switch-model

Switch model

Run the helper to change the default model. The user's switch request authorizes the config edit. Resolve script paths relative to this skill directory. Requires Python 3.11+ and FireConnect. On this machine, codex-switch-model runs the same helper from the terminal.

1. Use a familiar alias, or list exact IDs with python3 scripts/switch_model.py --list --search NAME. --list reports each model's id, display name, context window, vision support, price, and whether FireConnect curates it (in_codex_catalog).
2. Run the switch, for example python3 scripts/switch_model.py kimi. Any other serverless model works by short or full id, for example qwen3p8-max or accounts/fireworks/models/qwen3p8-max.
3. Report the saved model, provider, effort, and session restart requirement.

codex-switch-model --list
codex-switch-model --list --search glm
codex-switch-model kimi
codex-switch-model kimi fast
codex-switch-model glm flash
codex-switch-model deepseek pro
codex-switch-model deepseek-flash-latest
codex-switch-model --effort medium kimi
codex-switch-model --dry-run kimi
codex-switch-model --status
codex-switch-model grok
codex-switch-model gpt

Use $switch-model kimi inside Codex. Also recognize the user's $swtich-model typo.

FireConnect owns the Fireworks model catalog. This helper ships no catalog of its own. It stages FireConnect's own fireconnect codex on output in a private temporary home, then reads the generated Codex catalog and model list from that directory. Aliases point at Fireworks latest routers. The helper edits that catalog in exactly one way: it restores the max reasoning tier described below. Models present in FireConnect's catalog keep that curated metadata. Any other model from Fireworks' model API gets a generic catalog entry, so pinned versions such as kimi-k3 and models FireConnect does not curate, such as qwen3p8-max, are selectable by name. Generic entries use Fireworks' canonical full id in both config.toml and the catalog, a low/medium/high reasoning ladder plus max on the deep-tier families, the reported context length, and a conservative 65,536-token window when Fireworks reports none. Claiming a window that is too large delays compaction until the server rejects the request, so the fallback errs small. The catalog can include region-only deployments such as kimi-k3-us and glm-5p3-flash-us.

FireConnect curates reasoning tiers by hand and grants max to only a few base refs, so DeepSeek, GLM, and Kimi families arrive with a three-tier ladder. The helper restores max on those families. Fireworks documents the deep tier for DeepSeek V4/V4.1, Kimi K3, and GLM 5.2; a live probe confirmed it for GLM 5.3. Measured cost: max uses two to twelve times the output tokens of low and rarely improves accuracy on small coding tasks. Prefer low or high, and reach for max only when a task needs deep deliberation. The list merges FireConnect's curated entries with every chat-plus-tools model on Fireworks' model API. It excludes MiniMax, non-tool models, and the FireRouter BYOK router. MiniMax currently conflicts with Codex tool-message ordering. Missing or ambiguous IDs fail without changing the config. Do not invent model IDs or claim every serverless model supports Codex. Fireworks' model API lists more models than it serves, so a generic switch first runs one tiny validation request (developer plus user message, at most 64 output tokens). Broken models fail with the server's reason before any file changes; dry runs and curated switches run no inference. Example limits found on 2026-09-18: qwen3p8-max rejects Codex's developer-role messages in its chat template, while nemotron-lightning-3p5-30b-a3b passes.

Kimi K3 works in CLI sessions but currently fails in Codex desktop sessions. Fireworks rejects any tool schema that sets type beside $ref, and a desktop app tool sends that shape. The turn ends with stream disconnected before completion: JSON Schema not supported: when using $ref, type should be defined in the referenced schema instead of the parent schema. DeepSeek and GLM accept that shape. Verified on 2026-09-17: kimi-fast-latest failed in this desktop thread, passed in two CLI sessions, and a direct Fireworks probe reproduced the exact rejection. The warning covers the whole Kimi family, including pinned ids. Use Kimi in the CLI, or DeepSeek/GLM in the desktop, until the app or Fireworks fixes it.