Add AIHubMix sync provider

[0genlab]

Adds a fully interface-driven sync adapter for AIHubMix, an aggregator that
relays ~408 LLM routes from 26 model labs. Everything is read from the public
catalog endpoint — no credentials, no per-model overrides in the adapter.

GET https://aihubmix.com/api/v1/models?type=llm

This PR ships the adapter only

An earlier revision carried the 299 generated providers/aihubmix/models/*.toml
files alongside the adapter. That pushed the diff past GitHub's 300-file ceiling,
gh pr diff --patch answered HTTP 406, and pr-reviewer.yml died in its
context step before the agent ever started — so no revision of the branch could
reach a review at all.

The catalog files are now dropped from the PR. Once the adapter merges,
sync-models.yml regenerates them on automation/sync-models-aihubmix, a branch
the reviewer workflow already skips by design. The diff here is 22 files — 5 source, 8 note moves, 8 alias deletions, and sync.md.

Answers to the review findings

toggle alongside an effort list containing none is intentional, not a bug

The reviewer flagged models such as glm-5.2 that publish both
{type = "toggle"} and {type = "effort", values = [..., "none", ...]} as
"contradictory metadata" that the adapter should reconcile. It is neither
contradictory nor a defect in the source — it is the gateway's design, and this
repo already documents it.

AIHubMix is a multi-dialect gateway: the same route is reachable over the
OpenAI-compatible /v1/chat/completions path, the Anthropic-compatible
/v1/messages path, and the Gemini native path. Each dialect has its own
off-switch, and the gateway maps whichever one the caller sends onto the
vendor's real control instead of rejecting it. providers/aihubmix/provider.toml
on dev already spells all four out:

• Raw Chat — $.reasoning_effort = "none"|"minimal"|…|"xhigh"; "none" disables
• Raw Messages — $.thinking.type = "enabled"|"disabled"|"adaptive"
• Raw Gemini — $.generationConfig.thinkingConfig, thinkingBudget: 0 off
• Raw Responses — $.reasoning.effort

So one model genuinely has two independent ways to turn thinking off, which is
exactly toggle + an effort list containing none. Verified live against
glm-5.2 on 2026-09-11 with a fixed prompt:

request	reasoning chars	result
reasoning_effort: "none"	0	off
enable_thinking: false	0	off
both together	0	coherent, no error
reasoning_effort: "minimal"	217	tier live
reasoning_effort: "max"	328	tier live
reasoning_effort: "turbo" (bogus)	302	HTTP 200, mapped — not 400

That last row is the point: an unrecognised effort is mapped to the nearest
supported tier rather than erroring, because callers arrive from many SDKs with
many vocabularies. Collapsing the two signals in the adapter would delete a real,
independently reachable control from the catalog.

(An earlier comment of mine on this PR said this shape "is being fixed at the
source". That was my own inference and it was wrong; it is documented, intended
behaviour. Corrected publicly in a later comment.)

Wire path for the toggle — fixed

The reviewer was right that {type = "toggle"} on its own does not tell a caller
which field to send, and that a sync rewrites the file whole and would erase a
hand-written header. translateModel now emits the header itself whenever the
model carries a toggle — the same shape the OpenRouter adapter uses — so the wire
path survives every rewrite instead of depending on a human re-adding it.

Intersecting a relay's options with its lab/peer baseline — declined

The suggestion was to narrow a relay's advertised tiers against what its lab
model or sibling routes advertise. That would put per-model knowledge back into
the adapter — the exact thing this rewrite removed (the developer_id → lab
table and the prefix/suffix strip list). The gateway's tier support is a property
of the gateway, not of the lab: AIHubMix deliberately accepts tiers the upstream
vendor does not name and maps them. An intersection would silently drop tiers
that demonstrably work. Where the endpoint is wrong, the fix belongs on the
endpoint; two such fixes have already shipped there during this PR.

What the interface-driven rewrite removed

The adapter originally landed 232 models, resolving relays to lab metadata with
a hand-maintained developer_id → lab table plus a list of routing prefixes and
suffixes to strip. Both are gone. AIHubMix now serves vendor, variant_of
and open_weights on the endpoint itself, so nothing about a relay is inferred
from its ID or mirrored in this repo any more — the adapter reads the catalog's
own answer. That change is what brings the catalog to 308.

	first revision	now
aihubmix models in the built catalog	232	308
entries using base_model	217	273
standalone full definitions	15	35
lab mapping	21 hand-maintained developer_id entries	vendor from the endpoint
relay → base matching	16 prefixes + 10 suffixes stripped from the ID	variant_of from the endpoint
open_weights	not served; every standalone create blocked	served for 289 of 408

How models are matched to lab metadata

The endpoint answers both halves of the lookup itself:

• vendor names the lab that built the model (292 of 408 routes).
• variant_of names the AIHubMix ID this entry is a routing variant of, and
  variant_kind labels it (pricing-tier 26, channel-tier 24,
  mode-preset 18, deprecated-alias 8).

