feat: split LLM Gateway into two provider catalogs

[smakosh]

What

LLM Gateway is splitting its opencode-facing catalog into two providers served by the same endpoint and API key:

• llmgateway → renamed to "DevPass (LLM Gateway)" — everything else is unchanged: same provider id, same aggregated root-model catalog (claude-sonnet-5, gpt-5.5, …) auto-routed by the gateway, so existing user configs keep working as-is.
• llmgateway-providers ("LLM Gateway") — new — one entry per upstream provider mapping, addressed the way the gateway accepts provider-pinned requests: provider/model-id (anthropic/claude-sonnet-5, azure/gpt-5.5, …), each with that specific deployment's pricing, context/output limits, and capabilities. Names carry the serving provider ("GPT-5.5 (Azure)") so duplicate deployments of the same model stay distinguishable.

Following the pattern of the original llmgateway sync (#2773), this PR ships the sync code plus a small seed catalog (7 representative generated entries covering the factored, full, duplicate-deployment, and capability-delta paths — dev now rejects providers with zero models, so the catalog can't start empty). The scheduled sync automation fills in the rest (~359 entries) once the gateway change is live.

Sync

The new catalog is synced by a second module in sync/providers/llmgateway.ts from GET /v1/models?mapped=true (gateway PR):

• Mapped entries resolve their canonical base_model from the root model id + family (the serving provider prefix is unrelated to the originating lab). In a local run against the gateway build, 283 of 359 generated files factor against existing models/ metadata; the disambiguating name override is deliberately kept on factored files and carried forward on resyncs.
• Each mapped entry ships exactly one provider mapping whose vision/tools/reasoning flags describe that specific deployment; those flags drive capability fields (as overrides on factored entries, with values equal to the base dropped), the deployment's declared reasoning_efforts translate into reasoning_options, and the gateway's served max_output takes precedence over inherited/authored output limits.
• Guard: a deployment without the mapped view ignores the query param and returns unprefixed root ids — parseModels throws in that case instead of syncing wrong ids into the provider-pinned catalog. Until the gateway PR deploys, a scheduled sync of llmgateway-providers fails loudly rather than producing a wrong PR.
• llmgateway/custom (BYO-model placeholder) and llmgateway/auto (the auto-router, already covered by the DevPass catalog) are excluded.
• Schema fix along the way: context_length is optional in the gateway response (absent on pseudo-models and some non-text mappings), so the shared Zod schema and both builders now tolerate that.

Verification

• bun validate passes with the seed catalog.
• Unit tests cover mapped factoring (root id + kept name), capability/limit overrides from the mapping and max_output, max_output precedence on resyncs, and the unprefixed-id refusal guard.
• A full sync run against a gateway build with the mapped view creates all 359 model files, after which bun validate still passes.

[smakosh]

Note on the failing review check: the reviewer bot can't fetch the diff because this PR exceeds GitHub's 300-file diff API limit (PullRequest.diff too_large — 364 files, 359 of them the generated providers/llmgateway-providers/ catalog). validate passes. Happy to split the code change and the generated catalog into separate PRs if you prefer.

[smakosh]

Restructured to sync-code only (7 files, matching how #2773 introduced the original llmgateway sync): dropped the 359 generated catalog files — the scheduled sync automation will populate providers/llmgateway-providers/ once the gateway's mapped view is live. This also unblocks the review check, which previously failed on GitHub's 300-file diff API limit.

