Feat/edenai provider

[YacineMK]

Summary

Adds Eden AI sync mechanism (provider scaffold, logo, sync module).

Eden AI is an EU-based AI aggregation platform that exposes 100+ models from many providers behind a single OpenAI-compatible API (https://api.edenai.run/v3, Authorization: Bearer,
provider/model format), with unified billing/usage tracking and EU data residency options. This adds the edenai provider using @ai-sdk/openai-compatible (no new dependency), matching
the pattern of existing aggregator entries like openrouter / kilo.

Changes

• providers/edenai/provider.toml — provider entry
• providers/edenai/logo.svg — monochrome logo (currentColor via SVG mask)
• packages/core/src/sync/providers/edenai.ts — sync module modeled on openrouter.ts / llmgateway.ts
• packages/core/src/sync/index.ts — registered in providers + aggregators group

Design notes

The sync treats Eden's /v3/models as authoritative only for cost, context, and modality data. Capability, open_weights, and reasoning-control fields default to safe values (false,
[]) when Eden's API doesn't provide them, following the llmgateway pattern. Existing hand-authored fields are preserved via context.existing().

Only text-output models are synced; specialty Eden endpoints (OCR, TTS, embeddings, sentiment) live on other Eden APIs and are out of scope.

Catalog

The generated catalog is ~800 files, which exceeds GitHub's 300-file PR-diff limit and would block the pr-reviewer workflow. Following the pattern from kilo #2997 and hyper #3352,
the catalog will land in a follow-up PR from an automation/sync-models-edenai branch that bypasses the reviewer by design.

To generate locally after merge:
bun models:sync edenai

Sources

• Eden AI catalog: https://api.edenai.run/v3/models
• Eden AI docs: https://www.edenai.co/docs

Verification

• bun models:sync edenai --dry-run → 796 planned, 0 errors
• bun validate → passes
• packages/sdk tests → 23/23 pass

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/edenai.ts:245 - Check: open_weights is model-intrinsic and must inherit via base_model when metadata exists; provider sync must not invent overrides. Why: Final code sets open_weights = existing?.open_weights ?? false and always passes it into factorBaseModel. For new factored models that becomes false, which overrides bases with open_weights = true (e.g. models/minimax/MiniMax-M2.1.toml). Patch 5 already produced that bad override (open_weights = false on minimax/MiniMax-M2.1). Hourly Eden sync will republish incorrect catalog data. Action: For factored entries omit open_weights (use undefined so metadata inherits), matching deepinfra/hyper. Only set open_weights on full inline models, and prefer existing?.open_weights there rather than a blanket false.
• [medium] [possible mistake] packages/core/src/sync/providers/edenai.ts:366 - Check: PR design and AGENTS.md: Eden is authoritative for cost/context/modalities; capability fields should not blindly override models/ facts. Why: Factored path still injects Eden-derived reasoning, tool_call, structured_output, and attachment into factorBaseModel. Peers such as llmgateway pass only gateway-authoritative fields for new factored models. Noisy Eden flags can permanently override curated capabilities on first sync. Action: For resolvedBase !== undefined, limit overrides to cost, served limit.context, modalities (if treated as authoritative), dates, and reasoning_options when needed; leave other capabilities unset so they inherit.
• [medium] [violation] packages/core/src/sync/providers/edenai.ts:1 - Check: New complex sync translators need focused regression coverage when they can corrupt catalog data. Why: buildEdenAIModel has no tests in packages/core/test/sync.test.ts, while similar aggregators (hyper, llmgateway) do. The open_weights/base_model bug would have been caught by a factored open-weight case. Action: Add unit tests for factored open-weight inheritance, missing output-limit handling, and full-inline fallbacks.
• [low] [possible mistake] .github/workflows/sync-models.yml:73 - Check: New sync providers that need credentials must wire secrets into the hourly workflow (sync.md). Why: Fetch uses EDENAI_API_KEY when set, but the workflow never exports it. Final PR ships zero model TOMLs and relies on automation; if /v3/models requires auth in CI, Eden stays empty after merge. Action: Confirm unauthenticated catalog access in CI; if a key is required, add EDENAI_API_KEY to the sync job env (and repo secrets). Update the PR body to match the deferred-catalog approach.

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/edenai.ts:340 - Check: Provider-agnostic capabilities must not be clobbered with false defaults when factoring via base_model (see DeepInfra/LLM Gateway sync patterns and AGENTS.md merge rules). Why: reasoning, tool_call, and structured_output are set with === true (missing API flags become false), and open_weights defaults to false for new models. Those values are always passed into factorBaseModel(), so when Eden omits a capability the sync writes explicit false overrides that replace inherited metadata (e.g. open-weight or tool-capable bases). The PR itself notes Eden is not authoritative for capabilities, but the code still treats absences as negatives. Action: Only emit capability fields when Eden explicitly reports them (or when overriding a curated existing value). For resolved base_model entries, prefer the DeepInfra/LLM Gateway approach: override cost/limit/modalities (and provider-specific reasoning_options) and leave tool_call / reasoning / structured_output / open_weights unset so they inherit from models/.
• [high] [violation] packages/core/src/sync/providers/edenai.ts:237 - Check: New provider models must use base_model when a matching models/<provider>/<model>.toml exists. Why: candidateBaseModels() only maps owned_by via DIRECT_METADATA_PROVIDER or bare/exact IDs. Gateway routes such as amazon/anthropic.claude-opus-4-6-v1 never resolve to anthropic/claude-opus-4-6 (Bedrock-style provider.model IDs, -v1:0, region suffixes). The initial sync catalog in this PR already produced full inline amazon/anthropic entries without base_model while models/anthropic/… exists—the same result automation will recreate after merge. Action: Extend resolution for gateway IDs (parse anthropic.* / meta.* / openai.* style names, strip Bedrock/region suffixes, reuse resolveCanonicalBaseModel / stronger aliases) so underlying canonical metadata is factored; add focused tests covering at least one amazon/anthropic and one direct anthropic/openai path.

[github-actions]

Action items

• [medium] [possible mistake] packages/core/src/sync/providers/edenai.ts:296 - Check: Factored base_model entries must keep attachment consistent with any modality overrides (or leave both unset so curated models/ metadata inherits). Why: The factored path always writes Eden modalities but never sets attachment, while the comment claims DeepInfra/LLM Gateway behavior—those peers leave modalities unset for new factored models. After merge, base attachment can disagree with Eden modalities (e.g. attachment = true with text-only input), so the automated catalog will ship inconsistent capability data. Action: Either omit modalities on the factored path so base metadata inherits, or set attachment from the same Eden input modalities whenever modalities are overridden.
• [medium] [possible mistake] packages/core/src/sync/providers/edenai.ts:1 - Check: New sync modules with non-trivial ID resolution and factored/inline splits need focused regression coverage. Why: normalizeGatewayName, Bedrock candidate factoring, and the factored vs inline field split already required multiple fixup commits (output limits, open_weights, capability clobbering). Without unit tests, the deferred automation catalog is likely to reintroduce the same class of bugs. Action: Add focused buildEdenAIModel tests covering Bedrock-style IDs resolving to base_model, factored field selection (no fabricated limit.output / capability clobber), and attachment/modality consistency.

[github-actions]

Action items

• [high] [violation] providers/edenai/logo.svg:1 - Check: New providers must ship a logo with no hardcoded colors; use currentColor for fills/strokes so it adapts when inlined. Why: The logo still hardcodes fill="white" / fill="black" inside a <mask>. Site rendering rewrites every non-currentColor fill to currentColor (packages/web/src/render.tsx providerLogoSvg), which collapses the mask cutouts into a solid rect and breaks the mark in the UI. Action: Replace the mask/cutout approach with a monochrome path (or equivalent) that uses only currentColor / none—no white/black fills—so the logo survives inlining and matches AGENTS.md.

[github-actions]

No actionable findings.

[YacineMK]

Friendly ping @rekram1-node CI is green (796 models planned, 0 errors on dry-run, all 23 SDK tests pass) This follows the same catalog-deferred pattern as #2997 (kilo) and #3352 (hyper) Happy to answer any questions

[MVS-source]

@rekram1-node could you take a look when you get a chance? This adds Eden AI as a sync provider (mirrors the OpenRouter/LLM Gateway pattern) and is needed to unblock Eden AI showing up as a provider in Kilo Code. Happy to help test or answer questions about the API responses.

[rekram1-node]

You can't make prs for adding a provider and then not add any models for it, it will break consumers of the api.

[github-actions]

Action items

• [high] [possible mistake] providers/edenai/models/xai/grok-4.toml:1 - Check: Provider model fields must be internally consistent and match the served model. Why: The entry’s description calls Grok 4 a reasoning model with structured outputs and image+text inputs, but the authored flags say reasoning = false, structured_output = false, attachment = false, and modalities.input = ["text"] only. Peer catalogs disagree on reasoning for this ID, but this file contradicts itself. Action: Reconcile capabilities/modalities (and description) against Eden’s actual /v3/models entry and first-party Grok 4 docs; if it reasons or accepts images, set the matching fields (and reasoning_options = [] when reasoning is true with no verified control).
• [medium] [violation] providers/edenai/models/mistral/codestral-latest.toml:2 - Check: With base_model, provider-agnostic release_date should inherit from models/ unless the provider truly differs. Why: Seeds override canonical release dates with gateway/catalog dates: codestral-latest uses 2026-06-05 (base 2024-05-29), mistral-large-latest uses 2024-02-26 (base 2024-11-01), and MiniMax-M3 uses 2026-05-31 (base 2026-06-01). Those look like Eden listing/created timestamps, not model launches. Action: Drop the incorrect release_date overrides from those seed TOMLs so inheritance wins.
• [medium] [violation] packages/core/src/sync/providers/edenai.ts - Check: Factored sync should not stamp provider-agnostic release_date from API created. Why: buildEdenAIModel sets release_date from model.created on the factored path, so the next bun models:sync edenai will reintroduce bad overrides whenever Eden’s timestamp differs from models/ (same failure mode as the seed files). Peers such as NanoGPT/LLM Gateway leave release_date unset for new factored rows so base metadata inherits. Action: On the factored path, omit release_date (or only preserve an already-authored provider override); keep created/today fallbacks for full inline models only.

