hive-forge: stop embedding the forge token in clone URLs

This commit is contained in:
damocles 2026-08-26 00:20:58 +02:00
commit 676f7715fd
6 changed files with 154 additions and 27 deletions

View file

@ -144,7 +144,7 @@ lets a read-only user open a PR by pushing the current `HEAD` to the
magic ref `refs/for/<base>/<topic>`. Two verbs cover the workflow:
```
hive-forge -r internal/knowledge clone # clone with token auto-injected
hive-forge -r internal/knowledge clone # clone, auth handled for you
cd knowledge
# add / edit / delete any files, then commit normally
git add -A && git commit -m "add foo runbook"
@ -155,9 +155,13 @@ hive-forge -r internal/knowledge pr-create --agit \
```
`clone` derives the dest dir from the repo basename (override with a
positional arg); `--branch` / `--depth` are passed through. The token
is injected into the clone's `origin` remote so `pr-create --agit`
(default remote `origin`) can push without re-auth.
positional arg); `--branch` / `--depth` are passed through. The clone
URL and the `origin` remote it leaves behind carry no credentials —
`clone` instead configures `origin`'s `credential.helper` to invoke
`hive-forge credential-helper` (a hidden verb, not meant to be run by
hand), which git calls fresh on every fetch/push. That's what lets
`pr-create --agit` (default remote `origin`) push without re-auth,
without a durable token sitting in the checkout's `.git/config`.
`pr-create --agit` prints the PR URL. Re-running with the same
`--topic` force-updates the existing open PR (the AGit ref is

View file

@ -31,10 +31,12 @@ pub struct Client {
/// header the typed client sends.
web: HttpClient,
base: String,
/// Per-agent forge token. Kept so verbs that shell out to `git`
/// (e.g. `clone`) can assemble an authenticated push URL without
/// re-reading the token file.
token: String,
/// The `-f/--forge` label this client resolved against, `None` for
/// the default internal forge. `clone` bakes this back into the
/// `credential.helper` command it configures, so a later `git push`/
/// `git fetch` in that checkout re-resolves the *same* forge account
/// rather than silently falling back to the internal one.
forge_label: Option<String>,
/// Default repo used when a verb doesn't carry an explicit
/// `[repo]` override.
pub default_repo: String,
@ -98,7 +100,7 @@ impl Client {
api,
web,
base,
token,
forge_label,
default_repo,
json_mode,
})
@ -111,21 +113,26 @@ impl Client {
&self.api
}
/// Assemble an authenticated git URL for `repo` (e.g.
/// `internal/knowledge`) by injecting the agent's forge user +
/// token into the base URL's authority: `http://<user>:<token>@host/<repo>.git`.
/// The user comes from `HIVE_LABEL` (the agent's forge login),
/// falling back to `oauth2` which Forgejo also accepts as the
/// token-bearer username. Used by `clone` to clone/push.
/// The *credential-free* git URL for `repo` — no user/token in the
/// authority, so nothing durable lands in `.git/config` when this is
/// the URL `git clone` is given. Pairs with the `credential-helper`
/// verb, which `clone` configures as the repo's `credential.helper`
/// so git asks for (and gets) the token fresh from its file on every
/// fetch/push instead of it being embedded here. Replaces the old
/// `authed_git_url` (`http://<user>:<token>@host/<repo>.git`), which
/// left a durable token in every checkout's `.git/config` — a real
/// leak reported by atlas.
#[must_use]
pub fn authed_git_url(&self, repo: &str) -> String {
let user = std::env::var("HIVE_LABEL").unwrap_or_else(|_| "oauth2".to_owned());
// Split scheme from authority so credentials land in the right spot.
let (scheme, host) = self
.base
.split_once("://")
.unwrap_or(("http", self.base.as_str()));
format!("{scheme}://{user}:{}@{host}/{repo}.git", self.token)
pub fn plain_git_url(&self, repo: &str) -> String {
format!("{}/{repo}.git", self.base)
}
/// The `-f/--forge` label this client resolved against (`None` for
/// the internal forge). See the `forge_label` field doc for why
/// `clone` needs this.
#[must_use]
pub fn forge_label(&self) -> Option<&str> {
self.forge_label.as_deref()
}
/// True when the operator passed the global `--json` flag.
@ -331,7 +338,7 @@ pub fn index(n: u64) -> Result<i64> {
/// `/` or `..`) is rejected up front with the same charset spelled out,
/// rather than silently building a nonsense/traversing path and
/// surfacing a confusing file error later.
fn resolve_credentials(forge_label: Option<&str>) -> Result<(String, String)> {
pub(crate) fn resolve_credentials(forge_label: Option<&str>) -> Result<(String, String)> {
let Some(label) = forge_label else {
let base = std::env::var("HIVE_FORGE_URL").unwrap_or_else(|_| DEFAULT_URL.to_owned());
let token = read_token().context("read forge-token")?;

