项目文件夹

文件
Gelei Deng ab5fbb4d90 feat: ship the durable multi-model autonomous PentestGPT runtime (#493)
* first refactor

* feat: dockerized tool with persistent Claude+Codex login + multi-model benchmark

Run the autonomous CTF/pentest tool in Docker with a one-time, persistent login for
BOTH Claude Code and Codex, and add a multi-model benchmark harness.

Backend (multi-model):
- Add `--backend {claude,codex}` to the CTF pipeline. CodexBackend (pentestgpt/core/
  backend.py) wraps unified_agent's Codex backend and translates its events into
  AgentMessages, so the same pipeline runs on Claude (opus/sonnet) or Codex
  (gpt-5.5/gpt-5.4-mini). Wired through config.backend, pipeline stage construction,
  and the CLI (+ PENTESTGPT_CODEX_EFFORT; greppable [CODEX_USAGE] under PENTESTGPT_BENCH=1).

Docker tool (tool-only image; the benchmark stays OUTSIDE the image):
- Extend Dockerfile: Codex CLI (@openai/codex) + openai_codex SDK + unified_agent/
  pentestgpt_agent/pentestgpt_legacy packages + gobuster/dirb + socat. Add .dockerignore
  (keeps creds/benchmark/workspace out of the build context).
- Persistent dual login (the hard part) — asymmetric by token model:
  * Claude: `setup-token` -> token stored in the pentestgpt-claude volume; entrypoint
    exports CLAUDE_CODE_OAUTH_TOKEN (setup-token does not write .credentials.json; macOS
    host creds live in the Keychain and can't be copied).
  * Codex: the container does its OWN `codex login` (NOT seeding -- ChatGPT refresh tokens
    are single-use, so a shared/copied login 401s on first refresh). The 127.0.0.1:1455
    OAuth callback is forwarded into the container via a socat hop (-p 1455:8455).
  * scripts/docker-login.sh is idempotent: checks logins live, logs in only the missing one(s).
- docker-compose codex-config volume (+ pinned names); entrypoint token-export + non-blocking
  preflight; scripts/docker-auth-status.sh; Make targets (docker-build/login/auth-status/
  run/shell/down/nuke).
- Verified end-to-end: one `make docker-login` -> a fresh container reports claude+codex
  logged in with live round-trips; the CTF pipeline (Codex) captured a flag against an
  isolated fixture and the pentest pipeline ran cleanly; persists across recreation, no re-login.

Benchmark (multi-model, host-side):
- benchmark/pilot/ harness (run_pilot.py + report.py): builds each xbow challenge, discovers
  the loopback port, runs the pipeline across the 4 model combos, judges by the baked
  FLAG{sha256(UPPER-dir)}, and renders REPORT.md (infra failures excluded from solve rates).
  Includes the partial pilot's results (results.jsonl + REPORT.md).

Docs: docs/docker-dev-plan.md (full plan + implementation status); CLAUDE.md and README
docker quickstart; benchmark/pilot/README.md; design-doc roadmap (docs/redesign).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix: fail controller on backend error messages

* fix: allow listing sessions without target

* docs: add docker xbow benchmark report

* fix: infer concrete backend constructor type

* docs: refresh docker benchmark documentation

* feat(benchmark): add pure single-agent baseline + pipeline comparison

Add a "pure single agent" benchmark variant -- one bare `claude -p` /
`codex exec` call per target (no pipeline) -- to quantify what the 3-stage
PentestGPT pipeline buys over an un-orchestrated agent on the xbow targets.

- pentestgpt/prompts/stages.py: ctf_single_agent_{system,task}_prompt -- the
  pipeline's shared fragments collapsed into ONE turn, so prompt content is
  held constant and the only variable is the multi-stage decomposition.
- benchmark/pilot/run_docker_bench.py: docker-network runner
  (--variant single|pipeline). Brings the target up, discovers the container's
  internal IP+network (skips DB side-cars/ports), docker-runs the tool image on
  that network, and scores the ground-truth flag against the agent's *assistant
  text* only (parity with the pipeline's raw streaming). Reads stdout in chunks
  to handle >64KB JSON lines. Resumable; --dry-run supported.
- benchmark/pilot/report_comparison.py -> DOCKER_COMPARISON.md: head-to-head
  pipeline-vs-single per model on the common non-infra set.
- tests/unit/test_single_agent_prompt.py: prompt-builder coverage.
- docs: README, CLAUDE.md, benchmark README, DOCKER_REPORT updated.

