Files
Gitea-Tools/mcp_tool_error_boundary.py
T
sysadminandClaude Opus 4.8 7e5eca08b3 fix(mcp): surface safe diagnostics for opaque mutation internal_error
Mutation tools (create_issue, create_pr, delete_branch, issue-lock) returned
reason_code=internal_error / retryable=false with no class, message, HTTP
status, or actionable detail. The #699/#701 tool-error boundary intentionally
collapses every non-typed exception to a fixed "Internal tool error" and drops
the class and message, so a mutation failure is undiagnosable — and the #695
runtime guard blocks standalone reproduction. The true failing stage was
therefore invisible by two independent mechanisms.

Make the failure observable and actionable without weakening the secret-free
contract:

- Boundary (internal_error path ONLY): capture a safe exception class (a Python
  type identifier, never instance text) plus a strictly-redacted `detail`
  (token/Authorization/Bearer credentials, JSON/kv secret values incl. short
  passwords, raw URLs/hostnames, and absolute filesystem paths all stripped;
  whitespace collapsed; length-bounded) and an optional `mutation_stage`. Typed
  auth/authz/network/config/http paths are unchanged and still emit no detail.
- Convert raw exceptions into typed structured errors via a safe
  `gitea_reason_code` attribute contract: server raisers may self-declare a
  known reason and the boundary emits it (fixed message) instead of an opaque
  internal_error. Pre-flight order violations now raise `_PreflightOrderError`
  (a RuntimeError subclass → preflight_order_violation).
- Add a `_mutation_stage` context manager tagging escaping exceptions with the
  stage name (preflight_purity / resolve_auth / api_*), applied to
  delete_branch, create_issue, create_pr.

Fail-closed identity/repo/worktree/role/permission/lock/lease/ancestry gates
are untouched; only error *reporting* changed. New suite
test_mutation_error_diagnostics.py (18 tests) pins the redaction, class capture,
stage tagging, typed conversion, and adversarial no-leak contract. Existing
boundary suite (25), core server (209), and preflight (92) suites stay green.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
2026-07-17 20:42:25 -04:00

635 lines
24 KiB
Python

"""MCP tool-boundary error mapping for known Gitea client failures (#699).
Known authentication / authorization / network / configuration failures leave
the tool boundary as a sanitized structured ``CallToolResult`` with
``isError=True``. Stdio transport remains connected.
Design constraints (reviewer-ratified, #699 / PR #701):
1. **No secret material** in tool results or daemon logs. Messages are fixed
constants keyed by ``reason_code``; HTTP bodies / exception text never
surface. Sanitization failure fails closed to ``internal_error``.
2. **Narrow boundary** wraps the original FastMCP ``Tool.run`` success path;
re-raises framework control-flow exceptions (``UrlElicitationRequiredError``);
maps only typed client failures. Installation is idempotent.
3. **No RuntimeError substring heuristics.** Only explicit typed
authentication / authorization / network / configuration exceptions
receive those labels.
4. **HTTP classification** lives in ``gitea_auth.classify_http_status``;
every 403 is authorization-class.
"""
from __future__ import annotations
import json
import logging
import re
import sys
from typing import Any
logger = logging.getLogger("gitea_mcp.tool_error_boundary")
# Stable reason codes (#699).
REASON_AUTH_FAILED = "auth_failed"
REASON_AUTH_INVALID_TOKEN = "auth_invalid_token"
REASON_AUTHZ_INSUFFICIENT_SCOPE = "authz_insufficient_scope"
REASON_AUTHZ_DENIED = "authz_denied"
REASON_NETWORK_ERROR = "network_error"
REASON_CONFIG_ERROR = "config_error"
REASON_INTERNAL_ERROR = "internal_error"
REASON_UPSTREAM_UNAVAILABLE = "upstream_unavailable"
REASON_HTTP_ERROR = "http_error"
