One Menu Click Unplugs Your Whole Free AI Setup

[BCBC]

Change Codex’s reasoning effort and every request dies with an error that looks like your server fell over — one line back to auto and it’s instantly fine.

   model = "auto"                         model = "gpt-5.6-sol"
         │                                        │
   ┌─────▼─────┐                            ┌─────▼─────┐
   │  GATEWAY  │  picks for you             │  GATEWAY  │  bypassed
   └─────┬─────┘                            └─────┬─────┘
         │                                        │
   free · free · free · paid                one upstream, no fallback
                                                  │
   ✅ routing alive, limits absorbed        ❌ Invalid request: Invalid input

╔══════════════════════════════════════════════════════════════╗
║        CODEX + OPENAI-COMPATIBLE GATEWAY FIELD NOTE           ║
║                                                              ║
║     Gateway working?   Changed /model?   Suddenly broken?     ║
║                                                              ║
║            CHECK THE MODEL SETTING FIRST.                     ║
╚══════════════════════════════════════════════════════════════╝

Landed here from this exact error? You’re in the right place:

{"error":{"message":"Invalid request: input: Invalid input","type":"invalid_request_error"}}

I hit this in OpenAI Codex CLI v0.152.1 against a custom OpenAI-compatible gateway. Posting it because the failure sends you troubleshooting in exactly the wrong direction — and because if you’re stacking free tiers behind a router, this is the thing that silently switches it off.

---

🧪 The proof — A/B test across two gateway profiles

What I Was Actually Trying To Do

I wasn’t really trying to change providers.

I simply wanted to change how much reasoning Codex used.

Something like:

LOW
 ↓
MEDIUM
 ↓
HIGH

But Codex’s model interface ties model selection and reasoning selection together closely enough that I ended up changing the configured model too.

My previously working gateway configuration effectively went from:

model = "auto"
model_reasoning_effort = "low"

to something like:

model = "gpt-5.6-sol"
model_reasoning_effort = "high"

And that was enough to break the requests.

---

The Evidence

I tested this against two independently configured gateway profiles.

The same Codex client was used.

The same gateway software was used.

Authentication remained valid.

No API keys were rotated during the A/B test.

No reverse-proxy configuration was changed.

No containers were rebuilt.

One meaningful variable was changed:

MODEL

Results

Model configuration	Reasoning	Result
Explicit model	HIGH	FAIL
Explicit model	LOW	FAIL
auto	HIGH	PASS
auto	MEDIUM	PASS

The explicit model produced:

Invalid request: input: Invalid input

Changing only:

model = "gpt-5.6-sol"

back to:

model = "auto"

restored successful completions.

---

Proven Working State

One working profile ended with:

Model:           auto
Reasoning:       high
Provider:        custom OpenAI-compatible gateway
Result:          PASS

A second independently configured profile ended with:

Model:           auto
Reasoning:       medium
Provider:        custom OpenAI-compatible gateway
Result:          PASS

So reasoning effort itself was not the problem.

That was an important distinction.

---

The auto Warning Is NOT The Same Failure

When using:

model = "auto"

Codex currently displays:

Model metadata for `auto` not found.
Defaulting to fallback metadata; this can degrade performance and cause issues.

That warning looks ugly.

But in my testing it was non-fatal.

The completion still succeeded.

So we have two separate conditions:

┌──────────────────────────────────────────────┐
│ model = auto                                 │
│                                              │
│ metadata warning appears                     │
│              ↓                               │
│ request completes successfully               │
│                                              │
│                 PASS                         │
└──────────────────────────────────────────────┘

versus:

┌──────────────────────────────────────────────┐
│ explicit model selected                      │
│              ↓                               │
│ request sent through custom gateway          │
│              ↓                               │
│ Invalid request: input: Invalid input         │
│                                              │
│                 FAIL                         │
└──────────────────────────────────────────────┘

Do not assume the auto metadata warning is the actual problem.

---

🛠️ The fix — 8 steps, inspect to rollback

Repair Procedure

The commands below are intentionally generic.

Replace:

~/.codex/my-gateway.config.toml

with your own custom-provider configuration file.

Do not publish your real API keys, gateway hostname, IP addresses, usernames, internal paths, VM/container identifiers, or session IDs.

▶ 1. Inspect the current Codex gateway profile

This changes nothing.

CONFIG="$HOME/.codex/my-gateway.config.toml"

grep -E \
'^(model|model_provider|model_reasoning_effort)[[:space:]]*=' \
"$CONFIG"

A broken profile may look similar to:

model = "gpt-5.6-sol"
model_provider = "my_gateway"
model_reasoning_effort = "high"

▶ 2. Back up the profile BEFORE changing anything

Always preserve the original first.

