feat(concentrate): add Concentrate provider

[AjayK47]

Registers Concentrate as an OpenAI-compatible aggregator with an initial set of eight models. The remaining catalog follows in separate batches so each change stays small enough to review.

The sync joins the public model list with each model's detail endpoint and derives pricing, limits, modalities, tool calling, structured output and attachment support from that data. Capabilities and limits come from the aggregate list/detail payloads rather than a single upstream route, so a capability-thin route cannot narrow what the gateway accepts, and output limits are capped by canonical models.dev metadata.

Three rules encode the gateway's serving reality rather than its catalog:

• Pricing defaults to the cheapest advertised route, but IDs whose live traffic is relayed elsewhere are pinned to the observed serving route, so published cost matches what callers are charged.
• The gateway-level capability is authoritative for structured output. A route can advertise json_schema while the relay in front of it does not expose the feature.
• Context pricing tiers at or above a model's own window are dropped, because such a band can never be entered.

Reasoning controls are authored per model from the host's declared route metadata rather than inferred from the gateway's normalization enum, which silently accepts unsupported effort values. Every published effort value is backed by a route that declares it, and narrower or empty sets carry a leading comment with the API evidence. New remote models are reported instead of created automatically so each one gets that review first.

[github-actions]

Action items

• [medium] [possible mistake] providers/concentrate/models/claude-sonnet-5.toml:4 - Check: Relay reasoning_options must match the lab + same-surface peer baseline for this host. Why: anthropic/claude-sonnet-5 and established relays (e.g. OpenRouter, CrossModel) use toggle + effort low|medium|high|xhigh|max. This file ships effort only, with no leading wire-path comment explaining a host-specific narrower surface, so callers may miss an on/off control Concentrate actually exposes (its own schema models thinking.types.enabled). Action: Verify Concentrate’s Chat API for this ID; if thinking can be disabled separately from effort, add toggle plus a top-of-file wire comment; if not, keep effort-only and document the exact API evidence in a leading comment.
• [medium] [possible mistake] providers/concentrate/models/deepseek-v3-0324.toml:2 - Check: Provider overrides of attachment / [modalities] must be real host deltas. Why: Lab deepseek/deepseek-v3-0324 is text-only with attachment = false, and other hosts of this checkpoint do not advertise PDF. Claiming attachment = true and input = ["text", "pdf"] is a unique capability expansion without PR-mapped evidence. Action: Confirm Concentrate accepts PDF (or other non-text) input on deepseek-v3-0324; if not, drop attachment and the modalities override so the lab text-only surface is inherited.
• [low] [possible mistake] providers/concentrate/models/grok-4.5.toml:10 - Check: Published cost should match the billed cheapest/serving route. Why: Input/output (2/6) and the 200k tier (4/12) match xAI/OpenRouter, but cache_read is 0.5 (tier 1) versus first-party/peer 0.3 (tier 0.6). That split is suspicious if the same route set is used. Action: Verify Concentrate’s cache-read price for grok-4.5 (and the tier) from the detail/pricing payload and correct or cite the route that bills these cache rates.
• [low] [possible mistake] providers/concentrate/provider.toml:5 - Check: Material catalog data should cite direct provider pricing/docs/API evidence in the PR. Why: New provider costs, modality deltas, and per-model reasoning sets cannot be reviewed from the PR body alone (it describes sync policy but does not map sources to these eight models’ prices/controls). Action: Add PR citations (and/or leading TOML comments) for pricing and reasoning surfaces, stating which claim each source supports.

[github-actions]

Checking a few remaining peer cost values for consistency.One clear reasoning-options mismatch remains against lab and peer baselines.## Action items

• [medium] [possible mistake] providers/concentrate/models/grok-4.5.toml:858 - Check: Relay reasoning_options must match the lab/same-surface peer baseline for the model, with extra effort levels only when this host documents them or they show a live, meaningful effect. Why: First-party xAI and established relays (OpenRouter, Vercel, Kilo, CrossModel, etc.) publish Grok 4.5 as low / medium / high only; xhigh is the Grok 4.6 surface. This file adds xhigh with a route-metadata note but no live probe (unlike claude-sonnet-5), so clients may be offered a level the gateway only normalizes. Action: Confirm reasoning_effort=xhigh is honored with a distinct effect on Concentrate for this ID (and cite that), or drop xhigh to match lab/peers ["low", "medium", "high"].

