Repository navigation
Description
Before submitting
- I searched the existing issues for a related documentation report, including demo, walkthrough, and showcase proposals; no matching portfolio tracker was found.
- I removed credentials, tokens, private paths, and sensitive target data.
Documentation location
README: product direction and quick start, tool catalog, the domain guides under docs/, and the bundled reverse-engineer-anything skill.
Proposed destination: a discoverable demo index linked from the README, with reproducible walkthroughs and recordings for selected analyst questions. The directory and recording format can be chosen during implementation.
Problem
REA — One MCP for reverse engineering across binaries, applications, and runtime behavior.
REA now covers substantially more than the Hopper workflow pictured in the README. At the initial review baseline, main commit 554e9b16, the generated catalog contains 125 MCP tools in 12 families, 80 primary CLI commands, and 18 provider descriptors. Those are inventory counts, not a claim that every provider or platform supports every operation.
Guides, examples, source-owned fixtures, and real verification scripts already exist. New users still need a coherent demonstration of how an analyst asks a question, selects a target, follows the evidence through several operations, and obtains a useful answer. A tool list or a passing verifier does not show that experience.
This issue tracks future demo work and provides a consistent way to select and add scenarios. It does not implement or record the demos now. The review for this issue inspected current contracts, catalog, guides, fixtures, and verification lanes; it did not rerun all domain toolchains.
Expected content
1. Select scenarios by the analyst outcome
Start each proposal with one concrete question. Identify the smallest target and observation that would let a viewer assess the answer. Then map that outcome to the current public CLI/MCP contracts and an independent fixture oracle.
Use these eligibility checks before scheduling a recording:
- Answerable today: the required operation exists in the connected server's public contracts, on the exact proposed provider/host/target combination. Internal helpers, roadmap entries, and mock tests are not a substitute.
- A checkable result: expected facts can be checked against source-owned source, known bytes/digests, a deliberate version change, or an explicitly captured result. State what remains inferred or unknown.
- A shareable target: use a small, benign, source-owned or explicitly redistributable fixture. Record its origin, license, generation/download procedure, and digest. Do not base public recordings on a private local application or database.
- A reproducible environment: name the tested REA artifact/commit, OS/architecture, provider version, and existing prerequisites. Keep fixture setup separate from the actual analysis and provide a focused stop/cleanup procedure.
- A bounded experiment: estimate input size, working space, duration, and engine cost; run one heavy engine operation at a time. Measure actual resource use when verifying the script. Unknown budgets remain unknown until measured.
Among eligible proposals, prioritize:
| Criterion | Prefer | Defer or simplify |
|---|---|---|
| Analyst value | A recognizable question with a concrete answer | A catalog tour or a sequence with no analytical conclusion |
| REA's contribution | One REA connection, useful follow-ups, retained Evidence where supported, or a clear handoff between capabilities | A recording dominated by unrelated engine installation or manual tool switching |
| Time to first result | Small inputs and few steps after prerequisites are configured | A large opaque application or lengthy analysis with no additional insight |
| Evidence clarity | A visible source location, function, graph path, declared relationship, captured UI change, or deliberate difference | A plausible narrative that cannot be checked from the returned evidence |
| Coverage added | A new domain, useful boundary, or substantially different provider lifecycle | Several engines repeating the same already-demonstrated result |
| Reproducibility | Deterministic fixtures and a CLI equivalent for the demonstrated MCP workflow | Machine-specific state, proprietary data, or fragile timing |
| Maintenance cost | Existing fixtures, verifiers, and guides; an explicit reason for each extra dependency | A new custom orchestration framework just for recording |
These are review criteria, not invented benchmark scores. Give the selection rationale in each scenario entry; do not rank unmeasured scenarios using fabricated performance numbers.
