openai / openai/codex

Codex incorrectly reports an MCP workflow as impossible without checking composable tools

Open
#45,306 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

app bug mcp model-behavior
Dominant language
Rust
Stars
125k
Forks
19.4k
PR merge metrics
PR metrics pending

Description

Summary

Codex incorrectly stated that placing a newly created Notion child page at a specific position was impossible through MCP.

The agent checked the limitation of one tool, move_pages, and generalized that limitation to the entire Notion MCP integration. The workflow was actually possible by composing create_pages with update_page using update_content, then fetching the result to verify it.

Reproduction

  1. Ask Codex to create a dated child page inside an existing Notion page at a specific chronological position.
  2. create_pages creates the child page, but the new page block appears at the end of the parent page.
  3. Codex observes that move_pages cannot reorder children within the same parent.
  4. Codex incorrectly concludes that MCP cannot place the page at the requested position.
  5. The correct solution is to update the parent page content: remove the appended page reference and insert the same page reference after the correct preceding date page.
  6. Fetching the parent page verifies that the page appears exactly once at the requested position.

Expected behavior

Before declaring a workflow impossible, Codex should inspect all relevant MCP write tools and consider whether multiple tools can be composed. It should distinguish a limitation of one tool from a limitation of the entire integration.

If the agent is uncertain, it should state the uncertainty and investigate instead of giving a definitive capability denial.

Actual behavior

Codex generalized the limitation of move_pages into a false statement that exact positioning was impossible through MCP. This caused the page to be created at the bottom of the document and initially led to an incorrect explanation of the available capability.

Why this matters

This is a false-negative capability judgment, not merely a failed tool call. A technically experienced user can challenge the answer, but a user who does not understand MCP may accept the incorrect limitation and stop using a paid feature that could have completed the task.

Requested improvement

Please improve MCP capability discovery and planning so that Codex:

  • inventories relevant read and write tools before declaring a workflow impossible
  • reasons about composable multi-tool workflows
  • avoids generalizing a limitation of one tool to the whole integration
  • verifies the final structure and placement before reporting success
  • clearly separates an unverified assumption from a confirmed product limitation

Environment: Codex desktop on macOS 26.5.1, Notion MCP, Plus plan.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by tracing the MCP capability-discovery and planning flow implicated by the reproduction, including how tool limitations are generalized. Check how multi-tool workflows and final verification are handled. Done means Codex inventories relevant tools, considers composition, distinguishes uncertainty from confirmed limitations, and verifies the resulting page position.

Written by the indexing model from the issue text.

Assessment

Tech stack
rust
Domain
ai
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
48/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.