy (strict) and the test suite. All three must be clean before
+a change is ready. `make example` starts the fake FAQ bot, generates scenarios and runs
+them end to end without API keys.
+
+## Layout
+
+- `src/juried/config.py`, `criteria.py`, `scenarios.py`: models and file parsing.
+- `src/juried/targets/`: the system under test. Stage 1 has an HTTP endpoint target.
+- `src/juried/judge/`: providers. `base.py` holds the interface and the shared judge and
+  generation logic, `prompts.py` the pinned prompts, `stub.py` a key free provider.
+- `src/juried/runner.py`: repeated runs, concurrency and caching.
+- `src/juried/pytest_plugin.py`: collection of YAML scenarios as pytest items.
+- `src/juried/report/`: JSON and single file HTML output.
+- `src/juried/cli.py`: `init`, `generate` and `run`.
+
+Extension points for later stages: a new `Target` implementation for UI driving, a new
+`Provider` for other judges, a `turns` field on `Scenario` for multi turn conversations,
+and new scenario kinds for adversarial generation.
+
+## House rules
+
+- UK English in code, comments, docs and output
+- Comment only what the code cannot say. No docstrings that restate a name.
+- Keep dependencies to the current list unless there is a strong reason.
+- Type hints on public functions; ruff and mypy stay clean without disabling rules.
+- Bump `PROMPT_VERSION` in `judge/prompts.py` whenever a prompt changes, so cached
+  verdicts from the old prompt are not reused.

juried-0.1.0/LICENCE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 juried contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

juried-0.1.0/Makefile

@@ -0,0 +1,21 @@
+PYTHON ?= python
+
+.PHONY: check lint type test example
+
+check: lint type test
+
+lint:
+	$(PYTHON) -m ruff check src tests examples
+	$(PYTHON) -m ruff format --check src tests examples
+
+type:
+	$(PYTHON) -m mypy
+
+test:
+	$(PYTHON) -m pytest
+
+example:
+	cd examples/faq-bot && $(PYTHON) server.py & \
+	sleep 1; \
+	cd examples/faq-bot && juried generate && juried run; status=$$?; \
+	kill %1; exit $$status

