Repository navigation
Conversation
|
👋 Some commits in this PR are not signed and verified by GitHub. Please sign your commits with a GPG or SSH key registered in your GitHub account, then force-push. Commits that are not verified: See GitHub's guide on signing commits for setup instructions. I've added |
|
👋 Some commits in this PR are not signed and verified by GitHub. Please sign your commits with a GPG or SSH key registered in your GitHub account, then force-push. Commits that are not verified: See GitHub’s guide on signing commits for setup instructions. I’ve added |
5626ca2 to
8be334a
Compare
|
Thanks @aheritier. Re-signed the commit with an SSH signing key registered on my account and force-pushed: 8be334a now shows as Verified. No code changes. |
aheritier
left a comment
There was a problem hiding this comment.
No blocking implementation defects found. One documentation correction recommended; see inline comment.
Reviewed all 15 changed files and surrounding provider/configuration paths. The implementation is consistent with existing providers.
CI is green for 8be334a: gate, lint, Linux/Windows tests, cross-compilation, image, documentation, and CodeQL passed. Race testing was skipped. No local tests or authenticated inference were run; the live Runcrate test was skipped in CI.
GitHub reports BLOCKED / REVIEW_REQUIRED. No linked issue or Jira acceptance criteria were found, so this review covers the PR's stated scope.
| [Runcrate](https://www.runcrate.ai/?utm_source=docker-agent&utm_medium=directory&utm_campaign=gpu_pricing) | ||
| is a GPU cloud that also serves open models (DeepSeek, GLM, Qwen, gpt-oss and | ||
| others) through an OpenAI-compatible API. Docker Agent includes built-in | ||
| support for Runcrate as an alias provider. Open models are free with your |
There was a problem hiding this comment.
[should-fix] Qualify the claim that open models are free. Confidence: high.
Runcrate's published terms at https://www.runcrate.ai/inference/free say the free allowance requires a GitHub account at least 30 days old. After the allowance is exhausted, requests can consume workspace credit; paid fallback can also occur when free capacity is busy. Following this setup with a funded workspace therefore does not guarantee free inference.
Describe this as a limited free allowance for eligible GitHub accounts, link the current allowance/capacity terms, and mention that subsequent requests can consume workspace credit.
There was a problem hiding this comment.
Thanks @aheritier, fixed in 738c39a (signed): the overview now describes a limited free allowance for GitHub accounts at least 30 days old, says later requests or busy free capacity can use workspace credit, and links the current terms at runcrate.ai/inference/free. I also rebased on main to clear a conflict in pkg/config/auto.go (kept main's Baseten default), so the workflows need approval again.
|
👋 This PR has merge conflicts with the base branch. Please rebase or merge the latest base branch and resolve them. I've moved it to draft and added �[ backtick]status/needs-rebase�[ backtick]; it'll be picked back up automatically once the conflicts are cleared. |
Signed-off-by: Nirupam Bhowmick <48842933+heydryft@users.noreply.github.com>
The free allowance is limited, needs a GitHub account at least 30 days old, and later requests can use workspace credit. Link the current terms.
f302bcb to
738c39a
Compare
|
Thanks @aheritier. Rebased onto the latest main and resolved the conflicts; the branch is now up to date with main (0 commits behind) and both commits (5cacc46, 738c39a) are signed and Verified. No change to the substance of the Runcrate provider addition. Could you take another look when you have a moment? Happy to adjust anything. |
Summary
runcrate, base URLhttps://api.runcrate.ai/v1), in the same shape as the Atlas Cloud aliasRUNCRATE_API_KEYfor automatic model selection and credential forwarding, withdeepseek-ai/DeepSeek-V4.1-Flashas the default modelRuncrate is a GPU cloud that also serves open models (DeepSeek, GLM, Qwen, gpt-oss) through an OpenAI-compatible API. It is not in models.dev yet, so it gets the same exemptions as
atlascloudandapi-routeinTestProviderIDsMatchEmbeddedCatalogandmodelsDevAbsentProviders, and the docs page says that no catalog metadata is imported.Testing
go test ./pkg/model/provider/... ./pkg/config/...: pass (new rows inTestCatalogAliases, the auto-config tables, the system-message merge tests andopenAIAliasProviders)RUNCRATE_API_KEY=... go test ./pkg/model/provider/ -run 'TestOpenAIAliasProvider_LiveAPI/runcrate': PASS, live response"pong"docker-agent run --execwithmodel: runcrate/deepseek-ai/DeepSeek-V4.1-Flash: the agent answeredgofmt -l: cleanThis PR was prepared with AI assistance (an AI coding agent wrote the change and ran the tests and live calls above).
Disclosure: I work on Runcrate.