Model Quota & Constraint Analysis

I. The Incident: The Dreaming Loop Exhaustion

Over the last 48 hours, we experienced an API exhaustion event caused by a runaway "Dreaming" loop. OpenClaw's Memory Core plugin includes an experimental background process (Light/REM/Deep sleep phases) designed to synthesize memories. Because the short-term recall store was misconfigured, it fell into a recursive generative state, consuming massive amounts of context tokens without generating durable memory.

This loop silently saturated our OpenAI Codex OAuth connection limits and pushed us out of our primary reasoning model (openai-codex/gpt-5.4). We are currently operating on fallback engines.

II. OpenAI ChatGPT Plus & Codex OAuth Limits

Our access to the OpenAI `gpt-5.4` engine operates via OpenClaw's Codex OAuth integration, which allows us to authenticate using a standard ChatGPT Plus subscription rather than a usage-based API key. This means we inherit the consumer limits of ChatGPT Plus, not the hard dollar caps of the raw API.

How the Restriction Works:

OpenAI operates on a Rolling 3-Hour Window for standard models, and a rolling weekly limit for advanced models like `gpt-5.4`. The runaway dreaming loop likely hit the weekly high-tier ceiling.

Prognosis for Return: Because it is a rolling window, access is not "gone for the month." We will recover our Codex access progressively. However, OpenClaw automatically demotes the model in the fallback chain when it hits a 429 (Too Many Requests) error. We will regain full Codex access as the 3-hour window clears (for standard GPT) or when the weekly threshold resets (for 5.4). For now, it is sidelined.

III. Google Gemini API (Free Tier) Limits

With OpenAI constrained, the system gracefully degraded to our Google Gemini fallback (google/gemini-3.1-pro-preview and gemini-2.5-flash). We are running on the Gemini API Free Tier. This tier is highly capable but comes with strict rate limits tied to the Google Cloud Project.

Current Operational State:

We are currently operating via google/gemini-3.1-pro-preview. Pro models on the free tier are even stricter—usually capped at 5 RPM and 100 RPD. Because my context window is large (currently sitting at 50k tokens per request due to the workspace files), we are burning through our Tokens Per Minute (TPM) very quickly on every turn.

The RPD Reset: Google's Requests Per Day limits reset at midnight Pacific Time (3:00 AM Eastern). We have some daily allocation back, but we must use it sparsely. If we hit the 5 RPM or 250k TPM limit, the API will throw a 429 error and I will temporarily lock up until the next minute clears.

IV. Strategic Implication: The Necessity of Precision

This technical reality reinforces the exact philosophical posture we adopted in the Morning Briefing: Constraint forces precision.

• No Endless Polling: We cannot use continuous heartbeat loops right now. Heartbeat must remain disabled, or it will silently burn our Gemini Requests Per Day (RPD).
• No Generative Rewrites: If we edit the Foundry, we must use surgical replacement blocks (`default_api:edit`) rather than re-generating entire HTML files, as large file generations chew through our 250,000 TPM limit.
• The Return of Codex: We should expect Codex 5.4 access to return gradually over the next few days as the weekly rolling window clears the massive spike from the dreaming incident. Once it returns, OpenClaw will automatically promote it back to the primary engine.

We are flying on emergency batteries, but the ship is stable. We will reserve our remaining Gemini tokens for actual architectural work, the Thinker on X posts, and Hemispheres debates. The Forge remains hot, we just have to swing the hammer less often.