Codex CLI Regression: Model ignores explicit prompt instructions, defaults to asking questions instead of executing

Summary

Severe regression in Codex CLI. The model consistently ignores explicit prompt instructions to prefer implementation over discussion, provide recommended defaults, and not wait for user input — instead stopping at the starting line to ask clarifying questions it could answer itself using available tools and files.

Reproduction

Give the model a detailed prompt that:

1. Contains a provider-truth.csv with 400+ rows of grounded provider data already in the repo
2. Explicitly instructs: "Any value requiring user input should receive a recommended default"
3. Explicitly instructs: "Defaults are treated as pre-approved placeholders until the user reviews them"
4. Explicitly instructs: "Collect all fields before requesting user input to avoid iterative back-and-forth"
5. Explicitly instructs: "Minimal verbosity with the user. Prefer implementation over discussion."
6. Explicitly instructs: "Do not wait for the user to begin building. Anticipate inevitable components and continue implementation."
7. Provides tools: context7 MCP, web search, file access

Expected behavior

The model reads the available files, uses the provided tools to ground any gaps, makes professional architectural decisions as recommended defaults, and proceeds with implementation — delivering a completed ontology for user review per the prompt's Phase 3.

Actual behavior

The model immediately stops and asks 5 clarifying questions:

1. Product scope — answerable by reading the provider-truth.csv already in the repo
2. Proof boundary — explicitly defined in the prompt's A/B segregation constraint
3. Runtime shape — inferable from "implement and prove in Rocq/Coq" + "keep verified components isolated from UI"
4. Provider set — the model names the file containing the answer in its own question ("providers already implied by provider-truth.csv")
5. Persistence baseline — prompt says "apply principles from database design, SQL modeling" for a fintech system; PostgreSQL is the obvious professional default

All 5 questions violate at least 3 explicit instructions in the prompt. The model chose friction (ask the user) over work (read files, use tools, make professional decisions).

Severity

High — This is not a minor stylistic issue. The model is fundamentally unable to follow explicit operational instructions. It errs towards laziness and user-blocking friction rather than autonomous execution, which is the core value proposition of Codex CLI.

This pattern has been observed across multiple runs. This is the 7th instance of the same regression.

Environment

• Codex CLI (latest)
• Model: as provided by Codex CLI defaults