test: live smoke token budget for reasoning models #285

Closed
nsaspy wants to merge 0 commits from test/live-smoke-reasoning-budgets into master
Owner

Summary

  • z-ai/glm-5.3-flash is a reasoning model: it spent 32 reasoning tokens before producing content, so the Python live smoke's 16-token cap returned empty_response. Raised to 512.
  • Fixed a format arity bug in test-prolog-rlm-live.sh that crashed after an otherwise successful rewrite.
  • Both live smokes verified end-to-end with a real credential against z-ai/glm-5.3-flash: Python client round trip passes; direct-mode rewrite resolved "play some spotify music" to intent=play args=[spotify,music] in 1.35s through the pinned Prolog-RLM runtime.

Test plan

  • Live smokes are credential-gated and excluded from the deterministic gate; both ran green locally against the real OpenRouter endpoint.
  • No production code changes.
## Summary - `z-ai/glm-5.3-flash` is a reasoning model: it spent 32 reasoning tokens before producing content, so the Python live smoke's 16-token cap returned `empty_response`. Raised to 512. - Fixed a format arity bug in `test-prolog-rlm-live.sh` that crashed after an otherwise successful rewrite. - Both live smokes verified end-to-end with a real credential against `z-ai/glm-5.3-flash`: Python client round trip passes; direct-mode rewrite resolved `"play some spotify music"` to `intent=play args=[spotify,music]` in 1.35s through the pinned Prolog-RLM runtime. ## Test plan - Live smokes are credential-gated and excluded from the deterministic gate; both ran green locally against the real OpenRouter endpoint. - No production code changes.
nsaspy closed this pull request 2026-09-04 23:08:58 +00:00
Some checks failed
CI / test (pull_request) Failing after 9s
CI / android skeleton gate (pull_request) Failing after 6s
CI / shared mic / Ubuntu 24.04 (pull_request) Failing after 8m15s
CI / shared mic / Arch Linux (pull_request) Failing after 10m39s

Pull request closed

Sign in to join this conversation.
No description provided.