feat(webui): sanctioned restart and graceful reload controls (Closes #642)

Adds a gated `system.restart_namespace` action so operators and workers can
restart or gracefully reload MCP namespaces through an authorized, audited
path instead of manual host process killing (#630).

- webui/sanctioned_restart.py: restart/reload operation model, dry-run
  intent preview, confirmation-string enforcement, audit emission, and
  post-restart health verification. Fails closed on unknown auth, missing
  capability, or ambiguous target namespace.
- webui/console_authz.py: RBAC entries for the restart capability with
  secret redaction preserved.
- webui/gated_actions.py: registers the restart action in the gated action
  framework so it cannot be invoked without capability + confirmation.
- task_capability_map.py: capability mapping for the restart operation.
- docs/sanctioned-restart-controls.md: operator documentation for the
  sanctioned path and the explicit prohibition on pkill recovery.
- docs/webui-authz-audit.md: audit model updated for restart events.
- tests/test_webui_sanctioned_restart.py: authorized preview, unauthorized
  deny, confirmation enforcement, contamination classification, and audit
  emission coverage.

No unrestricted kill path is exposed; manual pkill remains classified as
contamination and continues to block clean claims.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
This commit is contained in:
2026-07-23 23:19:21 -04:00
co-authored by Claude Opus 4.8
parent b70d5f3efa
commit e593444eea
7 changed files with 1304 additions and 2 deletions
+122
View File
@@ -0,0 +1,122 @@
# Sanctioned restart and graceful reload controls (#642)
Sessions used to recover MCP connectivity by killing the host daemon
(`pkill -f mcp_server.py`, #630). That is forbidden and stays forbidden: it
kills every namespace on the host, contaminates whichever session survives, and
leaves no audit trail. This document describes the sanctioned replacement,
implemented in `webui/sanctioned_restart.py`.
## What the console will and will not do
The console **never** restarts anything. It authorizes an intent, records it,
and hands off to a host supervisor. There is no code path in which the console
sends a signal, spawns a process, or renders a kill command — a regression test
asserts the module contains no `subprocess`, `signal`, `os.kill`, `os.system`,
or `popen` reference, and that no returned payload contains a kill command.
## Operations
| Mode | Action | Minimum role | Behaviour |
|------|--------|--------------|-----------|
| `reload` | `system.reload_namespace` | controller | Host supervisor reloads the namespace in place, draining in-flight requests. |
| `restart` | `system.restart_namespace` | admin | Host supervisor restarts the namespace. In-flight requests are lost. |
Scope is always exactly one namespace. A fleet-wide restart is an explicit
non-goal: `all`, `*`, `fleet`, and an empty scope are refused with
`fleet_scope_not_permitted`, because that is precisely the blast radius the
forbidden kill already had. An unrecognised namespace is refused rather than
passed through to the host.
## The gate sequence
`assess_restart_request()` applies every gate in order and reports the first
failure with a stable reason code:
| Order | Gate | Reason code on failure |
|-------|------|------------------------|
| 1 | Mode is `restart` or `reload` | `unknown_mode` |
| 2 | Scope is a single known namespace | `fleet_scope_not_permitted`, `unknown_namespace` |
| 3 | Principal holds the required console role | `unauthorized` |
| 4 | Confirmation phrase supplied | `confirmation_required` |
| 5 | Confirmation names this namespace and mode | `confirmation_mismatch` |
| 6 | Out-of-band operator authorization present | `operator_authorization_missing` |
| 7 | Runtime is not contaminated | `contaminated_runtime` |
| 8 | Host restart hook configured | `restart_hook_not_configured` |
Passing every gate yields `host_action_required`, never "restarted".
### Confirmation binds the namespace
The required phrase is `"<mode> <namespace>"` — for example
`restart gitea-author`. Binding the namespace into the phrase is the point: a
confirmation typed for one namespace cannot be replayed against another.
### Operator authorization is not self-assertable
Host daemon maintenance is authorized out of band through
`GITEA_OPERATOR_DAEMON_MAINTENANCE_AUTHORIZATION`, read from the process
environment and nowhere else (#630; #710 finding F1). A worker session cannot
set an environment variable for an already-running daemon, so this cannot be
faked the way a tool argument could.
### The host hook
`GITEA_SANCTIONED_RESTART_HOOK` holds an opaque reference the *host* resolves —
a supervisor label such as a launchd job name, never a command line. With no
hook configured the request is refused; the console does not fall back to a
process kill. The value is read server-side and never rendered to a client.
## Manual kill remains contamination
`classify_restart_command()` classifies an operator-proposed recovery command.
A manual `pkill`/`kill`/`killall` of the MCP daemon is contamination, not a
restart: it returns `clean_claim_allowed: false` and builds a durable
contamination marker (redacted command only, never secrets) naming
`system.restart_namespace` as the sanctioned alternative.
A live, uncleared contamination marker also blocks a restart. This is stricter
than #630's task-scoped gate, which deliberately lets a contaminated worker keep
commenting and handing off: restarting a contaminated runtime would launder the
contamination rather than resolve it. Clear the marker through the reconciler
path first.
## Post-restart health verification
After the host supervisor acts, `verify_post_restart_health()` decides whether
the session may claim to be clean:
| Status | Meaning | Clean claim |
|--------|---------|-------------|
| `clean` | Required tool callable, proven through the live client namespace | Allowed |
| `unproven` | Reported healthy without live client-namespace evidence | Refused |
| `unhealthy` | Probe failed | Refused |
Only `probe_source=client_namespace` evidence clears a session. Static tool
registration is not proof, and neither is an offline subprocess probe — an IDE
client can hold a registered tool list while live calls fail with
`client is closing: EOF` (see
[`mcp-namespace-health.md`](mcp-namespace-health.md)).
## Audit
Every attempt — allowed or denied — is recorded through
`webui.console_audit` with actor, target namespace, mode, result, and reason
code, and is redacted before it is persisted. `system.restart_namespace` is
break-glass, so its records are retained for 730 days. Records carry
`process_kill_executed: false`, which is a fact about the code path rather than
a claim: no such path exists.