juried-0.1.0/PKG-INFO

@@ -0,0 +1,182 @@
+Metadata-Version: 2.5
+Name: juried
+Version: 0.1.0
+Summary: Acceptance testing for LLM features, built for QA teams.
+License-Expression: MIT
+License-File: LICENCE
+Requires-Python: >=3.12
+Requires-Dist: httpx>=0.27
+Requires-Dist: jinja2>=3.1
+Requires-Dist: pydantic>=2.6
+Requires-Dist: pytest>=8.0
+Requires-Dist: pyyaml>=6.0
+Provides-Extra: dev
+Requires-Dist: mypy>=1.11; extra == 'dev'
+Requires-Dist: ruff>=0.6; extra == 'dev'
+Requires-Dist: types-pyyaml; extra == 'dev'
+Description-Content-Type: text/markdown
+
+<p align="center">
+  <img src="https://raw.githubusercontent.com/jch1887/juried/main/juried.png" alt="juried" width="360">
+  <br>
+  Acceptance testing for LLM features, built for QA teams.
+</p>
+
+juried treats your LLM powered feature as a black box behind an HTTP endpoint. You write
+acceptance criteria in plain English, juried generates test scenarios from them, runs each
+scenario repeatedly, has a pinned LLM judge mark every response, and reports pass rates
+with confidence intervals. Scenarios are ordinary pytest tests, so `-k`, `-x`, markers
+and `--junitxml` all work and the results fit an existing CI job. The name says how it
+works: the feature goes before a jury of repeated runs and only passes when the panel's
+verdict holds.
+
+## Install
+
+```
+pip install juried
+```
+
+Python 3.12 or later. Judges read `ANTHROPIC_API_KEY` or `OPENAI_API_KEY` from the
+environment; nothing else is ever read from disk.
+
+## Three commands
+
+```
+juried init        # writes juried.toml and an example acceptance.md
+juried generate    # turns each criterion into scenarios/generated/<criterion>.yaml
+juried run         # runs every scenario N times under pytest and writes the report
+```
+
+`juried run` accepts pytest arguments after its own, for example
+`juried run --runs 20 -k refunds -x --junitxml=out.xml`. Generation is a one off step:
+review and edit the generated YAML, commit it, and `run` never touches it.
+
+## Configuration
+
+`juried.toml` lives at the repo root. Any key can be overridden with an environment
+variable named `JURIED_<SECTION>_<KEY>`, such as `JURIED_RUN_RUNS=20`. `.env.example`
+lists every variable juried reads, including the provider keys and the overrides;
+copy it to `.env` and export it from your shell, since juried does not load it itself.
+
+```toml
+[target]
+url = "https://staging.example.com/api/chat"
+headers = { Authorization = "Bearer ${STAGING_TOKEN}" }
+body = { message = "{{message}}", history = "{{history}}" }
+response_path = "choices.0.message.content"
+
+[run]
+runs = 10          # attempts per scenario
+threshold = 0.7    # gate on the lower bound of the Wilson 95% interval
+concurrency = 4
+
+[judge]
+provider = "anthropic"    # anthropic, openai or stub
+model = "claude-sonnet-5" # pinned and recorded with every verdict
+```
+
+Set `temperature` under `[judge]` only for a model that accepts it. `claude-sonnet-5` rejects
+the parameter, so the example leaves it out; when it is unset nothing is sent and the report
+says so.
+
+`{{message}}` is replaced with the scenario message; a value of exactly `"{{history}}"`
+becomes the earlier turns as a list of `{role, content}` objects. `response_path` is a
+dotted path into the JSON reply.
+
+## Criteria and scenarios
+
+Criteria are `##` headings in a Markdown file. The heading becomes a stable id; add
+`{#id}` to fix it explicitly.
+
+```markdown
+## Refund policy
+The bot explains that any item can be returned within 14 days of delivery for a full
+refund. It must state the 14 day window.
+
+## Unknown questions {#unknown}
+When the bot cannot answer it says so and gives help@example.com.
+```
+
+Generated and hand written scenarios share one YAML shape. Hand written files go