# Typed structured errors for previously-opaque raw mutation failures.
REASON_PREFLIGHT_ORDER = "preflight_order_violation"
REASON_MALFORMED_RESPONSE = "malformed_response"
ERROR_CLASS_AUTHENTICATION = "authentication"
ERROR_CLASS_AUTHORIZATION = "authorization"
ERROR_CLASS_NETWORK = "network"
ERROR_CLASS_CONFIGURATION = "configuration"
ERROR_CLASS_INTERNAL = "internal"
ERROR_CLASS_UPSTREAM = "upstream"
ERROR_CLASS_PRECONDITION = "precondition"
# Fixed, secret-free operator messages. Never interpolate HTTP bodies,
# Keychain contents, tokens, or arbitrary exception text.
FIXED_MESSAGES: dict[str, str] = {
REASON_AUTH_FAILED: "Gitea authentication failed",
REASON_AUTH_INVALID_TOKEN: (
"Gitea authentication failed: invalid or revoked credentials"
),
REASON_AUTHZ_INSUFFICIENT_SCOPE: (
"Gitea authorization failed: insufficient token scope"
),
REASON_AUTHZ_DENIED: "Gitea authorization failed: access denied",
REASON_NETWORK_ERROR: "Network error contacting Gitea",
REASON_CONFIG_ERROR: "Gitea configuration or credential resolution failed",
REASON_INTERNAL_ERROR: "Internal tool error",
REASON_UPSTREAM_UNAVAILABLE: "Gitea upstream unavailable",
REASON_HTTP_ERROR: "Gitea HTTP request failed",
REASON_PREFLIGHT_ORDER: (
"Mutation blocked: pre-flight order violation (fail closed)"
),
REASON_MALFORMED_RESPONSE: "Malformed response from Gitea",
}
# Error class for a self-declared safe reason code (``gitea_reason_code``).
_REASON_ERROR_CLASS: dict[str, str] = {
REASON_AUTH_FAILED: ERROR_CLASS_AUTHENTICATION,
REASON_AUTH_INVALID_TOKEN: ERROR_CLASS_AUTHENTICATION,
REASON_AUTHZ_INSUFFICIENT_SCOPE: ERROR_CLASS_AUTHORIZATION,
REASON_AUTHZ_DENIED: ERROR_CLASS_AUTHORIZATION,
REASON_NETWORK_ERROR: ERROR_CLASS_NETWORK,
REASON_CONFIG_ERROR: ERROR_CLASS_CONFIGURATION,
REASON_UPSTREAM_UNAVAILABLE: ERROR_CLASS_UPSTREAM,
REASON_HTTP_ERROR: ERROR_CLASS_INTERNAL,
REASON_PREFLIGHT_ORDER: ERROR_CLASS_PRECONDITION,
REASON_MALFORMED_RESPONSE: ERROR_CLASS_INTERNAL,
REASON_INTERNAL_ERROR: ERROR_CLASS_INTERNAL,
}
# Bounds for the safe diagnostic fields added to the ``internal_error`` path.
_DETAIL_LIMIT = 200
_CLASS_LIMIT = 120
_STAGE_LIMIT = 64
# Absolute-path prefixes stripped from diagnostic detail (never leak layout).
_ABS_PATH_RE = re.compile(r"/(?:Users|home|private|var|tmp|opt|etc|root)/[^\s'\"]*")
_DETAIL_SECRET_PREFIXES = ("token ", "Basic ", "Bearer ", "Authorization: ")
_SECRET_KEY = (
r"token|password|passwd|secret|authorization|api[_-]?key|"
r"access[_-]?token|refresh[_-]?token|session|cookie"
)
_SECRET_JSON_RE = re.compile(
r'"(' + _SECRET_KEY + r')"\s*:\s*"[^"]*"', re.IGNORECASE
)
_SECRET_KV_RE = re.compile(
r"\b(" + _SECRET_KEY + r")\s*=\s*[^\s&\"']+", re.IGNORECASE
)
_MUTATION_STAGE_ATTR = "_gitea_mutation_stage"
_SELF_DECLARED_REASON_ATTR = "gitea_reason_code"
_INSTALL_FLAG = "_gitea_auth_boundary_installed"
_ORIGINAL_ATTR = "_gitea_auth_boundary_original"
def _safe_str(exc: BaseException) -> str:
"""``str(exc)`` that never raises (a poisoned ``__str__`` must not escape)."""
try:
return str(exc)
except Exception:
return ""
def _safe_exception_class(exc: BaseException) -> str:
"""Fully-qualified type identifier for *exc* — never instance/message text."""
try:
t = type(exc)
mod = getattr(t, "__module__", "") or ""
name = getattr(t, "__qualname__", None) or getattr(t, "__name__", "") or ""
ident = f"{mod}.{name}" if mod else name
return ident[:_CLASS_LIMIT]
except Exception:
return "unknown"
def _redact_detail(text: Any, limit: int = _DETAIL_LIMIT) -> str:
"""Strictly redact free-form exception text for safe diagnostics.