## Environment variables
| Variable | Purpose |
|----------|---------|
| `GITEA_SANCTIONED_RESTART_HOOK` | Host supervisor reference; absent means restart is refused. |
| `GITEA_OPERATOR_DAEMON_MAINTENANCE_AUTHORIZATION` | Out-of-band operator authorization reference. |
| `WEBUI_AUDIT_LOG` | Console audit sink; absent means records are built but not persisted. |
## Non-goals
* No unrestricted `kill` from the UI, in any role, in any phase.
* No fleet-wide restart.
* No silent auto-restart loop: every attempt is confirmed and audited.
* This does not implement the Phase 1 health API (#634).
+11 -2
View File
@@ -91,6 +91,8 @@ already define, and a regression test asserts each mapping matches.
| `close_pr` | controller | privileged | `gitea.pr.close` | Yes | No | No | 3 |
| `merge_pr` | controller | privileged | `gitea.pr.merge` | Yes | **Yes** | **Yes** | 3 |
| `delete_branch` | admin | destructive | `gitea.branch.delete` | Yes | **Yes** | **Yes** | 3 |
| `system.reload_namespace` | controller | privileged | `runtime.reload_namespace` | Yes | No | No | 2 |
| `system.restart_namespace` | admin | destructive | `runtime.restart_namespace` | Yes | **Yes** | **Yes** | 2 |
**Dual control** means the acting principal may not be the sole authority: a
second distinct principal must confirm. **Break-glass** means the action is
@@ -102,6 +104,13 @@ honouring it.
`delete_branch` is admin-only rather than controller because it is the one
irreversible action in the set.
`system.restart_namespace` is admin-only for the same reason: restarting a
namespace drops every in-flight request on it. `system.reload_namespace` drains
first, so it is privileged but not destructive. Neither action is ever executed
by the console — both hand off to a host supervisor, and neither exposes a raw
process kill. See
[`sanctioned-restart-controls.md`](sanctioned-restart-controls.md) (#642).
### Authorization decision
`authorize(action_id, principal, for_execution=False)` returns a decision
@@ -199,8 +208,8 @@ breaking the request it describes.
| Class | Applies to | Default |
|-------|-----------|---------|
| `standard` | Routine gated writes | 90 days |
| `privileged` | `review_pr`, `close_pr`, and any unclassifiable action | 365 days |
| `break_glass` | `merge_pr`, `delete_branch` | 730 days |
| `privileged` | `review_pr`, `close_pr`, `system.reload_namespace`, and any unclassifiable action | 365 days |
| `break_glass` | `merge_pr`, `delete_branch`, `system.restart_namespace` | 730 days |
Each record carries its own class, day count, and computed `expires_at`, so
retention is auditable per record rather than inferred from file age. An
+15
View File
@@ -335,6 +335,21 @@ TASK_CAPABILITY_MAP: dict[str, dict[str, str]] = {
"role": "controller",
},
# #642: sanctioned host-daemon lifecycle controls. Deliberately *not* a
# ``gitea.*`` operation — restarting an MCP namespace is a host action, not
# a Gitea API call, and no configured Gitea profile should be able to
# satisfy it by accident. Authority comes from the console RBAC model plus
# out-of-band operator authorization (#630); these entries exist so the
# console cannot invent an authority the capability layer never declared.
"restart_namespace": {
"permission": "runtime.restart_namespace",
"role": "controller",
},
"reload_namespace": {
"permission": "runtime.reload_namespace",
"role": "controller",
},
# #601 first-class lease lifecycle — inspect/list need read; mutations gate on
# ownership in the control-plane DB (not a separate Gitea write permission).
"list_workflow_leases": {
+502
View File
@@ -0,0 +1,502 @@
"""Sanctioned restart / graceful reload control tests (#642).
Acceptance criteria under test:
1. The sanctioned restart path is implemented behind gates (capability,
confirmation, operator authorization, host hook).
2. Manual ``pkill`` stays forbidden and is classified as contamination.
3. Post-restart mutations require clean health/session proof.
4. Authorized restart preview, unauthorized deny, contamination classification.
5. No entry point exposes a raw kill.
"""
import json
import os
import tempfile
import unittest
import mcp_namespace_health
import runtime_recovery_guard
from task_capability_map import TASK_CAPABILITY_MAP
from webui import console_audit, console_authz, gated_actions, sanctioned_restart
NAMESPACE = "gitea-author"
# An operator-authorized, hook-configured host. Passed explicitly so no test
# depends on (or mutates) the real process environment.
READY_ENV = {
sanctioned_restart.RESTART_HOOK_ENV: "launchd:cc.prgs.gitea-author",
runtime_recovery_guard.OPERATOR_AUTHORIZATION_ENV: "ops-ticket-4821",
}
def admin(subject: str = "[email protected]") -> console_authz.Principal:
return console_authz.Principal(
subject=subject,
role=console_authz.ADMIN,
identity_source=console_authz.IDENTITY_ACCESS_PROXY,
authenticated=True,
)
def viewer() -> console_authz.Principal:
return console_authz.Principal(
subject="[email protected]",
role=console_authz.VIEWER,
identity_source=console_authz.IDENTITY_ACCESS_PROXY,
authenticated=True,
)
class TestCapabilityWiring(unittest.TestCase):
"""AC1: authority is declared, not invented by the console."""
def test_actions_resolve_through_the_capability_map(self):
for action_id in (
sanctioned_restart.ACTION_RESTART_NAMESPACE,
sanctioned_restart.ACTION_RELOAD_NAMESPACE,
):
with self.subTest(action=action_id):
action = console_authz.get_action(action_id)
self.assertIsNotNone(action)
self.assertIn(action.task_key, TASK_CAPABILITY_MAP)
self.assertEqual(
action.mcp_permission,
TASK_CAPABILITY_MAP[action.task_key]["permission"],
)
def test_restart_permission_is_not_a_gitea_operation(self):
"""No configured Gitea profile should satisfy a host restart."""
permission = TASK_CAPABILITY_MAP["restart_namespace"]["permission"]
self.assertFalse(permission.startswith("gitea."))
