[hemaprathaban]

Adds RainDragon as an inference provider, with mappings for Gemma 4 26B and Gemma 4 31B.

Public sources for the values in this PR:

• API base URL and sign-up: https://inference.raindragon.ai
• Pricing and context length: https://inference.raindragon.ai/v1/models (public, no login needed). Gemma 4 26B $0.075 in / $0.35 out per M tokens, Gemma 4 31B $0.13 / $0.38, cached input $0.03/M.
• Privacy policy: https://inference.raindragon.ai/privacy
• Region: inference runs in the US (us-east).
• Rate limits: not published, so set to null as the schema allows.

Happy to change anything you need.

[madeburo]

Thanks for adding RainDragon! Registry validation passes for all 544 files, and the model IDs, standard input/output prices, and provider-specific context windows match the public /v1/models response.

There are two discrepancies to resolve before merging:

1. Cached-input pricing for Gemma 4 26B — mappings/raindragon/gemma-4-26b.yaml sets cache_read_per_million: 0.02, but the public pricing API currently reports cached_input_price_per_million: 0.03 for google/gemma-4-26B-A4B-it. Please align the mapping with the public USD-per-million rate, or provide an authoritative source explaining when the $0.02 rate applies. The 31B cached-input price already matches.

2. Prompt/response retention — providers/raindragon.yaml describes prompts and responses as deleted after 24 hours, while the published privacy policy, particularly sections 1 and 4, says prompt/completion content is processed in memory and is not retained after request processing. Please clarify the applicable policy and make the registry description and public documentation consistent. This is a documentation discrepancy; I am not treating the public policy as independent verification of backend retention behavior.

Unknown rate limits being null are fine. I have not tested authenticated inference, so this review does not certify end-to-end API behavior. Happy to re-review once the two metadata discrepancies are resolved.

[hemaprathaban]

Hi, thank you for the review and for pointing out both discrepancies. Both have now been addressed. The production API returns the correct cached-input price of $0.02 per million tokens for Gemma 4 26B, and the privacy policy and provider description have been updated to consistently state that prompts and responses are retained for up to 24 hours before deletion. The PR metadata now matches the live API. When you have a chance, could you please re-review the changes and, if everything looks good, proceed with merging the PR so RainDragon can be added to the registry? Thanks again

…

On Sat, Oct 24, 2026 at 10:17 PM Umid ***@***.***> wrote: ***@***.**** requested changes on this pull request. Thanks for adding RainDragon! Registry validation passes for all 544 files, and the model IDs, standard input/output prices, and provider-specific context windows match the public /v1/models response. There are two discrepancies to resolve before merging: 1. *Cached-input pricing for Gemma 4 26B* — mappings/raindragon/gemma-4-26b.yaml sets cache_read_per_million: 0.02, but the public pricing API <https://inference.raindragon.ai/api/models> currently reports cached_input_price_per_million: 0.03 for google/gemma-4-26B-A4B-it. Please align the mapping with the public USD-per-million rate, or provide an authoritative source explaining when the $0.02 rate applies. The 31B cached-input price already matches. 2. *Prompt/response retention* — providers/raindragon.yaml describes prompts and responses as deleted after 24 hours, while the published privacy policy <https://inference.raindragon.ai/privacy>, particularly sections 1 and 4, says prompt/completion content is processed in memory and is not retained after request processing. Please clarify the applicable policy and make the registry description and public documentation consistent. This is a documentation discrepancy; I am not treating the public policy as independent verification of backend retention behavior. Unknown rate limits being null are fine. I have not tested authenticated inference, so this review does not certify end-to-end API behavior. Happy to re-review once the two metadata discrepancies are resolved. — Reply to this email directly, view it on GitHub <#8?email_source=notifications&email_token=ABE3HYWG56FNTXKMLYBTMLT5VTMKNA5CNFSNUABKM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UKJSXM2LFO4XTKNRTHEYDCOBUG4Y2M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#pullrequestreview-5639018471>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABE3HYUXSSQJ4DEFEQA4OJD5VTMKNAVCNFSNUABGKJSXA33TNF2G64TZHMYTEMZRG44DSOBUGU5US43TOVSTWNJXGQYDOMBUGIYDFILWAI> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/ABE3HYUJAGOAF3DYQYYENJL5VTMKNA5CNFSNUABKM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UKJSXM2LFO4XTKNRTHEYDCOBUG4Y2M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KUZTPN52GK4S7NFXXG> and Android <https://github.com/notifications/mobile/android/ABE3HYSRU4VLMQRFWKUWUYD5VTMKNA5CNFSNUABKM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UKJSXM2LFO4XTKNRTHEYDCOBUG4Y2M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2K4ZTPN52GK4S7MFXGI4TPNFSA>. Download it today! You are receiving this because you authored the thread.Message ID: ***@***.***>

[madeburo]

Thanks for the fixes and the follow-up. Both concerns from our previous review are now resolved:

• The public pricing API reports $0.02 per million cached-input tokens for Gemma 4 26B, matching the mapping; the 31B rate remains $0.03. The model IDs, standard input/output prices, and provider-specific context windows also match the public APIs.
• The published privacy policy now consistently states that prompt/completion content is retained for up to 24 hours and then deleted, matching the provider description.

Registry validation passed locally, including a trial merge with the current main branch (579 files), and GitHub's Validate Registry check has passed for this PR. I also fixed a maintainer-side CI issue in this branch: fork PRs now publish the validation report in the Actions summary instead of failing while trying to write a PR comment with a read-only token. The validator and its mandatory failure gate remain in place.

Approving this registry addition. Thank you for resolving the discrepancies and keeping the public documentation and registry metadata aligned!