feat(provider/aimlapi): add aimlapi.com LLM support

[Lookoff-AIMLAPI]

What type of change does this PR introduce?

• Bugfix
• Feature
• Refactor
• Documentation
• Not Sure?

Does this PR introduce breaking changes?

• Yes
• No

List any relevant issue numbers:

None. Shape copied from leon-ai#579 and leon-ai#580 (the MiniMax provider)
and from the Celeris
provider, which is the most recently added one.

Description:

Adds aimlapi.com as a remote LLM provider. It is an aggregator — one
OpenAI-compatible Chat Completions endpoint in front of many upstream vendors —
so a single key covers the OpenAI, Anthropic, Google, DeepSeek and Moonshot
families a user would otherwise configure one at a time.

The provider itself is a thin subclass of AISDKRemoteLLMProvider with
flavor: 'openai-compatible'. No base-class change was needed.

Two commits, deliberately separated:

Commit	Contents
18eb95a feat(provider/aimlapi): add aimlapi.com LLM support	the provider, config, schema, samples, catalog entries, tests
6488453 chore(aimlapi): fork-only placement — do not send upstream	list ordering only — drop this commit before offering the provider upstream

Gateway behaviour that shaped the code

1. The gateway returns HTTP 400 for a null-valued optional field that the
upstream OpenAI API accepts. Verified today against
POST https://api.aimlapi.com/v1/chat/completions:

null → 400	null → 200
temperature, top_p, seed, reasoning_effort, tools, tool_choice, response_format, stream, stream_options, parallel_tool_calls, max_tokens, max_completion_tokens	stop, presence_penalty, frequency_penalty, n, user, logprobs, logit_bias, verbosity

An OpenAI SDK that serialises an unset optional parameter as literal null
therefore fails on every real call while a mocked test suite stays green.
Leon's base class already omits unset options (typeof x === 'number' ? {...} : {}), so this PR does not change that behaviour — it pins it. The captured
request body for a default call is exactly:

{"model":"openai/gpt-5.6-sol","messages":[{"role":"system","content":"…"},{"role":"user","content":"…"}]}

test/agent/unit/aimlapi-llm-provider.spec.ts asserts no top-level field is
null, and separately that temperature, seed and max_tokens are
forwarded when they are set — so a future refactor to x ?? null fails here
instead of in production.

2. The reasoning switch differs per upstream vendor. As with Celeris, the
empty buildProviderOptions stops the generic compatible adapter from attaching
a vendor field this gateway would reject, and the catalog entries expose auto
reasoning only.

Model ids

Every id was checked against the live catalog
(GET https://api.aimlapi.com/v1/models?include=all, filtered to
type == "openai/chat-completions"; 936 rows / 785 distinct ids / 353 chat
models) and made to answer a real completion. Nothing was copied from another
aggregator's list.

The dotted spelling is used deliberately: anthropic/claude-sonnet-4-6,
claude-opus-4-7 and claude-opus-4-8 (dashed) publish only streaming in
capabilities, while the identical dotted ids publish the full set.

Catalog id	HTTP	model echoed back
openai/gpt-5.6-sol	200	gpt-5.6-sol (vendor prefix dropped)
anthropic/claude-opus-4.8	200	anthropic/claude-opus-4.8
anthropic/claude-sonnet-4.6	200	anthropic/claude-sonnet-4.6
google/gemini-3.8-flash	200	google/gemini-3.8-flash
deepseek/deepseek-v4-pro	200	deepseek-v4-pro-202606 (dated snapshot)
moonshot/kimi-k3	200	kimi-k3 (vendor prefix dropped)

Three ids echo back a name that is not byte-identical to the request. Leon does
not record or pin "the model that actually served the request", so this is
cosmetic here — noted because it would matter for a host that does. No existing
id was removed by this PR, so the "never delete on catalog absence alone" rule
did not come into play.

Attribution

AISDKRemoteLLMProvider already exposes a headers hook, so this uses the
mechanism the repo has rather than adding one. HTTP-Referer and X-Title
follow the OpenRouter convention and identify Leon, not the gateway:

HTTP-Referer:         https://github.com/leon-ai/leon
X-Title:              Leon
X-AIMLAPI-Source:     agent/leon
X-AIMLAPI-Partner-ID: part_leon

The Base URL is user-configurable, so the headers are scoped to the official
host: buildAIMLAPIHeaders() returns {} for any other hostname, so a
self-hosted gateway or a third-party proxy fronting the same schema never
receives them. A fresh object is returned per call, and a test mutates the
returned object to prove the module constant cannot be modified through it.
Another test asserts the partner id matches ^part_[A-Za-z0-9]{1,64}$ — a
malformed id is not rejected at request time, it just silently stops counting,
so an assertion is the only thing that catches a typo.

