GiviLoop 0.4.0 — Browser choices and quieter reviews

GiviLoop 0.4.0

Double-check code with a reviewer you choose, keep the evidence, and use existing chat access instead of an additional model API call where applicable. This release adds browser choices and quieter background startup to the CLI/MCP review workflow.

Changes

• Browser adapters for Claude, DeepSeek and Gemini, with separate persistent profiles, provider-aware login/check commands and CLI/MCP delivery. No model API keys or remote API inference adapters were added.
• Real anonymous Gemini reviews and signed-in Claude Free/DeepSeek reviews now complement the existing ChatGPT authenticated/anonymous checks. New adapters accept inline text and use the website's current/default model; automated ZIP uploads and model selection remain ChatGPT-only.
• Background Chrome starts without a startup window, then creates a minimized target without activating it. Two real Claude reviews on macOS completed without the owned Chrome processes becoming foreground; a ChatGPT review also completed. Setup, human verification and uploads can still show a window. Some Linux desktops restore Chrome during input; it is minimized again after sending.
• Correct handling of maximized windows, initial Claude/Gemini cookie choices and response actions outside the assistant text. Completion checks distinguish answer text from thinking status, user messages and code-copy controls. Uncertain submissions are not automatically retried.
• Repeatable source-checkout acceptance script: node scripts/web-acceptance.mjs --provider NAME. It sends synthetic buggy and corrected code, checks delivery and source preservation, and saves responses for assessment. It is a development script, not an installed CLI command.

DeepSeek: delivery is not review accuracy

DeepSeek completed both real requests, identified the two deliberately introduced regressions, and saved its answers through MCP. Its corrected-code review nevertheless claimed a bug using negative quantities, while the supplied contract explicitly required positive quantities. That is a false positive, not a failed login or broken response transfer.

It also used the wrong property name in one reproduction and suggested a NUL-separated cache key that collides for arbitrary string identifiers. Those suggestions were rejected; no generated patch was applied. Claude passed the same clean-code control. These few synthetic cases are not a provider ranking or a model benchmark. Findings from every provider require checking against source, contracts and tests. Detailed evidence.

Scope and known limits

The optional website adapters remain experimental. Websites can change their UI, require login or verification, impose quotas or reject automation. Headless ChatGPT remains blocked in the live checks; use --background. Specific subscription models, long web reasoning, authenticated Gemini, Windows live browser behavior and every macOS fullscreen/Spaces arrangement are not certified.

Existing local runtimes and manual transfer remain available. DwarfStar's native path was exercised with synthetic weights; trained-model review remains unvalidated on the available hardware. MLX's upstream server is experimental. Local engines.

Avoiding a separately billed API review can avoid that API token charge; GiviLoop does not measure total token or monetary savings. Chat plans, quotas and provider terms still apply. No provider authorization is claimed or established by this release. Costs and access.

Validation and installation

The release package is installed in a clean consumer and checked with 156 unit tests, 42 browser tests and a production dependency audit. CI also runs the unit suite on Node 20/22/24 on Linux/macOS. The attached release-checks.json, TAP logs and SHA256SUMS identify the actual checked archive. Browser fixtures are independent of user accounts; real website results are documented separately.

Download giviloop-0.4.0.tgz from this GitHub release, then:

npm install -g ./giviloop-0.4.0.tgz
givi --version
givi help

Node.js 20+ is required. givi-mcp starts the MCP stdio server. Restart an existing MCP server after upgrading. There is no npm registry publication; install the GitHub tarball or build from source. Existing profiles and runs remain in place. The 0.3.x tags/assets are unchanged.

Start with the README or Double Check recipe. The linked video was recorded with 0.3.1 and remains a demonstration of the workflow, not a new recording for 0.4.0.

Next

Collect first-review friction, independently confirmed findings, false positives and repeat use before adding more providers. Guided setup, structured finding evidence and a targeted recheck are proposed next steps, not features included in this release. Adoption plan and roadmap.

Verified commit: 49fb1da. Successful CI run. The attached package is the exact CI artifact; checksums cover the archive, report and both test logs.