[Bug] OpenCode Go bills gpt-5.6-terra requests as gpt-5.6-luna (Codex Memories background jobs) #42342

Description

Using the OpenCode Go plan as the provider for OpenAI Codex desktop
(model_provider = "opencode-go", base_url = https://opencode.ai/zen/go/v1,
wire_api = responses, main model deepseek-v4-flash), Codex's built-in
Memories feature runs background jobs (memory_stage1,
memory_consolidate_global) after each session. Those jobs POST
{"model": "gpt-5.6-terra", ...} (reasoning_effort=medium) to
/responses with cwd ~/.codex/memories.

On the OpenCode workspace usage page, these requests are billed/displayed as
gpt-5.6-luna (go) — a model the user never selected and that never
appears as a request anywhere in the local Codex logs (zero
gpt-5.6-luna requests locally). So the gateway is either silently
substituting/renaming the model, or the usage page is mislabeling it, and the
result is charges the user cannot attribute to anything they did.

Examples from today (Asia/Shanghai, all from local memory_stage1 /
memory_consolidate_global jobs whose request log shows model=gpt-5.6-terra):

• 21:17 — gpt-5.6-luna (go) $0.0496 (stage-1 summary job)
• 22:26 — gpt-5.6-luna (go) 612,201 in / 3,956 out, $0.2826 (global memory consolidation)

Local Codex request log (~/.codex/logs_2.sqlite):
POST to https://opencode.ai/zen/go/v1/responses: {"model":"gpt-5.6-terra", ...}

Plugins

None relevant (plain Codex desktop + opencode-go provider).

OpenCode version

Go plan gateway (opencode.ai/zen/go/v1); usage page opencode.ai/workspace/.../usage.

Steps to reproduce

1. Subscribe to OpenCode Go; configure Codex desktop with
   model_provider = "opencode-go" (zen/go/v1, responses API), main model
   deepseek-v4-flash.
2. Keep Codex Memories enabled (~/.codex/config.toml:
   [memories] generate_memories = true).
3. Finish a Codex session (triggers memory_stage1) and let global
   consolidation (memory_consolidate_global) run.
4. Open the workspace usage page → new records appear under
   gpt-5.6-luna (go).
5. Check local Codex logs → requests were sent with model=gpt-5.6-terra;
   no gpt-5.6-luna request exists locally.

Screenshot and/or share link

Can share usage-page screenshots and log excerpts on request.

Operating System

Windows 11 x64

Terminal

Codex Desktop (26.803.81509) / Codex CLI 0.147.0-alpha

Expected behavior

• Either gpt-5.6-terra is available on the Go plan and is billed/displayed
  as gpt-5.6-terra, or the gateway rejects it (4xx) instead of silently
  executing and billing a different model (gpt-5.6-luna).
• The usage page should show the actual requested/billed model so users can
  trace unexpected charges.
• Ideally, server-side enforcement (per-key model allowlist/blocklist) so
  hardcoded client-side model usage like this can be stopped without changing
  the client (related to [FEATURE]: Zen model disable toggle is not enforced at the gateway - Codex keeps auto-using gpt-5.6-luna and billing continues #41697, which covers the disable-toggle gap for
  gpt-5.6-luna; this report adds the distinct model substitution
  behavior and the Codex Memories background caller).

Happy to provide workspace/session details to maintainers for debugging.