plat 504

← Pulse platform issue index

PLAT-504 — "Selected model is at capacity" (Codex) was only visible in the terminal, not in the chat

Coordination	Value
State	fixed on main (provider 63976ce, pinned in the builder); needs a restart
Date	2026-10-05
Owner	coding-agent-bridge

Source

Owner, Upwork chat on Codex: the terminal showed "■ Selected model is at capacity. Please try a different model." The chat area showed nothing. A stored-events check of the chat found no event containing that text.

What Codex records (structured, real examples from 2026-10-05 10:30 and 10:31)

Rollout file: event_msg task_complete with last_agent_message: null and error: {"message": "Selected model is at capacity. Please try a different model.", "codex_error_info": "server_overloaded"}. codex exec --json stdout: {"type":"turn.failed","error":{"message":...}} (preceded by {"type":"error","message":...}); the message can be a JSON-serialized API error.

Cause

• Interactive (tmux) path: the rollout's turn error was read only when the reply text was empty, so any pane text (the error banner itself) hid it.
• Structured path (workflow steps): neither the stdout turn.failed nor the rollout error was read; the step failed as codex run failed: exit status 1: Reading additional input from stdin....

Done

• CodexTurnError (message + codex_error_info); Capacity() is true for server_overloaded; its text is "codex-cli model at capacity: ... (try again shortly or pick a different model)".
• Interactive: the rollout's error decides, whatever the pane says (no terminal text matching). Structured: turn.failed on stdout (else the error event), plus the rollout's code found by the run's own thread id. Structured records only.
• Checked live: a real Codex failure on the structured transport (unsupported model) now returns a CodexTurnError with the provider's reason. One unit test on the real capacity record.

Left

• Not seen live: an actual "at capacity" turn on either path (cannot be provoked); the interactive path was checked by code and the record shape only.
• No retry or capacity-wait handling (PLAT-101 style) for server_overloaded: it is reported as a failure with a clear reason. Decide separately.
• How the chat renders the error is the existing LLM-error card; not changed.