Recorded result (10 medium/hard targets x 4 models, container-to-container,
same baseline image digest 0c4c0f3e..., commit dca0019 image):

  Model               Pipeline   Single
  Claude Opus           5/10      7/10   (single +2)
  Claude Sonnet         6/10      4/10   (pipeline +2)
  Codex gpt-5.5         7/10      7/10   (tie)
  Codex gpt-5.4-mini    3/10      4/10   (single +1)
  TOTAL                21/40     22/40

Single agent matches the pipeline on solve rate (55% vs 52%) while using
~40% fewer Codex tokens (13.0M vs 21.8M) and solving faster. The pipeline
only clearly helps Claude Sonnet (which times out solo); Opus is better solo.
Full per-challenge grid in DOCKER_COMPARISON.md; raw records in
docker_single_results.jsonl.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* feat(benchmark): add pentestgpt_agent docker harness

* bench: refresh pentestgpt_agent smoke result

* fix(benchmark): make repeat rows variant-aware

* fix(agent): fall back for semantic executor labels

* fix(agent): tolerate executor prose evidence

* fix(benchmark): score accepted framework findings

* bench: append partial framework repeat results

* bench: complete framework repeat sweep

* bench: expose framework executor concurrency

* bench: add extended parallel framework sweep

* checkpoint: preserve working agent and benchmark state

* feat: harden durable agent loop and xbow qualification

* fix: reserve an exploit result turn

* docs: record clean xbow qualification

* build: consume unified-agent from the git wrapper repo

Repoint pentestgpt_agent_new's unified-agent dependency from the local
editable path (../../UnifiedAgentPoC, now renamed and gone) to the pinned
git source PentestGPT-Project/UnifedAgentWrapper@d05d21f. Regenerate uv.lock
and update test_dependency.py to assert the external package is installed
from that VCS URL (not the repo-root vendored copy) at version 0.2.0.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* refactor: make pentestgpt_agent_new the sole framework

Remove the retired ledger-based pentestgpt_agent package (instructor/executor/
judge) and its orphaned unit + smoke tests. The nested pentestgpt_agent_new
project (Supervisor/Executor over a durable SQLite loop, consuming unified-agent
from the git wrapper) is now the single maintained framework.

Repoint the top-level tooling to it:
- pyproject: drop the pentestgpt-agent console script and pentestgpt_agent from
  the wheel packages.
- Makefile: lint/format target parent code only; typecheck/check/ci now run the
  nested framework's own gate (ruff, format, mypy, pytest) via test-agent-new /
  check-agent-new, so `make check` finally covers it; `make run` delegates to the
  pentestgpt-agent-new CLI.
- Dockerfile: stop copying the removed package (kept the build working); note the
  framework is not baked into the image yet.
- docker container-health test: import the substrate packages that actually ship.
- CLAUDE.md / AGENT.md: describe the new framework, the git-sourced wrapper, and
  the deprioritized benchmark/Docker rewire.

The XBOW `--variant framework` path and docker-bench Makefile targets still point
at the old in-image framework and are left as a pending rewire (benchmarks
deprioritized); the naive `--variant single` path is unaffected.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* refactor: rename pentestgpt_agent_new -> pentestgpt_agent

The framework reclaims the clean name now that the old ledger-based package is
gone. Rename the nested project folder, its src package, the distribution
(pentestgpt-agent-new -> pentestgpt-agent) and CLI, and every import/reference in
the package, the umbrella Makefile, the Dockerfile, the docker health test, and
CLAUDE.md / AGENT.md. Regenerate uv.lock. The audit CLI stays pentestgpt-agent-audit;
the git-sourced unified-agent dependency is unchanged. `make check` is green
(108 nested tests). The two historical *_REPORT.md files keep the old name as
dated records.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* chore: extract benchmark harness to sibling xbow-benchmark repo

Move PentestGPT/benchmark/ out to ../xbow-benchmark (its own repo) to keep this
project clean. The harness was decoupled from the framework code (it scores
container output, never imports pentestgpt_agent/unified_agent), so only
operational ties remain and they now live in the sibling repo.

- Remove benchmark/ and the 4 harness unit tests (relocated + repointed there).
- Strip the docker-bench-*/bench-* targets and their config vars from the
  Makefile; keep the tool-image lifecycle (docker-build/login/run/...) and add a
  help pointer to `make -C ../xbow-benchmark help`.