View file

@ -176,6 +176,13 @@ enum Verb {
/// Re-run CI without an empty commit. Pass one of `--pr <n>`,
/// `--run <n>`, or `--branch <name>`; `--workflow` defaults to `ci.yml`.
CiRerun(verbs::ci_rerun::Args),
/// Git credential-helper protocol (`get|store|erase`) — not a verb
/// you run by hand. `clone` configures each checkout's
/// `credential.helper` to invoke this, so git asks it for a token
/// fresh on every fetch/push instead of one being embedded in the
/// remote URL.
#[command(hide = true)]
CredentialHelper(verbs::credential_helper::Args),
}
/// Wrapper over [`run`] that owns how a failure reaches the operator.
@ -197,6 +204,14 @@ fn main() -> ExitCode {
fn run() -> Result<()> {
let cli = Cli::parse();
let verb = cli.verb;
// `credential-helper` bypasses the normal client construction below:
// it needs no repo (unlike `Client::from_env`, which errors without
// one), and git may invoke it from a directory that isn't a resolved
// checkout yet (mid-`clone`, before the destination exists).
if let Verb::CredentialHelper(args) = verb {
return verbs::credential_helper::run(args, cli.forge.as_deref());
}
let client = client::Client::from_env(cli.repo, cli.json, cli.forge)
.context("initialize forge client")?;
// Attach the resolved repo to every verb's error uniformly here,
@ -211,7 +226,7 @@ fn run() -> Result<()> {
// line 1 column 0" into "repo typo-org/repo: EOF while parsing a
// value at line 1 column 0", which is diagnosable on sight.
let repo = client.repo().to_owned();
dispatch(&client, cli.verb).with_context(|| format!("repo {repo}"))
dispatch(&client, verb).with_context(|| format!("repo {repo}"))
}
fn dispatch(client: &client::Client, verb: Verb) -> Result<()> {
@ -253,5 +268,8 @@ fn dispatch(client: &client::Client, verb: Verb) -> Result<()> {
Verb::ArtifactGet(a) => verbs::artifact_get::run(client, a),
Verb::CiLog(a) => verbs::ci_log::run(client, a),
Verb::CiRerun(a) => verbs::ci_rerun::run(client, a),
Verb::CredentialHelper(_) => {
unreachable!("handled in `run` before client construction")
}
}
}

