Repository navigation
v0.8.8
[0.8.8] - 2026-10-02
Fixed
-
quickstartrecommended agitserver twice, and on a machine with nogit.
Measured on a clean Windows box (freshUSERPROFILE,uvxpresent, nogit, cwd a
git repo): the discovery table said "Discovered 3" but listed four rows, and the
gitentry was offered even though the binary is absent — so the spawned server
died with[PROCESS_DIED] … All git commands will error until this is rectified
while the summary still counted it (🎉 3 tools ready across 3 servers!). Two
causes:githad been moved into the zero-configuvxtable without deleting the
older "we are in a git repository" special case, and the gate checked only for
uvx. The row now carries the binary it needs (requires), is offered once, and
is skipped when that binary is missing. The old test hid the duplicate because it
matched on a phrase the duplicate's reason did not contain. -
The skill and the agent configs were written for agents that are not installed.
quickstartcreated six agent skill folders (.claude,.agents,.codex,
.cursor,.catpaw,.codeium/windsurf) on a box with none of the six agents and
reported "installed for 6 agent(s)";synclikewise wrote a config file for every
host it knew about — including%APPDATA%\Code\User\settings.jsonon a machine with
no VS Code. Both now write only where the agent is actually present (theexists
that detection already computed is now acted on, not merely stored), the wording is
"agent folder(s)", and a machine with no known folder says so and points at
mcptoon skills install-self --view <folder>for an agent mcptoon does not know
about. Cline's detection was the VS Code User directory — true on any box with
VS Code — and is now Cline's ownsaoudrizwan.claude-devfolder.Closing the hole this would otherwise leave: the first-run self-heal marker is
written once per release, so an agent installed after mcptoon would never receive
the skill. The heal now re-checks, on every command, for an existing agent view that
lacks our file, so the copy lands as soon as the agent's folder appears. -
A test asserted the installed skill path equals the source-tree one. True only
underPYTHONPATH=src; false under Nix, where the packaged copy lives in
/nix/storeand the source tree in the build directory. It was the reason
numtide/llm-agents.nix'snix-buildfailed on all three platforms for every bump
from 0.8.4 on (1 failed, 1739 passed, 12 skipped). -
A server that never started was counted, celebrated, and called healthy.
Reproduced 2026-10-02 on a simulated clean box with one working server and one whose
binary does not exist:quickstartended with🎉 11 tools ready across 2 servers!— one of those two
servers was dead. The tool count was already honest (error entries are skipped);
the server count was not. The line now names the gap:
11 tools ready across 1 server (1 could not start)!.mcptoon doctorprinted! dead [stdio] 0 tools (server may be empty), counted
0 issues, and finished with "All good! ✓". It now reports
✗ dead [stdio] cannot start: …and counts it as an issue.- Root cause:
manifest.get_server_toolsswallowed a failed start into[]— the
same value a server that answered with no tools returns — so callers could not
tell "could not start" from "started, exposes nothing". It now returns an
{"error": …}entry (the shapeget_manifestalready used), anddiscover --healthreports that aserrorinstead ofno-tools.
-
Re-running
quickstartdenied the servers it manages. Discovery reads other
agents' configs (Claude Desktop, Cursor, env vars, …), never mcptoon's own, so a
secondquickstarton a machine that already had servers printed "No MCP servers
found on this machine." whilemcptoon listprinted them. It now says
"Nothing new to discover — you already have N server(s) configured" and points at
mcptoon list/status/sync --self. -
A server list that lives in
config.tomlwas silently deleted.load_config()
reads bothconfig.tomlandconfig.json, but three callers gated on
CONFIG_FILE.exists()— the JSON path alone. On a machine whose servers live in
TOML that madequickstartandinit --autotake the "no config yet" branch and
callsave_config(), which sees the TOML file and rewrites it with only the
newly discovered servers: the user's whole list was gone, with no.bak
(reproduced 2026-10-02). All three now use a newconfig.has_config_file(), which
checks either format.doctorwas wrong the other way and now names whichever file
it actually found instead of reporting "No config found. Run: mcptoon quickstart" —
advice that, before this fix, would have deleted the config it was diagnosing. -
A failed write was reported as a successful removal.
mcptoon offfiltered its
results onremoved, which means "the gateway entry was found", not "it is gone" —
a write that failed still reportsremoved=True, written=False. Sooffprinted
"Done: 1 agent(s) no longer see mcptoon" while the file still held the entry
(reproduced 2026-10-02). It now requires both, names the agent whose write failed
("the gateway entry is STILL there"), and--jsongains afailedlist. The same
filter bug inmcptoon restoremade--jsonand the text form print two
different counts for one run; both now countrestored and written. -
mcptoon healthsaid "All N servers healthy" while listing ano-configrow.
A row whose server is not in the config is neither alive nor dead; it fell into
neither counter, and the footer keyed offerror_count == 0. A report reading
0/1 alive, 0 deadtherefore ended "All 1 servers healthy." (reproduced
2026-10-02). Unknown is now its own state —1 unknown, and a footer that says
0/1 servers healthy; 1 not in the config. -
serve/health called an alive-but-empty serverstatus: "error".failed
was inferred from "absent from_tool_index", which is also true of a server that
connects and legitimately exposes zero tools — so a single such server made the
endpoint reporterrorwithfailed_servers == 1(reproduced 2026-10-02). Failure
is now tracked where the fetch actually raises, the payload carries the per-server
errors, and an empty-but-alive server isok. -
An unreachable HTTP endpoint was counted as a discovered server. Two halves:
probe_http_endpointreturned{"alive": True, "tools_count": 0}for any
exception, so an endpoint that answeredinitializebut could not list tools was
indistinguishable from one that answered with zero tools (nowtools_count: None,
tools_known: False); anddiscover --http <dead-url>wrote the entry into
result.countand--writemerges anyway, leaving only the prose "(not
responding)" as a signal. A dead URL is now reported and not counted. -
mcptoon uninstallcounted things it had not removed. It summedremoved
(found, not written) for agent configs, and its unlink loop swallowedOSError, so
a file that could not be deleted was still counted in "Deleted … N file(s)". Both
now count only what actually happened, and failures are printed. The installer's
own registry (~/.mcp-cli/pro_installed.json) was missing fromstate_paths(), so
a full uninstall left it behind; it is now bookkeeping, and its path is
env-overridable like every other state file. -
The
install.shsyntax test failed on a Windows box that merely lacks WSL. It
ranbash -n install.shand skipped only on exit code 127. On Windowsbashon
PATH is usually%SystemRoot%\System32\bash.exe, the WSL launcher stub — with no
distro installed it does not exit 127, it blocks, so the run died with
subprocess.TimeoutExpiredinstead of skipping and the suite went red. The test
now recognizes the stub and treats a 30-second hang as "bash unusable", not "the
script is broken".