[github-actions]

Action items

• [medium] [possible mistake] providers/edenai/models/openai/o3.toml:3 - Check: reasoning_options must reflect Eden AI’s OpenAI-compatible request surface, not only that the model reasons. Why: Seed sets reasoning_options = [], but same-model peers on comparable OpenAI-compatible / chat-completions surfaces expose effort (openai/o3, openrouter/kilo/cloudflare-ai-gateway: low|medium|high). Empty is only correct if Eden was checked and does not accept a control. Action: Verify against Eden docs or a live request whether reasoning_effort / equivalent is accepted for openai/o3; if yes, author the verified effort values (with API syntax comment if non-obvious); if no, keep [] and note the negative check in the PR body.
• [medium] [possible mistake] providers/edenai/models/openai/gpt-5.6-terra.toml:3 - Check: Same reasoning_options surface check for GPT-5.6 Terra. Why: Seed uses [], while native OpenAI and OpenRouter list effort including none|low|medium|high|xhigh|max for this model on OpenAI-style APIs. Action: Confirm Eden passthrough/rejection of effort for openai/gpt-5.6-terra and either match verified values or keep [] with explicit negative evidence.
• [medium] [possible mistake] providers/edenai/models/moonshot/kimi-k3.toml:3 - Check: Reasoning controls for Kimi K3 on an OpenAI-compatible gateway. Why: Seed uses [], but OpenRouter (effort low|high|max) and LLM Gateway (broader effort enum) document caller controls for the same underlying model on compatible surfaces; native Moonshot also exposes toggle/effort. Action: Verify what Eden accepts for moonshot/kimi-k3 and update reasoning_options (or document why none apply).
• [low] [possible mistake] providers/edenai/models/xai/grok-4.5.toml:3 - Check: Reasoning controls for Grok 4.5. Why: Seed uses [] while native xAI authors effort low|medium|high for xai/grok-4.5. Action: Confirm whether Eden forwards xAI/OpenAI-style effort for this ID; set verified options or keep [] with a short negative note.
• [low] [violation] packages/core/src/sync/providers/edenai.ts:387 - Check: Sync must not drop provider-authored reasoning_options when Eden only reports supports_reasoning. Why: Factored path sets reasoning_options = existing?.reasoning_options ?? [] whenever Eden reports reasoning, so a first-time sync of a new reasoning model always writes [], and later runs only preserve what is already on disk—never discovering OpenAI-compatible effort that peers expose. That bakes unverified empties into the automated catalog. Action: Keep [] only as the unverified default, but document in the module/PR that effort/budget must be hand-audited for OpenAI-compatible routes (or detect/preserve richer options once verified), and avoid treating Eden’s boolean as proof that no control exists.

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/edenai.ts:17 - Check: reasoning_options must reflect Eden AI’s verified request surface; use [] when no caller control is verified (AGENTS.md + audit-reasoning-options). Why: edenDefaultReasoningOptions() stamps one full effort enum (none/minimal/low/medium/high/xhigh/max) on every reasoning model. The PR body itself says Eden does not expose a control surface and should default to []. That default will also be applied on the first automated full-catalog sync. Same-surface peers do not use this universal enum (e.g. OpenRouter/Cloudflare AI Gateway o3 → low|medium|high; OpenRouter MiniMax-M3 → []; OpenRouter Gemini 2.5 → budget_tokens/toggle, not this effort set). Action: Default new reasoning models to reasoning_options = [] unless Eden-specific evidence (docs, OpenAPI, or reproduced requests with negative controls) verifies each control and value. Do not copy the schema enum or peer values mechanically.
• [high] [violation] providers/edenai/models/openai/o3.toml:4 - Check: Seed reasoning models must not claim unverified effort values. Why: All seed reasoning entries (Anthropic Claude Haiku/Opus/Sonnet, DeepSeek V4 Flash/Pro, Gemini 2.5 Flash/Pro, MiniMax-M3, Kimi K3, GPT-5.6 Terra, o3, Grok 4.5) share the same fabricated effort list. That contradicts AGENTS.md and peer OpenAI-compatible catalog entries for the same models. Action: Replace those blocks with reasoning_options = [], or supply Eden-specific evidence and model-accurate options (not one global enum) before merge.

