You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[Phase E] Extend the onboarding pipeline beyond Claude to the Gemini and Copilot chains #2131
As a maintainer of a multi-provider engine,
I want new Gemini and Copilot models detected, gated and promoted by the same pipeline,
so that every provider chain in engine-models.sh stays current without hand-run rollouts.
Acceptance Criteria
Detection also polls Gemini's model list and diffs it against the gemini:* chains in scripts/lib/engine-models.sh, opening onboarding records the same way as for Claude.
Copilot models are detected from a source named in the policy file and diffed against the copilot:* chains. If no machine-readable source exists, the PR documents this and Copilot detection is disabled with that reason recorded, rather than guessed.
Pricing, preflight, quality gate, promotion and watch steps accept a provider argument. Provider-specific parts (pricing rows, liveness invocation, chain-variable names) are isolated and covered by bats for each provider.
Existing Claude behaviour is unchanged (the Claude pipeline tests still pass).
Chains today: gemini:triage|action = gemini-3.8-flash,gemini-3.1-pro-preview; gemini:deep|audit|single = gemini-3.1-pro-preview,gemini-3.8-flash; Copilot has a single model per key (see the duck/copilot rule ~line 160 of scripts/lib/engine-models.sh).
Activity