LoopCheck 0.5.2 — Readiness and real planning

LoopCheck 0.5.2

The first screen now separates live Strands planning from the disclosed keyless walkthrough. The live path connects an unchanged public MDN shopping-list sample and accepts a natural-language goal; it does not supply preset browser checks. Once requirements exist, checking the current change is the primary action. Contextual guidance explains whether to approve, generate missing checks, inspect evidence or recheck.

A read-only readiness check reports source access, preview reachability, model configuration and remaining recorded budget. It makes no provider request and does not approve a change. Public preview links resolve against the deployed host instead of a judge's own localhost.

The actual browser exercise exposed blocked inline JavaScript in the newly mounted MDN samples. Only the exact inline assets of the two pinned samples receive CSP hashes; the main application policy is unchanged. A real browser regression test verifies addition, cleared input and deletion through the production middleware.

Planner improvements simplify the initial observation and keep full evidence in the run while avoiding duplicated provider context. Tool validation failures are retained for diagnosis. Provider errors use bounded categories without response bodies or credentials. Planning drafts are not mislabeled as historical checks; they still cannot approve a change.

Verification and all unsuccessful live attempts are recorded in the 0.5.2 report. Release readiness, AWS access, public video and Devpost submission are independent gates. Read the exact commit's Actions results before deployment.

Upgrade: back up your existing data directory, update the source, retain the private .env and restart. No database migration beyond 0.5.0 is introduced. Existing requirements, evidence and recorded budget remain available. No human-efficiency percentage or prize outcome is claimed.

LoopCheck 0.5.1 — Keyless guided walkthrough

LoopCheck 0.5.1

Try the complete requirements-to-regression workflow without providing a model key. Each session gets an isolated cart copy and reviews preset requirements before real browser checks run. Add an unimplemented coupon requirement, see acceptance remain incomplete, inject a sample price regression, inspect its evidence and recheck the repair.

Source edits and requirements in this walkthrough are scripted presets, explicitly disclosed in the interface. It demonstrates real verification, not autonomous coding or agent planning. Real-project AI planning still uses the existing server-side Strands/Qwen configuration.

• Inline full reproduction evidence provides an alternative to clipboard transfer.
• Current requirement coverage remains visible when an older result is opened.
• Session ownership, source/contract guards, restart persistence and rendered UI behavior have focused regression tests.

Local verification passed 37 tests and 12 frozen cases. The focused tests passed again after the history-display correction. This release is published from its successful Linux verification commit; see that commit's Actions run for complete test and container artifacts. Walkthrough, retained failures and limits.

Back up the data directory and follow README.en.md. No additional database migration is introduced beyond 0.5.0. AWS hosting, a publicly hosted submission video and final contest submission remain independent delivery steps. No human-efficiency percentage is claimed.

LoopCheck 0.5.0 — Versioned requirements

LoopCheck 0.5.0

Save business requirements before their controls exist, then keep those requirements usable across daily code changes. Uncovered requirements now prevent acceptance even when all existing browser checks pass.

• Review edits, retirement, restoration and rebuilt checks with immutable requirement versions.
• Distinguish current results from historical evidence; return full reproduction steps through the existing three MCP tools.
• Explore simple dialogs with bounded, fresh-browser action replay. Planner retries can correct locators while preserving assertions and expected values.
• Preserve legacy projects and evidence with a SQLite backup before migration.

Verification

Local Windows verification passed 35 tests and all 12 frozen mutation cases. The latest live Strands/Qwen round passed four goals on two pinned MDN examples; all five rounds (0/4, 1/4, 2/4, 0/4, 4/4) remain available. This is iterative debugging on a small sample set, not a generalization or human-efficiency claim.

The 63.64-second English-captioned clip records an automated cart regression gate. Scripted feature changes, fault injection and editing are disclosed; it is not a continuous human or autonomous coding-agent demonstration.

Install and upgrade

Use README.en.md and docs/UPGRADE-0.5.md. Back up the data directory before upgrading. Local folder access and MCP stay in the native local edition; public mode accepts only bundled samples. Never put model keys in the browser, repository or reports.

AWS hosted access, a YouTube/Vimeo submission video and Devpost submission are separate delivery steps. Check the version-specific CI run before relying on Linux/container support.

LoopCheck 0.4.0 — Real-project acceptance companion

LoopCheck 0.4 connects running local web projects, preserves human-approved browser requirements, and returns regression evidence to the existing coding AI over MCP.

Verified

• Linux CI: 25 tests passed; 12 frozen mutation cases met their expected outcomes (6 deliberately injected faults and 6 non-fault changes).
• Fresh Docker build and public-mode browser gate passed under a 768 MiB limit; peak 324.27 MiB, no OOM.
• Original source: baseline passed, an injected arithmetic regression failed, and restoration passed the unchanged checks.
• CI used no provider calls. Separate live Strands/Qwen and MCP records are included in the source.

Tested executable commit: 9af20f8. The release commit adds documentation and evidence only.
CI: https://github.com/WeiR-h/loopcheck/actions/runs/34462275115

Download

• Source ZIP: includes MIT license, English README, setup scripts, tests, architecture and disclosed evidence. See README.en.md for setup and model configuration.
• MP4 + SRT: 127-second English-captioned evidence walkthrough; edited real browser clips and UI stills, not a continuous desktop recording.
• SHA256SUMS.txt: checksums for all three assets.

Limits and remaining delivery

No human-efficiency percentage or unknown-project reliability rate is claimed. AWS hosted judge access, YouTube/Vimeo publication and Devpost submission are pending. CI memory figures are not an AWS instance capacity or live-model load test. Native startup supports local projects; public Docker mode exposes bundled samples only. The local MCP bridge never approves requirements or edits source. AgentCore is not deployed.