feat(server): authenticate server-side git clone and fetch with a company-secret GitHub token (#10720)

## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work
> - Repo-only project workspaces are materialized by a server-side `git
clone`, and isolated `git_worktree` runs refresh their base ref with
server-side `git fetch`
> - Both operations run outside the agent process with no credentials,
so private GitHub repositories can never be cloned or refreshed —
agent-scoped credential env bindings do not reach them
> - The company secret store already has a well-known GitHub token
convention (`GITHUB_TOKEN` / `GH_TOKEN` / `PAPERCLIP_GITHUB_TOKEN`,
consumed by the external-object provider for API reads), but nothing
server-side consults it for git
> - This pull request resolves that token per run and authenticates the
managed clone and every base-ref refresh with it through an ephemeral
credential helper
> - The benefit is that isolated workspaces work on private repositories
with one company secret, while public repositories and self-hosted
ambient git configuration keep working unchanged

## Linked Issues or Issue Description

**Subsystem affected**

Server workspace materialization (`server/src/services/heartbeat.ts`)
and execution-workspace realization
(`server/src/services/workspace-runtime.ts`).

**Problem or motivation**

A project workspace configured with only a private GitHub `repoUrl`
cannot be used for isolated `git_worktree` runs: the managed `git clone`
runs with a sanitized, credential-less environment, and plain git cannot
consume a bare token env variable without a credential helper. There is
no way to give the server a git credential — storing a `GH_TOKEN`
company secret has no effect on server-side git, and a credential-less
private clone hangs on a terminal prompt until the ten-minute clone
timeout. Base-ref refreshes (`git fetch`) during worktree realization
have the same gap.

**Proposed solution**

A `git-credentials` module resolves a token per run — company secret by
well-known name (`GITHUB_TOKEN`, `GH_TOKEN`, `PAPERCLIP_GITHUB_TOKEN`),
then `GITHUB_TOKEN`/`GH_TOKEN` in the server process environment for
self-hosted deployments, then none — and builds a git invocation that
authenticates via an inline credential helper. The token travels in an
env variable; it never appears in argv, URLs, or on disk. Only
`https://github.com` remotes are authenticated; everything else keeps
ambient behavior. The provider is a single factory seam so a future
brokered credential source can replace it without touching call sites.

**Alternatives considered**

- A GitHub OAuth "connect your account" flow: heavier product surface,
needs app registration and callback custody; out of scope for a server
credential and better served by a dedicated connector later. The
provider seam keeps that path open.
- `gh auth setup-git`: writes helper configuration to disk and requires
a global token env; rejected in favor of per-invocation config with no
persistent state.
- Embedding the token in the clone URL: leaks into argv, error messages,
and `.git/config`; rejected.

## What Changed

- New `server/src/services/git-credentials.ts`:
`createGitRemoteAuthProvider` (memoized per run, one secret resolution
and one audit event), `buildGitAuthInvocation` (helper-reset + inline
helper, `x-access-token` username, `GIT_TERMINAL_PROMPT=0`),
`isGitHubHttpsRemoteUrl` host gating (rejects ssh/GHES/http/other
hosts/userinfo URLs), `describeGitAuthFailure`, and the canonical
`scrubGitCredentialText`. Secret resolutions pass a `system` consumer
access context so they are recorded as secret access events.
- `ensureManagedProjectWorkspace` (now exported) accepts an optional
auth provider; the clone env spreads the token after
`sanitizeRuntimeServiceBaseEnv` (which strips `PAPERCLIP_*`), always
sets `GIT_TERMINAL_PROMPT=0`, distinguishes "credential rejected" from
"no credential configured — add a GITHUB_TOKEN or GH_TOKEN company
secret" in the error, and removes the partially created directory on
clone failure so a timeout-killed clone cannot be adopted as a broken
checkout by the next run.
- `refreshRemoteTrackingBaseRef` (now exported) captures the remote URL
it already looked up, asks the provider for an invocation, and
attributes failed authenticated fetches to the credential in a scrubbed
warning. The optional provider threads through `detectDefaultBranch`,
`resolveAuthoritativeBaseRef`, `inspectExecutionWorkspaceBaseDrift`,
`realizeExecutionWorkspace`, and
`ensurePersistedExecutionWorkspaceAvailable`; heartbeat builds one
provider per run for both the anchor-resolution clone path and workspace
realization/restore.
- `github-external-object-provider.ts` imports the shared secret-name
list; `isGitHubDotCom` is exported from `github-fetch.ts`.
- Docs: "Private repositories and repo-only project workspaces" section
in the execution-workspaces guide, cross-linked from the secrets deploy
doc.

