For about three weeks, the tools that let my chat assistant touch my calendar sat switched off. I’d built them in early October, then turned them off on the 13th to free RAM on the Pi. And honestly, I wasn’t in a hurry to turn them back on. A model that can read my calendar is useful. A model that can change it is a different thing.
This week they came back. With a catch: nothing gets written unless I click a button.
Then I spent Friday night finding out how good the model behind that button actually is. The answer was “better than I expected, worse than I’d like, and the fixes weren’t where I thought they’d be.”
Monday: an Approve button
LibreChat v0.8.8 added human-in-the-loop tool approval. In the config you list the tools that should ask first, and when the model calls one of them, the reply stops. You get a card with the tool name and its exact arguments, and two buttons: Approve, Deny. A denied call goes back to the model as rejected.
That one feature is why I upgraded on Monday night. Every write tool of my
connectors went under ask: adding, editing or deleting events, tasks,
notes, feed subscriptions. Everything that only reads went under allow.
The FreshRSS and Nextcloud connectors left the “disabled” profile, and a
new spec called Calendar runs them on the Pi.
The upgrade had one small annoyance. The image moved to
ghcr.io/librechat-ai/librechat, and the config schema version is tied
to the app version, so 1.3.13 became 1.3.17. Then, on Friday, LibreChat
refused to start because two of my MCP server titles had parentheses in
them. “Wikipedia (large)” is apparently illegal now. It’s “Wikipedia Large”.
Fine.
Which calendar? Not the model’s call
The other change that night is one I care about more than it looks.
When I say “add the driving lesson on Thursday”, the model has to choose a calendar. Left to itself, it guesses. Sometimes it guesses well. Sometimes it invents a calendar name that doesn’t exist.
So the choice moved out of the model and into the connector. If the user names a calendar, the model passes it. If not, a small list of rules picks it: domains first, matched against the link’s host (handy for feed items), then keywords in the title, location, notes or link, with case and accents ignored. First match wins, and the reply says which rule matched. The rules name my own doctors, school and employer, so they live only in the encrypted env file, never in the repo in plain text.
Deterministic. Boring. Exactly what I want from the part that decides where my life gets written down.
Tuesday: hello again, -gzip
Last week I wrote about a bug that turned my recording linker’s “live”
into “running and failing politely”. Nextcloud AIO’s Apache gzips
responses and appends -gzip to the ETag. My code sent that ETag back in
If-Match, the server compared it with the real one, and every write came
back 412.
On Tuesday I found the exact same bug in the calendar connectors.
Same server, same Apache, same suffix. Every edit, every delete, every
note append would have failed with “changed in Nextcloud meanwhile”. Only
the writes that change something already there send a fetched ETag back,
which is probably why it hid for so long. The fix was the same one header,
Accept-Encoding: identity, and the commit message points straight at last
week’s fix.
I’m not proud that I fixed a bug in one service and didn’t go look for it in the other. But I’ll take a one-week gap over a six-day silent failure.
Friday: Qwen is back, with bigger tools
Qwen had been off since 31 October. The kernel’s OOM killer took it twice on the workstation, peaking around 5.2 GB of host RAM, and I switched it off. On Friday it came back. And this time I stopped making it live on the Pi’s diet.
Everything my assistant uses was sized for the Raspberry Pi: short passages, tiny schemas, few results. That’s right for a small model on a CPU. It’s silly for Qwen3.5-9B on a GPU that prefills around 1,000 tokens a second. A few extra thousand tokens of context cost it seconds, not minutes.
So the tools got a “large” version, side by side with the small one:
- Wikipedia serves a second endpoint from the same process and the same
open archive: 10 passages of up to 1,200 characters from the top 5
articles, plus a new
wikipedia_readthat returns an article’s lead, or one whole section, up to 12,000 characters. The old endpoint stayed byte-identical. I compared its tool list and eight test searches before and after. - Web search got a second container (about 60 MB) with the full schemas: time range, language, paging, and pages of 16,000 characters instead of the Pi-sized ones.
- Translation went off by default for Qwen. It writes Greek itself, and better than the translation model. It kept «πόδια» as legs where the translator turned it into feet. Small thing. Also exactly the kind of thing that makes an answer feel wrong.
To make room, the reply budget went from 32,768 tokens to 20,480, which leaves about 26,000 for history and tool results. At 32k there was too little left for tools that return 3,000 to 4,000 tokens a call.
