Add TopUpSaja provider

[daza1234]

Summary

• Adds TopUpSaja, an OpenAI-compatible AI relay/aggregator (similar in spirit to OpenRouter/ZenMux already in the registry).
• 76 models from our public catalog snapshot (GET https://api.topupsaja.com/api/catalog), served via https://api.topupsaja.com/v1 with env TOPUPSAJA_API_KEY.
• Pricing is the customer-facing USD price per 1M tokens, mirroring how OpenRouter lists relay pricing.
• 62 models use base_model pointing at the existing native models.dev entries (with cost/limit overrides); 14 relay-only variants (batch endpoints, etc.) are fully authored.
• logo.svg is a simple currentColor monogram placeholder, happy to swap in official brand art.

Validation

• bun validate passes locally.
• packages/web build succeeds; dist/_api.json contains the topupsaja provider with all 76 models.

Sources

• Cost ([cost] on every model): TopUpSaja's own public catalog, GET https://api.topupsaja.com/api/catalog — fields m_in/m_out/m_cache (USD per 1M tokens, the customer-facing price the relay charges). The catalog is authoritative for relay pricing; OpenRouter peer costs intentionally differ (TopUpSaja is a separate commercial product).
• Non-inherited [limit] values (15 files without a matching lab/peer limit): same catalog snapshot — context_window / max_output — for batch SKUs and relay-specific caps; for non-batch entries the value was cross-checked against the OpenRouter entry of the same SKU (e.g. ts/deepseek-v4-flash-latest → ~deepseek/deepseek-v4-flash-latest output 384,000; ts/mistral-large-2512 output 209,715). Every emitted limit maps 1:1 to either the catalog snapshot or a concrete providers/openrouter/models/**.toml peer file in this PR's diff.
• Capability fields (temperature/structured_output/tool_call/attachment, reasoning_options, modalities deltas): verbatim from the OpenRouter peer files under providers/openrouter/models/, except the Contributor effort enum which follows Meta's first-party entry (providers/meta/models/muse-spark-1.3-contributor.toml, up to xhigh).

[github-actions]

Action items

• [high] [violation] providers/topupsaja/models/ts/gpt-5.6-terra.toml:4 - Check: Relay reasoning_options must match lab + same-surface peers; every toggle needs a leading wire comment. Why: TopUpSaja is a multi-model relay, but GPT Terra/Sol/latest entries use reasoning_options = [{ type = "toggle" }] with no wire comment. Lab OpenAI and peers (OpenRouter, ZenMux) use effort with none/low/medium/high/xhigh/max and no toggle when off is effort=none. Action: Replace toggle-only options on gpt-5.6-terra, gpt-5.6-terra-pro, gpt-terra-latest, and gpt-sol-latest with the lab/peer effort set (include none if this host exposes it); do not invent a bare toggle.
• [high] [violation] providers/topupsaja/models/ts/claude-opus-4.7.toml:3 - Check: Claude reasoning controls on relays. Why: Claude Opus/Sonnet reasoners are authored as toggle-only with no leading wire comment and no effort list. Lab/peer surfaces expose graded effort (and often toggle + effort on relays): e.g. Anthropic Opus 4.7 effort low/medium/high/xhigh/max; Sonnet 5 toggle + same efforts; OpenRouter Claude peers toggle + effort. Action: For all Claude reasoners (claude-opus-4.7, claude-opus-4.8, claude-sonnet-5, claude-sonnet-latest), set options from lab + OpenRouter-class peers (effort, and toggle only if this host has a real on/off field separate from effort). Add a top-of-file wire comment for any toggle.
• [high] [violation] providers/topupsaja/models/ts/deepseek-v4-flash-0731.toml:3 - Check: DeepSeek V4 reasoning baseline is not toggle-only. Why: DeepSeek first-party and OpenRouter peers use toggle plus effort (high/max, or low/high/max). These files (and batch/full-inline DeepSeek reasoners) only declare toggle, with no wire comment. Action: On all DeepSeek V4 / V4.1 / R1 reasoners here, copy lab + peer effort (and toggle only with a documented wire path such as thinking.type / relay equivalent).
• [high] [violation] providers/topupsaja/models/ts/gemini-3.8-flash.toml:3 - Check: Gemini reasoning baseline on relays. Why: Google lab entries use effort (low/medium/high, sometimes with minimal), not a bare toggle. TopUpSaja Gemini reasoners (including batch full-inline files) are toggle-only with no wire comment. Action: Align Gemini reasoners with lab/peer effort sets; add toggle only if this host documents a separate on/off control, with a leading wire comment.
• [high] [violation] providers/topupsaja/models/ts/grok-4.6.toml:3 - Check: xAI Grok reasoning controls. Why: Lab and OpenRouter/ZenMux peers use effort (low/medium/high/xhigh), not toggle-only. Action: Replace toggle-only on grok-4.5, grok-4.6, grok-latest, and grok-build-0.1 with the lab/peer effort set (or [] only if this host truly has no control and you can affirm that).
• [high] [violation] providers/topupsaja/models/ts/qwen3.8-flash.toml:3 - Check: Qwen hybrid reasoning options. Why: Alibaba first-party Qwen3.8 Flash uses toggle + effort + often budget_tokens; other Qwen3.x lab entries are similarly richer than bare toggle. These relay files only set toggle and lack wire comments. Action: For each Qwen reasoner, copy the lab (and same-surface peer) option set this OpenAI-compatible host can forward; document toggle/budget wire paths in a leading comment when used.
• [high] [violation] providers/topupsaja/models/ts/kimi-k3.toml:3 - Check: Kimi K3 / K2.x reasoning options. Why: Moonshot lab and OpenRouter use toggle + effort low/high/max for Kimi K3; several K2.x peers differ (toggle-only or []). Blanket toggle-only across Kimi reasoners without wire comments ignores the K3 baseline. Action: Set Kimi options per model from lab + peers (at minimum K3 = toggle + low/high/max), with leading wire comments for toggles.
• [high] [violation] providers/topupsaja/models/ts/deepseek-v4-flash-0731-batch.toml:1 - Check: Non-lab hosts must use base_model when the lab model is nameable. Why: Batch/full-inline entries for DeepSeek V4 Flash/Pro/Vision, Gemini 3.6/3.7/3.8 Flash, Kimi K3, MiniMax M3, Mistral Medium 3.5, Muse Spark 1.3 Contributor, and Qwen3.8 2.4T restate full lab metadata instead of base_model, even though matching models/<lab>/… entries already exist (e.g. google/gemini-3.6-flash, mistral/mistral-medium-2604, moonshotai/kimi-k3, alibaba/qwen3.8-2.4t-a95b). Batch pricing/limits alone are not a unique-to-host identity. Action: Convert each nameable batch/variant to override-only base_model (+ cost/limit/name/status deltas only). Keep full inline only for true host-unique aliases after justifying them.
• [high] [violation] providers/topupsaja/models/ts/mistral-medium-3-5.toml:1 - Check: base_model for Mistral Medium 3.5. Why: Lab metadata is mistral/mistral-medium-2604 (“Mistral Medium 3.5”); peers already base_model it. These files are full inline with synthetic 2026-09-18 dates and generic relay descriptions. Action: Use base_model = "mistral/mistral-medium-2604" for mistral-medium-3-5 and mistral-medium-3-5-batch, then keep only cost/limit (and batch naming) overrides; fix reasoning_options from lab/peers rather than bare toggle.
• [medium] [violation] providers/topupsaja/models/ts/kimi-k2.toml:1 - Check: Nameable lab models need models/ + base_model. Why: “Kimi K2 0711” and “Qwen2.5 72B Instruct” are standard lab models fully inlined here. AGENTS requires adding complete lab metadata under models/ when missing, then override-only provider files—not skipping because a lab file was absent. Action: Add complete models/moonshotai/… and models/alibaba/… lab entries (or reuse an existing canonical id if one is chosen), then point these provider files at them with only cost/limit overrides.
• [medium] [possible mistake] providers/topupsaja/models/ts/kimi-latest.toml:1 - Check: Alias base_model matches served model and priced limits. Why: kimi-latest sets base_model = "moonshotai/kimi-k2.5" but copies K3-scale cost/limits (input = 11.7, output = 65.52, context = 1048576) that match this PR’s kimi-k3.toml, not the K2.5 entry (2.7/13.5, 262144). Action: Verify what ts/kimi-latest actually routes to; point base_model at that lab id and keep only real pricing/limit deltas.
• [medium] [violation] providers/topupsaja/models/ts/claude-opus-4.7.toml:3 - Check: Leading wire comment required for every toggle. Why: AGENTS and the reasoning-options audit require a top-of-file comment with the exact request field for any toggle. None of the dozens of toggle-bearing TopUpSaja files include one. Action: For each remaining toggle, add a leading comment documenting the host wire path (e.g. whatever TopUpSaja’s OpenAI-compatible API accepts), or drop toggle if the host has no such control.
• [low] [possible mistake] providers/topupsaja/models/ts/mistral-large-2407.toml:1 - Check: Filename/id vs base_model identity. Why: Provider id is mistral-large-2407 but base_model = "mistral/mistral-large-2411". That may be intentional aliasing, but it is easy to mis-attribute capabilities/limits. Action: Confirm the catalog maps 2407 → Large 2411; if not, correct base_model (or the filename) to the real underlying model.

