fix: anchor github-url host to GITHUB_SERVER_URL for GHES support

[phorcys420]

What

github-url was validated against a regex hardcoded to github.com, so on GitHub Enterprise Server every issue/PR URL (https://github.example.com/owner/repo/pull/1) failed validation and the action exited before commenting.

This anchors the host to the runner-provided GITHUB_SERVER_URL instead. The anchor stays runner-controlled, never user-controlled, so a workflow that templates user input into github-url still cannot redirect the action to an attacker-chosen host.

Fixes #39.

How

• parseGithubItemURL / deriveCommentKey take an optional serverURL (default https://github.com); GITHUB_SERVER_URL is read at the action.ts edge, matching how GITHUB_WORKFLOW is already handled.
• The matcher is built with the built-in RegExp.escape(normalizeBaseUrl(serverURL)), so host metacharacters (the dots in github.com) stay literal. No hand-rolled escaping, no URL parsing.
• Host anchoring semantics are unchanged (scheme + host must equal the server; extra path segments and non-server hosts rejected).
• The parseGithubURL error message now reflects the resolved server URL.

Tests

bun test (208 pass), typecheck, lint, format:check all clean. New coverage: dotcom still parses, GHES parses when GITHUB_SERVER_URL matches, host-mismatch is rejected (the security case), a dot-in-host isn't treated as a regex wildcard, and deriveCommentKey derives owner/repo#n on GHES.

Implementation plan

Fix: github-url host regex hardcoded to github.com (issue #39)

Make agents-chat-action work on GitHub Enterprise Server by anchoring the
github-url validation to the runner-provided GITHUB_SERVER_URL instead of
the literal github.com, without weakening the security property (the host
anchor stays runner-controlled, never user-controlled).

Repo: coder/agents-chat-action · Branch: phorcys/ghes-github-url

Approach (decisions locked)

• Build one regex from the server URL, escaped with the built-in
  RegExp.escape. No hand-rolled escape helper, no URL parsing. The server
  origin is escaped and interpolated into a single anchored pattern.
  RegExp.escape is Stage 4 and ships in Node 24 (the action runtime).
• Plumb serverURL in as an optional param (default https://github.com),
  reading process.env.GITHUB_SERVER_URL at the action.ts edge, consistent
  with how GITHUB_WORKFLOW is already read there.
• deriveCommentKey is GHES-aware too, so markers resolve to
  owner/repo#n on enterprise hosts instead of falling back to the raw URL.
• Host anchoring semantics preserved and unchanged: the escaped literal
  keeps host matching case-sensitive, exactly as the current github.com
  literal does. Scheme + host must equal the server; extra path segments and
  non-server hosts are rejected.
• Error messages reflect the real server URL; no extensive docs rewrite.

Changes

src/comment.ts

Replace the module-level GITHUB_URL_REGEX constant with a builder that anchors
to the escaped server origin, and thread serverURL through the two consumers.

const DEFAULT_GITHUB_SERVER_URL = "https://github.com";

// Anchored issue/PR URL matcher for a given GitHub server. RegExp.escape
// neutralizes metacharacters in the host (e.g. the dots in github.com) so it
// matches literally. Anchored at both ends; extra path segments are rejected.
// The trailing (?:[?#].*)? keeps query strings and fragments tolerated.
function githubURLRegex(serverURL: string): RegExp {
	const base = RegExp.escape(normalizeBaseUrl(serverURL));
	return new RegExp(
		`^${base}/([^/]+)/([^/]+)/(?:issues|pull)/(\\d+)/?(?:[?#].*)?$`,
	);
}

• parseGithubItemURL(input, serverURL = DEFAULT_GITHUB_SERVER_URL): bail on
  empty input, then githubURLRegex(serverURL).exec(input); groups stay
  owner=1, repo=2, number=3, returned via tuple destructuring.
• normalizeBaseUrl still trims a trailing slash / query / fragment off
  serverURL before it is escaped, so a stray GITHUB_SERVER_URL=https://host/
  does not double the slash in the pattern.
• Runtime check (must verify in the workspace): confirm RegExp.escape
  exists in the pinned Bun test runtime. It is guaranteed on Node 24 (action
  runtime) but Bun's JSC support is version-dependent. If the pinned Bun lacks
  it, fall back to the generic-host capture-and-compare variant (static regex
  with (https:\/\/[^/]+) origin group compared to normalizeBaseUrl(serverURL)),
  which needs no escaping and no RegExp.escape. Decide based on the test run.
• deriveCommentKey(...): accept serverURL in its param object and forward it
  to parseGithubItemURL.
• Update the doc comments on both to say the host is anchored to the current
  GitHub server (GITHUB_SERVER_URL) rather than github.com, preserving the
  "user-controlled github-url cannot redirect to an attacker host" note.
• normalizeBaseUrl is already imported; reused for the server origin so a
  trailing slash / query / fragment on GITHUB_SERVER_URL is tolerated.

src/action.ts

• Add a small edge read, e.g. process.env.GITHUB_SERVER_URL || undefined,
  used in the three sites that currently hardcode dotcom:
  • parseGithubURL() → pass to parseGithubItemURL, and rebuild the error
    message to reference the resolved server URL instead of the hardcoded
    https://github.com/... example and the "rejects non-github.com hosts"
    phrasing.
  • commentOnIssue() → pass serverURL into deriveCommentKey.
  • handleFailure() → pass serverURL into deriveCommentKey.
• Passing undefined triggers the helper default (https://github.com), so
  local/non-Actions callers and existing behavior are unchanged.

src/comment.test.ts

Add parseGithubItemURL coverage (currently only exercised indirectly via
deriveCommentKey):

• dotcom URL parses with the default server (no serverURL arg).
• GHES URL (e.g. https://github.acme.example/owner/repo/pull/1) parses when
  serverURL matches.
• host that does not match serverURL is rejected (returns undefined) — the
  security-relevant case.
• a deriveCommentKey case passing a GHES serverURL yields owner/repo#n
  (not the raw-URL fallback).

The existing deriveCommentKey "falls back to raw URL for non-github.com host"
test still passes: with the default server, code.acme.com doesn't match.

Docs

No meaningful changes (per scope). The README.md github-url row and the
api_error troubleshooting line still read fine; leaving them avoids churn on
an edge case. (Can revisit if you want a one-line mention.)

Validation

In a workspace (dogfood template, coder org):

1. Clone coder/agents-chat-action, bun install.
2. bun test — all pass, new cases included.
3. bun run typecheck.
4. bun run lint (Biome).
5. bun run build if the built dist/ is committed and expected to stay in
   sync (verify whether the repo commits build output before touching it).

Delivery

• Commit on phorcys/ghes-github-url, author
  57866459+phorcys420@users.noreply.github.com.
• Conventional commit, scope covering changed files, e.g.
  fix(src): anchor github-url host to GITHUB_SERVER_URL for GHES.
• No push until you approve. Then open a PR referencing github-url host regex is hardcoded to github.com, blocking GitHub Enterprise Server #39 with the
  Coder Agents disclosure and this plan in a collapsible section.

Open / to confirm during implementation

• Whether dist/ is committed (affects step 5 and the diff size).
• Exact final error-message wording in parseGithubURL().

---

🤖 Opened by Coder Agents on behalf of @phorcys420.