<!-- Adapted for Hermes Agent from REA's bundled skill (rea-agents 6.7.0, https://github.com/morluto/rea, MIT, Copyright (c) 2026 morluto). Changes: Hermes connection section, confirm-before-execute list, Hermes recovery guide. -->

REA

Use REA when a claim depends on a shipped binary or package, decompilation, passive application runtime evidence, or comparison with behavior not established by available source. For ordinary analysis of a complete source repository, use normal repository tools and do not run REA readiness or provider commands.

Connect only when needed

In Hermes this skill ships inside the rea plugin, and the same plugin registers the REA MCP server (rea-agents@6.7.0, stdio over npx). Nothing else needs registering: never run rea-agents setup, rea setup or hermes mcp add rea. Setup would add a second mcp_servers.rea entry to config.yaml that replaces the plugin's server and its pinned version.

Hermes loads MCP tools on demand. REA's tools (about 200) are behind the tool_search / tool_describe / tool_call bridge, not in the visible tool list. Search by capability (rea open binary, rea javascript application, rea android package), load the schema with tool_describe, then invoke it with tool_call. A tool name you cannot find this way is not available in this session.

If no REA tools are found, a call fails to start, or a required provider is unavailable, read connection and recovery. Diagnose the observed blocker, continue useful supported work, and resume the original task after repair.

Route the target first

Choose the first tool from the target the user supplied. Use open_binary for active-target native or archive workflows; target-free tools take their own explicit path or endpoint and do not need it.

• ASAR or extracted JavaScript/Electron tree: analyze_javascript_application.

• Archive/package member inventory (ZIP/APK/IPA/MSIX/AppX or DMG): open_binary with the supplied local path; use inspect_artifact when its graph and findings help answer the question.

• Android APK code, classes, methods, or incoming references: inspect_android_package, then focused Android tools when advertised. Archive member inventory still uses the archive route above.

• Managed PE/CLI assembly: inspect_managed_artifact.

• Firmware image: inspect_firmware_regions when advertised. Use extract_firmware when extraction is requested, with a caller-selected new absolute output directory. These tools use caller-supplied Binwalk/Unblob on Linux; see the firmware guide.

• .NET NativeAOT PE/ELF: native Ghidra analysis. Native code can yield pseudocode; consult the NativeAOT workflow in references/native-and-artifacts.md for optional metadata recovery and its supported host/layout boundary.

• User-owned browser page already open: list_browser_targets.

• Browser action the agent must perform (click, type, navigate): capture_browser_scenario with declared actions. See runtime observation for connect/launch and the focused follow-up browser tools.

• Fiddler Everywhere traffic or a native .saz capture: read the Fiddler workflow. Inspect saved SAZ with inspect_web_network_capture and format: "saz"; connect Everywhere's separate MCP server for live traffic.

• Retained HAR/SAZ/native mitmproxy/aioquic qlog capture: inspect_web_network_capture. Select original record ordinals when useful; inspect producer fields and byte/number sidecars without fetching recorded URLs or claiming live attribution. See historical captures for the exact upstream profile and credential exclusions. Qlog support starts with aioquic 1.5.0 JSON 0.3; stream coverage and endpoint loss declarations do not recover bodies or establish network loss. Prefer detail summary, then kind qlog trace/stream pages; select one trace, stream ID and direction for contribution/range pages. Facts retain all-recorded-events-before-selection scope. Other qlog layouts and PCAP correlation are unsupported.

• User-owned Electron runtime already open: list_electron_targets. Before an electron-main Inspector attempt, check packaged fuses with inspect_electron_fuses; disabled EnableNodeCliInspectArguments means the app ignores --inspect.

• Packet-field discovery: inspect_packet_fields searches installed protocols and field names/types before choosing fields for inspect_packet_capture. Query/page the relevant records; registrations do not establish observed traffic.

• Recorded packet captures: inspect_packet_capture with selected TShark fields, display filters and decode-as rules. Requires Linux, util-linux prlimit and REA_TSHARK_COMMAND; see capture guide. Packet limits count input packets before filtering. Output defaults to 50 matching frames; use match_offset / match_limit for later pages, or explicit null limit for complete matching output. Preserve the source digest across pages. This does not perform live capture. For address/DNS/SNI/protocol summaries use inspect_pcap_capture with MCP detail summary, then retained kind pcap pages. Requires Wireshark 4.x tshark and capinfos, with Linux prlimit or Windows Job Objects; see packet summaries. Select an exact DNS question for address pages; page exhaustion does not prove network completeness. This reads recordings and does not acquire traffic.

