Personal runners: threat model and controls

A personal runner is the preloop CLI running as a service on a developer's own machine. In 0.17.0 it can run flow jobs on the harnesses installed there (for example GitHub Copilot CLI on the developer's own seat) and, after explicit opt-in on the host, accept remote sessions started from the Preloop console. That makes the laptop reachable from Preloop. This page states what that means, which controls apply, and what wave 1 does not prevent.

The controls are numbered T1 to T12. Pull requests that implement them cite these ids and ship a test per id.

Scope of wave 1 (0.17.0)

• Harness inventory published by the runner (#1480).
• Flows routed to a runner by harness and model (#1481).
• Remote sessions for copilot_cli only, started from the web console (#1482 runner protocol, #1483 HTTP API and console).
• Workspaces: authorized directories on the host, or temporary checkouts of a tracker repository with a short-lived clone credential (#1484).
• Install prompt after preloop login and the account runner policy (#1479, EE #182).

Deferred to 0.18.0: system versus user service split (#1488), runner sharing to teams (EE #183), MDM enrollment and device credential (#1344 track), more session harnesses (#1489, #1490).

Workstation correlation

Persistent runners report the same account-salted HMAC-SHA256 workstation fingerprint as preloop agents discover --report. This self-reported correlation key can link a runner to discovery evidence on the same machine and account. It is not verified device identity or attestation. Different account salts produce different fingerprints. Ephemeral runners omit this key.

Both commands use the platform machine ID, falling back to a random ID saved in the CLI configuration directory. To use an administrative reference instead, pass the same --device-ref to preloop runner fg and preloop agents discover --report. The runner saves that reference locally in runner.json for service restarts; only its salted hash is reported. Pass a new reference to replace it. The runner refreshes the account salt every five minutes and sends the refreshed fingerprint on subsequent heartbeats.

Runner API responses expose the fingerprint only to the registrant and same-account users with manage_agents. Older clients can still register without it, and older servers without the discovery salt endpoint remain usable.

Assets and actors

Asset	Where it lives
Source code in authorized directories	Host disk
Harness login (Copilot OAuth token, GitHub tokens, keychain items)	Host user session, owned by the harness
Runner token	~/.preloop/ on the host, rotatable with preloop runner rotate-token
Clone credential for a tracker checkout	Runner process memory, for one clone
Session events, approvals, audit	Preloop server

Actor	Trust
Host user (runner owner)	Trusted for their own machine; decides what is enabled
Account owner or admin	May start sessions on members' runners in wave 1 (T2)
Other account members	No session access to someone else's runner
The model and the harness	Untrusted output; may be steered by prompt injection in code or issues
A party holding a stolen Preloop session	Treated as the actor it impersonates; limited by T2 to T12

Controls

T1. Per-harness opt-in on the host

Remote sessions are off for every harness until the host user runs preloop runner sessions enable <harness> on that machine. The server can never set sessions_enabled; it only reads it from the inventory. A harness can also be disabled entirely in ~/.preloop/runner.json ("harnesses": {"<id>": {"enabled": false}}); disabled harnesses are still reported for compliance but are ignored by routing and sessions. preloop runner sessions disable <harness> revokes at once. Both changes are audited as runner.sessions_enabled and runner.sessions_disabled with the host user as actor. Runner rejects with harness_not_enabled_for_sessions. Implemented by #1482 (command and runner check) and #1480 (inventory field).

Consent text printed by preloop runner sessions enable copilot_cli, which then asks Enable? [y/N] (default no):

Allow remote Copilot CLI sessions on this machine?

Your account owner, account admins and you will be able to start GitHub Copilot
CLI sessions here from Preloop, using your Copilot seat and login. Sessions run
as you, only in directories you authorize with `preloop runner dirs add` or in
temporary checkouts. Every tool call goes through Preloop approvals and is
audited. You get a notification when a session starts.

Turn off any time: preloop runner sessions disable copilot_cli

T2. Who may start a session

Wave 1 rule: the runner owner (registered_by_user_id) and account owners and admins. Everyone else gets 403 not_runner_owner. The check is the authorizer action runner:session_start (ACTION_RUNNER_SESSION_START) so team sharing (EE #183) plugs in later without changing the endpoint. The admin path is disclosed to the host user in the T1 consent text and in the T3 notice, which names the actor. Implemented by #1483.

T3. Host-visible notice

When the runner accepts a session_start, it writes a runner log line and raises a desktop notification as the logged-in user: Preloop: <actor> started a <harness> session in <label> (Windows toast, macOS user notification, Linux notify-send when present). A failed notification is logged and does not block the session. Implemented by #1482.

T4. Workspace containment

A session runs either in an authorized directory or in a temporary checkout. Authorized directories are listed only on the host (preloop runner dirs add|remove|list); the server sees {id, label, mode, harnesses}, never the path. Paths are resolved with realpath when loaded and again at use; symlinks and Windows reparse points cannot escape; comparison is case-insensitive on Windows; the filesystem root, a whole home directory and UNC admin shares are refused. Unknown or out-of-scope ids fail with workspace_not_authorized. Temporary checkouts live under ~/.preloop/host-workspaces/sessions/<id> and are deleted when the session ends. Implemented by #1484.

T5. Short-lived clone credentials

For a tracker checkout the server mints a credential immediately before session_start: a GitHub App installation token scoped to the one repository with contents: read (at most 60 minutes), or a fresh Bitbucket Cloud OAuth access token (about 2 hours). Long-lived secrets are never sent to a laptop: GitHub PAT trackers fail with checkout_requires_app_or_oauth, Bitbucket access tokens and app passwords fail with checkout_requires_oauth. The credential travels only inside the authenticated runner websocket message, is never stored server-side, never logged, never written to .git/config or any file, is passed to git through an in-memory askpass channel and is zeroed after the clone. Minting is audited as runner_session.checkout_credential_minted (provider, repository, expiry; never the token). Implemented by #1484.

