feat(hyper): add Charm Hyper provider and sync module

[oskarkocol]

supersedes #2983

Context

Our friends at charm land shipped a gateway.

I'm adding them as a provider.
And also closing: #2747
Their prices are accessible via API therefore adding automation:

curl -X GET https://hyper.charm.land/v1/models \
     -H "Content-Type: application/json"

Tests

happy tests:

bun test packages/core/test/sync.test.ts --filter "Hyper"

dry run also happy

bun models:sync hyper --dry-run

sync in sync and happy

bun models:sync hyper

[github-actions]

Based on my analysis of the PR, the diff, the schema, sync code, base model files, and existing provider pricing patterns, here are my findings:

Action items

• [medium] [possible mistake] providers/hyper/models/qwen3.6-flash.toml:11-12, providers/hyper/models/qwen3.6-max.toml:11-12, providers/hyper/models/qwen3.6-plus.toml:12-13, packages/core/src/sync/providers/hyper.ts:354-355 - Check: Verify Hyper API cache_hit/cache_create field semantics match the standard OpenAI convention used in the sync mapping. Why: The sync maps pricing.cache_hit → cost.cache_read and pricing.cache_create → cost.cache_write (standard convention), but the resulting values are inverted from every other provider in the catalog. For all three Qwen3.6 models, cache_read (1.25, 2.5, 2.5) exceeds input (1, 2, 2) while cache_write (0.1, 0.2, 0.2) is far below input — the exact inverse of the 10%/125% ratio documented in AGENTS.md and used by alibaba for the same underlying models (e.g., alibaba's qwen3.6-plus >256K tier: cache_read=0.20, cache_write=2.50, versus Hyper's cache_read=2.5, cache_write=0.2 with identical input=2/output=6). This strongly suggests Hyper's API returns these values under swapped field names, or the values are genuinely inverted. Action: Confirm against Hyper's API documentation whether cache_hit is the read cost (standard) or write cost (non-standard). If the API uses non-standard naming, swap the mapping in buildCost; otherwise confirm the unusual pricing is intentional and cite the source.
• [low] [violation] .pr-review/pull-request.json:4 - Check: PR body must cite direct provider sources for material data changes. Why: AGENTS.md (Citations, recommended) asks data-changing PRs to link to the provider's pricing page, model docs, or API reference in the PR body, with each citation stating what it supports. The PR body only references a superseded PR and test/sync command output; it does not link to Hyper's pricing page, API docs, or model documentation. This matters because the cache pricing values above cannot be verified without a source, and the schema relaxation (removing Provider.env .min(1)) also lacks documented justification. Action: Add first-party citations to the PR body (e.g., Hyper API reference, pricing page) and state which fields each URL supports.

[github-actions]

Action items

• [high] [violation] providers/hyper/provider.toml:2 - Check: Provider env must list auth environment variables; schema requires a non-empty array. Why: This PR sets env = [] and removes Provider.env’s .min(1) in packages/core/src/schema.ts, weakening a global invariant for every provider. No other provider ships an empty env, and README treats env as the auth-key surface for clients. An OpenAI-compatible gateway almost certainly needs a documented key name; empty env leaves consumers unable to configure credentials. Action: Revert the schema change, set the real Hyper API key env var(s) (e.g. whatever Charm documents), or provide first-party evidence that Hyper accepts unauthenticated inference and justify empty env without relaxing validation repo-wide.
• [medium] [violation] providers/hyper/logo.svg:1 - Check: New-provider logo must be compliant SVG (currentColor, no fixed size, preferably square viewBox). Why: The asset is a wide wordmark (viewBox="68 50 1800 526" plus fixed width/height), not a square icon mark. That breaks the contribution logo checklist and will render poorly wherever provider logos are shown as square badges. Action: Replace with a square icon-only SVG using currentColor, no fixed dimensions, and a square viewBox (e.g. 0 0 24 24).
• [medium] [possible mistake] providers/hyper/models/deepseek-v4-flash.toml:10 - Check: Synced costs should be clean per-1M USD values. Why: Several models store float noise (0.20000010875000002, 4.790500000000001, 0.9499997000000001, etc.). Final buildCost in packages/core/src/sync/providers/hyper.ts passes API numbers through without rounding, unlike other sync modules, so bad floats will keep being rewritten on every sync. Action: Round catalog prices in the Hyper sync (same style as other providers) and re-sync so TOML costs are stable decimals.
• [low] [possible mistake] providers/hyper/models/deepseek-v4-flash.toml:5 - Check: Niche-provider reasoning_options should document the exact request syntax and only claim verified controls. Why: Effort options are asserted from /v1/models reasoning.effort_levels, but there is no adjacent API comment (or PR evidence) for the request field/values Hyper accepts (e.g. reasoning_effort vs another path). AGENTS asks niche providers to record that syntax next to options. Action: Add the verified Hyper request syntax next to each [[reasoning_options]] entry (or as a leading provider comment), and drop any effort/toggle claim not backed by Hyper’s API surface.

