Second Look v0.6.0 — bring your existing AI tool

Second Look 0.6 lets you evaluate work from the AI tool you already use. Freeze a task with prepare, save an external response, and compare up to four candidates with assess—without a new API account.

• Portable snapshots bind original intent, source, acceptance checks and response IDs. Rebuild requests omit old writable source.
• Local Chromium checks compare behavior, regressions and screenshots across every candidate pair. Originals stay intact.
• Imported model labels and optional costs/tokens are owner-reported and unverified. Missing usage stays unknown; identical-source evidence reuse is marked.
• share exports portable HTML/JSON while omitting private labels, source, prompts and local paths.

The supported execution surface is trusted static HTML/CSS/JS. External generation and response saving are manual. There is no new arbitrary-repository/Python adapter or verified two-model performance claim. Existing Claude CLI live workflows remain available.

Walkthrough · Dated research · Hand-authored import demo

Install the wheel attached below, then run python -m playwright install chromium and secondlook doctor. No PyPI publication is implied. The source archive and sdist are also attached, with checksums and a release verification record.

Validation: 232 local tests passed; all four GitHub CI jobs passed (Linux Python 3.11/3.12/3.13, macOS Python 3.12). A fresh wheel install exercised doctor/init/prepare/assess/share outside the checkout. The public site and import report passed Chromium checks at 1440, 390 and 320 px.

Second Look 0.5.0 — first public beta

Give old AI work a new model without losing the original problem. Second Look preserves intent, selects worthwhile revisits under a budget, and compares patches, reframed revisions and clean-slate implementations against frozen checks.

Open the no-key demo · Quickstart · 한국어

This first public beta includes local installation diagnostics, acceptance capture without JSON for simple text requirements, portable sharing, and a persistent problem library. Default shared summaries exclude source, prompts, private paths and raw receipts; intent/screenshots are opt-in.

python3 -m venv .venv
source .venv/bin/activate
python -m pip install 'https://github.com/mandu5/secondlook/releases/download/v0.5.0/secondlook_revisit-0.5.0-py3-none-any.whl'
python -m playwright install chromium
secondlook doctor
secondlook demo --offline --output ./first-look

Open first-look/report.html. The synthetic reference demonstration uses real browser checks and zero model calls (3/8 → 8/8). See the actual model trials and evaluation correction for measured live evidence.

Execution currently supports trusted static HTML/CSS/JS through Claude Code CLI on macOS/Linux. Other providers, backends and arbitrary repository execution are not supported. Budget stops are best effort; reported costs are API-equivalent. This release makes no SOTA, causal model-uplift or general token-savings claim.

Validation: 202 local tests, 20 relevant tests after a tooltip-only correction, clean installed-wheel smoke, and 1440/390/320-pixel browser checks. Independent review found no Critical/Important issue; its single minor presentation issue was fixed. No paid model calls were added for this release. Public gates passed: Linux Python 3.11/3.12/3.13 and macOS Python 3.12 each passed 202 tests. Main and tag CI are green. Eleven deployed files and all four initial release assets matched local bytes over anonymous HTTPS. Public desktop/mobile report interactions and installation directly from the release URL were verified. See the attached release-verification.json for exact commit, hashes and CI URLs.

MIT license. Compare downloaded wheel/source archives against SHA256SUMS. The named source ZIP contains the full audited public tree; the Python sdist contains packaging inputs. Private pre-release history and raw personal-project evidence are excluded.