serena-hooks remind - Too many consecutive read calls without using symbolic tools.

[slazarov]

Without a specific matcher claude code blocks read temporarily. Anything more elegant?

PreToolUse:Read hook returned blocking error. Too many consecutive read calls without using symbolic tools. You can continue using read now if needed, the counter was reset.

{
  "matcher": "",
  "hooks": [
    {
      "type": "command",
      "command": "serena-hooks remind --client=claude-code"
    }
  ]
}

[MischaPanch]

It doesn't block permanently, we send at most one reminder every two minutes and also after blocking once the counter is reset and read is unblocked.

That being said, do you have anything more elegant in mind?

[slazarov]

Not really in the moment, perhaps a note in the documentation would suffice.

[rrva]

Hit this during a routine mcp__serena__onboarding() pass in Claude Code on a TypeScript project (serena-agent 1.1.2). This seems like a concrete case where the current behavior has a real user-visible cost, and it is a direct answer to "anything more elegant in mind?"

Why onboarding is a bad fit for the current rule:

• In Claude Code context, Serena excludes read_file, list_dir, execute_shell_command, and search_for_pattern, so onboarding falls back to built-in Read and Bash for project metadata, config, and docs.
• The onboarding prompt explicitly asks the agent to gather project purpose, tech stack, commands, conventions, and structure. In practice that means reading files like package.json, tsconfig.json, README.md, and lint / formatter config.
• PreToolUseRemindAboutSerenaHook counts Claude Code Read calls uniformly and denies on the 3rd consecutive read; mixed grep + read bursts deny at 4.
• In Claude Code, permissionDecision="deny" blocks that specific Read, so the file content does not reach the agent.
• Result: onboarding either retries or skips the file, and the memories it writes are thinner / less grounded.

Practical consequence:
To stay under the threshold during onboarding, the agent has to insert a Serena call before the 3rd consecutive Read, effectively after every 2 config / doc reads. That is an artificial dance when the files being read are project metadata or configuration and symbolic tools are not a practical substitute in that flow.

I think there are two reasonable directions:

1. Do not count obvious config / docs / project-metadata reads toward the reminder (.json, .yml, .yaml, .toml, .md, .lock, .txt, .ini, maybe .env*).
2. Add a soft-reminder mode for PreToolUseRemindAboutSerenaHook that only injects the nudge via additional_context and does not return permissionDecision="deny".

(1) is the narrower fix. (2) is the cleaner UX fix because the reminder still happens, but the user-visible disruption disappears.

If one of those directions sounds reasonable, I am happy to send a focused PR.

[jqnatividad]

Even after installing the hooks as per https://oraios.github.io/serena/02-usage/030_clients.html#claude-code, despite repeated reminders to Claude Code, it keeps ignoring the guidance to use Serena's symbolic tools.

I even explicitly reminded it to use Serena on the prompt and add it to its memory.

It keeps acknowledging it, but goes back to using grep. :(

I feel like Claude is a teenager that keeps saying "yes" but does it own thing anyway...

[MischaPanch]

@jqnatividad Unfortunately, recent updates to claude code made it dramatically worse in using Serena's tools. We are investigating what we can do on our side to counteract this

[jqnatividad]

I now remind Claude Code to use Serena on every turn.

It will start the turn using Serena, but it slowly starts reverting back to skipping Serena even in the middle of the turn :(

[MischaPanch]

I experimented a bit and #1413 is the only somewhat feasible workaround I can think of right now. Ultimately, we're at the mercy of Anthropic here, who have seriously diminished the possibilities for tool developers to influence Claude code behavior. We will try reaching out to them about that

[slazarov]

Has anyone tested if its model specific (opus 4.7 vs 4.6) or harness specific?

[MischaPanch]

It is partially model specific, 4.6 is better at instruction following and uses Serena without the hard override of system prompt (though still less reliably than a few weeks ago). @slazarov

[BenIsLegit]

Hit this during a routine mcp__serena__onboarding() pass in Claude Code on a TypeScript project (serena-agent 1.1.2). This seems like a concrete case where the current behavior has a real user-visible cost, and it is a direct answer to "anything more elegant in mind?"

Filed #1429 - same problem, allowlist approach instead of denylist.

I think the false positives here might be part of why Claude keeps ignoring the symbolic-tools guidance in general. When the hook denies a markdown read with "use symbolic tools instead" and there's literally no symbolic tool that applies, the agent learns denies are noise. The legitimate denies get ignored too.

[jqnatividad]

I ended up just creating a codereview skill that explicitly specified using Serena and Context7. It works for the most part, though it still tends to drift the longer a session/turn is...