[github-actions]

Action items

• [high] [violation] packages/core/src/schema.ts:350 - Check: Provider env must list auth environment variable names (README Contributing; schema previously required .min(1)). Why: The PR sets env = [] on Charm Hyper and drops the non-empty env constraint for every provider. Clients lose a documented API-key hook, and empty env becomes valid catalog-wide. Action: Restore env: z.array(z.string()).min(1, ...), and set Hyper’s env to the real auth var(s) (e.g. HYPER_API_KEY or whatever Charm documents)—do not relax the global schema for one provider.
• [high] [possible mistake] packages/core/src/sync/providers/hyper.ts:121 - Check: reasoning / reasoning_options must reflect Hyper’s inference surface; absence of effort metadata is not proof the model does not reason (AGENTS.md; chutes/crossmodel patterns). Why: reasoning: model.reasoning !== undefined forces reasoning = false whenever /v1/models omits the reasoning object. After resync this overrides base metadata for known reasoning models (minimax-m2.7, gemma-4-26b-a4b-it, glm-5, kimi-k2.5, kimi-k2.7-code, qwen3.6-*, qwen3.7-*), which is likely wrong if Hyper still serves thinking behavior without advertising effort levels. Action: Only set reasoning = true (and options) when the API positively advertises reasoning; when the field is absent, leave reasoning unset so base_model is inherited, or map always-on reasoning to reasoning = true + reasoning_options = []. Re-sync and fix the affected TOMLs; cite Hyper evidence for any intentional reasoning = false overrides.
• [medium] [possible mistake] packages/core/src/sync/providers/hyper.ts:67 - Check: Sync should add models the provider serves when a matching models/ entry can be resolved (sync.md; venice/chutes/openrouter). Why: translateModel returns undefined unless a local Hyper TOML already has base_model, so new Hyper catalog IDs are never created and are treated as absent (deleted if present without base_model). Aliases/resolve*BaseModel were removed in the simplify commit. Action: Resolve base_model for unknown IDs (aliases + metadata lookup) and only skip when no canonical metadata exists; keep requiring hand-authored files only when no models/ match is possible.
• [low] [possible mistake] providers/hyper/models/deepseek-v4-flash.toml:10 - Check: Synced costs should be stable per-1M USD values (other sync modules round). Why: Several models store float noise (0.20000010875000002, 1.52425/4.790500000000001, 0.9499997000000001, etc.), which indicates missing rounding in buildCost and hurts diffs/reviews. Action: Round API prices (e.g. to 6 decimal places like other providers) and re-sync so TOMLs get clean numbers.