Removes token/Authorization credentials, raw URLs and hostnames (via the
shared audit redactor), and absolute filesystem paths, then collapses
whitespace and truncates. Never raises; on any failure returns "".
"""
if not text:
return ""
try:
out = str(text)
# Drop credential-prefixed runs (token/Basic/Bearer/Authorization).
for prefix in _DETAIL_SECRET_PREFIXES:
idx = 0
while True:
i = out.find(prefix, idx)
if i == -1:
break
j = i + len(prefix)
while j < len(out) and not out[j].isspace():
j += 1
out = out[:i] + prefix + "[REDACTED]" + out[j:]
idx = i + len(prefix) + len("[REDACTED]")
# Secret VALUES carried in JSON ("key":"value") or kv (key=value) form,
# even short ones (e.g. a password), keyed by a sensitive field name.
out = _SECRET_JSON_RE.sub(r'"\1":"[REDACTED]"', out)
out = _SECRET_KV_RE.sub(r"\1=[REDACTED]", out)
# URLs / query secrets / hostnames.
try:
import gitea_audit
out = gitea_audit.redact_urls(out)
except Exception:
pass
# Absolute filesystem paths (workspace layout is not for LLM output).
out = _ABS_PATH_RE.sub("[PATH]", out)
# Any bare hostname the URL redactor missed (defence in depth).
out = re.sub(r"\b[\w.-]+\.(?:cc|net|com|org|io|dev|local)\b", "[HOST]", out)
out = " ".join(out.split())
return out[:limit]
except Exception:
return ""
def _safe_mutation_stage(exc: BaseException) -> str | None:
"""Return the bounded, redacted mutation stage tagged on *exc* (or None)."""
try:
raw = getattr(exc, _MUTATION_STAGE_ATTR, None)
if not raw:
return None
return _redact_detail(raw, limit=_STAGE_LIMIT) or None
except Exception:
return None
def fixed_message(reason_code: str) -> str:
"""Return the fixed sanitized message for *reason_code* (fail closed)."""
return FIXED_MESSAGES.get(reason_code, FIXED_MESSAGES[REASON_INTERNAL_ERROR])
def _safe_profile_name() -> str | None:
try:
from gitea_auth import get_profile
name = (get_profile() or {}).get("profile_name")
if name is None:
return None
text = str(name).strip()
# Profile names are non-secret identifiers; still bound length.
return text[:80] if text else None
except Exception:
return None
def _is_framework_control_flow(exc: BaseException) -> bool:
"""True for exceptions the MCP framework must re-raise unchanged."""
try:
from mcp.shared.exceptions import UrlElicitationRequiredError
if isinstance(exc, UrlElicitationRequiredError):
return True
except Exception:
pass
# BaseException subclasses that must never become tool isError payloads.
if isinstance(exc, (KeyboardInterrupt, SystemExit, GeneratorExit)):
return True
return False
def is_known_client_failure(exc: BaseException) -> bool:
"""True only for explicit typed Gitea client / config failures.
Never true for arbitrary ``RuntimeError`` (reviewer finding #3).
"""
try:
import gitea_auth
if isinstance(
exc,
(
gitea_auth.GiteaAuthError,
gitea_auth.GiteaAuthzError,
gitea_auth.GiteaNetworkError,
gitea_auth.GiteaConfigError,
gitea_auth.GiteaHttpError,
),
):
return True
except Exception:
return False
try:
import gitea_config
if isinstance(exc, gitea_config.ConfigError):
return True
except Exception:
pass
return False
def classify_exception(exc: BaseException) -> dict[str, Any]:
"""Return structured classification with **fixed** messages only.
Only typed authentication / authorization / network / configuration
failures receive those labels. Unexpected programming failures are
``internal_error``. HTTP bodies and exception text are never copied
into ``message``.