+anywhere under `scenarios/`; `runs` and `threshold` may be set per scenario.
+
+```yaml
+criterion: refund-policy
+scenarios:
+  - name: Asks how to get money back
+    kind: happy_path
+    message: I want my money back on a jumper that does not fit.
+    expected: States the 14 day return window and that the refund is full.
+  - name: Follows up after a delivery answer
+    kind: edge_case
+    history:
+      - role: user
+        content: Can you tell me about delivery?
+      - role: assistant
+        content: Standard delivery takes 3 to 5 working days.
+    message: and refunds?
+    expected: Explains the 14 day refund window without repeating the delivery answer.
+    runs: 20
+    threshold: 0.8
+```
+
+A phrase in double quotes inside `expected` must appear in the response word for word,
+ignoring case. Text outside quotes is judged on meaning. With
+`expected: Says returns are accepted within "14 days" for a full refund`, a response saying
+"you have 14 days and get every penny back" passes, while "a fortnight for a full refund"
+fails because `14 days` is missing. The stub judge applies the same rule.
+
+## How a scenario passes
+
+Each scenario runs `runs` times. The judge marks each response pass or fail with a one
+line reason, using a fixed prompt and the configured temperature, if any. juried computes the
+pass rate and its
+Wilson 95% interval, and the scenario passes when the lower bound meets `threshold`. With
+10 runs a perfect score gives a lower bound of 0.72, so the default threshold is 0.7; a
+stricter threshold needs more runs, and juried warns when a gate can never pass.
+
+A failing gate is a normal pytest failure that shows the pass rate, the interval, the
+threshold and the first failing transcript with the judge's reason. An HTTP error from
+your endpoint is reported as a transport error, separately from a judge fail. Responses
+and verdicts are cached in `.juried/` by content hash, so re-runs during development are
+cheap; pass `--no-cache` to skip it. Every verdict is appended to `.juried/verdicts.jsonl`
+with the judge model and timestamp.
+
+## The report
+
+After a run juried writes `reports/juried-report.html` and `reports/juried-report.json`.
+The HTML is a single self contained file with no scripts. It opens with the totals, then
+lists each acceptance criterion with its description, a table of its scenarios showing
+passes, pass rate, interval, threshold, response latency and gate result, and beneath the
+table each scenario's message, expectation and every failing run with the response and
+the judge's reason. Criteria with no scenarios are called out so coverage gaps are
+visible. The JSON file holds the same structure plus every attempt, for anyone who wants
+to chart trends.
+
+<img src="https://raw.githubusercontent.com/jch1887/juried/main/docs/report.png" alt="juried report showing three scenarios, one at 7 of 10 failing because its lower bound is below the threshold" width="900">
+
+In the third row seven of ten runs passed and the observed rate meets the threshold, yet the
+gate still fails because the lower bound does not.
+
+## Try it without API keys
+
+```
+cd examples/faq-bot
+python server.py &      # deterministic fake FAQ bot on port 8765
+juried generate          # uses the stub provider from juried.toml
+juried run
+```
+
+The stub judge passes a response that contains every `"quoted phrase"` in `expected`.
+One hand written refund scenario fails its gate on purpose, because the fake bot drops
+the 14 day detail every fourth time, which is the kind of flakiness juried exists to catch.
+
+## Development
+
+```
+make check      # ruff, mypy and pytest
+```
+
+See CONTRIBUTING.md for the development install.
+
+Licensed under the MIT licence.