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/llmgateway.ts:434 - Check: Mapped brand-new factored path must apply deployment capability flags. Why: PR/docs say each mapped entry’s providers[0].vision/tools/reasoning describe that deployment and the brand-new path should use them, but the canonical-base branch only passes { name, limit, cost }. ~283/359 first-sync files factor here, so deployment-specific capability deltas never land and base reasoning/tool_call/attachment are always inherited. Action: On brand-new factored creates, pass mapping-derived attachment, tool_call, and reasoning (and matching reasoning_options when reasoning) into factorBaseModel so they override base when they differ.
• [high] [violation] packages/core/src/sync/providers/llmgateway.ts:434 - Check: Factored models should inherit base limit.output unless the API reports a real max output. Why: Brand-new factored uses output: model.max_output ?? context. When max_output is absent, output is forced to full context (e.g. 200k), overriding lab defaults (e.g. 64k). Aggregated buildLLMGatewayModel correctly leaves factored output undefined so base inherits. Action: Use only model.max_output when present; otherwise omit output on factored creates (same as the aggregated path).
• [high] [possible mistake] packages/core/src/sync/providers/llmgateway.ts:353 - Check: Mapped sync should treat gateway max_output as authoritative for deployment limits. Why: output is existing?.limit?.output ?? model.max_output ?? context. existing is the resolved merge, so base/inherited output always wins and API max_output is ignored on updates—opposite of OpenRouter (max_completion_tokens ?? existing ?? context) and the PR goal of per-deployment limits. Action: Prefer model.max_output when present; only fall back to existing/context when the API omits it.
• [medium] [violation] packages/core/src/sync/providers/llmgateway.ts:434 - Check: Relay reasoners must not get empty reasoning_options when the host exposes controls. Why: Brand-new mapped models never set reasoning_options; the runner fills [] for reasoning = true. This host documents effort/budget controls (provider.toml), and sibling providers/llmgateway already has curated options for the same API. Empty means “no caller control,” not “unknown.” First population will stamp incorrect [] across the catalog. Action: Seed reasoning_options from the sibling root model (providers/llmgateway/models/<rootID>.toml) when present; otherwise set the real gateway control set for that model—do not rely on default [].
• [medium] [violation] packages/core/test/sync.test.ts:35 - Check: New sync behavior needs focused regression tests. Why: buildLLMGatewayMappedModel and the mapped parseModels guard are large new paths; only buildLLMGatewayModel is covered. Bugs above (root-ID factoring, mapping flags, max_output, unprefixed-ID refusal) can ship unnoticed. Action: Add tests for mapped factoring (root ID + name), capability/limit overrides from providers[0]/max_output, and refusal when the mapped view returns unprefixed IDs.

[smakosh]

Addressed the review items in f5678d9:

• Capability flags on factored creates: brand-new factored entries now pass the mapping's vision/tools/reasoning (and structured_output) into factorBaseModel, so deployment deltas land as overrides and values equal to the base are dropped as before.
• Factored output limit: model.max_output now wins whenever the gateway reports it (creates and resyncs, matching OpenRouter's precedence); when absent, factored creates inherit the base's limit.output and only fall back to context when the base metadata declares no output (output is required downstream).
• reasoning_options: rather than seeding from the sibling llmgateway files, the mapped view exposes each deployment's exact accepted reasoning_efforts, so those translate directly into [{ type = "effort", values = [...] }] — deployment-accurate (e.g. Anthropic mappings get low…max, Azure OpenAI none…xhigh). Mappings that declare no effort values keep the runner's [] behavior, same as the rest of the catalog.
• Tests: added unit tests for mapped root-ID factoring (name kept), capability/limit overrides from providers[0]/max_output, max_output preference on resyncs, and the unprefixed-id refusal guard.

Verified by regenerating the full catalog locally against the gateway build: 359 files, bun validate passes, and the sync test suite passes (the one failing test, DeepInfra preserves live modalities…, fails identically on upstream dev without this branch).

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/llmgateway.ts:167 - Check: Sync modules must not treat a failed/empty upstream payload as an authoritative empty catalog (sync.md deletion behavior). Why: data.every((model) => model.id.includes("/")) is vacuously true for [], so a 200 with no models (or only llmgateway/* / non-text rows that are later filtered out) passes the mapped-view guard and the runner deletes every local llmgateway-providers model. Action: Fail closed when the mapped payload is empty, or when the post-filter set is empty (e.g. require data.length > 0 and a non-empty filtered result before proceeding).
• [medium] [violation] packages/core/src/sync/providers/llmgateway.ts:375 - Check: When deployment flags are treated as authoritative, resyncs must refresh them (not only first create). Why: buildLLMGatewayMappedModel applies mapping.vision / tools / reasoning, reasoning_efforts → reasoning_options, and structured_outputs only on the brand-new branches. Existing factored/full paths still freeze prior attachment / tool_call / reasoning / options and only refresh cost+limit, so a wrong first write or a later API change never corrects the catalog—despite the PR treating those mapping fields as the reliable per-deployment source. Action: On existing paths, refresh capability fields and reasoning_options from the mapping when present (same approach as OpenRouter/live peers), while still preserving true hand overrides only when the API omits the signal.
• [medium] [possible mistake] packages/core/src/sync/providers/llmgateway.ts:358 - Check: Mapped entries are documented as carrying exactly one provider mapping. Why: The code always uses model.providers?.[0] with no length check. Missing/empty providers silently falls back to noisy supported_parameters; multiple mappings silently pick an arbitrary first row for capabilities. Action: Require providers?.length === 1 (or explicitly define merge/skip behavior) before trusting providers[0].

