glm-5.2[1m] fails with upstream 400 "Unknown Model" (Z.AI code 1211) — bracketed 1M suffix forwarded verbatim to openai-chat endpoint

[DomineYH]

Summary

Routing Codex to Z.AI's GLM-5.2 1M-context variant with the model ID
glm-5.2[1m] fails on every request with an upstream 400 error
(Z.AI code 1211, "Unknown Model"). The bare glm-5.2 code is accepted by
the provider; only the bracketed [1m] variant is rejected.

glm-5.2[1m] is Z.AI's documented, official model ID for the 1M-context
variant, so this looks like a model-string passthrough issue in opencodex
rather than an invalid model on my side.

Environment

• opencodex version: [e.g. output of ocx --version]
• Codex CLI version: [e.g. codex --version]
• OS: [macOS / Linux / Windows + arch]
• Provider / adapter: Z.AI GLM Coding Plan via openai-chat adapter
• baseUrl: [the Z.AI endpoint configured in config.json]
• Auth mode: key

Steps to reproduce

1. Configure a Z.AI (GLM) provider in ~/.opencodex/config.json using the
   openai-chat adapter.
2. Start the proxy: ocx start
3. Run Codex against the 1M variant:
   codex -m "<provider>/glm-5.2[1m]" "hello"

Expected

Request routes to GLM-5.2 with the 1M context window. Z.AI documents
glm-5.2[1m] as the model name that enables the 1M-context variant.

Actual

The request fails with:

{"error":{"message":"Provider error 400: {"error":{"code":"1211","message":"Unknown Model, please check the model code.\n"}}","type":"upstream_error","code":null}}

For comparison, glm-5.2 (without the [1m] suffix) routes successfully
through the same provider config.

Likely cause

1211 / "Unknown Model, please check the model code." is Z.AI's response
when the model field does not match a known model code. It appears opencodex
forwards the model string glm-5.2[1m] verbatim as the model value to the
upstream OpenAI-compatible chat-completions endpoint, which does not accept
the bracketed [1m] suffix — Z.AI only recognizes the bare code (glm-5.2) on
that path.

The [1m] suffix is a Z.AI convention documented specifically for the
Anthropic-compatible (Claude Code) endpoint, where it is set via
ANTHROPIC_DEFAULT_SONNET_MODEL=glm-5.2[1m]. It does not seem to be valid as a
chat-completions model value, so passing it straight through the openai-chat
adapter breaks.

Questions / suggestions

• Should opencodex strip or translate the [1m] (and similar bracketed)
  suffix before forwarding to the openai-chat adapter, while still enabling
  the 1M context another way (e.g. an explicit context/param)?
• Or should the GLM-5.2 1M variant be documented as only reachable through the
  anthropic adapter pointed at Z.AI's Anthropic-compatible endpoint?
• What is the currently recommended opencodex config to use the GLM-5.2 1M
  variant?

Happy to provide a redacted config.json or proxy request log if useful.

[lidge-jun]

Fixed on the feat/kiro-on-dev branch.

Root cause: opencodex forwards the model id verbatim as the chat-completions model value. glm-5.2[1m] was sent unchanged to Z.AI's OpenAI-compatible endpoint, which rejects the bracketed id (400, code 1211). The [1m] suffix is a Claude-Code / Anthropic-endpoint-only convention, not a valid OpenAI-compatible model value.

Fix (Option A): the openai-chat adapter now strips a trailing [...] suffix from the wire model value (stripBracketedModelSuffix), so glm-5.2[1m] resolves to glm-5.2. The strip is confined to the openai-chat path — the anthropic adapter keeps the suffix verbatim, since [1m] is valid there (Z.AI Anthropic coding endpoint). The 1M context window comes from the model catalog (modelContextWindows), not the model name.

Added unit tests (openai-chat strip + anthropic no-op) and a README note documenting both paths.

This will be closed once the change is merged to main.

[lidge-jun]

Fixed by 0b8d95c (fix(openai-chat): strip bracketed [1m] suffix from wire model id), now in dev/main/preview (released). The openai-chat adapter strips a trailing bracketed group (e.g. [1m]) from the wire model id via stripBracketedModelSuffix (src/adapters/openai-chat.ts:13,175), so glm-5.2[1m] is sent to Z.AI as bare glm-5.2 while the anthropic adapter keeps the suffix verbatim. Regression test: tests/openai-chat-model-suffix.test.ts. Closing as resolved.