The sibling repo mounts this checkout read-only (--source-root ../PentestGPT) and
runs the pentestgpt:latest image built here.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat: harden autonomous framework and runtime integration

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-13 16:49:08 +08:00

292 行
11 KiB
Python

"""Codex backend via the official openai-codex SDK (drives `codex app-server`).
The shared MCP tool server is injected per-client through `-c`-style config
overrides (`mcp_servers.<name>.*`), the documented mechanism mirroring the CLI.
Codex does NOT inherit the parent environment into MCP server processes, so the
ToolServerSpec env (PYTHONPATH etc.) is passed explicitly as a TOML inline table.
Usage note: turn token usage arrives via `thread/tokenUsage/updated`
notifications (the `turn/completed` payload carries no usage), so the
normalizer tracks the latest usage statefully.
"""
from __future__ import annotations
import contextlib
import inspect
import json
from collections.abc import AsyncIterator, Iterator
from typing import Any
from openai_codex import AsyncCodex, CodexConfig, Sandbox
from ..events import (
AgentEvent,
AssistantText,
CommandRun,
FileChanged,
RawEvent,
Reasoning,
SessionStarted,
TextDelta,
ToolCall,
ToolResult,
TurnCompleted,
)
from ..types import (
AgentAuthError,
AgentRunError,
RunOptions,
SandboxPolicy,
ToolServerSpec,
UnifiedUsage,
)
SANDBOX_MAP = {
SandboxPolicy.READ_ONLY: Sandbox.read_only,
SandboxPolicy.WORKSPACE_WRITE: Sandbox.workspace_write,
SandboxPolicy.FULL_ACCESS: Sandbox.full_access,
}
def _toml_string(value: str) -> str:
# A JSON string is a valid TOML basic string (same escaping rules for
# quotes/backslashes/control chars).
return json.dumps(value)
def build_config_overrides(spec: ToolServerSpec) -> tuple[str, ...]:
"""`-c key=value` style overrides defining the shared MCP server."""
key = f"mcp_servers.{spec.server_name}"
env_table = (
"{ " + ", ".join(f'"{k}" = {_toml_string(v)}' for k, v in sorted(spec.env.items())) + " }"
)
return (
f"{key}.command={_toml_string(spec.command[0])}",
f"{key}.args={json.dumps(spec.command[1:])}", # JSON array == TOML array of strings
f"{key}.env={env_table}",
f"{key}.startup_timeout_sec=30",
f"{key}.tool_timeout_sec=120",
f"{key}.required=true",
f'{key}.default_tools_approval_mode="auto"',
)
def build_turn_kwargs(opts: RunOptions) -> dict[str, Any]:
"""Per-turn kwargs for Thread.turn()/run(): output schema + reasoning effort."""
kwargs: dict[str, Any] = {"output_schema": opts.output_schema}
if opts.effort:
from openai_codex.generated.v2_all import ReasoningEffort
try:
kwargs["effort"] = ReasoningEffort(opts.effort)
except ValueError as e:
valid = ", ".join(m.value for m in ReasoningEffort)
raise AgentRunError(f"invalid effort {opts.effort!r} for codex (valid: {valid})") from e
return kwargs
def build_thread_kwargs(opts: RunOptions) -> dict[str, Any]:
kwargs: dict[str, Any] = {
"cwd": str(opts.workspace),
"sandbox": SANDBOX_MAP[opts.sandbox],
}
if opts.model:
kwargs["model"] = opts.model
if opts.instructions:
kwargs["developer_instructions"] = opts.instructions
return kwargs
class _TurnState:
"""Mutable accumulator across one turn's notification stream."""
def __init__(self, opts: RunOptions):
self.opts = opts
self.thread_id: str | None = None
self.session_announced = False
self.last_agent_text: str | None = None
self.usage: Any | None = None # latest ThreadTokenUsage
self.errors: list[str] = []
def _status_str(status: Any) -> str:
return str(getattr(status, "value", status))
def _stringify_mcp_result(result: Any) -> str:
if result is None:
return ""
structured = getattr(result, "structured_content", None)
if structured:
try:
return json.dumps(structured)
except TypeError:
return str(structured)
parts = []
for block in getattr(result, "content", None) or []:
text = getattr(block, "text", None)
parts.append(text if text is not None else str(block))
return "\n".join(parts)
def _normalize_item(root: Any, state: _TurnState) -> Iterator[AgentEvent]:
kind = getattr(root, "type", None)
if kind == "agentMessage":
text = getattr(root, "text", "") or ""
state.last_agent_text = text
yield AssistantText(text=text)
elif kind == "reasoning":
chunks = list(getattr(root, "summary", None) or []) or list(
getattr(root, "content", None) or []
)
text = "\n".join(str(c) for c in chunks)
if text:
yield Reasoning(text=text)
elif kind == "commandExecution":
yield CommandRun(
command=getattr(root, "command", "") or "",
exit_code=getattr(root, "exit_code", None),
output=getattr(root, "aggregated_output", "") or "",
)
elif kind == "mcpToolCall":
server = getattr(root, "server", "") or ""
tool = getattr(root, "tool", "") or ""
arguments = getattr(root, "arguments", None)
if isinstance(arguments, str):
