era-agent

# ERA ## Evolve & Reverse Engineer Anything **Reverse. Analyze. Evolve. Repeat.** ERA is an open-source reverse-engineering agent built on the original **[Reverse Engineer Anything (REA)](https://github.com/morluto/rea)** engine. REA does the reverse engineering; ERA adds the workflow around it: 1. **Analysis Mode**: one command (`era analyze`) routes a target to the right REA workflow, records every REA execution, normalizes REA Evidence into facts and writes a report. Untrusted executables run in a sandbox. 2. **ERA Studio**: a desktop and browser workbench for all of it, with a Monaco editor, diff viewer, terminal and an in-app [Documentation](#documentation) section. 3. **ERA Project Lab**: Review, Modify and Rebuild for projects you are authorized to work on. A coding agent works in an isolated copy; you approve a verified diff, and can roll it back. 4. **Multi-Agent Engineering**: architect, implementer and reviewer roles across Codex, Claude Code and your own command agents, with consent, cost and exposure controls. 5. **ERA Improvement Lab**: a controlled system in which REA reverse engineers ERA itself; a maintainer approves what is integrated. 6. **Measurement**: hand-written ground-truth benchmarks, version comparison and ERA-versus-direct-REA overhead. Every reverse-engineering fact ERA reports comes from a recorded REA execution. ERA never re-implements REA, and everything REA offers stays available through `era rea …`. [ERA and REA](#era-and-rea) separates what REA provides from what ERA adds. > **Status: v0.2.0.** The previous release is v0.1.0 (the CLI, still available). The desktop app is an > **unsigned development build** for macOS arm64 that you build from source; there is no signed or notarized download. Verified on macOS 27 (arm64), Node 24 and REA 6.4.0: see > [what is verified](docs/SUPPORT_AND_VALIDATION.md), the [release notes](docs/release-notes-v0.2.0.md), > [limitations](#limitations) and the [security summary](docs/ERA_V020_REMEDIATION.md). An independent > audit found release-blocking issues that are fixed; those fixes have not been independently > re-reviewed. ERA has no security certification. ## ERA Studio ERA Studio is the local desktop and browser interface to ERA's backend. Everything it shows comes from the same control plane the command line uses; the screenshots below are real captures (see [how they are made](docs/assets/screenshots/README.md): the coding agent in them is a scripted stand-in, everything else is real).  * **Analyze** software with the original REA engine and read the evidence, facts and structure. * **Review, Modify, Rebuild** projects in an isolated copy, with tests, regression detection and a verified diff. * **Edit and inspect** with Monaco, a side-by-side diff viewer, Git status and a sandboxed terminal. * **Configure agents**: Codex, Claude Code and custom command agents, with per-project consent and cost caps. * **Stay in control**: nothing reaches your project without your approval; approved changes can be rolled back. * **Improve ERA itself** through the Improvement Lab, and **read the docs offline** inside the app. | | | |---|---| |  |  | | **Analysis**: REA evidence and structure | **Diff and approval**: verified change, tests, your decision | |  |  | | **Multi-agent**: providers, consent, cost | **Improvement Lab**: ERA analyzing itself | All ten views, including project management, Project Lab, the editor, settings and the documentation: [screenshot gallery](docs/assets/screenshots/README.md). ```bash git clone --branch v0.2.0 https://github.com/dav1dngy/era-agent.git && cd era-agent npm ci && npm ci --prefix apps/studio node bin/era.mjs studio doctor # what Studio needs on this machine node bin/era.mjs studio app # desktop shell; `studio serve --open` runs it in a browser ``` macOS arm64 is the validated platform. Windows and Linux Studio, a debugger, an extension marketplace and SSO/RBAC are not provided. Start with [Getting started](docs/getting-started.md), the [Studio