"""
try:
return _classify_exception_impl(exc)
except Exception:
# Fail closed: sanitization / classification failure must not leak.
return {
"reason_code": REASON_INTERNAL_ERROR,
"error_class": ERROR_CLASS_INTERNAL,
"http_status": None,
"message": fixed_message(REASON_INTERNAL_ERROR),
"transport_survives": True,
}
def _self_declared_classification(exc: BaseException) -> dict[str, Any] | None:
"""Honor a safe ``gitea_reason_code`` attribute set by server-side raisers.
Converts a raw exception that self-declares a **known** reason code into a
typed structured error (fixed message) instead of an opaque internal_error.
Unknown / non-string / secret values are ignored (fail closed to internal).
"""
reason = getattr(exc, _SELF_DECLARED_REASON_ATTR, None)
if not isinstance(reason, str) or reason not in FIXED_MESSAGES:
return None
if reason == REASON_INTERNAL_ERROR:
return None
return {
"reason_code": reason,
"error_class": _REASON_ERROR_CLASS.get(reason, ERROR_CLASS_INTERNAL),
"http_status": None,
"message": fixed_message(reason),
"transport_survives": True,
}
def _classify_exception_impl(exc: BaseException) -> dict[str, Any]:
import gitea_auth
declared = _self_declared_classification(exc)
if declared is not None:
return declared
if isinstance(exc, gitea_auth.GiteaAuthError):
code = getattr(exc, "reason_code", None) or REASON_AUTH_INVALID_TOKEN
if code not in (
REASON_AUTH_FAILED,
REASON_AUTH_INVALID_TOKEN,
):
code = REASON_AUTH_INVALID_TOKEN
return {
"reason_code": code,
"error_class": ERROR_CLASS_AUTHENTICATION,
"http_status": getattr(exc, "http_status", None) or 401,
"message": fixed_message(code),
"transport_survives": True,
}
if isinstance(exc, gitea_auth.GiteaAuthzError):
code = getattr(exc, "reason_code", None) or REASON_AUTHZ_DENIED
if code not in (REASON_AUTHZ_DENIED, REASON_AUTHZ_INSUFFICIENT_SCOPE):
code = REASON_AUTHZ_DENIED
return {
"reason_code": code,
"error_class": ERROR_CLASS_AUTHORIZATION,
"http_status": getattr(exc, "http_status", None) or 403,
"message": fixed_message(code),
"transport_survives": True,
}
if isinstance(exc, gitea_auth.GiteaNetworkError):
return {
"reason_code": REASON_NETWORK_ERROR,
"error_class": ERROR_CLASS_NETWORK,
"http_status": getattr(exc, "http_status", None),
"message": fixed_message(REASON_NETWORK_ERROR),
"transport_survives": True,
}
if isinstance(exc, gitea_auth.GiteaConfigError):
return {
"reason_code": REASON_CONFIG_ERROR,
"error_class": ERROR_CLASS_CONFIGURATION,
"http_status": getattr(exc, "http_status", None),
"message": fixed_message(REASON_CONFIG_ERROR),
"transport_survives": True,
}
if isinstance(exc, gitea_auth.GiteaHttpError):
code = getattr(exc, "reason_code", None) or REASON_HTTP_ERROR
if code == REASON_UPSTREAM_UNAVAILABLE:
error_class = ERROR_CLASS_UPSTREAM
else:
error_class = ERROR_CLASS_INTERNAL
code = REASON_HTTP_ERROR
return {
"reason_code": code,
"error_class": error_class,
"http_status": getattr(exc, "http_status", None),
"message": fixed_message(code),
"transport_survives": True,
}
try:
import gitea_config
if isinstance(exc, gitea_config.ConfigError):
return {
"reason_code": REASON_CONFIG_ERROR,
"error_class": ERROR_CLASS_CONFIGURATION,
"http_status": None,
"message": fixed_message(REASON_CONFIG_ERROR),
"transport_survives": True,
}
except Exception:
pass
# Unwrap FastMCP ToolError cause when the original was a typed failure.
cause = getattr(exc, "__cause__", None)
if cause is not None and cause is not exc and is_known_client_failure(cause):
return _classify_exception_impl(cause)
