Skip to main content

OpenCode stand evidence (gateway request metrics)

Generated: 2026-10-02 21:23 UTC by scripts/compat/stand_evidence.py (IMP-82).

FieldValue
Data sourcesynthetic disposable PostgreSQL (seeded fixtures), run from imp-82 lane
Connectionpostgres@127.0.0.1:15439/yabanin (read-only session)
Driverpsycopg2
Windowlast 30 day(s), all orgs
RLS region (app.region)sg

Scope (all endpoints)​

endpointrequestsfirst daylast day
chat.completions112026-10-022026-10-02
responses22026-10-022026-10-02

chat.completions (POST /v1/chat/completions) totals​

  • Requests observed: 11 (observed).
  • Streaming (proxy: first_token_ms recorded on streamed responses): 7/11 (64%) — observed.
  • TTFT p50/p95: 704 / 1057 ms; avg duration 1087 ms.
  • HTTP >=400 rows: 2; rows with stored error text: 2 (counts only — error text is never printed).

Models (requested -> used)​

model_requestedmodel_usedrequestsstreamed (proxy)
yabanin/glm-5.3-flashglm-5.3-flash85
yabanin/glm-5.3glm-5.321
yabanin/qwen-localqwen3-local11

Errors by HTTP status​

status_coderequests
4291
5001

Tool calls​

  • Requests carrying tool calls: 3; tool-call rows: 5 (observed).
  • Top tools: bash (2), edit (1), grep (1), read (1).

User-Agent​

unavailable — request_logs.user_agent does not exist in this schema (client attribution is RM-80, not shipped yet). No user-agent evidence can be produced from this database; this is a schema fact, not a compatibility result.

Reading this report​

Status vocabulary: observed (rows found in the queried window), not_observed (queried, zero rows), unavailable (the schema cannot answer), not_run (the script could not run).

These are gateway-side request metrics. They are not a live CLI interoperability test: a live OpenCode/Codex/Claude run against the gateway stays not verified in docs/compatibility.md until such a run is recorded. not_observed here is never reported as verified anywhere.

Columns read: metadata/aggregates of request_logs and tool_calls only. Prompt/response content (response, upstream_request, upstream_response, messages, tool_call_invocations.arguments) and raw error text are never selected or printed; the rendered report is scanned for API-key/DSN patterns before writing.