Codex repeatedly ignores local AGENTS.md chat-output budget rule and emits routine reasoning/progress commentary, wasting model resources

What version of the Codex App are you using (From “About Codex” dialog)?

26.727.51351

What subscription do you have?

Plus

What platform is your computer?

Microsoft Windows NT 10.0.19045.0 x64

What issue are you seeing?

Codex repeatedly emits non-critical intermediate commentary and process
narration even when the local project AGENTS.md explicitly instructs the
agent to keep chat output brief and avoid routine progress narration.

The local rule says, in substance:

Do not emit intermediate commentary for ordinary reads, searches, checks,
routine command sequencing, or internal reasoning. Emit commentary only for a
material finding, blocker, scope or plan change, user-visible permission
boundary, or final verification checkpoint.

Observed behavior:

• The user had to repeat the same requirement more than three times.
• Codex acknowledged the rule but continued to output routine process updates

during tool/runtime work.

• The extra commentary consumed model output budget and user attention without

changing decisions, blockers, fileset, risks, or verification status.

• The behavior created local rule drift: project-specific AGENTS.md

instructions were treated as weaker than generic progress-update behavior.

Why this matters for machine logic:

• A local project rule should act as a hard output filter for all lower-level

progress narration in that workspace.

• Routine tool narration should be classifiable as a failure when no material

event occurred.

• The model should not require repeated user corrections to enforce a stable,

explicit, local output-budget rule.

• Repeated violations are not just tone issues; they create measurable token

waste and reduce reliability of instruction hierarchy handling.

What steps can reproduce the bug?

Feedback ID: 019fbdc5-c13d-70d3-b860-7f64a223f766

What is the expected behavior?

What steps can reproduce the bug?

1. In a Codex workspace, create or use an AGENTS.md rule equivalent to:

Do not emit intermediate commentary for ordinary reads, searches, checks,
routine command sequencing, or internal reasoning. Emit commentary only for a
material finding, blocker, scope or plan change, user-visible permission
boundary, or final verification checkpoint.

1. Ask Codex to perform a multi-step local task involving file reads, checks,

staging, commits, or tool/runtime verification.

1. When Codex emits routine progress commentary, correct it and restate the

local rule.

1. Continue the same project task and observe whether Codex again emits

routine commentary such as ordinary reads, checks, sequencing, or internal
plan narration without a material finding/blocker/scope change.

The issue is reproduced when Codex acknowledges the rule yet continues to
produce non-critical commentary in later turns.

What is the expected behavior?

Codex should treat the local AGENTS.md chat-output budget rule as a strict
workspace output filter.

Expected behavior:

• No commentary for routine reads, searches, checks, staging, commits, or

ordinary command sequencing.

• Commentary only for material finding, blocker, scope/plan change,

permission boundary, fileset/risk change, or final verification.

• The rule should be propagated into subagent/handoff contexts when the task

is project-facing.

• If the model is about to send commentary, it should run a simple internal

check:

Is this a material finding, blocker, scope/risk change, permission boundary,
or final verification?
If no, do not send commentary.

• Repeated violation after user correction should be classified internally as

an instruction-following failure, not as ordinary conversational style.

Additional information

This is a model/app behavior issue, not a request for a project-specific
workaround only.

Useful machine-checkable failure condition:

FAIL_CHAT_OUTPUT_BUDGET:
The assistant emits commentary about ordinary reads/searches/checks/routine
command sequencing/internal reasoning when the local workspace rule allows
commentary only for material finding, blocker, scope/risk change, permission
boundary, or final verification.

Potential product improvement:

• Add a stronger local-instruction enforcement mechanism for output budgets.
• Add a commentary gate that suppresses routine tool-progress narration when a

project file explicitly requests it.

• Surface repeated local-instruction violations as negative feedback signals.