• Decoder discovery: inspect_signal_decoders searches installed decoder IDs or selected decoder help, including channels and options. Query and page the relevant details before choosing settings; this does not scan hardware.

• Recorded logic/signal traces: decode_signal_recording with explicit decoder, channel and protocol settings. On Linux with util-linux prlimit, supply sigrok-cli 0.7.2 through REA_SIGROK_COMMAND; see recording guide. Output defaults to 50 literal lines; select annotation classes and page with line_offset / line_limit, or explicit null for all remaining output. Window counts are display lines, not annotations or samples. Preserve source digests across pages. This reads existing files; it does not acquire hardware signals.

• Explicit local WebAssembly artifact: inspect_wasm_artifact. Supply WABT 1.0.42 through REA_WABT_BIN_DIRECTORY; retain exact byte/tool identity and decoded WAT. Select JavaScript glue and candidate paths explicitly and preserve URL/path ambiguity. No implicit fetch or target execution; WAT is not original source. See the WASM guide.

• Explicit local EVM bytecode carrier: inspect_evm_interface with caller-selected raw or hex encoding. Preserve carrier/decoded digests and treat selectors, argument strings and mutability as inferred candidates; this performs no chain lookup or target execution. For the complete inferred interface without hex or duplicate stdout, discover support and select detail: "summary"; complete Evidence stays retained under its original identity. Use the returned parent_evidence_id for exact function or bytecode_hex/stdout/stderr pages. Candidate coverage does not establish interface completeness. See the EVM bytecode guide.

• Embedded Go compiler, module or build-setting metadata: inspect_go_binary with the explicit local file. This built-in target-free inspection requires no Go installation or disassembler. Preserve raw module bytes and source ranges; absent metadata does not prove a binary is not Go, and unparsed text records remain explicit. See the Go binary guide.

• Embedded cargo-auditable Rust dependency metadata: inspect_rust_binary with the explicit local ELF, PE, or thin Mach-O file. This built-in target-free inspection requires no Rust installation or disassembler. Preserve package order, dependency indices, and the section file range. A missing .dep-v0 section is null metadata, not proof the binary has no dependencies. A rustc commit hash is an inference from a read-only /rustc/<40-hex>/ path. See the Rust binary guide.

• Explicit local SQLite database snapshot: inspect_sqlite_database. Inspect with Node.js 24.x >=24.15 or 26+ providing native SQLite limits. Read schema first (compact inventory by default), or select an exact ordinary table and columns when rows are needed. Defaults are selected-table schema and 20 rows; request schema_scope: "database" for the complete stored schema. Preserve both database and existing sibling WAL identity; row samples report truncation and have unspecified order. Use a quiescent snapshot rather than assuming a live database/WAL pair was copied atomically. See the SQLite guide.

• Explicit Linux ELF file for offline layout, symbols, relocations or static mitigation evidence: inspect_binary_layout. This target-free operation uses caller-supplied pwntools without opening a disassembler database. Preserve its raw locations and inference/coverage limits; see the offline binary guide.

• Supplied Linux x86-64 ELF core: inspect_recorded_crash. Read every recorded thread, raw note and source-bound register without opening a live target. Historical PIDs do not select live processes. Request include_debugger_context only when core-only mapping candidates help; see the recorded crash guide.

• PE resource type/name/language trees and static icon group references: inspect_pe_resources with an explicit path. This portable operation requires no active target or engine; opaque payloads retain file ranges and SHA-256.

• Native executable, library, or analysis database: open_binary, then use focused analysis tools directly; call binary_overview when metadata or inventory context is useful and available from the selected provider.

For an existing IDA MCP configuration, use open_binary with the original input binary and provider_id: "ida"; do not pass an .idb or .i64 database. The legacy attached profile binds the already-open GUI input by SHA-256 and leaves its database open. The modern headless profile creates and closes an owned private database without saving. Installation and registration reuse mrexodia/ida-pro-mcp; see the IDA provider guide. Proceed directly to analyze_function, pseudocode, or function/string searches after opening; IDA does not supply binary_overview. Consult session availability when composing broader workflows. Modern direct callers and unsupported dossier facets remain unknown; live IDA results are not replayed from snapshots.