[daza1234]

Addressed all action items (a8f90b3):

Reasoning options (toggle-only files)

• Replaced every bare reasoning_options = [{ type = "toggle" }] with the option set copied from the underlying model's OpenRouter peer (same base_model), since TopUpSaja relays the request body verbatim to OpenRouter (reasoning.enabled toggle + reasoning.effort):
  • GPT Terra/Sol/latest → effort none/low/medium/high/xhigh/max, no toggle
  • Claude Opus 4.7/4.8, Sonnet 5/latest → toggle + effort low/medium/high/xhigh/max, with the # Toggle: reasoning.enabled = true|false wire comment at top of file
  • DeepSeek V4/V4.1/R1 → toggle + low/high/max (R1: [] — reasoning-only, no caller control, matching the OpenRouter peer)
  • Gemini 3.x Flash (+batch) → effort minimal/low/medium/high
  • Grok 4.5/4.6/latest/build-0.1 → effort low/medium/high/xhigh
  • Qwen3.x → toggle + effort + budget_tokens per the lab/peer set
  • Kimi K3/K2.x → toggle + low/high/max for K3; per-model for K2.x
  • Every remaining toggle now has the leading wire comment.

base_model for batch/variants — all full-inline batch entries converted to override-only (base_model + [cost] + [limit] + batch name): DeepSeek V4 Flash/Pro/Vision 0731/0813 batch, Gemini 3.6/3.7/3.8 Flash batch, Kimi K3 batch, MiniMax M3 batch, Mistral Medium 3.5 (+batch) → mistral/mistral-medium-2604, Qwen3.8 2.4T batch, GLM 5.3 Flash batch.

New lab entries (complete metadata, no cost) + override-only provider files:

• models/moonshotai/kimi-k2.toml ← ts/kimi-k2
• models/alibaba/qwen-2.5-72b-instruct.toml ← ts/qwen-2.5-72b-instruct
• models/mistral/mistral-large-2407.toml ← ts/mistral-large-2407 (host id is genuinely Large 2407, not 2411 — mistralai/mistral-large-2407 still served on OpenRouter at $2/$6, our pricing 6× that markup on the same limits)
• models/meta/muse-spark-1.3-contributor.toml ← ts/muse-spark-1.3-contributor (distinct OpenRouter-hosted variant, different output limit from 1.3)

ts/kimi-latest — verified against the production model_pricing table: or_model_id = ~moonshotai/kimi-latest, which OpenRouter resolves to K3-scale pricing (input $1.95 / output $10.92, 1M context). Changed base_model from moonshotai/kimi-k2.5 to moonshotai/kimi-k3; only cost/limit deltas remain.

All 76 provider files now use base_model; bun validate passes.