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/llmgateway.ts:379 - Check: Factored mapped resyncs must not invent limit.output when the gateway omits max_output and the file inherits output from base_model. Why: buildLLMGatewayMappedModel builds limit.output as model.max_output ?? existing?.limit?.output ?? context for the existing-factored path. Seeded factored files such as anthropic/claude-opus-4-8, anthropic/claude-sonnet-5, and */gpt-5.5 intentionally omit limit.output so they inherit lab output. On the next sync, if max_output is absent, this stamps output = context and permanently overrides the base (e.g. 128k → 1M). The brand-new path already avoids this via canonicalOutputLimit; the resync path does not. Action: Mirror the brand-new factored logic on resync: prefer max_output, else keep authored output, else leave output unset when the canonical base declares one (only fall back to context when the base has no output).
• [high] [violation] providers/llmgateway-providers/models/perplexity/sonar-pro.toml:1 - Check: Non-lab hosts must use base_model when the lab model is nameable and models/<lab>/<model>.toml exists. Why: Perplexity Sonar Pro already has complete lab metadata at models/perplexity/sonar-pro.toml, and the sibling aggregated catalog uses base_model = "perplexity/sonar-pro". This file is a full inline definition instead, and it also diverges from that lab entry (attachment = false vs true, text-only input vs text+image, output = 200_000 vs 8_192). Action: Replace the inline definition with an override-only file pointing at perplexity/sonar-pro (keep provider name/cost and only real deployment deltas). Fix the sync path that failed to factor this id (missing/wrong family or resolver miss) so future syncs do not re-expand it.
• [high] [violation] packages/core/src/sync/providers/llmgateway.ts:361 - Check: Multi-model relay reasoning_options must follow this host’s real controls and must not stamp [] just because effort metadata is missing when lab/same-surface peers expose controls. Why: Mapped sync only emits { type = "effort", values = reasoning_efforts } and otherwise leaves options unset, so the runner defaults new reasoners to []. Seeded embercloud/glm-5.1 therefore has reasoning_options = [] while first-party Zhipu/ZAI and the same gateway’s aggregated providers/llmgateway/models/glm-5.1.toml use toggle (plus interleaved). That is the empty-from-uncertainty anti-pattern for a reasoner on this unified gateway surface. Action: When reasoning_efforts is absent/empty, do not invent [] for models whose same-host/lab baseline is toggle (or other non-effort controls). Author toggle (with a leading wire comment) or preserve/copy the aggregated llmgateway options for that root model; update the glm-5.1 seed accordingly.
• [medium] [violation] packages/core/src/sync/providers/llmgateway.ts:400 - Check: Mapped sync treats deployment reasoning_efforts as authoritative for new entries and should keep them current on resync. Why: The existing-factored branch refreshes cost/limit/name but never passes reasoning_options from mapping.reasoning_efforts. After the first write, the runner only preserves whatever was already on disk, so later API effort changes never apply and a bad first stamp (e.g. []) is sticky. Action: On existing factored/full paths, set reasoning_options from reasoning_efforts when present; when absent, preserve existing options (so hand-curated toggle/budget is not wiped).
• [medium] [possible mistake] providers/llmgateway-providers/models/together-ai/kimi-k2.6.toml:7 - Check: Effort lists on this gateway should match the deployment’s real accepted values and same-host peers for the model. Why: This seed allows only effort = none while reasoning still inherits true from moonshotai/kimi-k2.6. The same gateway’s aggregated kimi-k2.6 entry documents a full none|minimal|low|medium|high|xhigh|max set, and the Moonshot lab entry uses toggle. ["none"] alone is an odd control surface (effectively “off only”) for a reasoner. Action: Verify Together’s mapped reasoning_efforts (and whether reasoning is actually enabled). Align to the real set—or set reasoning = false / drop the lone none effort if the deployment does not expose reasoning controls.

