Codex Agent Models

When you install the Codex target (mewkit init --target codex), every bundled agent ships with a concrete Codex model and reasoning effort pinned in its TOML. This mirrors how the Cursor and Claude Code toolkits pin a model per agent (see Cursor Agent Models and Claude Agent Models); session-level tier routing by task complexity is a separate layer (see Model Routing).

Codex lets a custom agent set model and model_reasoning_effort directly, and those override any inherited default — so the bundle assigns each agent the tier that fits its role rather than paying top-tier cost for bounded, wrapper-driven work.

How the mapping works

Each agent file at .codex/agents/<name>.toml carries both keys:

name = "planner"
model = "gpt-5.6-sol"
model_reasoning_effort = "high"
description = "..."
developer_instructions = """..."""

Agents are grouped into three model tiers by what the work actually demands:

model_reasoning_effort is tuned independently (low / medium / high) — a bounded agent can still need high effort, and a capable model can run at medium for routine work.

Use the Codex model id, not the API alias

Codex custom agents take the exact model id, which is not always the name used in API docs. Confirm what your account exposes with codex debug models before pinning:

gpt-5.6-sol   gpt-5.6-terra   gpt-5.6-luna   gpt-5.5   gpt-5.4   gpt-5.4-mini

Write gpt-5.6-sol — never the API-style alias gpt-5.6, which is not a Codex model id and will not resolve.

Core and delivery agents

Jira and Confluence agents

Integration agents run behind a validated CLI wrapper, so most are bounded enough for the mini tier. The exceptions are the ones that synthesize evidence or can damage a source-of-truth system.

Operating policy

1. Verify the model is exposed to your Codex account (codex debug models) before relying on a pinned tier — a model your plan cannot run will fail at spawn time.
2. Measure success by correct tool action, confirmation compliance, user correction rate, and escalation rate — not token cost alone.
3. Promote a mini agent to gpt-5.6-terra when it misroutes an action, mishandles an ambiguous query, or needs repeated remediation.
4. Destructive-action confirmation and wrapper validation are independent of model selection — lowering a tier never lowers a safety gate.

Changing an agent's model

Edit the two keys in the agent's TOML:

name = "jira-issue"
model = "gpt-5.4-mini"            # any model your Codex account exposes
model_reasoning_effort = "medium" # low | medium | high

Omitting both keys makes the agent inherit the Codex session default. The bundle pins them so behaviour is predictable across machines and accounts.