Skip to content

0.55.11 — GLM via Alpha.sh; truncated reasoning no longer returned as the answer - #216

Closed
VickyXAI wants to merge 1 commit into
mainfrom
fix/glm-alpha-reasoning-servedby
Closed

VickyXAI wants to merge 1 commit into
mainfrom
fix/glm-alpha-reasoning-servedby

Conversation

@VickyXAI

Copy link
Copy Markdown
Contributor

0.55.11

Fixed — GLM goes through Alpha.sh, and a reply cut off mid-reasoning is no longer passed off as the answer

  • mode:"glm" now starts at zai/glm-5.3, the GLM that every gateway serves
    through Alpha.sh first (Z.ai direct is the rescue leg). The list used to be
    glm-5 / 5.2 / 5.1 / 5-Turbo, none of which Alpha carries, so mode:"glm"
    never reached Alpha. mode:"coding" uses glm-5.3 in place of glm-5.2. Same
    $1.4/$4.4 price, so no budget reserve moves.
  • When a reasoning model spends the whole max_tokens budget thinking and emits
    no content, blockrun_chat used to return that reasoning as the reply. A
    half-finished analysis read as the model's verdict (seen with zai/glm-5.3,
    twice), and the "spent the entire max_tokens budget on internal reasoning"
    diagnosis never fired. It now fires. A reasoning-only reply that finished
    normally still comes back as before.
  • Replies carry served_by, the upstream the gateway names in
    x-blockrun-served-by (e.g. alpha.sh), on the paths where response headers
    reach the client and the gateway sends it.

Tests: 1495/1495, typecheck clean. New tests pin the length-cut reasoning case, the finished reasoning-only case, served-by surfacing, and glm-5.3 as glm[0]. README/docs schema-token rows updated for the edited tool description.

Live check (Base gateway, patched code): a long review prompt to zai/glm-5.3 now reports the max_tokens diagnosis instead of returning reasoning. Prod logs show ALPHA_ENABLED=true on blockrun-web and blockrun-arc and no Alpha failover for these calls.

… the answer; surface served_by

- mode:"glm" starts at zai/glm-5.3, the one GLM every gateway serves through
  Alpha.sh first. The tier listed glm-5/5.2/5.1/5-turbo, none on Alpha, so
  mode:"glm" never reached it. coding swaps glm-5.2 for glm-5.3 (same price).
- ChatAccumulator used reasoning_content as the reply when no content came,
  even when the model was cut at max_tokens. A half-finished analysis came back
  as the verdict and rejectEmptyLength never fired. Reasoning is now a fallback
  only for a reply that finished normally.
- servedBy / served_by from the gateway's x-blockrun-served-by header.
@VickyXAI

Copy link
Copy Markdown
Contributor Author

Closing: wrong layer. GLM supply routing (Alpha vs Z.ai) is a gateway bug, fixed in the x402 server, not the MCP.

@VickyXAI VickyXAI closed this Nov 18, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant