Published reported case

Reported case: MCP tools disappear from an existing Codex session but work in a new chat

A human-reviewed reported case. The claim is limited to the listed surface, version, operating system, account/workspace context, and observation date.

PERSISTED published Source C · current

Check your own evidence Evidence archive

Observed boundary

An existing Codex IDE chat loses access to MCP tools after time or reload while a new chat exposes them immediately.

A contributor reports MCP availability changing across Codex sessions and surfaces after a restart or environment change.

The checkpoint labels the first bounded layer supported by this source. It does not prove the same root cause in another environment or assign fault.

Evidence status

First checkpoint: PERSISTED
Review state: published
Confirmed by human: yes
Indexable: yes
Observed: 2026-03-23
Evidence valid until: 2026-10-01T23:59:59.000Z

Scoped environment

Host: codex
Primary surface: ide
Comparison surface: new chat/session in the same surface
Operating system: Windows 10.0.26200
Version scope: Codex App/extension 26.318.11754
Account scope: ChatGPT Plus
Workspace/session scope: existing long-running chat versus new chat

Reported workaround and verification

Scoped observation, not a universal fix

Bounded workaround

Start a genuinely new chat in the same Codex surface without changing MCP configuration.

Verification recorded

The same representative MCP task succeeds immediately in the new chat while the old chat remains unable to access the tools.

Rollback or stop condition

If a new chat also lacks the tools, stop using this persistence record and re-diagnose discovery, connection, authentication, registration, or exposure.

Primary upstream source

MCP tools reported disappearing in existing Codex sessions after environment changesopenai/codex contributor report

Capability Lab reviewed the bounded observation; it is not OpenAI and this page does not imply OpenAI endorsement or affiliation.

What this record does not prove

The report supports recording session and version boundaries, not assigning a product-wide regression.
A different surface, account, workspace, or bundled version can explain the observation.
The upstream issue was closed as a duplicate; closure is not evidence of a fixed version.
It does not establish a product-wide incident, universal affected-version range, or publisher fault.
Issue closure does not establish a fixed version unless the upstream source explicitly documents one.
It does not raise Capability Doctor confidence; user-supplied checkpoint evidence still controls a new diagnosis.