FigMirror

Least privilege is only indirectly suggested by the local Web UI, user-owned workspace, and explicit invocation; no permission manifest or sandbox constraints are shown, so support is thin. The normal flow requires the user to attach a figure, provide data, and issue a request, giving reasonably clear initial consent, but no per-iteration or file-write confirmation is documented. The README outlines the Orchestrator/Drawer/Reviewer path and local workspace, yet does not disclose model-provider transfers, network calls, telemetry, retention, or deletion. Sensitive-data, credential, and privacy handling are absent. Dependencies have minimum versions but no lockfile, hashes, audit evidence, or vulnerability-response policy. Installation, skill deployment, generated scripts/PDFs, and workspace writes are external effects, but their exact scope and side effects are not fully documented; rollback and uninstall instructions are absent. Paper authors and citation metadata are supplied, while rights and item-level provenance for reference figures, gallery content, and generated assets remain unclear. No red-line behavior is evidenced.

The README presents a consistent role-separated loop, shared state machine, and bounded iterations, but the supplied test configuration says the project ships with “dependencies=[],” directly conflicting with the three runtime dependencies in pyproject.toml; self-consistency is therefore only partial. Python requirements and minimum versions for Pillow, matplotlib, NumPy, and pytest are explicit, with a uv setup path adequate for ordinary dependency preparation, although versions are not locked and Claude/Codex, system-tool, and backend availability requirements are incomplete. The README merely points to separate troubleshooting material; the supplied evidence contains no concrete failure messages, diagnostic codes, or recovery procedures.

The material clearly distinguishes browser UI, skill-only use, Codex and Claude backends, and scenarios with or without a ready reference, including 2D and 3D scientific figures; audience and scenario coverage is thorough. Boundaries include editable matplotlib output, PDF export, visual iteration, and a hard iteration cap, while the roadmap identifies unreleased evaluation artifacts and incomplete backend hardening; unsupported plot types, data sizes, input formats, and quality limits are not systematically stated. Triggering has concrete prompt examples and explicit image/data prerequisites, but no ambiguity handling, false-trigger controls, or refusal conditions. Environment guidance covers Python 3.10+, uv, disk-cache placement, two backends, and a local port, but lacks an OS matrix, hardware requirements, offline behavior, and compatible backend-version ranges.

The README has clear navigation and well-separated quick start, method, results, roadmap, contribution, and citation sections, earning strong information-architecture credit. Installation covers the Web UI, skill-only setup, backend selection, port, and low-disk guidance, but uses an unpinned curl-to-shell command and the supplied material lacks uninstall steps, complete prerequisites, and installation verification. The FigMirror brand is mostly stable, while figcopy_serve, figcopy_runner, and FigCopy gallery terminology introduce legacy naming inconsistency. Commands, prompt examples, and visual cases are useful, though there is no full FAQ and linked detailed documents were not supplied. Limitations appear mainly as roadmap gaps rather than a systematic limitations section. No license is provided or identified. Dated milestones, version 0.1.0, and a roadmap offer some release history, but there is no formal changelog, tag policy, or compatibility commitment. Contribution paths and cited authors are visible; however, publisher identity is unverified and no named maintainer, support period, security contact, or update responsibility is established.

The intended deliverables—an editable matplotlib script, camera-ready PDF, previews, and iteration history—are concrete and directly usable, so output usability is strongly supported. Reference-conditioned Drawer/Reviewer iteration offers clear marginal value over repeated manual plotting edits and examples span several chart families, although effectiveness evidence is predominantly self-published. Cost-benefit support is thin: a hard iteration cap and local workflow are mentioned, but token usage, latency, compute needs, model fees, and human-review effort are not quantified.

Performance figures are tied to a cited paper and displayed result table, but reproducible PlotTwin-Bench artifacts are explicitly unreleased and the supplied source is insufficient to trace the algorithm and numerical results through implementation, so claim traceability remains weak. README, pyproject, and test configuration provide limited corroboration for naming, Python dependencies, and runtime layout, while also containing a dependency-comment contradiction; no independent source or substantive test evidence is supplied. The text generally distinguishes paper-reported results, product descriptions, completed milestones, and future roadmap items, supporting fact/inference separation, but promotional effectiveness statements are not consistently marked as unverified or self-evaluated.