Summary

• PR #2861 swapped crm-updater / crm-updater-haiku / strategist / next-steps from Kimi K3 to GLM 5.3 Flash for a staging-only comparison (all four -oss call sites).

• cedar-automation-bench PR #19's owner-review comparison bundle (13 CRM cases, 29 Strategist cases, scored across GLM/Sonnet-incumbent/Kimi) surfaced a clear, real-evidence signal that GLM underperforms Kimi on Strategist specifically:

  Model	Strategist scored pass rate	Unscored (environment_gap)
  Kimi K3	17/18 (94%)	11/29 runs
  Sonnet (incumbent)	16/25 (64%)	4/29 runs
  GLM 5.3 Flash	7/12 (58%)	17/29 runs

  GLM was both more likely to fail on the cases that could be scored, and far more likely to hit environment gaps in the first place.

• CRM updater told the opposite story in the same bundle , GLM went 15/15, beating both the Sonnet incumbent (10/15) and Kimi (11/15) , so crm-updater-oss, crm-updater-haiku-oss, and next-steps-oss are intentionally left on GLM. This PR touches only strategist-updates.ts.

Change

• strategist-updates.ts: GLM_MODEL → KIMI_MODEL at the strategist-oss call site, restoring the "do NOT set reasoningEffort: 'none'" comment (Kimi silently breaks tool-calling with that option set , see KIMI_MODEL's doc comment in models.ts).
• No other call site changed. No workload rename or PostHog flag change , strategist-oss remains model-agnostic infrastructure.

Test plan

• pnpm --filter @zero/server run types , clean
• pnpm deps:check , no violations
• Targeted vitest run across the 16 test files covering this module , 216/216 passed
• Staging traffic spot-check once deployed , confirm strategist-oss captures show kimi-k3 again (flag cedar-oss-system-subagents already active for isabelle/jesse, so this takes effect immediately on deploy)

🤖 Generated with Claude Code

https://claude.ai/code/session_012WwyQEcVnYQzkB4Fxsfpx6

Confidence Score: 5/5

The PR appears safe to merge because the model switch uses an already-supported routing configuration and preserves Kimi-compatible provider options.

Summary

This PR restores Kimi K3 for the OSS Strategist workload while leaving its workload identifier, feature-flag behavior, and non-OSS Sonnet route unchanged.

• Replaces GLM_MODEL with KIMI_MODEL at the strategist-oss model-resolution call site.
• Documents why Kimi is preferred for this workload.
• Preserves provider options that avoid Kimi’s known tool-calling incompatibility.

_{Reviews (1) · Last reviewed commit: "Revert strategist-oss from GLM 5.3 Flash..."}