GiviLoop 0.9.0 — Windowless reviews, eight MCP tools

GiviLoop 0.9.0

A smaller interface for a second code review, with web reviews that stay out of
your editor's way.

• Eight default MCP tools, including givi_review with your saved reviewer and
  givi_manage_review for recovery. --tools full retains advanced discovery.
• MCP web reviews and resume always use windowless ordinary Chrome. Visible
  overrides and givi_open are rejected before sending. Login, verification
  and ZIP uploads pause for manual CLI recovery; no visible fallback opens.
• Short CLI help for setup → review → status → report; givi help --all keeps
  the detailed reference. Explicit review IDs still isolate finding updates.
• Windows shutdown fallback targets only the Chrome process tree launched by
  GiviLoop. Release checks cover package installation, browser recovery and
  native desktop observation; deliberately visible tests run only in CI.
• A reproducible token benchmark reports measured agent usage. Avoiding a
  separate review API call does not establish lower total agent tokens or cost.

Upgrade

Install the attached checked package, then restart/reconnect your MCP client:

npm install -g ./giviloop-0.9.0.tgz

For a source checkout, update to v0.9.0, run npm ci and npm run build.
Existing review ledgers need no migration. Keep each review's runId when
recording findings. Old named review tools remain callable, but visible browser
controls are no longer available through MCP. Use the updated
tool reference and review skill.

Scope

Web integrations remain experimental and subject to provider quotas, login,
verification and terms. The checked browser fixtures do not guarantee access to
every live account or website revision. Local reviewers remain available.
See provider results and costs/access.

The HOL Guard contribution is optional and awaiting upstream integration. It
covers six direct CLI operations, not MCP; it is not a released protection layer.
See the precise scope.

The release attachments include the exact CI-tested installable package,
platform reports, test logs and SHA-256 checksums. Full results are linked from
the GitHub release.

Verified source: 7203e77. All 11 CI jobs passed. Attached reports record exact passing/skipped counts and package hashes. The installable tarball is the unmodified Linux CI artifact; the Windows-built archive was independently tested and differs only in line endings.

GiviLoop 0.8.2 — Explicit review identity for finding writes

GiviLoop 0.8.2 — Explicit review identity for finding writes

Community feedback asked whether a new automatic review could cause an agent to update findings in the wrong review. Reviews already had separate run IDs and ledgers, but omitting the run ID selected the latest run for both reads and writes. That default was ambiguous when finishing an earlier assessment after another review had started.

What changed

• givi findings add and update now require --run-id RUN_ID. The MCP tool givi_record_finding requires runId in its schema and at runtime. Missing or invalid IDs fail before a ledger is changed; there is no latest-run fallback for finding writes.
• A finding is addressed by (runId, findingId). IDs not present in the selected run are rejected. Titles/claims remain immutable and decisions append to the existing history.
• Automatic-review instructions and completion messages explicitly carry the returned run ID through reads, finding writes and reports. A new task with a changed snapshot creates its own review, leaving earlier findings intact. Same-task/snapshot deduplication is unchanged. Outstanding findings do not suspend new reviews; incomplete or failed delivery does.
• Read-only commands retain their latest-run convenience for manual use. Agents should pass the review ID throughout the whole assessment.

What was reproduced

With completed reviews A and B, and B marked latest, an omitted run ID allowed a finding intended for A to be added to B. An update carrying an ID found only in A was already rejected against B's ledger: this was not a demonstrated arbitrary cross-review overwrite. The patch removes implicit write selection entirely. Tests also deliberately give two ledgers the same local finding ID and verify that explicit run selection keeps their histories separate.

Upgrade

Existing ledgers need no migration. Scripts that omitted the run ID when adding or updating findings must now supply it. Preserve the original ID returned by the review; do not look up latest later. Restart the MCP server to load the new tool schema. To refresh the managed AGENTS.md instructions, run auto-review enable again (with --client codex when appropriate), after confirming the saved reviewer is the destination you intend to authorize. The runtime write guard applies even with older instructions.

