ORCH-118 (inventory-first, docs+tests only): publish an evidence-based map of
every place the orchestrator's control flow consumes (or can consume) an LLM
judgment, mark the control-path axis (C control-path vs P artifact-producer),
define "avoidable LLM control path" as a checkable two-bit predicate, classify
each call-site, and order the deterministic-replacement roadmap. Pin the map to
code with offline structural anti-drift tests.
- docs/architecture/llm-call-sites.md — map + machine-readable inventory block
+ control-path axis + classification + keep-LLM justifications + deterministic
non-agent paths (FR-1/FR-2/FR-3/FR-8).
- docs/architecture/llm-determinization-roadmap.md — ordered candidates BY ROLE,
savings sourced from agent_runs, recommended first slice = deployer staging
(FR-4). No fabricated follow-up Plane-IDs (R3/NFR-6).
- docs/architecture/llm-usage-policy.md — normative principle, keep/replace
criteria via the axis, definition of "avoidable LLM control path" (FR-5/FR-8).
- tests/test_llm_call_site_inventory.py — TC-01/02/03/04/05/06/09/12/13/14.
- tests/test_llm_determinization_docs.py — TC-07/08/11.
- CHANGELOG.md + docs/overview/tech-quality-security.md — golden-source sync (AC-8).
Avoidable LLM control paths = {tester, deployer}; control-path-keep = {reviewer};
not-control-path (P) = {analyst, architect, developer}. Single LLM transport =
launcher._spawn (S0); no alternative transport (TC-12). Runtime untouched:
STAGE_TRANSITIONS / QG_CHECKS / check_* / machine-verdict keys / DB schema are
byte-for-byte; no replacement runners implemented (FR-7). Full suite: 2081 passed.
Refs: ORCH-118
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
449 lines
18 KiB
Python
449 lines
18 KiB
Python
# ORCH-118 (FR-6 / AC-1, AC-3, AC-6, AC-7, AC-10): structural anti-drift tests for
|
|
# the LLM call-site map (docs/architecture/llm-call-sites.md).
|
|
#
|
|
# UNIT OF INVENTORY = an *LLM consultation* (a control-path consumes an LLM
|
|
# judgment), NOT "a process is spawned / Claude CLI exists" (R4, BRD §0). The
|
|
# discriminator of every check below is therefore **"consults an LLM"**, never
|
|
# "spawns a subprocess": the orchestrator spawns dozens of deterministic tools
|
|
# (git / pytest / docker / ssh / scanners / staging_check.py) and those are
|
|
# explicitly NOT matched — otherwise the test would degenerate into "count every
|
|
# Popen" (the corruption these tests exist to avoid).
|
|
#
|
|
# These tests are fully offline and deterministic: no network, no LLM, no
|
|
# subprocess-to-a-model. They read repository source + the canonical machine
|
|
# blocks embedded in the map and assert the map stays in sync with the code.
|
|
#
|
|
# The map carries two machine-readable blocks the parser keys on:
|
|
# * ORCH-118-INVENTORY-BLOCK — the call-site table (8 columns, D2)
|
|
# * ORCH-118-KEEP-JUSTIFICATION-BLOCK — keep-LLM named-judgment list (TC-05)
|
|
# Both are human-readable markdown AND machine-parseable (stdlib split), per
|
|
# ADR-001 D2/D5 (no brittle prose regex).
|
|
|
|
import re
|
|
from pathlib import Path
|
|
|
|
from src.qg.checks import QG_CHECKS
|
|
from src.stages import STAGE_TRANSITIONS
|
|
|
|
REPO_ROOT = Path(__file__).resolve().parents[1]
|
|
SRC = REPO_ROOT / "src"
|
|
WATCHDOG = REPO_ROOT / "watchdog"
|
|
AGENTS_DIR = REPO_ROOT / ".openclaw" / "agents"
|
|
MAP = REPO_ROOT / "docs" / "architecture" / "llm-call-sites.md"