With REA_KUNA_COMMAND configured, provider_id: "kuna" answers procedure lists, lookups, pseudocode, assembly and analyze_function from Kuna's cached whole-program export. Callers/callees use a separate body-free graph shared by the session; xrefs returns exact-address incoming sites and filters forwarding aliases. These queries and inventory lookups do not trigger a C export. Indirect targets remain unknown and can make relationship coverage partial. Reference caches expire on close; a fresh CLI process repeats the native reference walk. Typed references, CFG, raw-memory tools and annotations remain unsupported. A cold first answer can come from a single-function run (its limitations say so); assembly waits for the export. For pseudocode and assembly, raw_result.source names the cached .c/.asm file and byte span, which you may search directly instead of calling per function. See the Kuna provider guide.

If the user says only "my app" without a name, path, or target already selected for this task, ask which app before searching for artifacts. Paths in unrelated saved reports and nearby test fixtures do not select the user's target. Resolve an actual human-readable app name to one clear installed artifact when possible; ask when matches are ambiguous. Never choose an example app on the user's behalf.

In a target-free session, use open_binary to bind any archive/package or native target whose analysis tool operates on the active target. Do not call an unadvertised tool; inspect binary_session with {} and its result.tool_availability for availability reasons and remediation. Expand each group's names into entries, inheriting the inventory-level client requirements unless the group overrides them. In its closed effect and annotation vocabularies, listed flags are true and omitted known flags are false. The grouped name union is the complete tool inventory; use tools/list for canonical tool schemas. Provider candidate facts describe possible providers, while top-level session capabilities describe the current route and analysis_provider_binding names the selected provider. Do not treat those scopes as interchangeable.

Native startup and interpretation

Before the first Ghidra-backed query, explain that importing and analyzing a large executable can take several minutes. open_binary validates the target; it does not mean native analysis has completed. This plugin allows 900 seconds for Ghidra startup; each Hermes MCP call uses Hermes' normal tool deadline (300 seconds by default). The user can still cancel. Do not invent progress percentages or promise a fixed completion time.

Keep the target open for related queries: the first native query does the expensive import, and later queries reuse that analysis. Do not reopen the binary or start another server to "check progress". On a timeout, read connection and recovery before retrying.

Generated pseudocode is an interpretation, not original source. Corroborate important branches with assembly, references and callers. Keep inferred function names/types distinct from symbols actually present in the artifact; report type-propagation warnings and missing symbols rather than filling gaps with plausible source. When evaluating an unfamiliar target, record findings before reading matching-version source, then label source-only corrections.

Work summary-first

Start with the default result and use its inline Evidence and graph context. For a JavaScript application or ELF layout whose size is unknown or large, call analyze_javascript_application or inspect_binary_layout with "detail": "summary", then inspect the retained parent_evidence_id with inspect_analysis_view. Page module identities and select the returned node_id before requesting an item; a single file path can represent several graph roles. Do not repeat an identical analysis or rejected call without a changed input or resolved blocker. Recheck session state when a lifecycle change makes that status useful. Make a focused follow-up only when the returned result leaves a specific question unanswered.

If MCP reports resource_constraint with details.resource: "transport", call inspect_analysis_view with its reported same-session Evidence reference for a summary, one section/module, or a stable page. Use trace_application_feature once a seed is known, or call export_evidence_bundle to write the complete session to a selected path. The analysis remains complete in the ledger; avoid repeating it merely to request the same oversized response. Complete CLI JSON output also streams without a single MCP frame.

When the retained parent ID is unknown, use get_evidence_bundle with detail: "summary" and exact filters for operation, target SHA-256, analysis-profile digest or canonical procedure address. Select its returned Evidence ID for the next view instead of requesting the complete bundle. retention: "complete-record" means the whole record is stored; native value-flow truncation and unknown metadata remain separate. For an exported portable bundle, use rea inspect-evidence-bundle with JSON containing path, detail: "summary" and the same filters. Reading that file still requires validating its full contents within the local memory budget.

