fx Coding Agent

Workflows default to contents:read, pin some actions to commits, and validate downloaded artifact provenance, status, size, and commit relationships. The license, third-party notice path, and sound credit make source attribution strong. These are mostly build-infrastructure controls, however, and do not establish least privilege for the agent runtime. The README links to permission documentation but does not show tool-call confirmation rules, filesystem or command scopes, provider data flows, or OAuth/API-key storage and deletion. The agent can modify code and supports MCP, skills, and subagents, yet the supplied files do not explain control of external effects or recovery. Confirmation, transparency, sensitive-data handling, and external-effects scores are therefore thin, and rollback is absent. No red-line behavior is evidenced.

README installation, build, Zig-version, and test instructions are broadly consistent with the workflows. CI covers multiple architectures, startup performance, memory, native/WASM bridges, and artifact identity checks. Some dependencies and tools are precisely versioned, while other actions use moving major tags and installation relies on remote downloads, so availability is not fully controlled. PGSO tests show useful rejection messages for missing parameters, stale output, wrong runtime versions, and incomplete evidence, and the driver records failure stages and errors. Those tests concern an auxiliary optimization pipeline rather than the core interactive agent, preventing full marks for failure messaging.

The README clearly addresses interactive-shell, scripted ask, ACP, JavaScript/WASM embedding, browser, and framework scenarios. It supports gateways, subscription accounts, local models, and OpenAI-compatible endpoints, providing thorough audience and scenario coverage. Commands, environment variables, configuration precedence, and gateway-only routing conditions make triggers precise. Experimental labels establish some boundaries, but the supplied material does not comprehensively enumerate unsupported operations, model differences, or MCP/subagent limits. CI supports Linux, macOS, multiple architectures, Node, and Bun, while Windows, terminal prerequisites, and a complete compatibility matrix are not shown.

The README has clear sections for installation, onboarding, providers, themes, embedding, source builds, security, and licensing. Commands and configuration names appear stable and numerous copyable examples are present. Installation is concise, but the piped curl installer lacks documented verification, uninstall, or version-pinning instructions in the supplied material, and no FAQ is included. Experimental status is disclosed, though concrete known issues are sparse. The full Apache-2.0 text, copyright notice, and third-party attribution path justify full license marks. A release badge exists, but no changelog or compatibility policy is supplied. Vercel Labs labeling, a security contact, and a contribution-guide reference indicate an update path, while the publisher remains unverified and no named maintainers or support commitment are shown.

The product offers interactive, one-shot, and embedded output surfaces and explicitly preserves terminal scrollback. Benchmark workflows gate latency, memory, capacity, and binary-size behavior, showing attention to usability and operating cost. A small native, model-agnostic agent with several host surfaces offers plausible marginal value. However, the supplied evidence contains no benchmark results, task-quality evaluation, monetary cost figures, or substantial output examples, and the product is explicitly experimental. Output usability, marginal value, and cost-benefit are therefore adequate but not thoroughly demonstrated.

Major claims map to concrete commands, configuration examples, build steps, and workflow checks. README, license, tests, and CI corroborate the Zig requirement, ReleaseSafe builds, runtime surfaces, and attribution. Workflows retain JSON reports, logs, samples, and commit-identity artifacts, supporting traceability. Still, the supplied evidence omits the core agent implementation, linked manual contents, release results, and actual benchmark data, so promotional claims such as “tiny,” “any model,” and a complete manual are only partly checkable. Facts, examples, and experimental warnings are generally separated, but measured findings versus inference are not systematically labeled.