GiviLoop 0.6.0 — Quiet reviews and saved preferences

GiviLoop 0.6.0 — keep the review out of your way

23 September 2026. This release improves day-to-day use of Double Check: saved preferences, one-command reviews and browser attention at a time you choose. Web adapters remain experimental and subject to provider access, quotas and terms.

Changes

• setup saves provider/model/profile and background preferences per project. review prepares and sends selected files or current Git changes using them. ask, prepare and archive still do not send without explicit --send.
• Auto web delivery defaults to a minimized browser. Login, cookie setup, challenges and uploads pause as needs-attention before submission, instead of deliberately showing Chrome. Explicit foreground mode remains available.
• status, open, resume and cancel, also exposed through MCP. Resume requires an unchanged request proven unsent; it refuses uncertain, submitted or completed requests. Cancellation cannot retract a prompt.
• One MCP process reuses healthy quiet Chrome sessions, with a fresh chat for each review: at most two profiles, exclusive access and 60-second idle expiry. Release idle sessions explicitly before manual login. Failed/cancelled sessions close; stdin EOF/disconnect closes retained browsers. CLI runs still close their browser after each review.

Compatibility

Automatic web delivery now defaults to auto, not prefill. Specify --mode prefill if you only want the composer filled. Background is the default for auto; --foreground explicitly allows visible interactions and ZIP uploads. setup --background restores quiet mode after a saved foreground preference. Restart the MCP server after updating. See commands and controls.

Validation

The independent reviewer added regressions for quiet attention, preference isolation, safe resume, archive integrity, cooperative cancellation and browser-session lifecycle. Real-Chrome fixtures additionally cover two reviews through one browser, separate prompts, release of idle sessions, cancellation after submission, stdin EOF cleanup and relative repository paths.

A live signed-in Claude test on macOS arm64 completed two public synthetic reviews in 19.3 and 15.8 seconds. One Chrome process served both; a foreground-application observer sampled about every 20 ms and recorded zero foreground transitions to that process. The second answer concerned only its new request. The profile was released after MCP disconnect. This is one desktop observation, not a guarantee for fullscreen/Spaces or other window managers.

The first example reproduced the known empty-array reduce bug; the second was a clean identity function whose finite-number contract was accepted. These are workflow controls, not a comparative model benchmark. An initial test-harness result-schema error occurred after a separately successful saved answer; it was corrected before the two-review measurement. No uncertain request was resumed automatically.

Package and CI test totals are recorded in the release's release-checks.json and TAP assets. The installed tarball is tested in a clean consumer, including real-browser fixtures and dependency audit.

Limits

Quiet means GiviLoop does not deliberately show Chrome for attention; input/window-manager behavior can still cause a transient flash. Headless ChatGPT remains unsupported in practice. Verification can recur; GiviLoop does not bypass it. Browser success does not establish provider authorization. Local inference and manual transfer remain alternatives.

MCP session reuse is limited to one server process, not a daemon shared with CLI commands. Persistent browsers retain resources for up to one idle minute; release them sooner when needed. Local-runtime integrations keep their previously documented validation limits, including pending DwarfStar trained-model validation. This release does not claim new live coverage of every provider/model.

The README demo remains the real 0.5.0 recording, labeled accordingly. Avoiding a separate token-billed API review is still distinct from unmeasured total token savings.

Verified commit: 3dfad0b. Successful CI run: 194 unit tests, 49 browser tests, zero production dependency vulnerabilities. The tarball below is the exact CI artifact. quiet-live-validation.json records the separate macOS live observation.