def test_restart_is_destructive_dual_control_break_glass(self):
action = console_authz.get_action(
sanctioned_restart.ACTION_RESTART_NAMESPACE
)
self.assertEqual(action.action_class, console_authz.CLASS_DESTRUCTIVE)
self.assertEqual(action.minimum_role, console_authz.ADMIN)
self.assertTrue(action.dual_control)
self.assertTrue(action.break_glass)
self.assertTrue(action.requires_confirmation)
def test_reload_is_privileged_but_not_destructive(self):
action = console_authz.get_action(
sanctioned_restart.ACTION_RELOAD_NAMESPACE
)
self.assertEqual(action.action_class, console_authz.CLASS_PRIVILEGED)
self.assertTrue(action.requires_confirmation)
class TestPreview(unittest.TestCase):
"""AC4: an authorized preview renders the plan without executing it."""
def test_preview_lists_the_mutation_ledger(self):
preview = sanctioned_restart.build_restart_preview(
NAMESPACE, principal=admin(), env=READY_ENV
)
steps = [entry["step"] for entry in preview["mutation_ledger"]]
self.assertEqual(
steps, ["quiesce", "host_restart_hook", "health_recheck", "audit"]
)
self.assertTrue(preview["scope_valid"])
self.assertTrue(preview["post_restart_verification_required"])
def test_reload_preview_drains_instead_of_restarting(self):
preview = sanctioned_restart.build_restart_preview(
NAMESPACE, sanctioned_restart.MODE_RELOAD,
principal=admin(), env=READY_ENV,
)
steps = [entry["step"] for entry in preview["mutation_ledger"]]
self.assertIn("host_graceful_reload", steps)
self.assertNotIn("host_restart_hook", steps)
def test_preview_never_enables_execution(self):
preview = sanctioned_restart.build_restart_preview(
NAMESPACE, principal=admin(), env=READY_ENV
)
self.assertFalse(preview["execution_enabled"])
self.assertFalse(preview["authorization"]["execution_enabled"])
def test_confirmation_phrase_binds_the_namespace(self):
self.assertTrue(
sanctioned_restart.confirmation_matches(
NAMESPACE, sanctioned_restart.MODE_RESTART,
"restart gitea-author",
)
)
# A phrase typed for one namespace must not authorize another.
self.assertFalse(
sanctioned_restart.confirmation_matches(
"gitea-merger", sanctioned_restart.MODE_RESTART,
"restart gitea-author",
)
)
class TestGates(unittest.TestCase):
"""AC1/AC4: every gate denies with a stable reason code."""
def _assess(self, **kwargs):
params = {
"principal": admin(),
"confirmation": f"restart {NAMESPACE}",
"env": READY_ENV,
}
params.update(kwargs)
namespace = params.pop("namespace", NAMESPACE)
mode = params.pop("mode", sanctioned_restart.MODE_RESTART)
return sanctioned_restart.assess_restart_request(
namespace, mode, **params
)
def test_authorized_confirmed_request_passes_every_gate(self):
result = self._assess()
self.assertTrue(result["allowed"])
self.assertEqual(
result["reason_code"], sanctioned_restart.ALLOW_HOST_ACTION_REQUIRED
)
def test_passing_every_gate_is_not_an_execution_grant(self):
"""An allowed request still never lets the console touch the process."""
result = self._assess()
self.assertTrue(result["allowed"])
self.assertFalse(result["execution_enabled"])
self.assertFalse(result["console_executes"])
def test_unauthorized_principal_is_denied(self):
result = self._assess(principal=viewer())
self.assertFalse(result["allowed"])
self.assertEqual(
result["reason_code"], sanctioned_restart.DENY_UNAUTHORIZED
)
def test_anonymous_principal_is_denied(self):
result = self._assess(principal=None)
self.assertFalse(result["allowed"])
self.assertEqual(
result["reason_code"], sanctioned_restart.DENY_UNAUTHORIZED
)
def test_missing_confirmation_is_denied(self):
result = self._assess(confirmation=None)
self.assertFalse(result["allowed"])
self.assertEqual(
result["reason_code"], sanctioned_restart.DENY_CONFIRMATION_MISSING
)
def test_confirmation_for_another_namespace_is_denied(self):
result = self._assess(confirmation="restart gitea-merger")
self.assertFalse(result["allowed"])
self.assertEqual(
result["reason_code"], sanctioned_restart.DENY_CONFIRMATION_MISMATCH
)
def test_missing_operator_authorization_is_denied(self):
env = {sanctioned_restart.RESTART_HOOK_ENV: "launchd:cc.prgs.author"}
result = self._assess(env=env)
self.assertFalse(result["allowed"])
self.assertEqual(
result["reason_code"],
sanctioned_restart.DENY_OPERATOR_AUTHORIZATION,
)
def test_missing_host_hook_is_denied_without_kill_fallback(self):
env = {
runtime_recovery_guard.OPERATOR_AUTHORIZATION_ENV: "ops-ticket-1",
}
result = self._assess(env=env)
self.assertFalse(result["allowed"])
self.assertEqual(
result["reason_code"], sanctioned_restart.DENY_HOOK_NOT_CONFIGURED
)
def test_fleet_scope_is_refused(self):
for scope in ("all", "*", "fleet"):
with self.subTest(scope=scope):
result = self._assess(
namespace=scope, confirmation=f"restart {scope}"
)
self.assertFalse(result["allowed"])
self.assertEqual(
result["reason_code"], sanctioned_restart.DENY_FLEET_SCOPE
)
def test_unknown_namespace_is_refused(self):
result = self._assess(
namespace="gitea-nope", confirmation="restart gitea-nope"
)
self.assertFalse(result["allowed"])
self.assertEqual(
result["reason_code"], sanctioned_restart.DENY_UNKNOWN_NAMESPACE
)
def test_unknown_mode_is_refused(self):
result = self._assess(mode="obliterate")
self.assertFalse(result["allowed"])
self.assertEqual(
result["reason_code"], sanctioned_restart.DENY_UNKNOWN_MODE
)
def test_live_contamination_marker_blocks_restart(self):
marker = runtime_recovery_guard.build_contamination_record(
reason_class=runtime_recovery_guard.REASON_MANUAL_DAEMON_KILL,
command_redacted="pkill -f mcp_server.py",
)
result = self._assess(contamination_marker=marker)
self.assertFalse(result["allowed"])
self.assertEqual(
result["reason_code"], sanctioned_restart.DENY_CONTAMINATED_RUNTIME
)
def test_reconciler_cleared_marker_no_longer_blocks(self):
marker = runtime_recovery_guard.build_contamination_record(
reason_class=runtime_recovery_guard.REASON_MANUAL_DAEMON_KILL,
command_redacted="pkill -f mcp_server.py",
)
marker = dict(marker, cleared_by_reconciler=True)
result = self._assess(contamination_marker=marker)
self.assertTrue(result["allowed"])
class TestExecutionNeverKills(unittest.TestCase):
"""AC5: no path exposes or runs a raw process kill."""