Validation

The regression failed against 0.8.1 and passes with the patch. Tests cover CLI and actual MCP transport with two completed reviews, refusal of omitted/invalid IDs without changing either ledger, explicit updates to an older review after latest changes, simulated equal finding IDs across runs, and two automatic reviews while earlier findings remain unverified. Automatic transport tests use a loopback model fixture; no new live-provider or model-quality result is claimed.

The installed package passes 225 unit/integration tests and 53 real-browser fixture tests; npm audit reports zero production dependency vulnerabilities. The release artifact is taken from successful CI for the published commit. Existing web-provider, desktop and instruction-driven activation limits remain unchanged.

Verified commit: 5d2e785. Successful CI run: 225 unit tests, 53 browser tests, zero production dependency vulnerabilities. The tarball below is the exact CI artifact. finding-run-regression-before.tap intentionally records the regression failing against 0.8.1. finding-run-regression-after.tap records the 36 passing focused tests after the fix. Finding writes now require an explicit review ID; existing ledgers need no migration.

GiviLoop 0.8.1 — Verified automatic invocation in Codex

GiviLoop 0.8.1 — Verified automatic invocation in Codex

The automatic double check has now run from fresh Codex CLI sessions given an ordinary coding task, without mentioning GiviLoop or review. The test found a real setup gap: a connected MCP server could still be blocked by Codex's tool approval policy. givi auto-review enable --client codex now installs the project-local server configuration and approvals for the six workflow tools. Other tools remain on prompt; global settings, sandboxing and project trust are unchanged. Setup and recovery.

Live acceptance, 2026-09-23

• macOS arm64, Node 22.21.1, Chrome 153.0.8010.53, Codex CLI 0.154.0-alpha.6.2 supplied with the installed VS Code extension. This exercised the actual CLI, not an interactive VS Code chat session. Signed-in Claude web was the reviewer; its model identity was not independently verified.
• Public Git fixture: add an arithmetic mean function, preserve the existing sum, and test empty, singleton, mixed-sign and fractional inputs. The prompt did not request a second review. Project trust was supplied for this controlled fixture through the client's per-invocation configuration; production setup leaves trust to the user/client.
• The client loaded the product-installed .codex/config.toml and AGENTS.md, edited the files, ran the tests, called givi_auto_review once, read the response and exported the report. No MCP server/tool permission overrides were supplied by the test. A previous successful session used explicit per-tool test settings; it was superseded by this installed-configuration check.
• In one installed-configuration run, the host dismissed reviewer advice about validating invalid inputs because the task explicitly guaranteed finite numeric inputs, and saved that assessment. This is a small example of checking the review against the contract, not evidence of general model accuracy.
• The final observed run completed in 84.335 seconds for the entire coding task. One owned Chrome process, zero foreground transitions to it, zero visible owned browser-window observations, and a released browser profile on disconnect. An AppKit/CoreGraphics observer sampled approximately every 20 ms and recorded changes to foreground/window-owner IDs, without recording window titles or screen content. This is a sampled observation on this Mac, not proof of no sub-sample flashes or behavior on other desktops.

The first experiment was blocked by the host with MCP tool call requires approval, but approval policy is never, before submission. The scoped installer addresses that permission gap. One intermediate desktop observer failed to start while its executable was compiling; its host invocation succeeded, but it was excluded from window/focus evidence and replaced by the completed observed run above.

Browser decision

A real windowless Chromium target was tested. The first launch could fill and submit without an OS window, but reopening the same profile failed in the login/resume regression. A bounded retry of Chromium's specific startup error did not fix that case. The prototype was removed; this release retains the tested minimized-browser transport. No headless fallback, CAPTCHA bypass or hidden retry was introduced.

The observed Mac run stayed out of the way. Universal Chrome invisibility is not a supported guarantee. Login, cookie choices and human verification still pause with needs-attention for explicit recovery.

Automated validation

The installed release package passes 221 unit/integration tests and 53 real-browser fixture tests. npm audit reports zero production dependency vulnerabilities. New checks cover scoped Codex approvals, idempotent installation/removal, preservation of unrelated settings, refusal to overwrite unmanaged permissions or symlinks, and CLI flag validation. CI runs the unit suite on Node 20/22/24 for Ubuntu/macOS and the installed package/browser suite on Linux.

