Bridge implementations
REA bridge programs connect analysis flows to external tools and native operating-system services. They cover Android, native analysis, process ownership, packet capture, debugging integrations, and managed agent environments. Each bridge turns a platform-specific action into a bounded observation while preserving target identity, ownership, limits, and evidence for the systems above it.
This page is the implementation view of native runtime and bridges. The Ghidra integration and debugger and disassembler integrations pages describe the larger provider workflows.
Bridge responsibilities
The bridge family covers five related jobs. Some bridges query an analysis tool. Others inspect an existing artifact, supervise a live process, or keep a managed computer available for a caller. Their shared aim is to turn a tool-specific or platform-specific action into a bounded REA result.
| Bridge group | System aim | Main result |
|---|---|---|
| Android and managed-code bridges | Expose package, class, method, and reference facts from JADX or native managed-code analysis | Static metadata and derived analysis evidence |
| Native analysis bridges | Connect Ghidra, Hopper, Cutter, LLDB, and GDB to native inspection flows | Native image, function, call, instruction, and runtime observations |
| Process bridges | Identify and supervise processes across host boundaries | Stable ownership and cleanup decisions |
| Capture bridges | Read existing mitmproxy or packet recordings without acquiring traffic | Bounded, redacted artifact evidence |
| Agent environment bridges | Deploy a managed computer, run a bundle task, capture a desktop application, and clean up local sandboxes | Connection data, solve facts, capture state, and teardown status |
The bridge is not the product-facing API. The provider or service above it decides which capability is available and maps the result into the shared evidence model. This keeps tool limits and tool-specific gaps visible.
References: bridge/…/ReaJadxBridge.java• L47-65 bridge/…/ReaGhidraBridge.java• L154-187 bridge/…/rea_lldb_tracer.py• L1-9 bridge/…/capture.py• L1-5 bridge/…/solve.py• L1-9
Agent environment bridge
The agent environment bridge starts one bundle task on a local managed computer and exposes the connection data needed by the caller. The computer helper writes one JSON protocol line with the computer address, MCP URL, control token, and sandbox ID, then keeps the environment alive until its input closes or it receives a stop signal. Progress and build output stay on stderr so they cannot corrupt the control response. (bridge/…/computer.py• L1-9, bridge/…/computer.py• L20-25, bridge/…/computer.py• L53-93.)
Its capture controller can start a desktop-application capture, wait for named checkpoints, stop the capture, run a REA command, and export a bounded file inventory. Capture control uses a private request channel and progress log, so start, checkpoint, and stop commands can coordinate with a long-running capture without closing its input. Long-running REA commands use a private job folder, an owner token, and a supervisor that can cancel the child process. Startup and command failures remain explicit.
The file inventory reports regular files with relative paths, sizes, and SHA-256 digests. It does not follow directory symlinks. It records skipped links and other non-regular entries so a caller does not mistake an incomplete walk for a complete bundle view. A missing root fails instead of producing an empty successful inventory.
The solve bridge runs one bundle task in a fresh local agent environment and returns the selected verifier's outcome. It keeps checks from the requested task separate from checks for other tasks, reports trajectory and package-version data when available, and includes teardown state so a caller can distinguish a completed solve from an unconfirmed cleanup.
The cleanup helper is a second ownership boundary for stopped solves. It terminates sandboxes listed by agentenv, checks Docker resources and marked host processes, removes only a confirmed work folder, and leaves residue in place when cleanup cannot be verified.
This diagram shows the bridge lifecycle from bundle deployment or capture start to controlled observation and cleanup.
