Supercov

Coverage is described as running in an isolated workspace without rewriting project files, and the shown CI grants only contents: read. However, the tool executes a user-supplied test command, while quality scoring sends source content to the external Jev/TypeSafe AI service, so least privilege is not comprehensive. The README distinguishes local coverage, package-manager network access, and the paid quality service, but does not specify uploaded fields, retention, deletion, or key-protection practices. Cleanup has a dry-run and retention option, yet there is no built-in rollback or per-edit confirmation for tests written by an agent. Dependencies mix exact pins with broad ranges, workflow actions use version tags rather than commit digests, and no lockfile, vulnerability scan, or supply-chain policy is supplied. Repository, organization, issue tracker, and contributor copyright attribution are present, but publisher identity remains unknown and responsibility is not tied to named maintainers.

The README, ecosystem manifests, CI workflows, and example broadly agree on version 1.0.1, invocation patterns, and the core test loop, with extensive static test scripts. A deduction applies because the Python metadata says Pre-Alpha while the documentation presents broad multi-language and multi-platform availability, creating tension about maturity. Dependencies and environment requirements are detailed and platform packages share a version, but registries, an external API, test frameworks, and several language toolchains remain availability dependencies without documented degradation behavior. The loop is said to stop on failed tests, incomplete evidence, or ambiguous scope, but no implemented error-message code or representative failure output is included, so failure handling is adequate rather than thorough.

The material directly addresses interactive coding agents, leftover-token or overnight loops, and automated software factories, with distinct quality, coverage, and assertion-mapping scenarios. Boundaries are explicit: Supercov supplies signals and evidence but does not host, schedule, or replace the agent, and account/network requirements differ between quality and coverage. Copyable prompts precisely constrain edits to tests, define stop conditions, select bounded targets, and require reruns and diffs. Detailed matrices cover languages, runners, operating systems, architectures, minimum versions, installation channels, and attribution boundaries, justifying full scores here.

The README has strong information hierarchy, progressive command examples, documentation links, and support matrices. Installation covers npm, PyPI, RubyGems, Cargo, Go, Homebrew, and release archives with prerequisites. Naming, module paths, and version 1.0.1 are consistent across Cargo, npm, and native optional packages. Examples are substantial and the tutorial has a three-platform workflow, but no actual FAQ content is supplied and the linked troubleshooting material is not included. Limitations, attribution differences, completeness blockers, and coming-soon languages are explicit. The full MIT text agrees with all manifests. A version and bump script exist, but no changelog, release history, or migration guidance is provided. An issue tracker and Supercorp support statement create an update path, though the copyright names only contributors and no specific maintainer or response commitment is identified.

Outputs are concise, paginated, and optionally JSON, and can resolve files, lines, decisions, gaps, and run-to-run differences; the examples show artifacts an agent could consume directly. Claimed marginal value includes structural gap queries, MC/DC, per-test attribution, assertion maps, and immutable run records, but those broad benefits are largely self-described and the supplied worked example covers only one small session branch. Cost disclosure includes an approximate per-megabyte quality charge, content caching, and free local coverage, but the supplied material contains no benchmarks, representative runtime overhead, or storage-growth measurements. Marginal value and cost-benefit therefore receive solid but not full scores.

Quality findings are described as traceable to files and questions, while coverage can be queried by run, source location, and diff; workflows, test scripts, and before/after tutorial tests corroborate part of the narrative. Broad claims about all languages, platforms, exact attribution, and identical cross-registry binaries are not individually backed by included implementation code, artifact digests, or test results, and CI configuration alone does not prove passing execution. Fact and inference are separated particularly well: the documentation labels assertion scoring as agent-assessed, distinguishes aggregate from exact attribution, separates the external quality service from local coverage, and preserves incomplete evidence as a blocker rather than presenting it as a successful result.