cassette-fn

Record real calls once, replay them deterministically in tests, while everything around the call, including your control flow, keeps running for real. This is the Python port of the JavaScript cassette-fn package.

The problem

Testing anything that calls an LLM (or any other slow, expensive, or nondeterministic function) usually means one of three bad options: pay for a real API call on every test run and accept nondeterminism, hand-roll a mock object that quietly drifts from the real provider response shape, or reach for an HTTP-level VCR library. The HTTP option looks appealing but it freezes the whole call site: it intercepts the request before your code ever sees it, so on replay your retry logic and control flow never actually execute, and the test ends up verifying the recording instead of your code. cassette_fn records at the function boundary you choose instead, so only the one call that actually talks to a provider is faked.

Install

pip install cassette-fn

Usage

from cassette_fn import tape

# The one function that actually talks to a provider. Real LLM SDK calls are
# keyword-based, e.g. client.messages.create(model=..., max_tokens=..., messages=...);
# wrap() records and replays both positional and keyword arguments.
def call_model(prompt, model="claude-opus-5", max_tokens=1024):
    response = requests.post(
        "https://api.example.com/v1/complete",
        json={"prompt": prompt, "model": model, "max_tokens": max_tokens},
    )
    return response.json()

t = tape(dir=".tapes", name="greeting-test")
model = t.wrap(call_model)

# First run (no cassette yet): calls through for real, records the result.
# Every run after that (cassette exists): replays from disk, call_model never runs.
result = model("say hi", model="claude-opus-5", max_tokens=1024)

t.save()  # writes .tapes/greeting-test.json (no-op in replay/off mode)
print(t.stats())  # Stats(hits=0, misses=1, recorded=1, mode='record')

Commit the .tapes/*.json file next to your tests. Delete it and re-run once (with a real API key available) whenever you need to re-record.

API

tape(*, dir=".tapes", name="default", mode="auto", normalize=None) -> Tape

• dir (str) - cassette directory. Default ".tapes".
• name (str) - cassette file basename. File is <dir>/<name>.json. Default "default".
• mode ("auto" | "record" | "replay" | "off") - default "auto".
  • "auto": resolved once, at tape() construction, to "replay" if the cassette file already exists on disk, otherwise to "record". This is a one-time decision for the whole Tape instance, not re-checked on every call.
  • "record": always call through to the wrapped function and (over)write the cassette on save().
  • "replay": never call through. A lookup miss raises a CassetteError naming the missing key and the cassette path.
  • "off": pass every call straight through, record nothing, save() is a no-op.
  • If the CASSETTE_FN_MODE environment variable is set, it overrides mode entirely. This is the same environment variable name the JavaScript version reads, so both implementations honor one setting.
• normalize (Callable[[list, dict], tuple]) - applied to a call's positional argument list and keyword argument dict before they are hashed into a lookup key, so callers can strip volatile fields (timestamps, request ids, an API key) that would otherwise make every call miss. It receives (args, kwargs) and must return (normalized_args, normalized_kwargs). Default: identity, i.e. lambda args, kwargs: (args, kwargs).

Returns a Tape.

Tape.wrap(fn) -> Callable

Wraps any ordinary synchronous callable, called with any mix of positional and keyword arguments (e.g. wrapped("say hi", model="claude-opus-5", max_tokens=1024)), since essentially every real LLM SDK call is keyword-based. Each call:

1. Computes normalized_args, normalized_kwargs = normalize(args, kwargs), then:
   • if there are no keyword arguments, key = sha256(json.dumps(normalized_args, separators=(",", ":"), ensure_ascii=False)).hexdigest()[:16]. This is exactly the JavaScript implementation's key computation for the same positional arguments, so a cassette recorded by one implementation replays in the other (see "How it works").
   • if there are keyword arguments, they are folded in with their keys sorted so that argument order never affects the key: key = sha256(json.dumps([normalized_args, sorted_kwargs], separators=(",", ":"), ensure_ascii=False)).hexdigest()[:16]. f(a=1, b=2) and f(b=2, a=1) produce the same key.
2. In replay mode: looks up key, returns a deep copy of the stored result (so a caller mutating the returned value can never corrupt the cassette). Repeated identical calls replay the recorded entries for that key in the order they were recorded; once the list is exhausted, further calls keep repeating the last entry. A recorded thrown call is replayed as a freshly constructed RuntimeError with the same message (see "How it works" for why RuntimeError and not the original exception type). A missing key raises CassetteError immediately, naming the key and the cassette path.
3. In record mode: calls fn(*args, **kwargs) for real and appends {"args": ..., "kwargs": ..., "result": ...} (or, on a raised exception, {"args": ..., "kwargs": ..., "error": {"message": ..., "name": ...}}) to the ordered list under key, then returns (or re-raises) the real outcome. The "kwargs" key is omitted entirely when the call had no keyword arguments, so a purely positional call produces the same entry shape as the JavaScript version's cassettes. The recorded args, kwargs, and result are all deep-copied at record time, so mutating the caller's argument objects (including a messages=[...] list) after the call, or the returned result, can never corrupt the cassette.
4. In "off" mode: calls fn(*args, **kwargs) directly, no key is computed, nothing is recorded.

