feat(aihubmix): sync the model list with the gateway catalog

[0genlab]

Summary

Realigns the AIHubMix plugin's predefined model list with the gateway catalog (/api/v1/models?type=llm):

• +31 chat models — every top-100 catalog entry the plugin was missing
• −11 retired models that the gateway no longer serves
• −37 free-tier aliases (*-free)
• +10 already-shipped models that had never been registered in _position.yaml

Result: 214 LLM entries, every one of them registered in _position.yaml, no orphaned files, no duplicates. manifest.yaml 0.0.46 → 0.0.47.

Added (31)

Family	Models
OpenAI	gpt-6-astra, gpt-5.6-sol-disc, gpt-5.5-pro, ox-alpha
Gemini	gemini-3.1-flash-lite, gemini-3.1-flash-lite-nothink, gemini-3.1-flash-image, gemini-3.1-flash-lite-image, gemini-3-pro-image
DeepSeek	deepseek-v4.1-flash, deepseek-v4-flash-vision-exp, deepseek-v4-flash-0731-fast
GLM	coding-glm-5.3, coding-glm-5.2, glm-5.2-fast-preview
Qwen	qwen3.8-max-2026-09-02, qwen3.8-flash
Hunyuan / Kimi	hy4-preview, hy3-preview, coding-kimi-k3
Other	ernie-5.1, mercury-2.5-preview, mai-thinking-1, agnes-3.0-flash, agnes-2.5-flash, agnes-2.5-pro, agnes-2.5-pro-alpha, longcat-2.0, muse-spark-1.1, muse-spark-1.2, muse-spark-1.3

Validation

Two sources, each used for what it is authoritative about:

• Pricing, model IDs and modality features — the catalog API /api/v1/models?type=llm. Every price was diffed against it (31/31 exact, unit 0.000001 throughout).
• Parameter rules — the AiHubMix public projection, https://aihubmix.com/model-data/index.json (schema 2.0.0), plus its per-model detail files. It states, per model and per protocol, the officially supported enum, the server-side default and the context/output limits.

Every rule was then probed live against the production gateway, because an HTTP 200 does not prove a parameter took effect. Reasoning knobs were verified by comparing usage.completion_tokens_details.reasoning_tokens and the presence of message.reasoning_content across values.

Two defects were caught this way:

• gpt-6-astra — reasoning_effort: minimal returns HTTP 400 listing the accepted values. The rule exposes low/medium/high/xhigh. The projection's enum also has max, but the plugin dispatches this id over chat/completions, and on that surface max spends exactly as many reasoning tokens as medium (6 vs 6 over four runs each) while xhigh really escalates (21–43). On /responses max does work (215–380 against 143–184 for xhigh), so this is a surface-specific drift, not a bad enum. The model also gets the enable_stream opt-out used by the other new OpenAI models (streaming requires a KYC-verified org).
• ernie-5.1 — a boolean thinking is rejected upstream (json: cannot unmarshal bool into struct field ... model.Thinking), the same class of bug fixed for hy3 in feat(aihubmix): add qwen3.7-plus/step-3.7-flash and fix hy3 reasoning control #3427. It exposes reasoning_effort instead, verified effective (none → 0 reasoning tokens).

26 of the 31 returned a full completion on the test account. Five did not and are included anyway because they are listed in the catalog's top 100, with schemas following their validated family siblings: coding-kimi-k3 (no_available_channel on three attempts), hy3-preview (HTTP 502, Tencent upstream reported abnormal) and muse-spark-1.1/1.2/1.3 (HTTP 403 — the test account is access-restricted for that family).

Removed — retired models (11)

Each is absent from the catalog and confirmed unusable against the live gateway:

Model	Gateway response
gpt-4.5-preview	model_retired — "has been retired… use gpt-4.1 instead"
gemini-2.0-flash	"no longer available… update your code to use gemini-3.6-flash"
gemini-2.5-flash-preview-04-17	not found for API version v1beta
gemini-2.5-flash-preview-04-17-nothink	"This model ID is incorrect"
deepseek-v3.2-speciale	upstream 404 (twice)
deepseek-ai/DeepSeek-V3-0324	no_available_channel
aihubmix-DeepSeek-R1	no_available_channel
aihubmix-Llama-3-3-70B-Instruct	no_available_channel
chutesai/Llama-4-Maverick-17B-128E-Instruct-FP8	no_available_channel
chutesai/Llama-4-Scout-17B-16E-Instruct	no_available_channel
meta-llama/llama-4-scout-17b-16e-instruct	model_not_found

Older models that are missing from the catalog but still answer were deliberately kept: claude-3-5-haiku/sonnet-20241022, claude-opus-4-20250514, claude-opus-4-1-20250805, claude-sonnet-4-20250514, DeepSeek-V3.2-Exp(-Think), deepseek-reasoner, deepseek-v3.2-fast, deepseek-v4-flash-think, doubao-seed-2-0-lite, gemini-3-pro-preview, gemini-3.1-flash-lite-preview, xiaomi-mimo-v2.5(-pro), gemini-embedding-2-preview.

