API proxy token usage reports 0 input and cache tokens for an Anthropic-compatible Z.AI endpoint #75271

This issue was posted by Claude Code using claude-opus-5-5 on behalf of David.

The AWF api-proxy's token usage table reports Input and Cache Read as 0 for every request to Z.AI's Anthropic-compatible endpoint. The Claude Code CLI's own result for the same session reports the real counts. Output tokens match.

Versions: gh-aw v0.86.2 (compiler and github/gh-aw-actions/setup), ghcr.io/github/gh-aw-firewall/api-proxy:0.27.44, Claude Code 2.1.227 in the threat-detection job.

Setup: threat detection uses engine: claude with ANTHROPIC_BASE_URL: https://api.z.ai/api/anthropic and model glm-5.3-flash. The AWF config routes it through apiProxy with targets.anthropic.host: api.z.ai.

Evidence: run https://github.com/pydantic/pydantic-ai/actions/runs/42064093118, job detection (127393142350).

Claude Code's result line (2026-11-05T16:58:10Z):

"modelUsage":{"glm-5.3-flash":{"inputTokens":38719,"outputTokens":4058,"cacheReadInputTokens":112640,"cacheCreationInputTokens":0}}

The api-proxy table parsed from /tmp/gh-aw/sandbox/firewall/logs/api-proxy-logs/token-usage.jsonl in the same job (2026-11-05T16:58:14Z):

# | Alias | Input | Output | Cache Read | Cache Write | Duration
1 | glm53 | 0 | 135   | 0 | 0 | 3.0s
2 | glm53 | 0 | 63    | 0 | 0 | 5.7s
3 | glm53 | 0 | 1,087 | 0 | 0 | 24.6s
4 | glm53 | 0 | 234   | 0 | 0 | 7.9s
5 | glm53 | 0 | 1,659 | 0 | 0 | 34.4s
6 | glm53 | 0 | 880   | 0 | 0 | 21.2s
Total |   | 0 | 4,058 | 0 | 0 | 1m37s

Output totals agree (4,058). Input is 0 against 38,719, and cache read is 0 against 112,640.

Expected: the table and token-usage.jsonl carry the input and cache-read tokens the provider returned.

Impact: token summaries and anything built on token-usage.jsonl undercount Z.AI runs. This run's summary shows 4,058 of 155,417 tokens.