Files
LLeMbas/src/lembas/services/library/skills.py
T
Jaroslav Beneš 1eba860d39 Knowledge, notes, memory and skills, and a harness to make them used
Four places a model can reach for, differing in who writes a record and how it
gets in front of the model.

**Knowledge** is uploaded by a person and searched by the model. It goes through
`services/files.py:prepare` — the same pipeline as a chat attachment — so the
same PDF produces the same text whichever way it arrived, and `Document` carries
the same content columns as `Attachment` for the same reason.

**Notes** are written by the model and edited by you. Too long to inject, so
they are searched.

**Memory** is short facts, and every one of them goes into every request. That
single decision is where the rest of its design comes from: records are capped
short, the block has a budget, there is no search tool because the model is
already looking at them, and they are not shareable — a record about a person is
not content to hand round.

**Skills** are saved procedures. Only the name and description are injected; the
body is fetched when the model decides one applies, which is what makes a
hundred skills affordable. A model may write and revise its own — the safety
story is not a gate but a record: every revision is kept, attributed and
revertible. A model that has just read a hostile page can save a skill that
outlives the conversation, and the honest mitigation is that it is visible and
undoable rather than that it was prevented.

**The harness** is why any of it gets used. A model handed a tools array
ignores it and answers from recall, because nothing in the request suggests
otherwise. `services/harness.py` assembles a preamble from what this chat
actually has: when to reach for each tool, the memories, the skill index.

This is an exception to "system prompts are precedence, not concatenation", and
a deliberate one. That rule governs the three *authored* layers and is
untouched — exactly one still wins. The harness is a different axis: it
describes the machinery rather than the behaviour, nobody authored it, and there
is nothing for it to disagree with. It is prepended to whichever authored prompt
won, in one system message, since several endpoints reject a second.

Supporting changes:

- **Sharing**, in one helper. `visible_to()` is the only definition of who can
  see a library item and every listing and tool goes through it. Sharing grants
  *reading*; two people editing one note with no history and no merge is worse
  than copying it. **Administrators do not bypass this** — they bypass
  permissions elsewhere because an admin can grant themselves those anyway, but
  reading somebody's private notes is a different act.
- **FTS5**, created by `db/migrations.py:ensure_fts` with the triggers an
  external-content index needs. Idempotent, like the column sync beside it.
  Terms are ANDed and then ORed: the caller is usually a model writing a whole
  question, and requiring every word loses the match on one absent term.
- **The attach button is a menu** — file, image, a web page, or a document from
  the library. Attaching a document copies it, because history must not change
  when a document is edited later.
- **A URL fetcher with an SSRF guard.** This server can reach the router, the
  other services on the box and LLeMbas itself, and the address can come from a
  model. Private ranges are refused *after resolution* and redirects are followed
  by hand so every hop is checked. An admin can open it deliberately.
- **Model capabilities split** into protocol support and a toggle per built-in
  tool. Rows predating the split have no `tool_*` keys, and absent counts as on
  when `tools` is on — otherwise an upgrade silently takes web search away from
  every model already configured for it.

Also fixes the test fixture, which built the schema with `create_all` and so ran
against a database without the FTS tables production has; it now runs
`sync_schema`, the same path startup takes.

430 tests, ruff clean.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-21 19:43:57 +02:00