sequenceDiagram
participant Host as REA host
participant Bridge as Agent environment bridge
participant Computer as Managed computer
Host->>Bridge: Bundle task or capture command
Bridge->>Computer: Deploy task or start capture
Bridge-->>Host: Connection data, progress, or error
Host-->>Bridge: Checkpoint, stop, poll, or inventory request
Bridge->>Computer: Read bounded result and control state
Host-->>Bridge: Close, cancel, or stop signal
Bridge->>Computer: Tear down task or capture
Bridge-->>Host: Exit with cleanup status
References: bridge/…/computer.py• L1-9 bridge/…/computer.py• L20-93 bridge/…/test_computer.py• L85-175 bridge/…/capture.py• L1-13 bridge/…/capture.py• L166-290 bridge/…/capture.py• L386-471 bridge/…/capture.py• L474-509 bridge/…/solve.py• L1-9 bridge/…/solve.py• L60-127 bridge/…/solve_cleanup.py• L21-59 bridge/…/test_capture.py• L15-60 bridge/…/test_solve.py• L14-68 bridge/…/test_solve_cleanup.py• L52-104
Android and managed-code bridges
The Android bridge gives the static Android workflow a small metadata boundary over the selected JADX engine. It loads an APK into the retained analysis session and exposes package, manifest, class, method, and reference facts without running the application.
The bridge can inspect the APK's standard root DEX files directly for native method declarations. It also reports whether the root manifest requires split APKs, which can limit what the selected package contains. Class and method queries use exact identities and an explicit overload when needed. Java text is preferred for method bodies, with a selected smali block as a fallback when Java output cannot establish the requested body.
Reference results preserve direct DEX invoke sites and distinguish APK-defined targets from external prototypes when the source can be inspected. Load failures include their observed kind, while response budgets and partial coverage remain visible. Callers can therefore distinguish complete evidence from evidence limited by output space, memory, cancellation, or analysis failure.
This diagram shows how an Android package becomes bounded static evidence.
flowchart LR
APK[APK] --> JADX[JADX session]
JADX --> Facts[Manifest, class, method, and reference facts]
Facts --> Evidence[Bounded provider evidence]
References: bridge/…/ReaJadxBridge.java• L47-120 bridge/…/ReaJadxBridge.java• L207-289 bridge/…/ReaJadxBridge.java• L292-369 bridge/…/ReaJadxBridge.java• L373-469 bridge/…/ReaJadxBridge.java• L471-557 bridge/…/ReaJadxBridge.java• L559-635 bridge/…/ReaJadxBridge.java• L684-740 bridge/…/ReaJadxBridge.java• L866-963
Native analysis bridges
Native analysis bridges let REA use established reverse-engineering tools while keeping request shape and evidence meaning under REA control. They support read-only inspection and selected annotations when the provider and transport allow them.
Ghidra bridge
The Ghidra bridge is a headless session boundary for native image analysis. It admits a matching imported program and serves bounded requests for image layout, bytes, symbols, procedures, strings, references, decompilation, instruction facts, data types, and function analysis.
The session reports analysis state, extensions, profile data, and capability limits. Before analysis is served, architecture-specific preparation can establish the correct entry context for DOS COM images or install a committed calling-convention policy for supported MIPS and RISC-V images. These preparation results are published only after the full update and self-check succeed. Unix sessions can admit selected database edits. The supported Windows transport remains read-only. The bridge publishes only the capabilities that the current session can safely provide.
Function analysis combines listing facts with decompiler-derived views. These are analysis observations rather than source or runtime proof. Large functions can return partial value-flow data with explicit omission information, and failed decompilation does not erase listing facts that remain available.
Address and file-offset selection stay tied to the digest-matched target source. Architecture policy and special image preparation publish their status only after their full update succeeds. This keeps calling conventions, DOS COM entry context, and raw-image entry context tied to the admitted target.
See the related Ghidra integration page for the provider workflow and analysis features.
References: bridge/…/ReaGhidraBridge.java• L128-287 bridge/…/ReaGhidraBridge.java• L290-536 bridge/…/ReaGhidraBridge.java• L568-629 bridge/…/ReaGhidraMipsSoftFloat.java• L36-101 bridge/…/ReaGhidraPrepareCom.java• L8-35 bridge/…/ReaGhidraPrepareCom.java• L37-76 bridge/…/ReaGhidraRiscvCallingConvention.java• L37-126
NativeAOT bridge
The NativeAOT bridge extends native image analysis for supported .NET NativeAOT images. It finds the runtime directory, rehydrates metadata into derived memory, recovers method tables and type links, and reports which frozen objects received committed annotations.