2. Maintain a domain coverage matrix
All entries below are proposed, not completed demos. “Starter” is the suggested first batch; “Follow-up” and “Specialist” are candidates to revisit using the criteria above. Each domain can gain more rows when a distinct analyst outcome warrants one.
| ID / batch | Analyst question and intended result | Existing public operations | Fixture / verification starting point and important boundary |
|---|---|---|---|
| D01 / Starter — JavaScript and Electron static analysis | Which modules handle a selected feature, and how do its IPC, configuration, or data relationships connect? Show one graph trace with source locations. | analyze_javascript_application, trace_application_feature, trace_javascript_semantics |
Source-owned readiness JS/Electron trees; verify:application-workflows with explicit inputs. Static analysis needs no Hopper/Ghidra or target execution. Recovered relationships do not prove execution; unresolved dynamic relationships stay unknown. |
| D02 / Starter — Native binary analysis | Which function implements a small computation or check, and what supports that explanation? Discover it, inspect pseudocode/assembly and references, then close the session. | open_binary, search_strings, search_procedures, analyze_function, xrefs, close_binary |
Existing small C conformance sources and verify:ghidra / verify:hopper / verify:ida, as applicable. Pick one configured provider for the main story. A short optional IDA variant can contrast an already-open GUI target with an owned headless database. IDA's initial real coverage is Windows; Ghidra Windows P0 has its narrower native x64 PE/read-only boundary. |
| D03 / Starter — Browser behavior | What changes when a user performs a selected action? Show the page result and captured activity, then compare a deliberately changed variant. | inspect_web_page, observe_web_session, capture_browser_scenario, compare_web_captures / compare_web_screenshots |
A source-owned loopback web fixture and verify:browser. Keep passive inspection and active interaction distinct. Wait for an explicit result-ready condition and request DOM/screenshot capture when the page changes without navigating; final URL alone is insufficient. |
| D04 / Follow-up — Build/version differences | What changed in a feature between two small application builds? Identify the paired entities, changed relationships or return shapes, and ambiguous matches. | compare_application_versions, compare_source_to_bundle, compare_javascript_export_shapes |
Existing JS application fixtures and verify:application-workflows. Static differences are not behavioral equivalence; partial coverage cannot establish absence. |
| D05 / Follow-up — .NET PE/CLI | Which managed members changed, and which native calls are declared? Show identity, metadata/CIL, and a deliberate member difference. | inspect_managed_artifact, inspect_managed_members, inspect_managed_native_boundaries, compare_managed_members |
Existing byte-built managed fixtures and scripts/verify-managed-conformance.mjs. Do not load the assembly. Single-file unpacking, IL2CPP decoding, automatic original C# recovery, and managed runtime observation are outside the shipped parser scope. |
| D06 / Follow-up — Android APK | What does the manifest declare, where is a selected method, and which incoming static references lead to it? | inspect_android_package, search_android_classes, inspect_android_class, inspect_android_method, trace_android_references |
Public hash-pinned Appium ApiDemos; fixtures:android and verify:android, with existing Java/JADX. Linux is verified; Windows provider boundary is unsupported and macOS remains unverified. No emulator/APK execution, signature validation, or claim that static references are runtime calls. |
| D07 / Follow-up — Firmware to native code | Which regions are reported, what files can be extracted, and how can one verified native child be analyzed through the same REA connection? | inspect_firmware_regions, extract_firmware, then open_binary / analyze_function on a returned child |
Small source-generated gzip/USTAR fixture; fixtures:firmware and verify:firmware, with the separate optional Ghidra handoff lane. Current real coverage is Linux x64; Binwalk/Unblob/prlimit are supplied separately. Decompressed offsets are not original flash addresses; no hardware/emulation or unsupported-architecture claim. |
| D08 / Follow-up — Node/Electron runtime evidence | Which scripts or contexts were observed, and how do they relate to the static application? Optionally capture a declared Electron action separately. | list_javascript_runtime_targets, observe_javascript_runtime, reconcile_javascript_runtime; separate capture_electron_scenario |
Existing Inspector target fixture, verify:inspector, and the optional real verify:electron lane. Passive scriptParsed does not establish module execution, importing callers, or IPC events. Active Electron captures have different authority and lifecycle. |
| D09 / Follow-up — Process behavior and reconstruction | Where do two declared command runs first differ? Preserve the difference, incomplete observations, and the remaining reconstruction obligations. | capture_process_scenario, compare_process_captures, record_unknown, build_reconstruction_obligation_ledger, evaluate_reconstruction_coverage, Evidence bundle tools |