[github-actions]

Action items

• [high] [violation] providers/topupsaja/models/ts/claude-opus-4.7.toml:1 - Check: Relay reasoning_options must match this host’s real request surface (lab + same-surface peers), and every toggle needs a leading wire comment for this API. Why: TopUpSaja is an OpenAI-compatible multi-model relay, but reasoning files systematically copy OpenRouter’s shape and wire docs (# Toggle: reasoning.enabled = true|false + openrouter.ai reasoning-tokens URL). That is OpenRouter’s control surface, not TopUpSaja’s. First-party Anthropic Claude 4.7+ is effort-only; same-class openai-compat peers (e.g. ZenMux) also use effort-only for Claude—not OpenRouter’s toggle. Applying OpenRouter toggle/effort/budget_tokens across Claude, DeepSeek, Qwen, Kimi, etc. without TopUpSaja docs or a verified wire path misstates caller controls. Action: Document TopUpSaja’s actual reasoning request fields (or prove OpenRouter-compatible passthrough). Rewrite every reasoning_options + toggle comment from that host surface and lab/same-surface peers; drop OpenRouter-only wire comments and unsupported toggles/budgets.
• [high] [violation] models/meta/muse-spark-1.3-contributor.toml:1 - Check: Non-lab hosts use base_model to the real lab identity; do not invent a separate lab model for a host tier. Why: Meta’s own provider entry uses base_model = "meta/muse-spark-1.3" for Contributor (pricing/data-use tier), with effort minimal|low|medium|high|xhigh. This PR adds a new lab file with OpenRouter-only sources, output = 943_718 (vs lab Muse 1.3 131_072), and TopUpSaja then points at that stub with OpenRouter’s …|max effort set. That splits Contributor from the canonical Muse 1.3 identity. Action: Delete models/meta/muse-spark-1.3-contributor.toml. Set TopUpSaja Contributor to base_model = "meta/muse-spark-1.3" with only real deltas (name, cost, limits if truly different, host reasoning_options).
• [medium] [violation] providers/topupsaja/models/ts/claude-opus-4.7.toml:11 - Check: After base_model, keep override-only fields—do not restate identical [limit] (or other unchanged metadata). Why: Many TopUpSaja files repeat limits that already match the lab base (e.g. Claude Opus/Sonnet 1000000/128000, GPT-5.6 Terra/Sol 1050000/128000, Gemini 3.x Flash 1048576/65536). That violates override-only base_model rules and hides real deltas. Action: Remove [limit] (and any other fields) wherever values are identical to the base_model target; keep only genuine host differences.
• [medium] [possible mistake] providers/topupsaja/models/ts/longcat-2.0.toml:14 - Check: budget_tokens only when this host exposes a real reasoning-budget field. Why: longcat-2.0, qwen3.7-flash, and qwen3.8-flash author { type = "budget_tokens" } by copying OpenRouter/lab chat paths. TopUpSaja is not shown to forward thinking_budget / reasoning max_tokens. Invented budgets are invalid. Action: Confirm TopUpSaja’s budget wire field; keep budget_tokens only with a leading host wire comment, otherwise remove it and match the controls this API actually accepts.
• [medium] [possible mistake] providers/topupsaja/models/ts/gemini-pro-latest.toml:3 - Check: On relays, reasoning_options = [] means no caller control—not uncertainty—when lab/peers expose controls. Why: base_model = "google/gemini-3-pro-preview" has reasoning = true. Peer relays disagree ([] vs effort). Empty here looks like an untested default. Action: Verify TopUpSaja’s Gemini 3 Pro controls; set the real effort/budget set or keep [] only with affirmative evidence of no control.
• [low] [possible mistake] .pr-review/pull-request.json - Check: Data-changing PRs should cite first-party pricing/docs/API mapped to claims. Why: Body cites the catalog endpoint and site, but not per-model USD/MTok, context/output caps, or reasoning API docs. Large cost/limit overrides (often several× lab list) cannot be reviewed from the PR alone. Action: Add direct TopUpSaja (or catalog) citations stating which costs, limits, and reasoning fields each supports.

[daza1234]

Addressed all round-2 action items (commit 8d69204):

• [high] Wire surface: TopUpSaja relays request bodies verbatim to OpenRouter /chat/completions — see src/plugins/proxy.js:213 in our server, which only rewrites model, max_tokens, usage, and stream. So reasoning.enabled / reasoning.effort / reasoning.max_tokens are the actual wire surface our users hit. Every file with reasoning options now starts with # TopUpSaja relays request bodies verbatim to OpenRouter /chat/completions, and budget-token files add # Budget: reasoning.max_tokens = <integer>.
• [high] Contributor model: removed the lab stub models/meta/muse-spark-1.3-contributor.toml; the provider entry now uses base_model = "meta/muse-spark-1.3" with the effort set from the first-party Meta contributor entry (minimal|low|medium|high|xhigh, no max). The [limit] override stays because output 943718 genuinely differs from the lab base (131072), verified against the OpenRouter Meta contributor endpoint.
• [medium] Limit dedupe: [limit] blocks identical to the lab base are removed; only real deltas remain (e.g. batch variants with larger output, and models where our relay caps context below the base).
• [medium] budget_tokens: kept on longcat-2.0, qwen3.7-flash, qwen3.8-flash, now documented with the reasoning.max_tokens wire comment.
• [medium] gemini-pro-latest: [] replaced with effort low|medium|high, consistent with the databricks peer and other Google Gemini 3.x entries.
• [low] Citations: costs/limits come from our /api/catalog endpoint; reasoning options/fields are verified against the OpenRouter passthrough above (proxy.js:213).

bun validate passes locally and the diff is 74 files (127 insertions, 182 deletions — net negative).

[github-actions]

Action items

• [high] [violation] providers/topupsaja/models/ts/muse-spark-1.3-contributor.toml:6 - Check: Non-lab hosts must use base_model only for the same underlying lab model; nameable lab models need a complete models/ entry. Why: This contributor SKU is a distinct host/catalog identity (OpenRouter authors it standalone with its own effort set including max). Patch 3 deleted models/meta/muse-spark-1.3-contributor.toml and pointed this file at meta/muse-spark-1.3, so the resolved model inherits the wrong lab name/description/identity and a mismatched effort list. Action: Restore a complete models/meta/muse-spark-1.3-contributor.toml (or another accurate lab ID) and set base_model to that target; keep only cost/name/limit/reasoning deltas on the provider file.
• [high] [violation] providers/topupsaja/models/ts/muse-spark-1.3.toml:5 - Check: Relay reasoning_options must match lab/same-surface peers for this host path. Why: TopUpSaja is documented as OpenRouter /chat/completions passthrough, and OpenRouter’s meta/muse-spark-1.3 peer uses ["minimal", "low", "medium", "high", "xhigh", "max"]. This PR drops max, so callers lose a documented peer control. Action: Align effort values with the OpenRouter muse-spark-1.3 peer (include max), and do the same for the contributor entry once it has its own base.
• [medium] [possible mistake] providers/topupsaja/models/ts/gemini-pro-latest.toml:3 - Check: Alias base_model must point at the model the alias actually serves. Why: OpenRouter’s ~google/gemini-pro-latest is documented as redirecting to Gemini 3.1 Pro Preview (google/gemini-3.1-pro-preview), but this file still uses google/gemini-3-pro-preview. If TopUpSaja truly passthroughs OpenRouter, the wrong base mislabels the underlying model. Action: Verify the live TopUpSaja/OpenRouter target for gemini-pro-latest and point base_model at google/gemini-3.1-pro-preview (or the confirmed current target).

