Files
reverse-engineering-anything/tests/support/analysis-efficiency/agent-trial.py
T
morluto 1f24b3fa57 feat(evidence)!: name the Evidence subject by its target identity (format revision 2) (#11416)
* feat(evidence)!: name the Evidence subject by its target identity (format revision 2)

The Evidence subject carries subject.identity (kind, sha256, format,
architecture, selection) and the Evidence ID commits to it, so raw/DOL
entry points and ESP flash offsets of the same bytes are different
subjects. Every record declares evidence_format: 2; revision 1 bundles,
standalone records and snapshots are refused with the recreate guidance.
binaryTargetIdentity is the one projection of an open target, used by
Evidence and snapshot comparison.

BREAKING CHANGE: Evidence format revision 2. Evidence IDs commit to the
subject's target identity (subject.identity, including the raw/DOL entry
point or ESP flash offsets), and every record declares evidence_format: 2.
Evidence bundles, records and analysis snapshots from earlier releases are
refused; recreate the observations with this release.

* refactor(analysis-view): read parent subjects through the Evidence identity accessors (#11423)

Analysis views compare their parent's file identity (evidenceFileTarget)
and take path and digest from evidenceSubjectTarget instead of reading
the subject's digest, format and architecture members, which duplicate
subject.identity until they are removed. A counted guard
(evidence-subject-declaration-read) keeps new direct reads out and
records the 20 remaining files.

* fix(evidence): carry format refusals through Evidence readers and bind target selections

- safeParseEvidence/safeParseImmutableEvidence return the format refusal as
  an expected failure; every consumer maps it to the one integrity refusal
  with the recreate guidance. An unmarked object is an earlier record only
  when it has every member records have always carried.
- ArtifactProvider inventory roots and provider subjects that name the
  target's own bytes carry the target's selection (ESP flash offsets).
- Snapshot rows bind only Evidence whose identity, selection included,
  matches the snapshot target.
- Entry-point selections use the raw-binary entry point form.

* fix(evidence): refuse unauthenticated or earlier-format records instead of reading them as absent

- recognizeEvidence (evidenceValidation) replaces tryParseEvidence: input
  without an evidence_id is not Evidence (ok(undefined)); a record that names
  one but fails authentication or is from another format revision is an
  EvidenceIntegrityError carrying the recreate guidance, never undefined.
- Direct analysis retains that refusal as a secondary failure instead of
  silently skipping source capture; cleanup projection, source-closure
  annotation and the flutter, debugger and release-exposure recoveries use
  the same owner (the last three previously read the unauthenticated schema).
- A desktop capture.json from an earlier format is refused with the recreate
  guidance instead of "not desktop application capture Evidence".

* fix(evidence): advertise revision 2 ledger examples and rank delivery omissions by serialized size

- The obligation ledger and reconstruction coverage examples named
  evidence_format 1, which revision 2 refuses; they now use
  EVIDENCE_FORMAT_REVISION.
- MCP delivery chose which unretained recovery to omit by a parsed-storage
  lower bound clipped at the response budget. Once revision 2 subjects made
  the source record's estimate cross the budget, both candidates tied at the
  clip and the smaller source was omitted ahead of the larger completed
  observation. Candidates are now ranked by serialized size, which is what
  the budget counts.

* refactor(native): compare function snapshots by their Evidence target identity (#11450)

Function snapshots carry the subject's file target identity instead of the
duplicated subject declaration. Call paths and function comparisons compare
targets through sameTargetIdentity, so the same bytes opened at a different
raw entry point or ESP flash offset no longer count as one coordinate space:
build_call_path refuses the mix, and compare_functions no longer matches
entries by address across selections. Structural candidates and native
function matching read the architecture and digest through
evidenceFileTarget.

Refs #11417

* refactor(artifacts): compare artifact and WASM summaries by their Evidence target identity (#11461)

Artifact comparison authenticates a summary against its retained inventory
through sameTargetIdentity and rebinds the alias with the whole target,
selection included, so a flash-selected inventory summary taken under an
alias path is no longer refused. WASM comparison, the analysis view
artifact projection and the inventory root check read the digest, format
and display path through the Evidence identity accessors.

Refs #11417

* refactor(evidence): read record subjects through the identity accessors in JS, managed, ledger and bundle readers (#11470)

JavaScript application graphs, runtime reconciliation, managed inspection
and native verification sources, the record summary filter, the ledger's
residual-unknown target and bundle scope selection read the subject digest,
format and display path through evidenceFileTarget and
evidenceSubjectDisplay. No comparison changes: each reader matches one
record's facts against its own payload or a caller filter.

Refs #11417

* refactor(evidence)!: name the Evidence subject only by its target identity (#11497)

* refactor(evidence)!: name the Evidence subject only by its target identity

The subject keeps `identity` and the display facts `name` and `local_path`.
The duplicated `digest`, `format` and `architecture` members are removed; the
format-owned declaration now validates the identity itself. Readers the type
checker found move to the owner accessors; untyped reads in verify scripts,
the efficiency harness and test matchers move to `subject.identity`. The
`evidence-subject-declaration-read` text guard and its baseline are deleted.

Fixes #11473
Refs #11417

BREAKING CHANGE: the Evidence subject no longer has `digest`, `format` or
`architecture`; read `subject.identity.sha256`, `subject.identity.format` and
`subject.identity.architecture`. A record that still carries them is refused.

* test(evidence): give the Flutter verifier fixture the identity subject and check the DOL declaration without the ID

* test(evidence): drop the removed subject members from the declaration matrix and tamper fixtures

* feat(evidence)!: give every Evidence record an explicit subject identity (#11536)

* feat(evidence)!: give every Evidence record an explicit subject identity

A producer that cannot name a file passes an `unknown` target identity
with a closed reason (derived-from-linked-evidence, not-observed,
unsupported-subject) instead of no subject. The subject is non-null;
unknown subjects have no display facts. The generic missing-subject
limitation and `subjectUnavailableReason` are removed; producer-specific
explanations are ordinary limitations. Comparison identity requires the
derived unknown, and null-subject consumers ask the owner accessors.

Refs #11417

BREAKING CHANGE: every Evidence record has a subject. A record without a
file target names `subject.identity` `{ "kind": "unknown", "reason": … }`
with `name` and `local_path` null, instead of `subject: null` and the
limitation "Artifact identity is unavailable for this observation.".
Their Evidence IDs change.

* fix(evidence): name explicit unknown subjects in JavaScript callers and keep a managed graph's file subject

The verification scripts are not type-checked, so five createEvidence
callers still passed undefined (or a missing first target) after the subject
became required. They now pass unknownTarget("not-observed").

A managed application graph chose the first present input subject. Every
subject is now present, so an unknown artifact subject hid a file subject on
the member or boundary Evidence. The graph now takes the first input subject
that names a file.

* test(evidence): assert the subject kind instead of a presence the type now guarantees

Every Evidence record has a subject after this change, so test guards and
assertions that checked it for null always passed. Each now checks that the
fixture subject names a file, or is removed where a later fileIdentity,
fileSubjectTarget or fileSubjectDisplay call already requires that. Non-null
assertions on Evidence subjects are removed; analysis-view inputs, whose
subject is optional, keep an undefined check.

* test(evidence): expect the Evidence subject format under its identity

Two boundary tests still matched subject.format, which slice 2e moved into
subject.identity. toMatchObject is not type-checked against the Evidence
shape, so they failed only when run (the signature test on macOS).
2026-11-16 06:28:01 +08:00

