Use REA when the user wants to understand how an app or feature works, compare app versions, decompile code, or build a similar feature.
For a website already open in a user-owned Chrome-family browser, start with
list_browser_targets and inspect_web_page instead of opening a binary. Use
analyze_web_bundle, observe_web_session, discover_webmcp_tools, capture
comparison, or screenshot tools only for the corresponding investigation. Ask
the user to approve the exact page origins and loopback CDP endpoint before the
first call. Browser observation is disabled unless the operator configured
REA_BROWSER_OBSERVE_ENABLED, REA_BROWSER_CDP_ENDPOINTS_JSON, and
REA_BROWSER_ALLOWED_ORIGINS_JSON before starting REA.
Browser inspection is passive. Never claim that it clicked, navigated, evaluated page JavaScript, captured prior activity, or contained the page's network. Query values, credentials, cookies, authorization headers, storage values, and raw WebSocket or JSON values are deliberately absent. Accessibility text, bounded redacted console primitives, value-free JSON/WebSocket shapes, script sources, storage key names, source-map fetches, and screenshot pixels each require their documented independent opt-in or approval. Treat truncation, policy filtering, and attach-window coverage as explicit limitations, and cite the returned Evidence v2 IDs. WebMCP declarations are page-declared untrusted; REA inventories them but never invokes them.
Electron file:// pages use list_electron_targets and
inspect_electron_page with a separate loopback endpoint and canonical
filesystem-root authority. Never claim that Electron inspection invoked IPC,
evaluated renderer JavaScript, navigated, or escaped an approved root.
Understand the request
Identify the app and what the user wants to understand or build. Do not ask for information they already provided.
If the app is missing, ask which app they want to reverse engineer. If the app is known but the goal is unclear, ask what they want to understand or build, and offer to start with an overview. Never require the user to supply a program path, address, architecture, or reverse-engineering terminology.
Notes is only a documentation example. Never select an app unless the user names it or confirms it.
Ensure REA is ready
- Run
npx -y rea-agents doctor. - If setup is needed, tell the user REA needs to install its local binary-analysis tools. Do not lead with implementation details or assume the user knows reverse-engineering products.
- Before installing external software, obtain approval and identify what will be installed. If deeper analysis needs Hopper, describe it as REA's local analysis engine and note that it is a separate Mac app with its own license. Then run
npx -y rea-agents setup --yes. - If macOS or an installer requests human input, tell the user exactly what needs attention. After they finish, rerun setup and doctor.
- If setup registers a new MCP server, tell the user to restart their agent to load all REA tools. Direct CLI commands remain available before restart.
Locate the app
Accept a human-readable app name. Search macOS application locations and system metadata. If one clear match is found, continue without asking for a path. If several apps match, show their names and locations and ask which one the user means. If none match, ask where the app is installed.
REA accepts a .app bundle directly. Do not expose its internal Contents/MacOS path unless it helps explain an error.
Investigate
Briefly tell the user what you will investigate. Open the app with open_binary, begin with binary_overview, and narrow the investigation around the requested feature. Use decompilation, strings, names, callers, callees, and cross-references as needed.
For applications, ZIP/APK/IPA packages, Electron ASAR archives, or DMGs, call
inventory_artifact before extraction. Follow its deterministic occurrence
pages and cite graph manifest IDs. extract_artifact requires explicit user
approval, an absent absolute output root, and selected occurrence IDs; never
extract every entry implicitly. Symlinks and encrypted entries are inventory
facts, not extractable files. Inventory traverses discovered ASARs within the
same graph and shared limits, including ASARs inside an approved mounted DMG.
Native DMG traversal is macOS-only, read-only, and disabled by default. Use it
only when the user approves that inventory call with
native_mount_approved: true and the operator has separately enabled
REA_ARTIFACT_NATIVE_MOUNT_ENABLED=true. Without both gates, retain the DMG
root-hash-only result. Never imply that approval changes extraction authority.
Explain conclusions in plain language. Point to the relevant decompiled code, strings, names, and connections so the user can see how the explanation was reached. Do not claim to recover original source code or automatically clone an application.
Search uses bounded deterministic pages. Prefer literal mode; use regex mode only
when regex semantics are needed, and continue from next_offset while
has_more is true. Treat nullable permissions and explicit unavailable metadata
as unknown, never as false.