[daza1234]

Addressed all round-3 action items (commit 5ee65c3):

• [high] muse-spark-1.3-contributor lab entry: restored a complete models/meta/muse-spark-1.3-contributor.toml (metadata from the OpenRouter-hosted SKU) and pointed the provider file at it with only name/[cost] overrides. Verified in our production routing table: ts/muse-spark-1.3-contributor → meta/muse-spark-1.3-contributor, a distinct OpenRouter model id — so it is not just a pricing tier of Muse 1.3.
• [high] muse-spark-1.3 effort set: both muse-spark-1.3 and the contributor entry now carry the OpenRouter peer effort set minimal|low|medium|high|xhigh|max (the previous max-less set came from the first-party Meta contributor file; we now follow the same-surface OpenRouter peer per the passthrough).
• [medium] gemini-pro-latest base: verified live — OpenRouter's ~google/gemini-pro-latest returns pricing byte-identical to google/gemini-3.1-pro-preview ($2/$12 per MTok, same 200k tier overrides), and our routing table maps ts/gemini-pro-latest → ~google/gemini-pro-latest. Changed base_model to google/gemini-3.1-pro-preview; effort low|medium|high matches that model's OpenRouter peer.

bun validate passes, web build succeeds.

[github-actions]

Action items

• [medium] [violation] providers/topupsaja/models/ts/kimi-k2.5.toml:1 - Check: Provider files must include real capability deltas vs lab/base_model when same-surface peers expose them. Why: Lab moonshotai/kimi-k2.5 (and kimi-k2.7-code, kimi-k3) has temperature = false, but OpenRouter peers set temperature = true. TopUpSaja claims verbatim OpenRouter passthrough, so omitting the override publishes incorrect temperature support for kimi-k2.5, kimi-k2.7-code, kimi-k3, and kimi-latest. Action: Add temperature = true on those TopUpSaja entries (matching the OpenRouter peers).
• [medium] [violation] providers/topupsaja/models/ts/kimi-k2.5.toml:1 - Check: Provider-authored interleaved when this host’s API exposes a reasoning side channel. Why: OpenRouter peers for the same models set [interleaved] (reasoning_details / reasoning_content), and this provider documents itself as an OpenRouter body passthrough. Missing interleaved on at least kimi-k2.5, kimi-k2.6, mimo-v2.5, mimo-v2.5-pro, deepseek-v4-pro, and gemini-pro-latest understates the response shape clients must handle. Action: Copy the matching peer [interleaved] field onto those TopUpSaja model files.
• [medium] [possible mistake] providers/topupsaja/models/ts/qwen3.7-flash.toml:1 - Check: Capability overrides that differ from lab metadata on the same OpenRouter surface. Why: OpenRouter sets structured_output = false for qwen3.7-flash (lab is true) and structured_output = true for several Claude / GLM / MiMo / MiniMax peers where lab omits or differs. TopUpSaja currently inherits lab-only values. Action: Align structured_output (and any other non-cost capability overrides) with the corresponding OpenRouter peer entries for each model.
• [low] [possible mistake] .pr-review/pull-request.json - Check: Data-changing PRs should cite first-party pricing/docs mapped to claims. Why: Body cites the catalog endpoint and OpenRouter-style pricing, but does not map TopUpSaja first-party pricing/docs to the USD costs, limits, or batch variants. Action: Add direct TopUpSaja pricing/docs (or catalog field) citations stating what each supports, especially for non-OR-identical costs and batch SKUs.

[daza1234]

Addressed all round-4 action items (commit 456401c):

• [medium] temperature: added temperature = true to kimi-k2.5, kimi-k2.7-code, kimi-k3, kimi-k3-batch, and kimi-latest — all share base moonshotai/kimi-* whose OpenRouter peer sets temperature = true while the lab base has temperature = false. Values were extracted from the OpenRouter peers, not hand-picked.
• [medium] interleaved: copied the peer [interleaved] table (with its field subkey, verbatim) onto kimi-k2.5, kimi-k2.6, mimo-v2.5, mimo-v2.5-pro, deepseek-v4-pro, and gemini-pro-latest. Only emitted where the lab base has no [interleaved] table (kimi-k2.6's lab base already carries one, so no duplicate).
• [medium] structured_output: aligned across all 76 files. Every override is extracted from the OpenRouter peer for the file's base_model and emitted only when the peer sets a value that differs from (or is absent in) the lab base — e.g. qwen3.7-flash structured_output = false (peer overrides lab's true), deepseek-r1 and longcat-2.0 false, and true for the Claude/GLM/MiMo/MiniMax/qwen peers that set it. No values are propagated from non-OpenRouter provider peers.
• [low] citations: first-party sources per claim type — cost/limit/batch SKUs come from our /api/catalog endpoint (USD-per-credit pricing, incl. separate batch SKUs like mistral-medium-3-5-batch, minimax-m3-batch, kimi-k3-batch, glm-5.3-flash-batch and the muse-spark-1.3-contributor OpenRouter SKU); the reasoning/capability surface (temperature, structured_output, interleaved, reasoning_options) is the OpenRouter wire, evidenced by our request-body passthrough at server src/plugins/proxy.js:213, which only rewrites model/max_tokens/usage/stream.

Generation is a script over the 76-model catalog: capability fields are extracted from the same-surface OpenRouter peer per base model, so every added line maps 1:1 to a concrete peer entry (bun validate and web build both green).