guide](docs/ERA_STUDIO.md), [installation](docs/installation.md) and [troubleshooting](docs/troubleshooting.md); the same pages are in the app under **Documentation**. --- ## What works Verified on macOS 27 arm64, Node 24.20, REA 6.4.0, Hopper 6.6.0-demo: | Capability | Command | REA engine call | Verified | |---|---|---|---| | JavaScript application analysis (modules, imports, exports, entrypoints, URL literals) | `era analyze <dir>` | `analyze-javascript-application` | ✅ | | Single JavaScript file | `era analyze <file.js>` | same, on a staged directory | ✅ | | Mach-O structure | `era analyze <binary>` | `inspect-macho` | ✅ | | Native decompilation of exported functions | `era analyze <binary>` | `function … --snapshot` | ✅ with Hopper (Ghidra/IDA supported by REA, not tested) | | Sandboxed runtime capture (no network, writes only to scratch) | `era analyze <binary> --runtime` | `capture-process` under `sandbox-exec` | ✅ macOS | | Archive inventory (ZIP) | `era analyze <file.zip>` | `inspect-artifact` | ✅ (APK/IPA/DMG routed, not tested) | | Explanations: local deterministic, or an opt-in model | `era explain <analysis>` | none (reads ERA facts) | ✅ local; remote/command providers tested with fakes | | Every original REA command, MCP server, agent setup | `era rea …`, `era rea mcp`, `era rea setup` | passthrough | ✅ (real MCP session tested) | | Improvement Lab: self-analysis, experiments, approval, rebase, rollback, cancel, recover | `era self-analyze`, `era improve …`, `era rollback` | REA self-analysis | ✅ | | Coding agents | `era improve implement --agent …` | — | `manual` ✅ real; `claude-code`, `codex`, `command` ✅ adapters tested with fake CLIs, not run against real models | | **Project Lab**: review, modify or rebuild an authorized *external* project in an isolated copy, with human approval and rollback | `era project …` | REA for JavaScript projects and for Rebuild references | ✅ with scripted *mock* agents + the projects' own tests; **real Codex** validated (small tasks); **Claude Code not validated** (signed out on the dev machine) | | Benchmarks, version comparison, REA overhead | `era benchmark [--versions …] [--rea-baseline]` | via `era analyze` / direct REA | ✅ | | Status, history, JSONL logs, static dashboard | `era status`, `era history`, `era dashboard` | `providers` | ✅ | **Not supported through `era analyze`** (available in REA via `era rea …`): ELF/PE analysis, Android, firmware, browser and Electron runtime, .NET. On Linux, Mach-O inspection, Hopper and the runtime sandbox are unavailable. ## ERA and REA | | REA (upstream, unmodified) | ERA (this repository) | |---|---|---| | Reverse-engineering operations (127 CLI commands, 170 MCP tools) | ✅ provides | uses; never re-implements | | MCP server and agent registration (Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI, VS Code, …) | ✅ `rea mcp`, `rea setup` | reuses through `era rea …` | | Evidence, unknowns, evidence bundles | ✅ | stores REA Evidence unmodified next to run records | | Target routing, normalized facts, reports | — | ✅ | | Sandboxed runtime capture | REA's capture is explicitly not a sandbox | ✅ macOS `sandbox-exec` wrapper | | Ground-truth benchmarks and version comparison | — | ✅ | | Self-analysis and the Improvement Lab | — | ✅ | | Coding-agent and reasoning-provider adapters | — | ✅ | | ERA Studio (desktop and browser workbench) | — | ✅ | | Project Lab: Review, Modify, Rebuild with approval and rollback | — | ✅ | | Multi-Agent Engineering (architect, implementer, reviewer) | — | ✅ | All of `src/` is ERA-specific. REA is the `rea-agents@6.4.0` dependency (MIT, © morluto). See [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md), [docs/original-rea-analysis.md](docs/original-rea-analysis.md) and [docs/feature-gap-analysis.md](docs/feature-gap-analysis.md). ## Install There are three different things you can install; pick the one you need. | You want | Source | Includes ERA Studio | |---|---|---| | **ERA with Studio** (v0.2.0): CLI plus the desktop shell and the browser workbench | the default branch `main`, or the tag `v0.2.0` | yes | | **The CLI only** | the `era-agent-0.2.0.tgz` attached to the release (v0.2.0), or the tag `v0.1.0` for the previous CLI-only release | no | | A **desktop application** | built locally with `npm run package --prefix apps/studio` (unsigned, macOS arm64). No signed or notarized download is published | yes | The CLI tarball and the tag `v0.1.0` do **not** contain `apps/studio`; installing only the CLI and then running a Studio command fails. ```bash # CLI only (the previous CLI-only release, v0.1.0) git clone --branch v0.1.0 https://github.com/dav1dngy/era-agent.git && cd era-agent npm install && npm test && node bin/era.mjs status # v0.2.0 with ERA Studio git clone --branch v0.2.0 https://github.com/dav1dngy/era-agent.git && cd era-agent npm ci && npm ci --prefix apps/studio node bin/era.mjs studio doctor # checks what Studio needs on this machine node bin/era.mjs studio app # desktop shell; or: node bin/era.mjs studio serve --open (browser) ``` Put `era` on your PATH with `npm link` (optional). Studio has an in-app **Documentation** section (sidebar, ⌘K, and F1 in the desktop app) that works offline. Requirements: Node.js `^22.19 || ^24.11 || >=26` (REA's supported range; verified on 24.20 only), git; Python 3 for the Studio terminal. Native decompilation needs Hopper, Ghidra 12.1.x (a JDK 21 on `PATH` and `GHIDRA_INSTALL_DIR`; the desktop app imports only your shell's `PATH`, so start it from that shell) or IDA Pro MCP. ERA needs no API keys. Details, state locations and troubleshooting: [docs/installation.md](docs/installation.md). What has been verified, on which platform and revision: [docs/SUPPORT_AND_VALIDATION.md](docs/SUPPORT_AND_VALIDATION.md). ## Analyze software ```bash era analyze benchmarks/datasets/js-app era analyze benchmarks/datasets/native-hello/hello era analyze benchmarks/datasets/native-hello/hello --runtime --arg demo # sandboxed execution era analyze benchmarks/datasets/archive/js-app.zip --json era explain <analysis-id> # local explanation; no data leaves the machine era inspect path/to/target # classification + engine status era rea doctor # any original REA command ``` Excerpt of a real report ([full report](examples/sample-analysis-report.md)), decompiled by Hopper through REA from `clang -O1` output: ```c ### _greet @ 0x100000460 r8 = 0x1505; // 5381: djb2 hash seed do { r8 = r8 + r8 * 0x20 + r9; // h = h * 33 + c ... printf("Hello, %s! checksum=%u\n", r19, r8); ``` Static analysis only reads files. `--runtime` executes the target inside a macOS sandbox (no network, writes only to the analysis scratch directory). Without a sandbox ERA refuses unless you pass `--unsandboxed`. ## ERA Improvement Lab ```bash era self-analyze # REA analysis of ERA's packed artifact + CLI runtime era improve propose [--finding ID] # REA self-analysis → one hypothesis + isolated worktree era improve implement <cycle> --agent claude-code|codex|command # or edit the worktree yourself era improve evaluate <cycle> # tests + benchmarks + REA self-analysis → rejected | awaiting-approval era improve show <cycle> # reviewable change proposal (report + patch) era improve approve <cycle> # maintainer: merge into main and tag era-vN era improve decline <cycle> | rebase <cycle> | cancel <cycle> | recover era rollback era-v2 # restore any version as a new commit era agents # coding agents, reasoning providers, REA ``` **Trust model (since the independent audit).