Make POST /__aimock/reset the canonical full reset and converge llmock.reset()

[jpr5]

What

POST /__aimock/reset/fixtures was never a fixtures reset. It clears the fixtures array, the journal (entries and per-test fixture match-counts, across every testId), videoStates, openRouterVideoJobs, veoVideoJobs, grokVideoJobs, falJobs, falQueueStates, the Gemini interaction and event-id counters, and re-zeroes the aimock_fixtures_loaded gauge. POST /__aimock/reset does exactly the same thing — and it was the one marked deprecated. The deprecation pointed at the dishonest name.

This inverts it and converges the in-process path onto the same implementation.

• POST /__aimock/reset is canonical: plain { "reset": true }, no Deprecation header, no deprecated/deprecation body fields, no log warning.
• POST /__aimock/reset/fixtures is the deprecated alias, carrying the identical deprecation mechanics the old route used.
• performFixturesReset → performFullReset, exported and shared.
• /reset/journal and DELETE /fixtures are untouched.
• aimock-pytest's reset() and reset_fixtures() both target the canonical route.

LLMock.reset() was not equivalent to the HTTP full reset — it left veoVideoJobs, grokVideoJobs and both Gemini counters populated. It now calls the same performFullReset, so the two cannot drift again.

One deliberate, commented difference: reset() also clears the search / rerank / moderation fixtures. Those are registrable only through LLMock; the control API has no route that creates them.

Compatibility

/reset/fixtures reset semantics are unchanged — it clears exactly what it always did, so existing callers keep working. Its response body is additively extended: deprecated and deprecation now accompany reset, plus a Deprecation: true header. A caller doing strict equality on the old { "reset": true } body must relax that assertion; a caller reading body.reset is unaffected.

LLMock.reset() clears state that is module-global, not per-instance — the Gemini counters and the fal.ai job/queue maps. With two LLMock instances live in one process, a.reset() rewinds the Gemini id sequence b is mid-way through, so b re-emits aimock-int-0/evt_1, ids it already handed out. The fal maps were already global; the counters are newly reached. Documented in the reset() docblock and the CHANGELOG.

Release follow-up (required, not done here)

aimock-pytest calls POST /__aimock/reset, but _version.py pins AIMOCK_VERSION = "1.38.0" — the published, pre-fix server, where /reset is the deprecated alias. The client therefore trips a deprecation on every reset until the pin moves. Bump AIMOCK_VERSION to the first npm release containing this change before publishing the next aimock-pytest. The Python suite cannot catch this on its own because tests/conftest.py redirects AIMOCK_CLI_PATH to the local build.

Coverage

Every store performFullReset touches has a guard, and every guard was verified to fail under a mutation that removes exactly what it claims to protect — then reverted and re-confirmed green.

Store / behavior	Mutation	Verbatim RED
/reset carries no deprecation	re-add header + body fields	expected 'true' to be undefined
/reset/fixtures signals deprecation	strip the signal	expected undefined to be 'true'
deprecation message direction	point the message back at itself	Expected: "…use POST /__aimock/reset (full reset)…" / Received: "…migrate to POST /__aimock/reset/fixtures…"
log warning, both directions	add a warn to canonical /reset	expected "warn" to not be called at all, but actually been called 1 times
grokVideoJobs	drop .clear()	expected 200 to be 404
veoVideoJobs	drop .clear()	expected 200 to be 404
videoStates	drop .clear()	expected 200 to be 404
falJobs	drop .clear()	expected 200 to be 404
falQueueStates	drop .clear()	expected 200 to be 404
Gemini interaction counter	drop resetInteractionCounter()	expected 'aimock-int-2' to be 'aimock-int-0'
Gemini event-id counter	drop resetEventIdCounter()	expected 'evt_6' to be 'evt_1'
fixture match-counts, all testIds	clearEntries() + clearMatchCounts("__default__")	expected 2 to be +0
search/rerank/moderation clear	delete the three lines	expected [ { title: 'Weather Report', …(2) } ] to have a length of +0 but got 1
pre-start (targets === null) path	move the global clears below the early return	expected 200 to be 404
aimock_fixtures_loaded re-zero	drop the setGauge	expected '# TYPE aimock_fixtures_loaded gauge\n…' to contain 'aimock_fixtures_loaded{} 0'
Python client's route choice	point reset() back at /reset/fixtures	assert 'http://…/__aimock/reset/fixtures'.endswith('/__aimock/reset')

The match-count guard drives both the default scope and x-test-id: tenant-a, using two sequenceIndex siblings and re-adding the same fixture objects after the reset (counts are keyed by object identity). A default-only clear leaves every other tenant parked mid-sequence, so it is asserted explicitly.

The Python guards spy on requests.post, let the real call go out, and assert on the server's own deprecation signal — the only thing that distinguishes two routes with identical reset semantics. They anchor that discriminator by asserting the alias does signal deprecation, keyed on use POST /__aimock/reset (full reset), the half of the message a self-referential variant would not contain.

Suites

• Node, pnpm build && pnpm test: 172 files, 5079 passed, 0 failed, 0 skipped. The build matters — without it cli.test.ts is describe.skipIf(!CLI_AVAILABLE) and 45 tests silently skip.
• Python, pytest tests/ -q against the locally built dist/cli.js: 22 passed, 0 failed, 0 skipped, on CPython 3.12.13 (bare python3 on the dev box is 3.9.6, below requires-python = ">=3.10"). Installed editable so the tests exercise the edited source.
• prettier --check, eslint, tsc --noEmit, commitlint all clean.
• pnpm test:drift not run — it exercises live provider APIs and is unrelated to this change.

Docs

The control-API page states what each route clears, and corrects a claim with a spend consequence: a full reset does not universally yield no_fixture_match. In replay mode the next request fails 404 and in strict mode 503 (both carrying code: "no_fixture_match"), but in record mode with a provider key configured it is proxied to the real provider — live upstream calls and real spend, not an error.

Deliberately not done

• aimock-pytest's reset_fixtures() is as dishonestly named as the route was — it is a full reset. Kept as an alias for reset(); renaming a published Python API does not belong in this change.
• ~18 prose occurrences of "a fixtures reset" remain across the video handlers and router.ts. Several are log strings asserted verbatim by tests, so rewording them is a behavior-adjacent change for its own PR. The phrase describes the concept accurately; only the dangling performFixturesReset symbol reference was a real defect, and that is fixed.

🤖 Generated with Claude Code

https://claude.ai/code/session_017ATs7KdGhg9zhE1AkEMmo6

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@copilotkit/aimock@358

commit: 1101d2b