|
|
|
|
# The single allowed transport (S0): launcher._spawn builds + launches the Claude CLI.
|
|
TRANSPORT_FILE = "src/agents/launcher.py"
|
|
|
|
# Roles split by control-path axis (§0-bis). Ground truth lives in src/qg/checks.py
|
|
# (the deterministic consumers) — the map must mirror it.
|
|
P_ROLES = frozenset({"analyst", "architect", "developer"})
|
|
C_ROLES = frozenset({"reviewer", "tester", "deployer"})
|
|
AGENT_ROLES = P_ROLES | C_ROLES
|
|
|
|
# P-roles are consumed by deterministic gates that judge an ARTIFACT (file
|
|
# presence / CI) independently of any LLM self-report.
|
|
P_CONSUMERS = frozenset(
|
|
{"check_analysis_complete", "check_architecture_done", "check_ci_green"}
|
|
)
|
|
# C-roles are consumed by verdict-parsers that READ a machine-verdict the LLM wrote
|
|
# — the LLM judgment branches the control flow (PASS->advance / FAIL->rollback).
|
|
C_CONSUMERS = frozenset(
|
|
{
|
|
"check_reviewer_verdict",
|
|
"_parse_tests_verdict",
|
|
"_parse_staging_status",
|
|
"_parse_deploy_status",
|
|
}
|
|
)
|
|
|
|
ALLOWED_CLASSES = frozenset(
|
|
{
|
|
"keep-LLM",
|
|
"replace-deterministic-now",
|
|
"replace-later/risky",
|
|
"needs-hybrid-fallback",
|
|
"already-deterministic",
|
|
}
|
|
)
|
|
|
|
AGENT_IDS = frozenset({"A1", "A2", "A3", "A4", "A5", "A6"})
|
|
ALL_IDS = frozenset({"S0", "A1", "A2", "A3", "A4", "A5", "A6", "D1", "D2"})
|
|
|
|
# Deterministic leaf modules / routing that must NOT consult an LLM (FR-3 / AC-3).
|
|
DETERMINISTIC_MODULES = (
|
|
"stages",
|
|
"stage_engine",
|
|
"serial_gate",
|
|
"merge_gate",
|
|
"coverage_gate",
|
|
"security_gate",
|
|
"staging_verdict",
|
|
"review_parse",
|
|
"error_classifier",
|
|
"frontmatter",
|
|
"self_deploy",
|
|
"post_deploy",
|
|
"transition_lease",
|
|
"reconciler",
|
|
"job_reaper",
|
|
)
|
|
|
|
# Alternative-LLM-transport signatures forbidden anywhere in src/** + watchdog/**
|
|
# (TC-12 / FR-6f): an LLM SDK import or a direct Anthropic/Claude HTTP endpoint.
|
|
_FORBIDDEN_TRANSPORT_RE = (
|
|
re.compile(r"^\s*(?:from|import)\s+anthropic\b", re.M),
|
|
re.compile(r"^\s*(?:from|import)\s+openai\b", re.M),
|
|
re.compile(r"api\.anthropic\.com"),
|
|
re.compile(r"/v1/messages"),
|
|
)
|
|
|
|
# Frozen snapshot of the runtime contract (TC-09 / FR-7 / AC-7). ORCH-118 is
|
|
# docs+tests-only; if this drifts, the map task touched the stage machine / gates.
|
|
EXPECTED_STAGE_AGENTS = frozenset(
|
|
{"analyst", "architect", "developer", "reviewer", "tester", "deployer"}
|
|
)
|
|
EXPECTED_QG_CHECKS = frozenset(
|
|
{
|
|
"check_analysis_approved",
|
|
"check_analysis_complete",
|
|
"check_architecture_done",
|
|
"check_ci_green",
|
|
"check_review_approved",
|
|
"check_tests_passed",
|
|
"check_reviewer_verdict",
|
|
"check_tests_local",
|
|
"check_deploy_status",
|
|
"check_staging_status",
|
|
"check_branch_mergeable",
|
|
"check_staging_image_fresh",
|
|
"check_security_gate",
|
|
"check_coverage_gate",
|
|
}
|
|
)
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# Helpers (stdlib only)
|
|
# ---------------------------------------------------------------------------
|
|
def _src_py_files() -> list[Path]:
|
|
return sorted(SRC.glob("**/*.py"))
|
|
|
|
|
|
def _src_and_watchdog_py_files() -> list[Path]:
|
|
files = list(SRC.glob("**/*.py"))
|
|
if WATCHDOG.is_dir():
|
|
files.extend(WATCHDOG.glob("**/*.py"))
|
|
return sorted(files)
|
|
|
|
|
|
def _rel(p: Path) -> str:
|
|
return p.relative_to(REPO_ROOT).as_posix()
|
|
|
|
|
|
def _function_body(source: str, name: str) -> str:
|
|
"""Return the source text of ``def <name>`` up to (excluding) the next
|
|
same-or-lower-indent def/class/decorator. Robust to line drift."""
|
|
lines = source.splitlines()
|
|
start = None
|
|
indent = 0
|
|
for i, line in enumerate(lines):
|
|
stripped = line.lstrip()
|
|
if stripped.startswith(f"def {name}("):
|
|
start = i
|
|
indent = len(line) - len(stripped)
|
|
break
|
|
assert start is not None, f"def {name}( not found in source"
|
|
body = [lines[start]]
|
|
for line in lines[start + 1 :]:
|
|
if not line.strip():
|
|
body.append(line)
|
|
continue
|
|
cur_indent = len(line) - len(line.lstrip())
|
|
head = line.lstrip()
|
|
if cur_indent <= indent and head.startswith(("def ", "class ", "@")):
|
|
break
|
|
body.append(line)
|
|
return "\n".join(body)
|
|
|
|
|
|
def _extract_block(text: str, name: str) -> str:
|
|
start = f"<!-- {name}:START -->"
|
|
end = f"<!-- {name}:END -->"
|
|
assert start in text, f"missing block start marker {start!r} in map"
|
|
assert end in text, f"missing block end marker {end!r} in map"
|
|
return text.split(start, 1)[1].split(end, 1)[0]
|
|
|
|
|
|
def _parse_pipe_table(block: str) -> list[dict]:
|
|
"""Parse a GitHub-style pipe table into a list of {column: value} dicts."""
|
|
header = None
|
|
rows: list[dict] = []
|
|
for raw in block.splitlines():
|
|
line = raw.strip()
|
|
if not line.startswith("|"):
|
|
continue
|
|
cells = [c.strip() for c in line.strip("|").split("|")]
|
|
joined = "".join(cells)
|
|
if joined and set(joined) <= set("-: "):
|
|
continue # separator row |---|---|
|
|
if header is None:
|
|
header = [c.lower() for c in cells]
|
|
continue
|
|
rows.append(dict(zip(header, cells)))
|
|
return rows
|
|
|
|
|
|
def _inventory_rows() -> list[dict]:
|
|
block = _extract_block(MAP.read_text(encoding="utf-8"), "ORCH-118-INVENTORY-BLOCK")
|
|
rows = _parse_pipe_table(block)
|
|
assert rows, "inventory block parsed to zero rows"
|
|
return rows
|
|
|
|
|
|
def _agent_rows() -> list[dict]:
|
|
return [r for r in _inventory_rows() if r["id"] in AGENT_IDS]
|
|
|
|
|
|
def _by_role() -> dict[str, dict]:
|
|
return {r["role"]: r for r in _agent_rows()}
|
|
|
|
|
|
def _parse_justifications() -> dict[str, str]:
|
|
"""Parse the keep-LLM named-judgment list: ``- role: justification text``."""
|
|
block = _extract_block(
|
|
MAP.read_text(encoding="utf-8"), "ORCH-118-KEEP-JUSTIFICATION-BLOCK"
|
|
)
|
|
out: dict[str, str] = {}
|
|
for raw in block.splitlines():
|
|
line = raw.strip()
|
|
m = re.match(r"^[-*]\s*([A-Za-z_-]+)\s*:\s*(.+)$", line)
|
|
if m:
|
|
out[m.group(1).strip()] = m.group(2).strip()
|
|
return out
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# TC-01 — single LLM-consultation transport (necessary, completed by TC-12).
