Add lanseq provider

[CHYYX]

Adds Lanseq as an OpenAI-compatible inference provider.

Initial live model:

• Qwen3.8-27B INT4

Lanseq provides inference for open-weight models from Hong Kong / APAC infrastructure. Additional model SKUs and capacity can be added based on demand.

The initial endpoint has been validated end-to-end through OpenCode using its custom OpenAI-compatible provider path.

API:
https://api.lanseq.cloud/v1

First-party documentation:
https://api.lanseq.cloud/docs

The documentation supports the provider-specific metadata in this PR:

• Pricing: input $0.25 / 1M tokens, cached input $0.045 / 1M tokens, output $1.99 / 1M tokens.
• Limits: 70,000-token context and 8,192-token maximum output.
• Reasoning control: reasoning_effort accepts none, low, medium, and xhigh.

[github-actions]

Action items

• [high] [violation] providers/lanseq/models/qwen3.8-27b-int4.toml:1 - Check: Provider base_model overrides must keep only real deltas; resolved metadata must stay accurate for this SKU. Why: The file drops description while overriding modalities to text-only and attachment = false. Merge inherits the lab description (“Dense 27B vision-language model… image and video understanding”), which contradicts this INT4 text-only deployment. An earlier revision had a correct SKU-specific description and it was removed. Action: Restore a provider-local description that states this is a text-only INT4 deployment of Qwen3.8-27B (and keep only other true deltas).
• [high] [possible mistake] providers/lanseq/provider.toml:4 - Check: OpenAI-compatible providers need a stable public api (and useful doc) for catalog consumers. Why: api and doc point at https://api.172.98.22.239.sslip.io/... — an IP + sslip.io hostname that looks ephemeral, not a durable provider endpoint. Catalog clients would send traffic/API keys there. Action: Replace with the provider’s stable production base URL and documentation URL, or confirm in the PR body that this sslip.io host is the intentional long-lived public API and document how it is operated.
• [low] [possible mistake] providers/lanseq/models/qwen3.8-27b-int4.toml:9 - Check: Material cost/limit claims should cite first-party pricing or API docs. Why: cost (0.25 / 1.99 / cache_read 0.045) and reduced limits (70k / 8,192) have no mapped citations; the PR body only lists the same sslip.io API/docs URLs without stating what they support. Action: Add a leading TOML comment and/or PR-body bullets mapping each price and limit to a direct provider source (and currency conversion note if not USD-native).
• [low] [possible mistake] providers/lanseq/logo.svg:1 - Check: New provider logos should use currentColor, no fixed size/hardcoded color, and preferably a square viewBox. Why: Final mark is a wide wordmark (viewBox="0 0 320 80") with system-font text; it meets color rules but is a poor fit for small square logo slots (the prior monogram was closer to the preferred shape). Action: Prefer a square path-based monogram/currentColor mark (e.g. 0 0 24 24) instead of a wide text wordmark.

[CHYYX]

Clarification on the Lanseq API and documentation URLs:

• https://api.172.98.22.239.sslip.io/v1 is the intentional public Lanseq API endpoint for this initial provider deployment, not a temporary validation URL.
• The endpoint is operated by Lanseq on our Hong Kong/APAC inference infrastructure and is intended to remain available for catalog/API consumers.
• TLS is enabled and the endpoint has been validated end-to-end through OpenCode using its custom OpenAI-compatible provider path.
• https://api.172.98.22.239.sslip.io/docs is the first-party Lanseq API documentation for this deployment.
• The documentation specifies the current SKU pricing (input $0.25/M tokens, cached input $0.045/M tokens, output $1.99/M tokens) and limits (70,000 context, 8,192 max output), as reflected in the provider model metadata.
• Lanseq provides inference for open-weight models; Qwen3.8-27B INT4 is the initial live model, with additional model SKUs and capacity to be added based on demand.

We understand that the sslip.io hostname is IP-derived. It is intentionally long-lived for this initial deployment; we can migrate the catalog entry to a branded Lanseq domain in a follow-up when that production domain is available.