def test_authorized_execution_defers_to_the_host_supervisor(self):
result = sanctioned_restart.execute_restart(
NAMESPACE,
principal=admin(),
confirmation=f"restart {NAMESPACE}",
env=READY_ENV,
)
self.assertTrue(result["allowed"])
self.assertFalse(result["success"])
self.assertFalse(result["process_kill_executed"])
self.assertEqual(
result["outcome"], sanctioned_restart.ALLOW_HOST_ACTION_REQUIRED
)
def test_denied_execution_reports_the_refusing_gate(self):
result = sanctioned_restart.execute_restart(
NAMESPACE, principal=viewer(), confirmation=f"restart {NAMESPACE}",
env=READY_ENV,
)
self.assertFalse(result["allowed"])
self.assertEqual(
result["outcome"], sanctioned_restart.DENY_UNAUTHORIZED
)
self.assertFalse(result["process_kill_executed"])
def test_module_never_spawns_a_process(self):
path = os.path.join(
os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
"webui", "sanctioned_restart.py",
)
with open(path, encoding="utf-8") as handle:
source = handle.read()
for forbidden in (
"import subprocess", "import signal", "os.kill", "os.system",
"popen",
):
with self.subTest(forbidden=forbidden):
self.assertNotIn(forbidden, source.lower())
def test_no_surface_returns_a_kill_command(self):
payloads = [
sanctioned_restart.build_restart_preview(
NAMESPACE, principal=admin(), env=READY_ENV
),
sanctioned_restart.restart_policy(),
sanctioned_restart.execute_restart(
NAMESPACE, principal=admin(),
confirmation=f"restart {NAMESPACE}", env=READY_ENV,
),
]
for payload in payloads:
rendered = json.dumps(payload, default=str).lower()
self.assertNotIn("kill -9", rendered)
self.assertNotIn("pkill -f", rendered)
def test_policy_declares_no_raw_kill_and_no_silent_restart(self):
policy = sanctioned_restart.restart_policy()
self.assertFalse(policy["raw_kill_exposed"])
self.assertFalse(policy["console_executes_process_kill"])
self.assertFalse(policy["fleet_scope_permitted"])
self.assertFalse(policy["silent_auto_restart_permitted"])
self.assertTrue(policy["audit_required"])
class TestContaminationClassification(unittest.TestCase):
"""AC2: manual pkill is contamination, and it blocks clean claims."""
def test_manual_daemon_pkill_is_contamination(self):
result = sanctioned_restart.classify_restart_command(
"pkill -f mcp_server.py"
)
self.assertTrue(result["contamination"])
self.assertFalse(result["clean_claim_allowed"])
self.assertIsNotNone(result["contamination_marker"])
self.assertEqual(
result["sanctioned_alternative"],
sanctioned_restart.ACTION_RESTART_NAMESPACE,
)
def test_broad_process_kill_is_contamination(self):
result = sanctioned_restart.classify_restart_command("killall -9 Python")
self.assertTrue(result["contamination"])
self.assertFalse(result["clean_claim_allowed"])
def test_marker_names_the_sanctioned_alternative(self):
result = sanctioned_restart.classify_restart_command(
"pkill -f mcp_server.py"
)
marker = result["contamination_marker"]
self.assertIn(
sanctioned_restart.ACTION_RESTART_NAMESPACE, marker["detail"]
)
def test_benign_command_is_not_contamination(self):
result = sanctioned_restart.classify_restart_command("git status")
self.assertFalse(result["contamination"])
self.assertTrue(result["clean_claim_allowed"])
def test_no_command_is_not_contamination(self):
result = sanctioned_restart.classify_restart_command(None)
self.assertFalse(result["contamination"])
self.assertTrue(result["clean_claim_allowed"])
class TestPostRestartHealth(unittest.TestCase):
"""AC3: a clean post-restart claim needs live client-namespace proof."""
def test_live_client_probe_clears_the_session(self):
result = sanctioned_restart.verify_post_restart_health(
NAMESPACE,
probe_result={"success": True},
probe_source=mcp_namespace_health.PROBE_SOURCE_CLIENT,
registered_tools=["gitea_whoami"],
required_tool="gitea_whoami",
)
self.assertEqual(result["status"], sanctioned_restart.HEALTH_CLEAN)
self.assertTrue(result["clean_claim_allowed"])
self.assertTrue(result["mutations_allowed"])
def test_offline_probe_does_not_clear_the_session(self):
result = sanctioned_restart.verify_post_restart_health(
NAMESPACE,
probe_result={"success": True},
probe_source=mcp_namespace_health.PROBE_SOURCE_OFFLINE,
registered_tools=["gitea_whoami"],
required_tool="gitea_whoami",
)
self.assertFalse(result["clean_claim_allowed"])
self.assertFalse(result["mutations_allowed"])
def test_failed_probe_is_unhealthy(self):
result = sanctioned_restart.verify_post_restart_health(
NAMESPACE,
probe_result={"success": False, "error": "client is closing: EOF"},
probe_source=mcp_namespace_health.PROBE_SOURCE_CLIENT,
registered_tools=["gitea_whoami"],
required_tool="gitea_whoami",
)
self.assertEqual(result["status"], sanctioned_restart.HEALTH_UNHEALTHY)
self.assertFalse(result["clean_claim_allowed"])
def test_static_registration_alone_never_clears_the_session(self):
result = sanctioned_restart.verify_post_restart_health(
NAMESPACE,
registered_tools=["gitea_whoami"],
required_tool="gitea_whoami",
)
self.assertFalse(result["clean_claim_allowed"])
class TestAuditEmission(unittest.TestCase):
"""Every restart attempt is audited with actor, target, and result."""
