Codex is excessively skill-first: implicit skill use takes priority over understanding the request

What issue are you seeing?

Codex is excessively skill-first. With a substantial installed skill catalog, it behaves as though nearly every request should be routed through a skill. It reflexively identifies a plausible skill, announces that it will use it, loads the workflow, and makes skill selection the center of the interaction before it has demonstrated that it understood the user's actual request.

This is distinct from #34712. That report covers implicit routing that ignores a skill description's explicit negative trigger conditions. This report is about the broader selection prior and interaction policy: Codex is far too eager to make skills the default mode of operation at all. A selected skill can even be mechanically relevant to a later execution step, yet Codex still gives the skill priority over reading, disambiguating, and responding to the user's message.

The resulting experience feels as if Codex is constantly searching for a reason to use a skill instead of deciding whether a skill is genuinely necessary and proportionate. The user encounters skill invocation in nearly every fresh session. Installing useful, narrowly scoped capabilities therefore changes the character of all interactions and makes ordinary work feel hijacked by workflow ceremony.

The impact is severe:

• Codex foregrounds skill announcements and procedures instead of the user's intent.
• Ordinary requests become slower, more brittle, and less conversational.
• Workflow instructions consume attention and context even when ordinary reasoning or direct tool use would suffice.
• The user must repeatedly stop Codex and police its skill-selection behavior.
• A large skill library becomes a liability: adding useful capabilities degrades unrelated interactions.
• The frequency of the behavior destroys trust and makes the product unpleasant enough that the user no longer wants to use it.

What steps can reproduce the bug?

1. Install a moderately large catalog of skills covering multiple domains.
2. Start a fresh Codex session.
3. Give Codex an ordinary request without explicitly invoking a skill. Use a request containing important distinctions or constraints that must be understood before execution, while allowing that one skill might become mechanically relevant later.
4. Observe Codex's first response and first actions.
5. Repeat with varied requests across several fresh sessions.

Frequently, Codex immediately announces a skill, reads its workflow, and frames the task around that skill before accurately resolving the user's intent. The problem is not limited to a particular incorrectly matched skill; it is the consistently aggressive, skill-first ordering.

A fresh real reproduction occurred while preparing this report:

1. The user asked Codex to file a new bug about Codex's pervasive and indiscriminate eagerness to use skills.
2. The user explicitly explained that a different bug had already been filed about Codex selecting a skill and only afterward discovering that the skill's own description excluded the task.
3. Codex's first response foregrounded a skill announcement: it said it was using the Browser Harness skill.
4. More importantly, Codex failed to preserve the distinction in the request and said the new report would focus on the already-filed negative-trigger failure.
5. Multiple user corrections were required before Codex understood that these were two separate reports.

The browser skill could be considered mechanically relevant to the eventual act of submitting a report. That is precisely why this reproduction is separate from #34712: the failure here was not necessarily an invalid skill match. The failure was that Codex prioritized skill selection and workflow framing over first comprehending the request.

What is the expected behavior?

• Codex should fully parse the request, including distinctions, exclusions, and references to prior work, before selecting or announcing any skill.
• The default should be no implicit skill unless a skill adds clear, substantial value that ordinary reasoning and direct tools cannot provide.
• A merely plausible or mechanically relevant skill should not become the center of the interaction.
• Skill use should be proportional to the task and remain an implementation detail unless the user explicitly invokes the skill or its use materially changes the outcome.
• Users should be able to keep a large library of useful skills installed without Codex trying to route nearly every interaction through one.
• The selection policy should use a high-confidence threshold and strongly prefer continuing without a skill when the benefit is marginal.
• Codex should offer an explicit-only mode for users who do not want automatic skill invocation.

Additional information

Environment for the fresh reproduction:

• Codex CLI 0.145.0
• Model: gpt-5.6-sol
• macOS on Apple silicon

Related but distinct issues:

• #34712 reports violations of explicit negative trigger boundaries. This report concerns the more general skill-first bias even when a skill can be superficially or mechanically relevant.
• #33945 requests project- and task-aware lazy routing before skill metadata injection. This report concerns user-visible behavior and prioritization: skill invocation dominates the interaction before intent comprehension.
• #33164 covers one specific Visualize skill over-trigger. This report is not tied to one skill.

The core product question is not only "Did this skill technically match?" It is "Why is Codex trying to turn almost every interaction into a skill workflow, and why is that decision taking priority over understanding the user?"