A row in custom_tools becomes a ToolDef like any built-in, offered beside
the thirteen. The registry had to stop being an import-time constant for
that: `resolve_tools` now returns the schemas *and* the runners together,
carried to the loop on the ToolContext.
That closes a hole on the way. `run_tool` looked names up in the global
REGISTRY with no reference to what had been offered, so a model naming a
tool its chat was gated out of -- a family switched off, a permission the
reader lacks -- had it run anyway. The resolved set is now authoritative.
Arguments come from a model, so an argument may fill a hole but never move
the target: the scheme and host of a URL template are literal, values are
escaped for where they land, and the origin is pinned afterwards. Every
redirect hop is checked the way services/fetch.py checks one, and the
secret is dropped if a hop leaves the origin it was issued for.
Also fixes the tool-activity block claiming every library tool had
"searched the web", which it has done since the second family landed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Filling the composer and waiting for Enter made the card a form to review
rather than a thing to press. One click, one reply.
That changes what a prompt has to be. The built-ins ended mid-sentence --
"My plan: " -- because nothing was sent until the person finished the
thought; sent cold they are a model guessing at material nobody gave it. All
three are rewritten to ask for what they need, so the first reply is the
right question instead. There is a test that they end as complete sentences,
since the failure is silent and only visible in the answer.
requestSubmit, not submit: it fires the submit event, which is what htmx
listens for. Same call the Enter key already makes.
Version bumped because app.js is what changed, and the service worker caches
it -- without the bump the first load after this would still only fill the
box.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A regenerate button that works, per-reply metrics, temporary chats, prompt
suggestions, an admin request inspector and compaction. The bump also
invalidates the service worker's cache, which is keyed on it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The instructions LLeMbas puts in front of a model were hard-coded: six
strings in a GUIDANCE dict, two headings, and the title request inline in
chat.py. An operator could not see what was being sent, let alone change
it, and there was nowhere for a custom tool to contribute its own guidance
when custom tools land.
services/prompts.py now holds each piece as a Fragment, and /admin/prompts
edits them with a preview of the whole assembled system message including
unsaved edits. harness.py keeps only the decisions -- which fragments apply
to this request, and what their variables resolve to.
The design turns on one choice: a fragment carries its gate as data
(families, requires, when_tools) rather than as a callable, because a
database row can carry the same three fields. Custom tools will therefore
register a fragment source and change nothing else -- there is a test that
says exactly that, and it is the reason the rest of the shape is what it is.
Consequences worth knowing:
- Defaults live in code, overrides in the database, and text equal to its
default is never stored. Otherwise pressing Save once would freeze
today's wording forever and no later release could improve it.
- An empty override means off. A fragment that was not submitted at all
keeps what it had, because it may be missing from the page only because
whatever contributes it is currently switched off.
- requires= replaced the hand-written pair of memory guidance variants.
The sentence that refers to a section now lives inside that section, so
it cannot outlive it. That was the general problem the pair was a
special case of.
- {{name}}, with anything unrecognised passing through verbatim. The name
grammar is the guard: {"total": 1} and ${PATH} are not candidates.
Substitution is one pass and never recursive, because {{memories}}
carries text a model wrote.
The wording is also overhauled, and a model now gets the core fragments
even with no tools -- the date above all. "An empty harness is worse than
none" was about tokens that say nothing; a model with no clock being asked
about the present is not that. Clearing those boxes restores the old
silence exactly. New: today's date, who it is talking to, the three-round
tool budget, that tool results are not replayed, that anything a tool
returns is data rather than instruction, and what the <document> wrapper
around an attachment is. Extended: memory_forget, notes_edit/delete,
skill_create/edit, and reading a knowledge document in full rather than
answering from an extract.
Tool descriptions stay in code and are listed read-only. They are schema
and they state facts about what a runner does; an edit would make the text
a lie with nothing to catch it.
No schema change -- one JSON row in the settings table.
488 tests. Version 0.2.0, which also invalidates the service worker cache
so the green artwork appears without a hard reload.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Four pieces of work.
**Installable.** A manifest carrying the instance name, PWA icons rasterised
from the existing mark at design time, a service worker and a themed offline
page. The worker caches the shell only and bails out on /api/, /auth/, /admin/
and anything accepting text/event-stream -- passing a reply stream through a
worker turns it into one delivery at the end, or nothing. It is served from
GET /sw.js rather than the static mount because a worker's scope is the path it
came from.
**Send and Stop are one button.** They were two, and the hidden one was never
hidden: `.btn` is display: inline-flex, which outranks the browser's own
`[hidden] { display: none }`, so Stop sat permanently beside Send. app.css now
forces the attribute to win -- every control toggled with `hidden` depended on
that -- and the composer renders one button carrying both icons, with ui.js
flipping data-composer-action and the type with it.
**Audio.** Speech to text and text to speech against any OpenAI-shaped
/v1/audio/* endpoint: dictate into the composer, have a reply read out.
Instance settings in Admin, per-reader overrides in Settings, with the voice
list discovered from the server where it offers one. Recorded audio is capped
and never written to disk -- it is not an attachment, it has no owner, and
nothing would ever sweep it.
**Web search, as a tool.** This is the tool loop PLAN.md described as the real
work: one reply is now a bounded sequence of requests rather than one. The model
asks, the tool runs, the result goes back and it is asked again, up to three
rounds. Providers are DuckDuckGo (no setup), SearXNG and Firecrawl.