## Verification

- `cd server && npx vitest run src/__tests__/git-credentials.test.ts` —
resolution chain order and precedence, env fallback, memoization,
audited access context, host-gating matrix, invocation shape (token
absent from argv), scrubber, failure descriptions, and a real-git `git
credential fill` round trip that proves the helper executes and answers
with the env-carried token (no network).
- `cd server && npx vitest run
src/__tests__/heartbeat-managed-clone-credentials.test.ts` — clones
behave byte-identically with no provider or a null-returning provider
(local repos, no network), authenticated-failure errors name the
credential, non-auth failures do not mention credentials, partial clone
directories are removed, pre-existing non-git directories keep the
"Using it as-is" path, and the sanitizer spread order keeps the token
env alive.
- `cd server && npx vitest run src/__tests__/workspace-runtime.test.ts`
— new `refreshRemoteTrackingBaseRef` cases: provider offered the remote
URL and null keeps behavior identical; failed authenticated fetch
warning names the credential; unauthenticated failure warning stays
credential-free.
- `pnpm --filter @paperclipai/server typecheck` is clean.
- Manual (optional, networked): store a `GH_TOKEN` company secret,
configure a repo-only project workspace pointing at a private GitHub
repository, run an isolated-workspace issue — the managed clone succeeds
and the worktree run proceeds.

## Risks

- Every new parameter is optional; with no provider the git invocations
are byte-identical to before. Public repos and ambient credential
helpers keep working whenever no token resolves.
- Precedence change when a token exists: a stored company secret now
wins over ambient helpers for `https://github.com` remotes (the helper
list is reset for that invocation). The rejected-credential error names
the secret so an operator can fix or remove it.
- `GIT_TERMINAL_PROMPT=0` on the managed clone is the one always-on
change: a credential-less private clone now fails fast with a clear
message instead of hanging until the ten-minute timeout (it could only
ever "succeed" interactively on a TTY dev server).
- The token is scoped to the git process env for one invocation; it is
never written to agent env, run context, disk, or logs, and error text
is scrubbed of URL userinfo.
- No migrations, no image changes (git ships in the image).

## Model Used

Claude Fable 5 (`claude-fable-5`, extended thinking, agentic tool use
via Claude Code CLI).

## Checklist

- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge
This commit is contained in:
Devin Foley authored and GitHub committed 2026-08-02 20:27:09 -07:00
1 parent 97590ff8c4
commit e0c2448267
10 files changed
+835 -28

No files matched your search

+7
View File
@@ -51,6 +51,13 @@ Project env applies to every issue run in that project. When a project env key
matches an agent env key, the project value wins before Paperclip injects its
own `PAPERCLIP_*` runtime variables.
Separately from env bindings, the **server itself** consumes a company secret
named `GITHUB_TOKEN`, `GH_TOKEN`, or `PAPERCLIP_GITHUB_TOKEN` (by name, no
binding needed) to authenticate server-side git operations — cloning private
GitHub repos for repo-only project workspaces and refreshing worktree base
refs. See
[Execution workspaces](../guides/board-operator/execution-workspaces-and-runtime-services.md#private-repositories-and-repo-only-project-workspaces).
## User-Specific Secrets
User-specific secrets let a shared agent or project declare a slot such as
@@ -77,6 +77,32 @@ Some workspaces need heavy one-time setup — seeding a database, warming caches
- **Provisioning failed** — the command failed; the workspace detail links to the runtime logs for the failing operation.
- While the command runs, the runtime service shows a **Provisioning…** state before it transitions to starting/running.
## Private repositories and repo-only project workspaces
A project workspace can be **repo-only**: a `Repo URL` with no local path. The server then
materializes a managed checkout on demand (`git clone` into a managed directory) and, for
isolated `git_worktree` runs, refreshes the base ref (`git fetch`) before preparing each
worktree. Both operations run on the server, outside any agent process — so agent-scoped
credential env bindings do not apply to them.
For **private GitHub repositories**, store a token as a **company secret** named one of
`GITHUB_TOKEN`, `GH_TOKEN`, or `PAPERCLIP_GITHUB_TOKEN` (checked in that order; Settings →
Secrets). The server resolves it per run and authenticates managed clones and base-ref
fetches with it. Details and caveats:
- Scope: only `https://github.com/...` repo URLs are authenticated this way. SSH URLs, GitHub
Enterprise hosts, and other providers keep ambient behavior (system git config/credential
helpers on the server host). URLs that embed their own credentials are never overridden.
- Fallback: with no matching company secret, the server falls back to a `GITHUB_TOKEN` or
`GH_TOKEN` variable in the **server process environment** (useful for self-hosted single-tenant
deployments), then to unauthenticated access — public repos keep working with no setup.
- The token never appears in command lines, URLs, or on disk; it is passed to git through an
ephemeral credential helper. Each resolution is recorded as a secret access event.
- This is separate from the **agent push credential**: agents pushing branches/PRs still need
`GH_TOKEN`/`GITHUB_TOKEN` bound at agent or project scope (see
[deploy/secrets](../../deploy/secrets.md)) so the token reaches the agent process env. The
same company secret can back both uses via a binding.
## Cross-run persistence (no-remote-git contract)
Code state moves between runs through the local execution-workspace cwd alone — not through a git remote.