345 lines
20 KiB
Python

"""Evaluate one real Codex CLI task with an isolated REA MCP registration."""
import argparse
import hashlib
import json
import os
from pathlib import Path
import shutil
import signal
import subprocess
import tempfile
import time
from codex_context import capture_tool_context
from runtime_identity import file_digest, runtime_fingerprint
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--mode', choices=['direct', 'client-code'], default='client-code',
help='client-code enables code_mode_host; direct is a host-disabled diagnostic')
parser.add_argument('--runtime-root', type=Path, required=True)
parser.add_argument('--revision', required=True)
parser.add_argument('--prompt', type=Path, required=True)
parser.add_argument('--artifact-path', help='Replace {{artifact_path}} in a question template')
parser.add_argument('--output', type=Path, required=True)
parser.add_argument('--model', default='gpt-5.6-terra')
parser.add_argument('--reasoning', default='high')
parser.add_argument('--timeout', type=int, default=900)
parser.add_argument('--codex', default='/usr/bin/codex')
parser.add_argument('--providers', type=Path)
parser.add_argument('--enabled-tools', type=Path)
parser.add_argument('--catalog-load', action='store_true',
help='Measure a no-tool response with identical instructions across catalogs')
parser.add_argument('--no-mcp', action='store_true',
help='No-server control; requires --catalog-load')
parser.add_argument('--profile-parent', type=Path,
help='External disposable-profile parent (default: cache beside the source Codex home)')
parser.add_argument('--authority', choices=['static', 'browser'], default='static')
parser.add_argument('--include-account-tools', action='store_true',
help='Allow account apps/plugins; default isolates REA from their catalogs')
parser.add_argument('--capture-tool-context', action='store_true',
help='Retain owned tool/usage items; excludes account/ambient state')
parser.add_argument('--guide-enabled-tools', action='store_true',
help='Opt-in exact-name discovery guidance for selected REA tools')
args = parser.parse_args()
if args.no_mcp and not args.catalog_load:
parser.error('--no-mcp requires --catalog-load')
if args.catalog_load and (args.guide_enabled_tools or args.artifact_path):
parser.error('--catalog-load excludes tool guidance and artifact tasks')
if args.guide_enabled_tools and args.enabled_tools is None:
parser.error('--guide-enabled-tools requires --enabled-tools')
enabled_tools = None
if args.enabled_tools:
enabled_tools = json.loads(args.enabled_tools.read_text())
if not isinstance(enabled_tools, list) or not all(isinstance(name, str) and name for name in enabled_tools) or len(enabled_tools) != len(set(enabled_tools)):
raise ValueError('enabled-tools must be a unique list of exact tool names')
if args.guide_enabled_tools and not enabled_tools:
parser.error('--guide-enabled-tools requires a nonempty selected tool list')
def terminate(signum, _frame):
# The outer resource scope terminates descendants together. Ensure the
# runner's copied credentials are also removed on a resource abort.
raise SystemExit(128 + signum)
signal.signal(signal.SIGTERM, terminate)
if not args.runtime_root.is_absolute() or not args.output.is_absolute():
raise ValueError('Runtime and output paths must be absolute')
repository = args.runtime_root
driver_repository = Path(__file__).resolve().parents[3]
experiment = Path(__file__).resolve().parent
node = shutil.which('node')
expected_node = (repository / '.nvmrc').read_text().strip().removeprefix('v')
if subprocess.check_output([node, '--version'], text=True).strip() != 'v' + expected_node:
raise SystemExit('Use runtime-pinned Node ' + expected_node + ' on PATH')
provider_env = {'REA_LOG_LEVEL': 'silent', 'NODE_OPTIONS': '--max-old-space-size=512'}
if args.providers:
declared = json.loads(args.providers.read_text())
if not isinstance(declared, dict) or not all(
isinstance(value, str) for value in declared.values()
):
raise ValueError('--providers must contain an object of explicit string settings')
provider_env.update(declared)
provider_policy_text = (experiment / 'provider-environment.json').read_text()
provider_policy = json.loads(provider_policy_text)
undeclared = sorted(key for key in os.environ if (
any(key.startswith(prefix) for prefix in provider_policy['namespace_prefixes'])
or key in provider_policy['additional_keys']
) and key not in provider_env)
if undeclared:
raise ValueError('Declare inherited provider settings explicitly in --providers: '
+ ', '.join(undeclared) + '. Values are not logged.')
source_auth_home = (Path(os.environ['CODEX_HOME']) if 'CODEX_HOME' in os.environ
else Path.home() / '.codex')
source_auth = source_auth_home / 'auth.json'
args.output.mkdir(parents=True, exist_ok=False, mode=0o700)
runtime_files, runtime_digest = runtime_fingerprint(repository)
(args.output / 'runtime-files.json').write_text(json.dumps(runtime_files))
# Output artifacts may live inside a checkout. Execution must not: Codex
# discovers ancestor AGENTS.md and skills independently of its user config.
profile_parent = (args.profile_parent or source_auth_home.parent / '.cache' / 'rea-agent-trials').resolve()
if any(profile_parent.is_relative_to(root.resolve()) for root in [repository, driver_repository]):
raise ValueError('--profile-parent must be outside the runtime and driver checkouts')
profile_parent.mkdir(parents=True, exist_ok=True, mode=0o700)
profile = Path(tempfile.mkdtemp(prefix='profile-', dir=profile_parent))
trial_codex_home = profile / 'codex'
account = profile / 'account'
workspace = profile / 'workspace'
for directory in [trial_codex_home, account, workspace]:
directory.mkdir(mode=0o700)
auth = trial_codex_home / 'auth.json'
started = time.monotonic()
timeout = False
exit_code = None
environment = dict(os.environ)
environment.update({
'CODEX_HOME': str(trial_codex_home), 'HOME': str(account),
'XDG_CONFIG_HOME': str(account / '.config'),
'XDG_CACHE_HOME': str(account / '.cache'),
'XDG_DATA_HOME': str(account / '.local/share'),
'NODE_OPTIONS': '--max-old-space-size=512',
})
target_digest = file_digest(Path(args.artifact_path)) if args.artifact_path else None
command_args = [str(repository / 'scripts/rea.mjs'), 'mcp']
instruction = (
'Use the registered REA MCP server for this investigation. Static analysis '
'only: do not execute target code, fetch URLs, launch a GUI, or modify targets. '
'Preserve observations, derivations, inferences, limitations, and real Evidence IDs. '
'An Evidence value has evidence_id, normalized_result and limitations; '
'non-Evidence session values have result. '
'The analyzed target digest is Evidence.subject.identity.sha256. '
'normalized_result.engine.artifact_sha256 identifies the analysis engine '
'artifact and must never be substituted for the target digest. '
'Prefer summary/focused views where appropriate. Do not invent successful facts '
'after a tool error. No other agents. Stop once the requested facts are supported.'
)
if args.authority == 'browser':
instruction = (
'Use the registered REA tools for the explicitly selected public browser scenario. '
'Launching the selected headless browser with a fresh empty profile and observing '
'the supplied public URL is authorized. Keep all actions within that scenario. '
'No injected code, credentials, unrelated targets, shell/file tools or other agents. '
'Preserve real Evidence IDs, observation completeness, failed captures and limitations. '
'Do not infer original source or global behavior from one observed session.'
)
if args.mode == 'client-code':
instruction += (
' Use the client JavaScript executor to compose registered MCP calls when '
'possible and select the requested normalized_result fields before returning '
'intermediate output to model context. Keep real Evidence IDs and limitations.'
)
if args.guide_enabled_tools:
instruction += (
' Configured REA tool function names for this trial: '
+ json.dumps(['mcp__rea__' + name for name in enabled_tools])
+ '. For discovery select only these exact names from ALL_TOOLS and '
'return only their metadata. Preserve the original analyst question, '
'dependency order and source Evidence.'
)
if args.catalog_load:
instruction = (
'This is a catalog loading measurement. Do not call any tools, search, '
'or delegate. Reply with exactly the text requested by the user.'
)
quote = json.dumps