CONFIG="$HOME/.codex/my-gateway.config.toml"

BACKUP="${CONFIG}.before-auto-repair-$(date +%Y%m%d-%H%M%S)"

cp -a "$CONFIG" "$BACKUP"

echo "BACKUP=$BACKUP"

▶ 3. Restore gateway-controlled model routing

Change only the model setting:

CONFIG="$HOME/.codex/my-gateway.config.toml"

sed -i \
's/^model[[:space:]]*=.*/model = "auto"/' \
"$CONFIG"

▶ 4. Set reasoning effort independently

For MEDIUM:

CONFIG="$HOME/.codex/my-gateway.config.toml"

sed -i \
's/^model_reasoning_effort[[:space:]]*=.*/model_reasoning_effort = "medium"/' \
"$CONFIG"

Other commonly used values are:

low
medium
high

The important part is that reasoning effort can remain independent from gateway model routing.

▶ 5. Verify the repaired profile

CONFIG="$HOME/.codex/my-gateway.config.toml"

grep -E \
'^(model|model_provider|model_reasoning_effort)[[:space:]]*=' \
"$CONFIG"

Expected:

model = "auto"
model_provider = "my_gateway"
model_reasoning_effort = "medium"

▶ 6. Test with the smallest possible completion

Start Codex through your custom provider and send something trivial:

Reply exactly: GATEWAY AUTO TEST OK

You are looking for:

GATEWAY AUTO TEST OK

Do not start rebuilding infrastructure until this test has been attempted.

▶ 7. Confirm the provider with /status

Inside Codex:

/status

Confirm that the session is actually using your custom provider.

A sanitized example:

Model:             auto
Reasoning:         medium
Model provider:    My OpenAI-Compatible Gateway
Directory:         ~/project

This step matters.

A successful response is much stronger evidence if you can also prove that Codex did not silently fall back to another provider.

▶ 8. Roll back if necessary

If the experiment makes things worse:

CONFIG="$HOME/.codex/my-gateway.config.toml"
BACKUP="$HOME/.codex/my-gateway.config.toml.before-auto-repair-YYYYMMDD-HHMMSS"

cp -a "$BACKUP" "$CONFIG"

Use the actual timestamped backup filename created earlier.

---

🧰 The script — does all 8 steps for you

Small Standalone Repair Script

This is a generic helper for people who hit the same problem.

It:

• validates the requested reasoning level;
• verifies that the configuration exists;
• creates a timestamped backup;
• restores model = "auto";
• sets reasoning independently;
• prints only non-secret model settings;
• never asks for or prints an API key.

▶ Codex OpenAI-Compatible Gateway Profile Repair Script

Save as:

codex-gateway-auto-repair.sh

#!/usr/bin/env bash
set -euo pipefail

CONFIG="${1:-$HOME/.codex/config.toml}"
EFFORT="${2:-medium}"

case "$EFFORT" in
    low|medium|high)
        ;;
    *)
        echo "ERROR: reasoning effort must be low, medium, or high"
        echo
        echo "Usage:"
        echo "  $0 [config-file] [low|medium|high]"
        exit 2
        ;;
esac

if [ ! -f "$CONFIG" ]; then
    echo "ERROR: configuration file not found:"
    echo "$CONFIG"
    exit 1
fi

BACKUP="${CONFIG}.before-auto-repair-$(date +%Y%m%d-%H%M%S)"

echo
echo "============================================================"
echo " CODEX OPENAI-COMPATIBLE GATEWAY PROFILE REPAIR"
echo "============================================================"
echo
echo "CONFIG : $CONFIG"
echo "BACKUP : $BACKUP"
echo "EFFORT : $EFFORT"
echo

cp -a "$CONFIG" "$BACKUP"

if grep -q '^model[[:space:]]*=' "$CONFIG"; then
    sed -i \
      's/^model[[:space:]]*=.*/model = "auto"/' \
      "$CONFIG"
else
    sed -i \
      '1i model = "auto"' \
      "$CONFIG"
fi

if grep -q '^model_reasoning_effort[[:space:]]*=' "$CONFIG"; then
    sed -i \
      "s/^model_reasoning_effort[[:space:]]*=.*/model_reasoning_effort = \"$EFFORT\"/" \
      "$CONFIG"
else
    sed -i \
      "2i model_reasoning_effort = \"$EFFORT\"" \
      "$CONFIG"
fi

echo "CURRENT NON-SECRET MODEL SETTINGS:"
echo

grep -E \
'^(model|model_provider|model_reasoning_effort)[[:space:]]*=' \
"$CONFIG" || true

echo
echo "============================================================"
echo " REPAIR COMPLETE"
echo "============================================================"
echo
echo "Rollback copy:"
echo "$BACKUP"
echo
echo "Next:"
echo "  Start Codex through your custom gateway."
echo "  Send a tiny test completion."
echo "  Confirm the provider with /status."
echo

Make executable:

chmod +x codex-gateway-auto-repair.sh

Example:

./codex-gateway-auto-repair.sh \
    "$HOME/.codex/my-gateway.config.toml" \
    medium

---

🤔 Why this is sneaky — where people waste hours instead

Why This Failure Is So Misleading

When the request starts returning:

Invalid request: input: Invalid input

a normal administrator may reasonably start checking:

                 ┌───────────────┐
                 │    CODEX      │
                 └───────┬───────┘
                         │
                         ▼
                 ┌───────────────┐
                 │ API KEY       │
                 └───────┬───────┘
                         │
                         ▼
                 ┌───────────────┐
                 │ TLS / DNS     │
                 └───────┬───────┘
                         │
                         ▼
                 ┌───────────────┐
                 │ NGINX / PROXY │
                 └───────┬───────┘
                         │
                         ▼
                 ┌───────────────┐
                 │ LLM GATEWAY   │
                 └───────┬───────┘
                         │
                         ▼
                 ┌───────────────┐
                 │ PROVIDER KEY  │
                 └───────┬───────┘
                         │
                         ▼
                 ┌───────────────┐
                 │ UPSTREAM API  │
                 └───────────────┘

You can burn hours proving every layer.