[github-actions]

Action items

• [high] [possible mistake] providers/lanseq/provider.toml:4 - Check: Provider api/doc must be real, usable public endpoints for an OpenAI-compatible host. Why: Both point at https://api.172.98.22.239.sslip.io/... (raw IP via sslip.io). That is ephemeral infrastructure, not a durable provider domain, so catalog consumers get an unstable/untrusted base URL and docs link. Action: Replace api and doc with the provider’s stable public hostname (or clearly document why this IP-based endpoint is the intended production surface), and cite the live docs/API that list models.
• [medium] [possible mistake] providers/lanseq/models/qwen3.8-27b-int4.toml:6 - Check: Relay reasoning_options must match this host’s request surface (lab/peer baseline only when this host exposes the same controls). Why: Host is a third-party OpenAI-compatible relay of alibaba/qwen3.8-27b. The set none|low|medium|xhigh matches some peers, but the PR body does not document Lanseq’s wire field or verified values (earlier drafts explicitly blocked shipping options until API controls were verified). Action: Confirm Lanseq accepts these efforts (and the exact field, e.g. reasoning_effort), then keep only the verified set; add a short PR note or leading TOML comment with the wire path/evidence. Do not invent levels this host rejects.
• [medium] [violation] providers/lanseq/logo.svg:1 - Check: New providers need a compliant logo.svg (currentColor, no fixed size/hardcoded color; preferably square viewBox). Why: Final logo is a wide wordmark (viewBox="0 0 320 80") rendered as plain <text> in Arial. That fails the preferred square icon shape and is fragile as a catalog mark (font-dependent, poor at small sizes). The earlier monogram paths were closer to the guideline. Action: Restore a geometric monogram (or other path-based mark) with currentColor and a square viewBox (e.g. 0 0 24 24), no fixed width/height or hardcoded colors.
• [low] [possible mistake] providers/lanseq/models/qwen3.8-27b-int4.toml:10 - Check: Material cost/limit claims should be backed by provider pricing or docs in the PR. Why: cost (0.25 / 1.99 / 0.045) and reduced limit (70_000 / 8_192) have no mapped citations in the PR body, so reviewers cannot confirm operator-supplied numbers. Action: Add first-party pricing/docs (or operator price sheet) links in the PR body and state what each supports (input/output/cache, context/output caps).

[github-actions]

Action items

• [high] [possible mistake] providers/lanseq/provider.toml:4 - Check: Provider api/doc must be the durable public OpenAI-compatible base and documentation URLs for this host. Why: Both point at https://api.172.98.22.239.sslip.io/... (raw public IP via sslip.io). That looks like a temporary or self-hosted endpoint, not a stable provider hostname, and risks publishing a non-durable catalog entry that clients will hard-code. Earlier commits also noted verified public API/docs were still blocked. Action: Confirm this is the intentional long-lived public base URL; if not, replace api and doc with the production host and documentation URLs (and note them in the PR body).
• [low] [possible mistake] providers/lanseq/models/qwen3.8-27b-int4.toml:6 - Check: Relay reasoning_options must match this host’s wire controls (or lab + same-surface peers with host confirmation). Why: effort = none|low|medium|xhigh matches several Qwen3.8-27B peers, but the PR never documents that Lanseq accepts those values; an earlier commit explicitly deferred reasoning_options until public API controls were verified. Action: Cite Lanseq docs or a live probe for the exact request field/values, or adjust reasoning_options to what this endpoint actually exposes.
• [low] [possible mistake] providers/lanseq/models/qwen3.8-27b-int4.toml:11 - Check: Material pricing/limit data should be backed by first-party evidence in the PR. Why: Costs (0.25 / 1.99 / 0.045) and reduced limits (70_000 / 8_192) are provider-specific and cannot be inherited from the lab entry; the PR body only links the API/docs host and does not map sources to these figures. Action: Add PR (or leading TOML) citations stating what supports input/output/cache pricing and the INT4 context/output limits.