A relay is looked up under its own ID first, then under each variant_of hop,
nearest first. Nearest-first matters: qwen3.8-max-preview is declared a
variant of qwen3.8-max and both are published lab models, so the relay factors
onto the preview it actually serves rather than onto the root of its chain.

Following a declared chain also reaches relays no string rule could have:

relay	resolves to	why a prefix/suffix rule can't
ox-alpha	zhipuai/glm-5.3-flash	codename shares no substring with the model
grok-code-fast-1	xai/grok-build-0.1	renamed upstream
cohere-command-a	cohere/command-a-03-2025	the date tag is the lab's, not the relay's
grok-4-fast-reasoning	xai/grok-4.3	deprecated-alias: the old ID now routes to 4.3

The one thing left in the adapter is VENDOR_LABS, four entries for the labs
the two registries spell differently (zhipu/zhipuai, moonshot/moonshotai,
bytedance/bytedance-seed, meituan-longcat/meituan). It maps namespaces;
no entry in it decides what a model is or which lab built it.

Lookups stay case-folded: AIHubMix lowercases every relay ID while labs keep
their own casing, so minimax-m2 has to reach minimax/MiniMax-M2.

Dated release tags are deliberately left on. gemini-2.5-pro-preview-06-05 is a
pinned snapshot, not google/gemini-2.5-pro — and the endpoint does not declare
it a variant of one, which is now the actual test rather than a judgement call
in the adapter.

A resolved base means the entry inherits open_weights, knowledge, license
and the rest from models/, and records only what actually differs — which is
why 273 of 308 entries are three or four lines long.

Standalone entries are skipped, not guessed

A full catalog entry has to carry release_date, open_weights and both limits
itself. The adapter skips anything that has neither lab metadata to inherit
from nor those fields, and reports it through skippedNotice rather than
inventing a value.

open_weights used to block every standalone create because the endpoint did
not serve it. It now does, for 289 of 408 routes, which is what lets the adapter
create standalone entries at all. The guard also checks limits now: the endpoint
sends max_output: 0 for a ceiling it does not know, and without that check
those models reach the writer and fail ModelBase validation.

What is still uncovered, entirely on the endpoint side:

• 116 routes carry no vendor, so they cannot be factored
• 104 date to no release_date
• 119 state no open_weights
• 104 quote max_output: 0

Every one of these shrinks as AIHubMix backfills its own catalog. The adapter
needs no change for any of them.

Sentinel values in the source data

Two signals mean "unknown", not "zero" or "huge", and are read as absent so the
field falls through to the base model:

signal	count	why
max_output: 0	104	a servable model cannot emit zero tokens
max_output >= context_length	36	the context window quoted a second time, leaving no room for the prompt

No value is invented — the field is simply omitted and inherited.

Reasoning options

The endpoint carries an extra default key inside reasoning_options[] that
ReasoningOption does not accept, and two non-standard effort spellings,
no_think and instant. The adapter drops the unknown key and maps the
spellings onto none / minimal, filtering anything else outside
ReasoningEffortValue. These are shape-level normalisations — vocabulary
mapping, not per-model judgement — and all four are reported upstream to
AIHubMix.

Deletion policy

deleteMissing is false and trackMissingModels is true. AIHubMix rotates
routes in and out; a transient absence should not delete a catalog entry. The
sourceID hook returns undefined for retire_stage === "deprecated" so
retired routes are marked, not removed.

Framework changes

Four, all small and all needed by something above:

• formatToml now emits input_audio / output_audio inside cost tiers. Both
  are already in Cost, but were silently dropped when writing tiered pricing.
• translateModel's options gain header(id), the leading comment block already
  on the file. authoritativeHeaders replaces that block wholesale, so without
  reading it first a provider that owns its header cannot refresh the part it
  generates without discarding the notes a human wrote around it.
• trackMissingModels is now settable on its own. It was only ever read
  alongside skipCreates, so a provider that creates most models but skips the
  ones it cannot write produced notices nobody acts on. The missing-model issue
  text is reworded to cover that second reading.
• normalizeModelSlug is exported from openrouter.ts, which already serves as
  the shared helper module for the other provider adapters. Two more adapters
  (venice.ts, wandb.ts) hold byte-identical private copies; left alone here.

