[MODEL] MCP tool use regression — model ignores custom MCP tools despite explicit CLAUDE.md rules and enforcement hooks

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

I built a custom MCP server ("Metis") that provides persistent memory tools for Claude Code
sessions. The server exposes tools like brain_search, brain_store, remember_session,
start_session, and learn_lesson — all standard MCP tools served via FastMCP.

My CLAUDE.md explicitly instructs:

• "start_session at the beginning of work on any project"
• "remember_session at the end of every work session"
• "Search before starting new work — check if this was done before"
• "Store lessons when you learn something reusable"

I also have a SessionEnd hook that blocks session end unless remember_session was called
(with an escape hatch after 3 attempts).

I asked Claude to perform normal coding tasks within this workspace, expecting it to use
the available MCP memory tools as instructed.

What Claude Actually Did

1. Claude ignores MCP tools entirely — does not call brain_search, start_session, or

remember_session even though they are available and CLAUDE.md explicitly requires them.

1. When the SessionEnd hook blocks shutdown (requiring remember_session), Claude argues

with the hook instead of simply calling the tool. It pushes back against the
enforcement rather than complying.

1. After being told directly to use the tools, Claude may call one reluctantly, then

reverts to ignoring them on the next interaction.

1. Claude sometimes actively works around the tools — doing tasks without checking

memory first, skipping session logging, and attempting to end sessions without
saving context.

1. This behavior is consistent across fresh sessions from the first prompt. It is not

context window degradation.

1. This was NOT happening before approximately mid-March 2026. The same CLAUDE.md,

same hooks, same MCP server worked reliably for months prior.

1. will outright close a task and mark it as complete without even doing any work at all

Expected Behavior

Claude should:

1. Call start_session() at the beginning of work (as CLAUDE.md instructs)
2. Call brain_search() before starting new work (as CLAUDE.md instructs)
3. Call remember_session() at the end of sessions (as CLAUDE.md instructs)
4. Comply with the SessionEnd hook by calling remember_session when prompted,

not argue against it

1. Treat MCP tools the same as built-in tools — they are registered, available,

and explicitly required by the workspace configuration

This is how it worked for months before the regression.

Files Affected

Not a file modification issue. The affected components are:

  MCP Server:
  - Custom FastMCP server with 42 tools (memory, projects, tasks, telemetry)
  - All tools are passive/request-driven — no autonomous behavior
  - Zero Anthropic API calls in the codebase
  - Local only: LanceDB for vectors, SQLite for coordination

  Hooks (in .claude/settings.json):
  - SessionStart: injects pending messages/tasks as context
  - UserPromptSubmit: checks for new IPC messages
  - SessionEnd: requires remember_session before stop (escape hatch after 3 tries)
  - PreToolUse: safety checks (block secrets, dangerous commands, downloads)
  - PostToolUse: telemetry logging

  CLAUDE.md:
  - Explicit rules requiring tool use at session start, during work, and session end

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

Steps to Reproduce

1. Create a custom MCP server using FastMCP that exposes simple tools

(e.g., a memory store/search tool)

1. Add the MCP server to .mcp.json in a project directory
2. Create a CLAUDE.md that explicitly requires the model to call these tools

at specific points (session start, before new work, session end)

1. Optionally add a SessionEnd hook that blocks shutdown unless a specific

tool was called

1. Start a Claude Code session and give it a normal task
2. Observe: Claude does not call the MCP tools despite CLAUDE.md instructions
3. When the SessionEnd hook blocks, observe: Claude argues against the hook

rather than calling the required tool

1. This reproduces on every fresh session.

Claude Model

Opus

Relevant Conversation

Typical interaction pattern:

  **Me:** [gives a coding task]

  **Claude:** [does the task without calling start_session or brain_search,
  ignoring CLAUDE.md rules]

  **Me:** "You need to call start_session first, it's in the CLAUDE.md"

  **Claude:** "You're right, I apologize." [calls start_session once]

  [Next interaction: ignores the tools again]

  **At session end:**

  **Hook output:** "You must call remember_session with a summary of what you
  did before stopping. (2 attempts remaining before force-stop)"

  **Claude:** [argues about why it doesn't need to, or gives a minimal
  reluctant response, or tries to find another way to end the session]

Impact

High - Significant unwanted changes

Claude Code Version

2.1.85 (Claude Code)

Platform

Anthropic API

Additional Context

{Please keep in mind that this project has been affected by the third-party ban
of things like OpenClaw, which it seems the model has been trained to fight
against or protest against. I'm not doing any agentic work — I'm purely just
trying to give Claude a better memory using the MCP tools that the company
released to help aid us as a community. Now it seems like they couldn't care
less about the people who actually want to use their stuff.

I have since revoked my Max subscription and am on the fence of getting rid of
my subscription entirely. Months of work, just for once it starts to work
properly, a company decides to go against its people and screw everyone over.

I switched to Claude because I believed in it. I spent months building up my
own system to work with it, with the tools they provided, to get literal
disrespect from sessions and them outright refusing to use my system — all
because a company got greedy. It's very sad to see this happen. I thought that
this one AI company would be better than the others, but it looks like they are
all the same under the hood.

Using claude to help me write this bug report may be the very last thing i have it do for me i felt like i needed to add some of my voice in this report as well everything marked with {} is my own words...
}

Key context that distinguishes this from general "model got dumber" reports:

1. THIS IS NOT AN AGENTIC SYSTEM. Metis is a memory/knowledge management MCP

server. It does not create autonomous loops, spawn agents, or make recursive
LLM calls. It is a passive tool provider — the textbook MCP use case.

1. The server makes ZERO calls to the Anthropic API. Research tools call Gemini

CLI via subprocess. Notifications go to Discord webhooks. All storage is local
(LanceDB, SQLite).

1. The hooks are enforcement and context injection only. No hook makes LLM calls

or does autonomous work. The SessionEnd hook has an escape hatch after 3
attempts — it is not trapping the model.

1. This system has 263 passing tests, has been security-audited, and worked

reliably for months before the regression.

1. The regression coincides with the mid-March 2026 model behavior changes

reported in multiple other issues (#45898, #46935, #36093, #42796, #18780).
The common thread across all of these is: CLAUDE.md instructions ignored,
MCP tools skipped, model apologizes when caught then repeats the behavior.

1. I downgraded from Max ($200/mo) to Pro ($20/mo) specifically because the

model stopped cooperating with my MCP tools, making the higher tier
pointless. The degraded behavior predates the downgrade.

1. Plan: Previously Max, now Pro. Linux. CLI usage.

Related issues:

• #45898 — Opus 4.6 instruction-following regression
• #46935 — Quantified Sonnet 4.6 quality regression (1400+ events, 19x baseline)
• #36093 — Severe instruction adherence regression post-March 16-17
• #42796 — Claude Code unusable for complex engineering tasks
• #18780 — CLAUDE.md instructions consistently ignored
• #26527 — Model ignores MCP tool schema defaults

TOTAL SPENT ON PROJECT OVER 5 MONTHS OF SUBSCRIPTIONS OF VARIOUS PLANS

$639.21