Removed — free-tier aliases (37)

All *-free entries (gpt-4.1-free, gemini-3-flash-preview-free, coding-glm-5-free, the nemotron *-free variants, …). They draw on one account-wide free quota and start returning quota errors the moment it is used up, which makes them unreliable as predefined models — 9 of them already failed that way while probing this change.

Also registered (10)

Live models that shipped without a _position.yaml entry: gemini-2.5-pro, gemini-2.5-pro-preview-03-25, gemini-2.5-pro-preview-05-06, both *-search variants, gemini-2.5-flash-image, minimax-m2.5, minimax-m2.5-highspeed, qwen3-vl-235b-a22b-instruct, qwen-plus-latest.

Intentionally not added (7)

Top-100 entries that are not chat models and are not served on /v1/chat/completions:

• mai-image-2.5, mai-image-2.5-flash, mai-image-2.5-pro, mai-image-2.6, mai-image-2.6-flash — image generation (Requested path is not found); they belong to the image endpoint / tools/aihubmix_image.
• gpt-audio-1.5 — requires an audio modalities field the OpenAI-compatible path cannot send.
• gpt-4o-transcribe-diarize — speech-to-text (The requested operation is unsupported); models/speech2text/ is the right home.

Gemini image models

Review flagged that the three new image models would have shipped text-only, and that was correct. llm.py routes every gemini* id to GoogleLargeLanguageModel, which only asks for IMAGE output when the id is listed in google.py's IMAGE_GENERATION_MODELS; all three were missing from it, so _set_response_modalities and _set_image_config were both skipped.

They are now in the allowlist, and each YAML exposes aspect_ratio / resolution with the options that model actually accepts — probed live against the gateway, dimensions read back off the returned JPEG:

Model	Resolution	Aspect ratio
gemini-3-pro-image	1K / 2K / 4K	base set; 4:1 and 8:1 return 400
gemini-3.1-flash-image	1K / 2K / 4K	base set + 4:1, 8:1
gemini-3.1-flash-lite-image	1K only — 2K/4K return 400	base set + 4:1, 8:1

_set_image_config's hardcoded ratio allowlist also had to learn 4:1 and 8:1, otherwise the new YAML options would have been dropped before the request left the plugin. No include_thoughts rule was added: _set_thinking_config returns early for IMAGE_GENERATION_MODELS, so that control would be inert.

Parameter rules realigned to the public projection

A later pass rechecked all 23 affected YAMLs against the projection rather than against vendor docs and sibling files, and reconciled every difference.

reasoning_effort now carries each model's own documented enum and default:

Model	Options	Default
gpt-6-astra	low / medium / high / xhigh / max	medium
gpt-5.5-pro	medium / high / xhigh (rule was missing)	medium
coding-glm-5.2	none / minimal / low / medium / high / xhigh / max	max
coding-glm-5.3	low / medium / high / max	medium
coding-kimi-k3	low / medium / high / max	medium
deepseek-v4.1-flash, deepseek-v4-flash-vision-exp, deepseek-v4-flash-0731-fast	none / low / medium / high / max	medium
glm-5.2-fast-preview	none / minimal / low / medium / high / xhigh / max (rule was missing)	max
qwen3.8-max, qwen3.8-max-2026-09-02	none / minimal / low / medium / high / xhigh / max	xhigh
qwen3.8-flash	low / medium / xhigh	medium
hy3-preview	no_think / low / high	no_think
hy4-preview	none / high	high
mercury-2.5-preview	instant / low / medium / high (rule was missing)	medium
ox-alpha	low / medium / high / max (rule was missing)	medium
muse-spark-1.1 / 1.2 / 1.3	minimal / low / medium / high / xhigh (rule was missing)	medium

Reading rule: the projection's values is the documented enum and its default is the value the server applies when the field is omitted — itself an accepted wire value, confirmed live on coding-glm-5.3 (medium), qwen3.8-max-2026-09-02 and hy3. So each YAML exposes values ∪ {default} and defaults to the projection default.

Limits follow the projection's specs: qwen3.8-max* context 991000 → 1000000 and max output 128000 → 131072; qwen3.8-flash and glm-5.2* max output 128000 → 131072; hy3-preview / hy4-preview gain the max_tokens rule they were missing (128000 / 64000).

