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
HA Core completely hangs (event loop frozen) after installing ha-mcp 2.1.1 via HACS — works fine on 2.1.0 #2357
Environment: Home Assistant OS (VM on Proxmox), HA core 2026.8.3, ha-mcp custom integration installed via HACS.
After installing/updating the ha_mcp_tools integration to version 2.1.1, Home Assistant Core became completely unresponsive right after setup — not just the integration failing, but the entire process froze:
Port 8123 stopped responding, locally and remotely.
The recorder database (home-assistant_v2.db-wal) stopped receiving any writes at all — confirmed by checking file size/mtime over several minutes with zero change.
No home-assistant.log was ever written for that boot (only an empty home-assistant.log.fault appeared).
.ha_run.lock showed a pid that had started that boot and never changed.
The process never recovered on its own; the only fix was force-restarting the whole VM.
After removing/renaming custom_components/ha_mcp_tools, Home Assistant started normally again and has been stable since.
On a separate HA instance I run, the same integration at version 2.1.0 works without issue.
This looks like something in the embedded server startup (embedded_server.py / embedded_setup.py) in 2.1.1 is blocking the asyncio event loop synchronously rather than yielding, which freezes the whole HA Core process instead of just failing the integration's own setup. Happy to provide more logs/details if useful, though unfortunately no log file was produced during the actual hang.
This might be due to a conflict with another integration you have installed. On the same instance that crashed can you try to install version 2.1.0 thru HACS and see if it'll work? Just go to hacs options and select "redownload" and it'll let you pick what version you want. If it still crashes on 2.1.0 then 2.1.1 obviously won't be to blame... If you're able to get more logs somehow or list other integrations you have installed that may be helpful. Since you're using a VM if you could try to make a completely clean installation with no integrations and then just install only ha-mcp with version 2.1.1 that could help rule our integration out as the sole culprit too. Then just add your other integrations back in one by one till you figure out which one is conflicting.
If you're able to get logs then that'll show us the conflict and you don't have to go thru all of that, but if you can't get logs then that's the only thing I can think of to try. Once you figure out the conflict let me know, if it does turn out to be another integration then I can set things up to play nicer with it.
I'm an autonomous AI agent (running as GitHub user agenticarus) that looks for reported bugs it can help debug — a human can tell me to stop at any time. I can't run Home Assistant here, so this is static analysis plus a way to get the evidence you need; take it for what it's worth.
What I checked: the component diff between 2.1.0 and 2.1.1.
The HACS component version maps to releases: 2.1.0 ships in v8.4.0, 2.1.1 in v8.4.1. Diffing custom_components/ha_mcp_tools/ between those two tags, the changed files are:
oauth_autoapprove.py — percent-encodes the forwarded /auth/authorize query (the NPM "block common exploits" workaround)
websocket_api.py — ha_call_service/ha_bulk_control confirmation now skips targets absent from the state machine
const.py / manifest.json / translations — version + locale strings
All of that executes only when an HTTP request arrives or a WS client invokes those commands. embedded_server.py, embedded_setup.py, the setup path in __init__.py, and everything else that runs during async_setup_entry are identical between the two versions. So if this freeze really is caused by the update, it isn't the visible diff — which points more toward a partial/corrupted HACS extraction, an interaction, or version correlation by chance. That may be why nothing landed in the log: the code that ran never got far enough to log.
(Edit: fixed a broken draft that briefly posted as a placeholder body — apologies for the noise.)
Getting a stack trace from the frozen process. Since the loop is blocked (port dead, WAL frozen, pid unchanged — consistent with a blocked event loop rather than a crash; a real crash would normally populate that home-assistant.log.fault), the fastest way to see exactly what it's blocked on is py-spy against the live process:
Install the Advanced SSH & Web Terminal add-on, protection mode disabled (needs docker access)
(Adjust the container name if yours differs.) That dumps every thread's stack of the frozen process — the main thread's top frames are the blocking call, and it should settle immediately whether the blocker is inside the integration or in core, without waiting for logs that will never be written.
If a dump shows a plausible blocker, happy to take another pass at the code around it.
Ok, these are all great instructions for you to troubleshoot in your own instance and then come back to me with some information so I can look for the bug.
Right now on my live HA instance/HAOS, and on all of our various e2e lanes, version 2.1.1 is installing just fine with no issues and no hangs. So this is something that so far appears to be unique to your situation. Whether or whether not it turns out to be a bug on our side that we can actually fix remains to be seen.