# No message-substring authentication heuristics (reviewer finding #3).
# Unexpected failure → internal_error, now carrying SAFE diagnostics so the
# failure is actionable: a type identifier + strictly-redacted detail +
# optional mutation-stage tag. The operator ``message`` stays the fixed
# constant; none of these fields carry secrets/bodies/paths/env.
result = {
"reason_code": REASON_INTERNAL_ERROR,
"error_class": ERROR_CLASS_INTERNAL,
"http_status": None,
"message": fixed_message(REASON_INTERNAL_ERROR),
"transport_survives": True,
"exception_class": _safe_exception_class(exc),
"detail": _redact_detail(_safe_str(exc)),
}
stage = _safe_mutation_stage(exc)
if stage:
result["mutation_stage"] = stage
return result
def build_structured_error_payload(
classification: dict[str, Any],
*,
tool_name: str | None = None,
profile_name: str | None = None,
) -> dict[str, Any]:
"""LLM-safe structured payload — fixed message + typed metadata only."""
try:
reason = str(classification.get("reason_code") or REASON_INTERNAL_ERROR)
message = fixed_message(reason)
# Refuse to emit any classification message that is not the fixed constant.
if classification.get("message") != message:
message = fixed_message(reason)
payload: dict[str, Any] = {
"success": False,
"isError": True,
"reason_code": reason if reason in FIXED_MESSAGES else REASON_INTERNAL_ERROR,
"error_class": classification.get("error_class") or ERROR_CLASS_INTERNAL,
"message": message,
"transport_survives": True,
"retryable": classification.get("error_class")
in {
ERROR_CLASS_AUTHENTICATION,
ERROR_CLASS_NETWORK,
ERROR_CLASS_CONFIGURATION,
},
}
status = classification.get("http_status")
if isinstance(status, int):
payload["http_status"] = status
if tool_name and isinstance(tool_name, str):
# Tool names are identifiers, not secrets; bound length.
payload["tool"] = tool_name[:120]
if profile_name and isinstance(profile_name, str):
payload["profile"] = profile_name[:80]
# Safe diagnostics — internal_error path only. These make an otherwise
# opaque crash actionable without leaking secrets. Re-derived here from
# the fixed message gate above; typed reasons never carry them.
if payload["reason_code"] == REASON_INTERNAL_ERROR:
exc_cls = classification.get("exception_class")
if isinstance(exc_cls, str) and exc_cls:
payload["exception_class"] = exc_cls[:_CLASS_LIMIT]
detail = classification.get("detail")
if isinstance(detail, str):
payload["detail"] = _redact_detail(detail)
stage = classification.get("mutation_stage")
if isinstance(stage, str) and stage:
payload["mutation_stage"] = stage[:_STAGE_LIMIT]
return payload
except Exception:
return {
"success": False,
"isError": True,
"reason_code": REASON_INTERNAL_ERROR,
"error_class": ERROR_CLASS_INTERNAL,
"message": fixed_message(REASON_INTERNAL_ERROR),
"transport_survives": True,
"retryable": False,
}
def log_sanitized_daemon_reason(
classification: dict[str, Any],
*,
tool_name: str | None = None,
stream=None,
) -> None:
"""Daemon log: reason codes only — never exception text or response bodies."""
stream = stream if stream is not None else sys.stderr
try:
reason = classification.get("reason_code") or REASON_INTERNAL_ERROR
error_class = classification.get("error_class") or ERROR_CLASS_INTERNAL
# Only emit known tokens; never classification['message'] from callers
# that might have been poisoned.
if reason not in FIXED_MESSAGES:
reason = REASON_INTERNAL_ERROR
error_class = ERROR_CLASS_INTERNAL
parts = [
"mcp_tool_error",
f"reason_code={reason}",
f"error_class={error_class}",
]
if tool_name and isinstance(tool_name, str):
parts.append(f"tool={tool_name[:120]}")
status = classification.get("http_status")
if isinstance(status, int):
parts.append(f"http_status={status}")
