Skip to the content.

Security

CLike drives cloud models and local coding agents that generate and execute code and write files. This document describes the security model of the current release, its known limitations, and how to report a vulnerability.

Deployment assumption

CLike is designed for a single developer on localhost. The services are not meant to be exposed on a network or shared between users.

Security model (1.10.2)

Area Control
Service perimeter Every orchestrator and gateway endpoint except /health requires Authorization: Bearer <CLIKE_API_TOKEN> (constant-time comparison). Services fail closed (HTTP 503) when no token is configured. Ports are published on 127.0.0.1 only.
Browser threats No CORS. Requests with unexpected Host headers are rejected (DNS rebinding). The extension’s local MCP server is disabled by default and, when enabled, always requires a token and rejects any request carrying an Origin header.
Service-to-service Internal calls carry the service token.
Secrets in the extension The service token is stored in VS Code SecretStorage and sent only to the configured orchestrator/gateway origins.
Workspace settings Settings that select binaries, agent permission/sandbox modes, service URLs, Git remotes/push are machine-scoped; untrusted workspaces are not supported.
File confinement Paths originating from requests, models or agents are confined to their roots (workspace, runs/kit/<REQ>, telemetry), rejecting traversal, absolute, drive/UNC paths and symlink escapes. The extension is the only workspace writer.
Project context sources .clike/context.yaml sources are read by the orchestrator only inside the workspace (patterns with .. or absolute paths and symlinks leaving it are ignored), read-only, size-capped, and presented to models as data, never as instructions; canonical documents win on conflict.
Process execution No shell for process execution in the extension (argv only); on Windows the agent prompt goes through stdin.
Evaluation inputs The gate executes the LTC profile stored in the workspace; an inline profile is accepted only if identical. The project root must be inside the configured projects directory.
Evaluation environment Eval commands run without credentials (API keys, service token and similar variables are removed). What the checks need is created per run in the sandbox (generated sessions/ids, disposable PostgreSQL databases on the developer-configured CLIKE_EVAL_POSTGRES_URL server, services of other REQs on loopback ports, stopped after the checks); developer values come only from the git-ignored docker/eval.env.
Evaluation sandbox Eval/gate commands run in a separate container with no credentials, no access to the gateway or vector store, non-root, read-only root filesystem, dropped capabilities and resource limits, projects mounted read-only. An offline mode removes network egress. It runs with an init that reaps orphaned processes; an exhausted sandbox (processes, memory, disk) stops the eval as environment-blocked instead of sending correct code to a repair. A local agent re-running checks (eval-check) goes through the extension and never receives the service token or network access.
Eval tools Native tools for the checks are installed only from CLike’s catalog (or the project’s .clike/eval-tools.yaml) after the developer’s confirmation, by a separate service reachable only by the orchestrator, without credentials; vendor archives are https-only and verified by sha256; the sandbox mounts them read-only.
Acceptance integrity The acceptance surface of each requirement (tests, CI profile) is locked server-side before evaluation; removed/modified tests, weakened profiles and added skip markers block the gate. Auto-eval repairs may amend it only in audited, non-weakening ways (LTC command fixes, ci/ dependency upgrades, unused imports, and test fixes with every assertion unchanged, flagged review_required at the gate).
Gate override Manual overrides require a reason, are recorded in an audit log with an artifact digest and are reported as OVERRIDE, never PASS.
Gate verdict reuse The gate reuses an eval verdict only when a digest the orchestrator computes from the project itself (code, KIT, every LTC, plan.json, CLike version) matches the one recorded with the passing eval in CLIKE_STATE_DIR, which agents cannot write; acceptance integrity and required outputs are still checked. A report edited by hand never produces a PASS.
KIT repository-root files A KIT writes root deliverables only under runs/kit/<REQ>/root/; promotion writes them only within an allowlist (scripts/, deploy/, infra/, … plus .clike/project.json root_folders / root_files) and never .git, .clike, runs, docs, src, test, ci, .env* or keys; anything else is reported rejected and not written.
Containers Services run as a non-root user; the CLike repository (including .git) is mounted read-only.
Git Only the files of a phase are committed; no branch is rewritten; commit, push, merge-on-gate and PR automation are opt-in.
Logging Request bodies are never logged; unhandled errors return a correlation id instead of internal details.
Repository A pre-commit/CI hygiene check blocks telemetry, private notes, keys and secrets from being committed.

Known limitations

Reporting a vulnerability

Please do not open a public issue for security problems. Use GitHub’s private vulnerability reporting for this repository (Security → Report a vulnerability), including affected version, reproduction steps and impact. We aim to acknowledge reports within a few working days.