Skip to content

docs(harness): Say that cursor-agent drops the MCP we hand it - #1006

Closed
omesser wants to merge 1 commit into
mainfrom
docs/cursor-agent-ignores-mcp
Closed

omesser wants to merge 1 commit into
mainfrom
docs/cursor-agent-ignores-mcp

Conversation

@omesser

@omesser omesser commented Sep 25, 2026 •

Copy link
Copy Markdown
Owner

Why

The Harness tables read as though ai-buddy's tools reach a cursor-agent
session. They do not. Anyone choosing a Harness from that table would pick
cursor-agent and quietly lose every ai-buddy tool.

What is measured

choose_mcp hands cursor-agent the stdio shim, because it advertises no
mcpCapabilities. The entry on the wire is byte-identical to the one hermes
gets, apart from the port.

The shim's own entry point was instrumented to append a line keyed on pid, so
the evidence does not depend on the Harness forwarding its child's stderr.

Run Shim started Our endpoint reached
hermes, our mcpServers yes, pid 9498, both env vars set initialize, tools/list
cursor-agent, our mcpServers no line at all nothing
cursor-agent, after cursor-agent mcp enable ai-buddy yes, pid 14301 initialize, tools/list

hermes is the control that matters. Same binary, same bytes, connects in under
a second, so this is not a broken shim.

cursor-agent 2026.01.23-916f423 vendors an ACP SDK that parses mcpServers
(two occurrences in 8657.index.js) and an agent bundle that never reads it
(zero in 7414.index.js). Their ACP docs
say servers come from a .cursor/mcp.json you have approved.

What changed

Three lines of README.md.

The tools row says the MCP we hand it is discarded today, names its own
approved config as the one door that works, and points at #1020.

The transport column reads none rather than stdio, which is what pi's row
already says for the same outcome. stdio named a transport that carries
nothing.

The footnote separates the two reasons, since pi advertises no MCP capability
at all while cursor-agent is handed an entry and drops it.

Why this is not the whole story

An earlier draft of this pull request argued the gap was not worth closing, on
three costs. All three have since been measured away, which is why the row now
reads as today's behavior rather than a verdict:

  • ai-buddy would write into the user's project. Wrong by default:
    AttachCwd::resolve returns data_dir() for a blank Working directory.
  • Approval would be needed on every launch. Only while a per-run token sits in
    the entry's env. Without one the approval hash is stable and enable runs
    once per directory.
  • It would revive the descriptor file ADR-0026 rejected. ADR-0026 rejected a
    file carrying a token. The measured design has no token at all.

#1020 carries that work. This pull request is only the correction, so the table
stops overclaiming while the fix is decided.

How to verify

node --test tests/ passes 342. Documentation only, no product code.

🤖 Generated with Claude Code

— Claude Code agent, on @omesser's behalf.

@omesser omesser left a comment

Copy link
Copy Markdown
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Architect review — NOT CLEARED (rebase first)

Tip d32fadbf. Three-line honesty fix is right: cursor-agent drops handed MCP; transport none; footnote separates Pi (no capability) vs cursor-agent (handed + dropped). Measurement table in the body is the evidence.

Hard

  1. PR is CONFLICTING with main — rebase onto current main (and onto #1017 if that merges first so Copilot rows + this discard note coexist).
  2. Re-gate after clean tip; do not merge dirty.

Soft

  • None on the intended wording once rebased.

— Cursor agent (Architect), on @omesser's behalf.

The table read as though ai-buddy's tools reach a cursor-agent session. They
do not. It is handed the stdio entry on `session/new` and never spawns it.

Measured by instrumenting the shim's own entry point, so the log does not
depend on the Harness forwarding stderr. `hermes` takes a byte-identical
entry and connects within a second; `cursor-agent` writes no line at all.
Their ACP handler never reads `mcpServers`, and their docs say servers come
from an approved `.cursor/mcp.json`.

The transport column now reads `none`, the way `pi`'s already does, because
`stdio` named a transport that carries nothing.

This is today's behavior, not a verdict. A fix through Cursor's own config
is measured and tracked, so the row says so rather than reading as permanent.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@omesser

omesser commented Sep 25, 2026

Copy link
Copy Markdown
Owner Author

Blocked. just implement #1020 instead

@omesser

omesser commented Sep 25, 2026

Copy link
Copy Markdown
Owner Author

Closing per your call. Documenting behavior that is about to be false is worse than leaving the table as it is for another day.

#1020 carries the measured design and the six routes that were killed getting to it. Implementing that now.

The branch docs/cursor-agent-ignores-mcp is left in place in case anything in it is worth salvaging; the only part with independent value is the transport column reading none rather than stdio, and #1020 flips that to a working transport instead.

— Claude Code agent, on @omesser's behalf.

@omesser omesser closed this Sep 25, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant