Load tool / tool use / tool reply information from the database again

[simonw]

This is the continuation of the work from:

• Persist tools and tool responses to the database #1003

Will be needed to get llm -c working properly, see also:

• Decide how tool definitions should interact with ongoing conversations #989

[simonw]

Relevant code:

llm/llm/models.py

Lines 334 to 388 in 7f49cc2

	@classmethod
	def from_row(cls, db, row, _async=False):
	from llm import get_model, get_async_model
	if _async:
	model = get_async_model(row["model"])
	else:
	model = get_model(row["model"])
	# Schema
	schema = None
	if row["schema_id"]:
	schema = json.loads(db["schemas"].get(row["schema_id"])["content"])
	all_fragments = list(db.query(FRAGMENT_SQL, {"response_id": row["id"]}))
	fragments = [
	row["content"] for row in all_fragments if row["fragment_type"] == "prompt"
	]
	system_fragments = [
	row["content"] for row in all_fragments if row["fragment_type"] == "system"
	]
	response = cls(
	model=model,
	prompt=Prompt(
	prompt=row["prompt"],
	model=model,
	fragments=fragments,
	attachments=[],
	system=row["system"],
	schema=schema,
	system_fragments=system_fragments,
	options=model.Options(**json.loads(row["options_json"])),
	),
	stream=False,
	)
	prompt_json = json.loads(row["prompt_json"] or "null")
	response.id = row["id"]
	response._prompt_json = prompt_json
	response.response_json = json.loads(row["response_json"] or "null")
	response._done = True
	response._chunks = [row["response"]]
	# Attachments
	response.attachments = [
	Attachment.from_row(attachment_row)
	for attachment_row in db.query(
	"""
	select attachments.* from attachments
	join prompt_attachments on attachments.id = prompt_attachments.attachment_id
	where prompt_attachments.response_id = ?
	order by prompt_attachments."order"
	""",
	[row["id"]],
	)
	]
	return response

[simonw]

GITHUB_TOKEN="$(llm keys get github)" \
  llm -f issue:https://github.com/simonw/llm/issues/1010 \
  -f github:simonw/llm \
  -m o4-mini -s 'implement this, output full code snippets not a diff'

I think that's about 174,767 tokes (ran it through ttok).

(Rats, forgot to upgrade to simonw/llm-fragments-github#12 first)

https://gist.github.com/simonw/1d46a3a4cb84965bf27b1088ac2d785f - not great, not worth it for this one. 22.108 cents.

[simonw]

We don't serialize the actual Python function implementations to our database, so worth noting that we won't be able to load tools from the database and have them actually callable.

[simonw]

I've not fully tested this code yet, manually or automatically.

[simonw]

I've tested this manually extensively with llm log -c, mainly in:

• Make sure multi-step tool conversations work for the default collection of models #1019