The bridge keeps recovery changes inside one analysis transaction. It publishes the report only after discovery, metadata recovery, method-table analysis, calling-convention handling, and coverage checks finish. Derived memory is identified by its address, size, and digest, but it has no original file offset. Generated type identities and prototypes remain analysis inferences.
References: bridge/…/NativeAotExtension.java• L22-40 bridge/…/NativeAotExtension.java• L41-113 bridge/…/NativeAotExtension.java• L116-153
Hopper and Cutter bridges
The Hopper and Cutter bridges connect REA to an already running or REA-launched desktop disassembler. Each bridge binds requests to one target document and keeps tool API calls on the tool-required thread. Hopper selects the document whose executable or database path matches the target. Cutter tracks the active file with a document generation, so a display name or a later document change cannot silently change the session target.
The bridges provide bounded inventories, function views, searches, bytes, references, decompilation, and annotations where the tool exposes them. They refuse analysis-dependent operations while background analysis is active. Their results preserve incomplete coverage and unavailable capabilities when the public API cannot provide a full fact. Annotation changes invalidate derived views before later reads use them.
Their local transports use per-session capability checks and private endpoint publication. Hopper keeps the authenticated session on its dedicated Python thread. Cutter uses authenticated loopback IPC and reports whether a command completed, failed, or may have produced effects before a timeout. A command with unknown completion is not safe to retry automatically. Shutdown distinguishes a retained shared document from a document that REA closed.
Hopper and Cutter share the REA_BRIDGE_MAX_REQUEST_BYTES request ceiling. The default is 16 MiB. Values outside the admitted range are rejected rather than replaced with a fallback. This keeps the sender, bridge, and frame parser on one payload contract. Large replies fall back to bounded status or completion information instead of hiding whether a command ran.
References: bridge/hopper_bridge.py• L1-6 bridge/hopper_bridge.py• L17-31 bridge/hopper_bridge.py• L136-217 bridge/hopper_bridge.py• L982-1028 bridge/hopper_bridge.py• L1302-1471 bridge/…/rea_bridge.py• L1-6 bridge/…/rea_bridge.py• L37-67 bridge/…/rea_bridge.py• L461-609 bridge/…/rea_bridge.py• L612-845 bridge/…/rea_bridge.py• L879-1040 bridge/…/rea_bridge.py• L1095-1119
LLDB bridge
The LLDB bridge observes selected live native calls without evaluating expressions or running target code on behalf of the tracer. It launches one owned process, stops at requested functions or Objective-C methods, reads registers and limited memory, records locations and optional backtraces, drains target output, and ends the process when the request or a bound is complete.
The launch is held at entry while the host confirms process identity. The tracer checks the selected file before and after launch and compares module file identity and UUID observations. It publishes a stable identity observation, but it does not claim that LLDB proved the exact bytes mapped into memory.
Events, other stops, target output, and an optional observation journal use separate bounded records. The journal flushes each accepted JSON line so earlier observations can survive a later tracer failure. A failed launch reports whether a process was created and whether its exit was observed, so cleanup does not mistake a failed kill or an invalid process for a normal exit.
This diagram shows the ownership and observation boundary around one trace.
sequenceDiagram
participant Host as REA host
participant LLDB as LLDB tracer
participant Target as Owned target process
Host->>LLDB: Configured executable and breakpoints
LLDB->>LLDB: Check file and module identity
LLDB->>Target: Launch and stop at entry
LLDB-->>Host: PID for identity acknowledgement
Host-->>LLDB: Identity readable or unavailable
LLDB->>Target: Continue and observe bounded events
LLDB-->>Host: Events, output, stops, and cleanup state
See the related debugger and disassembler integrations page for the broader live-debugger model.
References: bridge/…/rea_lldb_tracer.py• L1-9 bridge/…/rea_lldb_tracer.py• L64-79 bridge/…/rea_lldb_tracer.py• L121-232 bridge/…/rea_lldb_tracer.py• L603-747 bridge/…/rea_lldb_tracer.py• L748-925 bridge/…/test_rea_lldb_tracer.py• L27-367 bridge/…/test_rea_lldb_tracer.py• L384-610
Native UI bridge
The Swift native UI bridge observes one selected macOS application window and performs a small set of actions on one selected accessibility element. It checks the running executable, launch time, window owner, window geometry, and executable digest before it reads the accessibility tree or captures a screenshot.