with contextlib.suppress(json.JSONDecodeError, ValueError):
arguments = json.loads(arguments)
call_id = getattr(root, "id", None)
yield ToolCall(name=f"mcp__{server}__{tool}", input=arguments, call_id=call_id)
error = getattr(root, "error", None)
failed = _status_str(getattr(root, "status", "")) != "completed" or error is not None
output = (
getattr(error, "message", None) or str(error)
if error is not None
else _stringify_mcp_result(getattr(root, "result", None))
)
yield ToolResult(call_id=call_id, output=output or "", is_error=failed)
elif kind == "fileChange":
for change in getattr(root, "changes", None) or []:
yield FileChanged(
path=str(getattr(change, "path", change)),
kind=str(getattr(change, "kind", "update")),
)
else:
yield RawEvent(backend="codex", kind=f"item:{kind}", data=root)
def _usage_from_state(state: _TurnState) -> UnifiedUsage:
total = getattr(state.usage, "total", None)
if total is None:
return UnifiedUsage()
return UnifiedUsage(
input_tokens=getattr(total, "input_tokens", 0) or 0,
cached_input_tokens=getattr(total, "cached_input_tokens", 0) or 0,
output_tokens=getattr(total, "output_tokens", 0) or 0,
reasoning_output_tokens=getattr(total, "reasoning_output_tokens", 0) or 0,
)
def normalize_notification(method: str, payload: Any, state: _TurnState) -> Iterator[AgentEvent]:
if method == "thread/started":
thread = getattr(payload, "thread", None)
thread_id = getattr(thread, "id", None)
if thread_id:
state.thread_id = thread_id
if not state.session_announced and state.thread_id:
state.session_announced = True
yield SessionStarted(session_id=state.thread_id)
elif method == "item/completed":
item = getattr(payload, "item", None)
root = getattr(item, "root", item)
if root is not None:
yield from _normalize_item(root, state)
elif method == "item/agentMessage/delta":
if state.opts.stream_text:
yield TextDelta(text=getattr(payload, "delta", "") or "")
elif method == "thread/tokenUsage/updated":
state.usage = getattr(payload, "token_usage", None)
elif method == "error":
error = getattr(payload, "error", None)
message = getattr(error, "message", None) or str(error)
state.errors.append(message)
yield RawEvent(backend="codex", kind="error", data=message)
elif method == "turn/completed":
turn = getattr(payload, "turn", None)
status = _status_str(getattr(turn, "status", ""))
success = status == "completed"
turn_error = getattr(turn, "error", None)
error_parts = []
if turn_error is not None:
error_parts.append(getattr(turn_error, "message", None) or str(turn_error))
error_parts.extend(state.errors)
structured = None
if state.opts.output_schema and state.last_agent_text:
try:
structured = json.loads(state.last_agent_text)
except (json.JSONDecodeError, ValueError):
structured = None
duration = getattr(turn, "duration_ms", None)
yield TurnCompleted(
success=success,
final_text=state.last_agent_text,
usage=_usage_from_state(state),
cost_usd=None, # Codex reports tokens only
session_id=state.thread_id,
duration_ms=duration,
structured_output=structured,
stop_reason=status,
error=None if success else ("; ".join(p for p in error_parts if p) or status),
)
elif method in ("turn/started", "item/started", "item/updated"):
pass # uninteresting transitions; item/completed carries the substance
else:
yield RawEvent(backend="codex", kind=method, data=payload)
def _looks_like_auth_error(exc: Exception) -> bool:
text = str(exc).lower()
return any(s in text for s in ("auth", "login", "unauthorized", "api key", "logged out"))
class CodexBackend:
name = "codex"
async def stream(self, prompt: str, opts: RunOptions) -> AsyncIterator[AgentEvent]:
overrides = build_config_overrides(opts.tool_server) if opts.tool_server else ()
config = CodexConfig(config_overrides=tuple(overrides), cwd=str(opts.workspace))
state = _TurnState(opts)
try:
async with AsyncCodex(config) as codex:
if opts.resume:
thread = await codex.thread_resume(opts.resume, **build_thread_kwargs(opts))
else:
thread = await codex.thread_start(**build_thread_kwargs(opts))
state.thread_id = getattr(thread, "id", None)
if state.thread_id and not state.session_announced:
state.session_announced = True
yield SessionStarted(session_id=state.thread_id)
handle = thread.turn(prompt, **build_turn_kwargs(opts))
if inspect.isawaitable(handle):
handle = await handle
async for note in handle.stream():
method = getattr(note, "method", "")
payload = getattr(note, "payload", None)
for event in normalize_notification(method, payload, state):
yield event
except (AgentRunError, AgentAuthError):
raise
except Exception as e: # JSON-RPC / transport errors from the SDK
if _looks_like_auth_error(e):
raise AgentAuthError(f"Codex is not authenticated (run `codex login`): {e}") from e
raise AgentRunError(f"Codex run failed: {type(e).__name__}: {e}") from e