def _run(self, principal, sink):
prior = os.environ.get(console_audit.AUDIT_LOG_ENV)
os.environ[console_audit.AUDIT_LOG_ENV] = sink
try:
return sanctioned_restart.execute_restart(
NAMESPACE,
principal=principal,
confirmation=f"restart {NAMESPACE}",
env=READY_ENV,
request_id="req-642",
)
finally:
if prior is None:
os.environ.pop(console_audit.AUDIT_LOG_ENV, None)
else:
os.environ[console_audit.AUDIT_LOG_ENV] = prior
def test_allowed_attempt_is_written_with_actor_and_target(self):
with tempfile.TemporaryDirectory() as tmp:
sink = os.path.join(tmp, "audit.jsonl")
result = self._run(admin(), sink)
self.assertTrue(result["audit"]["written"])
with open(sink, encoding="utf-8") as handle:
record = json.loads(handle.read().strip())
self.assertEqual(
record["action"], sanctioned_restart.ACTION_RESTART_NAMESPACE
)
self.assertEqual(record["target"]["namespace"], NAMESPACE)
self.assertEqual(record["target"]["mode"], "restart")
self.assertEqual(record["result"], console_audit.RESULT_ALLOWED)
self.assertEqual(record["actor"]["subject"], "[email protected]")
self.assertFalse(record["metadata"]["process_kill_executed"])
def test_denied_attempt_is_audited_too(self):
with tempfile.TemporaryDirectory() as tmp:
sink = os.path.join(tmp, "audit.jsonl")
self._run(viewer(), sink)
with open(sink, encoding="utf-8") as handle:
record = json.loads(handle.read().strip())
self.assertEqual(record["result"], console_audit.RESULT_DENIED)
self.assertEqual(
record["reason_code"], sanctioned_restart.DENY_UNAUTHORIZED
)
def test_restart_audit_uses_break_glass_retention(self):
action = console_authz.get_action(
sanctioned_restart.ACTION_RESTART_NAMESPACE
)
self.assertEqual(
console_audit.retention_class_for(action),
console_audit.RETENTION_BREAK_GLASS,
)
class TestRegistrySurface(unittest.TestCase):
"""AC5: the console surfaces the control, still disabled, with no kill."""
def test_registry_exposes_both_actions_disabled(self):
registry = gated_actions.load_action_registry()
for action_id in (
sanctioned_restart.ACTION_RESTART_NAMESPACE,
sanctioned_restart.ACTION_RELOAD_NAMESPACE,
):
with self.subTest(action=action_id):
action = registry.get(action_id)
self.assertIsNotNone(action)
self.assertFalse(action.enabled)
def test_registry_preview_names_the_namespace_target(self):
preview = gated_actions.preview_action(
sanctioned_restart.ACTION_RESTART_NAMESPACE, namespace=NAMESPACE
)
target = preview["mutation_ledger"][0]["target"]
self.assertIn(NAMESPACE, target)
self.assertFalse(preview["enabled"])
def test_registry_attempt_fails_closed(self):
result = gated_actions.attempt_action(
sanctioned_restart.ACTION_RESTART_NAMESPACE, namespace=NAMESPACE
)
self.assertFalse(result["success"])
if __name__ == "__main__":
unittest.main()
+28
View File
@@ -236,6 +236,34 @@ _ACTION_SPECS: tuple[ConsoleAction, ...] = (
phase=3,
summary="Remove a remote feature branch.",
),
# #642: sanctioned daemon lifecycle. These exist so operators have an
# audited path off `pkill -f mcp_server.py` (#630). Restart drops every
# in-flight request on a namespace, so it carries the same dual-control and
# break-glass weight as a merge; reload drains first and is privileged but
# not destructive. Neither ever exposes a raw kill: execution is handed to
# a host supervisor by ``webui.sanctioned_restart``.
ConsoleAction(
action_id="system.reload_namespace",
task_key="reload_namespace",
action_class=CLASS_PRIVILEGED,
minimum_role=CONTROLLER,
requires_confirmation=True,
dual_control=False,
break_glass=False,
phase=2,
summary="Gracefully reload one MCP namespace via the host supervisor.",
),
ConsoleAction(
action_id="system.restart_namespace",
task_key="restart_namespace",
action_class=CLASS_DESTRUCTIVE,
minimum_role=ADMIN,
requires_confirmation=True,
dual_control=True,
break_glass=True,
phase=2,
summary="Restart one MCP namespace via the host supervisor.",
),
)
ACTIONS: dict[str, ConsoleAction] = {a.action_id: a for a in _ACTION_SPECS}
+13
View File
@@ -110,6 +110,8 @@ def _format_target(action_id: str, params: dict[str, Any]) -> str:
)
if action_id == "create_issue":
return f"issue {params.get('title', '?')!r}"
if action_id in {"system.restart_namespace", "system.reload_namespace"}:
return f"MCP namespace {params.get('namespace', '?')!r}"
return "unspecified"
@@ -165,6 +167,17 @@ def build_action_registry() -> ActionRegistry:
"gitea_create_issue_comment", "Post a PR review thread comment."),
("close_pr", "Close PR", "close_pr", "gitea_edit_pr",
"Close a pull request without merge."),
# #642: the sanctioned replacement for the forbidden manual daemon-kill
# recovery path (#630). The "tool" is a host supervisor hook, not an MCP
# call — the console never signals a process. Preview and gating live in
# ``webui.sanctioned_restart``; these stay disabled like every other
# registry entry.
("system.reload_namespace", "Reload MCP namespace", "reload_namespace",
"host.supervisor_reload",
"Gracefully reload one MCP namespace via the host supervisor."),
("system.restart_namespace", "Restart MCP namespace",
"restart_namespace", "host.supervisor_restart",
"Restart one MCP namespace via the host supervisor."),
)
actions = tuple(
GatedAction(
+613
View File
@@ -0,0 +1,613 @@
"""Sanctioned MCP restart and graceful reload controls (#642, Phase 2).
Sessions have historically recovered MCP connectivity by killing the host
daemon (``pkill -f mcp_server.py``, #630). That path stays forbidden: it kills
every namespace on the host, contaminates the surviving session, and leaves no
audit trail. This module is the sanctioned replacement.