Window selection uses the complete window list, so an off-screen window remains a candidate. The bridge still rejects a missing window, a foreign owner, or a non-window-layer match. It returns a bounded tree with element paths, roles, values, actions, child counts, and gaps where the host API could not read a child list.
JSON-inexact accessibility numbers and unsupported value types become explicit gaps instead of false scalar values. The bridge also distinguishes an absent attribute from a failed read and an unavailable action list from a successful empty list. A click, scroll, or text entry acts on a validated element path and checks that the element still belongs to the selected process.
References: bridge/…/ReaNativeUI.swift• L7-46 bridge/…/ReaNativeUI.swift• L48-134 bridge/…/ReaNativeUI.swift• L135-189 bridge/…/NativeUIValues.swift• L4-39 bridge/…/NativeUIChildren.swift• L3-26
GDB and pwndbg bridge
The pwndbg bridge starts GDB in batch mode to inspect a recorded core context. It disables automatic loading and network-assisted symbol behavior, loads only the selected core snapshot, and returns bounded debugger observations through a fixed Python entry point.
Before it asks pwndbg for mappings, it validates the recorded architecture and confirms that the inferior is a core connection with no selected executable. The historical PID remains recorded metadata and is never used to find or control a live process. A setup guard failure produces no mapping query.
The sibling pwntools adapters inspect copies supplied by REA and classify dependency, format, resource, and output failures without launching a target process. The layout adapter reports structural ELF facts for a fixed set of architecture and byte-order profiles. It preserves source coordinates where they can be proved. Runtime load state and incomplete inventories remain unknown. The shared decoder runtime applies owned CPU, address-space, file-size, cache, memory, and reply limits.
References: bridge/…/core_context.py• L1-65 bridge/…/core_context.py• L68-114 bridge/…/test_core_context.py• L16-124 bridge/…/layout_impl.py• L1-28 bridge/…/layout_impl.py• L110-134 bridge/…/layout_impl.py• L188-223 bridge/…/layout_impl.py• L324-350 bridge/…/test_layout.py• L20-115 bridge/…/recorded_crash_impl.py• L1-19 bridge/…/recorded_crash_impl.py• L37-63 bridge/…/recorded_crash_impl.py• L125-163 bridge/…/decoder_runtime.py• L1-62
Process ownership bridges
Process bridges answer a safety question before REA reads or controls a process: is the observed PID still the process selected by the request? They make process ownership explicit so PID reuse or replacement does not turn an old identity into a new target.
The Swift process helper reads Darwin process snapshots and run-token data. A snapshot supplies process identity, credentials, lifecycle, parent and process-group relations, and a display name. The helper compares process identity before and after token reads. It reports readable or unavailable observations with a reason when the environment is hidden, malformed, ambiguous, or changes during the read.
Code-signing restrictions are applied at the native boundary even when process arguments are incomplete or malformed. A readable non-empty environment without REA's token remains a readable non-match. The helper feeds the wider ownership and cleanup system. It does not itself kill or adopt a process.
This diagram shows how the helper checks identity around a process observation.
sequenceDiagram
participant Host as REA host
participant Helper as Swift process helper
participant OS as Darwin process services
Host->>Helper: PIDs and mode
Helper->>OS: Snapshot process identities
Helper->>OS: Read token or identity data
Helper->>OS: Snapshot identities again
Helper-->>Host: Readable or unavailable observations
References: bridge/…/ProcessRunTokenReader.swift• L4-42 bridge/…/ProcessRunTokenReader.swift• L50-108 bridge/…/ProcessRunTokenReader.swift• L117-176 bridge/…/ProcessRunTokenReader.swift• L225-334 bridge/…/ProcessRunTokenReader.swift• L336-441 bridge/…/ProcessRunTokenReaderMain.swift• L11-32
Capture bridges
Capture bridges inspect recordings through the native library that created them, then project the result into safe evidence. They do not open a network interface, start a proxy listener, send target requests, or treat decoded content as proof of live behavior.
