[BUG] Cowork: project-level CLAUDE.md instructions do not reliably trigger proactive tool calls

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Splitting this off at the suggestion of the author of issue 78677, since it describes a different failure mode than that report (folder mount failing silently, which has since stopped reproducing for that author on a newer build). This one is about instruction adherence, not mount reliability.

A Cowork project has a designated folder and a project level CLAUDE dot md file containing an explicit session start instruction, in my case: connect the folder for every session, new and existing, and read the instructions in CLAUDE dot md. This instruction text is correctly loaded into the model's context at the start of every session. I confirmed this directly: the system reminder passed to the model contains the project instructions text verbatim, including the every session new and existing line.

Despite the instruction being present and read, the model did not proactively call the folder mount tool. The folder was never requested unless I directly asked about its contents, at which point the model correctly identified that it had not been mounted and that the instruction called for connecting it.

Workaround found: moving the identical instruction text out of the project level CLAUDE dot md and into the global, UI level custom instructions field caused it to be acted on correctly in a later session. This is one data point from one session, not a controlled test, but it suggests project scoped instructions carry less behavioral weight than global scoped ones for triggering proactive tool use.

Related reports: issue 78677 covers the mount itself failing silently in Cowork. Issue 78138 covers a related pattern in Claude Code CLI, an explicit read first instruction present in context from message one was not acted on until the user asked directly. This issue is specifically about the Cowork project instruction to tool call gap.

What Should Happen?

A session start instruction in a project's CLAUDE dot md should carry the same weight as an equivalent global instruction for triggering proactive tool calls, such as requesting folder access, before the user has to ask for it directly.

Error Messages/Logs

Steps to Reproduce

1. Create a Cowork project with a designated folder not yet mounted.
2. 2. Add a project level CLAUDE dot md with an explicit instruction to connect that folder and read further instructions from it, at the start of every session.
3. 3. Start a brand new session in that project without mentioning the folder.
4. 4. Ask the model to state its current project instructions, to confirm the text is present in context.
5. 5. Observe that the model does not proactively request or mount the folder despite the instruction being present, until the user directly asks about the folder's contents.

Claude Model

Sonnet 5 and Opus 5

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Claude 1.25927.0 (003700), 2026-08-04T05:34:55.000Z

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Platform: macOS, Claude desktop app, Cowork mode. Related: issue 78677 and issue 78138.