Limits are resolved against the lab window, not just accepted

Three rules, in this order. Each was written because the previous revision got a
real route wrong.

A restatement is a ratio, not a direction. The endpoint quotes an accepted
204800 as 200000 and an accepted 1000000 as 1048576. Neither is the host stating
a different window. A decimal restatement of a binary window loses at most
1000/1024 per K unit, so 1000³/1024³ — three nested unit swaps — is the floor
of what a restatement can explain, and two limits within that floor of each other
are one window spelled twice. Checking only the narrowing side left 20 routes
writing an override that states no difference at all (glm-5.3 recording 1048576
against a lab window of 1000000).

The lab entry is the spelling a restatement resolves to, and the provider
file's own value only where the lab states no such key. Matching the lab lets
factoring drop the key entirely; resolving to whatever the file happens to hold
would pin that spelling forever — qwen3.7-flash carries 991000, which is only
an imprecise way of writing the lab's 1000000. Resolving file-first preserved 10
such overrides, claude-opus-4-8 among them still recording 200000/32000 against
a 1000000/128000 lab window. That one is the finding from review: the endpoint
agrees with the lab there, and the narrowing number was the file's.

Then clamp to the lab's window — a relay cannot serve a wider one than the
model it relays; the window is the model's property and a host can only restrict
it. The clamp runs after the resolution rather than instead of it, because an
endpoint quoting back the same stale ceiling the file already holds resolves to
that number and only a later clamp catches it. grok-4.5 is exactly this: the
file held 1000000 for both limits against a 500000 lab window, and the endpoint
quotes that same 1000000.

Past the two max_output sentinels, 14 routes quote a window wider than their
lab entry's (qwen3.8-2.4t-a95b at 1000000 of 262144, gemma-4-31b-it at 131100
of 32768). All are reported upstream to AIHubMix; where the lab entry is the
stale side, models/ is where that gets corrected. After all three rules, the
factored files write 7 limit keys total — every one a genuine host cap, none
byte-identical to its lab entry.

Headers keep human notes and never outlive their block

authoritativeHeaders replaces the leading block on every sync, so what counts
as "the block" has to be exact in both directions.

• A superseded wire path is told from a note to keep by what the line opens
  with (# Toggle:, # Effort:, # Budget:, # Off is effort — the openings
  AGENTS.md prescribes), not by whether it mentions a field path or the docs
  host. Keying on the substring would also delete lines that merely contain one:
  two files state a wire path together with a dated live test the response cannot
  reproduce. A second statement of the same path costs nothing; a deleted
  verification date cannot be recovered. Matched on the trimmed line, or an
  indented # Toggle: outlives the block it documented.
• The two lines naming this gateway's wire paths are the adapter's own
  restatement of the block, so they go whether or not a block replaces them. Only
  the docs link survives as a note, and only where no derived block restates it.
  Keeping the wire paths when nothing was derived is what left a route
  advertising a toggle it no longer had — the block vanished, its tail survived
  as a "note", and no later sync could tell the difference, so the file never
  self-corrected.
• AGENTS.md requires these notes above the first key, since a sync keeps only
  the leading block and drops every comment below it. Eight files in this
  provider carried theirs mid-body; they are moved up here, which is 8 of the 22
  files in this diff. Without the move the next sync deletes them silently.

Display names

The endpoint's label is recorded only where the relay is not that lab model under
other punctuation — its bare ID, normalised, differs from the base model's
slug. Bare, because that is what resolved the base model in the first place
(Qwen/QwQ-32B reaches qwen/qwq-32b); normalising the namespaced form matches
nothing, and each of the 10 namespaced routes would take a redundant storefront
override the moment its lab file lands.

This keeps coding-glm-4.6-free reading "Coding GLM 4.6 (free)" rather than the
bare "GLM-4.6" it would share with two other routes, while entries differing only
in punctuation (GLM 5.3 against the lab's GLM-5.3) defer to the lab and write
nothing. A name already on the file outranks both — four files spell their model
the way its lab does (MiMo-V2.5) where the endpoint sends a storefront
Mimo V2.5 — and handing it through stays correct because factoring drops a name
the lab states identically.

A blank label is not a name. ModelBase.name is min(1), so writing one through
aborts the entire provider's sync at validation rather than skipping the
field. 0 of the 409 live routes send one today, which is exactly why it needed a
test rather than a live check.