Fresh-host invocation is now verified for the client/configuration above. It remains instruction-driven: a host that ignores or overrides AGENTS.md, disables the tool or does not trust the project will not perform the review. Other frontends are not automatically certified by this CLI test. Web adapters remain experimental and provider terms/quotas still apply; successful delivery does not establish contractual permission. Total token savings remain unmeasured. The existing video is still labeled as a 0.5.0 recording.

Verified commit: ec7fa1a. Successful CI run: 221 unit tests, 53 browser tests, zero production dependency vulnerabilities. The tarball below is the exact CI artifact. automatic-host-validation.json records the fresh Codex CLI invocation and sampled macOS desktop observations. automatic-host-assessment.md contains a contract-based assessment from another fresh client session on the public example. Invocation remains instruction-driven; the sampled desktop result is not a universal invisibility guarantee.

GiviLoop 0.8.0 — Opt-in automatic double check at task completion

GiviLoop 0.8.0 — Automatic double check at task completion

An agent can now request a second opinion when it finishes a code task, without a separate review prompt. givi auto-review enable installs a project AGENTS.md rule and pins the locally configured reviewer. With MCP connected and instructions loaded, the host calls givi_auto_review, verifies concrete findings and exports evidence. Setup and scope.

Behavior

• Opt-in per checkout; setup alone sends nothing automatically. Disable removes the managed rule while preserving other project instructions and saved evidence.
• Selected changed source/test files only, with bounded context and best-effort redaction. Failed checks and selections without eligible code skip transmission. Checks are the host's attestation, not tests executed by GiviLoop.
• Persistent once-per-task and unchanged-snapshot guards shared by CLI/MCP. Review-driven fixes cannot trigger a loop when the task ID is preserved. Concurrent operations do not submit twice.
• Quiet browser delivery, no provider fallback. Login, verification, cancellation, failures and interrupted attempts suspend automatic sends; explicit recovery never silently resubmits uncertain requests.
• Source changes during or after the review are flagged. Successful delivery remains pending independent verification; there is no automatic clean-code verdict or blind application of suggestions.
• Ordinary run cleanup preserves records referenced by automatic-review history. The history has explicit size/count bounds.

Live validation, 2026-09-23

The locally installed tarball passed 214 unit/integration tests and 53 real-browser fixture tests, with zero production dependency vulnerabilities reported by npm audit. The release assets attach the corresponding CI package and test evidence after the release commit passes CI.

A public synthetic discount-allocation change was reviewed through the actual MCP tool and signed-in Claude browser session on macOS arm64. Two happy-path tests passed before submission. The live reviewer found that splitDiscount(100, 3) returned [33, 33, 33], losing one cent despite the explicit sum-preservation contract. An independent local reproduction confirmed the result; the exported report recorded one confirmed finding.

Delivery completed in 18.068 seconds. Calls with the same task and the same snapshot were deduplicated. The focus observer sampled approximately every 20 ms, recorded no foreground transition to the owned Chrome process, and the dedicated profile was released on MCP disconnect. This observation does not guarantee zero flashes on other desktops.

This was a seeded public example, not evidence of general review accuracy or adoption. A test harness explicitly invoked the host lifecycle call: an unattended coding-agent session choosing to invoke the new rule has not been validated. The underlying MCP/CLI and real-browser fixture tests cover successful delivery, attention/resume, deduplication, concurrent calls and failure handling. Local runtime integration tests use a loopback fixture; they do not establish new trained-model results.

Remaining limits

Automatic invocation depends on the host loading and following AGENTS.md; this is not an IDE hook or file watcher. MCP configuration and a new agent session are required. Other clients need their equivalent instruction mechanism. The host remains responsible for selecting task scope, reporting checks honestly and verifying the review.

Source extension/path filtering deliberately omits configuration, docs and generated files. Whitespace edits in code remain eligible; there is no semantic formatting detector. Current changes relative to HEAD are covered, not already committed branch changes. Additional callers/contracts can require an explicit focused review. Browser adapters remain experimental; provider terms, accounts, quotas and verification still apply. Headless ChatGPT is unsupported in practice. Total token savings remain unmeasured. DwarfStar trained-model validation is still pending. No new model-accuracy ranking, external pilot or automatic telemetry was introduced.