[smakosh]

Round 2 addressed in 48bf376:

• Factored resync output: now mirrors the create path — served max_output wins when reported; otherwise the authored output is kept, and when neither exists the file keeps inheriting the base's limit.output (context fallback only when the base declares none). Covered by a test asserting no limit override is invented on such resyncs.
• sonar-pro factoring: root cause was the canonical resolver — perplexity was missing from CANONICAL_PROVIDER_PREFIXES, so Sonar ids never matched their lab metadata (this also lets OpenRouter's perplexity/* ids factor). The seed is now override-only (base_model = "perplexity/sonar-pro" + name/cost and the real deployment deltas).
• reasoning_options: three-level fallback now — deployment-declared reasoning_efforts win; else non-empty curation on the file; else the aggregated llmgateway catalog's curated options for the same root model (same gateway surface). A curated [] counts as unknown, so a bad first stamp heals on resync. Additionally the runner no longer stamps [] onto factored reasoners whose base metadata declares reasoning_options — the empty array would shadow the base's real controls at generate time (the existing preserveReasoningOptions default-[] test still passes; the new behavior only kicks in when the base declares options). Across the full catalog this drops empty-[] files from 128 to 36 (models with no declared efforts, no sibling curation, and no base options). embercloud/glm-5.1 now inherits toggle.
• together-ai/kimi-k2.6: verified against the gateway's catalog — Together's deployment accepts any effort string without validating it and no tier changes reasoning length; the only honoured control is none via the thinking switch. That is a plain on/off switch, so a ["none"]-only effort list now translates to { type = "toggle" }, matching the Moonshot lab entry.
• Resync effort refresh: existing factored/full paths now update reasoning_options from reasoning_efforts when present and preserve curated options otherwise.

Verified with a full local regeneration (359 files, bun validate green) and the sync test suite (117 pass; the DeepInfra modalities test fails identically on dev without this branch).

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/llmgateway.ts:167 - Check: Mapped sync must not authoritatively delete the catalog on an empty or fully filtered response. Why: data.every((m) => m.id.includes("/")) is true for [], and after filtering llmgateway/* / non-text the desired set can also be empty. With default deleteMissing, that wipes the seed catalog (and later the full mapped tree) instead of failing loudly like the unprefixed-id guard. Action: Reject empty raw data and an empty post-filter set (throw), same spirit as the unprefixed-id check.
• [medium] [possible mistake] providers/llmgateway-providers/models/openai/gpt-5.5.toml:8 - Check: Provider cost must include context tiers when this host/deployment uses them. Why: First-party providers/openai/models/gpt-5.5.toml and same-surface providers/llmgateway/models/gpt-5.5.toml both author [[cost.tiers]] at 272k (10/45/1). Mapped cost only preserves existing?.cost?.tiers and never invents tiers, so these seeds (and any later first-create sync) publish flat 5/30 pricing and understate large-context cost. Action: Verify gateway OpenAI (and Azure, if identical) pricing; if tiers apply, add the 272k tier to openai/gpt-5.5.toml and azure/gpt-5.5.toml so resync preserves them.
• [medium] [violation] providers/llmgateway-providers/models/perplexity/sonar-pro.toml:3 - Check: attachment and [modalities].input must stay consistent after base_model merge. Why: Seed sets attachment = false while inheriting lab modalities.input = ["text", "image"], so the resolved model claims image input but no attachments. buildLLMGatewayMappedModel only maps mapping.vision → attachment and never overrides modalities when vision is off, so this will recur for other vision-capable bases. Action: When vision is false, also override modalities (drop non-text inputs, or match architecture); fix the sonar-pro seed the same way.
• [medium] [violation] providers/llmgateway-providers/models/embercloud/glm-5.1.toml:5 - Check: Every toggle needs a leading top-of-file wire-path comment (AGENTS.md → Reasoning options). Why: embercloud/glm-5.1.toml and together-ai/kimi-k2.6.toml ship type = "toggle" with no leading comment; provider.toml docs are not a substitute for the per-file requirement, and sync preserves only leading headers. Action: Add a leading comment on both toggle seeds documenting the exact gateway control (e.g. thinking / reasoning_effort=none path used for on/off).
• [medium] [possible mistake] providers/llmgateway-providers/models/embercloud/glm-5.1.toml:1 - Check: Same-host reasoners that emit a reasoning side channel should set interleaved. Why: Lab and same-endpoint peers (providers/zhipuai/models/glm-5.1.toml, providers/llmgateway/models/glm-5.1.toml, and the kimi-k2.6 counterparts) set [interleaved] field = "reasoning_content". Mapped seeds omit it, and brand-new factored sync never copies it, so clients miss the side channel on this API. Action: Add interleaved to the glm-5.1 and kimi-k2.6 seeds (and consider preserving/copying it in mapped sync from sibling aggregated entries when present).