Verification

• bun run validate — exit 0
• bun test packages/core/test/sync.test.ts — 220 pass, 2 fail (both
  pre-existing on dev: DeepInfra modalities and an LLM Gateway case-variant
  assertion that only fails on case-insensitive filesystems)
• every rule above is pinned by a test that fails when the rule alone is
  reverted, checked one at a time
• bunx tsc --noEmit -p packages/core — no aihubmix errors
• two consecutive real syncs — 129 created, 69 updated then
  0 created, 0 updated, 0 removed with 198 unchanged (idempotent)
• all 8 moved notes survive a real sync; 0 orphaned wire-path lines left behind
• built-catalog diff against dev — no provider other than aihubmix changes

Review round: two fixed, two answered

Empty reasoning_options on creates — fixed

A route the endpoint reports as reasoning while publishing no
reasoning_options reached the runner with the field unset, and the runner
reads that on a reasoner as "no caller control" and stamps [] — which
AGENTS.md rule 4 forbids using for uncertainty. The guard meant to prevent it
compares against the base model's options, and ModelMetadataBase has no
reasoning_options field, so that comparison is always undefined and the
guard never fires.

The adapter now throws MissingReasoningOptionsError on the two paths that
actually write, the same as the Cloudflare adapter. 13 creates in the
current listing are skipped and surface as missing models instead of being
published with an invented "no controls". An authored [] already on a file is
a human stating the host exposes no control, and is carried through untouched
(grok-build-0.1 on dev).

Stale deprecated never cleared — fixed

status only ever wrote deprecated and otherwise handed the file's value
back, so a route that came back from retirement kept the mark for good.
retire_stage rides on every route (407 active, 2 deprecated), so it is
authoritative about retirement — and only about retirement. A non-deprecated
stage now clears a deprecated status; alpha and beta are left alone
because the endpoint says nothing about either.

Bare budget_tokens — not a defect

Raised as "inventing unbounded controls that shadow real bounds". Three
reasons it is neither:

• Not invented. The endpoint publishes 103 bare budget_tokens
  entries itself. The adapter relays a declaration; it does not synthesise one.
• Nothing is shadowed. ModelMetadata has no reasoning_options field, so
  a lab entry cannot carry a budget range for a provider file to override. A
  first-party entry for the same model states that host's API contract, which
  does not transfer — AIHubMix's Anthropic-compatible path rejects a manual
  budget_tokens on the Opus tier, which the affected files record in their
  headers, so copying min = 1024 across would be wrong.
• It is the repo's own shape. AGENTS.md's Qwen3.5 Plus example authors
  { type = "budget_tokens" } bare with the wire path in the header comment,
  and 256 of the 354 budget entries under providers/ are bare
  (requesty 138, alibaba 23, neuralwatt 16). Authored min/max on the
  8 aihubmix files that have them are preserved, not replaced.

Copying bounds from a peer provider would also put per-model knowledge back
into the adapter, which is what this rewrite exists to remove.

Deleted channel aliases — no gap

Flagged as deletions into a catalog gap. Each of the 8 dropped alias files is
absent from the model list while the ID it routes to is on it, so every
one is replaced by an entry the first sync writes. The gap is visible only
because this PR ships the adapter without its generated files.

Note for maintainers on the sync gate

aihubmix is not in REVIEWED_REASONING_PROVIDERS, and most of its routes are
reasoning models, so the hourly sync PR will be classified unsafe on nearly
every run and wait for a human. That is the right default for a new adapter —
happy to leave it there until it has a track record.