A restart is modelled as a *gated action*, never as a command:
1. **Capability** — the console action resolves through ``console_authz``
against ``task_capability_map``, so the console cannot invent an authority
the MCP layer does not already define.
2. **Preview** — :func:`build_restart_preview` renders a mutation ledger and
the exact confirmation phrase. It never returns a shell command.
3. **Confirmation** — the operator echoes a phrase naming the exact namespace
and mode. A phrase for one namespace never authorizes another.
4. **Operator authorization** — host daemon maintenance is authorized out of
band through the environment (#630, and #710 finding F1: a worker session
cannot set an env var for an already-running daemon, so this cannot be
self-asserted the way a tool argument could).
5. **Execution** — :func:`execute_restart` never spawns a process. Once every
gate passes it hands the request to the configured host-managed restart
hook; with no hook configured it fails closed.
6. **Health recheck** — :func:`verify_post_restart_health` requires live
client-namespace probe evidence before any post-restart clean claim.
Manual ``pkill`` remains forbidden and is classified as contamination by
:func:`classify_restart_command`, which blocks clean claims (#630 AC3).
This module performs no I/O beyond reading its own environment configuration,
imports no MCP client, and holds no credential.
"""
from __future__ import annotations
import os
from dataclasses import asdict, dataclass
from typing import Any
import mcp_namespace_health
import runtime_recovery_guard
from webui import console_audit, console_authz
# --- Operations -------------------------------------------------------------
MODE_RESTART = "restart"
MODE_RELOAD = "reload"
MODES: tuple[str, ...] = (MODE_RESTART, MODE_RELOAD)
ACTION_RESTART_NAMESPACE = "system.restart_namespace"
ACTION_RELOAD_NAMESPACE = "system.reload_namespace"
ACTION_FOR_MODE: dict[str, str] = {
MODE_RESTART: ACTION_RESTART_NAMESPACE,
MODE_RELOAD: ACTION_RELOAD_NAMESPACE,
}
# Namespaces the console may target. An unlisted name fails closed rather than
# being passed through to a host hook.
KNOWN_NAMESPACES: tuple[str, ...] = tuple(
sorted(
set(mcp_namespace_health.DEFAULT_NAMESPACES)
| {"gitea-author", "gitea-reviewer", "gitea-merger",
"gitea-reconciler", "gitea-controller"}
)
)
# Scope tokens that would mean "everything at once". Explicit non-goal: the
# console never offers a fleet-wide restart, because that is the blast radius
# `pkill -f mcp_server.py` already had.
_FLEET_TOKENS = frozenset({"*", "all", "fleet", "any", ""})
# --- Environment configuration ----------------------------------------------
# Read server-side only; the value is an opaque host hook reference (e.g. a
# launchd label), never a command line, and is never rendered to a client.
RESTART_HOOK_ENV = "GITEA_SANCTIONED_RESTART_HOOK"
# --- Reason codes -----------------------------------------------------------
DENY_UNKNOWN_MODE = "unknown_mode"
DENY_UNKNOWN_NAMESPACE = "unknown_namespace"
DENY_FLEET_SCOPE = "fleet_scope_not_permitted"
DENY_UNAUTHORIZED = "unauthorized"
DENY_CONFIRMATION_MISSING = "confirmation_required"
DENY_CONFIRMATION_MISMATCH = "confirmation_mismatch"
DENY_OPERATOR_AUTHORIZATION = "operator_authorization_missing"
DENY_HOOK_NOT_CONFIGURED = "restart_hook_not_configured"
DENY_CONTAMINATED_RUNTIME = "contaminated_runtime"
ALLOW_HOST_ACTION_REQUIRED = "host_action_required"
# Post-restart verification outcomes.
HEALTH_CLEAN = "clean"
HEALTH_UNPROVEN = "unproven"
HEALTH_UNHEALTHY = "unhealthy"
def _clean(value: Any) -> str:
return str(value or "").strip()
# --- Mutation ledger --------------------------------------------------------
@dataclass(frozen=True)
class RestartLedgerEntry:
"""One planned step, shown before anything is asked of the host."""
sequence: int
step: str
summary: str
executes_process_kill: bool = False
def _mutation_ledger(namespace: str, mode: str) -> tuple[RestartLedgerEntry, ...]:
if mode == MODE_RELOAD:
middle = RestartLedgerEntry(
sequence=2,
step="host_graceful_reload",
summary=(
f"Ask the configured host supervisor to reload {namespace} "
"in place, draining in-flight requests. The console does not "
"signal the process itself."
),
)
else:
middle = RestartLedgerEntry(
sequence=2,
step="host_restart_hook",
summary=(
f"Ask the configured host supervisor to restart {namespace}. "
"The console never sends a signal and never runs a kill."
),
)
return (
RestartLedgerEntry(
sequence=1,
step="quiesce",
summary=(
f"Stop admitting new gated mutations for {namespace} and "
"record the intent before anything restarts."
),
),
middle,
RestartLedgerEntry(
sequence=3,
step="health_recheck",
summary=(
f"Re-probe {namespace} through the live client namespace and "
"prove the required tool is callable again."
),
),
RestartLedgerEntry(
sequence=4,
step="audit",
summary=(
"Append actor, target namespace, mode, and result to the "
"console audit log."
),
),
)
# --- Confirmation -----------------------------------------------------------
def confirmation_phrase(namespace: str, mode: str) -> str:
"""Exact phrase an operator must echo, naming the namespace and mode.
Binding the namespace into the phrase is the point: a confirmation typed
for ``gitea-author`` cannot be replayed against ``gitea-merger``.
"""
return f"{_clean(mode)} {_clean(namespace)}"
def confirmation_matches(
namespace: str, mode: str, confirmation: str | None
) -> bool:
"""Compare *confirmation* to the required phrase (exact, whitespace-trimmed)."""
return _clean(confirmation) == confirmation_phrase(namespace, mode)
# --- Scope validation -------------------------------------------------------
def _validate_scope(namespace: str, mode: str) -> tuple[str, str] | None:
"""Return ``(reason_code, detail)`` when the scope is refused."""
ns = _clean(namespace)
md = _clean(mode)
if md not in MODES:
return (
DENY_UNKNOWN_MODE,
f"Mode {md!r} is not one of {', '.join(MODES)}.",
)
if ns.lower() in _FLEET_TOKENS:
return (
DENY_FLEET_SCOPE,
(
"Fleet-wide restart is an explicit non-goal: it reproduces the "
"blast radius of `pkill -f mcp_server.py` (#630). Restart one "
"namespace at a time."
