Blocked, not rejected
---
title: Blocked, not rejected
published: true
tags: tooling, aiagents, claudecode, buildinpublic
canonical_url: thekilted.dev
---
{% embed youtube.com %}
The most capable-looking command in a fifty-seven-item tool catalogue turned out to be five
function calls to a server this setup doesn't run. No trial was run, because narrating what a
missing server would do isn't a test. The verdict needed a name of its own, and blocked fit.
SuperClaude is a configuration layer for Claude Code, installed as an add-on rather than built in:
slash commands, agents and behavioural modes, fifty-seven items in the catalogue in total. One of
them read as the single most capable thing in it.
pm-agent, invoked as /sc:pm, read on paper like a full session-management layer. Cross-session
memory that survives between conversations. Checkpointing every thirty minutes. Self-evaluation, so
the agent could ask itself whether it was actually done before reporting done. The kind of feature
list that makes a reader comparing AI coding tools stop and read it twice.
The check
The catalogue's own summary of /sc:pm never says how any of that actually works. Reading the full
source (agents/pm-agent.md and commands/pm.md in the SuperClaude repository) does.
Every one of those mechanisms turned out to be a literal call to a function on a Serena server.
Serena is an MCP (Model Context Protocol) server, a separate program that Claude Code connects to
for extra tools. The calls are list_memories(), read_memory(), write_memory(),
think_about_task_adherence(), think_about_whether_you_are_done(). Cross-session memory was a
named function call to a server. So was self-evaluation: another named call to the same server.
None of those functions exist without a Serena MCP server actually connected, and at the time of
the review this setup's .mcp.json defined no servers at all. None, checked directly in the file
rather than assumed.
The dependency did not stop there. For full behaviour, /sc:pm also wanted seven more MCP
servers (eight, all told, all absent) loaded dynamically through a third-party Docker gateway,
airis-mcp-gateway. Reading more of the command's own source only lengthened the list of things
that would need to
exist first, never shortened it.
To run the same check on any tool, search its command files for calls to functions that need a
connected server, then compare that list with your own MCP configuration.
Source: SuperClaude-Org/SuperClaude_Framework
A repeat
The shape had shown up in the framework before. An earlier command, /reflect,
had already been excluded from an earlier pass for the identical reason. Its own file states
plainly, in its own "Will Not" section:
> Operate without proper Serena MCP integration and reflection tool access.
Same disqualifying issue, same missing server, a different command. Once is a fact about one
command. Twice is a fact about the framework's own design pattern: describe a capability in the
documentation, implement it as a call to infrastructure the reader is never told they need.
A second objection
Set the missing server aside, and there was still a second, unrelated problem. The command's own
file declared itself the default operating foundation:
> The DEFAULT operating foundation that runs automatically at every session start... the default
> entry point for all interactions.
Auto-delegating every request to sub-agents before a human ever saw it. That was a takeover of a
workflow already in place, duplicating this setup's own working four-file lifecycle (NOTES.md,
MEMORY.md, CLAUDE.md, DECISIONS.md).
A different project in the same workspace had already declined the exact same shape for the
framework's Deep Research mode: an ungated feature that wanted to become the default way every
session behaved, rather than a tool invoked on request. This setup read that rule, then reached the
same objection on pm-agent and checked it against that record. The missing server was the
independent finding; the second objection leaned on that earlier record.
The verdict
No trial was run. On paper the command looked capable, but a trial without the missing server
would only narrate what write_memory() would do, with nothing real underneath the narration.
A trial without the server would be a plausible performance of a test.
This setup keeps a standing rule for exactly that failure mode: trust the artifact over the summary.
Reading pm-agent's actual source instead of its one-line catalogue description is what caught the gap
before a fake trial could run.
So the verdict on record was not rejected. It was blocked.
A blocked verdict names the specific missing thing and stays open to being wrong about it. If a
Serena MCP server is ever installed here for an unrelated reason, this is worth a second look.
The session-takeover design would still need weighing on its own merits. A smaller, scoped-down
version of the trial was floated too and turned down; that was the project owner's call, and no
second missing dependency lay behind it. A rejected verdict would have closed that door for good.
