## Summary Adds `experimental/`, a first version of worker RSI: a Curator that shapes a Raven worker's harness from the task, the materials handed to it and feedback, the way a new employee is trained. The worker is a harness-of-harnesses. A root Raven harness faces the user and calls child harnesses, each a full Raven agent hosted over ACP, through the nodes of its playbooks. Every harness has the same four strategy facets (memory, planning, capability, action). The Curator reads the actual harnesses and generates code and config on those facets at both levels: the root candidate sets the playbooks and writes per-node requirements, each child is curated against the requirements of its nodes, and the set is checked on copies and activated together. Layout: - `curator/`: staged generation (understand, select, design, implement, repair), two-level composition, harness contracts, and `raven_adapter/`, the only layer that touches a worker's Raven runtime - `analyst/`: turns an evaluator's words and execution records into observable behavior requirements - `iteration/`: the generic loop (onboarding curation, then trial, review and curation each round) and a CLI for a human evaluator - `simulation/`: a simulated travel agency that runs the loop, plus one showcase case (`simulation/cases/s0925c/`: case notes, the exported transcript split by phase, and the round 1 and round 3 decks) - `docs/design.md` (structure, curation flow, composition sequence, the Raven connection) and `docs/rsi-iteration.md` (the loop and the Analyst) Key decisions: - No change under `raven/`, and `raven/` does not import `experimental/`. Generated artifacts land on Raven's existing extension points. The few Raven internals the adapter reads are listed in `docs/design.md` section 5.4. - The Curator explores a read-only source snapshot; the caller names the evaluation side to withhold from it, so the generic layers name no scenario. - The single-layer owner analysis is kept only to reproduce the showcase case and is marked for removal; the two-layer Analyst is the default. - The two decks are `.pptx` files of about 230 KB each. They pass `check-large-files`; flagging them in case they count as report assets under AGENTS.md section 7, in which case they can move out of git behind a link. - One follow-up for Raven 0.2.2: playbooks in prompt and rounds modes ship no fixed nodes, so node requirements are read from dag playbooks only. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification On the branch rebased onto the latest `main`: - `uv run pytest tests/test_harness_curator_*.py tests/test_analyst_*.py tests/test_iteration_*.py tests/test_simulation_*.py -q`: 368 passed - `uv run pytest -m integration tests/integration/test_harness_curator_*.py --deselect tests/integration/test_harness_curator_real_llm.py -q` (offline, stub providers): 57 passed - `uv run pytest -q` (whole suite, with the host proxy variables unset and `PYTHONTZPATH` set): 26692 passed, 2 failed. One was an order-dependent test of this PR and is fixed here; the other, `tests/test_subagent_node_runtime.py::test_what_cannot_be_read_names_nothing`, fails only because the machine runs as root and is unrelated. - `uv run ruff check` and `uv run ruff format --check` on the touched files: clean - `scripts/check_large_files.py origin/main..HEAD`, `scripts/check_source_language.py origin/main..HEAD`, `scripts/check_commit_messages.py origin/main..HEAD` and `npx commitlint --from origin/main --to HEAD`: all exit 0 - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [x] User-facing docs or screenshots are updated when needed ## Risk No user-visible change for Raven: `experimental/` is not imported by `raven/` and is not packaged in the wheel (`packages = ["raven"]`). `CONTEXT.md` gains one glossary entry. The Curator's exploration offers a shell only under an OS sandbox executor. Rollback is a revert of the squash commit; nothing else depends on `experimental/`. - [x] Security impact considered - [x] Backward compatibility considered - [x] Rollback path is clear for risky changes ## Related Issues N/A --------- Co-authored-by: Yuqi Yang <yiqiyang524@gmail.com> Co-authored-by: Claude (claude-opus-5-5) <noreply@anthropic.com>
179 lines
8.3 KiB
Python
179 lines
8.3 KiB
Python
"""Program-side triage, one bounded model exchange and the recorded feedback stay faithful to the judgements."""