|
|
# ---------------------------------------------------------------------------
|
|
def test_tc01_single_llm_transport():
|
|
"""Exactly one place in src/** assembles+launches the Claude CLI, matched by
|
|
the CONJUNCTION of transport signals (CLAUDE_BIN AND --system-prompt AND a
|
|
process launcher) — and it is launcher._spawn. The conjunction is mandatory:
|
|
bare CLAUDE_BIN would false-positive on preflight.py (existence check) and
|
|
config.py (path literal), neither of which consults an LLM (ADR D5a)."""
|
|
hits = []
|
|
for f in _src_py_files():
|
|
text = f.read_text(encoding="utf-8")
|
|
launches = ("Popen" in text) or ('"bash"' in text) or ("'bash'" in text)
|
|
if "--system-prompt" in text and "CLAUDE_BIN" in text and launches:
|
|
hits.append(_rel(f))
|
|
assert hits == [TRANSPORT_FILE], (
|
|
"expected the single LLM-transport to be launcher._spawn; got: " + repr(hits)
|
|
)
|
|
# The transport assembly lives inside _spawn specifically.
|
|
launcher = (SRC / "agents" / "launcher.py").read_text(encoding="utf-8")
|
|
assert "--system-prompt" in _function_body(launcher, "_spawn"), (
|
|
"--system-prompt is not inside def _spawn — transport moved?"
|
|
)
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# TC-12 — no alternative LLM transport (FR-6f / AC-1, AC-6).
|
|
# ---------------------------------------------------------------------------
|
|
def test_tc12_no_alternative_llm_transport():
|
|
"""No LLM-SDK import, no direct Anthropic/Claude HTTP endpoint, and no SECOND
|
|
--system-prompt-bearing subprocess builder anywhere in src/** + watchdog/**.
|
|
Closes the gap 'one _spawn green, a new consultation grew next to it'."""
|
|
sdk_offenders = []
|
|
for f in _src_and_watchdog_py_files():
|
|
text = f.read_text(encoding="utf-8")
|
|
for rx in _FORBIDDEN_TRANSPORT_RE:
|
|
if rx.search(text):
|
|
sdk_offenders.append(f"{_rel(f)}: {rx.pattern}")
|
|
assert not sdk_offenders, (
|
|
"alternative LLM transport found (allowed transport = S0/launcher._spawn "
|
|
"only):\n" + "\n".join(sdk_offenders)
|
|
)
|
|
# No second --system-prompt builder outside the allowlisted transport file.
|
|
second_builders = [
|
|
_rel(f)
|
|
for f in _src_and_watchdog_py_files()
|
|
if "--system-prompt" in f.read_text(encoding="utf-8")
|
|
and _rel(f) != TRANSPORT_FILE
|
|
]
|
|
assert second_builders == [], (
|
|
"a second --system-prompt subprocess builder appeared: " + repr(second_builders)
|
|
)
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# TC-02 — deterministic modules carry no LLM consultation (FR-6b / AC-3).
|
|
# ---------------------------------------------------------------------------
|
|
def test_tc02_deterministic_modules_no_llm_consultation():
|
|
"""The listed routing/leaf modules do not consult an LLM (no _spawn transport,
|
|
no alternative transport). Their git/pytest/docker/ssh/scanner subprocesses are
|
|
deterministic TOOLS, not LLM consultations — discriminator is 'consults LLM',
|
|
not 'spawns subprocess'."""
|
|
offenders = []
|
|
for mod in DETERMINISTIC_MODULES:
|
|
path = SRC / f"{mod}.py"
|
|
assert path.is_file(), f"deterministic module missing: {path}"
|
|
text = path.read_text(encoding="utf-8")
|
|
if "--system-prompt" in text:
|
|
offenders.append(f"{mod}: builds --system-prompt (LLM transport)")
|
|
for rx in _FORBIDDEN_TRANSPORT_RE:
|
|
if rx.search(text):
|
|
offenders.append(f"{mod}: {rx.pattern}")
|
|
assert not offenders, "LLM consultation found in deterministic path:\n" + "\n".join(
|
|
offenders
|
|
)
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# TC-03 — prompt files on disk match the map, both ways (FR-6c / AC-1).