[github-actions]

Action items

• [high] [possible mistake] providers/lanseq/provider.toml:4 - Check: Provider api/doc should be durable first-party public endpoints suitable for the catalog. Why: Both point at api.172.98.22.239.sslip.io (raw IP via free sslip.io DNS). That pattern is typically ephemeral/dev infrastructure, not a stable provider host; catalog clients would hard-code API keys and traffic to an address that can change or disappear without a branded domain. Earlier WIP commits in this PR also treated a verified public API/docs URL as blocked. Action: Replace api and doc with a stable Lanseq-controlled hostname (or document why this IP/sslip endpoint is the permanent public API), and keep docs at a URL that lists models/pricing for ongoing verification.
• [low] [possible mistake] providers/lanseq/models/qwen3.8-27b-int4.toml:16 - Check: Material cost/limit claims need direct, mappable evidence (PR body and/or leading TOML comment). Why: [cost] (0.25 / 1.99 / 0.045) and [limit] (70_000 / 8_192) are operator-specific and not supported by a first-party pricing page or a clear claim→source map; the only cited source is the same IP-hosted /docs URL, placed mid-file after keys (AGENTS.md requires source/rationale comments above the first key). Action: Add a leading top-of-file comment (and preferably PR-body bullets) that map each price and limit to a durable first-party source, or correct the values if the docs differ.

[github-actions]

Action items

• [high] [violation] providers/lanseq/provider.toml:4 - Check: New public providers must publish a durable api/doc suitable for the catalog (stable first-party host, not ephemeral infrastructure). Why: api and doc are https://api.172.98.22.239.sslip.io/... — an IP embedded in free dynamic DNS. That is not a stable public provider identity; catalog clients would ship API keys and traffic to a disposable host, and the linked “docs” cannot serve as lasting evidence. Local-only providers intentionally use 127.0.0.1; this is neither local nor a proper public SaaS endpoint. Action: Replace api and doc with a stable first-party domain (and matching docs that list models/pricing), or withdraw the provider until that exists.
• [low] [possible mistake] providers/lanseq/models/qwen3.8-27b-int4.toml:3 - Check: Material cost/limit claims should cite reviewable first-party evidence mapped to each claim. Why: Pricing (0.25 / 1.99 / 0.045), limits (70_000 / 8_192), and the reasoning_effort enum are supported only by the same ephemeral sslip.io docs URL in the leading comment and PR body, so reviewers cannot treat them as durable first-party sources. Action: After a stable docs/pricing page exists, cite it in the PR body and leading TOML comment with what each URL supports (prices, context/output caps, accepted reasoning_effort values).

[github-actions]

Action items

• [medium] [violation] providers/lanseq/models/qwen3.8-27b-int4.toml:2 - Check: Leading source comments must match the provider’s current first-party api/doc endpoints. Why: The model file still cites https://api.172.98.22.239.sslip.io/docs for pricing, limits, and reasoning_effort, but providers/lanseq/provider.toml was moved to https://api.lanseq.cloud. That leaves the only evidence for cost, context/output limits, and the unusual none|low|medium|xhigh effort set pointing at a retired host. Action: Update the leading comment block to https://api.lanseq.cloud/docs (or another current first-party page) and map each claim (input/cache_read/output prices, 70k/8k limits, accepted/rejected reasoning_effort values) to that source.
• [low] [possible mistake] providers/lanseq/provider.toml:4 - Check: PR evidence for material data changes should cite the same endpoints the catalog ships. Why: The PR body still lists the sslip.io API/docs URLs while the final provider entry uses api.lanseq.cloud. Reviewers cannot reconcile the body with the merged metadata. Action: Update the PR body API/docs links to the lanseq.cloud URLs (and briefly note what each supports: pricing, limits, reasoning controls).

[github-actions]

Action items

• [medium] [violation] providers/lanseq/models/qwen3.8-27b-int4.toml:27 - Check: Override-only base_model files must not restate identical inherited fields. Why: Lab models/alibaba/qwen3.8-27b.toml already has modalities.output = ["text"]. Restating it is a redundant override; only the text-only input delta is needed (objects deep-merge). Action: Drop output = ["text"] from [modalities] and keep only input = ["text"].
• [low] [possible mistake] .pr-review/pull-request.json - Check: Data-changing PRs should cite first-party docs in the PR body, mapped to claims. Why: The PR body still points at https://api.172.98.22.239.sslip.io/... while the final provider files use https://api.lanseq.cloud/..., so reviewers cannot match the body citations to the shipped API/doc/pricing claims. Action: Update the PR body to the final api.lanseq.cloud API and docs URLs and briefly map them to pricing, limits, and reasoning_effort values.

[github-actions]

No actionable findings.