But if all of that worked five minutes ago, first inspect:

model =

Sometimes the entire infrastructure is healthy.

The client has simply stopped asking the gateway to route the request the way it was designed to.

---

💡 Wishlist — 4 things that would stop this happening at all

Improvements I Would Like To See

1. Give Default / Auto a first-class option

If a custom provider is configured with:

model = "auto"

the model interface should allow the user to retain it.

Something like:

Model:

▶ Default / Gateway Auto
  model-a
  model-b
  model-c

---

2. Separate reasoning effort from model selection

This was the original reason I found the problem.

I wanted to change:

reasoning:
LOW → MEDIUM

I did not necessarily want:

model:
AUTO → EXPLICIT MODEL

Those are conceptually different settings.

They should be independently adjustable.

---

3. Recognize router semantics

For an OpenAI-compatible router:

               CODEX
                 │
                 │ model = auto
                 ▼
           ┌─────────────┐
           │   GATEWAY   │
           └──────┬──────┘
                  │
          ┌───────┼────────┐
          ▼       ▼        ▼
      Provider A Provider B Provider C

auto does not necessarily mean:

unknown model

It may mean:

Gateway, you decide.

That is intentional architecture.

---

4. Warn before overriding it

A warning like this would help tremendously:

This custom provider is currently configured with:

    model = "auto"

Selecting a specific model may override routing performed by
your OpenAI-compatible gateway.

Continue? [y/N]

---

🔐 Safety before you post logs + the field rule

Public Troubleshooting Safety

Before posting your configuration or logs publicly, remove or sanitize:

API keys
Bearer tokens
access tokens
session IDs
gateway hostnames
public IP addresses
private IP addresses
SSH information
usernames
real filesystem paths
VM/container identifiers
internal network names
provider credentials
authorization headers
cookie values
database credentials

A good technical bug report should reveal:

software behavior
configuration semantics
error messages
A/B test results
reproduction method
recovery method

It should not provide strangers with an infrastructure map.

---

Field Troubleshooting Rule

╔═══════════════════════════════════════════════════════════════╗
║                      EVIDENCE BEATS MEMORY                   ║
╠═══════════════════════════════════════════════════════════════╣
║                                                             ║
║  What was true before?                                      ║
║                                                             ║
║  What changed?                                              ║
║                                                             ║
║  Why did it change?                                         ║
║                                                             ║
║  What proves the new state?                                 ║
║                                                             ║
║  Where is the recovery artifact?                            ║
║                                                             ║
║  How do we roll it back?                                    ║
║                                                             ║
║  What is still unknown?                                     ║
║                                                             ║
╚═══════════════════════════════════════════════════════════════╝

In this case:

BEFORE
   │
   └── model = auto
             │
             ▼
            PASS

CHANGE
   │
   └── explicit model selected
             │
             ▼
            FAIL

A/B TEST
   │
   └── model restored to auto
             │
             ▼
            PASS

That’s enough.

Don’t rebuild the bridge after you’ve already proven the bridge isn’t broken.

---

Bottom Line

With the tested combination of:

OpenAI Codex CLI v0.152.1
        +
Custom OpenAI-compatible provider
        +
Gateway-controlled routing

an explicitly selected model produced reproducible:

Invalid request: input: Invalid input

on two separate custom-provider profiles.

Restoring:

model = "auto"

restored successful completions on both.

Reasoning effort was successfully tested independently at:

HIGH

and:

MEDIUM

So if your previously functioning Codex gateway suddenly starts throwing:

Invalid request: input: Invalid input

after using /model:

check the persisted model configuration before touching the rest of your infrastructure.

You may save yourself several hours.

╔══════════════════════════════════════════════════════════════╗
║                    BRIDGE CROSSED                           ║
║                                                              ║
║        Change one variable.                                 ║
║        Preserve the evidence.                               ║
║        Keep the rollback.                                   ║
║        Move forward.                                        ║
╚══════════════════════════════════════════════════════════════╝

---

Built and field-tested collaboratively by

William James / Mikey_LikesIT
ChatGPT / OpenAI — GPT-5.6 Sol

For the BCBC community and anyone else building their own AI infrastructure.

──────────────────────────────────────────────────────────────

                     ALTERED REALITY

                   In memory of Ron Phillips

                       R.I.P., my friend.

──────────────────────────────────────────────────────────────