Knobs the projection does not expose were dropped, each after confirming on the gateway that the 200 was empty:

Removed	Why
gpt-5.5-pro temperature, top_p	reasoning-only model with no chat-completions surface; accepted and dropped
glm-5.2-fast-preview boolean thinking	thinking: false still returns reasoning_content; only the official object form disables it, which parameter_rules cannot express — reasoning_effort: none is the working off switch
glm-5.2-fast-preview thinking_budget	no budget_tokens option in the projection

agnes-2.5-flash / agnes-3.0-flash gained the projection's official thinking toggle instead: a boolean enable_thinking, which llm.py sends as chat_template_kwargs.enable_thinking. Verified end to end — true → reasoning tokens and reasoning_content, false → neither.

ernie-5.1 keeps its reasoning_effort rule even though the projection lists only thinking_budget for it: the enum is verified effective on the gateway (none → 0 reasoning tokens) while a boolean thinking is rejected upstream.

Enums taken from the projection's per-field definitions

The pass above read only specs.reasoning_options — the projection's top-level summary of a model's reasoning knobs. The authoritative enum actually lives one level deeper, in domains[].capabilities[].protocols[].fields[]: each protocol lists its own fields with enum, default and an override.status (model-specific-enum when the set is narrowed for this model). A final pass compares every model against the fields[] of the protocol it actually routes to in llm.py, and each change below is backed by gateway behaviour, not just an HTTP 200.

gpt-5.6-sol-disc reasoning_mode is restored — dropping it above was wrong. It is official-model-level on openai.responses, and a bogus value 400s with Supported values are: 'standard' and 'pro'. mode: pro spends 69 reasoning tokens where standard spends 0. Two sibling fields were missing entirely and are added:

Added	Enum	Evidence
reasoning_summary	auto / concise / detailed	streaming with detailed produced 79 response.reasoning_summary_text.delta events, which openai_response.py already parses
reasoning_context	auto / current_turn / all_turns	bogus value 400s naming the three; left with no default so the server-side all_turns still applies

glm-5.2 gained the reasoning_effort it was missing (none…max, default max): none → 0 reasoning tokens and no reasoning_content, minimal → 294, max → 243.

The three deepseek flash models swap an invented enum for the official toggle. Their reasoning_effort listed a none that the projection's fields[] does not contain (low/high/max, default medium); it is gone. The off switch is the field the projection really marks official-model-level, a thinking object — shipped as a boolean because llm.py now normalises it into {"type": "enabled"｜"disabled"}. Verified: enabled → 141/136/177 reasoning tokens, disabled → none.

json_schema is now offered only where the gateway honours it. Probed with a strict schema and max_tokens: 400:

Model	Result	Action
deepseek-v4.1-flash, deepseek-v4-flash-vision-exp	400 This response_format type is unavailable now	rule, structured-output flag and the json_schema response_format option removed
coding-glm-5.2, coding-glm-5.3	200 but ignored — plain Hello!, schema never applied	same
glm-5.2, glm-5.2-fast-preview, deepseek-v4-flash-0731-fast	{"greeting": "Hello"} — enforced	kept

gemini-3.1-flash-lite-nothink deliberately keeps no reasoning_effort even though the projection lists one: minimal, high and a bogus value all return 200 with zero reasoning tokens. That enum belongs to the base model, not the -nothink id.

Fields present in the projection but deliberately not adopted: tool_choice (the plugin sets it itself, and injecting it with no tools risks 400s); service_tier, prompt_cache_retention and prompt_cache_options.* (nested and/or transport-billing knobs a flat parameter_rules entry cannot express); messages[].content[].image_url.detail and input[].phase (per-message, not per-model); single-value enums with no useful choice (mai-thinking-1 reasoning_display: ['encrypted'], ernie-5.1 web_search: ['auto']); and everything whose verdict is spec-only rather than official-model-level.

Every rule re-checked against gateway behaviour

A final sweep called each rule instead of trusting the projection: the full reasoning_effort enum of 13 models, json_schema on 25, and the thinking switch on 11.

The first finding is about method. reasoning_effort is not validated on the chat-completions surface — reasoning_effort: "zzz-not-a-level" returns 200 and the model reasons anyway on all 16 models tried, so a 200 says nothing and the reasoning-token count is the only usable evidence. Only the OpenAI-native surface validates: gpt-5.5-pro 400s with Supported values are: 'medium', 'high', and 'xhigh' — exactly the three its rule already offers, and none, minimal and low are each rejected by name.