** The stable supervisor runs every REA execution used to judge a candidate. Candidate code (its tests, its fact extraction, its self-analysis plan, its CLI) runs only inside a macOS sandbox: no network, writes only to its own checkout and private temp, no reads of ground truth, ERA's state or credential stores, and no secret environment variables. Approval merges exactly the evaluated commit. Without a sandbox, evaluation refuses unless `--trust-candidate`. See [docs/audit-remediation.md](docs/audit-remediation.md). Controls: iteration, time, REA-run and API-cost limits with hard caps. One experiment at a time. Cancellation kills the whole process tree. A watchdog ensures no subprocess outlives its deadline or its parent. A protected evaluator (experiments cannot edit the tests, rules, datasets or acceptance criteria that judge them). No merge without maintainer approval. Refused in CI. Full description: [docs/improvement-lab.md](docs/improvement-lab.md). ### Experiments so far | # | Experiment | Evidence | Decision | |---|---|---|---| | 1 | Infer CommonJS export names from REA's semantic graph | F1 0.956 → 0.987, but REA self-analysis findings 7 → 9 | **rejected** (main untouched) | | 2 | Split REA self-analysis into size-balanced passes | coverage finding resolved, findings 7 → 6 | accepted → era-v2 | | 3 | Retry #1 on era-v2 | F1 0.956 → 0.987, findings 6 → 5 | accepted → era-v3 | | 4 | Byte-bounded REA passes with import placeholders (campaign 2) | findings 26 → 7, complete REA coverage | first evaluation passed; **rejected** on re-evaluation (a Hopper-flaky test); retried as #5 | Cycles 2 and 3 ran before the approval gate existed and were merged automatically; cycle 1 was rejected. Cycles 1–5 were evaluated by the pre-audit evaluator: candidate code ran unsandboxed and self-analysis evidence came from the candidate's own pipeline. The independent audit re-checked the recorded raw outputs (49/49 digests match; cycle 5's findings reproduce), but those evaluations predate the supervisor/sandbox model below. Cycle 5's "104 → 107 tests" means the baseline had 1 failing test (REA coverage) that the candidate fixed. | 5 | Retry of #4 after the flaky test fix | findings 27 → 9, tests 104 → 107 (0 failed), accuracy unchanged | accepted, **approved by the maintainer → era-v4** | Every record (cycle JSON, REA run records with output digests, self-analysis reports, ledger) is in [`experiments/`](experiments/). The narrative is in [docs/self-improvement.md](docs/self-improvement.md). ## Benchmarks Measured on macOS 27 arm64, REA 6.4.0, Hopper 6.6.0-demo, 3 repetitions per case, quiet host (`benchmarks/results/versions-v1-v2-v3.json`, `benchmarks/results/main-with-rea-baseline.json`): | Metric | era-v1 | era-v2 | era-v3 | main (RC) | |---|---|---|---|---| | Accuracy (F1) | 0.956 | 0.956 | 0.987 | 0.987 | | Precision | 0.988 | 0.988 | 0.988 | 0.988 | | Completeness (recall) | 0.933 | 0.933 | 0.987 | 0.987 | | Execution time (ms, Σ medians) | 37,351 | 44,534 | 47,968 | 51,458 | | CPU time (s, Σ medians) | 33.18 | 35.26 | 34.19 | 34.06 | | Peak memory (MB RSS) | 389.6 | 389.6 | 389.8 | 389.5 | | Error rate | 0.000 | 0.000 | 0.000 | 0.000 | | Reliability | 1.000 | 1.000 | 1.000 | 1.000 | | REA executions per pass | 10 | 10 | 10 | 10 | Wall time is dominated by Hopper start-up and varies by up to ~25% between identical runs. The rise from era-v1 to main is within that noise for the native cases, but CPU time is flat. Accuracy is deterministic. **ERA vs direct REA** (same REA commands, median wall ms; `--rea-baseline`): | Case | REA direct | ERA analyze | Overhead | |---|---|---|---| | js-commonjs-app | 1,786 | 3,555 | +1,769 ms | | js-esm-app | 1,757 | 3,466 | +1,709 ms | | native-macho | 20,928 | 19,727 | −1,201 ms (Hopper noise) | | native-macho-runtime | 25,014 | 22,339 | −2,675 ms (Hopper noise) | | zip-archive | 713 | 2,371 | +1,658 ms | ERA added a fixed ~1.7 s per analysis in this run, mostly the `rea providers` probe. Since the audit remediation the probe runs only for Mach-O targets (a quick, informal measurement on the ZIP fixture: ~2.37 s → ~1.45 s; the table has not been re-measured). ERA is **not** faster than REA; the negative native figures are run-to-run Hopper variance. Methodology: [docs/benchmark-methodology.md](docs/benchmark-methodology.md). ## ERA Project Lab Project Lab applies ERA to **external projects you are authorized to work on**. Three modes: * **Review**: read-only architecture, security and maintainability review with evidence, severity and file:line locations (REA for JavaScript projects, static checks, an optional read-only coding-agent review whose findings are accepted only if they point at real files, lines and quoted code). * **Modify**: "fix this bug", "add search", "refactor": ERA establishes a baseline with the project's own tests, plans, lets a configured coding agent change an **isolated copy**, re-runs the tests, detects regressions (including deleted tests), and produces a reviewable diff. You approve before anything touches the original; `rollback` undoes it. * **Rebuild**: from an authorized reference (for example a program you wrote or may analyze), ERA gathers REA evidence and observed behavior, writes a specification (no decompiled code is handed to the implementer), has an agent build an independent project, and compares it with the reference behavior. ```bash era project create demo --source ~/code/shop --mode modify --task "Fix the cart total" \ --test "npm test --silent" --accept-fail "node --test acceptance/cart.test.js" --agent claude-code era project modify demo && era project show demo era project approve demo # confirmation, then the verified diff is applied era project rollback demo ``` ### Multi-agent engineering Modify can coordinate **several coding agents by role**: an architecture agent designs (read-only), an implementation agent changes the isolated copy, an independent review agent inspects the verified diff (read-only), and the implementer makes bounded corrections that ERA re-tests. ERA itself runs the tests, validates the review findings against the real code and decides; you approve. Roles are configuration (`claude-only`, `codex-only`, `claude+codex`, or any agent per role), and remote providers need your explicit, recorded consent before any source code is sent. ```bash era agents # providers, capabilities, availability, sign-in era project create demo --source ~/code/shop --task "Fix the cart total" --team codex-only \ --allow-remote codex --accept-fail "node --test acceptance/cart.test.js" era project team demo && era project modify demo ``` **Validated for real:** OpenAI Codex (all roles, including a real end-to-end run and a real review that caught a planted flaw). **Not validated:** Claude Code. The CLI on the development machine was not signed in, so every Claude-specific path was tested only against scripted agents. Evidence and gaps: [provider-validation.md](docs/provider-validation.md); design and security: [multi-agent.md](docs/multi-agent.md); setup: [provider-configuration.md](docs/provider-configuration.md); tests: [multi-agent-testing.md](docs/multi-agent-testing.md). Per-project state, locks, limits and cost caps; cancellation, interruption and resume; a sandbox for the project's own code. Docs: [project-lab.md](docs/project-lab.md), [project-configuration.md](docs/project-configuration.md), [project-lab-testing.md](docs/project-lab-testing.md). **Limits:** macOS-focused sandbox; Project Lab has run with scripted mock agents and, for Codex only, real multi-agent runs on small projects (Claude Code is not validated); it does not recover source code and does not support arbitrary formats. ## Connect AI agents ERA keeps three roles separate: REA does the reverse engineering, a **coding agent** edits ERA in Improvement Lab experiments, and an optional **reasoning provider** explains analyses. **Use REA's tools from Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI or VS Code (MCP).** REA ships the MCP server and the client registration; ERA forwards to them: ```bash era rea setup --dry-run # show which installed clients REA would configure era rea setup # register REA's MCP server (REA asks before changing configs) era rea mcp doctor # validate MCP startup and tool listing ``` Clients then see REA's tools (for example `analyze_javascript_application`, `inspect_macho`). ERA's own commands are CLI-only in v0.1.0; there is no separate ERA MCP server. **Use Claude Code or Codex as the Improvement Lab coding agent.** ```bash era agents # what is installed era improve implement <cycle> --agent claude-code # claude -p, acceptEdits, --max-budget-usd era improve implement <cycle> --agent codex # codex exec -s workspace-write ``` Budgets are configured in `era.config.json` (`improvement.budget`). Both adapters are tested only against fake CLIs in this release. ## Providers | Role | Default | Options | |---|---|---| | Reverse-engineering engine | REA | — | | Coding agent | `manual` | `command`, `claude-code` (`claude -p`, `--max-budget-usd`), `codex` (`codex exec -s workspace-write`) | | Reasoning provider | `none` (local, deterministic) | `command` (for example a local model), `claude-code`; remote providers need `reasoning.allowRemote` **and** `--allow-remote` | ## Configuration ERA needs no API keys and never stores provider credentials; each coding-agent CLI keeps its own sign-in. Settings come from, in order of precedence, environment variables ([`.env.example`](.env.example) lists them), `era.config.json` in the working directory, and defaults. ```bash era status # environment, REA resolution, native providers, agents era agents # coding agents, reasoning providers and team presets era project check <id> # everything that would stop or weaken a run, with fixes ``` Reference: [providers](docs/provider-configuration.md) (Claude Code, Codex, custom command agents, spending and consent), [projects](docs/project-configuration.md) (commands, acceptance checks, protected paths, limits), [installation](docs/installation.md#configuration). Studio stores its own settings (theme, mode, permissions, consent, cost caps) under *Settings*. ## Security Authorized targets only. Analyzed artifacts and REA output are untrusted data, fenced in prompts and never followed as instructions. Runtime capture is sandboxed. Secrets are redacted from logs and prompts. Least-privilege CI with SHA-pinned actions, CodeQL, dependency audit, SBOM and provenance attestation. See [SECURITY.md](SECURITY.md), [Studio security](docs/ERA_STUDIO_SECURITY.md) and the [audit remediation](docs/ERA_V020_REMEDIATION.md). Remote coding agents receive project content only with your per-project consent, and credential-like files are withheld. Nothing is sent anywhere by default: remote consent is empty, unmetered agents are refused, and there is no telemetry. ## Testing ```bash npm test # node:test suite, about 3 minutes (real REA and sandbox where available) npm run lint && npm run check # ESLint; syntax, secret and local-path scan of tracked files npm run test:desktop --prefix apps/studio # real Electron scenarios npm run package --prefix apps/studio && npm run test:package --prefix apps/studio # unsigned app and its smoke test npm run clean-install-test # the packed CLI, installed in an empty project ``` [docs/testing.md](docs/testing.md) explains what is real and what is mocked; [docs/SUPPORT_AND_VALIDATION.md](docs/SUPPORT_AND_VALIDATION.md) records results per platform and revision. ## Troubleshooting [docs/troubleshooting.md](docs/troubleshooting.md) covers installation and launch, native analysis (Ghidra, Hopper), agents and approvals, and the Improvement Lab; it is also under **Documentation → Troubleshooting** in the app. ## Documentation | Topic | Page | |---|---| | Getting started, installation | [getting-started](docs/getting-started.md), [installation](docs/installation.md) | | ERA Studio | [Studio guide](docs/ERA_STUDIO.md), [architecture](docs/ERA_STUDIO_ARCHITECTURE.md) | | REA integration and reverse engineering | [REA integration](docs/rea-integration.md), [what ERA adds](docs/original-rea-analysis.md) | | Project Lab | [project-lab](docs/project-lab.md), [configuration](docs/project-configuration.md), [screenshot rebuild](docs/screenshot-rebuild.md) | | Multi-Agent Engineering | [multi-agent](docs/multi-agent.md), [providers](docs/provider-configuration.md) | | Improvement Lab | [improvement-lab](docs/improvement-lab.md), [self-improvement](docs/self-improvement.md) | | Security and privacy | [SECURITY.md](SECURITY.md), [Studio security](docs/ERA_STUDIO_SECURITY.md) | | Evidence, audits, releases | [support and validation](docs/SUPPORT_AND_VALIDATION.md), [audit](docs/ERA_V020_INDEPENDENT_AUDIT.md), [remediation](docs/ERA_V020_REMEDIATION.md), [release notes](docs/release-notes-v0.2.0.md), [index](docs/index.md) | ## Limitations * **Experimental.