Fork-only placement commit (6488453)

Both lists it touches are hand-ordered and both feed the first choice a new user
sees: getLLMModelCatalogProviders() derives the setup wizard's provider order
from the catalog's array order, and LLM_PROVIDER_ACCOUNT_CONFIGS is the order
used wherever providers are listed. The commit moves aimlapi.com to the front of
each and changes nothing else.

Leon already has a featured concept — recommended, used exactly once per
provider — so openai/gpt-5.6-sol carries it in the provider commit like every
other provider's lead model. No badge mechanism was invented.

Before / after getLLMModelCatalogProviders():

celeris, openrouter, openai, anthropic, zai, minimax, moonshotai
aimlapi, celeris, openrouter, openai, anthropic, zai, minimax, moonshotai

Checks

Run on Node 26.7.0 / pnpm 11.20.0.

• pnpm exec tsc --noEmit --pretty false — exit 0
• pnpm exec eslint on all 14 changed files — exit 0, no output
• pnpm run build:server — exit 0 (build:app builds the Electron desktop
  shell and was not run; nothing in this PR touches it)
• pnpm run test:agent:unit — 76 passed (baseline 70; +6 from the new spec)
• pnpm run test:core:unit — 71 passed / 1 failed, identical to a pristine
  checkout. The failure is network-helper.spec.ts > retries only the stalled remainder of a parallel range (expected […(8)] to have a length of 6), which
  is timing-dependent: it passed on the first pristine run and failed on every
  subsequent run, pristine tree included. Unrelated to this change.
• pnpm run test:controlled:unit — 7 passed (baseline 7)
• pnpm run test:controlled:e2e — 1 failed, identical to a pristine
  checkout: ENOENT … /.leon/tmp/controlled-skill-smoke-*.json, an environment
  prerequisite of the full postinstall setup, which was skipped here.

Real inference, through the provider

Not a mock and not a raw curl — AIMLAPILLMProvider.runChatCompletion() was
called directly with a real key in LEON_AIMLAPI_API_KEY.

aimlapi/openai/gpt-5.6-sol, plain completion:

{"choices":[{"finish_reason":"stop","message":{"content":"LEON AIMLAPI OK"}}],
 "usage":{"prompt_tokens":29,"completion_tokens":9}}

aimlapi/openai/gpt-5.6-sol, tool calling:

{"choices":[{"finish_reason":"tool-calls","message":{"content":"","tool_calls":[
  {"id":"call_qikIdtV1kLkiXK8LP3f1Ln0e","type":"function",
   "function":{"name":"get_weather","arguments":"{\"city\":\"Paris\"}"}}]}}],
 "usage":{"prompt_tokens":151,"completion_tokens":17}}

aimlapi/anthropic/claude-sonnet-4.6, plain completion — confirming the
aggregator path works across vendors:

{"choices":[{"finish_reason":"stop","message":{"content":"LEON AIMLAPI OK"}}],
 "usage":{"prompt_tokens":28,"completion_tokens":10}}

Not verified

• The setup wizard was not driven interactively; its provider list, model list
  and "(Recommended)" suffix were verified by evaluating the exact functions it
  reads (getLLMModelCatalogProviders, getLLMModelCatalogEntries,
  getLLMProviderAccountConfig).
• pnpm run test:agent:e2e requires the full postinstall environment and live
  keys for every provider in the matrix; the aimlapi row was added to
  PROVIDER_MATRIX but that suite was not run.
• Streaming (shouldStream: true) was exercised only through the existing unit
  coverage, not against the live endpoint.

One repo rule worth flagging before this goes upstream

scripts/commit-msg.js:23 enumerates every valid provider scope. aimlapi is
added to that alternation in the provider commit, exactly as leon-ai#579
did for
minimax, so feat(provider/aimlapi): … passes the hook. The placement
commit's scope, chore(aimlapi), is not in the enum and would be rejected by
that hook — which is harmless, since that commit is meant to be dropped before
the provider change is offered upstream.