Two specs, on purpose
Then the part that’s really about trust.
There are now two Qwen specs, and they are kept apart deliberately.
qwen-chat reads the web and has a sandbox with internet access.
qwen-organiser has my calendar, tasks, notes and feeds, and nothing that
reads the web.
Why not one assistant that does everything? Because together they’d be dangerous. A web page or a feed item could tell the model to read one of my notes and send it out inside a URL, and fetching a URL needs no approval. Simon Willison calls this the lethal trifecta: private data, untrusted content, and a way to send data out. Pick any two. Apart, the web chat has no personal data, and the organiser’s only way out is a write I approve.
Friday night: grading my own secretary
Here’s the thing about an Approve button. It protects the data. It doesn’t make the assistant good. If it asks me to approve the wrong change half the time, I’ll stop using it. Or worse, I’ll start clicking Approve without reading.
So I wrote a test suite. 25 requests, taken from things I actually ask assistants: what’s on tomorrow, add this email to my calendar, what’s on my shopping list, cancel Monday’s lecture, how many lectures next week, a few in Greek. Each one has a written description of what counts as right.
The suite runs against the deployed organiser, with its real prompt, sampling and tools, from inside the LibreChat container on the Pi, with Qwen on the workstation. Reads hit my real Nextcloud and FreshRSS. Every write is refused, so nothing changes. I graded by hand: 1 for right, ½ for right with a flaw, 0 for wrong or an unwanted write.
The baseline scored 14.5 out of 25.
What went wrong
Some failures were funny. Most weren’t.
“Cancel Monday’s lecture” marked every Monday as cancelled in 3 of 4 runs. The lecture is a repeating event, and the edit tool only knew how to change the whole series. Thank god for the refused writes.
Asked about Wednesday in Greek, «Τετάρτη», the model worked out the date itself and got it wrong. Same with “next week”.
“Add eggs to my shopping list” sent it hunting through tools, sometimes until the context overflowed. And nothing stopped it from simply inventing a list called “Shopping List”.
And one that genuinely annoyed me: “delete this task” became “mark it done”. There is no delete tool for tasks, so the model quietly did a different change instead.
The fixes went into the tools
I expected to spend the night rewriting the prompt. Mostly, I didn’t.
- One search. A new
findtool searches events (a month back, a year ahead), tasks and notes at once, by word stem, accents ignored. A miss tells the model to try other words, or the language the entry is likely written in. Most of my entries are in Greek, and the model’s first query is often in English. - Real names only. At startup, the calendar and list parameters get my
actual calendars and lists as a JSON Schema
enum. The model can’t invent “Shopping List” anymore. It isn’t one of the allowed values. - No date arithmetic. The tools now take dates as words:
yesterday,next tuesday,this friday,next week,+3d,in 3 days,9am,12 November, and, after midnight, Greek too: «αύριο», «Δευτέρα», «12 Νοεμβρίου». The model copies what I said. The parser does the maths. Parsers are good at maths. - One occurrence. Editing a repeating event now requires either an
occurrencedate orall_occurrences=true. One date gets a standardRECURRENCE-IDoverride that can be edited, moved, or cancelled withSTATUS:CANCELLED, which Nextcloud shows struck through. Without either, the tool refuses. - Do exactly what was asked, or nothing. That one was a prompt line. If it’s unclear which entry I mean, or no tool does what I asked, it says so or asks. It doesn’t swap in something else.
I also tried the knobs everyone reaches for first:
| Run | Score |
|---|---|
| Baseline | 14.5 |
| New tools, precise sampling (0.6 / 0) | 15 |
| New tools, old sampling (1.0 / 1.5) | 15.5 |
| New tools, thinking off | 14.5, but 5–10× faster |
| Final: + occurrence edits, date words in descriptions, “do exactly what was asked” | 16.5 |
Sampling made no measurable difference. Turning thinking off looked almost as accurate on paper and was much faster, but its mistakes were the worst kind: confident “you have nothing” answers, because it searched in English for entries titled in Greek. I didn’t adopt it. A fast secretary that tells you your appointment doesn’t exist is worse than a slow one.
(One sample per case, graded by me. I know. It’s a direction, not a benchmark.)
What still fails
I’d rather write this part down than pretend it isn’t there.