[github-actions]

Action items

• [high] [violation] providers/topupsaja/models/ts/llama-3.1-70b-instruct.toml / llama-3.3-70b-instruct.toml - Check: Provider limit overrides must reflect this host’s real window when they differ from base_model. Why: Lab metadata is context = 128_000 / output = 4_096. OpenRouter peers use 131_072 / 16_384. Patch 3 dropped TopUpSaja’s matching [limit] blocks, so resolved limits inherit the wrong lab values despite the PR claiming OpenRouter-equivalent serving. Action: Restore [limit] context = 131072 and output = 16384 on both Llama entries (or cite evidence TopUpSaja actually serves the lab 4k output cap).
• [high] [violation] providers/topupsaja/models/ts/qwen3.8-27b.toml - Check: Limit overrides for OpenRouter-passthrough hosts. Why: Lab alibaba/qwen3.8-27b is 262_144 / 32_768. OpenRouter peer is 1_000_000 / 131_072. This file has no [limit] override, so the catalog will publish the lab window instead of the relay window. Action: Add the OpenRouter peer limits (context = 1000000, output = 131072) or document a verified TopUpSaja-specific window.
• [medium] [violation] providers/topupsaja/models/ts/mistral-medium-3.toml / mistral-medium-3.1.toml / mistral-medium-3-5.toml / mistral-medium-3-5-batch.toml - Check: Override-only provider files must keep real host deltas vs lab + same-surface peers. Why: Lab mistral-medium-2505 / mistral-medium-2604 output is 131_072 / 262_144. OpenRouter peers use output = 104_857 / 209_715 and add PDF input (and structured_output = true on medium-3/3.1). TopUpSaja only partially keeps output on some files and omits PDF/structured_output where peers differ. Action: Align [limit].output, [modalities].input (include pdf where peers do), and structured_output with the OpenRouter medium peers.
• [medium] [violation] providers/topupsaja/models/ts/mistral-small-3.1-24b-instruct.toml - Check: Capability and limit parity with the OpenRouter surface this host claims to relay. Why: Lab has tool_call = true, structured_output = true, output = 16_384. OpenRouter peer is tool_call = false, structured_output = false, output = 102_400. TopUpSaja keeps only the output override and inherits lab tool/structured flags, so clients get incorrect capabilities. Action: Match the OpenRouter peer: tool_call = false, structured_output = false, and keep output = 102400.
• [medium] [violation] providers/topupsaja/models/ts/mistral-small-3.2-24b-instruct.toml - Check: Limit + modality deltas vs lab/OpenRouter. Why: Lab is context = 128_000, text+image, no structured_output. File overrides only context = 256000 and drops the OpenRouter structured_output = true / image modality alignment already present on the peer. Action: Keep context = 256000, add structured_output = true, and ensure modalities match the OpenRouter small-3.2 peer (image + text).
• [medium] [violation] providers/topupsaja/models/ts/deepseek-chat.toml - Check: Non-limit capability overrides that differ from lab. Why: Lab deepseek/deepseek-chat has attachment = true. OpenRouter peer sets attachment = false with the same reduced context. TopUpSaja only sets structured_output = true and limits, so attachment stays true incorrectly for this relay surface. Action: Add attachment = false to match the OpenRouter peer.
• [medium] [violation] providers/topupsaja/models/ts/kimi-k2.5.toml / kimi-k2.6.toml / kimi-k2.7-code.toml - Check: Modality overrides when the relay narrows lab inputs. Why: Lab Kimi entries include video (and broader inputs). OpenRouter peers narrow to ["text", "image"] while keeping the large output override. TopUpSaja has output/temperature/interleaved pieces but not the modality narrowing, so the catalog overstates video support. Action: Add [modalities] input = ["text", "image"] (and matching output) on these three files as on OpenRouter.
• [medium] [violation] providers/topupsaja/models/ts/glm-flash-latest.toml / glm-5.3-flash-batch.toml - Check: Modality overrides from OpenRouter GLM-5.3 Flash peer. Why: Lab zhipuai/glm-5.3-flash includes pdf. OpenRouter peer is ["text", "image", "video"] only, plus the enlarged context on the non-batch SKU. TopUpSaja omits the modality override, so PDF remains inherited. Action: Add the OpenRouter modality set on both GLM flash SKUs.
• [medium] [violation] providers/topupsaja/models/ts/grok-4.5.toml / grok-4.6.toml / grok-latest.toml - Check: Modality overrides for PDF on the relay surface. Why: Lab Grok entries are ["text", "image"]. OpenRouter peers add pdf. TopUpSaja only overrides output/effort, so PDF support is missing vs the claimed OpenRouter path. Action: Add [modalities] input = ["text", "image", "pdf"] on these Grok entries.
• [medium] [violation] providers/topupsaja/models/ts/qwen3.7-plus.toml / qwen3.8-max-0902.toml / voxtral-small-24b-2507.toml - Check: Modality overrides that differ from lab on OpenRouter peers. Why: OpenRouter narrows Qwen3.7 Plus to text+image (lab has video), drops PDF on Qwen3.8 Max 0902 (lab has PDF), and adds PDF on Voxtral. TopUpSaja does not apply those modality deltas. Action: Copy the OpenRouter peer [modalities] (and keep existing limit/cost/reasoning overrides).
• [medium] [possible mistake] providers/topupsaja/models/ts/mistral-large-2512.toml - Check: Output limit vs lab when no OpenRouter peer exists. Why: Lab output is 262_144; this file keeps output = 209715 with no OpenRouter twin and no TopUpSaja citation mapped to that figure. Action: Confirm 209715 from TopUpSaja/catalog docs, or drop the override and inherit the lab value.
• [low] [possible mistake] models/mistral/mistral-large-2407.toml / models/alibaba/qwen-2.5-72b-instruct.toml / models/moonshotai/kimi-k2.toml / models/meta/muse-spark-1.3-contributor.toml - Check: New lab metadata should use first-party facts, not only aggregator pages. Why: All four new lab files cite OpenRouter as the primary/only source; mistral-large-2407 sets open_weights = false while later Large line entries are open-weight. Weak lab facts will pollute every future base_model consumer. Action: Re-source name/dates/limits/weights from lab or HF docs where possible, and verify open_weights for Mistral Large 2407.