Removed — the gateway does not honour them:

Removed	Evidence
mercury-2.5-preview reasoning_effort	inert across six calls; instant, the level that should spend the least, spent the most (864 / 872 reasoning tokens against 794 for high)
ernie-5.1, ox-alpha json_schema	both answer a schema-constrained request in prose (Hello! I am using English.); json_object works, so response_format keeps that option only
mai-thinking-1 json_schema + response_format	every structured format 400s with Structured `response_format` is not enabled for model

Restored — the gateway honours them better than the YAML claimed. glm-5.2-fast-preview gets thinking and thinking_budget back: dropping them earlier was based on the projection plus a raw boolean that the vendor rejected, but llm.py now sends the official object, and with it the switch spends no reasoning tokens when disabled (333 when enabled) while a budget of 64 spends exactly 64. The same measured wording is now on glm-5.2's two rules.

coding-glm-5.2 keeps its enum — the levels do separate (low ~20 tokens against 57–59 for high/xhigh/max) — but none does not stop thinking: ~40 reasoning tokens and a reasoning_content body across two samples. The help text says so, and it joins the hy3-preview / hy4-preview drift already filed as a separate task.

Not verifiable this round, so left exactly as the projection has them: coding-kimi-k3 (400 cannot be served at the moment), hy3-preview json_schema (502 upstream, twice) and muse-spark-1.1/1.2/1.3 (403 on this key).

Plugin plumbing fixed alongside

The projection also exposed two defects in llm.py that no YAML change could reach.

gpt-5.5-pro was going over the wrong protocol. Its projection entry lists only openai.responses and anthropic.messages — unlike gpt-5.5, it has no chat-completions surface. RESPONSE_SERIES_COMPATIBILITY = ("gpt-5-codex", "gpt-5-pro", "gpt-5.6", "o3-pro") matches by prefix and "gpt-5.5-pro" starts with none of them, so the plugin sent it to chat completions. It now routes through openai_response.py like its siblings; its rules (reasoning_effort, max_tokens, response_format, json_schema) are exactly the ones that surface translates.

The thinking toggle was broken on 24 models. parameter_rules can only express scalars, so those YAMLs ship a boolean thinking — but every vendor's chat-completions field is an object. The bare boolean fails on the gateway:

Model	"thinking": false
doubao-seed-2-0-pro, doubao-seed-2-0-lite	400 Mismatch type openapi.Thinking with value bool
kimi-k2.5	400 'thinking' must be json_object
coding-glm-5	400 JSON parse error: ... ChatCompletionRequest$Thinking
DeepSeek-V3.2-Exp-Think, ernie-5.0-thinking-exp	400 cannot unmarshal bool into ... model.Thinking
glm-5.2	200 — and ignored: 531 reasoning tokens still billed, thinking never stopped

llm.py now converts it to the official {"type": "enabled"｜"disabled"} on the OpenAI-compatible branch only, so anthropic.py keeps consuming the boolean itself and Gemini keeps going through google.py. With the object form those same calls return 200 and thinking actually stops — glm-5.2 645 reasoning tokens → none, doubao-seed-2-0-pro → 0, coding-glm-5 → 14.

gemini-3-pro-preview needed the third variant: google.py's _set_thinking_config reads thinking_mode, so its thinking rule was simply never read. Renamed.

Documentation

The README stopped at the configuration screenshot, so it now also links the API-key console, the step-by-step Dify integration guide (with its Chinese version), and the AiHubMix site and docs root. The links live in the README rather than in the manifest.yaml / provider/aihubmix.yaml descriptions, because no model plugin in this repo puts a URL in either field and both render as plain text.

Release Notes

0.0.47 - catalog realignment: 31 new models, 11 retired ones removed, parameter rules rebuilt from AIHubMix's published schema

• Adds 31 models from the AIHubMix top-100 catalog that the plugin was missing, including gpt-6-astra, gpt-5.6-sol-disc, gpt-5.5-pro, gemini-3.1-flash-lite, deepseek-v4.1-flash, coding-glm-5.3, qwen3.8-flash, hy4-preview, ernie-5.1 and three Gemini image models.
• Removes 11 models the gateway no longer serves, so they stop appearing in the model picker and failing at call time.
• Removes 37 *-free aliases, which the gateway does not serve.
• Registers 10 already-shipped models that were missing from the model list and were therefore invisible in Dify.
• Rebuilds every model's parameter rules from AIHubMix's published model schema: reasoning-effort and thinking options now match what each model actually accepts, and knobs the gateway ignores or rejects are no longer offered.
• Fixes thinking being impossible to switch off on OpenAI-compatible models (the boolean was sent in a form the gateway ignores) and on gemini-3-pro-preview (the parameter was never read).
• Fixes the three Gemini image models being unable to return an image.
• Structured output (json_schema) is now advertised only on the models where the gateway honours it.
• The README now links the AiHubMix Dify integration guide, the API-key console, and the AiHubMix site and documentation.