[smakosh]

Round 3 addressed in 545d19a:

• Empty-response wipe: both syncs now throw on an empty raw response, and the mapped sync also throws when filtering leaves nothing — same spirit as the unprefixed-id guard, so the delete-missing pass can never be fed an empty desired set. Test added covering both.
• attachment vs modalities: factored creates from a vision-less deployment now override modalities with the deployment's own (text-only) inputs, so attachment = false can't coexist with inherited image input. The sonar-pro seed is regenerated accordingly (input = ["text"]).
• interleaved: mapped entries now copy the interleaved side-channel from the aggregated llmgateway entry for the same root model when the deployment reasons (same wire surface; existing curation wins on resyncs). glm-5.1 and kimi-k2.6 seeds carry field = "reasoning_content".
• Toggle comments: both toggle seeds now have the leading wire-path header (reasoning_effort = "none" → deployment thinking switch; headers survive resync).
• gpt-5.5 tiers — good catch, and it went deeper than the seeds: the gateway's own catalog supports context-tiered billing and gpt-5.6-sol declares exactly this 272k tier, but gpt-5.5 predates that support and was never backfilled — i.e. the gateway currently bills long-context gpt-5.5 flat. We're fixing the gateway catalog (10/45, cache read 1.00 over 272k, both OpenAI and Azure), and the seeds here now author the same [[cost.tiers]] so resync preserves them.

bun validate green, 118 sync tests pass (DeepInfra modalities failure is pre-existing on dev).

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/llmgateway.ts:420 - Check: Optional context_length must not author limit.context = 0. Why: Mapped (and aggregated) translation does const reported = model.context_length ?? 0 then context = reported > 0 ? reported : existing?.limit?.context ?? reported. On a brand-new model with missing/zero context_length, existing is undefined, so context becomes 0 and is written into [limit] (and often output via the same fallback). That overrides a real base context with an unusable window and can pass schema (min(0)). Action: When no positive served context exists, leave limit.context unset on factored creates (inherit base) and do not fall back to 0; refuse or skip unfactored creates that lack a positive context rather than publishing zero limits.
• [medium] [violation] packages/core/src/sync/providers/llmgateway.ts:547 - Check: vision === false must not leave image/pdf input modalities. Why: Brand-new factored path sets modalities: mapping?.vision === false ? defaultModalities(model) : undefined, which trusts architecture.input_modalities. If the API keeps model-level image/pdf inputs while the deployment flag is vision: false, the file still gets attachment = false with image/pdf input—the contradiction the comment says to prevent. The full (unfactored) create path never applies this guard at all (attachment from mapping.vision, modalities always from architecture). Action: When mapping.vision === false, force non-image input (e.g. text-only, or strip image/pdf) on both factored and full create paths; add a test where architecture still lists image but vision is false.

[smakosh]

Round 4 addressed in 7ddf9e9:

• Zero limits: a missing/zero context_length can no longer be authored as limit.context = 0 in either builder. Factored entries (creates and resyncs) leave context unset and inherit the base; unfactored creates without a positive served context are skipped via translateModel returning undefined (and reported through sourceID) instead of publishing unusable limits. The output fallbacks follow the same rule, so no path can cascade a 0 into limit.output either.
• vision vs modalities: vision: false now forces non-image input from the mapping itself (deploymentModalities strips image/pdf, falling back to ["text"]) rather than trusting the model-level architecture, on the factored create path, the unfactored create path, and the existing-full fallback alike.
• Tests: skipped-create (both builders), factored create without served context authors no limit at all, and a vision-false-with-image-architecture case asserting text-only input on both create paths.

Also verified end-to-end: running the sync over the committed seed files leaves all 7 byte-identical (leading toggle headers and the authored [[cost.tiers]] survive resync) while creating the other 352, and bun validate passes. 121 sync tests pass (DeepInfra modalities failure remains pre-existing on dev).

[rekram1-node]

These are the provider names you want to go with? They seem a bit interesting.... Not intuitive.