[github-actions]

Action items

• [high] [violation] providers/concentrate/provider.toml:1 - Check: New provider.toml must include required doc. Why: Schema requires a non-empty doc URL for every provider; this file only sets name, env, npm, and api, so generation/bun validate will reject the provider. Action: Add doc pointing at Concentrate’s public model/API docs (e.g. models list or chat-completions reference).
• [medium] [violation] packages/core/src/sync/providers/concentrate.ts:360 - Check: Output limit must stay a finite positive number when capping against canonical metadata. Why: buildConcentrateModel always passes canonicalOutputLimit(baseModel), which returns undefined when the lab entry has no numeric limit.output. That disables the default parameter and makes Math.min(max_tokens, undefined) → NaN, so sync can write an invalid limit.output. Action: Only apply the canonical ceiling when it is a number (e.g. outputLimit === undefined ? model.summary.max_tokens : Math.min(...)), and cover the undefined path in the unit tests.

[AjayK47]

Both of these look like false positives:

• provider.toml already sets doc = "https://concentrate.ai/docs", and bun validate passes on this branch.
• The NaN case cannot happen — canonicalOutputLimit returning undefined triggers the default parameter, so Math.min gets max_tokens rather than undefined. Checked both paths: undefined ceiling gives limit.output = 64000, a 32000 ceiling caps to 32000.

No changes for those two. Earlier findings are addressed: claude-sonnet-5 now publishes the none off switch, grok-4.5 is back to low/medium/high, and source links were added per model.

[github-actions]

Checking the logo content more carefully and whether initial model files omit sync-derived limits that should be present.

[github-actions]

No actionable findings.

[github-actions]

Action items

• [critical] [violation] providers/concentrate/logo.svg:1 - Check: New provider logos must use currentColor with no hardcoded colors. Why: The second commit sets fill="#ffffff", which breaks the logo requirement and will not theme correctly (and can disappear on light backgrounds). Action: Restore fill="currentColor" (or equivalent path fills) with no fixed colors/size attributes; keep a square viewBox.

[github-actions]

No actionable findings.

[rekram1-node]

I dont see interleaved being set anywhere, normally it'd be reasoning_content

[AjayK47]

Good catch ! it's optional so nothing flagged it missing. Fixed for all 7, plus added a test so a future sync can't silently drop it again.

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/concentrate.ts:407 - Check: Sync must preserve hand-authored provider-only fields the API does not own (interleaved, and peers also keep status / provider / experimental). Why: buildConcentrateModel passes reasoning_options through existing but never interleaved. Seven reasoning models set [interleaved] field = "reasoning_content"; the next bun models:sync concentrate will drop those blocks because other aggregator modules (e.g. OpenRouter) explicitly re-emit interleaved: existing?.interleaved and Concentrate does not. Action: Include interleaved: existing?.interleaved (and the same pattern for any other hand-authored provider-only keys you intend to keep) in the factorBaseModel overrides, and cover it in packages/core/test/concentrate.test.ts.

[github-actions]

Action items

• [medium] [possible mistake] providers/concentrate/models/claude-opus-4-5.toml:18 - Check: interleaved must describe this host’s actual reasoning side channel (true, or field = reasoning_content / reasoning_details). Why: Commit 4 states Concentrate’s OpenAPI returns reasoning on reasoning, and that reasoning_content is only a compatibility alias that is not the field that carries content. Publishing field = "reasoning_content" on all seven reasoners therefore mislabels the wire key for catalog consumers that do not apply the AI SDK’s reasoning_content ?? reasoning fallback. Action: Switch those blocks to interleaved = true (general support when the key is not one of the two enum values), or keep a named field only after confirming the populated response key is actually reasoning_content / reasoning_details and documenting that in the leading comment.

[github-actions]

No actionable findings.