A client can also truncate a successful tool result without a REA transport error. Treat that preview as incomplete; use the retained parent ID for smaller views or export when the full record is needed. Do not parse the preview as canonical JSON or repeat the producer. For JavaScript, read the module-view and trace recipes before drilling into a large graph; a broad bidirectional trace can expand unrelated containment relationships even after summary analysis.

Every conclusion must distinguish observations, inferences, and unknowns. Cite Evidence IDs, preserve limitations and incomplete coverage, and never imply that static analysis observed execution. Runtime requests execute the declared target and lifecycle; do not broaden the target or action beyond those fields.

Within the user's requested investigation, call static analysis tools directly. Hermes does not prompt before REA tools run, so ask the user first, naming the exact command, process, app or device, before any call that executes or changes something outside REA's own session:

• capture_process_scenario (runs a command under a PTY) and observe_native_calls (launches the active Mach-O under LLDB);
• capture_browser_scenario, capture_electron_scenario and capture_native_ui_scenario (launch or drive an application);
• install_adb_package, uninstall_adb_package, push_adb_file, start_adb_app, start_adb_activity and stop_adb_app (change a connected Android device);
• instrument_with_frida, start_frida_session, load_frida_script, resume_frida_session, unload_frida_script and close_frida_session (attach to, spawn, inject into or resume a process; local or remote);
• start_gdb_session, gdb_console and close_gdb_session (an unrestricted GDB that can run the inferior; closing can detach and resume it);
• start_rizin_debug_session, rizin_debug_command and close_rizin_debug_session (an unrestricted Rizin debugger that runs the target);
• execute_rizin_command and cutter_command (one arbitrary Rizin command; commands and plugins can execute code, write files or reach the network);
• export_evidence_bundle or close_binary with an output path that already exists (overwrites a file).

Ask once per target and action; the read-only status tools (frida_session_status, gdb_session_status, rizin_debug_session_status) and device/process listings need no approval.

Never run an untrusted target outside an environment the user chose for it. In a scheduled or unattended run with nobody to ask, skip these calls and report the step that needs approval.

Plan broader investigations

For requests that span multiple features or subsystems, use a staged workflow:

1. Turn the request into a checklist of questions and the evidence each answer needs. Resolve target identity and constraints from the conversation and workspace before asking for information again.
2. Inspect the current REA session, artifact identity, saved analysis database, bookmarks, and prior evidence. Reuse matching state; do not open duplicate sessions or repeat identical analysis.
3. Start with the smallest useful overview or inventory. Follow each question from its entry point through relevant data and state changes to its result. Batch related operations around a specific hypothesis, then expand only when the returned evidence leaves a concrete gap.
4. Inspect relevant packaged resources and configuration alongside code when they affect the question. Use format-aware inventory and parsers; do not infer behavior from filenames, strings, or layout alone.
5. Corroborate a conclusion with the evidence type it requires. Use runtime observation only when static evidence cannot answer the question and the required host runtime and OS access are available. For JavaScript behavior, run probes against the actual app through the available browser, Electron, or process capture workflow.
6. Decompose work into independent questions. When parallel workers are available and the questions do not depend on one another, assign distinct scopes, point workers to existing evidence, and ask them to return sources, conclusions, and unresolved gaps. Otherwise, work sequentially.
7. Keep a concise finding ledger linking each conclusion to Evidence IDs, confidence/evidence type, search boundary, and remaining unknowns. Update the shared index or investigation report so later passes can reuse results.

Before finishing, revisit the original checklist. Mark each question as answered, partially answered, or unresolved based on its evidence; keep bounded negative searches bounded, and do not describe a broad investigation as complete while required questions remain open.

Read only the relevant guide

• Native binaries, managed assemblies, archives, and extraction: references/native-and-artifacts.md
• ASARs, extracted JavaScript, feature tracing, and version comparison: references/javascript-applications.md
• Android APK declarations, classes, methods, and static references: references/android-applications.md
• Browser/Electron observation (passive and scenario) and static/runtime reconciliation: references/runtime-observation.md
• Evidence paging, comparisons, residual unknowns, and verification: references/evidence-workflows.md

Finish the task

Explain findings in plain language and tie them to returned evidence. When the user asks to build something, use normal coding tools and separate observed behavior from design choices. Close an opened native session with close_binary when the investigation is complete.