Every successful analysis result is Evidence v2. Cite evidence IDs and preserve limitations and residual unknowns. Process capture is disabled by default, requires per-call approval plus operator policy, and uses host networking only when the operator explicitly permits it. It is behavioral evidence, not a security sandbox.
capture_process_scenario produces Process Capture v4 only. If a boundary
reports Process Capture v3, tell the user or calling agent to rerun the original
scenario; v3 cannot be upgraded because it lacks required manifest and
settlement evidence. Treat the v4 manifest commitments as compatibility and
provenance evidence, and distinguish root exit from descendant settlement.
When comparing stored captures, set max_capture_age_ms when the task requires
fresh behavioral evidence. Different scenario and executable digests are
allowed, but schema and comparison-contract digests must be compatible.
Do not let unanswered questions disappear. Use record_unknown only with
explicit approval, attach supporting and contradicting evidence IDs, and record
the authority/environment still required. Use update_unknown with the current
expected_revision; stale updates must be re-read, not retried blindly. A
verified resolution needs qualifying observed evidence. Inference, withdrawn,
and out-of-scope dispositions are never substitutes for observed behavior.
Set unknown_registry_approved: true on trace_feature or
capture_process_scenario only when the user approves durable automatic
recording of their bounded residuals. The same flag on a direct operation
records typed provider unavailability, and on compare_process_captures
records an observed disagreement as a contradiction.
Use compare_artifacts with one record or bounded arrays of
inventory_artifact Evidence pages per side. Pages must share one manifest;
collect all node, occurrence, and edge pages for exhaustive comparison. It
compares stable occurrence paths, content, metadata, and graph relations,
cites both evidence sets on every delta, and paginates changes. Incomplete
inventories are truncated or unknown, never equivalent. Set
unknown_registry_approved: true only with approval to preserve the
disagreement or missing evidence as a residual unknown.
Use compare_functions with explicit analyze_function Evidence page sets;
it does not perform fuzzy whole-binary matching. Collect every pseudocode,
assembly, collection, and CFG page when exhaustive comparison matters.
Absolute addresses are volatile only in CFG topology; pseudocode constants are
never stripped. Omitted assembly, truncated scans, unavailable reference kinds,
and cross-provider text remain unknown rather than equal.
Use compare_bundles for canonical Evidence v2 bundle membership and
residual-unknown history changes. Pair cross-version observations explicitly;
the tool never guesses record identity. One-sided records prove only bundle
inclusion or omission, not behavioral absence. Use the returned canonical
bundle digests to anchor paginated reports.
Use find_changed_behavior to combine existing comparison Evidence. Runtime
process differences are observed changes; artifact and function differences
remain static candidates, not causal proof. Supply complete comparison pages.
For an automatic two-version artifact investigation, use its
investigation_run mode with explicit write approval and a workspace beneath
REA_EVIDENCE_ROOTS_JSON. It checkpoints inventory, comparison, and report
Evidence in monotonic CAS-linked revisions. Repeating the same content and
budgets resumes or reuses the run. This automatic mode does not execute either
version or perform fuzzy function matching, so its behavior status remains
unknown without separate controlled runtime Evidence.
Use build_call_path with explicit analyze_function Evidence groups from one
artifact and provider. Select endpoints by exact address. Missing dossiers,
incomplete callee pages, or a depth frontier make absence unknown; found paths
remain valid and cite each contributing dossier.
Use correlate_static_and_runtime only with explicit mappings between exact
static and runtime comparison findings. A matching pattern is a hypothesis,
never proof of causality. Declare side alignment; unmapped similarities are not
correlated.
Use verify_reconstruction with a finite typed specification and canonical
Evidence bundle. Pass means all declared claims passed with comparable
authority, not global implementation equivalence. Missing, limited, active-
unknown, or incompatible evidence stays unknown; observed differences fail.
Build
When requested, use normal coding tools to build a version suited to the user's project, stack, interface, and requirements. Keep the implementation tied to what the investigation established, and distinguish observed behavior from assumptions or design choices.
Human input and cleanup
Hopper or macOS may show a window that needs human input. Tell the user what appeared and ask them to handle it; do not guess or take over unrelated UI. Call close_binary when the investigation is complete.