Fix Codex default_permissions Without Disabling the Sandbox

Yan Gao — Fri, 07 Aug 2026 09:35:08 +0000

The Codex error saying that permission profiles exist without default_permissions is not a request to disable safety. It means the effective configuration contains profile definitions but does not select a default profile.

The fastest fix is to decide which configuration model the run should use.

Path 1: keep the legacy sandbox model

Use this path when the run already relies on sandbox_mode, approval_policy, or [sandbox_workspace_write] settings.

Keep those settings together and remove unused [permissions.*] blocks. An explicit --sandbox flag also selects this model for the run.

Path 2: use permission profiles

Use this path when you want a built-in or named profile. Remove loaded legacy sandbox_mode settings, set default_permissions, and provide the matching profile block.

default_permissions = "project-edit"

[permissions.project-edit]
extends = ":workspace"

Start from a bounded profile. Add filesystem or network exceptions only after the task proves it needs them.

No-prompt mode is not full access

approval_policy = "never" suppresses approval prompts. It does not widen the selected sandbox or permission profile. A workspace-bounded run remains workspace-bounded.

This distinction matters for unattended work. Removing prompts can make a bounded workflow repeatable, but it is not a substitute for defining the boundary.

Verify the effective run

Do not trust one TOML file in isolation. Record the Codex version, inspect every loaded config layer, resolve the named profile from the intended project root, and test write and network behavior with harmless operations.

A bounded profile-resolution probe on Windows is:

codex sandbox -P project-edit -C C:\absolute\project\path -- powershell -NoProfile -Command "Get-Location"

This proves the profile name resolves. It does not prove write or network behavior, so test those separately.

The complete decision table, Windows sandbox guidance, failure signatures, and six-step boundary checklist are in the canonical TGWise guide:

https://tgwise.com/guides/codex-windows-permissions/

Disclosure: I maintain TGWise Agent Lab.

Why Claude Code Ignores CLAUDE.md: Loading, Adherence, and Enforcement

Yan Gao — Fri, 07 Aug 2026 06:34:14 +0000

When Claude Code appears to ignore CLAUDE.md, rewriting the file is often the wrong first move.

The symptom can come from three different layers:

The instruction never entered the active session.
The instruction loaded, but the model did not follow it in the task.
The instruction describes an outcome that should have been enforced outside the model.

These layers need different fixes.

1. Prove the file loaded

Run /context and inspect the Memory files section. If the intended CLAUDE.md is absent, investigate the working directory, filename, exclusions, nested scope, and active settings sources.

Do not strengthen the wording of a file that is not in context.

2. Check scope and conflicts

A loaded instruction can still be irrelevant to the current path or conflict with user, project, nested, or path-scoped instructions.

Replace vague rules with observable behavior. A useful rule names:

when it applies;
the action to take;
the artifact or command that proves completion; and
the condition that blocks completion.

For example, "follow our API standards" is difficult to verify. "When editing src/api/**, run npm run test:api and report the failing test if it does not pass" is testable.

3. Test behavior instead of recall

Do not ask the model whether it read the file. Correctly repeating a rule proves recall, not compliance.

Run one bounded task where the rule creates a visible difference. Inspect the diff, command output, or generated artifact outside the chat. If the result matters, repeat the task in a fresh session.

4. Move critical boundaries out of prose

If a failure must never occur, CLAUDE.md should explain the boundary but should not be its only control.

Use the mechanism that can reject the bad outcome:

permissions or a sandbox for access boundaries;
PreToolUse hooks for command checks;
formatters and linters for deterministic style;
tests for behavior;
CI for release gates; and
human approval for irreversible external actions.

The practical rule is simple: use instructions for orientation and use tooling for enforcement.

The complete six-step diagnostic, current Anthropic sources, compaction checks, and a local decision tool are available in the original TGWise guide:

https://tgwise.com/guides/claude-code-ignoring-claude-md/