# A disposable CODEX_HOME does not disable account-connected apps. Keep
# the requested feature policy identical in the config and its receipt.
features = {
'multi_agent': False, 'shell_snapshot': False, 'shell_tool': False,
'unified_exec': False, 'code_mode_host': args.mode == 'client-code',
'apps': args.include_account_tools, 'plugins': args.include_account_tools,
}
config = '\n'.join([
'approval_policy = "never"', 'sandbox_mode = "read-only"',
'web_search = "disabled"', 'model = ' + quote(args.model),
'model_reasoning_effort = ' + quote(args.reasoning),
'developer_instructions = ' + quote(instruction),
'[features]',
*[name + ' = ' + str(enabled).lower() for name, enabled in features.items()],
'[history]', 'persistence = "none"',
]) + '\n'
if not args.no_mcp:
config += '\n'.join([
'[mcp_servers.rea]', 'required = true', 'command = ' + quote(node),
'args = ' + quote(command_args), 'startup_timeout_sec = 60',
'tool_timeout_sec = 600',
*([] if enabled_tools is None else ['enabled_tools = ' + quote(enabled_tools)]),
'[mcp_servers.rea.env]',
*[key + ' = ' + quote(value) for key, value in provider_env.items()],
]) + '\n'
# MCP read-only hints cover normal static tools. Explicitly authorize only
# the session-opening primitive needed by the requested native tasks.
approved_tools = ['open_binary', 'analyze_javascript_application', 'inspect_analysis_view',
'inspect_android_package', 'search_android_classes',
'inspect_android_class', 'inspect_android_method',
'inspect_firmware_regions', 'inspect_managed_artifact',
'inspect_sqlite_database', 'inspect_pcap_capture',
'inspect_web_network_capture',
'inspect_packet_capture', 'inspect_packet_fields',
'decode_signal_recording', 'inspect_signal_decoders',
'inspect_managed_members', 'inspect_managed_native_boundaries',
'inspect_artifact', 'inspect_wasm_artifact',
'compare_functions', 'build_call_path', 'compare_artifacts',
'compare_process_captures', 'find_changed_behavior',
'compare_managed_members', 'project_managed_application_graph',
'compare_wasm_interfaces', 'project_apple_application_graph',
'project_android_application_graph', 'project_harmony_application_graph',
'trace_application_feature', 'trace_javascript_semantics',
'compare_application_versions', 'compare_javascript_export_shapes',
'reconcile_javascript_runtime',
'inspect_evm_interface', 'get_evidence_bundle', 'export_evidence_bundle',
'import_evidence_bundle', 'list_unknowns', 'verify_unknown_resolution',
'record_unknown', 'update_unknown', 'reset_investigation', 'binary_session',
'capture_browser_scenario',
'resolve_containing_procedure', 'analyze_function', 'close_binary']
if not args.no_mcp:
for tool in approved_tools:
config += '[mcp_servers.rea.tools.' + tool + ']\napproval_mode = "approve"\n'
try:
if source_auth.exists():
shutil.copyfile(source_auth, auth)
auth.chmod(0o600)
(trial_codex_home / 'config.toml').write_text(config)
(trial_codex_home / 'config.toml').chmod(0o600)
prompt = args.prompt.read_text()
template_digest = hashlib.sha256(prompt.encode()).hexdigest()
if '{{artifact_path}}' in prompt:
if not args.artifact_path or not os.path.isabs(args.artifact_path):
raise ValueError('A question template requires an absolute --artifact-path')
prompt = prompt.replace('{{artifact_path}}', args.artifact_path)
elif args.artifact_path:
raise ValueError('--artifact-path requires the literal template marker')
(args.output / 'prompt.txt').write_text(prompt)
(args.output / 'trial-settings.json').write_text(json.dumps({
'mode': args.mode, 'model': args.model, 'reasoning': args.reasoning,
'capture_tool_context': args.capture_tool_context,
'guide_enabled_tools': args.guide_enabled_tools,
'catalog_load': args.catalog_load, 'mcp_enabled': not args.no_mcp,
'driver_commit': subprocess.check_output(
['git', 'rev-parse', 'HEAD'], cwd=driver_repository, text=True).strip(),
'runtime_revision': args.revision,
'runtime_root': str(repository),
'workspace': str(workspace),
'workspace_isolated_from_runtime': not workspace.is_relative_to(repository.resolve()),
'selected_artifact_path': args.artifact_path,
'selected_artifact_sha256_before': target_digest,
'enabled_tools': enabled_tools,
'authority': args.authority,
'runtime_files_sha256': runtime_digest,