The mitmproxy bridge reads native tnetstring flow records from a private snapshot. It keeps producer bytes and display values separate, removes transport credentials and caller-supplied sensitive values, and reports record locations and redaction facts. URL user information is removed only when the recorded value has a valid authority under the same URL rules used by the provider. Rejected source text remains unchanged.
Packet capture inspection follows the same boundary at the provider layer. The existing TShark process reads a private capture copy under host resource limits. The packet bridge admits caller-selected TShark and capinfos tools only after version checks, so results stay tied to the selected capture and tool identity.
References: bridge/…/capture.py• L1-43 bridge/…/capture.py• L45-203 bridge/…/capture.py• L206-322 bridge/…/capture.py• L324-386 bridge/…/test_capture.py• L46-263 src/…/PcapProvider.ts• L219-263 src/…/PcapProvider.ts• L277-289
Bridge lifecycle and trust boundary
Every bridge sits between REA-owned control and an external runtime. The surrounding provider supplies private request and result paths, selects the executable or tool, checks identity before and after use, and joins cleanup before reporting completion.
The bridge applies rules that the external runtime does not know. These rules include fixed protocol shapes, authentication, target binding, output limits, resource limits, cancellation, and explicit unavailable states. A successful response means that the bridge observed and checked a bounded result. It does not mean that the external tool is a full sandbox or that an inferred fact is runtime truth.
| Boundary concern | Bridge behavior |
|---|---|
| Target selection | Bind analysis sessions and live processes to a path, digest, document, PID identity, or session token |
| Authentication | Use private descriptors, bearer tokens, local sockets, or authenticated loopback frames where a separate tool process is involved |
| Resource control | Limit input, output, time, memory, trace size, frame size, or retained events before unbounded data reaches the host |
| Partial results | Keep completed observations and label truncation, unknown coverage, or unavailable capabilities |
| Cleanup | Stop owned processes and remove private endpoints; preserve failure ownership when cleanup cannot be confirmed |
References: bridge/…/computer.py• L20-25 bridge/…/computer.py• L67-93 bridge/…/capture.py• L386-471 bridge/…/solve_cleanup.py• L21-59 bridge/…/ReaGhidraBridge.java• L198-269 bridge/…/ReaGhidraBridge.java• L271-449 bridge/hopper_bridge.py• L178-217 bridge/hopper_bridge.py• L1302-1471 bridge/…/rea_bridge.py• L612-845
Change checklist
When these files or sections are changed, remember to consider these:
- Keep bridge capability names aligned with provider schemas and each tool's advertised capability set.
- Keep target identity checks aligned with owner process and cleanup paths. A new launch path must not create an unowned child.
- Keep partial, truncated, unavailable, and failed outcomes distinct when changing result fields or frame limits.
- Keep authentication descriptors and endpoint publication atomic when changing local socket or loopback transports.
- Keep external-tool version and extension ABI checks aligned with the packaged bridge code.
- Keep
REA_BRIDGE_MAX_REQUEST_BYTESvalidation, its allowed range, and client and bridge framing rules aligned across Hopper, Cutter, and callers. - Keep Cutter document-generation checks, completion receipts, timeout state, and bounded recovery replies aligned when changing command dispatch or response encoding.
- Keep Hopper target-document binding, dedicated-thread dispatch, analysis-complete gates, cache invalidation, and shutdown ownership aligned when changing tool operations.
- Keep the agent computer's single stdout protocol line separate from stderr progress and keep EOF and signal paths joined to teardown and exit status.
- Keep Android load-state reporting, exact class and overload identity, root manifest selection, native declaration coverage, and reference output limits aligned when changing JADX tools.
- Keep NativeAOT source revision reporting, supported architecture admission, recovery transaction, derived-memory digest, coverage counts, and partial diagnostics aligned when changing metadata recovery.
- Keep Ghidra profile declarations, DOS COM preparation, architecture-specific calling-convention reports, prototype storage checks, and session capability data aligned when changing architecture policy.
- Keep LLDB PID publication, identity acknowledgement, journal limits, and final process retirement aligned when changing launch or trace failure handling.
