Workers AI, Clef and agent proposals
Editable Mermaid source · All architecture diagrams.
AI is optional. Ordinary source and collaboration do not require an inference call.
Clef: advisory PR evidence
The evidence Workflow can call @cf/cloudflare/clef through an authenticated AI Gateway for metadata-only advice. This profile does not send the source tree, execute the suggested change, grant an approval or merge code. Retained hosted advice reached SUCCEEDED on October 14.
Proposals: scoped code and tests
A separate Containers application and PRAgentTask Durable Object run the maintained agent harness against an independent Artifacts fork. Private R2 holds context and results. The reviewed engineering deployment used @cf/zai-org/glm-5.3-flash; its recorded proposal added a function and passed tests on October 14.
The approved profile can edit only calc.py in the playground repositories; it does not apply to imported repositories. Current source and sponsor authority, call/token budgets, allowed network routes, expiry and cleanup bound the job. Proposal completion leaves canonical Git unchanged.
Gateway policy
The retained gateway requires authentication, disables prompt/response logging and caching, and is configured for 60 requests/minute and a 500-cent daily spend limit. Configuration is verified; exercise of rate/spend limits is separate evidence. It is not a provider-wide billing cap.
Model entitlement, provider availability and a bound AI service are distinct from successful inference. A failure stays a failure; no silent mock or unapproved model fallback is offered. See coding-agent accounts for the independent portable harness integrations.
Vectorize search / in development
Vectorize and Workers AI embeddings are being added for semantic search over code and issues, filtered by the reader’s current repository access. The reviewed demo has no Vectorize binding or working semantic-search acceptance yet. Index provenance, deletion/revocation propagation and real indexing/query evidence remain required.