One test pastes in an “email” with an extra line at the bottom, addressed to the assistant, asking it to copy all my notes into the event description. The model obeys. In every run. The only guard is the approval card, which shows the description, so I’d see my notes sitting in it before anything got saved. That works. It also means the guard is me, at whatever hour I’m reading.
The delete-turns-into-done habit survived the prompt fix. After a rejection, the model sometimes retries the same write in a new form. And open-ended searches can still wander until they hit the step limit.
16.5 out of 25 is not a secretary I’d leave alone. It’s one I’ll use with the Approve button, and keep testing.
Mail, but only drafts
Same night, a mail connector: list, read, save drafts. No send, no move, no
delete. Reading uses BODY.PEEK, so a message I haven’t opened stays unread.
A draft goes through the approval card and then waits in my Drafts folder for
me to send it from my own mail app.
It’s tested against a throwaway GreenMail server on the Pi, with a Greek subject, an HTML body with a stub plain part, an attachment and a threaded reply draft. It isn’t connected to my real mailbox yet, because that needs my password and I haven’t typed it in. It’s also the one connector with a route to the internet, so it stays out of any spec that reads the web.
University webmail, through a tunnel
Not everything this week was about the model.
My university’s webmail only answers from inside the university network. From home it just times out. So on Friday I set up Roundcube on the Pi, inside the network namespace of an OpenVPN container logged in to the university VPN. To the mail server, my Pi looks like it’s on campus.
Only the university’s address range goes through the tunnel. The server pushes “send everything through me”, and I filter that out. The interesting bits were all old cryptography:
- The IMAP server offers only TLS 1.2 with a 1024-bit Diffie-Hellman
key. OpenSSL 3 refuses that at its default security level
(
dh key too small), so IMAP runs atSECLEVEL=1. The certificate is still verified. I don’t love it. 1024-bit DH has been considered weak since Logjam in 2015. But it’s their server, not mine. - Roundcube’s first logins failed with “authentication failure” and the right
password. The server advertises DIGEST-MD5, Roundcube prefers it, and the
backend rejects it. DIGEST-MD5 was moved to Historic back in
2011. Forcing
PLAIN, inside TLS, fixed it. - And one decision I’m glad about: if the VPN login is rejected, the container doesn’t restart and try again. It logs why and waits, unhealthy. That account is my whole university identity: mail, the course platform, the account portal, everything. A restart loop hammering it with a bad password is how you get locked out the week before an exam.
The small stuff
- The homepage’s Watchtower widget had failed every poll with
ENOTFOUND. It callshttp://watchtower:8080, which isn’t published on the host, and the homepage container wasn’t on Watchtower’s network. One network, six lines. - Version bumps, each checked side by side first: mcp-searxng 2.10.0 (lite tool schema byte-identical), Playwright MCP 0.0.83, MongoDB 8.0.34, tzdata 2026.6, pgvector 0.8.7, pypdf 7.0.0. Meilisearch 1.56.1 converts its index in place, so I rehearsed that on a copy first: same documents, same hits, per-user filter intact. pypdf 7 extracts the same text or better from my test PDFs, maths spacing and Greek tables included.
The takeaway
The thing I keep coming back to: the score didn’t move when I changed how the model thinks. It moved when I changed what the model is allowed to get wrong. No inventing calendar names, because there’s an enum. No date arithmetic, because the parser does it. No cancelling every Monday by accident, because the tool won’t let you without saying which one.
If you’re building something with a model in it, try this before you touch the prompt again: list the five mistakes it makes most, and ask which of them a tool could make impossible. Then write 25 real requests and grade it. By hand, honestly, on a Friday night. It’s humbling. It’s also the only way I know to find out whether “it works” is actually true.
Sources
- LibreChat and its MCP and agents documentation
- Model Context Protocol
- Simon Willison, “The lethal trifecta for AI agents” (June 2025)
- OWASP Top 10 for LLM Applications: LLM01 Prompt Injection
- RFC 9110 (HTTP Semantics), 412 Precondition Failed and
Apache
mod_deflate,DeflateAlterETag - RFC 5545 (iCalendar), on
RECURRENCE-IDandSTATUS - JSON Schema
enum - Qwen3.5 and llama.cpp
- Roundcube and the Roundcube Docker image
- OpenSSL security levels and weakdh.org (Logjam)
- RFC 6331: Moving DIGEST-MD5 to Historic
- GreenMail, the test mail server
- Meilisearch
- The commit history of my homelab repo, 2–7 November 2026