Change Type

• Documentation / non-plugin change
• Non-LLM plugin (tools, extensions, datasource, etc.)
• LLM plugin

Screenshots / Videos

Before	After
183 selectable models - 11 of them retired, 37 of them *-free aliases the gateway refuses, and 10 shipped models invisible because they were never registered	214 models, all served by the gateway, all registered
Thinking could not be turned off on OpenAI-compatible models or on gemini-3-pro-preview: the request returned 200 and kept billing reasoning tokens	thinking: false actually stops reasoning (glm-5.2 645 reasoning tokens -> 0)
The three Gemini image models could not return an image	gemini-3-pro-image, gemini-3.1-flash-image, gemini-3.1-flash-lite-image return images

LLM Plugin Checklist

Areas affected by this change (check all that apply)

• Message flow (system messages, user / assistant turn-taking)
• Tool interaction flow (multi-round usage, Agent App and Agent Node)
• Multimodal input (images, PDFs, audio, video, etc.)
• Multimodal output (images, audio, video, etc.)
• Structured output (JSON, XML, etc.)
• Token consumption metrics
• Other LLM functionality (reasoning, grounding, prompt caching, etc.)
• New models / model parameter fixes

Version

• Bumped top-level version in manifest.yaml (not the one under meta) - 0.0.46 -> 0.0.47
• dify_plugin>=0.10.2 is declared in pyproject.toml and locked in uv.lock (the pin this plugin already uses; the template's >=0.3.0,<0.6.0 range predates the current SDK)

Testing

• Local deployment - Dify version: not run
• SaaS (cloud.dify.ai)

This change is schema and metadata only: it adds and removes YAML model definitions and fixes three parameter-plumbing bugs in llm.py / google.py. Validation was done directly against the production AIHubMix gateway plus the plugin's own unit tests, not through a Dify UI run - happy to record a local deployment run if a maintainer wants that before merge.

Against the gateway - every rule was exercised with a real call and judged on observed behaviour rather than on HTTP 200: reasoning-token counts (usage.completion_tokens_details.reasoning_tokens), the presence of reasoning_content, the returned body shape for json_schema, and the error text returned for deliberately bogus enum values. The per-model results are in the tables above. Four models could not be reached and fall back to the published schema: muse-spark-1.1/1.2/1.3 (403, key-level restriction) and hy3-preview (502 from the Tencent upstream).

Unit tests

tests/test_top100_models.py — asserts every added model parses, declares llm/chat, uses the standard pricing unit and is registered in _position.yaml; that no *-free file or position entry remains; that each retired model is gone from both; that the three image models are in IMAGE_GENERATION_MODELS and expose exactly the validated aspect-ratio/resolution options with no inert thinking switch; plus regression assertions for the two parameter defects above. Four further tests pin the realignment: reasoning_effort options/defaults and the context/output limits against a table transcribed from the projection, the absence of the inert knobs, and the agnes thinking toggle. Three more cover the plumbing: the responses-API routing of gpt-5.5-pro, the boolean→object thinking conversion (the helper is lifted out of llm.py and exercised directly, so it runs without dify_plugin), and the Gemini rule rename. Three more pin the fields-level pass: that gpt-5.6-sol-disc exposes every official responses reasoning field with the projection's enum and that openai_response.py translates each one, that the deepseek flash trio turn thinking off through the official field, and that json_schema is offered only where the gateway honours it. The sweep above adds a fourth: the restored glm-5.2-fast-preview switch, mercury's inert enum staying out, and mai-thinking-1 keeping no structured-output surface. Review follow-ups add three more: the registry cross-check (every _position.yaml id has a schema behind it and every llm schema is registered, compared on the declared model: id rather than the filename, with the totals 31 added / 214 registered pinned and duplicate ids rejected), the structured-output feature flag now tracking the json_schema rule across the 17 schemas this PR gives it, and gpt-6-astra asserting the reduced reasoning_effort enum. tests/test_qwen38_max_config.py is a pre-existing module this PR touches: the catalog now reports 1,000,000 context / 131,072 max output for qwen3.8-max (was 991,000 / 128,000), so its limit assertions were refreshed — pricing is unchanged.

The whole test job was reproduced locally against this branch, packaged the way CI packages it:

$ dify plugin package models/aihubmix --output_path test-models-aihubmix-package.difypkg
$ unzip -q test-models-aihubmix-package.difypkg -d unpacked && cd unpacked
$ UV_PROJECT_ENVIRONMENT=.ci-test-uv-venv uv sync --all-groups --frozen --python 3.12
$ PLUGIN_FILE_PATH=$(realpath ../test-models-aihubmix-package.difypkg) \
    UV_PROJECT_ENVIRONMENT=.ci-test-uv-venv uv run --frozen --with pytest python -m pytest
33 passed in 40.72s

That covers all four test modules, including the two pre-existing ones that need dify_plugin (test_september_models.py, test_opus5_parameters.py). The other steps of the job were run by hand too: manifest author is langgenius, 0.0.47 is not yet on the Marketplace (0.0.46 is), no venv* directory, pyproject.toml present, uv sync --frozen reports no lockfile drift, and the Release Notes section above extracts non-empty. Only the two steps that need repository secrets or the CI-cloned .scripts/ toolkit were skipped (Check Packaging, Check Plugin Install).