[smakosh]

These are the provider names you want to go with? They seem a bit interesting.... Not intuitive.

We want to have two providers:

• LLM Gateway
• DevPass By LLM Gateway

one is model IDs only and the other has models mapped by providers.

[github-actions]

Action items

• [medium] [violation] providers/llmgateway-providers/logo.svg:1 - Check: New provider logo.svg must use currentColor with no fixed size or hardcoded colors. Why: The root <svg> sets width="24" height="24", which violates the new-provider logo rules in AGENTS.md (merge blocker). Action: Drop the fixed width/height attributes and keep a square viewBox with currentColor fills only (same shape as the compliant examples in AGENTS.md).
• [medium] [possible mistake] packages/core/src/sync/providers/llmgateway.ts (mapped parseModels / buildLLMGatewayMappedModel) - Check: Mapped catalog entries are defined as one upstream mapping whose vision/tools/reasoning/reasoning_efforts drive capabilities and reasoning_options. Why: parseModels only rejects empty/unprefixed payloads; it still accepts mapped IDs with missing or empty providers, and the builder then silently falls back to noisy supported_parameters / sibling defaults. That can author wrong capability overrides or [] reasoning controls for the provider-pinned catalog. Action: Fail the mapped sync when any kept model lacks exactly one providers[] entry (same fail-loud style as the unprefixed-id guard), and add a unit test for that case.

[github-actions]

Action items

• [medium] [violation] packages/core/src/sync/providers/llmgateway.ts - Check: Override-only base_model files must not restate or invent identical/synthetic lab metadata on resync. Why: Brand-new mapped factored creates intentionally author only a disambiguating name (no description), but the existing-factored path does existing.description ?? describeModel(...). On the first automation resync, describeModel synthesizes a generic blurb (e.g. “Flagship reasoning model…”) that differs from the lab description and gets written as a sticky override on every name-pinned mapped entry. Action: On the existing-factored path, pass only description: existing.description (no describeModel fallback), matching other factored providers (e.g. merge-gateway/cortecs); leave description unset so the lab text keeps inheriting.
• [medium] [violation] packages/core/src/sync/providers/llmgateway.ts - Check: Every toggle reasoning control needs a leading top-of-file wire-path comment (AGENTS.md → Reasoning options). Why: reasoning_efforts: ["none"] is translated to { type: "toggle" } with no mechanism to attach a header. The seed toggle files include the required comment, but bulk creates from the scheduled mapped sync will emit headerless toggles; sync only preserves an existing header, it never creates one. Action: When emitting toggle (none-only efforts or sibling toggle), ensure new files get a leading wire comment (e.g. default header for this provider’s $.reasoning_effort = "none" path), or avoid writing toggle until that header can be authored.

[github-actions]

Action items

• [medium] [violation] packages/core/src/sync/providers/llmgateway.ts - Check: Mapped factored resyncs must keep deployment capability flags aligned with providers[] (same authority as brand-new creates / merge-gateway). Why: On existing base_model entries the builder freezes resolved reasoning / attachment / tool_call / structured_output / modalities from the prior file, but still refreshes reasoning_options from live mapping.reasoning + efforts. If a deployment later reports reasoning: false (or the reverse), options are cleared or filled while the boolean stays stuck, so preserveReasoningOptions can stamp [] on a non-reasoner or drop controls on a reasoner. Hourly resyncs will not correct the seed catalog or any later drift. Action: On the existing-factored path, apply mapping flags the same way as the create path (reasoning/vision/tools/structured_outputs and vision-driven modalities), or freeze reasoning_options with the other curated fields; do not refresh options from a mapping.reasoning value you are not also writing.
• [low] [possible mistake] providers/llmgateway-providers/models/together-ai/kimi-k2.6.toml:8 - Check: Provider-pinned capabilities/limits should match the upstream Together deployment this id represents. Why: Seed sets tool_call = false, structured_output = false, and limit.output = 32_768, while first-party providers/togetherai/models/moonshotai/Kimi-K2.6.toml has tools/structured output enabled and output = 131_000, and the seed’s unit prices match that Together row. That combination looks like a bad mapping dump rather than a real Together pin. Action: Re-check providers[].tools / structured_outputs / max_output for together-ai/kimi-k2.6 on the mapped catalog and correct the seed (or cite why this pin is intentionally weaker than Together’s own API).