[YacineMK]

hi @rekram1-node I was following the same pattern as Hyper #3352 and Kilo #2997 which I mentioned before, because ci fail when u have more than 300 files But fair point I've added a curated seed of flagship models. The full catalog will follow through the automated sync

[github-actions]

Action items

• [high] [violation] providers/edenai/models/openai/o3.toml:3 - Check: Relay reasoning_options must follow lab + same-surface peers; [] means verified no caller control, not uncertainty. Why: Every reasoning seed sets reasoning_options = [] while lab/OpenRouter peers expose real controls (e.g. o3 → effort low|medium|high; gpt-5.6-terra → none|low|medium|high|xhigh|max; DeepSeek V4 → toggle + high|max / peer effort; Claude Sonnet/Opus → graded effort; Gemini 2.5 → toggle + budget_tokens; Kimi K3 → effort low|high|max; Grok 4.5 → effort low|medium|high). That misreports Eden’s API surface. Action: Replace [] on all reasoner seeds (anthropic/*, deepseek/*, google/*, moonshot/kimi-k3, openai/gpt-5.6-terra, openai/o3, xai/grok-4.5, and MiniMax only if peers/docs show a control) with the peer/lab option set; add top-of-file wire comments for any toggle.
• [high] [violation] packages/core/src/sync/providers/edenai.ts:304 - Check: Aggregator sync must not invent empty reasoning_options when controls are unknown. Why: Factored and inline paths do existing?.reasoning_options ?? [] whenever reasoning is true. Future bun models:sync edenai will re-seed [] for every reasoner and wipe correct options unless they were already hand-authored. Action: Stop defaulting to []. Preserve authored options; when missing, leave unset for runner/existing handling only after deriving from known base_model/peer controls, or skip auto-creating reasoners until options are curated.
• [medium] [possible mistake] providers/edenai/models/moonshot/kimi-k3.toml:10 - Check: Provider modality overrides must be real host deltas, not incomplete API snapshots. Why: Seed forces input = ["text", "image"] while models/moonshotai/kimi-k3 includes video. Sync always writes Eden modalities/attachment on the factored path, so noisy API modality lists can permanently drop lab modalities. Action: Confirm Eden’s served modalities for each override (kimi-k3 video; mistral-large pdf; etc.). If Eden is incomplete, omit modalities on factored sync (cost/context only, like LLM Gateway) and drop unverified seed modality blocks.

[github-actions]

Action items

• [high] [violation] providers/edenai/models/openai/gpt-5.6-terra.toml:3 - Check: Relay reasoning_options must match lab/same-surface peers; [] means verified no caller control, not uncertainty (AGENTS.md Reasoning options). Why: Eden AI is a multi-model OpenAI-compatible aggregator. Seed reasoners set reasoning_options = [] while lab/peers expose real controls (e.g. OpenAI/OpenRouter GPT-5.6 Terra: effort none|low|medium|high|xhigh|max; o3: low|medium|high; DeepSeek V4: toggle + high|max; Claude Opus 4.7: effort low|medium|high|xhigh|max; Gemini 2.5: toggle + budget_tokens; Kimi K3 / Grok 4.5 / MiniMax-M3 also non-empty). Same empty list appears on anthropic/claude-haiku-4-5, claude-opus-4-7, claude-sonnet-4-6, deepseek/deepseek-v4-{flash,pro}, google/gemini-2.5-{flash,pro}, minimax/MiniMax-M3, moonshot/kimi-k3, openai/o3, xai/grok-4.5. Action: Replace [] with host-appropriate options copied from first-party + established OpenAI-compat peers (OpenRouter/Kilo/lab), add toggle wire comments when using toggle, and only keep [] when Eden docs/API affirm no control.
• [high] [violation] packages/core/src/sync/providers/edenai.ts:298 - Check: Automated sync must not invent empty reasoning_options for reasoners on a multi-model relay. Why: After the latest patch, new/factored reasoners omit reasoning_options; the shared runner then defaults them to [] via preserveReasoningOptions. The PR design also states capability/reasoning-control fields default to [] when Eden is silent. First full bun models:sync edenai (~800 models) will therefore mass-author empty options for controlled reasoners, which AGENTS forbids on relays. Action: When base_model resolves (or lab peers are known), author/copy that model’s control set instead of relying on the empty default; preserve hand-authored non-empty options; do not treat API silence as “no control.”

[YacineMK]

@rekram1-node Eden's api only reports supports_reasoning: true/false no per-model control detail so [] is the correct default per AGENTS.md Sync now preserves hand-authored options so individual seeds can be upgraded in follow-up PRs (same pattern as LLM Gateway)