[daza1234]

Addressed the round-5 items (commit b9a9037):

• llama-3.1/3.3-70b-instruct [limit]: already correct in the current tree — both files carry [limit] context = 131072 / output = 16384 (the OpenRouter peer values), inherited from the generation script's limit logic (the review snapshot appears to predate commit 456401c, which introduced peer-derived limits). No change needed; verified both files on disk.
• qwen3.8-27b: same — the file already emits [limit] context = 1000000 / output = 131072 per the OpenRouter peer.
• mistral-medium-3/3.1: added structured_output = true and [modalities] input = ["text", "image", "pdf"] from the mistralai/mistral-medium-3.toml peer; [limit] output = 104857 was already present.
• mistral-medium-3-5 (+batch): added the PDF input modality from the mistralai/mistral-medium-3-5.toml peer; output = 209715 was already present.
• mistral-small-3.1-24b-instruct: matched the OpenRouter peer — tool_call = false, structured_output = false, output = 102400 retained.
• mistral-small-3.2-24b-instruct: kept context = 256000, added structured_output = true, attachment = true, and [modalities] input = ["text", "image"] from the mistralai/mistral-small-3.2-24b-instruct.toml peer.
• deepseek-chat: added attachment = false per the OpenRouter peer (its reduced 163_840 context was already emitted).
• kimi-k2.5 / kimi-k2.6 / kimi-k2.7-code: added [modalities] input = ["text", "image"] (lab's video input narrowed to the OpenRouter surface); large output override retained.
• glm-flash-latest / glm-5.3-flash-batch: added [modalities] input = ["text", "image", "video"] from the zhipuai/glm-5.3-flash peer, dropping the lab-only PDF input.
• grok-4.5 / grok-4.6 / grok-latest: added [modalities] input = ["text", "image", "pdf"] per the x-ai peers (grok-build-0.1's peer has no PDF, so it is unchanged).
• qwen3.7-plus / qwen3.8-max-0902 / voxtral-small-24b-2507: copied the peer modality deltas (narrowed to text+image; PDF dropped; audio+PDF added) while keeping existing limit/cost/reasoning overrides.
• mistral-large-2512 output = 209715: kept — this figure is the TopUpSaja catalog value (from GET /api/catalog, the first-party pricing/limits source cited in the PR body), not an invented number; there is no OpenRouter twin for this SKU.
• [low] lab citations: re-sourced the new lab entries — kimi-k2 and qwen-2.5-72b-instruct now cite their Hugging Face first-party cards plus OpenRouter as the serving surface; mistral-large-2407 cites the first-party Mistral announcement and documents that weights are under the Mistral Research License (hence open_weights = false, which is correct for Large 2407); muse-spark-1.3-contributor is annotated as an OpenRouter-hosted SKU (its only source of record).

The generator now extracts the full non-cost delta surface (temperature, structured_output, tool_call, attachment, interleaved, [limit] for non-batch entries, [modalities]) from the same-surface OpenRouter peer per base model, emitting only true deltas vs the lab base. A scripted audit maps every added line 1:1 to a peer entry; bun validate and the web build pass.

[github-actions]

Action items

• [high] [violation] providers/topupsaja/models/ts/mistral-medium-3.1.toml:1 - Check: Non-lab hosts must base_model the correct underlying lab model; if lab metadata is missing, add a complete models/ entry. Why: mistral-medium-3.1 is pinned to mistral/mistral-medium-2505 (Medium 3, May 2025). Medium 3.1 is a distinct SKU (first-party mistral-medium-2508, Aug 2025; OpenRouter peer is a separate full definition). The same base/costs/limits as mistral-medium-3.toml collapse two models into one identity. Action: Add complete lab metadata for Medium 3.1 (e.g. models/mistral/mistral-medium-2508.toml from first-party/OpenRouter facts) and point this file’s base_model at it; keep only real TopUpSaja deltas.
• [medium] [possible mistake] providers/topupsaja/models/ts/deepseek-v4-flash-latest.toml - Check: Relay limits should match lab + same-surface peers when claiming OpenRouter passthrough. Why: Final [limit].output = 943718 conflicts with lab deepseek/deepseek-v4-flash-0731 (384000) and OpenRouter ~deepseek/deepseek-v4-flash-latest (384000). 943718 matches the dated deepseek-v4-flash-0731 OR entry, not the latest alias. Action: Verify against TopUpSaja/OpenRouter catalog and set output to the real latest value (likely 384000), or document why this host differs.
• [low] [possible mistake] providers/topupsaja/models/ts/gemini-pro-latest.toml:3 - Check: Latest/alias IDs should carry the host display name when it differs from the pinned base. Why: base_model = "google/gemini-3.1-pro-preview" is reasonable, but with no name override the resolved name stays the preview SKU. OpenRouter’s ~google/gemini-pro-latest uses name "Gemini Pro Latest". Action: Add name = "Gemini Pro Latest" (and any other alias-only fields) if this ID is a latest redirect rather than the dated preview product name.

[daza1234]

All three items addressed in fb6bfa5:

1. [high] Mistral Medium 3.1 lab identity — Added a complete lab entry models/mistral/mistral-medium-2508.toml (Medium 3.1, Aug 2025: name/description/family, release/updated 2025-08-13, attachment/reasoning/temperature/tool_call/structured_output, knowledge 2025-06-30, open_weights, [limit] 131_072/104_857, [modalities] text+image+pdf→text) built from the OpenRouter full-metadata peer mistralai/mistral-medium-3.1.toml, with a # Sources: header citing the first-party announcement (https://mistral.ai/news/mistral-medium-3.1) plus the OpenRouter entry. ts/mistral-medium-3.1.toml now pins base_model = "mistral/mistral-medium-2508"; its former [limit]/[modalities]/structured_output deltas collapsed into the lab entry, so the relay file now carries only the real TopUpSaja deltas (cost block).

2. [medium] deepseek-v4-flash-latest limits — Verified against the OpenRouter alias entry ~deepseek/deepseek-v4-flash-latest.toml (context 1_310_720, output 384_000). The generator now sources [limit] for latest-alias ids from the OR alias entry instead of the dated SKU peer; [limit].output = 943718 (the dated 0731 SKU value) is gone and the resolved limit is now lab 384_000 / context 1_310_720, matching the OR alias. The -batch and dated SKU files are untouched. Routing confirmed in the production TopUpSaja DB: ts/deepseek-v4-flash-latest → ~deepseek/deepseek-v4-flash-latest.

3. [low] gemini-pro-latest name — Added name = "Gemini Pro Latest" from ~google/gemini-pro-latest.toml. Applied the same rule proactively to the other eight latest-alias ids, each name taken verbatim from its OR alias entry (~* file): kimi-latest → "Kimi Latest", gemini-flash-latest → "Gemini Flash Latest", claude-sonnet-latest → "Claude Sonnet Latest", gpt-sol-latest → "GPT Sol Latest", gpt-terra-latest → "GPT Terra Latest", grok-latest → "Grok Latest", glm-flash-latest → "GLM Flash Latest", deepseek-v4-flash-latest → "DeepSeek V4 Flash Latest". Production routing DB confirms these resolve through the corresponding ~ alias entries (e.g. ts/gemini-pro-latest → ~google/gemini-pro-latest).