|
|
# ---------------------------------------------------------------------------
|
|
def test_tc03_prompt_files_match_map():
|
|
on_disk = {p.stem for p in AGENTS_DIR.glob("*.md")}
|
|
in_map = {r["role"] for r in _agent_rows()}
|
|
assert on_disk == set(AGENT_ROLES), (
|
|
f"prompt files on disk drifted from the 6 canonical roles: {on_disk}"
|
|
)
|
|
assert in_map == on_disk, (
|
|
f"map agent roles {in_map} != prompt files on disk {on_disk}"
|
|
)
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# TC-04 — totality + axis-consistent classification (FR-6d / FR-2 / AC-2).
|
|
# ---------------------------------------------------------------------------
|
|
def test_tc04_classification_total_and_axis_consistent():
|
|
rows = _inventory_rows()
|
|
ids = [r["id"] for r in rows]
|
|
assert len(ids) == len(set(ids)), f"duplicate call-site ids: {ids}"
|
|
assert set(ids) == set(ALL_IDS), f"call-site id set drifted: {set(ids)}"
|
|
|
|
for r in rows:
|
|
cls = r["classification"]
|
|
if r["id"] == "S0":
|
|
assert cls in ("-", "—"), f"S0 (transport) must not be classified: {cls!r}"
|
|
else:
|
|
assert cls in ALLOWED_CLASSES or cls in ("-", "—"), (
|
|
f"{r['id']} class out of taxonomy: {cls!r}"
|
|
)
|
|
|
|
# Class is DERIVED from the axis (not postulated): P->keep; C+!avoidable->keep;
|
|
# C+avoidable->replace-*/hybrid.
|
|
for r in _agent_rows():
|
|
axis = r["axis"].upper()
|
|
avoidable = r["avoidable"].lower()
|
|
cls = r["classification"]
|
|
if axis == "P":
|
|
assert cls == "keep-LLM", f"{r['role']} is P but class {cls!r}"
|
|
elif axis == "C" and avoidable == "no":
|
|
assert cls == "keep-LLM", f"{r['role']} is C-keep but class {cls!r}"
|
|
elif axis == "C" and avoidable == "yes":
|
|
assert cls in {
|
|
"replace-deterministic-now",
|
|
"replace-later/risky",
|
|
"needs-hybrid-fallback",
|
|
}, f"{r['role']} is avoidable but class {cls!r}"
|
|
else:
|
|
raise AssertionError(f"{r['role']}: bad axis/avoidable {axis!r}/{avoidable!r}")
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# TC-05 — keep-LLM requires a named judgment; C-keep states non-derivability.
|
|
# ---------------------------------------------------------------------------
|
|
def test_tc05_keep_llm_named_judgment():
|
|
keep_roles = {r["role"] for r in _agent_rows() if r["classification"] == "keep-LLM"}
|
|
assert keep_roles == {"analyst", "architect", "developer", "reviewer"}, (
|
|
f"keep-LLM role set drifted: {keep_roles}"
|
|
)
|
|
just = _parse_justifications()
|
|
for role in keep_roles:
|
|
assert just.get(role, "").strip(), f"keep-LLM role {role} has no named judgment"
|
|
# reviewer is C-keep: its justification must explain NON-derivability of the verdict.
|
|
assert "deriv" in just["reviewer"].lower(), (
|
|
"reviewer (C-keep) justification must state the verdict is NOT derivable "
|
|
"from an exit-code"
|
|
)
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# TC-06 — capability != consultation: D1/D2 intercepted before _spawn (FR-6e).