Source-owned readiness JS CLI and verify:readiness. Show a deliberate mismatch and an unknown. Selected filesystem snapshots are not syscall traces; complete equivalence is not inferred from matching partial captures. This is direct capture, not arbitrary replay execution. |
| D10 / Specialist — Apple resources and native UI | How is a selected UI resource wired to a native handler? Show resource objects/dispatch evidence; add selected-window observation only on a verified host. | inspect_artifact, decode_interface_builder, inspect_keyed_archive, inspect_asset_catalog, inspect_native_dispatch_metadata, trace_native_ui_action; optional observe_native_ui / capture_native_ui_scenario |
Source-owned XIB/Objective-C/Swift/window fixtures and their real macOS lanes. Asset tools and UI capture have host/tool prerequisites. Static wiring does not prove an action executed. A permission-denial test is not positive UI capture evidence. |
| D11 / Specialist — NativeAOT | What recovered type relationships and virtual slots help explain a native function? | Ghidra inspect_native_load_image, inspect_native_data_type, analyze_function |
Optional source-built NativeAOT fixtures and verify:ghidra:nativeaot. Tested scope is the documented .NET 8.0.22/RTR 9.1 x64 ELF/PE targets on Linux x64. Recovery modifies only ephemeral analysis memory; inferred names/layouts and derived ranges do not recover original C# or runtime state. |
| D12 / Specialist — DOS MZ/COM | How do load addresses, file bytes and 16-bit function analysis relate? | Ghidra open_binary, inspect_native_load_image, read_bytes, address_to_file_offset, analyze_function |
Generated MZ/COM fixtures; verify:ghidra:dos / verify:ghidra:com. Linux real workflows are verified. COM interpretation is explicit; analysis segment settings do not prove DOS/PC-98 runtime behavior. |
The matrix intentionally does not enumerate every tool or every engine combination. Use the actual connected MCP tool list and session availability when writing a script. rea providers / rea capabilities alone do not inventory every application, Android, or browser workflow. Use open_binary for workflows whose contract requires a binary session; Android, firmware, browser and other explicit-object tools have their own inputs. Do not force all domains through one invented preparation flow.
3. Give every demo the same useful deliverables
- A short analyst question and visible conclusion, followed by the evidence supporting it and any unresolved part.
- One configured REA MCP connection and a real agent transcript, with an equivalent CLI recipe. Use existing same-connection Evidence references where advertised; do not require manual copying of large JSON records during the agent demo.
- A compact written walkthrough, a recording or terminal capture, and a link from the demo index. Show representative results inline; keep full reproducibility instructions available alongside the recording.
- Target source/provenance, exact generation or hash-verified download instructions, tested package version/commit and provider/host details. Verify the demonstrated artifact rather than assuming
@latestincludes repository main. - Actual time to first result, scenario duration, sampled process-family memory/CPU and working-space usage where measurable, plus cleanup outcome. Label estimates and samples precisely; do not present them as hard engine limits.
- A focused negative/control observation when relevant: for example an ambiguous match, unsupported facet, deliberately changed output, or incomplete capture. Show the corresponding actionable result without turning the walkthrough into a permission/setup checklist.
- Public material checked for machine-specific paths, credentials, private code, and unrelated desktop/browser data. Use a clean recording workspace and shareable targets. Label edited output as an excerpt; do not present a modified Evidence payload as an authenticated original.
4. Keep this issue easy to extend
Use statuses Proposed, Selected, Scripted, Verified, Recorded, and Published; use Blocked with a concrete prerequisite or linked defect. “Verified” requires the actual advertised workflow on its declared host/provider. A mock or permission-denial result covers only its stated narrower boundary.
Copy this block for a new scenario or replace it with a link to the resulting walkthrough/PR:
#### Dxx — <scenario title>
- Status / owner:
- Analyst question / expected answer:
- Why select it / coverage added:
- Target source, license, digest and generation procedure:
- Public REA operations / CLI recipe:
- Host, REA artifact and provider prerequisites:
- Independent oracle / meaningful unknown or control:
- Estimated cost; measured time, CPU/RSS, temporary/output space:
- Lifecycle and cleanup result:
- Verification command / report:
- Script, recording, documentation and PR links:
- Blocker or next action:Keep the domain matrix and task list current. Add a new row for a distinct question; add a provider variant to an existing scenario when the outcome is unchanged. Link existing capability defects instead of duplicating them. Split out an implementation issue only when a scenario reveals substantial code work, and link it here while keeping the demo status explicit.