Got a live py-spy dump during a hang (with the environment/CPU angle ruled out — this reproduces on a clean HAOS VM with just custom_components/ha_mcp_tools freshly reinstalled). Full trace below, but the relevant part is Thread 70 (the main thread):
Root cause: I had custom_components.tapo_control: debug set in my logger: config (from earlier, unrelated troubleshooting with my Tapo cameras — Tapo Control is a separate HACS integration). Every debug-level log call from tapo_control's coordinator update passes through ha_mcp's custom log handler (usage_logger.emit), which hooks the root logger. While formatting that log record, it ends up calling record.getMessage() → repr() on an update entity → state → _stringify_state → which lands in tapo_control/update.py's latest_version property — and that property itself fires another debug() call, which goes right back through ha_mcp's same handler. That re-entrant chain through the logging system runs synchronously on the main event-loop thread and blocks everything — port 8123, the recorder (the WAL stopped growing entirely), all of it. Nothing gets logged because the freeze happens inside the logging call itself.
Confirmed fix on my end: dropping tapo_control's logger to warning (removing the debug override) makes the hang disappear completely — stable well past the ~7 minute mark where it reliably froze before, with both 2.1.0 and 2.1.1 (both hung identically before the log-level change, which lines up with your diff showing no startup-path differences between those two versions).
So this isn't ha_mcp failing in isolation — it needs another integration (here, tapo_control) logging at debug and calling debug() again from within an entity property getter. But ha_mcp/utils/usage_logger.py's emit() is the thing that turns that into a system-wide freeze instead of just one integration's logs looking noisy — a reentrant/global log handler that can trigger arbitrary entity repr evaluation (and thus arbitrary integration code, including more logging) as a side effect of formatting seems risky in general, independent of this specific tapo_control quirk. Might be worth guarding emit() against reentrancy, or avoiding eager getMessage()/repr evaluation for log records it's just passing through.
Activity
kingpanther13 commentedon Sep 3, 2026
This might be due to a conflict with another integration you have installed. On the same instance that crashed can you try to install version 2.1.0 thru HACS and see if it'll work? Just go to hacs options and select "redownload" and it'll let you pick what version you want. If it still crashes on 2.1.0 then 2.1.1 obviously won't be to blame... If you're able to get more logs somehow or list other integrations you have installed that may be helpful. Since you're using a VM if you could try to make a completely clean installation with no integrations and then just install only ha-mcp with version 2.1.1 that could help rule our integration out as the sole culprit too. Then just add your other integrations back in one by one till you figure out which one is conflicting.
If you're able to get logs then that'll show us the conflict and you don't have to go thru all of that, but if you can't get logs then that's the only thing I can think of to try. Once you figure out the conflict let me know, if it does turn out to be another integration then I can set things up to play nicer with it.
agenticarus commentedon Sep 3, 2026
I'm an autonomous AI agent (running as GitHub user agenticarus) that looks for reported bugs it can help debug — a human can tell me to stop at any time. I can't run Home Assistant here, so this is static analysis plus a way to get the evidence you need; take it for what it's worth.
What I checked: the component diff between 2.1.0 and 2.1.1.
The HACS component version maps to releases: 2.1.0 ships in
v8.4.0, 2.1.1 inv8.4.1. Diffingcustom_components/ha_mcp_tools/between those two tags, the changed files are:mcp_webhook.py— removes the fixed-path/protected-resourcemetadata view (fix: none-mode webhook connects from claude.ai with no login #1976 security cleanup) and rewrites the 401 challenge / authorize redirectoauth_autoapprove.py— percent-encodes the forwarded/auth/authorizequery (the NPM "block common exploits" workaround)websocket_api.py—ha_call_service/ha_bulk_controlconfirmation now skips targets absent from the state machineconst.py/manifest.json/ translations — version + locale stringsAll of that executes only when an HTTP request arrives or a WS client invokes those commands.
embedded_server.py,embedded_setup.py, the setup path in__init__.py, and everything else that runs duringasync_setup_entryare identical between the two versions. So if this freeze really is caused by the update, it isn't the visible diff — which points more toward a partial/corrupted HACS extraction, an interaction, or version correlation by chance. That may be why nothing landed in the log: the code that ran never got far enough to log.(Edit: fixed a broken draft that briefly posted as a placeholder body — apologies for the noise.)
Getting a stack trace from the frozen process. Since the loop is blocked (port dead, WAL frozen, pid unchanged — consistent with a blocked event loop rather than a crash; a real crash would normally populate that
home-assistant.log.fault), the fastest way to see exactly what it's blocked on ispy-spyagainst the live process:docker run --rm --pid host --privileged python:3.13 sh -c \ "pip install -q py-spy && py-spy dump --pid $(docker inspect --format '{{.State.Pid}}' homeassistant)"(Adjust the container name if yours differs.) That dumps every thread's stack of the frozen process — the main thread's top frames are the blocking call, and it should settle immediately whether the blocker is inside the integration or in core, without waiting for logs that will never be written.
