Breaking

Codex users circulate an instruction-debt audit before switching to GPT-6 Astra

September 7, 2026 at 04:47 EDT

Software Dev & Coding
AI Agents

As OpenAI’s GPT-6 Astra reaches Codex, developers are circulating a read-only “instruction debt audit” that inspects an entire agent setup for wasted context, conflicting instructions, bad skill triggers, unclear permissions, and broken stopping rules—before the new model is put to work.

September 2026 · Codex

Audit the harness before GPT-6 Astra

A read-only instruction-debt audit is circulating for Codex users moving to GPT-6 Astra. It flags wasted context, conflicting rules, bad skill triggers, unclear permissions, and broken stop conditions — and it applies no file changes.

How often agent files go wrong

Share of 100 popular repositories with each smell. Column height is proportional.

62%

Lint leakage

42%

Context bloat

35%

Skill leakage

Each full block is 10 percentage points; the shorter top block on skill leakage is the remaining 5%.

Layers inspected

Skill-list context cap

32 KiB

Default AGENTS.md limit

Updated Codex harness vs GPT-5.6 Sol

Mind2Web task completion. Column height scales with the 1.9x speedup.

1.0x

GPT-5.6 Sol

1.9x

GPT-6 Astra + Codex

Takeaway: the model upgrade is faster in the harness, but leftover instructions still occupy the window on every turn.

What the audit does

1. MapAGENTS.md, skills, hooks, permissions

→

2. InspectFive layers, including stop rules

→

3. Walk throughFive paper scenarios, no live runs

→

4. ProposeExact diffs; apply nothing

After a cleanup

Users who ran a pass say about half their rules were fighting each other, and tightening the stack changed day-to-day agent behavior.

Limits

The audit is a prompt, not an official scanner. It must not widen permissions, and it should not treat older rules as dead just because the model is newer.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.