The README video remains the clearly labeled 0.5.0 recording. The new automatic flow is documented in text; no replacement video is claimed.

Verified commit: a2158f7. Successful CI run: 214 unit tests, 53 browser tests, zero production dependency vulnerabilities. The tarball below is the exact CI artifact. automatic-live-validation.json records the separate macOS live test. automatic-double-check-report.md contains its independently assessed finding on a public seeded example. Unattended host invocation remains unvalidated.

GiviLoop 0.7.0 — First-use demo and shareable Double Check reports

GiviLoop 0.7.0 — try it, verify it, share the result

23 September 2026. This release connects the first-use demo to a repeatable before-commit review and a portable report. Web integrations retain their documented experimental status and provider access/terms limits.

New workflow

• givi demo --offline: an explicitly authored public example, no account/browser/network delivery, a fixed independent reproduction and a Markdown report.
• givi demo: the same public example through your configured web/local reviewer. Only bundled source is sent. The actual response and deterministic reproduction stay distinct; the known control is not a model-quality score.
• givi demo --finish: complete a paused demo after login/resume without sending another prompt. It checks example integrity and never executes a modified verifier or model-generated code. Project sources, preferences and the selected project review remain unchanged.
• givi report and MCP givi_export_report: local Markdown export with effective verdicts, stale warnings, recorded evidence, referenced source hashes and request/response fingerprints. Empty/unassessed findings and unsuccessful runs remain explicit. Raw source, requests and complete answers are omitted. Inspect evidence before sharing; nothing is posted automatically.
• A before-commit recipe and voluntary feedback fields for actual second use. Reports survive automatic run retention; regenerate them after changes because exported files are snapshots.

Validation and scope

New unit/integration cases cover offline execution, unchanged project state, tampered verifiers, provider-option errors, idempotent finish, stale/empty/failed reports, redaction, Markdown rendering, symlink protection, retention and MCP export. Real-Chrome fixtures cover live-shaped demo delivery plus login pause → resume → finish with one submission.

A real signed-in Claude demo completed with its actual response saved, the known empty-array bug independently reproduced and all four bundled candidate-fix cases passing. This does not certify other model claims or measure total token savings. The source was unchanged. Release assets include the package/CI results and a sanitized live-validation summary.

A second real Claude review examined the new implementation. An unexpected persisted verdict was independently reproduced as a misleading-report bug, then fixed by validating ledger decision states. A claim about an unhandled CLI JSON stack trace was disproved by a targeted test: the existing error boundary already catches it. The exported report preserves both decisions and explicitly notes the frozen pre-fix source and missing original CLI context.

The current Git review scope is working-tree/staged changes relative to HEAD, not an entire branch. The calling agent still verifies real-project findings under its repository permissions. The only verifier executed by the new demo is the installed, fixed public example; the report exporter runs no tests.

No new providers, automatic code fixes, model-output execution, community posts or telemetry were added. The five-user pilot kit is ready; recruitment, external trials and actual adoption measurements remain pending. The README video remains the labeled 0.5.0 recording.

Verified commit: cb5efa4. Successful CI run: 202 unit tests, 51 browser tests, zero production dependency vulnerabilities. The tarball below is the exact CI artifact. adoption-live-validation.json records the separate macOS live demo and implementation review. double-check-report.md contains the independently assessed findings.

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.

GiviLoop 0.5.0 — Guided setup and review evidence

GiviLoop 0.5.0

Get to a first review with guided setup, then preserve why advice was accepted or rejected. This release adds onboarding and evidence recording to the existing CLI/MCP workflow.

• givi setup checks prerequisites, offers provider/login/access choices, generates a local MCP configuration snippet and can run an opt-in public example. It does not modify editor settings, install models or choose an implicit provider for later commands.
• givi findings add|update|list and givi_record_finding / givi_list_findings keep claims, confirmed/dismissed/unverified assessments, reasons, evidence, file hashes and decision history. Runs with findings survive automatic retention cleanup.
• Changed referenced files or review content make prior conclusions stale and effectively unverified. Hashes capture files when an assessment is recorded; unreferenced context/environment changes are not detected.
• givi recheck / givi_prepare_recheck create a new request for one finding with current source and parent links. Inspect it and explicitly send it; no automatic submission, patch execution or fixed verdict.
• A new video shows real setup, a live Claude review, independent example assertions, evidence recorded through MCP and a targeted recheck. The dismissed false claim is an author-supplied control, not a claim attributed to Claude. Only the temporary demo copy is corrected.

