Skip to content

GiviLoop 0.3.0 — Double Check and existing chat access

Choose a tag to compare

View all tags
@vgflutter vgflutter released this 21 Sep 21:36
v0.3.0
b0c6bb8
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

GiviLoop 0.3.0

Double-check your code. Put your existing chat access to work.

This release packages the existing second-review loop for community use: select code or Git changes, obtain a review, preserve its request and response, and return the findings to your coding agent for verification. It does not add a dedicated Double Check command or an autonomous verification engine.

Included

  • CLI and ten MCP tools for preparing context, asking a reviewer and reading a saved response. Double Check recipes start with analysis before edits.
  • Optional ChatGPT background integration using native Chrome and the dedicated authenticated profile. This resolved the observed visible/background 403 and completed real CLI/MCP reviews. Human verification is handed back to the account holder and repeated challenges stop.
  • Automatic local adapters for Ollama, DwarfStar, llama.cpp, LM Studio and MLX-LM, with explicit models, bounded output and loopback-only connections.
  • Request/response association, run locking and retention, cancellation cleanup, incomplete-output rejection and explicit errors for uncertain submission. An uncertain send is not retried automatically.
  • Manual web transfer, focused context budgets, best-effort redaction, troubleshooting, a reproducible example and documentation of costs and access conditions.

Release scope and known limits

The distribution contains tested CLI/MCP and local review workflows. The optional ChatGPT browser adapter remains experimental. A website can refuse access or change its interface. Headless remains blocked in the tested session; exact subscription-model selection, long web reasoning and live ZIP upload are not validated. There is no promise of unattended website availability or one-time-only verification.

Native DwarfStar server/GPU compatibility was tested with a synthetic model; trained-model review still requires suitable hardware. MLX-LM's upstream server is experimental. Real trained-weight reviews completed for Ollama, llama.cpp, LM Studio and MLX, including incorrect advice documented in the validation reports.

A web review makes no separately billed model API call through GiviLoop, which can avoid that extra API expense. Chat plans and quotas still apply. Total token savings are not measured. OpenAI's European terms prohibit automatic output extraction; a technical success, existing subscription or MIT disclaimer does not establish permission for the automatic browser path. Costs and access.

Validation

The installed package is checked with the complete unit and isolated Chrome suites and a production dependency audit. GitHub Actions covers Node 20/22/24 on Linux/macOS, plus the Chrome/package job. Release artifacts include the package verification report and test logs; consult those and the commit's CI result for the executed checks.

Real authenticated browser reviews completed through CLI and MCP in 20.4 and 26.3 seconds respectively. Eleven regression cases independently checked the two proposed fixes. These small known-bug fixtures demonstrate the transfer/verification workflow, not a benchmark of model quality or token savings. Browser evidence · Local inference evidence.

Install

Download giviloop-0.3.0.tgz from this GitHub release, then:

npm install -g ./giviloop-0.3.0.tgz
givi --version
givi help

Node.js 20+ is required. givi-mcp starts the MCP stdio server. Google Chrome is required only for browser automation; local inference uses a separately installed runtime and model. There is no npm registry publication for this release.

Start with the Double Check recipe or the README quickstart. Add .giviloop/ to the reviewed repository's ignore rules; run artifacts can contain source and prompts.

Next

Collect real usage and false positives before expanding scope. Structured finding evidence and rechecks are proposed follow-ups, not features promised by this release. Roadmap.

Executed release checks

All seven CI jobs passed: Node 20/22/24 on Linux/macOS and the Linux Chrome/package job. The installed archive passed 144 unit tests, 15 Chrome tests and an audit with zero known vulnerabilities. Linux CI and local macOS produced the same package SHA-256.

The final README web quickstart completed in 21.5 seconds without another human verification request. Its saved response was read through MCP; the empty-array defect was reproduced and the proposed fix checked independently on four cases. The example source stayed unchanged. Web token usage remains unknown.

Assets 7
Loading