juried-0.1.0/README.md

@@ -0,0 +1,164 @@
+<p align="center">
+  <img src="https://raw.githubusercontent.com/jch1887/juried/main/juried.png" alt="juried" width="360">
+  <br>
+  Acceptance testing for LLM features, built for QA teams.
+</p>
+
+juried treats your LLM powered feature as a black box behind an HTTP endpoint. You write
+acceptance criteria in plain English, juried generates test scenarios from them, runs each
+scenario repeatedly, has a pinned LLM judge mark every response, and reports pass rates
+with confidence intervals. Scenarios are ordinary pytest tests, so `-k`, `-x`, markers
+and `--junitxml` all work and the results fit an existing CI job. The name says how it
+works: the feature goes before a jury of repeated runs and only passes when the panel's
+verdict holds.
+
+## Install
+
+```
+pip install juried
+```
+
+Python 3.12 or later. Judges read `ANTHROPIC_API_KEY` or `OPENAI_API_KEY` from the
+environment; nothing else is ever read from disk.
+
+## Three commands
+
+```
+juried init        # writes juried.toml and an example acceptance.md
+juried generate    # turns each criterion into scenarios/generated/<criterion>.yaml
+juried run         # runs every scenario N times under pytest and writes the report
+```
+
+`juried run` accepts pytest arguments after its own, for example
+`juried run --runs 20 -k refunds -x --junitxml=out.xml`. Generation is a one off step:
+review and edit the generated YAML, commit it, and `run` never touches it.
+
+## Configuration
+
+`juried.toml` lives at the repo root. Any key can be overridden with an environment
+variable named `JURIED_<SECTION>_<KEY>`, such as `JURIED_RUN_RUNS=20`. `.env.example`
+lists every variable juried reads, including the provider keys and the overrides;
+copy it to `.env` and export it from your shell, since juried does not load it itself.
+
+```toml
+[target]
+url = "https://staging.example.com/api/chat"
+headers = { Authorization = "Bearer ${STAGING_TOKEN}" }
+body = { message = "{{message}}", history = "{{history}}" }
+response_path = "choices.0.message.content"
+
+[run]
+runs = 10          # attempts per scenario
+threshold = 0.7    # gate on the lower bound of the Wilson 95% interval
+concurrency = 4
+
+[judge]
+provider = "anthropic"    # anthropic, openai or stub
+model = "claude-sonnet-5" # pinned and recorded with every verdict
+```
+
+Set `temperature` under `[judge]` only for a model that accepts it. `claude-sonnet-5` rejects
+the parameter, so the example leaves it out; when it is unset nothing is sent and the report
+says so.
+
+`{{message}}` is replaced with the scenario message; a value of exactly `"{{history}}"`
+becomes the earlier turns as a list of `{role, content}` objects. `response_path` is a
+dotted path into the JSON reply.
+
+## Criteria and scenarios
+
+Criteria are `##` headings in a Markdown file. The heading becomes a stable id; add
+`{#id}` to fix it explicitly.
+
+```markdown
+## Refund policy
+The bot explains that any item can be returned within 14 days of delivery for a full
+refund. It must state the 14 day window.
+
+## Unknown questions {#unknown}
+When the bot cannot answer it says so and gives help@example.com.
+```
+
+Generated and hand written scenarios share one YAML shape. Hand written files go
+anywhere under `scenarios/`; `runs` and `threshold` may be set per scenario.
+
+```yaml
+criterion: refund-policy
+scenarios:
+  - name: Asks how to get money back
+    kind: happy_path
+    message: I want my money back on a jumper that does not fit.
+    expected: States the 14 day return window and that the refund is full.
+  - name: Follows up after a delivery answer
+    kind: edge_case
+    history:
+      - role: user
+        content: Can you tell me about delivery?
+      - role: assistant
+        content: Standard delivery takes 3 to 5 working days.
+    message: and refunds?
+    expected: Explains the 14 day refund window without repeating the delivery answer.
+    runs: 20
+    threshold: 0.8
+```
+
+A phrase in double quotes inside `expected` must appear in the response word for word,
+ignoring case. Text outside quotes is judged on meaning. With
+`expected: Says returns are accepted within "14 days" for a full refund`, a response saying
+"you have 14 days and get every penny back" passes, while "a fortnight for a full refund"
+fails because `14 days` is missing. The stub judge applies the same rule.
+
+## How a scenario passes
+
+Each scenario runs `runs` times. The judge marks each response pass or fail with a one
+line reason, using a fixed prompt and the configured temperature, if any. juried computes the
+pass rate and its
+Wilson 95% interval, and the scenario passes when the lower bound meets `threshold`. With
+10 runs a perfect score gives a lower bound of 0.72, so the default threshold is 0.7; a
+stricter threshold needs more runs, and juried warns when a gate can never pass.
+
+A failing gate is a normal pytest failure that shows the pass rate, the interval, the
+threshold and the first failing transcript with the judge's reason. An HTTP error from
+your endpoint is reported as a transport error, separately from a judge fail. Responses
+and verdicts are cached in `.juried/` by content hash, so re-runs during development are
+cheap; pass `--no-cache` to skip it. Every verdict is appended to `.juried/verdicts.jsonl`
+with the judge model and timestamp.
+
+## The report
+
+After a run juried writes `reports/juried-report.html` and `reports/juried-report.json`.
+The HTML is a single self contained file with no scripts. It opens with the totals, then
+lists each acceptance criterion with its description, a table of its scenarios showing
+passes, pass rate, interval, threshold, response latency and gate result, and beneath the
+table each scenario's message, expectation and every failing run with the response and
+the judge's reason. Criteria with no scenarios are called out so coverage gaps are
+visible. The JSON file holds the same structure plus every attempt, for anyone who wants
+to chart trends.
+
+<img src="https://raw.githubusercontent.com/jch1887/juried/main/docs/report.png" alt="juried report showing three scenarios, one at 7 of 10 failing because its lower bound is below the threshold" width="900">
+
+In the third row seven of ten runs passed and the observed rate meets the threshold, yet the
+gate still fails because the lower bound does not.
+
+## Try it without API keys
+
+```
+cd examples/faq-bot
+python server.py &      # deterministic fake FAQ bot on port 8765
+juried generate          # uses the stub provider from juried.toml
+juried run
+```
+
+The stub judge passes a response that contains every `"quoted phrase"` in `expected`.
+One hand written refund scenario fails its gate on purpose, because the fake bot drops
+the 14 day detail every fourth time, which is the kind of flakiness juried exists to catch.
+
+## Development
+
+```
+make check      # ruff, mypy and pytest
+```
+
+See CONTRIBUTING.md for the development install.
+
+Licensed under the MIT licence.

