When the CRM updater uses the OSS/Understudy route, conversation-field updates run on GLM 5.3 Flash instead of Kimi K3, including the shared crm-updater-haiku branch.Not checked
prod
When next-steps task updates use the OSS/Understudy route, the next-steps agent runs on GLM 5.3 Flash instead of Kimi K3 and can complete its real tool-call loop.Not checked
prod
When the strategist uses the OSS/Understudy route, strategist updates run on GLM 5.3 Flash instead of Kimi K3 while retaining the existing multi-step tool workflow.Not checked
prod
Libra has verdicts on 0 of 3 tracked behaviors on prod; 3 are still being checked. Libra checks hourly for 3 days after each deploy.
Summary
Adds GLM_MODEL = 'glm-5.3-flash' to services/llm/models.ts (routed via Understudy exactly like SOL_MODEL/KIMI_MODEL, family: 'openai', never direct).
Swaps the three -oss call sites currently on Kimi K3 over to GLM: crm-updater (which crm-updater-haiku shares a model branch with), strategist, next-steps.
Staging-only comparison , no workload rename or PostHog flag change needed. -oss workload names are model-agnostic, and cedar-oss-system-subagents is already active for <email> / <email>, so this takes effect immediately on deploy to staging.
Why GLM, and why it's safe here
Same rigor applied after the earlier Sol reasoning_effort incident (PR #2850): confirmed with standalone scripts against the real Understudy gateway, not just a typecheck.
glm-5.3-flash produces real tool_calls (not text-embedded pseudo-XML) both with and without an explicit reasoning_effort: 'none'.
Completes a clean multi-step tool-call → tool-result → final-answer loop.
Unlike Kimi K3 (which silently breaks tool-calling if reasoningEffort: 'none' is set) GLM has no known sensitivity to that provider option either way, so the existing THINKING_DISABLED_PROVIDER_OPTIONS at each call site is left unchanged.
Test plan
pnpm --filter @zero/server run types , clean
pnpm deps:check , no violations
Targeted vitest run across the 16 test files covering these three modules , 216/216 passed
Standalone script confirmation of GLM tool-calling behavior against the live Understudy gateway (both with and without reasoning_effort: 'none')
Staging traffic spot-check once deployed (flag is already on for isabelle/jesse)
The PR appears safe to merge as a staging-only model substitution with unchanged routing and workload contracts.
Summary
This PR introduces the Understudy-routed glm-5.3-flash model and switches three staging OSS agent paths from Kimi K3 to GLM while retaining their existing feature flags, workloads, provider options, and tool restrictions.
Adds and re-exports the shared GLM_MODEL identifier.
Updates CRM field updater, strategist, and next-steps OSS model selection.
Leaves direct Anthropic routes and workload-based observability unchanged.