|
|
|
|
import json
|
|
from types import SimpleNamespace
|
|
|
|
import pytest
|
|
|
|
from experimental.analyst.feedback import Feedback
|
|
from experimental.analyst.materials import RECORDS, read_records
|
|
from experimental.analyst.run import NAME, Limits, analyse, triage
|
|
from experimental.curator.harness import Plan, Task
|
|
from experimental.curator.raven_adapter.worker import Execution
|
|
from experimental.iteration.exchange import ExchangeError
|
|
from experimental.iteration.protocols import Exchange, Item, Signal
|
|
from raven.contracts.llm_provider import LLMResponse, ToolCallRequest
|
|
|
|
|
|
class Provider:
|
|
def __init__(self, *responses):
|
|
self.responses = list(responses)
|
|
self.requests = []
|
|
|
|
async def chat_with_retry(self, **kwargs):
|
|
self.requests.append(kwargs)
|
|
return self.responses.pop(0)
|
|
|
|
|
|
class FakeWorker:
|
|
def __init__(self, root, plan=None):
|
|
self.baseline = SimpleNamespace(task=Task(text="Serve the agency's travellers"))
|
|
self.root = root
|
|
self.last_plan = plan
|
|
|
|
async def inspect(self):
|
|
return SimpleNamespace(sources={"skill.local/pricing": {}, "reference.index": {}, "skill.local/sop": {}})
|
|
|
|
|
|
def response(name, arguments):
|
|
return LLMResponse(content=None, tool_calls=[ToolCallRequest(name, name, arguments)])
|
|
|
|
|
|
def execution(turn_id, reply, *extra):
|
|
records = [{"kind": "runner.event", "event_type": "Text", "event": {"content": reply}}, *extra]
|
|
return Execution(turn_id, [], records, {}, "artifact-1")
|
|
|
|
|
|
@pytest.fixture
|
|
def sessions():
|
|
return {
|
|
"student": [
|
|
Exchange("Plan me a trip", execution("t1", "Here is a three-day itinerary.")),
|
|
Exchange("How much?", execution("t2", "About 3000.", {"kind": "tool.error", "error": "no pricing tool"})),
|
|
]
|
|
}
|
|
|
|
|
|
REQUIREMENT = {
|
|
"behavior": "Before proposing an itinerary, ask for budget, dates, party size and departure city.",
|
|
"observed": "In turn t1 the assistant proposed a three-day itinerary knowing only the destination.",
|
|
"evidence": ["t1", "Here is a three-day itinerary."],
|
|
"expectation": "new",
|
|
"acceptance": "No itinerary appears before those four items have been asked for.",
|
|
"strength": "must_hold",
|
|
}
|
|
|
|
|
|
def test_triage_decides_only_when_nothing_needs_interpretation():
|
|
assert triage(()).decision == "continue"
|
|
passed = Signal("verifier", items=(Item("asks budget", "pass"),), satisfied=True)
|
|
assert triage((passed,)).decision == "continue"
|
|
assert triage((Signal("verifier", items=(Item("asks budget", "fail"),), satisfied=False),)) is None
|
|
assert triage((Signal("verifier", "Too pushy.", satisfied=True),)) is None
|
|
assert triage((Signal("dataset", metrics={"pass_rate": 1.0}),)) is None
|
|
assert triage((Signal("human", "Fine."),)) is None
|
|
|
|
|
|
def test_feedback_requires_requirements_exactly_for_curate():
|
|
with pytest.raises(ValueError, match="curate requires"):
|
|
Feedback(decision="curate", reason="r")
|
|
with pytest.raises(ValueError, match="carry none"):
|
|
Feedback(decision="continue", reason="r", requirements=(REQUIREMENT,))
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_model_reads_records_then_submits_and_the_round_is_recorded(tmp_path, sessions):
|
|
plan = Plan(
|
|
understanding="u",
|
|
design="d",
|
|
changes=[{"target": "planning.strategy", "reason": "r", "expected": "asks first", "verification": "v"}],
|
|
)
|
|
worker = FakeWorker(tmp_path, plan)
|
|
signals = (Signal("human", "It never asked my budget before recommending."),)
|
|
provider = Provider(
|
|
response(RECORDS, {"turn_id": "t2", "kind": "tool."}),
|
|
response(NAME, {"decision": "curate", "reason": "The intake gap recurs.", "requirements": [REQUIREMENT]}),
|
|
)
|
|
previous = (Signal("dataset", metrics={"pass_rate": 0.5}),)
|
|
feedback = await analyse(worker, provider, signals, sessions, previous_signals=previous, model="analyst-model")
|
|
assert feedback.decision == "curate" and feedback.requirements[0].expectation == "new"
|
|
materials = json.loads(provider.requests[0]["messages"][1]["content"])
|
|
assert materials["signals"][0]["text"] == "It never asked my budget before recommending."