** Improvement-Lab cycles so far were implemented by a coding agent working interactively (`manual` mode), not by the `claude-code` or `codex` adapters. The adapters are tested against scripted CLIs; Codex was additionally validated in real runs (see [docs/provider-validation.md](docs/provider-validation.md)); Claude Code was never validated against the real CLI. * Native analysis covers Mach-O only. Hopper's public API exposes no complete function bodies, and Hopper occasionally hangs (ERA's watchdog contains it). * **REA limits observed:** a fixed JavaScript result budget (ERA splits its self-analysis around it); no retained call-argument string literals; CommonJS member exports are reported as `default` (ERA infers member names and marks them as inferred). * `sandbox-exec` is deprecated by Apple but functional. For project commands, terminals, the Git panel and Lab candidates the profile denies network and writes, denies signalling other processes, and closes the home directory and the shared temporary folders for reads (toolchains on `PATH`, ERA's own code and the working copy stay readable). Other locations (for example external volumes) remain readable, so it is still not a general confidentiality sandbox: analyze hostile binaries on an isolated host. Runtime capture of a reference binary uses the same profile; the coding-agent CLIs do not run under it. There is no sandbox on Linux, so Lab evaluation fails closed there. * `era rea …` passthrough runs REA directly: no ERA run records, sandbox, watchdog or redaction. Process-group containment does not cover deliberately detached descendants or GUI providers such as Hopper. * Lab benchmarks evaluate ERA's interpretation of supervisor-run REA Evidence. The candidate cannot read ground truth, but its extraction code produces the facts. Review the patch. * Raw REA evidence stays private (`.era/runs/`; 198 MB for cycle 5 alone). Published exports carry its SHA-256 digests, and anonymized files keep the original Evidence IDs. * Benchmarks use five small authored fixtures; scores are not representative of arbitrary software. No type checking (JSDoc/TypeScript) yet. * ERA does not modify, retrain or fine-tune any foundation model. "Self-improvement" means changes to ERA's own source code. ## Roadmap * **v0.2.0**: ERA Studio, Project Lab and configurable multi-agent engineering. Next: independent re-review of the audit fixes, broader platform validation, and (for a distributed app) signing and notarization. * **v0.3.0**: multi-agent verification, advanced workflows, team collaboration, role-based permissions. ## Acknowledgements ERA would not exist without **Reverse Engineer Anything (REA)** by [morluto](https://github.com/morluto) and its contributors (<https://github.com/morluto/rea>). REA provides every reverse-engineering capability ERA uses. ERA is an **independent open-source project**. It is not affiliated with or endorsed by the REA authors, and it does not modify or redistribute REA's source (REA is installed from npm as a dependency). ## Development ERA was developed by [dav1dngy](https://github.com/dav1dngy) together with **Claude** (Anthropic) using [Claude Code](https://claude.com/claude-code), which wrote and tested most of the implementation and acted as the coding agent in the Improvement Lab experiments. Claude is credited as a co-author in the commit history. ## Contributing and license See [CONTRIBUTING.md](CONTRIBUTING.md), [docs/testing.md](docs/testing.md) and [docs/release-checklist.md](docs/release-checklist.md). ERA is MIT-licensed ([LICENSE](LICENSE)). It integrates **Reverse Engineer Anything** by morluto (MIT); see [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).