|
|
# ---------------------------------------------------------------------------
|
|
def test_tc06_capability_not_consultation():
|
|
launcher = (SRC / "agents" / "launcher.py").read_text(encoding="utf-8")
|
|
body = _function_body(launcher, "launch_job")
|
|
i_finalizer = body.find('"deploy-finalizer"')
|
|
i_monitor = body.find('"post-deploy-monitor"')
|
|
i_spawn = body.find("self._spawn(")
|
|
assert i_finalizer != -1, "deploy-finalizer guard not found in launch_job"
|
|
assert i_monitor != -1, "post-deploy-monitor guard not found in launch_job"
|
|
assert i_spawn != -1, "self._spawn( call not found in launch_job"
|
|
assert i_finalizer < i_spawn, "deploy-finalizer guard must precede _spawn"
|
|
assert i_monitor < i_spawn, "post-deploy-monitor guard must precede _spawn"
|
|
assert "return self._run_deploy_finalizer_job" in body
|
|
assert "return self._run_post_deploy_monitor_job" in body
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# TC-09 — runtime contract snapshot unchanged (FR-7 / AC-7).
|
|
# ---------------------------------------------------------------------------
|
|
def test_tc09_runtime_contract_snapshot():
|
|
agents = {t["agent"] for t in STAGE_TRANSITIONS.values() if t["agent"]}
|
|
assert agents == set(EXPECTED_STAGE_AGENTS), (
|
|
f"STAGE_TRANSITIONS agent set changed: {agents}"
|
|
)
|
|
assert set(QG_CHECKS) == set(EXPECTED_QG_CHECKS), (
|
|
f"QG_CHECKS name set changed: {set(QG_CHECKS)}"
|
|
)
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# TC-13 — control-path axis is consistent with the real consumer (FR-6g / AC-10).
|
|
# ---------------------------------------------------------------------------
|
|
def test_tc13_control_path_axis_correct():
|
|
checks_src = (SRC / "qg" / "checks.py").read_text(encoding="utf-8")
|
|
rows = _agent_rows()
|
|
for r in rows:
|
|
role = r["role"]
|
|
axis = r["axis"].upper()
|
|
consumer = r["output_consumer"].split(":")[0].strip()
|
|
assert re.search(rf"def {re.escape(consumer)}\(", checks_src), (
|
|
f"{role}: output_consumer {consumer!r} is not a def in src/qg/checks.py"
|
|
)
|
|
if role in P_ROLES:
|
|
assert axis == "P", f"{role} must be axis P, got {axis!r}"
|
|
assert consumer in P_CONSUMERS, (
|
|
f"{role} (P) consumer {consumer!r} is not a deterministic artifact gate"
|
|
)
|
|
elif role in C_ROLES:
|
|
assert axis == "C", f"{role} must be axis C, got {axis!r}"
|
|
assert consumer in C_CONSUMERS, (
|
|
f"{role} (C) consumer {consumer!r} is not a verdict-parser"
|
|
)
|
|
else:
|
|
raise AssertionError(f"unexpected agent role in map: {role!r}")
|
|
assert {r["role"] for r in rows if r["axis"].upper() == "P"} == set(P_ROLES)
|
|
assert {r["role"] for r in rows if r["axis"].upper() == "C"} == set(C_ROLES)
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# TC-14 — avoidable LLM control-path set is exactly {tester, deployer} (FR-6h).
|
|
# ---------------------------------------------------------------------------
|
|
def test_tc14_avoidable_set_fixed():
|
|
rows = _agent_rows()
|
|
by_role = {r["role"]: r for r in rows}
|
|
avoidable = {r["role"] for r in rows if r["avoidable"].lower() == "yes"}
|
|
assert avoidable == {"tester", "deployer"}, (
|
|
f"avoidable LLM control-path set drifted from {{tester, deployer}}: {avoidable}"
|
|
)
|
|
# reviewer: control path (C) but KEEP — verdict not derivable.
|
|
assert by_role["reviewer"]["axis"].upper() == "C"
|
|
assert by_role["reviewer"]["avoidable"].lower() == "no"
|
|
# analyst / architect / developer: NOT control path (P artifact-producer).
|
|
for role in ("analyst", "architect", "developer"):
|
|
assert by_role[role]["axis"].upper() == "P", f"{role} must be P (not control path)"
|
|
assert by_role[role]["avoidable"].lower() == "no", f"{role} must not be avoidable"
|