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 A previous successful session used explicit per-tool test settings; it was superseded by this installed-configuration check.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.