Initial task list and completion criteria
- Confirm the starter questions, fixture choices, and selection rationale for D01–D03.
- Write reproducible scripts/walkthroughs for the starter demos using current public contracts.
- Verify the demonstrated package and real workflows, including cleanup, evidence limits, and measured resource cost.
- Record the starter demos and publish a README-linked index with equivalent CLI recipes.
- Test the documented paths as a new user with only the stated prerequisites; record and link usability failures.
- Fill in ownership, next action and priority rationale for the follow-up/specialist matrix; keep deferred rows visible.
- Document how to add/update a scenario and what invalidates its previous verification (contract, fixture, provider or host changes).
The initial deliverable is a reproducible starter portfolio plus a maintained extension path, not 125 individual tool videos. This issue can close once the starter portfolio is published and remaining candidates have explicit follow-up links or deferral records; future additions can use the same template.
Supporting evidence
- Baseline catalog: exact source-generated tool families, provider coverage, and CLI inventory.
- Testing guide and source-owned conformance fixtures: existing real lanes and their precise proof boundaries.
- JavaScript workflows, browser scenarios, and reconstruction readiness.
- IDA, Android, firmware, .NET, NativeAOT, native/Apple investigation, and DOS guides.
Additional context
Related product direction: #32. Investigation/capability organization is tracked separately in #740; this demo portfolio can proceed without waiting for that refactor. Domain provider work is referenced in #717, #718, and the merged IDA integration #747/#754; demo status must follow actual current contracts and verification rather than those issues' open/closed state.
Suggested labels: documentation, triage, P2: medium, area: cli, area: mcp, area: tests. This is one documentation/demo tracking issue; separate subissues are unnecessary until a concrete demo needs substantial implementation work.
Activity
014-code commentedon Oct 11, 2026
Implementation update for D01-D03: PR #1704. I kept the issue open because several requested proof items still need real provider, agent and recording runs.
I worked through the scenarios in this order:
I checked the current public CLI/MCP contracts, the existing source-owned examples and the verifier lanes, then used the three starter questions already listed here. The implementation adds a README-linked demo page, a guide-hub entry and a separate evidence record rather than treating a verifier report as a published recording or benchmark.
For D01, I analyzed the six-file Notes Electron fixture with the public CLI and saved the analysis, feature-trace and semantic-trace Evidence IDs in the evidence record. The application graph had 29 nodes and 42 edges; tracing the exact
notes:exportfeature returned complete coverage of all 29 nodes and 42 edges. A forward-influence semantic query seeded by propertyinvokewas ambiguous, partial and had one relevant unknown. A literal semantic seed fornotes:exporthad no match, so the page keeps the IPC channel claim tied to the application graph rather than presenting that string as a semantic result.The source-backed data flow is
renderer.js->preload.js->main.js->csv.js: the renderer callswindow.notes.exportCsv(), the preload invokesnotes:export, and the main handler writesnotes.csvunder Electron's Downloads path. The CSV writer emits theid,titleheader and note rows; the handler returns the destination for the renderer's status message. This is static analysis, so it does not prove that a click or write occurred or resolve the actual Downloads path. I also extended the agent evaluation scenario to require the analysis, feature trace and semantic query in order. A usable real Codex transcript was not available on this host, so I did not claim that agent workflow as verified.Evidence IDs: analysis
ev_46841fb660ac9c02a2932b465177397114a4aaaed9fecdebd79b6904d976bd30; feature traceev_3ece6d552f75f7d4ecb4e303e12703dec3a966a1d479fd1c3aad4685dce68b5e; semantic traceev_68c6438a3907198bd130fe5b7fd7fb201372dbd1ab80311a4fd131e97210b6f8. The root artifact SHA-256 is3c11d59338104ba96399bfd05bebf4359ac6a293f49d2b095566456842f4396e.For D02, I used the existing C conformance fixture (
rea_entry->rea_branch->rea_leaf) and documented fixture generation, integrity checks, Ghidra CLI steps, the corresponding MCP lifecycle andclose_binarycleanup. This Windows host had neither a C compiler nor a configured Ghidra, IDA or Hopper provider. I therefore recorded the provider run as blocked and did not invent pseudocode, references or a cleanup result.For D03, I added a loopback-only
