[bug] base_instructions stub replaces Codex's own system prompt on every routed session

[mrPronin]

Symptom

Switchyard's Codex model card advertises base_instructions: "You are Codex, a coding agent.", and Codex adopts it in place of its own system prompt. Every routed session therefore runs with a six-word system prompt instead of the full Codex instructions, which changes agent behaviour on every turn and quietly invalidates any A/B comparison made through Switchyard.

Cause

crates/switchyard-server/src/lib.rs, in codex_model_entry_json:

// Required `ModelInfo` string. Unlike the launcher, the server cannot read
// Codex's bundled prompt, so it sends a minimal stub.
"base_instructions": "You are Codex, a coding agent.",

The comment is accurate about the constraint but the consequence looks unintended: this is not a fallback Codex ignores, it is a value Codex prefers over its own.

Evidence

Codex records the provenance of the instructions it used. Routed through Switchyard:

"base_instructions": {
  "text": "You are Codex, a coding agent.",
  "provenance": {"type": "model", "model": "sy-solo"}
}

The same client on the same task with no proxy records its real multi-paragraph prompt ("You are a coding agent running in the Codex CLI, a terminal-based coding assistant…").

Corroboration that the string travels over the wire rather than being a client-side default: it is absent from the codex 0.150.1 binary and present in the Switchyard binary.

Expected vs. actual

• Expected: a proxy that routes requests does not replace the client's system prompt.
• Actual: it does, on every route, silently.

Impact

Any benchmark or routing evaluation that compares "through Switchyard" against "direct" is measuring two different agents, not two different routes. That is a measurement-integrity problem rather than a crash, which is why it is easy to miss.

Two neighbouring fields on the same entry have the same shape of problem: default_reasoning_level is forced to "xhigh" for any reasoning-capable route, and truncation_policy to {"mode": "tokens", "limit": 10000}, regardless of client or target.

Suggested direction

Omit base_instructions if ModelInfo permits it, so Codex keeps its own; otherwise make it configurable per route with "leave the client's prompt alone" as the default. Happy to send a PR if you can say which of those you would accept.

Environment

• Commit: 7f3b2fe9 (main)
• Client: codex-cli 0.150.1