[github-actions]

Action items

• [medium] [violation] packages/core/src/sync/providers/llmgateway.ts - Check: Mapped resyncs must realign vision-related modalities in both directions (PR claims deployment flags stay authoritative on resyncs; attachment/image input must stay consistent). Why: Factored and full resync paths only force text-only modalities when mapping.vision === false; when vision later becomes true they keep existing.modalities. A prior no-vision override (e.g. input = ["text"]) therefore stays sticky after vision returns, so the file can keep text-only input while attachment flips back to true/inherited vision — the inverse of the contradiction the sync already tries to prevent. Hourly resync will not self-heal. Check: mapping?.vision === false ? deploymentModalities(...) : existing.modalities. Action: When mapping.vision === true (or whenever vision is present), recompute modalities via deploymentModalities(model, mapping.vision) or clear the override so base image/pdf inputs can inherit again; add a resync test for vision false → true.

[github-actions]

Action items

• [medium] [possible mistake] packages/core/src/sync/providers/openrouter.ts:37 - Check: Shared canonical prefix tables must not change unrelated providers without intent. Why: Adding perplexity to CANONICAL_PROVIDER_PREFIXES makes resolveCanonicalBaseModel("perplexity/…") succeed for every caller (OpenRouter, Kilo, Vercel, Merge Gateway, …). Today those hosts still ship full standalone Perplexity TOMLs (e.g. providers/openrouter/models/perplexity/sonar-pro.toml); the next automation sync will start factoring them against models/perplexity/*, which is an unstated cross-provider behavior change. LLM Gateway can resolve the same lab IDs via resolveModelMetadataBaseModel (exact models/ path match) without widening the OpenRouter prefix map. Action: Prefer resolveModelMetadataBaseModel (or a llmgateway-local alias) for mapped root IDs; only keep the shared prefix if you intentionally want OpenRouter/Kilo/etc. to factor Perplexity on the next sync and call that out.
• [medium] [violation] packages/core/src/sync/providers/llmgateway.ts:294 - Check: Missing/zero context_length must never be authored as limit.context = 0 (stated goal of the zero-context fix). Why: Both builders compute servedContext correctly, then the existing-full limit object still does context: context ?? reported (and output: … ?? reported). When the API omits/zeros context and the existing full row has no limit.context, reported is 0/undefined and the file is rewritten with limit.context = 0 (schema min(0) allows it). Factored paths avoid this; the full-model resync path undoes the guard. Action: Stop falling back to reported for authored limits—use only a positive context/servedOutput, and skip or preserve the existing full row when neither the API nor the file provides a usable context.

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/llmgateway.ts:493 - Check: Mapped reasoning_options must keep this host’s full control surface (effort and any real budget_tokens), and resync must not wipe hand-authored non-effort controls. Why: When providers[].reasoning_efforts is non-empty, the builder sets reasoning_options to only that effort list (or a lone toggle for ["none"]) and short-circuits past existing/sibling options. The same gateway surface already curates budget_tokens on aggregated peers (e.g. claude-sonnet-4-6, claude-opus-4-5, gemini-2.5-pro) and documents $.reasoning.max_tokens in provider.toml. The bulk mapped sync will therefore omit budgets on create and erase them on every later resync whenever efforts are present. DigitalOcean’s sync already preserves non-effort options when applying remote efforts. Action: When applying deployment efforts, merge them with non-effort options from existing and/or the aggregated sibling (keep budget_tokens; replace only the effort/toggle entry), and add a regression test for a sibling/existing budget_tokens + remote reasoning_efforts case.
• [medium] [possible mistake] packages/core/src/sync/providers/llmgateway.ts:500 - Check: Brand-new mapped creates should not drop known gateway context pricing tiers that the DevPass sibling already authors. Why: cost.tiers is only taken from existing?.cost?.tiers, never from the aggregated llmgateway sibling. GPT-5.5 seeds were hand-fixed with the 272k tier, but the scheduled full sync (~359 files) will create other tiered models (GPT-5.5-pro, Grok 4.x, Gemini, etc.) without those tiers, understating long-context price until someone re-authors them. Action: Seed cost.tiers from the aggregated sibling (or equivalent curated source) on create when the API does not supply tiers; preserve existing tiers on resync as today.

[github-actions]

I've finished reviewing the split provider, sync module, seeds, logo, and reasoning-option handling against the project rules.

[smakosh]

Should be ready now @rekram1-node