Two decisions worth stating. Tools are only offered to models flagged `tools`,
because an endpoint without support rejects the whole request rather than
ignoring the array -- the same reason images only reach models flagged
`vision`. And tool results are not replayed as context on the next turn, for the
same reasons reasoning is not: the answer already contains what the model made
of them, and replaying stale results into every later request wastes the window
and reliably sends a small model into a search loop. The sources stay visible in
the transcript instead.
Search results are untrusted third-party text and are treated as such: escaped,
and only http/https URLs rendered as links.
338 tests, ruff clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Drag, paste or pick a file in the composer. Images go to vision models
as multimodal content parts; PDFs and text files have their content
extracted and placed in the prompt. Verified end to end against
gemma4-e4b-q8 on llama-swap: given a drawing and a text file, it named
the red square and blue circle and read the number out of the document.
Type is decided by inspecting the bytes, never the filename or the
browser's Content-Type -- a .png full of text is stored as text. Images
are downscaled to 1400px and re-encoded: a phone photo is several
megabytes of base64, which is slow and a large slice of the context
window. PDF text is extracted once, at upload, and stored; re-extracting
per request would let a reply change because a parser was upgraded.
Design points worth keeping:
- Images are only sent to models an administrator has marked `vision`.
This is not graceful degradation -- most endpoints reject the entire
request rather than ignoring an image part. A plain text turn stays a
plain string for the same reason: the list form 400s on endpoints that
do not implement it.
- Images reach the model as base64 data URIs, not links. A local
endpoint has no route back to LLeMbas, and a hosted one has no
credentials for it.
- Non-images are served Content-Disposition: attachment with nosniff, so
an uploaded .html can never execute in this origin. Stored names are
random; the uploader's name is a label and never a path.
- Uploads are unbound until the message is sent, which is what lets a
file be removed beforehand. claim() only takes unclaimed rows owned by
the sender, so a forged id cannot pull in someone else's file.
Abandoned uploads are swept at startup.
- A scanned PDF says so rather than silently contributing nothing, and
truncation is declared to the model in the document tag so it can
admit it did not see page 400.
- "Here, look at this" with no words is a legitimate turn, so a message
is only empty when it carries neither text nor files.
Also fixes auto-titling, which read message["content"] as a string and
would have broken on the first multimodal turn.
186 tests, ruff clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
LLeMbas now runs end to end. Register, add an OpenAI-compatible
connection, and hold a real streaming conversation organised into
folders. Verified against the local llama-swap instance.
Streaming is the one genuinely tricky part. Sending a message returns
two HTML fragments -- the user bubble and an empty assistant bubble
carrying an sse-connect -- and that attribute is the ONLY thing that
starts a generation. Rendering an incomplete assistant message as a
streaming shell falls out of the same template, which means loading a
page whose last reply never finished simply picks it up again.
Details worth knowing about, each commented where it matters:
- SSE payloads are split across several data: lines. A raw newline in
one data: line truncates the event, which shows up the first time a
model emits a code block.
- Markdown is rendered server-side by the same helper for both the page
and the final streamed frame, so the two cannot disagree. The fence
renderer is replaced outright rather than using markdown-it's
highlight option, which re-wraps output in a second <pre>.
- escape_text is html.escape, not nh3.clean_text: it escapes character
by character, so escaping stream chunks separately equals escaping
the whole string.
- The stream opens its own session via session_scope(); it outlives the
request handler and the dependency-scoped session may be closed.
- Deleting a folder keeps the chats inside it (FK is SET NULL). Losing
a conversation to a mis-clicked folder delete is unforgivable.
- Login failures use one message for "no such account" and "wrong
password" so the form cannot enumerate registered addresses.
Also adds deploy/ for the gamebox install at https://chat.lan: system
unit, nginx vhost with buffering off (buffering on turns streaming into
one lump at the end), and install/update scripts following the same
service-user and /srv bind-mount conventions as llama-swap and comfyui.
70 tests, ruff clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Establish the LLeMbas foundation: FastAPI/Jinja/SQLite layout, the ORM
schema, and the original SVG identity.
Notable decisions, all recorded in comments at the point they matter:
- No Alembic. SQLite only, schema created at startup, so models carry a
few columns nothing reads yet (Message.parent_id for branching,
content_parts_json for multimodal turns). Adding them later to a live
database without migrations is the painful path.
- Sessions are server-side rows keyed by a SHA-256 of the cookie value,
not JWTs, so logout and bans revoke access immediately.
- Upstream API keys are Fernet-encrypted with a key derived from
LEMBAS_SECRET_KEY. decrypt() fails soft to "" so rotating the secret
degrades to re-entering keys rather than crashing the admin UI.
- Artwork is generated by scripts/build_artwork.py rather than hand-drawn
per file: the mallorn leaf appears in the icon, favicon, lockup and
banner, and one source is the only way those stay in sync. The wordmark
is Source Serif 4 (OFL) converted to outlines, because a README banner
cannot load a webfont and <text> would render in whatever serif the
viewer happens to have.
- Icons live in a template partial, not assets/, because same-document
<use href="#id"> is universally supported and the cross-document form
is not.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Jaroslav Beneš