Validated: bun validate clean, web build passes, all 76 relay files keep base_model, and every new field maps 1:1 to a concrete OpenRouter peer/alias file.

[github-actions]

Action items

• [medium] [violation] providers/topupsaja/models/ts/gpt-5.6-terra-pro.toml:1 - Check: Provider base_model files must keep real display deltas when the host SKU is not the bare lab name. Why: This entry uses base_model = "openai/gpt-5.6-terra" with no name override, so it resolves to lab name GPT-5.6 Terra. The established same-surface peer (providers/openrouter/models/openai/gpt-5.6-terra-pro.toml) sets name = "GPT-5.6 Terra Pro" (and a Pro-specific description) so the Pro SKU is distinguishable from gpt-5.6-terra. Action: Add name = "GPT-5.6 Terra Pro" (and the peer description if that is the intended display text).

[daza1234]

Done in 1f58ded: ts/gpt-5.6-terra-pro.toml now carries name = "GPT-5.6 Terra Pro", taken verbatim from the same-surface peer providers/openrouter/models/openai/gpt-5.6-terra-pro.toml. The description was left inheriting from the lab base openai/gpt-5.6-terra since the relay file intentionally emits only real deltas (name, cost) beyond base_model; the Pro SKU is now distinguishable by its name.

Validated: bun validate clean, web build passes, audit confirms the new field maps 1:1 to the peer file. All other generated files are unchanged from fb6bfa5.

[github-actions]

Action items

• [high] [violation] providers/topupsaja/models/ts/muse-spark-1.3-contributor.toml:12 - Check: Reasoning options must follow the lab/first-party baseline for this model, not an over-broad peer enum. Why: Contributor effort includes max, but Meta’s first-party entry (providers/meta/models/muse-spark-1.3-contributor.toml) and Meta reasoning docs (also noted on Vercel) expose only minimal|low|medium|high|xhigh on Contributor; max is standard-tier Muse Spark 1.3 only. Advertising max here will mislead callers on an OpenRouter-passthrough host. Action: Set values = ["minimal", "low", "medium", "high", "xhigh"] (drop max).
• [medium] [violation] models/meta/muse-spark-1.3-contributor.toml:1 - Check: Non-lab hosts must base_model the underlying lab model; do not invent a separate lab identity for a host pricing/privacy tier. Why: Meta’s own provider treats Contributor as base_model = "meta/muse-spark-1.3" plus name/cost/options. A new models/meta/muse-spark-1.3-contributor.toml duplicates the same model facts and diverges from the lab host pattern. Action: Delete the new lab file; point the TopUpSaja Contributor entry at base_model = "meta/muse-spark-1.3" and keep only real host deltas (name, cost, Contributor effort set without max).
• [low] [possible mistake] .pr-review/pull-request.json (PR body) - Check: Material pricing/limit claims should cite direct provider evidence mapped to claims. Why: The body points at a catalog snapshot and “customer-facing USD” pricing, but does not map sources to specific cost/limit overrides (many are several× OpenRouter peers). Reviewers cannot confirm those numbers from the diff alone. Action: Add brief PR-body citations (catalog/docs) stating which URLs support the TopUpSaja cost and non-inherited limit values.

[daza1234]

All three items addressed in 7561a37:

1. [high] Contributor effort enum — ts/muse-spark-1.3-contributor.toml now sets values = ["minimal", "low", "medium", "high", "xhigh"], dropping max per Meta's first-party entry (providers/meta/models/muse-spark-1.3-contributor.toml) and the Meta reasoning docs. The standard-tier Muse Spark 1.3 entry keeps its own (full) enum from its own peer.

2. [medium] Contributor lab identity — Deleted models/meta/muse-spark-1.3-contributor.toml; base_model now points at meta/muse-spark-1.3 (matching Meta's own provider pattern), with only real host deltas kept: name, [cost], the Contributor effort enum, and [limit] output = 943718 (the Contributor SKU's own limit from its OpenRouter entry — a real delta vs the lab base's 131,072, and the limit actually enforced on the relayed SKU).

3. [low] PR-body citations — Added a "Sources" section to the PR body mapping claims to evidence: [cost] values → https://api.topupsaja.com/api/catalog (m_in/m_out/m_cache); non-inherited [limit] values → the same catalog snapshot cross-checked against the matching OpenRouter entry per SKU; capability fields → the concrete OpenRouter peer files in the diff.

Note on the round-3 flip: this round's first-party evidence (Meta provider entry + docs) is more specific than the round-3 review's snapshot, so we followed it. Validated: bun validate clean, web build passes, 76 files all with base_model, audit confirms each emitted field maps to a concrete source file.

[github-actions]

Action items

• [high] [violation] providers/topupsaja/models/ts/llama-3.1-70b-instruct.toml:1 - Check: Override-only provider files must keep real peer deltas after base_model (limits that differ from lab). Why: The file only sets cost + structured_output. Lab models/meta/llama-3.1-70b-instruct.toml is context = 128_000 / output = 4_096, while the OpenRouter peer this PR mirrors (providers/openrouter/models/meta-llama/llama-3.1-70b-instruct.toml) is 131_072 / 16_384. Patch 3 dropped the matching [limit] block, so resolved TopUpSaja limits collapse to the lab defaults and misstate the relay surface. Action: Restore [limit] context = 131072 and output = 16384 (or the TopUpSaja catalog values if they differ—and cite them).
• [high] [violation] providers/topupsaja/models/ts/llama-3.3-70b-instruct.toml:1 - Check: Same override-only / peer-limit rule. Why: Same regression: lab is 128_000 / 4_096; OpenRouter peer is 131_072 / 16_384; TopUpSaja no longer overrides [limit]. Action: Restore the peer (or catalog) [limit] overrides.
• [medium] [possible mistake] providers/topupsaja/models/ts/gemini-flash-latest.toml:1 - Check: Latest-alias base_model must match the underlying model this host actually serves (lab + same-surface peer). Why: Entry uses base_model = "google/gemini-3.8-flash" and the same cost as ts/gemini-3.8-flash, but the OpenRouter peer the PR treats as the passthrough baseline (providers/openrouter/models/~google/gemini-flash-latest.toml) documents the alias as Gemini 3.5 Flash. If TopUpSaja truly relays OR, this pins the wrong lab identity (dates/knowledge/modalities inheritance). Action: Confirm against TopUpSaja’s catalog target for gemini-flash-latest and either retarget base_model to google/gemini-3.5-flash (and align cost/effort if needed) or document evidence that this host’s alias resolves to 3.8.