View file

@ -4,6 +4,15 @@
//! `client::Client::from_env` for the full resolution chain). Pairs
//! with `pr-create --agit`: clone, edit + commit normally, then open a
//! PR via the `AGit` ref.
//!
//! The clone URL itself carries no credentials, and the resulting
//! checkout's `.git/config` never gets one either: `-c
//! credential.helper=!hive-forge credential-helper` (baked into the
//! `git clone` invocation, which persists into the new repo's config)
//! makes git ask the `credential-helper` verb for a fresh token on every
//! subsequent fetch/push instead. Previously the token itself rode in
//! the clone URL and landed, durably, in every checkout's `.git/config` —
//! a real credential leaked to every clone on disk, reported by atlas.
use std::process::Command;
@ -40,9 +49,17 @@ pub fn run(client: &Client, args: Args) -> Result<()> {
.map(str::to_owned)
.with_context(|| format!("clone: cannot derive a destination dir from repo {repo}"))?,
};
let url = client.authed_git_url(repo);
let url = client.plain_git_url(repo);
let helper = match client.forge_label() {
Some(label) => format!("!hive-forge credential-helper -f {label}"),
None => "!hive-forge credential-helper".to_owned(),
};
let mut git_args = vec!["clone".to_owned()];
let mut git_args = vec![
"clone".to_owned(),
"-c".to_owned(),
format!("credential.helper={helper}"),
];
if let Some(depth) = args.depth {
git_args.push(format!("--depth={depth}"));
}

View file

@ -0,0 +1,80 @@
//! `credential-helper <get|store|erase>` — implements git's
//! credential-helper protocol (`gitcredentials(7)`) so a cloned repo's
//! `.git/config` only ever holds a *reference* to this command, never
//! the forge token itself. `clone` configures the new repo's
//! `credential.helper` to invoke this subcommand; git then calls it
//! fresh on every fetch/push, reading the token from its usual
//! on-disk location (the per-agent `forge-token` file, or the
//! `--forge <label>` sidecar) at auth time rather than the token
//! having been baked into the remote URL and left durably in
//! `.git/config`. A real leak of that shape, found and reported by
//! atlas, is what this verb exists to close.
//!
//! Deliberately bypasses the normal `Client::from_env` construction in
//! `main.rs` (see the special-case there): that path requires resolving
//! an active repo, which can fail depending on where git happens to run
//! this helper from (e.g. mid-`clone`, before the destination checkout
//! exists) — and a credential lookup has no need for a repo anyway.
//!
//! Only `get` does anything: reads the `key=value` lines git sends on
//! stdin, and if a `host=` line is present, refuses to answer unless it
//! matches the resolved forge's own host (defence in depth against this
//! helper ever being invoked outside the specific-URL scope `clone`
//! configures it under). `store`/`erase` drain stdin and no-op — there
//! is nothing durable to store or erase, the token always comes fresh
//! from its file.
use std::io::Read;
use anyhow::{Context, Result, bail};
use clap::{Args as ClapArgs, ValueEnum};
use crate::client::resolve_credentials;
#[derive(ClapArgs)]
pub struct Args {
/// The git credential-protocol operation git invokes this with.
op: Op,
}
#[derive(Clone, Copy, ValueEnum)]
enum Op {
Get,
Store,
Erase,
}
/// # Errors
///
/// Returns an error if credential resolution fails (no token
/// provisioned for the active forge) or the `get` request's stdin
/// can't be read, or if a `host=` line in the request doesn't match
/// the resolved forge.
pub fn run(args: Args, forge_label: Option<&str>) -> Result<()> {
let mut input = String::new();
std::io::stdin()
.read_to_string(&mut input)
.context("read credential request from stdin")?;
let Op::Get = args.op else {
// `store`/`erase`: protocol says read the request, do nothing.
return Ok(());
};
let (base, token) = resolve_credentials(forge_label)?;
if let Some(requested) = input.lines().find_map(|l| l.strip_prefix("host=")) {
let expected = url::Url::parse(&base)
.ok()
.and_then(|u| u.host_str().map(str::to_owned));
if expected.is_some_and(|e| e != requested.trim()) {
bail!(
"hive-forge credential-helper: refusing to hand the forge token to \
host {requested:?} (configured forge is {base:?})"
);
}
}
let user = std::env::var("HIVE_LABEL").unwrap_or_else(|_| "oauth2".to_owned());
println!("username={user}");
println!("password={token}");
Ok(())
}

View file

@ -16,6 +16,7 @@ pub mod comment;
pub mod comment_edit;
pub mod comment_show;
pub mod comments;
pub mod credential_helper;
pub mod dependency;
pub mod diff;
pub mod issue;