T6. Harness credentials are never touched

The runner executes the user's own installed, unmodified harness binary. It does not open the harness credential store or copy or forward its credentials. Inventory may query the harness's read-only login-status interface. Copilot's auth.getStatus response is reduced to isAuthenticated; the raw response and any account or token fields are never logged or published. Inventory reports only login_state and login_source; for env it checks that a variable name is set and never reads the value. The child environment is the existing host-exec allowlist. Any code that reads a harness credential store is a review blocker. Implemented by #1480 (inventory) and #1482 (session launch).

T7. Mandatory approvals and deny-tool defaults

Remote sessions are a prompt injection target: the model reads code, issues and tool output it did not write. A remote session starts only when the Preloop approval hook (preloop agents onboard "Copilot CLI" --approvals) is installed for the host user; otherwise the runner refuses (copilot_approval_hook_missing in the existing host-exec code). Every tool call is evaluated by Preloop policies and can require a console approval. The default deny-tool list from the host-exec profile applies, deny always wins over allow, and --allow-all-tools is never used without the hook. Implemented by #1482.

T8. Copilot specifics

• --no-remote is always passed, so the session cannot also be steered through GitHub's own remote control.
• --no-ask-user and non-interactive -p turns; each turn is one process resumed with --resume=<session uuid>.
• BYOK variables (for example COPILOT_PROVIDER_BASE_URL) are stripped by the environment allowlist, so a seat session really uses GitHub-hosted models under the user's Copilot plan and policies.
• Model-judged (assisted) permission approval is never enabled: the runner does not pass --assisted-approval (present in Copilot CLI 1.0.95 help, marked experimental) and does not switch to the /permissions assisted mode. Tool decisions come only from the Preloop approval hook (T7).

Implemented by #1482.

T9. Limits

Defaults: at most 2 concurrent remote sessions per runner (max_concurrent_reached), idle timeout 30 minutes (idle_timeout), a maximum session duration (max_duration), one turn in flight per session (409 turn_in_progress), event payloads bounded to 64 KB and redacted by the existing redaction. Implemented by #1482 and #1483.

T10. Kill switch and stop

The account kill switch (see Account kill switch) applies to remote sessions: halting flows or agents ends live sessions with killed_by_kill_switch and rejects new starts. The actor can stop a session from the console (graceful or kill); the host user can stop it locally (stopped_on_host). A stop mid-turn kills the harness process. Implemented by

1483 (server) and #1482 (runner).

T11. Audit

Every step is an audit event with resource_type = runner_session and details that always carry actor, runner id and name, host, harness, model and workspace kind and label: runner_session.start_requested, .started, .rejected, .turn_sent (turn id and text length, not the text), .stop_requested, .ended (with end_reason), .checkout_credential_minted, plus runner.sessions_enabled, runner.sessions_disabled and runner.harness_inventory_changed. Implemented by #1483 and #1482.

T12. Revocation and policy

Ways to cut access, from narrowest to widest: stop the session (T10); disable the harness for sessions on the host (T1); remove an authorized directory; preloop runner disable or stop on the host; rotate the runner token (preloop runner rotate-token) or delete the runner in the console; account kill switch. The account runner policy (runner_policy: optional, required or forbidden) decides whether the CLI asks, informs or stays silent after login. Implemented by #1479 (login prompt and field) and EE #182 (policy and compliance view).

Consent text shown after preloop login when the policy is optional (#1479):

Install a Preloop runner on this machine?

The runner starts at login, runs your Preloop flows locally and reports which
coding agents are installed here (names, versions, signed-in state; never your
credentials). Remote sessions stay off until you enable them per agent.

Install and start the runner now? [Y/n]

When the policy is required, the same text is printed without the question, followed by the account message. When it is forbidden, one line says so.

What leaves the machine

Sent to Preloop:

• Harness inventory: harness ids, versions, signed-in state, model ids, governance state. Never executable paths, argv, environment values or tokens.
• Authorized directory ids, labels, modes and harness lists. Never paths.
• Session events: agent messages, tool calls and results, usage, stderr, bounded and redacted (T9).
• Approval requests and decisions, turn summaries, audit events.

Not sent to Preloop: repository content as such. The harness itself sends what it needs to its own model provider (GitHub for Copilot) under the user's seat, as it would in a local session; tool results the agent produces can include file excerpts and are visible in the session events.

Laptop sleep and offline

• The runner reconnects after sleep or network loss and reports session_state for every live session, so delivery resumes.
• If the runner stays offline, the server ends its sessions with runner_offline; turns are not queued for an offline runner. Flows that need the harness follow their harness_fallback (queue, fallback_server or fail).
• Idle sessions end after the idle timeout (T9) whether or not the laptop is awake. A Copilot session file survives on disk, but a new turn requires a new session.
• Temporary checkouts are deleted when the session ends; if the runner was offline, cleanup happens on its next start.

What wave 1 does not stop

• A user can uninstall or stop the runner, or use a harness on a machine that never enrolled. Wave 1 detects this, it does not prevent it: the compliance view (EE #182) shows members as stale or never_connected and lists ungoverned harnesses. Prevention belongs to the MDM track (#1181 managed-agent deployment recipes, #1186 managed-config bundles).
• An account admin can start a session on a member's runner without asking first. The member has consented (T1) and is notified (T3); per-session approval by the host user is not in wave 1.
• Anything the harness does outside a Preloop session, and what the harness sends to its own model provider.
• A compromised host user account: the runner runs as that user and inherits its access.