Tape.save() -> None

Writes the cassette to disk, creating dir recursively if needed. No-op in "replay" and "off" mode.

Tape.stats() -> Stats

A frozen dataclass with fields:

• hits (int) - calls served from the cassette.
• misses (int) - calls not found in the cassette: every record-mode call (nothing is ever read back from the cache in record mode), plus replay-mode lookup misses. A replay miss increments misses before it raises, so a caller that catches the error still sees an accurate count.
• recorded (int) - entries written into the in-memory cassette during this session (always 0 in replay and off mode, since neither writes new entries).
• mode (str) - the resolved runtime mode ("record" | "replay" | "off"), never the literal "auto" input, since the whole point of resolving it once is to know which concrete behavior the instance is running.

Tape.keys() -> list[str]

Keys currently present in the in-memory cassette (loaded entries in replay mode, accumulated entries so far in record mode).

Cassette file format

For a purely positional call, unchanged from the JavaScript version, so a cassette file recorded by either implementation can be replayed by the other. Human-diffable on purpose, so cassettes commit cleanly to git:

{
  "version": 1,
  "name": "my-test",
  "entries": {
    "<key>": [ { "args": [ "..." ], "result": { "...": "..." } } ]
  }
}

An entry for a call made with keyword arguments carries a "kwargs" object as well:

{
  "args": [ "say hi" ],
  "kwargs": { "model": "claude-opus-5", "max_tokens": 1024 },
  "result": { "...": "..." }
}

The "kwargs" key is only present when the call actually had keyword arguments; an older entry (or one recorded from a purely positional call) has no "kwargs" key at all, and still loads and replays correctly. An entry for a call that raised looks like {"args": [...], "error": {"message": "...", "name": "..."}} (with "kwargs" added the same way, when present) instead of "result". Written with json.dump(data, indent=2, ensure_ascii=False).

How it works

cassette_fn wraps one function you choose, the one that actually makes a network call to a provider. It hashes the (optionally normalized) argument list into a short key, and stores or replays call outcomes under that key in a plain JSON file. Everything that calls the wrapped function is real code running for real on every test run; only the wrapped function itself is faked on replay.

The key is computed with json.dumps(..., separators=(",", ":"), ensure_ascii=False) rather than the default json.dumps(...), for two reasons that both matter for cross-implementation cassette portability:

• The default separators insert a space after , and :; JavaScript's JSON.stringify never does, so the default would produce a different serialized string, and therefore a different key, for the same arguments.
• json.dumps defaults to ensure_ascii=True, which escapes every non-ASCII codepoint (\uXXXX); JSON.stringify never escapes non-ASCII. Left at the default, any argument containing an accent, a curly quote, CJK text, or an emoji, which is the common case for real LLM prompts, would silently hash to a different key in Python than in JavaScript. ensure_ascii=False fixes this; it has no effect on ASCII-only arguments, since there is nothing to escape either way.

Both were verified by hand, not just read from the code: a cassette recorded with the real JavaScript implementation for a plain ASCII call was replayed successfully with this Python implementation, and separately a cassette recorded for a call containing accented Latin characters, a punctuation dash, CJK text, and an emoji was also replayed successfully, in both cases a hit (not a miss), confirming the two implementations produce identical keys for identical arguments, ASCII or not. The same construction extends naturally to keyword arguments; see Tape.wrap above.

A recorded exception is reconstructed on replay as a Python RuntimeError whose message matches the original. The original exception's type name is preserved as a .name attribute on the reconstructed RuntimeError (mirroring how the JavaScript version sets err.name), since Python has no general way to reconstruct an arbitrary exception class from a stored type name the way JavaScript can construct a plain Error and overwrite its .name. This is the one spot where this port intentionally does not reproduce the original exception type, only its message and recorded name.

Errors raised by cassette_fn itself (an invalid mode, a missing or unparseable cassette file, a malformed cassette structure, a replay lookup miss) are raised as CassetteError, a RuntimeError subclass, so except RuntimeError catches both cassette_fn's own errors and a reconstructed recorded exception.

Limits, honestly:

• The replay match key is the full argument list (after normalize). Two calls that differ only in a field you did not strip via normalize will not match, and you'll get a new recording ("record"/"auto") or a miss ("replay").
• "auto" mode's replay-vs-record decision is made once, at construction, from whether the cassette file exists at that moment. It does not re-check mid-run, and it does not merge new calls into an existing cassette; re-recording means deleting the file (or using mode="record") and running again.
• No HTTP interception and no provider SDK integration. You choose the function boundary; if you wrap too high (e.g. a whole multi-step agent function) you lose the fine-grained determinism this library is for, and if you wrap too low you may end up wrapping something that is not the actual network call.
• Unlike the JavaScript version, which wraps async functions, this port wraps ordinary synchronous callables. Async support is intentionally not provided; if you need to record around an async call site, call the async function yourself and wrap a synchronous function that drives it (e.g. one that calls asyncio.run(...) internally), or wrap only the synchronous part of the call that actually needs recording.
• Tape.wrap expects a plain callable that returns once with a full result. Wrapping a function that returns a generator or other lazy/streaming object will record the object itself, not its eventual contents, and will not behave correctly on replay.

Related

The JavaScript version of this package lives at the repository root: cassette-fn (JavaScript).

License

MIT