Incoming Request“Rename getUser to fetchUser in src/api.ts”Incoming requestPOST /responses · model=siftInspecting requestJev returns typed signals, not codeEvaluating routing strategypolicy.yaml · first match winsCODEX SIFTSmart routerjev-1.13.0task_kinddifficultyneeds_reasoningneeds_planninghigh_stakescross_cuttingawaiting requestSmall edit in one file≈300 msflashsmall asks & editsFLASHgpt-5.6-lunaeffort lowusage 0.17/1.35standbycrafteveryday workCRAFTgpt-5.6-terraeffort mediumusage 0.55/1.35standbyforgehard & high-stakesFORGEgpt-5.6-soleffort highusage 1.35/1.35standbyCodex TUIsimulated responseflash · Small edit in one fileRenamed getUser to fetchUser insrc/api.ts.
romanmeclazcke/codex-sift · MIT

CODEX SIFT

Model router for the Codex CLI

Each user turn goes to flash, craft or forge before Codex bills it. On twelve live prompts that cut estimated quota 53.7% versus always-forge.

Results

53.7% less quota on twelve prompts

Live Jev classified twelve representative Codex prompts into flash, craft or forge. Estimated quota fell from 16.2 to 7.5 against an always-forge baseline. The report measures routing cost. Answer quality is out of scope.

Estimated Codex quota, 12 prompts

53.7%saved vs always-forge

Always-forge baseline

every turn on gpt-5.6-sol

16.20

Sift routed

flash + craft + forge lanes

7.50

Measured by npm run e2e:report with live Jev answers. It measures routing and quota efficiency against the baseline. It does not claim answer quality. Read the full report.

12 / 12

turns classified by live Jev

0 degraded

11/12

landed on the expected lane

policy expectation

418 ms

average Jev latency

in this run

4 · 5 · 3

flash · craft · forge

lane mix

Lane usage per prompt

Gray is the always-forge cost (1.35). Blue is what the routed lane used.

  • flash · gpt-5.6-luna
  • craft · gpt-5.6-terra
  • forge · gpt-5.6-sol
Hover or focus a bar to see the prompt behind it.

How it works

A decision layer before Codex spends the turn

  1. you

    codex-sift

  2. Codex TUI

    model=sift

  3. Sift proxy

    127.0.0.1 · Jev + policy.yaml

  4. Upstream

    chatgpt.com/backend-api/codex or api.openai.com

model_provider

Intercepts the Responses call

Codex has no hook that changes the model before a turn is billed. Sift starts a loopback Responses proxy and launches Codex with process-local -c overrides. TUI, sessions and sandbox stay native.

TypeSafe Jev

Asks Jev six atomic questions

On every user turn Sift asks Jev for task_kind, difficulty, needs_reasoning, needs_planning, high_stakes and cross_cutting. Typed signals come back in about 300 ms. Jev never writes code or picks a model.

first match wins

Maps signals with policy.yaml

YAML rules turn those signals into flash, craft or forge. If Jev times out or the key is missing, Sift fails open to on_jev_down (default craft) and marks the decision degraded.

model + reasoning.effort

Rewrites model and effort only

Only model and reasoning.effort change. Your ChatGPT or API-key login is forwarded upstream. Tool-loop continuations stay pinned to the chosen lane.

What is not routed

  • Tool-loop continuations of the same user turn
  • codex-auto-review and compact/review prompts
  • Codex housekeeping that is not a /responses call
  • Invocations with -m / --model pointing at a real model

Sift is Codex-only. It is not an MCP server and does not patch ~/.codex/config.toml. Pin another model in /model and routing pauses until you pick Sift again.

Quick start

Install, set up, then run Codex through Sift

Install with npm install -g codex-sift, run codex-sift setup to assign models to flash, craft and forge, then start a session. Arguments match codex.

Install from npm, then assign a model to flash, craft and forge.

# recommendednpm install -g codex-siftcodex-sift setup# or without a global installnpx codex-sift setupnpx codex-sift

FAQ

Questions about Codex Sift

What is Codex Sift?
Codex Sift is a model router for the OpenAI Codex CLI. It sends each user turn to the flash, craft or forge lane before Codex bills the turn.
How does Codex Sift choose a model?
On every user turn Sift asks TypeSafe Jev for typed signals such as task_kind and difficulty. Your policy.yaml maps those signals to a lane. First matching rule wins.
Does Codex Sift change the Codex TUI?
No. Sift launches Codex through a loopback model_provider. Sessions, tools and sandbox stay native to Codex.
How much quota can routing save?
On twelve live Jev prompts in reports/e2e-routing.md, estimated quota fell from 16.2 to 7.5 versus an always-forge baseline, a 53.7% cut. That measures routing cost, not answer quality.
Who built Codex Sift?
Codex Sift was built by Roman Meclazcke (github.com/romanmeclazcke, x.com/romanmeclazcke_, linkedin.com/in/roman-meclazcke) and is released under the MIT license.
How do I install Codex Sift?
Install from npm with npm install -g codex-sift, then run codex-sift setup. You can also use npx codex-sift without a global install. The package is at npmjs.com/package/codex-sift.

Next step

Install Sift and keep the Codex TUI

Stop spending your strongest model on “explain this file”. Add the decision layer in front of Codex, then check the difference with codex-sift report. Built by Roman Meclazcke.

X · @romanmeclazcke_LinkedInGitHub

↑ Back to the request path