- Keep window selection, off-screen window handling, permission checks, and selected-window identity aligned when changing native UI capture.
- Keep JSON number gaps distinct from valid accessibility values when changing the native UI value projection.
- Keep Darwin process snapshots, run-token parsing, code-signing availability reasons, and before-and-after identity checks aligned when changing process ownership observations.
- Keep pwntools package profiles, admitted ELF architecture profiles, source-coordinate meaning, artifact-only scope, effective resource limits, and output-limit replies aligned across layout and core decoders.
- Keep pwndbg core setup order, validated architecture selection, executable and symbol discard, and mapping-query guards aligned when changing recorded-core inspection.
- Keep solve outcome selection tied to the requested verifier and keep teardown reports separate from task score and trajectory data.
- Keep sandbox cleanup ownership, Docker and host-process checks, work-folder removal, and residue reporting aligned when changing stopped-solve recovery.
- Keep capture control acknowledgements, progress offsets, exit status, and filesystem inventory treatment of symlinks and missing roots aligned when changing agent environment capture.
- Keep mitmproxy URL parsing, source-byte projection, sensitive-value redaction, and native parser bounds aligned when changing recorded-flow projection.
Important failure modes
| Trigger | Consequence | Guard |
|---|---|---|
| Bundle task fails or deploys no computer | The caller receives connection data that cannot reach a managed environment | Return a startup error and keep teardown in the same lifecycle |
| The computer helper writes progress on its protocol stream | The caller cannot parse the connection record | Reserve stdout for one flushed JSON line and send progress to stderr |
| EOF or a stop signal arrives while the computer is running | The managed environment can remain up or the helper can exit before cleanup | Join both stop paths to teardown and return a distinct status when resources remain |
| A long-running command loses its owner or is cancelled before start | A late start can run without a caller that can safely control it | Claim the job folder with an owner token before launch and require the same token for poll and cancel |
| Computer or sandbox teardown leaves a resource up | The bridge reports success while an owned environment remains | Inspect teardown results, Docker resources, and marked host processes before removing the work folder |
| Capture folder is already used or a stop claims it during startup | Two capture lifecycles can share state or start the wrong process | Claim the folder before launch and reject an existing capture |
| Capture inventory root is missing or contains a directory symlink | The caller can mistake an empty or incomplete list for a valid bundle inventory | Fail a missing root and disclose skipped links without following them |
| Selected APK, native image, or Ghidra program changes during admission | Results can describe a different artifact | Digest and session identity checks before serving analysis |
| Android load exhausts the worker JVM or native declaration output reaches its bound | Static metadata is unavailable or later declarations become unknown | Return a typed load failure and explicit partial coverage instead of a successful complete result |
| NativeAOT recovery fails after database changes begin | The analysis database can expose partial metadata as committed evidence | Keep recovery and coverage audit in one transaction and publish complete or partial status only after the audit |
| Ghidra architecture preparation or calling-convention self-check fails | The session can report native facts under the wrong entry context or storage policy | Refuse the session until the preparation report is written after all checks pass |
| Recorded core uses an unsupported architecture or leaves an executable selected | Mapping results could describe the debugger host or a live target instead of the snapshot | Validate the decoder profile and core-only state before querying pwndbg |
| Core setup guard fails or a mapping reply exceeds its output budget | The caller receives no safe mapping result or an incomplete reply | Query mappings only after setup succeeds and return an explicit unsupported or output-limit result |
| Hopper or Cutter document changes during a command | A query or edit can target the wrong document | Bind the session to the target document and keep tool calls on the required thread |
| Hopper or Cutter uses an invalid or different request ceiling | The bridge can reject valid work or use the wrong framing contract | Validate the setting before startup and confirm the live bridge ceiling |