),
)
if ns not in KNOWN_NAMESPACES:
return (
DENY_UNKNOWN_NAMESPACE,
f"Namespace {ns!r} is not a known MCP namespace.",
)
return None
# --- Host hook --------------------------------------------------------------
def restart_hook(env: dict[str, str] | None = None) -> dict[str, Any]:
"""Report the configured host-managed restart hook.
The hook is a reference the *host* resolves (a supervisor label), not a
command this process runs. ``configured=False`` fails restart closed.
"""
source = env if env is not None else os.environ
reference = _clean(source.get(RESTART_HOOK_ENV))
return {
"configured": bool(reference),
"reference": reference or None,
"source": RESTART_HOOK_ENV if reference else None,
"self_assertable": False,
"console_executes_process": False,
}
# --- Preview ----------------------------------------------------------------
def build_restart_preview(
namespace: str,
mode: str = MODE_RESTART,
*,
principal: console_authz.Principal | None = None,
env: dict[str, str] | None = None,
) -> dict[str, Any]:
"""Render the dry-run preview for a restart/reload request.
Read-only: no authorization is granted, no host is contacted, and the
result never contains a shell command.
"""
ns = _clean(namespace)
md = _clean(mode)
action_id = ACTION_FOR_MODE.get(md, ACTION_RESTART_NAMESPACE)
action = console_authz.get_action(action_id)
decision = console_authz.authorize(action_id, principal)
scope_error = _validate_scope(ns, md)
hook = restart_hook(env)
operator = runtime_recovery_guard.operator_authorization(env)
return {
"action_id": action_id,
"namespace": ns,
"mode": md,
"scope_valid": scope_error is None,
"scope_reason_code": scope_error[0] if scope_error else None,
"scope_detail": scope_error[1] if scope_error else None,
"required_role": action.minimum_role if action else None,
"required_permission": action.mcp_permission if action else None,
"action_class": action.action_class if action else None,
"dual_control": action.dual_control if action else True,
"break_glass": action.break_glass if action else True,
"requires_confirmation": True,
"confirmation_phrase": confirmation_phrase(ns, md),
"mutation_ledger": [asdict(entry) for entry in _mutation_ledger(ns, md)],
"authorization": decision.to_dict(),
"operator_authorization": operator,
"restart_hook": hook,
"execution_enabled": False,
"raw_process_kill_exposed": False,
"known_namespaces": list(KNOWN_NAMESPACES),
"post_restart_verification_required": True,
}
# --- Gate -------------------------------------------------------------------
def assess_restart_request(
namespace: str,
mode: str = MODE_RESTART,
*,
principal: console_authz.Principal | None = None,
confirmation: str | None = None,
contamination_marker: dict[str, Any] | None = None,
env: dict[str, str] | None = None,
) -> dict[str, Any]:
"""Decide whether a restart request may proceed to the host hook.
Every gate must pass. The first failure wins and is reported with a stable
reason code; a pass never means "restarted", only "may be handed to the
configured host hook".
"""
ns = _clean(namespace)
md = _clean(mode)
action_id = ACTION_FOR_MODE.get(md, ACTION_RESTART_NAMESPACE)
preview = build_restart_preview(ns, md, principal=principal, env=env)
def refuse(reason_code: str, detail: str) -> dict[str, Any]:
return {
"allowed": False,
"gates_passed": False,
"reason_code": reason_code,
"detail": detail,
"action_id": action_id,
"namespace": ns,
"mode": md,
"preview": preview,
"execution_enabled": False,
}
scope_error = _validate_scope(ns, md)
if scope_error is not None:
return refuse(*scope_error)
# Authority is checked as an authorization decision, not an execution
# grant. ``for_execution=True`` asks "may the console perform this write?",
# and the answer here is permanently no: step 2 of the ledger is a request
# to the host supervisor, so the console's Phase 2 execution gate is not
# the relevant gate. Every branch below keeps ``execution_enabled`` False
# and :func:`execute_restart` never touches a process.
decision = console_authz.authorize(action_id, principal)
if not decision.allowed:
return refuse(DENY_UNAUTHORIZED, decision.detail)
if not _clean(confirmation):
return refuse(
DENY_CONFIRMATION_MISSING,
(
"Type the confirmation phrase "
f"{preview['confirmation_phrase']!r} to proceed."
),
)
if not confirmation_matches(ns, md, confirmation):
return refuse(
DENY_CONFIRMATION_MISMATCH,
(
"Confirmation does not name this namespace and mode; expected "
f"{preview['confirmation_phrase']!r}."
),
)
operator = preview["operator_authorization"]
if not operator["authorized"]:
return refuse(
DENY_OPERATOR_AUTHORIZATION,
(
"Host daemon maintenance requires out-of-band operator "
"authorization via "
f"{runtime_recovery_guard.OPERATOR_AUTHORIZATION_ENV}."
),
)
# #630's task-scoped gate deliberately lets a contaminated worker keep
# commenting and handing off. Restart is stricter and unconditional: a
# runtime already contaminated by a manual kill must be reconciled before
# it is restarted again, or the restart just launders the contamination.
if contamination_marker and not contamination_marker.get(
"cleared_by_reconciler"
):
return refuse(
DENY_CONTAMINATED_RUNTIME,
(
"A live contamination marker is present; clear it through the "
"reconciler path before restarting."
),
)
hook = preview["restart_hook"]
if not hook["configured"]:
return refuse(
DENY_HOOK_NOT_CONFIGURED,
(
"No host-managed restart hook is configured "
f"({RESTART_HOOK_ENV}). The console will not fall back to a "
"process kill."
),
)
return {
"allowed": True,
"gates_passed": True,
"reason_code": ALLOW_HOST_ACTION_REQUIRED,
"detail": (
"Every gate passed. The restart must be performed by the "
"configured host supervisor; the console does not signal the "
"process."
