llm-gateway(glm): pin reasoning effort to high (var-overridable)

[aaronjmars]

What

scripts/llm-gateway.sh glm arm now pins Claude Code reasoning effort to high:

export CLAUDE_CODE_EFFORT_LEVEL="${GLM_REASONING_EFFORT:-high}"
export CLAUDE_CODE_ALWAYS_ENABLE_EFFORT=1

Why

GLM-5.3 / glm-5.3-flash are forced-thinking models (thinking cannot be disabled), and depth is the reasoning_effort dial: the 5.3 series supports low / high / max only. Without a param, the Anthropic-compat endpoint leaves depth model-decided (docs claim default max, empirically it lands between high and max). This pins a deterministic fleet default.

Verified honored on the live fleet endpoint https://api.z.ai/api/anthropic (2026-08-31), fixed hard prompt, glm-5.3:

effort	thinking chars	output tokens
low	2,356	2,045
high	15,251	7,959
max	32,717	15,645
no param	21,034	10,840

CLAUDE_CODE_ALWAYS_ENABLE_EFFORT=1 is required because Claude Code only sends the effort field for Claude model ids it recognizes; glm-* ids would silently drop it otherwise.

Override

Per-instance repo var GLM_REASONING_EFFORT (low / high / max) overrides the default. No new secrets; no new required vars.

Verification

• bash -n clean; source-test of the glm arm with GATEWAY=glm exports CLAUDE_CODE_EFFORT_LEVEL=high (and =max with the var set).
• Endpoint honoring verified directly with curl against the fleet key (table above).