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/hyper.ts:136 - Check: Provider attachment and modalities must stay consistent when overriding vision support. Why: buildHyperModel sets attachment from capabilities.vision but never writes modalities. That leaves multimodal bases with attachment = false while still inheriting image/video inputs—already visible on gemma-4-26b-a4b-it and llama-4-maverick-17b-128e-instruct-fp8 (and the reverse risk when vision is true but only a vision boolean is known). Consumers will disagree on whether files/images are accepted. Action: Derive both fields together (e.g. vision → attachment = true and modalities.input including image; no vision → attachment = false and text-only input unless another modality is verified). Update affected Hyper TOMLs and cover this in the Hyper sync tests.
• [medium] [possible mistake] packages/core/src/sync/providers/hyper.ts:137 - Check: reasoning / reasoning_options must reflect Hyper’s inference surface, not only upstream models/ metadata. Why: When the API omits reasoning, the sync still emits reasoning_options = [] if the base model reasons, so the catalog claims Hyper reasoning with no control. Hyper’s own /v1/models schema treats reasoning as optional metadata, so omission is a real signal; earlier revisions of this PR set reasoning = false for those rows. Action: Confirm against Hyper whether models without a reasoning object still reason on this provider. If not, set reasoning = false and drop reasoning_options; if they do reason with no control, keep reasoning_options = [] and cite that evidence in the PR body.
• [low] [possible mistake] providers/hyper/models/deepseek-v4-flash.toml:6 - Check: Niche-provider effort options should document the exact request syntax callers must send. Why: Several models advertise [[reasoning_options]] type = "effort" from catalog effort_levels, but neither the TOMLs nor the PR body state the Hyper request field/path (e.g. OpenAI-style reasoning_effort vs another payload). Without that, clients cannot use the declared controls reliably. Action: Add a short adjacent # API: ... comment on the effort options (or PR citation mapped to the claim) with the verified request shape and accepted values.

[github-actions]

Action items

• [high] [possible mistake] packages/core/src/sync/providers/hyper.ts:146 - Check: Provider reasoning must reflect whether the model reasons on this inference provider; absence of controls is reasoning_options = [], not reasoning = false. Why: buildHyperModel sets reasoning: model.reasoning != null, treating a missing Hyper reasoning object (effort metadata) as non-reasoning. That forces reasoning = false on base models that are reasoning = true (minimax/MiniMax-M2.7, alibaba/qwen3.6-*, alibaba/qwen3.7-*, moonshotai/kimi-k2.5, moonshotai/kimi-k2.7-code, zhipuai/glm-5, google/gemma-4-26b-a4b-it), while peers keep those models as reasoning with empty options. Hyper already advertises effort only for some IDs (e.g. kimi-k2.6 vs kimi-k2.5), so omission looks like “no control,” not “does not reason.” Action: Only set reasoning = true when the API confirms reasoning; when the object is absent, leave reasoning unset so base_model can inherit, and emit reasoning_options = [] for inherited reasoning models (restore the TOMLs accordingly). Do not keep reasoning = false unless Hyper evidence shows thinking is disabled for that ID.
• [medium] [possible mistake] packages/core/src/sync/providers/hyper.ts:127 - Check: Provider modality overrides should not drop capabilities the catalog already attributes to the model unless the provider API is authoritative for those modalities. Why: hyperModalities always rewrites input to either ["text"] or ["text","image"] from capabilities.vision. For vision-capable IDs that still have richer base modalities (e.g. kimi-k2.6, kimi-k2.7-code, qwen3.6-flash, qwen3.6-plus with video), sync permanently strips video (and any non-image attachment types). Action: When vision is true, preserve non-image base input modalities (or only add image instead of replacing the whole array); only force text-only when Hyper explicitly reports no vision.
• [medium] [possible mistake] packages/core/test/sync.test.ts:734 - Check: Tests should encode the intended reasoning merge behavior, not the incorrect false override. Why: The Hyper tests expect reasoning: false / missing reasoning_options when API reasoning metadata is omitted for minimax/MiniMax-M2.7, locking in the bad inference above. Action: Update tests so omitted API reasoning inherits base reasoning = true and yields reasoning_options = [] (or equivalent), and only assert reasoning = false when the API affirmatively marks a non-reasoning model.

[oskarkocol]

@rekram1-node the review bot is in a loop 😓 Should a missing reasoning object from Hyper strictly mean "inherits base reasoning" (with empty options)?

All the models in Hyper have reasoning enabled, but they don't expose it as a flag.

Also, for modalities: if a model supports vision, I'm merging "image" into the base modalities rather than overwriting the array so we don't accidentally drop "video". Sound good?

[lgh06]

Thanks for your efforts! can't wait for this PR to be merged.

[rekram1-node]

this lgtm, sorry it fell under radar