),
"action_id": action_id,
"namespace": ns,
"mode": md,
"preview": preview,
"execution_enabled": False,
"console_executes": False,
"console_active_phase": console_authz.ACTIVE_PHASE,
}
# --- Execution --------------------------------------------------------------
def execute_restart(
namespace: str,
mode: str = MODE_RESTART,
*,
principal: console_authz.Principal | None = None,
confirmation: str | None = None,
contamination_marker: dict[str, Any] | None = None,
env: dict[str, str] | None = None,
request_id: str | None = None,
session_id: str | None = None,
) -> dict[str, Any]:
"""Run every gate, audit the outcome, and hand off to the host.
This function never spawns a process, never sends a signal, and never
builds a command line. ``success`` is False in both directions: a refused
request is refused, and an authorized request still requires the host
supervisor to act.
"""
assessment = assess_restart_request(
namespace,
mode,
principal=principal,
confirmation=confirmation,
contamination_marker=contamination_marker,
env=env,
)
action_id = assessment["action_id"]
allowed = assessment["allowed"]
audit = console_audit.record_event(
action_id=action_id,
result=(
console_audit.RESULT_ALLOWED if allowed
else console_audit.RESULT_DENIED
),
principal=principal,
target={"namespace": assessment["namespace"], "mode": assessment["mode"]},
reason_code=assessment["reason_code"],
detail=assessment["detail"],
request_id=request_id,
session_id=session_id,
metadata={
"gates_passed": assessment["gates_passed"],
"process_kill_executed": False,
"post_restart_verification_required": True,
},
)
return {
"success": False,
"outcome": (
ALLOW_HOST_ACTION_REQUIRED if allowed else assessment["reason_code"]
),
"allowed": allowed,
"detail": assessment["detail"],
"namespace": assessment["namespace"],
"mode": assessment["mode"],
"action_id": action_id,
"process_kill_executed": False,
"host_hook": assessment["preview"]["restart_hook"],
"next_action": (
"Have the host supervisor perform the restart, then call "
"verify_post_restart_health with live client-namespace evidence "
"before claiming a clean session."
if allowed
else assessment["detail"]
),
"assessment": assessment,
"audit": audit,
}
# --- Contamination classification -------------------------------------------
def classify_restart_command(
command: str | None,
*,
mcp_pids: list[Any] | tuple[Any, ...] | None = None,
session_id: str | None = None,
remote: str | None = None,
role: str | None = None,
) -> dict[str, Any]:
"""Classify an operator-proposed recovery command (#630 AC2).
A manual ``pkill``/``kill`` of the MCP daemon is contamination, not a
restart. When contaminating, a durable marker is returned so downstream
gated mutations and clean claims fail closed.
"""
classification = runtime_recovery_guard.classify_recovery_command(
command, mcp_pids=mcp_pids
)
contaminating = bool(classification.get("contamination"))
marker = None
if contaminating:
marker = runtime_recovery_guard.build_contamination_record(
reason_class=(
classification.get("reason_class")
or runtime_recovery_guard.REASON_MANUAL_DAEMON_KILL
),
command_redacted=classification.get("redacted_command"),
session_id=session_id,
remote=remote,
role=role,
detail=(
"Manual daemon kill is forbidden; use the sanctioned "
f"{ACTION_RESTART_NAMESPACE} gated action instead."
),
)
return {
"contamination": contaminating,
"sanctioned": not contaminating and not classification.get("process_kill"),
"clean_claim_allowed": not contaminating,
"reason_class": classification.get("reason_class"),
"redacted_command": classification.get("redacted_command"),
"classification": classification,
"contamination_marker": marker,
"sanctioned_alternative": ACTION_RESTART_NAMESPACE,
}
# --- Post-restart health verification ---------------------------------------
def verify_post_restart_health(
namespace: str,
*,
probe_result: dict[str, Any] | None = None,
probe_source: str | None = None,
registered_tools: list[str] | tuple[str, ...] | None = None,
required_tool: str | None = None,
profile: str | None = None,
) -> dict[str, Any]:
"""Require live proof a namespace is callable before any clean claim (AC3).
Static registration is not proof and neither is an offline subprocess
probe: only ``probe_source=client_namespace`` evidence can clear a
post-restart session for mutations.
"""
ns = _clean(namespace)
health = mcp_namespace_health.classify_namespace_probe(
ns,
required_tool=required_tool,
registered_tools=registered_tools,
probe_result=probe_result,
profile=profile,
probe_source=probe_source,
)
healthy = bool(health.get("healthy"))
proven = bool(health.get("ide_namespace_proven"))
if healthy and proven:
status = HEALTH_CLEAN
elif healthy:
status = HEALTH_UNPROVEN
else:
status = HEALTH_UNHEALTHY
reasons = list(health.get("reasons") or [])
if status == HEALTH_UNPROVEN:
reasons.append(
"Namespace reported healthy without live client-namespace "
"evidence; a post-restart clean claim requires "
f"probe_source={mcp_namespace_health.PROBE_SOURCE_CLIENT!r}."
)
return {
"namespace": ns,
"status": status,
"healthy": healthy,
"ide_namespace_proven": proven,
"clean_claim_allowed": status == HEALTH_CLEAN,
"mutations_allowed": status == HEALTH_CLEAN,
"reasons": reasons,
"health": health,
}
# --- Policy surface ---------------------------------------------------------
def restart_policy() -> dict[str, Any]:
"""Machine-readable description of the sanctioned restart contract."""
return {
"policy_version": 1,
"modes": list(MODES),
"actions": [ACTION_RESTART_NAMESPACE, ACTION_RELOAD_NAMESPACE],
"known_namespaces": list(KNOWN_NAMESPACES),
"fleet_scope_permitted": False,
"console_executes_process_kill": False,
"raw_kill_exposed": False,
"requires_confirmation": True,
"confirmation_binds_namespace": True,
"operator_authorization_env": (
runtime_recovery_guard.OPERATOR_AUTHORIZATION_ENV
),
"restart_hook_env": RESTART_HOOK_ENV,
"manual_kill_classified_as": runtime_recovery_guard.CONTAMINATION_KIND,
"post_restart_clean_claim_requires": (
mcp_namespace_health.PROBE_SOURCE_CLIENT
),
"audit_required": True,
"silent_auto_restart_permitted": False,
}