Five phases turn source code into reviewable specifications
Reversa's Discovery workflow sends a codebase through 5 phases: reconnaissance, excavation, interpretation, generation, and review. Specialist instructions map modules, trace business rules, describe data, and produce documents such as C4 diagrams, state machines, questions, gaps, and code-to-spec matrices. Another coding agent is the main consumer, although maintainers can read the Markdown. This fits a project where missing system knowledge is a bigger problem than missing source code.
The README groups agents into 10 teams covering existing systems, new projects, feature work, migrations, bugs, documentation, and code-quality tasks. That breadth becomes useful after an extraction because later work can read the same files. It also expands the review surface. A team adopting Discovery alone should install only what it intends to use and check each output against the repository and the running system.
Version 1.3.3 supports many agents but supplies no model
The package identifies itself as version 1.3.3 and requires Node.js 18.20.2 or newer. Its installer detects supported environments and writes the matching entry file, such as CLAUDE.md, AGENTS.md, or .cursorrules. The README lists 14 agent environments, including Claude Code, Codex, Cursor, Gemini CLI, Hermes, GitHub Copilot, and Aider. Reversa supplies instructions and state management. The selected coding agent supplies the model and execution environment.
Reversa says it neither requests nor stores model API keys. Analysis still runs inside an authenticated coding-agent session, so prompts and source follow that agent's provider and deployment policy. Teams with confidential code need to approve the agent first, then inspect Reversa's generated files and exclusions. The framework does not turn a hosted agent into a local service.
What happened when we ran it
Our sandbox installed commit 0b5b5f2 in 80 seconds. Npm added 61 packages, the installed tree occupied 16 MB, and npm audit reported 0 known vulnerabilities. The repository contained 479 files, about 4,855 lines of source, and used 2.8 MB after checkout. Those are modest figures for an installer carrying a large catalog of Markdown-based agent instructions.
There was no build script or target, so our build step was skipped. The measured checkout also exposed no test script or target, and the repository had no tests directory, so the test step was skipped rather than passed. Two CI workflow files were present, but no Dockerfile was included. We verified installation and dependency audit only. We did not prove that the 5-phase analysis produces correct specifications on a real legacy system.
Two output directories do not create a hard sandbox
Reversa directs analysis writes to 2 areas, .reversa/ for state and _reversa_sdd/ for output, with named exceptions for engine instruction files and installed skills. A SHA-256 manifest supports updates without overwriting files the user changed. Checkpoints in state.json let an interrupted analysis resume, and the guided workflow asks for CONTINUAR between agents. These controls make a long investigation easier to inspect and stop.
The README also says to commit the project, keep it on a remote host, and make a folder copy before starting. That warning matters more than the guarantee in its immutability section. AI agents receive the write boundary through instructions, and the project acknowledges that agents can make mistakes. Run Discovery in a disposable branch or worktree and inspect git status after each phase.
One open report found 11 integrity failures in review output
Open issue 30 describes a Reversa 1.2.58 run where the final reviewer marked work complete despite inconsistent artifacts. The reporter's checker found 11 problems, including orphaned question references, missing required files, conflicting severities, and mismatched counts. The issue says Discovery is predominantly static by design. Confidence marks help readers locate inferred claims and gaps, but they cannot replace deterministic cross-checks.
Issue 26 documents an installation updated from 1.2.43 to 1.2.56 that reported the new version while retaining 48 of 64 skills. An August 31 commit says the installer now always installs all agents, so the report does not prove version 1.3.3 still fails. It does justify comparing installed skill directories and manifests after an update instead of trusting the displayed version.
September activity offsets the absent GitHub release
GitHub listed 1,593 stars and 23 combined issues and pull requests on September 12, 2026. The last push was September 8, only 4 days earlier, and recent commits added Hermes support plus installer policy work. The latest-release API returned no published GitHub release, while the package file says 1.3.3 and the visible tag list stops at 1.2.57. Pin the npm version instead of inferring it from the Releases page.
Documentation is available in English, Portuguese, and Spanish, yet some operational details trail the code. Issue 33, opened September 9, says the writer could not find documentation for reversa-config.json after an update made the project read-only. The CLI help we inspected is Portuguese. English-only teams should check policy files before an agent starts.
Reversa is a draft map, not proof of system behavior
Version 1.3.3 gives a team one vocabulary for extraction, uncertainty, planning, and later changes. Its best use is a structured first pass over a system whose original authors or documents are gone. The 5-phase flow can turn vague unease into a review queue. Our run had no test target, and issue 30 records inconsistent reviewer output, so every code reference, count, gap, and business rule still needs an owner before the draft guides production work.