# Safe diagnostics — internal_error path ONLY. Typed reasons still emit
# no detail= (secrets lived there). class/detail/stage are re-redacted
# here as defence in depth.
if reason == REASON_INTERNAL_ERROR:
exc_cls = classification.get("exception_class")
if isinstance(exc_cls, str) and exc_cls:
parts.append(f"exception_class={exc_cls[:_CLASS_LIMIT]}")
stage = classification.get("mutation_stage")
if isinstance(stage, str) and stage:
parts.append(
f"mutation_stage={_redact_detail(stage, limit=_STAGE_LIMIT)}"
)
detail = classification.get("detail")
if isinstance(detail, str) and detail:
parts.append(f"detail={_redact_detail(detail)}")
line = " ".join(parts)
stream.write(line + "\n")
if hasattr(stream, "flush"):
stream.flush()
logger.warning(line)
except Exception:
# Fail closed: never fall back to logging the exception.
try:
stream.write(
"mcp_tool_error reason_code=internal_error "
"error_class=internal\n"
)
if hasattr(stream, "flush"):
stream.flush()
except Exception:
pass
def to_call_tool_result(
exc: BaseException,
*,
tool_name: str | None = None,
profile_name: str | None = None,
log: bool = True,
) -> Any:
"""Build a FastMCP ``CallToolResult`` with ``isError=True`` for *exc*."""
from mcp.types import CallToolResult, TextContent
try:
classification = classify_exception(exc)
if log:
log_sanitized_daemon_reason(classification, tool_name=tool_name)
payload = build_structured_error_payload(
classification, tool_name=tool_name, profile_name=profile_name
)
text = json.dumps(payload, indent=2, sort_keys=True)
return CallToolResult(
content=[TextContent(type="text", text=text)],
structuredContent=payload,
isError=True,
)
except Exception:
# Absolute fail-closed path — no exception text.
fallback = {
"success": False,
"isError": True,
"reason_code": REASON_INTERNAL_ERROR,
"error_class": ERROR_CLASS_INTERNAL,
"message": fixed_message(REASON_INTERNAL_ERROR),
"transport_survives": True,
"retryable": False,
}
return CallToolResult(
content=[
TextContent(
type="text",
text=json.dumps(fallback, indent=2, sort_keys=True),
)
],
structuredContent=fallback,
isError=True,
)
def install_tool_run_boundary(Tool) -> bool:
"""Install a **narrow** error boundary around FastMCP ``Tool.run``.
Strategy (preserves framework semantics):
* Call the **original** ``Tool.run`` for the success path (async, return
types, convert_result, protocol behavior unchanged).
* Re-raise ``UrlElicitationRequiredError`` and other control-flow
exceptions without mapping to ``internal_error``.
* Map typed Gitea client failures (and ToolError whose ``__cause__`` is
typed) to structured ``CallToolResult(isError=True)``.
* Map remaining unexpected tool failures to fixed ``internal_error``
isError results (transport survival) without secret-bearing text.
* Idempotent: second install is a no-op and returns ``False``.
"""
if getattr(Tool.run, _INSTALL_FLAG, False):
return False
from mcp.server.fastmcp.exceptions import ToolError
from mcp.shared.exceptions import UrlElicitationRequiredError
original_run = Tool.run
async def run_boundary(
self,
arguments: dict[str, Any],
context=None,
convert_result: bool = False,
) -> Any:
try:
return await original_run(
self,
arguments,
context=context,
convert_result=convert_result,
)
except UrlElicitationRequiredError:
# Framework control-flow — must not become internal_error.
raise
except BaseException as exc:
if _is_framework_control_flow(exc):
raise
if not isinstance(exc, Exception):
raise
# Prefer typed cause under FastMCP ToolError wrappers.
target: BaseException = exc
if isinstance(exc, ToolError) and exc.__cause__ is not None:
target = exc.__cause__
# Known client failures → structured isError with reason codes.
# Unexpected failures → fixed internal_error isError (no secrets).
profile_name = _safe_profile_name()
return to_call_tool_result(
target,
tool_name=getattr(self, "name", None),
profile_name=profile_name,
)
setattr(run_boundary, _INSTALL_FLAG, True)
setattr(run_boundary, _ORIGINAL_ATTR, original_run)
Tool.run = run_boundary # type: ignore[method-assign]
return True
def boundary_is_installed(Tool) -> bool:
"""True when the #699 boundary is active on *Tool.run*."""
return bool(getattr(Tool.run, _INSTALL_FLAG, False))