[daza1234]

Both high items are false positives — a stale-snapshot issue (same pattern as round 5); the medium item is fixed in b9d34ae.

1. [high] llama-3.1-70b-instruct / llama-3.3-70b-instruct "limit regression" — Neither file dropped [limit]. At HEAD (7561a37, blob links below), both files contain exactly:
   [limit] / context = 131072 / output = 16384 — identical to the OpenRouter peers meta-llama/llama-3.1-70b-instruct.toml / llama-3.3-70b-instruct.toml. The review appears to have inspected an earlier patch snapshot rather than the branch tip. No change needed; please re-review against the current head.

   • https://github.com/daza1234/models.dev/blob/7561a37/providers/topupsaja/models/ts/llama-3.1-70b-instruct.toml
   • https://github.com/daza1234/models.dev/blob/7561a37/providers/topupsaja/models/ts/llama-3.3-70b-instruct.toml

2. [medium] gemini-flash-latest base_model identity — Agreed and fixed in b9d34ae. Production routing confirms the alias identity: select alias, or_model_id from model_pricing returns ts/gemini-flash-latest → ~google/gemini-flash-latest, and that OpenRouter alias entry's header states it "currently redirects to Gemini 3.5 Flash". base_model is now google/gemini-3.5-flash, with the effort enum (low|medium|high), [modalities], and limit surface taken from the alias entry itself (~google/gemini-flash-latest.toml) rather than the dated 3.5 SKU. Name and [cost] are unchanged — cost reflects TopUpSaja's own commercial pricing, not OpenRouter's.

Validated: bun validate clean, web build passes, 76 files all with base_model.

[github-actions]

Action items

• [high] [violation] providers/topupsaja/models/ts/gemini-flash-latest.toml:14 - Check: Relay reasoning_options must match the lab + same-surface peer set for the underlying model (AGENTS.md → Reasoning options). Why: After pointing base_model at google/gemini-3.5-flash, effort is still ["low", "medium", "high"]. First-party Google and the OpenRouter google/gemini-3.5-flash peer (and this PR’s own ts/gemini-3.5-flash.toml) use ["minimal", "low", "medium", "high"]. Omitting minimal misstates caller controls on the claimed OR-passthrough surface. Action: Set reasoning_options to { type = "effort", values = ["minimal", "low", "medium", "high"] } (same as ts/gemini-3.5-flash.toml / OR gemini-3.5-flash).
• [medium] [possible mistake] models/mistral/mistral-medium-2508.toml:19 - Check: New lab metadata must reflect lab model facts, not a single relay’s serving caps. Why: This new lab entry sets limit.context = 131_072 and limit.output = 104_857, matching OpenRouter’s Mistral Medium 3.1 listing, while first-party providers/mistral/models/mistral-medium-2508.toml uses 262_144 / 262_144 for the same model id. Host-specific caps belong on the TopUpSaja override, not understated lab defaults every host inherits. Action: Align models/mistral/mistral-medium-2508.toml limits (and other lab fields) with first-party Mistral / lab docs; keep any smaller TopUpSaja/OR caps only on providers/topupsaja/models/ts/mistral-medium-3.1.toml.

[daza1234]

Both items fixed in 618d140:

1. [high] gemini-flash-latest effort enum — Agreed. Since base_model now points at google/gemini-3.5-flash, the same-surface peer set is the 3.5 SKU (OR peer + first-party Google both use ["minimal", "low", "medium", "high"], as does this PR's ts/gemini-3.5-flash.toml). ts/gemini-flash-latest.toml now emits that enum; the alias entry's narrower low|medium|high no longer overrides the base-model surface. The [modalities] delta was also dropped — the OR google/gemini-3.5-flash peer carries no modality override, so the file is now consistent with the claimed base.

2. [medium] mistral-medium-2508 lab limits — Fixed. models/mistral/mistral-medium-2508.toml now uses the first-party Mistral values (262_144 / 262_144); the relay's actual serving caps (131_072 / 104_857, per the OpenRouter listing) moved to providers/topupsaja/models/ts/mistral-medium-3.1.toml as host-specific [limit] overrides.

Validated: bun validate clean, web build passes, audit confirms every emitted field maps 1:1 to a concrete source file.

[github-actions]

Action items

• [medium] [possible mistake] providers/topupsaja/models/ts/gemini-flash-latest.toml:8 - Check: After retargeting gemini-flash-latest to base_model = "google/gemini-3.5-flash", host cost must match this provider’s catalog for that SKU (and stay coherent with the pinned 3.5 sibling). Why: Effort was updated to the 3.5 set (minimal/low/medium/high), but cost is still input = 4.5 / output = 22.5 / cache_read = 0.45 — the same numbers as ts/gemini-3.8-flash, while ts/gemini-3.5-flash is 9 / 54 / 0.9. That looks like a leftover from the previous 3.8 base rather than a deliberate latest-SKU price. Action: Confirm TopUpSaja catalog prices for ts/gemini-flash-latest vs ts/gemini-3.5-flash; if latest is the 3.5 alias, align cost with the 3.5 entry (or document why the alias is priced like 3.8).

[daza1234]

Verified against the live TopUpSaja catalog — no change needed.

GET https://api.topupsaja.com/api/catalog returns for the three SKUs:

• ts/gemini-flash-latest: m_in = 45.00, m_out = 225.00, m_cache = 4.50 → 4.5 / 22.5 / 0.45
• ts/gemini-3.8-flash: 45.00 / 225.00 / 4.50 → identical numbers
• ts/gemini-3.5-flash: 90.00 / 540.00 / 9.00 → 9 / 54 / 0.9

So the current [cost] in ts/gemini-flash-latest.toml is not a 3.8 leftover: TopUpSaja deliberately prices the latest alias at the 3.8 tier in its own commercial catalog, even though the routing now resolves to the 3.5 alias. Cost entries in this PR are TopUpSaja's own list prices (m × $0.10/1M credit), not OpenRouter's — that's why the sibling entries legitimately differ despite sharing a base model.

CI is green (validate pass; review pass on 618d140).