🤖 Generated with Claude Code

[Copilot]

🟡 Changes recommended

The new Gemini image-generation IDs are not registered in the provider’s image-model configuration, preventing image output setup.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Adds 55 AIHubMix top-100 LLM catalog entries, registers them, and validates their schemas.

Changes:

• Adds model definitions across 11 families.
• Updates catalog ordering and regression tests.
• Bumps plugin version to 0.0.47.

File summaries

File	Description
models/aihubmix/models/llm/auto.yaml	Adds Smart Router.
models/aihubmix/models/llm/gpt-6-astra.yaml	Adds GPT-6 Astra.
models/aihubmix/models/llm/gpt-5.6-sol-disc.yaml	Adds GPT-5.6 Sol Disc.
models/aihubmix/models/llm/gpt-5.5-pro.yaml	Adds GPT-5.5 Pro.
models/aihubmix/models/llm/gpt-oss-20b-free.yaml	Adds free GPT-OSS.
models/aihubmix/models/llm/ox-alpha.yaml	Adds Ox Alpha.
models/aihubmix/models/llm/gemini-3.8-flash-free.yaml	Adds free Gemini 3.8 Flash.
models/aihubmix/models/llm/gemini-3.7-flash-free.yaml	Adds free Gemini 3.7 Flash.
models/aihubmix/models/llm/gemini-3.6-flash-free.yaml	Adds free Gemini 3.6 Flash.
models/aihubmix/models/llm/gemini-3.5-flash-lite-free.yaml	Adds free Gemini 3.5 Lite.
models/aihubmix/models/llm/gemini-3.1-flash-lite.yaml	Adds Gemini 3.1 Lite.
models/aihubmix/models/llm/gemini-3.1-flash-lite-nothink.yaml	Adds non-thinking variant.
models/aihubmix/models/llm/gemini-3.1-flash-image.yaml	Adds Flash Image.
models/aihubmix/models/llm/gemini-3.1-flash-lite-image.yaml	Adds Flash-Lite Image.
models/aihubmix/models/llm/gemini-3-pro-image.yaml	Adds Pro Image.
models/aihubmix/models/llm/deepseek-v4.1-flash.yaml	Adds DeepSeek V4.1 Flash.
models/aihubmix/models/llm/deepseek-v4-flash-vision-exp.yaml	Adds vision variant.
models/aihubmix/models/llm/deepseek-v4-flash-0731-fast.yaml	Adds fast dated variant.
models/aihubmix/models/llm/coding-glm-5.3.yaml	Adds Coding GLM 5.3.
models/aihubmix/models/llm/coding-glm-5.3-free.yaml	Adds free GLM 5.3.
models/aihubmix/models/llm/coding-glm-5.2.yaml	Adds Coding GLM 5.2.
models/aihubmix/models/llm/coding-glm-5.2-free.yaml	Adds free GLM 5.2.
models/aihubmix/models/llm/glm-5.2-fast-preview.yaml	Adds GLM fast preview.
models/aihubmix/models/llm/qwen3.8-max-2026-09-02.yaml	Adds dated Qwen Max.
models/aihubmix/models/llm/qwen3.8-flash.yaml	Adds Qwen Flash.
models/aihubmix/models/llm/hy4-preview.yaml	Adds Hunyuan 4 preview.
models/aihubmix/models/llm/hy3-free.yaml	Adds free Hunyuan 3.
models/aihubmix/models/llm/hy3-preview.yaml	Adds Hunyuan 3 preview.
models/aihubmix/models/llm/coding-kimi-k3.yaml	Adds Coding Kimi K3.
models/aihubmix/models/llm/coding-kimi-k3-free.yaml	Adds free Kimi K3.
models/aihubmix/models/llm/minimax-m2.7-free.yaml	Adds free MiniMax M2.7.
models/aihubmix/models/llm/ernie-5.1.yaml	Adds ERNIE reasoning controls.
models/aihubmix/models/llm/mercury-2.5-preview.yaml	Adds Mercury preview.
models/aihubmix/models/llm/mai-thinking-1.yaml	Adds MAI Thinking.
models/aihubmix/models/llm/agnes-3.0-flash.yaml	Adds Agnes 3.0 Flash.
models/aihubmix/models/llm/agnes-2.5-flash.yaml	Adds Agnes 2.5 Flash.
models/aihubmix/models/llm/agnes-2.5-pro.yaml	Adds Agnes 2.5 Pro.
models/aihubmix/models/llm/agnes-2.5-pro-alpha.yaml	Adds Agnes Pro Alpha.
models/aihubmix/models/llm/longcat-2.0.yaml	Adds LongCat 2.0.
models/aihubmix/models/llm/dots-3-note-preview-free.yaml	Adds free Dots preview.
models/aihubmix/models/llm/lfm-2.5-2.6b-free.yaml	Adds free LFM model.
models/aihubmix/models/llm/laguna-xs-2.1-free.yaml	Adds free Laguna XS.
models/aihubmix/models/llm/north-mini-code-free.yaml	Adds free North Mini Code.
models/aihubmix/models/llm/ling-3.0-flash-free.yaml	Adds free Ling Flash.
models/aihubmix/models/llm/ling-3.0-tiny-free.yaml	Adds free Ling Tiny.
models/aihubmix/models/llm/nemotron-3-super-120b-a12b-free.yaml	Adds free Nemotron Super.
models/aihubmix/models/llm/nemotron-3-ultra-550b-a55b-free.yaml	Adds free Nemotron Ultra.
models/aihubmix/models/llm/nemotron-3-nano-omni-30b-a3b-reasoning-free.yaml	Adds free Nano Omni.
models/aihubmix/models/llm/nemotron-3.5-lightning-free.yaml	Adds free Lightning.
models/aihubmix/models/llm/nemotron-3.5-content-safety-free.yaml	Adds safety model.
models/aihubmix/models/llm/nemotron-nano-9b-v2-free.yaml	Adds free Nano 9B.
models/aihubmix/models/llm/nemotron-nano-12b-v2-vl-free.yaml	Adds free Nano VL.
models/aihubmix/models/llm/muse-spark-1.1.yaml	Adds Muse Spark 1.1.
models/aihubmix/models/llm/muse-spark-1.2.yaml	Adds Muse Spark 1.2.
models/aihubmix/models/llm/muse-spark-1.3.yaml	Adds Muse Spark 1.3.
models/aihubmix/models/llm/_position.yaml	Registers models in catalog order.
models/aihubmix/tests/test_top100_models.py	Adds schema regressions.
models/aihubmix/manifest.yaml	Bumps plugin version.