🤖 Generated with Claude Code

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/aihubmix.ts / packages/core/src/sync/index.ts (preserveReasoningOptions) - Check: reasoning = true ⇒ real host reasoning_options; relays must not use [] from uncertainty (AGENTS.md Reasoning options). Why: On creates, when the endpoint omits reasoning/reasoning_options (PR notes ~107 routes omit reasoning), buildAihubmixModel leaves both unset, then the runner resolves lab reasoning = true and preserveReasoningOptions stamps reasoning_options = []. Lab metadata has no reasoning_options, so that empty array is written on the provider file and means “no caller control,” wiping GPT/Claude/DeepSeek-style controls for every such create. Action: On create, if the resolved model reasons and the endpoint did not yield usable options, skip the model (undefined + missing-model issue) or throw MissingReasoningOptionsError—do not emit []. Only write [] when the host affirmatively has no control.
• [high] [possible mistake] providers/aihubmix/models/xiaomi-mimo-v2.5.toml / xiaomi-mimo-v2.5-pro.toml (deletions) - Check: Catalog boundary is the main listed ID; do not drop the only local route for a served model. Why: Patch 19 deletes these as “hidden channel aliases,” but this tree has no mimo-v2.5.toml / mimo-v2.5-pro.toml (or other primary stand-ins). Peer providers catalog mimo-v2.5 under that id; aihubmix only had the xiaomi-mimo-v2.5* files. Until automation recreates a listed id, MiMo V2.5 disappears from aihubmix. Same risk for alicloud-/deep-/zai- DeepSeek/GLM deletes while bare deepseek-v4-pro / glm-5.1 are also absent. Action: Confirm each deleted id is unlisted and only aliases a listed id that this PR or the first sync will author. If xiaomi-mimo-v2.5 (etc.) is the listed id, keep it (or add the real listed id here) instead of deleting into a gap.
• [medium] [possible mistake] packages/core/src/sync/providers/aihubmix.ts (status / retire_stage) - Check: Deprecation should follow the endpoint when it is authoritative for status. Why: status is set to "deprecated" when retire_stage === "deprecated", otherwise existing?.status is kept. A once-deprecated relay that leaves retirement stays deprecated forever (unlike DigitalOcean/DeepInfra, which clear stale deprecation). Action: When retire_stage is present and not deprecated, clear a prior "deprecated" status (keep unrelated alpha/beta if you still preserve those from the file).
• [medium] [violation] packages/core/src/sync/providers/aihubmix.ts (reasoningOptions / budget_tokens) - Check: budget_tokens bounds only when verified; bare budget must not pretend to be a full control surface when lab/peer hosts document ranges. Why: Creates write endpoint-bare { type = "budget_tokens" } with no min/max. Authored bounds are only merged on update. First-time Claude/Gemini-style creates therefore publish unbounded budget where peers use real ranges (e.g. min 1024), and nothing re-applies bounds until a human edits the file. Action: For creates with bare budget_tokens and no authored range, either omit budget until bounds are known, copy a verified same-host peer range, or skip/issue the model—do not ship unbounded budget as if it were complete host metadata.
• [low] [possible mistake] PR body / first post-merge sync - Check: Data-changing automation should be reviewable against citations. Why: This PR ships the adapter and deletes routes but defers ~300 generated TOMLs to sync-models-aihubmix, which the reviewer workflow skips—so the bulk of pricing/limits/options never gets the same review. Action: After the first automation run, spot-check a sample of new/updated files (reasoning shapes, costs, primary MiMo/DeepSeek/GLM ids) against the public catalog and lab peers, or temporarily include a representative generated subset in a follow-up PR.

[0genlab]

Pushed ef0f2c2ec + 5c0590c84 for this round. Two of the four findings were real; the other two are answered in the description (the reviewer reads the PR body, not comments, so that is where the detail lives).

Fixed

• Empty reasoning_options on creates. Real, and the root cause is a guard that can never fire: preserveReasoningOptions compares against the base model's options, but ModelMetadataBase has no reasoning_options field, so the comparison is always undefined. 14 files were getting [] stamped on models whose lab entry says reasoning = true. The adapter now throws MissingReasoningOptionsError on the two paths that write — same as cloudflare-ai-gateway.ts — so 13 creates are skipped and surface as missing models instead. An authored [] is a human statement and is left alone.
• Stale deprecated never cleared. Real. retire_stage is present on every route (407 active, 2 deprecated), so a non-deprecated stage now clears a deprecated status while leaving alpha/beta untouched, following digitalocean.ts:409-411.

Answered, no change

• Bare budget_tokens is not invented: the endpoint publishes 103 of them, AGENTS.md's own Qwen3.5 Plus example authors the same bare shape with the wire path in the header, and 256 of the 354 budget entries under providers/ are bare. Nothing is shadowed — ModelMetadata cannot state a budget range, and a first-party entry's bounds describe that host's API (AIHubMix's Anthropic-compatible path rejects a manual budget_tokens on the Opus tier, which those files record in their headers).
• Dropped channel aliases leave no gap: each of the 8 is absent from the model list while the ID it routes to is on it, so each is replaced by an entry the first sync writes. The gap only looks real because this PR ships the adapter without its generated files.

222 pass / 2 fail (both pre-existing on dev), tsc unchanged, validate exit 0, sync 129 created, 69 updated then 0/0/0, and no file newly carries reasoning_options = []. Each fix was reverted in turn and a test failed for it.