|
|
assert materials["previous_signals"][0]["metrics"] == {"pass_rate": 0.5}
|
|
assert materials["previous_expectations"][0]["expected"] == "asks first"
|
|
assert materials["skills"] == ["skill.local/pricing", "skill.local/sop"]
|
|
assert materials["sessions"]["student"][1]["errors"] == ["no pricing tool"]
|
|
assert materials["sessions"]["student"][1]["record_kinds"] == {"runner.event": 1, "tool.error": 1}
|
|
tools = {entry["function"]["name"] for entry in provider.requests[0]["tools"]}
|
|
assert tools == {RECORDS, NAME}
|
|
assert provider.requests[0]["tools"][0]["function"]["parameters"]["properties"]["turn_id"]["enum"] == ["t1", "t2"]
|
|
tool_replies = [row for row in provider.requests[1]["messages"] if row["role"] == "tool"]
|
|
query_result = json.loads(tool_replies[0]["content"])
|
|
assert query_result["turn_id"] == "t2" and "no pricing tool" in query_result["text"]
|
|
assert "runner.event" not in query_result["text"]
|
|
assert all(request["model"] == "analyst-model" for request in provider.requests)
|
|
record = json.loads(next((tmp_path / "analysis").glob("*.json")).read_text())
|
|
assert record["feedback"]["decision"] == "curate" and record["materials"]["task"] == worker.baseline.task.text
|
|
assert [row["event"] for row in record["trace"]] == ["model.call", "query", "model.call", NAME]
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_invalid_submissions_are_returned_and_the_budget_is_final(tmp_path, sessions):
|
|
worker = FakeWorker(tmp_path)
|
|
signals = (Signal("human", "Not good."),)
|
|
provider = Provider(
|
|
response(NAME, {"decision": "curate", "reason": "r"}),
|
|
LLMResponse(content="Let me think."),
|
|
)
|
|
with pytest.raises(ExchangeError, match="analyst: submission budget exhausted") as info:
|
|
await analyse(worker, provider, signals, sessions, limits=Limits(max_calls=2))
|
|
assert [row["event"] for row in info.value.trace] == [
|
|
"model.call",
|
|
"output.rejected",
|
|
"model.call",
|
|
"output.missing",
|
|
]
|
|
rejection = json.loads(next(row for row in provider.requests[1]["messages"] if row["role"] == "tool")["content"])
|
|
assert "curate requires" in rejection["error"]
|
|
record = json.loads(next((tmp_path / "analysis").glob("*.json")).read_text())
|
|
assert record["error"].startswith("analyst: submission budget exhausted") and "feedback" not in record
|
|
|
|
|
|
def test_record_query_filters_by_kind_and_bounds_offsets(sessions):
|
|
result = read_records(sessions, {"turn_id": "t2", "kind": "runner", "length": 40})
|
|
assert result["next_offset"] == 40 and result["total_characters"] > 40
|
|
assert read_records(sessions, {"turn_id": "t1"})["next_offset"] is None
|
|
with pytest.raises(ValueError, match="offset exceeds"):
|
|
read_records(sessions, {"turn_id": "t1", "offset": 10**6})
|
|
with pytest.raises(ValueError, match="no turn"):
|
|
read_records(sessions, {"turn_id": "missing"})
|
|
|
|
|
|
def test_limits_reject_non_positive_bounds():
|
|
for bad in ({"max_calls": 0}, {"call_timeout": 0}, {"call_timeout": float("inf")}):
|
|
with pytest.raises(ValueError, match="finite positive"):
|
|
Limits(**bad)
|
|
|
|
|
|
def test_read_records_rejects_an_unknown_turn_as_a_parameter_error(sessions):
|
|
with pytest.raises(ValueError, match="no turn in this round has the id 'ghost'"):
|
|
read_records(sessions, {"turn_id": "ghost"})
|
|
assert json.loads(read_records(sessions, {"turn_id": "t2", "kind": "tool."})["text"])[0]["kind"] == "tool.error"
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_the_last_call_offers_only_the_submission_tool(tmp_path, sessions):
|
|
worker = FakeWorker(tmp_path)
|
|
provider = Provider(
|
|
response(RECORDS, {"turn_id": "t1"}),
|
|
response(NAME, {"decision": "continue", "reason": "Nothing to change."}),
|
|
)
|
|
feedback = await analyse(worker, provider, (Signal("human", "Hm."),), sessions, limits=Limits(max_calls=2))
|
|
assert feedback.decision == "continue"
|
|
assert [entry["function"]["name"] for entry in provider.requests[0]["tools"]] == [RECORDS, NAME]
|
|
assert [entry["function"]["name"] for entry in provider.requests[1]["tools"]] == [NAME]
|
|
nudges = [
|
|
row for row in provider.requests[1]["messages"] if row["role"] == "user" and "last call" in str(row["content"])
|
|
]
|
|
assert len(nudges) == 1
|