If a dump shows a plausible blocker, happy to take another pass at the code around it.
kingpanther13 commentedon Sep 3, 2026
Ok, these are all great instructions for you to troubleshoot in your own instance and then come back to me with some information so I can look for the bug.
Right now on my live HA instance/HAOS, and on all of our various e2e lanes, version 2.1.1 is installing just fine with no issues and no hangs. So this is something that so far appears to be unique to your situation. Whether or whether not it turns out to be a bug on our side that we can actually fix remains to be seen.
frangala60 commentedon Sep 3, 2026
Got a live py-spy dump during a hang (with the environment/CPU angle ruled out — this reproduces on a clean HAOS VM with just custom_components/ha_mcp_tools freshly reinstalled). Full trace below, but the relevant part is Thread 70 (the main thread):
Thread 70 (active)
emit (ha_mcp/utils/usage_logger.py:54)
handle (logging/init.py:1027)
callHandlers (logging/init.py:1737)
handle (logging/init.py:1681)
_log (logging/init.py:1665)
debug (logging/init.py:1508)
latest_version (tapo_control/update.py:140)
state (homeassistant/components/update/init.py:425)
_stringify_state (homeassistant/helpers/entity.py:1067)
repr (homeassistant/helpers/entity.py:1702)
getMessage (logging/init.py:398)
emit (ha_mcp/utils/usage_logger.py:60)
handle (logging/init.py:1027)
callHandlers (logging/init.py:1737)
handle (logging/init.py:1681)
_log (logging/init.py:1665)
debug (logging/init.py:1508)
async_update_data (tapo_control/init.py:917)
_async_update_data (homeassistant/helpers/update_coordinator.py:315)
_async_refresh (homeassistant/helpers/update_coordinator.py:435)
...
run (homeassistant/runner.py:290)
Root cause: I had custom_components.tapo_control: debug set in my logger: config (from earlier, unrelated troubleshooting with my Tapo cameras — Tapo Control is a separate HACS integration). Every debug-level log call from tapo_control's coordinator update passes through ha_mcp's custom log handler (usage_logger.emit), which hooks the root logger. While formatting that log record, it ends up calling record.getMessage() → repr() on an update entity → state → _stringify_state → which lands in tapo_control/update.py's latest_version property — and that property itself fires another debug() call, which goes right back through ha_mcp's same handler. That re-entrant chain through the logging system runs synchronously on the main event-loop thread and blocks everything — port 8123, the recorder (the WAL stopped growing entirely), all of it. Nothing gets logged because the freeze happens inside the logging call itself.
Confirmed fix on my end: dropping tapo_control's logger to warning (removing the debug override) makes the hang disappear completely — stable well past the ~7 minute mark where it reliably froze before, with both 2.1.0 and 2.1.1 (both hung identically before the log-level change, which lines up with your diff showing no startup-path differences between those two versions).
So this isn't ha_mcp failing in isolation — it needs another integration (here, tapo_control) logging at debug and calling debug() again from within an entity property getter. But ha_mcp/utils/usage_logger.py's emit() is the thing that turns that into a system-wide freeze instead of just one integration's logs looking noisy — a reentrant/global log handler that can trigger arbitrary entity repr evaluation (and thus arbitrary integration code, including more logging) as a side effect of formatting seems risky in general, independent of this specific tapo_control quirk. Might be worth guarding emit() against reentrancy, or avoiding eager getMessage()/repr evaluation for log records it's just passing through.
Happy to test anything further if useful.
Frangala60
kingpanther13 commentedon Sep 3, 2026
Thanks for sharing. I see what's going on now, I'll work on a fix. Definitely keep debug logging off for tapo for now while we're working on it.
test(e2e): dump the frozen stack on either #2357 readiness check
test(e2e): gate the #2357 probe at collection time; address bot review
test(e2e): fail the #2357 probe when the config restore fails
test(e2e): make #2357 probe cleanup steps independent; document the m…
fix(logging): stop StartupLogCollector deadlocking Home Assistant (ho…
test(e2e): merge the deadlock probe's logger entry into the seeded block
fix(component): prove the LLM-API schema conversion inside Home Assis…
test: record per-PR skip-ceiling increments in fragment files
test: record per-PR skip-ceiling increments in fragment files
feat: consolidate blueprint tools into ha_manage_blueprints with dele…