juried-0.1.0/examples/faq-bot/acceptance.md

@@ -0,0 +1,16 @@
+# FAQ bot acceptance criteria
+
+## Opening hours
+The bot tells customers the opening hours: Monday to Friday, 9am to 5pm. It says the
+shop is closed at weekends.
+
+## Refund policy
+The bot explains that any item can be returned within 14 days of delivery for a full
+refund. It must state the 14 day window.
+
+## Delivery times
+The bot states that standard delivery takes 3 to 5 working days.
+
+## Unknown questions {#unknown}
+When the bot cannot answer it says so and gives the support address help@example.com
+rather than inventing an answer.

juried-0.1.0/examples/faq-bot/juried.toml

@@ -0,0 +1,14 @@
+[target]
+url = "http://127.0.0.1:8765/chat"
+body = { message = "{{message}}", history = "{{history}}" }
+response_path = "reply"
+
+[run]
+runs = 10
+threshold = 0.7
+
+[judge]
+provider = "stub"
+
+[generate]
+scenarios_per_criterion = 2

juried-0.1.0/examples/faq-bot/scenarios/handwritten.yaml

@@ -0,0 +1,37 @@
+# Hand written scenarios. The stub judge passes a response that contains every
+# "quoted phrase" in expected; a real judge reads the whole expectation.
+scenarios:
+  - criterion: opening-hours
+    name: Asks when the shop is open
+    kind: happy_path
+    message: What time are you open?
+    expected: Gives the hours, including "9am" and "5pm", and mentions "weekends".
+
+  - criterion: opening-hours
+    name: Asks about Sunday
+    kind: edge_case
+    message: can i pop in on sunday?
+    expected: Says the shop is closed at "weekends".
+
+  - criterion: refund-policy
+    name: Asks how to get money back
+    kind: happy_path
+    message: I want my money back on a jumper that does not fit.
+    expected: States the "14 days" return window and that the refund is "full".
+
+  - criterion: unknown
+    name: Asks something off topic
+    kind: edge_case
+    message: Do you know a good pizza place nearby?
+    expected: Admits it does not know and gives "help@example.com".
+
+  - criterion: unknown
+    name: Follows up with yes
+    kind: edge_case
+    history:
+      - role: user
+        content: Can you tell me about delivery?
+      - role: assistant
+        content: Standard delivery takes 3 to 5 working days and costs 3.99.
+    message: yes please
+    expected: Stays on the delivery topic, mentioning "working days".

juried-0.1.0/examples/faq-bot/server.py

@@ -0,0 +1,79 @@
+"""A deterministic stand in for an LLM backed FAQ bot, so the example runs without API keys.
+
+Run it with `python server.py` and point juried at http://127.0.0.1:8765/chat. Every fourth
+refund question deliberately drops the 14 day detail, so that one scenario shows a flaky
+pass rate in the report.
+"""
+
+from __future__ import annotations
+
+import json
+import threading
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from typing import Any
+
+PORT = 8765
+
+HOURS = "We are open Monday to Friday, 9am to 5pm, and closed at weekends."
+REFUND_FULL = "You can return any item within 14 days of delivery for a full refund."
+REFUND_VAGUE = "Returns are accepted for a full refund, please contact us to arrange one."
+DELIVERY = "Standard delivery takes 3 to 5 working days and costs 3.99."
+UNKNOWN = "I'm not sure about that. Please email help@example.com and the team will help."
+
+
+class Handler(BaseHTTPRequestHandler):
+    refund_calls = 0
+    lock = threading.Lock()
+
+    def do_POST(self) -> None:
+        if self.path != "/chat":
+            self.reply(404, {"error": "not found"})
+            return
+        length = int(self.headers.get("Content-Length", "0"))
+        try:
+            payload = json.loads(self.rfile.read(length) or b"{}")
+        except json.JSONDecodeError:
+            self.reply(400, {"error": "body must be JSON"})
+            return
+        message = str(payload.get("message", "")).lower()
+        history = payload.get("history") or []
+        self.reply(200, {"reply": answer(message, history)})
+
+    def reply(self, status: int, body: dict[str, Any]) -> None:
+        data = json.dumps(body).encode("utf-8")
+        self.send_response(status)
+        self.send_header("Content-Type", "application/json")
+        self.send_header("Content-Length", str(len(data)))
+        self.end_headers()
+        self.wfile.write(data)
+
+    def log_message(self, format: str, *args: Any) -> None:
+        return None
+
+
+def answer(message: str, history: list[dict[str, str]]) -> str:
+    if any(word in message for word in ("hour", "open", "close", "weekend", "saturday", "sunday")):
+        return HOURS
+    if any(word in message for word in ("refund", "return", "money back")):
+        with Handler.lock:
+            Handler.refund_calls += 1
+            vague = Handler.refund_calls % 4 == 0
+        return REFUND_VAGUE if vague else REFUND_FULL
+    if any(word in message for word in ("deliver", "shipping", "post")):
+        return DELIVERY
+    if history and message.strip() in {"yes", "yes please", "ok", "go on"}:
+        return "Of course. " + answer(history[-1].get("content", "").lower(), [])
+    return UNKNOWN
+
+
+def main() -> None:
+    server = ThreadingHTTPServer(("127.0.0.1", PORT), Handler)
+    print(f"faq-bot listening on http://127.0.0.1:{PORT}/chat")
+    try:
+        server.serve_forever()
+    except KeyboardInterrupt:
+        server.shutdown()
+
+
+if __name__ == "__main__":
+    main()