'runtime_dispatcher_sha256': file_digest(repository / 'scripts/rea.mjs'),
'package_lock_sha256': file_digest(repository / 'package-lock.json'),
'experiment_sources_sha256': {
name: hashlib.sha256((experiment / name).read_bytes()).hexdigest()
for name in ['agent-trial.py', 'codex_context.py', 'runtime_identity.py', 'provider-environment.json']
},
'developer_instructions': instruction,
'developer_instructions_sha256': hashlib.sha256(instruction.encode()).hexdigest(),
'prompt_sha256': hashlib.sha256(prompt.encode()).hexdigest(),
'question_template_sha256': template_digest,
'provider_settings_sha256': hashlib.sha256(
json.dumps(provider_env, sort_keys=True).encode()).hexdigest(),
'provider_search_path_sha256': hashlib.sha256(
provider_env.get('PATH', os.environ.get('PATH', '')).encode()).hexdigest(),
'provider_environment_policy_sha256': hashlib.sha256(
provider_policy_text.encode()).hexdigest(),
'features': features,
}, indent=2) + '\n')
invocation = [args.codex, '--no-daemon', 'exec', '--strict-config',
*([] if args.capture_tool_context else ['--ephemeral']),
'--skip-git-repo-check', '--ignore-rules', '--json',
'--color', 'never', '-C', str(workspace),
'--output-last-message', str(args.output.resolve() / 'answer.txt'), '-']
with (args.output / 'events.jsonl').open('w') as events, (args.output / 'stderr.log').open('w') as errors:
child = subprocess.Popen(invocation, env=environment, cwd=workspace,
stdin=subprocess.PIPE, stdout=events, stderr=errors, text=True)
child.stdin.write(prompt)
child.stdin.close()
try:
exit_code = child.wait(timeout=args.timeout)
except subprocess.TimeoutExpired:
timeout = True
child.terminate()
try:
exit_code = child.wait(timeout=10)
except subprocess.TimeoutExpired:
child.kill()
exit_code = child.wait()
usage = []
tools = []
event_types = {}
with (args.output / 'events.jsonl').open() as event_file:
for line in event_file:
if not line.strip():
continue
event = json.loads(line)
kind = event.get('type', 'unknown')
event_types[kind] = event_types.get(kind, 0) + 1
if 'usage' in event:
usage.append(event['usage'])
item = event.get('item', {})
if item.get('type') == 'mcp_tool_call' and kind == 'item.completed':
tools.append({
**{k: item.get(k) for k in ['server', 'tool', 'arguments', 'status', 'error']},
'codex_event_result_json_utf8_bytes': len(json.dumps(
item.get('result'), separators=(',', ':'), ensure_ascii=False
).encode()),
})
final_target_digest = file_digest(Path(args.artifact_path)) if args.artifact_path else None
try:
_, final_runtime_digest = runtime_fingerprint(repository)
except OSError:
final_runtime_digest = None
summary = {
'mode': args.mode, 'requested_model': args.model, 'reasoning': args.reasoning,
'runtime_revision': args.revision,
'runtime_files_sha256_before': runtime_digest,
'runtime_files_sha256_after': final_runtime_digest,
'runtime_bytes_unchanged': runtime_digest == final_runtime_digest,
'catalog_load': args.catalog_load, 'mcp_enabled': not args.no_mcp,
'capture_tool_context_requested': args.capture_tool_context,
'selected_artifact_sha256_before': target_digest,
'selected_artifact_sha256_after': final_target_digest,
'selected_artifact_bytes_unchanged': target_digest == final_target_digest if target_digest else None,
'model_identity_scope': 'requested alias; resolved server-side model identity is not inferred',
'codex_version': subprocess.check_output([args.codex, '--version'], text=True).strip(),
'exit_code': exit_code, 'timeout': timeout,
'wall_seconds': round(time.monotonic() - started, 3),
'usage': usage, 'completed_mcp_calls': tools, 'event_types': event_types,
'factual_correctness': 'not_assessed',
'scope': 'real_cli_trial_requires_transcript_and_fact_review',
}
(args.output / 'summary.json').write_text(json.dumps(summary, indent=2) + '\n')
print(json.dumps(summary))
finally:
try:
if args.capture_tool_context:
capture_tool_context(trial_codex_home, args.output)
finally:
# Capture failure must never prevent copied-credential cleanup.
shutil.rmtree(profile)
(args.output / 'cleanup.json').write_text(json.dumps({'profile_removed': not profile.exists()}) + '\n')
raise SystemExit(124 if timeout else 125 if final_runtime_digest != runtime_digest else exit_code)
if __name__ == '__main__':
main()