feat(provider): add Vispark Lab with Vision Small/Medium/Large

[vansh-one]

Adds first-party provider Vispark (Vispark Lab) with 3 Vision chat models.

Provider (first-party lab, OpenAI-compatible):

• API https://api.lab.vispark.in/v1 — GET /v1/models, POST /v1/chat/completions, POST /v1/completions
• Auth via X-API-Key or Authorization Bearer; env VISPARK_LAB_API_KEY
• Platform https://lab.vispark.in/ — keys: https://lab.vispark.in/profile#api — Vision docs: https://lab.vispark.in/#vision

Models (IDs match live GET /v1/models exactly):

• vispark/vision-small — $1.05 in / $3.16 out per MTok, 1M ctx / 65K out
• vispark/vision-medium (default) — $4.21 / $12.63, 1M ctx / 65K out
• vispark/vision-large — $7.37 / $22.11, 1M ctx / 65K out
• Multimodal in (text, image, audio, video, pdf), text out; tools + structured output + temperature; attachment true; open_weights false

[github-actions]

Action items

• [medium] [possible mistake] providers/vispark/models/vispark/vision-small.toml:248 - Check: Costs must be USD/MTok; non-USD conversions need a top-of-file rate/date note (AGENTS.md → Cost). Why: Input/output prices (1.05/3.16, 4.21/12.63, 7.37/22.11) are atypical USD list prices and look like FX-converted figures, but comments only restate the same numbers with no source currency, rate, or date. Action: Confirm these are published USD/MTok (or convert correctly), document currency→USD rate/date in the leading comment if converted, and cite the pricing page each figure comes from.
• [medium] [possible mistake] providers/vispark/models/vispark/vision-small.toml:241 - Check: reasoning = true requires reasoning_options for this host’s real control surface; [] means always-on with no caller control, only with affirmative evidence (.opencode/skills/audit-reasoning-options/SKILL.md). Why: All three models set reasoning = true and reasoning_options = [] while marketing “reasoning,” but neither the PR body nor leading comments document that Lab’s chat API has no toggle/effort/budget (or always-on CoT). Empty options from uncertainty are invalid; inventing controls would also be wrong. Action: Verify against Lab docs or a live request: if no reasoning control exists and the model always reasons, keep [] and state that affirmatively in the leading comment + PR body; if a wire control exists, author the real option set (with toggle wire path if needed); if the model does not reason, set reasoning = false and drop reasoning_options.
• [low] [possible mistake] models/vispark/vision-small.toml:62 - Check: release_date / last_updated must be accurate first-public / last-change dates. Why: All three models share release_date = "2024-05-15" and last_updated = "2026-08-01" while being described as current foundational Vision models; a mid-2024 release for this catalog entry is hard to reconcile without a source. Action: Verify both dates against a first-party announcement or docs and correct them (or cite the source that supports 2024-05-15 / 2026-08-01).
• [low] [possible mistake] .pr-review/pull-request.json:3 - Check: Data-changing PRs should cite direct provider pricing/docs/API and map each link to what it supports. Why: The body lists endpoints and prices but does not map first-party sources to concrete claims (USD pricing, 1M/65K limits, multimodal inputs including audio/video/pdf, tools, structured output, temperature, reasoning behavior). Without that, material facts cannot be reviewed. Action: Add first-party links (pricing, Vision docs, model/API reference) and note which claim each supports.

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.

[rekram1-node]

i feel like some of the language here could be provocative? Is vispark the only indian lab? Ideally it's more about the models themselves rather than this marketing lingo...

[vansh-one]

Fair point — rewritten all three descriptions to neutral, capability-focused copy (modalities, context window, reasoning tier, tool use) in the latest commit. The lingo was carried over from our API's display strings; the catalog entries no longer make those claims.

[github-actions]

No actionable findings.