Review details

• Files reviewed: 58/58 changed files
• Comments generated: 1
• Review effort level: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🟡 Changes recommended

It omits 23 advertised free models and leaves three Gemini image models unable to request image output.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

• Files reviewed: 51/51 changed files
• Comments generated: 4
• Review effort level: Balanced

[Copilot]

🔵 Needs a closer look

The retired-model test does not validate registry removal when schema filenames differ from gateway model IDs.

Review details

Suppressed comments (1)

Previously missed (1) — in code that hasn't changed since the last review.

models/aihubmix/tests/test_top100_models.py:135

• These six values are schema filenames, but _position.yaml stores their different gateway model IDs (for example, deepseek-r1-aihubmix.yaml is registered as aihubmix-DeepSeek-R1). Therefore each model not in positions assertion is vacuously true and would not catch a stale registry entry. Track filename and model ID separately so the removal test covers both artifacts.

• Files reviewed: 52/52 changed files
• Comments generated: 0 new
• Review effort level: Balanced

[Copilot]

🟡 Changes recommended

An ineffective GPT-6 option and stale Qwen assertions must be corrected before approval.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (1)

Previously missed (1) — in code that hasn't changed since the last review.

models/aihubmix/tests/test_top100_models.py:142

• For these six aliases, the tuple contains the filename stem, but _position.yaml contains each schema's model value (for example aihubmix-DeepSeek-R1 and deepseek-ai/DeepSeek-V3-0324). Consequently the position assertion would still pass if any retired model ID were accidentally left registered. Track filename and model ID separately so the regression test covers both removals.

• Files reviewed: 55/55 changed files
• Comments generated: 2
• Review effort level: Balanced

[Copilot]

🔵 Needs a closer look

The existing Qwen 3.8 Max configuration test still asserts the old limits and will fail in the full test suite.

Review details

Suppressed comments (1)

models/aihubmix/models/llm/qwen3.8-max.yaml:15

• This change leaves tests/test_qwen38_max_config.py:11-12 asserting the old 991_000 context and 128_000 output limit, so the full AIHubMix test suite fails even though the new focused test passes. Update both existing expectations to 1_000_000 and 131_072.

• Files reviewed: 57/57 changed files
• Comments generated: 0 new
• Review effort level: Balanced

[Copilot]

🔵 Needs a closer look

The updated Qwen limits break an existing test, and several retirement assertions check filenames instead of registered model IDs.

