test(image_gen): pin the availability check to the dispatch chain

[juniperbevensee]

Tests only. No source change. The behaviour is already correct on main (fixed in e13104cfd). Revised after an independent audit (9 agents) found two false statements in my own docstring and one real coverage gap.

Why this exists

check_image_generation_requirements() gates whether image_generate is emitted into the tool schema (tools/registry.py:551-557), while _dispatch_to_plugin_provider() decides where a call goes. If the check is more permissive than the dispatch, the agent advertises a capability that fails on every call — a false claim about its own abilities.

They disagreed before e13104cfd: the check returned True if any registered provider was available, while dispatch falls through to the in-tree FAL path unless image_gen.provider is explicitly set (deliberately — an OPENAI_API_KEY present for other features must not silently opt you into a paid image backend).

There was a test — and it passes against the buggy code

My original claim of "zero tests" was wrong. tests/tools/test_image_generation_plugin_dispatch.py::test_requirements_ignore_unselected_paid_plugin already called this function. What makes it worth naming is that it is vacuous: it stubs check_fal_api_key and _read_configured_image_provider but never stubs _ensure_plugins_discovered and never registers a ready provider. In a keyless environment all eight real providers report is_available() == False, so the old any-provider body returned False too.

Verified by splicing in the genuine pre-e13104cfd body:

file	result against the buggy body
test_image_generation_plugin_dispatch.py	6 passed — cannot catch it
test_image_gen_availability.py (this PR)	4 failed — catches it

A test that cannot fail is not coverage. That's the stronger version of the point I was originally making.

The coverage gap the audit found

My docstring claimed these tests pinned the configured == "fal" short-circuit. They didn't. Three mutants on the one line e13104cfd actually added survived a fully green suite. Added test_configured_fal_short_circuits_even_with_a_ready_fal_provider, which kills two of them.

The third — narrowing the guard to if configured == "fal" alone — is not a gap: get_provider returns None for any non-str name (agent/image_gen_registry.py:69-70), so an unset provider can never resolve and the not configured half cannot change the return value. Redundant defence, indistinguishable by any test. Documented in the file so nobody "closes" it with an assertion that asserts nothing.

Mutation results — all six re-verified directly

mutation
delete the short-circuit	KILLED
drop == "fal"	KILLED
get_provider(configured) → any-provider scan	KILLED
drop .is_available()	KILLED
drop the _load_fal_client SDK probe	KILLED
FAL branch unconditional	KILLED
drop not configured	survives by construction — see above

Order-dependence fixed

tests/agent/test_empty_tool_name_loop_dampening.py:136-141 purges agent.*/tools.* from sys.modules without restoring, so module-level imports here bound to a different module object than the one the function re-imports at call time — and a fake subclassing the stale ABC failed register_provider's isinstance check. Modules and the fake are now resolved together at call time.

Before: 1 failed when run after that file. After: 18 passed. CI never saw it, because scripts/run_tests.sh spawns a subprocess per file — which is exactly why it would have sat there indefinitely.

Coverage

scenario	expected
keyed provider, image_gen.provider unset	False ← the regression
provider = "fal" + a ready plugin named fal	False ← the short-circuit
selected + available	True
selected but unavailable / not registered	False
a different provider ready	False
FAL key alone, no plugin, no provider set	True ← must not narrow
FAL key but fal-client absent	False ← falls through

121 passed across the four image-related files. Verified live on the agent host earlier: check_image_generation_requirements() returns False there with OPENROUTER_API_KEY set and no FAL key — Matilde was not misadvertising the tool.

Not addressed here

• tests/tools/test_image_generation_plugin_dispatch.py:117 — the vacuous neighbour still needs a discovery stub and a registered fake, or deletion.
• agent/learn_prompt.py:77 tells the model to reference image_generate in generated SKILL.md with no availability gate.
• tools/registry.py's last-good grace keeps a tool advertised briefly after its backend dies.

🤖 Generated with Claude Code