The host user/agent verifies claims. GiviLoop stores that assessment; it does not run or certify evidence, and an empty finding list does not prove clean code. Setup and evidence guide.

Existing web adapters remain experimental: website changes, authentication, verification and quotas still apply. Headless ChatGPT is unsupported in live checks; use background mode. This release does not establish provider authorization. Avoiding an extra model API call is distinct from unmeasured total token savings. The 0.4.0 provider limits and cost/access conditions still apply.

Install

Download the attached package, then:

npm install -g ./giviloop-0.5.0.tgz
givi setup --repo /path/to/project

Node.js 20+ is required. Restart existing MCP servers to expose the new tools. GitHub tarball/source distribution only; no npm registry publication.

The attached test report, TAP logs and checksums identify the exact CI-checked package. Validation includes the installed package, CLI/MCP evidence persistence, setup isolation and the browser fixture suite, plus Node 20/22/24 on Linux/macOS. The video uses a real website; automated browser fixtures use isolated synthetic pages.

Next: invite a small group to test real changes and report friction, independently verified findings, false positives and repeat use. Community outreach is separate; this release posts no messages to Reddit or other services.

Verified commit: c09979c. Successful CI run: 174 unit tests, 45 browser tests, zero production dependency vulnerabilities. The tarball below is the exact CI artifact. giviloop-double-check.mp4 is the new 74-second demo; demo-run.json records its provenance.

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.

GiviLoop 0.3.1 — browser cleanup and Double Check demo

GiviLoop 0.3.1

This patch fixes a completed browser review leaving the CLI process alive, and makes the first-use documentation shorter and reproducible.

Fixed

Chrome helpers can inherit the browser's stderr pipe and keep it open after Chrome exits. GiviLoop had already saved the answer and closed its browser, but Node could keep waiting for that pipe. Cleanup now closes GiviLoop's own pipe after browser shutdown. It does not terminate unrelated Chrome processes or helpers.

A focused regression reproduces the inherited-pipe case: the old implementation times out; the corrected caller exits successfully while the synthetic helper remains alive. Existing shutdown, cancellation, profile persistence and browser round-trip checks remain in the suite.

Documentation and examples

• Shorter README focused on Double Check and using existing chat access to avoid a separate API review charge where applicable.
• New recording of an actual browser review, saved response, MCP read and independent verification. The recording uses public example code and a dedicated terminal view, with no account UI or private conversations.
• node examples/double-check/verify.mjs reproduces the known bug and tests a candidate correction in memory; the source remains unchanged. The script does not execute reviewer output.
• Opt-in recording and live-check scripts, plus refreshed end-to-end results.

Install

Download giviloop-0.3.1.tgz from this release and run:

npm install -g ./giviloop-0.3.1.tgz
givi --version

Node.js 20+ is required. This is a GitHub distribution; there is no npm registry publication. The 0.3.0 tag and package remain unchanged.

The 0.3.0 integration limits still apply. Web access remains experimental, headless was blocked in the tested session, and technical success does not establish permission under provider terms. There is no measured total-token-savings claim. Costs and access.

Demo

Watch the 57-second demo. Real CLI review, MCP handoff and independent verification; no time compression or private account UI.

Executed checks

All seven CI jobs passed: Node 20/22/24 on Linux/macOS plus Linux Chrome/package checks. The installed archive passed 145 unit tests, 15 browser tests, the packaged example check and a dependency audit with zero known vulnerabilities. Linux and macOS generated the same package checksum.

Live authenticated CLI/MCP browser flows and four local-runtime flows completed. The recorded browser review took 20.5 seconds without another human verification request; its answer was read through MCP and its proposed fix checked on four cases. A separate Git-diff web review completed in 23.9 seconds and its correction was checked on seven cases. The validation report records incorrect model advice separately from successful transfer.

GiviLoop 0.3.0 — Double Check and existing chat access

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.