| Cutter UI dispatch times out or a command raises after tool work begins | The caller cannot know whether the command had effects and may corrupt later assumptions by retrying | Mark completion as unknown, poison the session after a UI timeout, and preserve a receipt for explicit acknowledgement |
| LLDB target identity changes during launch or launch ownership is unclear | Runtime events no longer belong to the selected executable, or cleanup can kill the wrong process | Check file and module identity before and after launch and report process creation and exit separately |
| LLDB trace, output, journal, or breakpoint locations exceed a bound | The caller cannot receive a complete result | Stop at the relevant limit and mark truncation or resource failure |
| Process PID is reused during token reading | Ownership can be assigned to a replacement process | Compare process identity before and after the read |
| Selected UI window is off screen or its value cannot be represented faithfully | A valid window can be reported as absent, or a value can become a misleading scalar | Select from the full window list and return explicit representation gaps |
| Native capture declares invalid or oversized data | The parser can fail or allocate beyond the intended budget | Use bounded readers and owned resource limits |
| Mitmproxy URL authority is invalid or ambiguous | Credentials can be removed from text that is not a valid authority, or unsafe text can be misclassified | Apply the provider URL boundary and retain rejected source text |
| Pwntools dependency, ELF/core format, or reply exceeds its profile | Static artifact evidence is unavailable or incomplete | Pin package versions, validate format ranges, apply resource limits, and return explicit failure kinds |
| Bridge output exceeds its frame or evidence budget | The caller cannot receive a complete result | Reject or return an explicit truncated or output-limit result |
Key files
| File | Why to read it |
|---|---|
| bridge/…/computer.py | Shows managed computer startup, the single-line protocol, signal and stdin ownership, and teardown status |
| bridge/…/test_computer.py | Shows the helper contract for EOF, signals, protocol readiness, and retained teardown resources without using a live computer |
| bridge/…/capture.py | Shows desktop capture control, long-running command ownership, cancellation, and bundle file inventory boundaries |
| bridge/…/solve.py | Shows one-task solve execution, selected-verifier outcome projection, trajectory facts, and cleanup reporting |
| bridge/…/solve_cleanup.py | Shows sandbox termination checks and safe removal of stopped-solve work folders |
| bridge/…/test_solve.py | Shows that outcome and checks come only from the requested verifier |
| bridge/…/test_solve_cleanup.py | Shows retry and residue rules for incomplete sandbox cleanup |
| bridge/…/ReaJadxBridge.java | Shows the Android metadata boundary, manifest and split-package projection, root DEX declaration inventory, exact overload identity, and bounded references |
| bridge/…/ReaGhidraBridge.java | Shows the main native analysis session, target admission, capabilities, architecture reports, function evidence, and observation limits |
| bridge/…/ReaGhidraMipsSoftFloat.java | Shows the MIPS soft-float calling-convention policy and its storage checks |
| bridge/…/ReaGhidraPrepareCom.java | Shows DOS COM entry and segment-context preparation before analysis |
| bridge/…/ReaGhidraRiscvCallingConvention.java | Shows the supported RISC-V calling-convention policy and storage checks |
| bridge/…/NativeAotExtension.java | Shows NativeAOT metadata recovery, derived-memory evidence, transaction status, and coverage reporting |
| bridge/hopper_bridge.py | Shows dedicated-thread tool access, target-document binding, authentication, analysis gates, request ceilings, and shutdown ownership |
| bridge/…/rea_bridge.py | Shows authenticated desktop-tool IPC, document generations, command completion state, response limits, and uncertain command completion |
| bridge/…/rea_lldb_tracer.py | Shows live call observation, launch identity checks, output capture, journaling, source locations, and bounded retention |
| bridge/…/ReaNativeUI.swift | Shows macOS window identity, off-screen window selection, accessibility actions, bounded tree capture, and screenshots |
| bridge/…/NativeUIValues.swift | Shows how accessibility scalar values and JSON representation gaps are reported |
| bridge/…/ProcessRunTokenReader.swift | Shows Darwin process identity, relation data, code-signing environment limits, run-token observations, and replacement detection |
| bridge/…/core_context.py | Shows the fixed core-only debugger boundary and recorded mapping observations |
| bridge/…/layout_impl.py | Shows offline ELF layout evidence, admitted architecture profiles, and structural limits |
| bridge/…/decoder_runtime.py | Shows shared Python resource limits and bounded reply handling |
| bridge/…/capture.py | Shows offline native-record decoding, URL authority handling, and sensitive-value projection |