208 lines
6.5 KiB
Python

"""Skills: named instructions the model can choose to follow.
Two fields carry the design.
`description` is what gets injected -- one line per skill, for every skill --
and is therefore the only thing the model has to go on when deciding whether a
skill is relevant. A description that does not say *when* to use the skill makes
it invisible in practice.
`body` is fetched only when the model decides to use it. That split is what
makes a hundred skills affordable: the index costs a line each, the instructions
cost nothing until wanted.
**A model may rewrite its own skills**, which is the point -- it is how it
learns a procedure once instead of being told every time. The safety story is
not a gate but a record: every write snapshots what was there first, so a change
can be read and undone. A skill written after reading a hostile web page is a
real risk, and the honest mitigation is that it is visible, attributed and
revertible rather than that it was somehow prevented.
"""
from __future__ import annotations
import logging
import re
from sqlalchemy import select
from sqlalchemy.orm import Session as DBSession
from lembas.db.models import AUTHOR_MODEL, AUTHOR_USER, Skill, SkillRevision, User
from lembas.services import sharing
from lembas.services.library.fts import search_ids
log = logging.getLogger(__name__)
INDEX = "skills_fts"
# A name the model can quote back without getting it wrong.
SKILL_NAME_PATTERN = re.compile(r"^[a-z0-9][a-z0-9-]{1,60}$")
MAX_DESCRIPTION_CHARS = 400
MAX_BODY_CHARS = 20_000
# The index goes into every request, so it has a ceiling like memory does.
MAX_INDEX_SKILLS = 60
class SkillError(Exception):
"""A rejected skill write, with a message fit for the model or the user."""
def slugify(name: str) -> str:
cleaned = re.sub(r"[^a-z0-9]+", "-", (name or "").strip().lower()).strip("-")
return cleaned[:60]
def visible(db: DBSession, user: User | None):
return select(Skill).where(sharing.visible_to(Skill, user))
def get(db: DBSession, skill_id: str, user: User | None) -> Skill | None:
skill = db.get(Skill, skill_id)
if skill is None or not sharing.can_read(db, skill, user):
return None
return skill
def by_name(db: DBSession, name: str, user: User | None) -> Skill | None:
"""Look one up the way the model refers to it."""
if user is None:
return None
return db.scalar(visible(db, user).where(Skill.name == slugify(name)))
def enabled_for(db: DBSession, user: User | None) -> list[Skill]:
"""Skills that should appear in the index, oldest first for a stable order."""
if user is None:
return []
return list(
db.scalars(
visible(db, user)
.where(Skill.enabled.is_(True))
.order_by(Skill.name)
.limit(MAX_INDEX_SKILLS)
)
)
def search(db: DBSession, user: User | None, needle: str, *, limit: int = 10) -> list[Skill]:
hits = search_ids(db, INDEX, needle, limit=limit * 4)
if not hits:
return []
order = {hit.id: position for position, hit in enumerate(hits)}
rows = list(db.scalars(visible(db, user).where(Skill.id.in_(list(order)))))
rows.sort(key=lambda skill: order.get(skill.id, len(order)))
return rows[:limit]
def snapshot(db: DBSession, skill: Skill, *, author: str, note: str = "") -> SkillRevision:
"""Record what a skill looked like before it is changed."""
revision = SkillRevision(
skill_id=skill.id,
description=skill.description,
body=skill.body,
author=author,
note=note[:200],
)
db.add(revision)
return revision
def create(
db: DBSession,
*,
owner: User,
name: str,
description: str,
body: str,
author: str = AUTHOR_USER,
) -> Skill:
slug = slugify(name)
if not SKILL_NAME_PATTERN.match(slug):
raise SkillError(
"A skill name must be two or more letters, numbers or hyphens, "
"such as 'weekly-report'."
)
if by_name(db, slug, owner) is not None:
raise SkillError(f"A skill called {slug!r} already exists. Edit it instead.")
if not description.strip():
raise SkillError(
"A skill needs a description saying when to use it — it is the only "
"thing shown until the skill is opened."
)
skill = Skill(
owner_id=owner.id,
name=slug,
description=description.strip()[:MAX_DESCRIPTION_CHARS],
body=body.strip()[:MAX_BODY_CHARS],
author=author if author in (AUTHOR_USER, AUTHOR_MODEL) else AUTHOR_USER,
)
db.add(skill)
db.commit()
log.info("skill %r created by %s", slug, author)
return skill
def update(
db: DBSession,
skill: Skill,
*,
description: str | None = None,
body: str | None = None,
enabled: bool | None = None,
author: str = AUTHOR_USER,
note: str = "",
) -> Skill:
"""Change a skill, keeping what it was.
The snapshot happens before the change and in the same transaction, so
there is no window where a skill has been rewritten with no record of what
it used to say.
"""
changing = (description is not None and description.strip() != skill.description) or (
body is not None and body.strip() != skill.body
)
if changing:
snapshot(db, skill, author=author, note=note)
if description is not None and description.strip():
skill.description = description.strip()[:MAX_DESCRIPTION_CHARS]
if body is not None:
skill.body = body.strip()[:MAX_BODY_CHARS]
if enabled is not None:
skill.enabled = enabled
if changing:
skill.author = author if author in (AUTHOR_USER, AUTHOR_MODEL) else skill.author
db.commit()
return skill
def revert(db: DBSession, skill: Skill, revision: SkillRevision, *, author: str) -> Skill:
"""Put a skill back to an earlier revision.
The revert is itself a change, so the current state is snapshotted first --
going back is undoable too.
"""
snapshot(db, skill, author=author, note="before revert")
skill.description = revision.description
skill.body = revision.body
db.commit()
return skill
def delete(db: DBSession, skill: Skill) -> None:
sharing.forget_resource(db, skill)
db.delete(skill)
db.commit()
def index_block(db: DBSession, user: User | None) -> str:
"""The one-line-per-skill listing that goes into the prompt."""
skills = enabled_for(db, user)
if not skills:
return ""
return "\n".join(f"- {skill.name}: {skill.description}" for skill in skills)