Bare gpt-reserve gets no catalog row outside codexDesktopAuthless: tool_mode falls back to direct, ~250k tokens of inlined tool schemas at boot #3844

Summary

The bare gpt-reserve wire id gets no catalog row outside codexDesktopAuthless mode. Codex then falls back to default model settings, which means tool_mode is unset (i.e. direct) instead of code_mode_only. Every MCP/plugin tool schema is inlined into the request instead of being deferred behind the exec namespace.

On a Codex Desktop install with the normal plugin set this is ~250k input tokens on turn 1, against a fallback model_context_window of 258400. The session compacts after nearly every action.

Environment

• opencodex 2.45.0 (also reproduced on 2.42.0)
• Codex 0.153.4, Codex Desktop 26.901.51231, macOS 15 (arm64)
• Normal loopback routing: root openai_base_url = http://127.0.0.1:10100/v1, codexDesktopAuthless unset
• model_catalog_json = ~/.codex/opencodex-catalog.json (39 models, no gpt-reserve)

Evidence

Turn 1 of a fresh Desktop thread, prompt "hi", from the rollout token_count event:

{"input_tokens": 252450, "cached_input_tokens": 3328, "output_tokens": 94,
 "model_context_window": 258400}

~/.opencodex/usage.jsonl confirms the wire id resolves fine upstream — the request itself succeeds:

{"model": "gpt-reserve", "resolvedModel": "gpt-5.6-luna",
 "requestedEffort": "max->xhigh", "status": 200, "totalTokens": 252544}

Minimal repro

Point Codex at a request sink and compare a catalogued slug against gpt-reserve:

codex exec --config openai_base_url='"http://127.0.0.1:10199/v1"' -m <slug> "hi"

gpt-reserve behaves identically to a nonsense slug. Breakdown of the 56 inlined tools under gpt-reserve (CLI only — Desktop is larger):

Note that [features] tool_search = true and tool_search_always_defer_mcp_tools = true are set in config.toml and do not help — the per-model tool_mode field is the gate.

Cause

createReserveCatalogProjection() in src/codex/catalog/reserve.ts returns early:

if (!isEffectiveCodexDesktopAuthless(config) || mainSelectors.length === 0) return undefined;

Two consequences for a user on normal openai_base_url routing:

1. The projection never fires, so no Reserve row is emitted at all.
2. Even when it does fire, the row is account-bound (<selector>/gpt-reserve per isReserveCatalogProjection), while the Desktop model picker sends the bare gpt-reserve.

native-models.ts is explicit that this is intended — "Reserve wire identity, not a globally available native catalog registration." But the account still serves the bare slug upstream (200 above), so a user can select it and silently lose both tool deferral and the correct context window with no warning anywhere.

Sync does recognise a hand-added bare row: observedReserveCatalogSource() captures it into the root opencodex_reserve_source field. It just never emits a model row from it outside the authless gate, so ocx sync strips the row on every run (including the auto-sync on ocx start).

Workaround

Clone the gpt-5.6-luna row with slug: "gpt-reserve" into opencodex-catalog.json. That restores tool_mode: code_mode_only, context_window: 300000, effective_context_window_percent: 95, auto_compact_token_limit: 270000, and the request drops to 120,912 chars with 0 inlined tools — identical to Luna. It has to be re-applied after every sync.

Suggested fix

Either:

• Emit a bare gpt-reserve row whenever a genuine reserve observation exists (opencodex_reserve_source is populated), independent of codexDesktopAuthless; or
• If Reserve is genuinely unsupported outside authless mode, warn on sync when the persisted Codex model selection is a slug with no catalog row — silently degrading to tool_mode: direct and a smaller window is very hard to attribute.

More generally: any slug the backend accepts but the catalog omits will hit this. A sync-time warning for "selected model has no catalog row" would catch the whole class.