Merge branch 'runner-egress-reset' · 9e4606f · flagon-io/g1t

606	606		the toolchains' download sites (`nodejs.org`, `go.dev`,
607	607		`static.rust-lang.org`), and the public container registries (Docker Hub,
608	608		GitHub's, Quay, and `mirror.gcr.io`, the mirror of Docker Hub that a job's
609		−	Engine asks first). A request anywhere else gets `403` with
610		−	the reason. To reach another host, someone with the Maintain [role](/guides/access-and-roles/) or
	609	+	Engine asks first). A request anywhere else gets `403`, and its status
	610	+	line names the host (`403 Forbidden by g1t guardrails: example.com is not
	611	+	an allowed domain`), so even a tool that reports only the status says which
	612	+	host to allow. To reach another host, someone with the Maintain [role](/guides/access-and-roles/) or
611	613		higher adds it to the project's
612	614		allowed domains under Settings → Guardrails; a project whose guardrails
613	615		turn the network restriction off runs its jobs with an open network.

70	70		and HTTPS (443) has any route out: SSH, raw TCP and UDP connections fail.
71	71		- Every HTTP and HTTPS request it makes is handed to g1t's runner Worker
72	72		before it leaves. The Worker forwards requests to allowed hosts and
73		−	answers every other with `403` and a line saying the host is not allowed
74		−	and where to allow it.
	73	+	answers every other with `403`. The refusal names the host three times
	74	+	over, since tools differ in what they show of a failed request: in its
	75	+	status line (`403 Forbidden by g1t guardrails: example.com is not an
	76	+	allowed domain`, which is what a tool that prints only the status shows),
	77	+	in an `x-g1t-blocked-host` header beside `x-g1t-guardrails: blocked`, and
	78	+	in a plain-text body saying where to allow it: Allowed domains for
	79	+	every sandbox, or Workflow-only domains for the jobs of one workflow
	80	+	and environment.
75	81		- To see the host of an HTTPS request, the connection is re-encrypted with
76	82		a certificate the sandbox is given when it starts, which g1t adds to the
77	83		sandbox's trusted certificates (and points Node.js, Python, curl, Cargo

50	50		const response = refusal("Evil.com:443");
51	51		assert.equal(response.status, 403);
52	52		assert.equal(response.headers.get("x-g1t-guardrails"), "blocked");
	53	+	assert.equal(response.headers.get("x-g1t-blocked-host"), "evil.com");
	54	+	// Tools that print only the status line still name the host.
	55	+	assert.equal(response.statusText, "Forbidden by g1t guardrails: evil.com is not an allowed domain");
53	56		const text = await response.text();
54		−	assert.match(text, /evil\.com is not on this project's allowed domains/);
55		−	assert.match(text, /Settings, Guardrails/);
	57	+	assert.match(text, /^g1t guardrails refused evil\.com: it is not on this project's allowed domains/);
	58	+	assert.match(text, /Settings, Guardrails: under Allowed domains/);
	59	+	assert.match(text, /Workflow-only domains/);
56	60		assert.equal(blockedStep("Evil.com"), "Blocked: evil.com (not an allowed domain)");
57	61		});
58	62

74	74		return [...hosts];
75	75		}
76	76
77		−	/** What a refused request gets back, so the agent knows why. */
	77	+	/**
	78	+	* What a refused request gets back. Many tools show only the status line
	79	+	* of a failed request (a GraphQL client prints its reason phrase, a
	80	+	* package manager "HTTP 403: Forbidden"), so the reason phrase names the
	81	+	* host too, and the body and a header say the rest.
	82	+	*/
78	83		export function refusal(host: string): Response {
79		−	const name = normalizeHost(host);
	84	+	// A reason phrase is printable ASCII; a host from a URL already is.
	85	+	const name = normalizeHost(host).replace(/[^\x21-\x7e]/g, "");
80	86		return new Response(
81		−	`g1t guardrails: ${name} is not on this project's allowed domains, so this sandbox cannot reach it. ` +
82		−	"A member can allow it under the project's Settings, Guardrails.\n",
83		−	{ status: 403, headers: { "content-type": "text/plain; charset=utf-8", "x-g1t-guardrails": "blocked" } },
	87	+	`g1t guardrails refused ${name}: it is not on this project's allowed domains, so this sandbox cannot reach it. ` +
	88	+	"A member can allow it in the project's Settings, Guardrails: under Allowed domains for every sandbox, " +
	89	+	"or under Workflow-only domains for the jobs of one workflow and environment.\n",
	90	+	{
	91	+	status: 403,
	92	+	statusText: `Forbidden by g1t guardrails: ${name} is not an allowed domain`,
	93	+	headers: { "content-type": "text/plain; charset=utf-8", "x-g1t-guardrails": "blocked", "x-g1t-blocked-host": name },
	94	+	},
84	95		);
85	96		}
86	97