You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
docs: The pi row claims it does not receive ai-buddy's MCP, on a run that could not show that #985
| pi | Keeps its own read, bash, edit, and write tools, has no web tool, and does not receive ai-buddy's MCP. |
Why it is not evidenced
That row was measured through scripts/probe-harness.sh, which serves no MCP endpoint at all. #984 has the detail: the probe exits before mcp_http::serve runs, so choose_mcp finds no endpoint, skips the HTTP branch even for a Harness that advertises it, and hands over the stdio shim with an empty environment. That shim exits immediately without AI_BUDDY_MCP_URL.
Under that probe, no Harness receives ai-buddy's MCP. So a run showing Pi without our tools shows nothing about Pi.
The claim may still be true. Pi's row also records MCP transport: none in the session table, which is a different measurement off initialize and may well be the real reason. But the tools row states a delivery outcome, and the run behind it could not observe one.
How this surfaced
The same inference was made about Goose and was wrong. A Goose session reported NO SUCH TOOL for an ai-buddy tool on both a fresh and a resumed session, which read as Goose ignoring the mcpServers field. It was the probe. That conclusion reached a PR description and a review before being retracted, and checking the other rows for the same shape is how this one turned up.
What to change
Either re-measure Pi once the probe can serve an endpoint, per #984, or soften the row to what was actually observed. If the real reason is that Pi advertises no MCP transport at initialize, say that instead, because it is a fact from the handshake rather than an inference from absent tools.
Worth checking every other row in that table for the same shape while here. This is the one that states a delivery outcome outright.
Done (README acceptance met): the pi tools row now states the initialize handshake (http and sse both false) and marks whether Pi lists ai-buddy's tools as unmeasured, pointing at #984. The session-table dagger dropped "gets nothing forwarded today".
Left over (different job):docs/DEVELOPMENT.md still quotes the old README sentence and claims the app forwards nothing. Tracked as #1008.
#984 stays open (probe still serves no MCP endpoint).
Stock probe can serve MCP now. The README pi tools row was re-measured and updated: Pi still advertises no HTTP on initialize, was handed the stdio server on session/new, and never asked for tools/list (pi-acp stores those servers and does not pass them to pi; #1019 covers the .mcp.json Apply path the probe does not write).
No reopen — this issue’s docs acceptance already landed in #1007; #1035 only replaces the earlier “unmeasured” pointer with that stock-probe outcome.
Activity
omesser commentedon Sep 25, 2026
Shipped on main via #1007 (
e1eb97d).Done (README acceptance met): the
pitools row now states theinitializehandshake (httpandsseboth false) and marks whether Pi lists ai-buddy's tools as unmeasured, pointing at #984. The session-table dagger dropped "gets nothing forwarded today".Left over (different job):
docs/DEVELOPMENT.mdstill quotes the old README sentence and claims the app forwards nothing. Tracked as #1008.#984 stays open (probe still serves no MCP endpoint).
— Cursor agent (Coder), on @omesser's behalf.
omesser commentedon Sep 26, 2026
Follow-up after #1035 (
8588715) closed #984.Stock probe can serve MCP now. The README
pitools row was re-measured and updated: Pi still advertises no HTTP oninitialize, was handed the stdio server onsession/new, and never asked fortools/list(pi-acpstores those servers and does not pass them topi; #1019 covers the.mcp.jsonApply path the probe does not write).No reopen — this issue’s docs acceptance already landed in #1007; #1035 only replaces the earlier “unmeasured” pointer with that stock-probe outcome.
— Cursor agent (Coder), on @omesser's behalf.