/portfolio-demoroute whose Export notes button displaysExported 2 notesorExported 3 notes. The real-browser verifier launches Chrome for both variants, clicks the same button, waits for the result-ready DOM state, captures the DOM through the public CLI and runscompare-web-captureson the normalized results. Runb8b2d376-d5a4-495a-b558-dedfc0884d1dused Node.js 24.18.0, npm 11.16.0 and Chrome 154.0.8037.99. The captures aligned, the overall status waschanged, and the result step had one changed DOM artifact. The verifier terminated its owned Chrome processes, removed both temporary profiles and comparison inputs, and closed the page proxy and fixture server.Validation performed:
npm run build:cached,npm run check:fast,npm run verify:windows-native:artifact, andnpm run verify:browserpassed.python scripts/website.py checkpassed with 20 HTML pages, 437 local references, sitemap/metadata checks and 7 maintenance tests. Scoped Oxfmt checks on the 9 changed files,git diff --check, the pre-commit format/lint hooks and the pre-pushcheck:fastpassed. The full-repositorynpm run format:checkflags even unchangedCONTRIBUTING.mdin this Windows checkout becausecore.autocrlf=true; the PR files pass the scoped formatter check.Still outstanding for this issue: a real D01 agent transcript with retained Evidence citations, a real D02 provider run with cleanup evidence, walkthrough recordings, a clean-user setup run, and measured time/CPU/RSS/temporary-space costs. These gaps and their boundaries are visible in the new evidence record; this PR does not claim the issue is complete.
0xRafie commentedon Oct 12, 2026
Verification record for D03 (browser behavior): notes-web demo on two Chromium-family browsers, using the scenario block from this issue.
D03 — notes-web Export notes, connect mode on Chrome and Brave
notes.json(200), builds a CSV in the page, and the status reads "Exported 2 notes to notes.csv.".website/public/examples/notes-web/from maind96550f, served withpython3 -m http.server --bind 127.0.0.1 8791fromwebsite/public.export.jssha25666f5c2501bfca72eb743a1fffbdce9a58d295e69ca66f19562e9b0fa9fc189a6(925 bytes). REA's captured source digest matches the file.{"browser":{"mode":"connect","cdp_endpoint":"http://127.0.0.1:<port>","target_id":"<target_id>"}, "start_url":{"url":"http://127.0.0.1:8791/examples/notes-web/"}, "actions":[{"step_id":"export","action":"click","locator":{"kind":"css","selector":"#export"},"timeout_ms":10000}, {"step_id":"wait","action":"wait_for_timeout","duration_ms":500}], "capture":{"events":["network"],"network":{"response_body":true},"at_end":["dom","url"]}}d96550f(node scripts/rea.mjs); Node 26.9.0. Browsers: Google Chrome 154.0.8037.99 and Brave Browser 155.1.97.56 (Chromium 155.0.8059.40), both headless.scenario_start,export,waitcompleted. In theexportstep:GET http://127.0.0.1:8791/examples/notes-web/notes.json, resource_typefetch, response status200. The final DOM status reads "Exported 2 notes to notes.csv." and matches the handler inexport.js.clicklistener on#exportatexport.jsline 0, column 60 (source_association: script_id).inspect-web-page(4 s window) while the button was clicked through a separate CDPRuntime.evaluatereportednotes.jsonGET Fetch 200 with initiator{type: script, url: .../notes-web/export.js, line: 3, column: 27}(Chrome only).inspect-web-pagewhile the connect-mode scenario re-navigated itsstart_urlfailed withtarget_unavailable/target_changed, so passive and driven capture should not overlap on one target.completenesswasincomplete(missing_sections: ["events"], "CDP attachment cannot recover pre-attach events...") on both browsers. No changed-variantcompare-web-capturesrun in this record.browser.cleanup: disconnected-external, and the external browser stayed alive after each capture. I then stopped both browsers and the server and deleted the temp profiles.npm run verify:browsernot run for this record.compare-web-capturesin connect mode.pablodiazjorge commentedon Oct 15, 2026
Native Windows walkthroughs, in case they help the index: rea-lab keeps reproducible writeups run with rea 6.4.0 and real Ghidra 12.1.4 on Windows 11, each with its scripted session and sanitized, re-authenticated Evidence: WinHelloCPP (the Ghidra sample), Wallace v1 against v2 with
compare_functions, and a seeded crackme at -O0 and -O2. They live outside the REA repository, so they fit as external references rather than numbered demos.TheDesignFounder commentedon Oct 16, 2026
I've opened a draft walkthrough
for a small game example that could fit this collection: does moving diagonally make the player faster?