This one stayed open on purpose.
---
Originally published at thekilted.dev.
Mentions
Blocked, not rejected --- title: Blocked, not rejected published: true tags: tooling, aiagents, claudecode, buildinpublic canonical_url: thekilted.dev --- {% embed youtube.com %} The most capable-looking command in a fifty-seven-item tool catalogue turned out to be five function calls to a server this setup doesn't run. No trial was run, because narrating what a missing server would do isn't a test. The verdict needed a name of its own, and blocked fit. SuperClaude is a configuration layer for Claude Code, installed as an add-on rather than built in: slash commands, agents and behavioural modes, fifty-seven items in the catalogue in total. One of them read as the single most capable thing in it. pm-agent, invoked as /sc:pm, read on paper like a full session-management layer. Cross-session memory that survives between conversations. Checkpointing every thirty minutes. Self-evaluation, so the agent could ask itself whether it was actually done before reporting done. The kind of feature list that makes a reader comparing AI coding tools stop and read it twice. The check The catalogue's own summary of /sc:pm never says how any of that actually works. Reading the full source (agents/pm-agent.md and commands/pm.md in the SuperClaude repository) does. Every one of those mechanisms turned out to be a literal call to a function on a Serena server. Serena is an MCP (Model Context Protocol) server, a separate program that Claude Code connects to for extra tools. The calls are list_memories(), read_memory(), write_memory(), think_about_task_adherence(), think_about_whether_you_are_done(). Cross-session memory was a named function call to a server. So was self-evaluation: another named call to the same server. None of those functions exist without a Serena MCP server actually connected, and at the time of the review this setup's .mcp.json defined no servers at all. None, checked directly in the file rather than assumed. The dependency did not stop there. For full behaviour, /sc:pm also wanted seven more MCP servers (eight, all told, all absent) loaded dynamically through a third-party Docker gateway, airis-mcp-gateway. Reading more of the command's own source only lengthened the list of things that would need to exist first, never shortened it. To run the same check on any tool, search its command files for calls to functions that need a connected server, then compare that list with your own MCP configuration. Source: SuperClaude-Org/SuperClaude_Framework A repeat The shape had shown up in the framework before. An earlier command, /reflect, had already been excluded from an earlier pass for the identical reason. Its own file states plainly, in its own "Will Not" section: > Operate without proper Serena MCP integration and reflection tool access. Same disqualifying issue, same missing server, a different command. Once is a fact about one command. Twice is a fact about the framework's own design pattern: describe a capability in the documentation, implement it as a call to infrastructure the reader is never told they need. A second objection Set the missing server aside, and there was still a second, unrelated problem. The command's own file declared itself the default operating foundation: > The DEFAULT operating foundation that runs automatically at every session start... the default > entry point for all interactions. Auto-delegating every request to sub-agents before a human ever saw it. That was a takeover of a workflow already in place, duplicating this setup's own working four-file lifecycle (NOTES.md, MEMORY.md, CLAUDE.md, DECISIONS.md). A different project in the same workspace had already declined the exact same shape for the framework's Deep Research mode: an ungated feature that wanted to become the default way every session behaved, rather than a tool invoked on request. This setup read that rule, then reached the same objection on pm-agent and checked it against that record. The missing server was the independent finding; the second objection leaned on that earlier record. The verdict No trial was run. On paper the command looked capable, but a trial without the missing server would only narrate what write_memory() would do, with nothing real underneath the narration. A trial without the server would be a plausible performance of a test. This setup keeps a standing rule for exactly that failure mode: trust the artifact over the summary. Reading pm-agent's actual source instead of its one-line catalogue description is what caught the gap before a fake trial could run. So the verdict on record was not rejected. It was blocked. A blocked verdict names the specific missing thing and stays open to being wrong about it. If a Serena MCP server is ever installed here for an unrelated reason, this is worth a second look. The session-takeover design would still need weighing on its own merits. A smaller, scoped-down version of the trial was floated too and turned down; that was the project owner's call, and no second missing dependency lay behind it. A rejected verdict would have closed that door for good. This one stayed open on purpose. --- Originally published at thekilted.dev.