PR #9 wired the entire dev-department to MiniMax-M2.7, including the six
sub-team leads. Per the model-tiering directive (leads → Claude Opus
subscription, ICs → MiniMax-M2.7), the leads must run `model: opus`
(the anthropic-oauth entry in config.yaml's runtime_config.models —
auth via CLAUDE_CODE_OAUTH_TOKEN, no base_url).
Changed: dev-lead, app-lead, core-lead, cp-lead, infra-lead, sdk-lead.
Unchanged: integration-tester, release-manager (ICs, stay on MiniMax);
fullstack-engineer + triage-operator already `model: opus`.
NB: molecule_runtime reads the MODEL_PROVIDER env var as the *picked
model id* (misnomer), at higher precedence than this `model:` field.
The deployed fix on the PC2 platform also sets per-workspace
workspace_secrets MODEL_PROVIDER=opus + CLAUDE_CODE_OAUTH_TOKEN and
blanks ANTHROPIC_AUTH_TOKEN/ANTHROPIC_BASE_URL (so the MiniMax-flavored
global_secrets don't leak in). The operator-host lead persona env files
(/etc/molecule-bootstrap/personas/<lead>/env) still carry the stale
MODEL_PROVIDER=claude-code (which made the claude CLI 404 on
`--model claude-code`) and need the same correction out of band — they
aren't tracked in any repo.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