Review details

Suppressed comments (2)

Previously missed (1) — in code that hasn't changed since the last review.

models/aihubmix/tests/test_top100_models.py:142

• These entries are filenames, not the model IDs stored in _position.yaml, so the second assertion never verifies removal of aihubmix-DeepSeek-R1, aihubmix-Llama-3-3-70B-Instruct, deepseek-ai/DeepSeek-V3-0324, or the three namespaced Llama IDs. Represent each retired item as (filename, model_id) and assert the filename is absent while the actual model ID is absent from positions.

models/aihubmix/models/llm/qwen3.8-max.yaml:15

• The full plugin test suite now has contradictory assertions: tests/test_qwen38_max_config.py:11-12 still requires this model's context and output limits to be 991_000 and 128_000. With these updated values that pre-existing test fails, even though the new projection test only covers the dated variant. Update the existing test to assert 1_000_000 and 131_072 (and include this base model in the projection table) before merging.

• Files reviewed: 57/57 changed files
• Comments generated: 0 new
• Review effort level: Balanced

[Copilot]

🔵 Needs a closer look

An existing Qwen regression test now fails, and supported structured-output metadata remains incomplete.

Review details

Suppressed comments (2)

Previously missed (1) — in code that hasn't changed since the last review.

models/aihubmix/tests/test_top100_models.py:388

• These three models keep working json_schema rules, but none of their schemas declares the structured-output feature. That contradicts this PR's own supported/unsupported distinction (the preceding branch removes and asserts the flag when schemas are not honored) and leaves Dify capability discovery reporting structured output as unsupported. Add the feature to these supported models and assert its presence here.

models/aihubmix/models/llm/qwen3.8-max.yaml:15

• This catalog update makes the existing tests/test_qwen38_max_config.py fail: lines 11–12 still require a 991_000 context and 128_000 maximum output. Update that regression test to the new 1_000_000 / 131_072 values (and ideally include this base model in the new projection tables) so the full AIHubMix test suite remains green.

• Files reviewed: 56/56 changed files
• Comments generated: 0 new
• Review effort level: Balanced

[Copilot]

🟡 Changes recommended

The existing Qwen regression test now fails, and catalog-count and retired-registration assertions need correction.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (2)

Previously missed (1) — in code that hasn't changed since the last review.

models/aihubmix/tests/test_top100_models.py:144

• For these six retired models, the filename differs from the model ID that _position.yaml stores (for example, deepseek-r1-aihubmix vs. aihubmix-DeepSeek-R1). As a result, assert model not in positions is vacuous and would still pass if the retired registry entries were reintroduced. Track both filename and model ID explicitly.

models/aihubmix/models/llm/qwen3.8-max.yaml:15

• These catalog updates leave the pre-existing tests/test_qwen38_max_config.py asserting 991_000 and 128_000, so the plugin's full test suite now fails even though the new test expects these values. Update that regression test to 1_000_000 and 131_072 in the same change.

• Files reviewed: 56/56 changed files
• Comments generated: 1
• Review effort level: Balanced

[Copilot]

🔵 Needs a closer look

The implementation contains 31 additions and 214 registrations, contradicting the advertised totals of 32 and 215.

Review details

Suppressed comments (1)

models/aihubmix/tests/test_top100_models.py:8

• NEW_MODELS contains 31 IDs, and the resulting _position.yaml contains 214 entries, while the PR contract states 32 additions and 215 total entries. The “Added (32)” table also names only 31 models, so this test currently codifies the mismatch instead of detecting the advertised total. Please either add the missing catalog model and register it, or correct the stated scope/counts and add an explicit total-count assertion.

• Files reviewed: 57/57 changed files
• Comments generated: 0 new
• Review effort level: Balanced

[0genlab]

Note for reviewers: the latest Copilot review (2026-09-11T13:29:46Z) is stale — it ran against a25fe937, one commit behind the current head.

Its finding was the count mismatch (31 additions / 214 registered ids vs the advertised 32 / 215). That is fixed in 5c98415c: the PR description now states 31 / 214, and test_every_new_model_is_registered pins both numbers explicitly, as the review asked:

assert len(NEW_MODELS) == 31
assert len(positions) == 214
assert len(positions) == len(set(positions))
...
assert set(positions) - declared == set()
assert declared - set(positions) == set()

The last two close the gap in both directions, so a schema added without a _position.yaml entry — or an entry with no schema — fails in the test rather than at runtime.

All 8 review threads on this PR are resolved, and the PyYAML-only suite is green (24 passed). I do not have permission to re-request a Copilot review on this repository, so the bot will pick up the current head on the next push.

🤖 Generated with Claude Code

[Copilot]

🔵 Needs a closer look

The large gateway-dependent catalog update includes unverified models and unresolved validation-test gaps.

Review details

Suppressed comments (1)

Previously missed (1) — in code that hasn't changed since the last review.

models/aihubmix/tests/test_top100_models.py:467

• The validation output in the PR description reports 22 passed, but this module now contains 23 uniquely named, non-skipped test functions, including this final regression test. That output therefore predates the current test file; please rerun the current suite and update the validation result so the latest plumbing change is actually verified.

• Files reviewed: 57/57 changed files
• Comments generated: 1
• Review effort level: Balanced

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.