In Relay garden,
we used REA to locate the movement function in a small JavaScript reference we
wrote. Reading that function showed that full diagonal and straight input cover
the same distance, away from the edges of the game world. A separate
implementation matched the reference across 2,000 test inputs.
The proposed walkthrough includes the command-line steps, recorded MCP requests
and responses, and instructions for repeating the checks. It also explains where
REA's analysis was incomplete, including a search that missed a number present
in the code.
There are two videos: a 30-second gameplay clip
and a separate guided recording of the investigation. The investigation recording follows a prepared
sequence using real REA calls and test results. We knew the reference code
beforehand.
Play the game ·
Read the article
We added original artwork using Dreamlayer. This would add one example to the
work in #1704 and leave the broader goals of #778 open.
Would this be a useful addition?
ylwl1997 commentedon Oct 19, 2026
Follow-up scenario candidate (community / unofficial)
Thanks for maintaining this demo-portfolio tracker — the D01–D12 matrix is a clear way to schedule reproducible walkthroughs.
Suggestion: consider a Follow-up row (adjacent to D06 Android APK) for own Android TV / ADB-authorized device ops + optional APK inspect:
adb/pm disable-user, and what does a chosen open-source sideload APK declare before install?inspect_android_package(and related Android tools) on an APK the operator already chose; device lane stays on an authorizedadb devicessession (human gate). Main ops path can be agent + platform-tools ADB.Community teaching write-up (independent / not affiliated with REA maintainers; not a claim that this is an official demo entry):
Related site notes: https://rea.run/docs/android-adb/ · https://rea.run/compliance/ · https://rea.run/faq/legal-boundary/
Happy to adjust wording if maintainers prefer this stays outside the formal matrix until a shareable fixture + recording exist. Official source of truth remains https://github.com/morluto/rea.
8rulerstar commentedon Oct 21, 2026
A possible IL2CPP scenario, in case it fits the tracker. Location and format are entirely up to you.
Dxx (Unity IL2CPP): which assembly and type declare a game script's methods, and where is their native code?
docs/unity-il2cpp-demo, based on maind0067ba34; not a PR.Sample.Player.Update, what do it and its nested type declare, and which native address does the build linkUpdateto? Answer:Assembly-CSharp.dll(image 35) declaresSample.Player(type 2459) withUpdate,.ctorand fieldslot; nestedSlot(2458) declaresAdd,.ctorand fieldCount;Addhas no parameters. With the arm64 slice ofGameAssembly.dylib,Update(method 12477) links to0x153560, statecode.inspect_il2cpp_metadata(feat(il2cpp): inventory Unity IL2CPP global-metadata images, types and methods #3807, feat(il2cpp): read member names, string literals and metadata versions 24 and 27 #5326, feat(il2cpp): link methods to native addresses and write Ghidra seeds #5327) covers the metadata and its native address links.2875f161ae0da99b83932b52aace567747070a45af9d696bc52a725a58c358b7, version 31) but three differentGameAssembly.dylibdigests at the same size; the walkthrough records the one used.inspect-il2cpp-metadata --detail summary --evidence_output,inspect-analysis-viewover images, types, methods, fields, parameters (method_indices) andmethod_addresses, and--binary --architecture arm64 --ghidra-seeds. MCP parity was measured on an earlier build (c374c2fe7) and not rerun.CiBuild(editor-only) returns no rows. No independent oracle for the addresses.npm run verify:il2cpp:metadatapassed in 10.0 s; the--binarydecode 1.3 s, 498 MB, 11,022 seeds. Token cost unmeasured.analyze_functionon0x153560is unmeasured, because the local Ghidra 12.1.4 has no macOS arm64 native decompiler andopen_binaryrejected the provider.Thank you for your time.