Select exact declaring types before managed signatures and CIL bodies are decoded, preserving caller choices in CLI/MCP Evidence. Keep unmatched names and selected coverage explicit; comparisons and graphs retain unknowns outside their observed scope. Reuse existing metadata ownership/accounting and advertise unique selections through the actual SDK Schema.
Validated with 229 focused tests across 21 files, 11 managed conformance controls, fullstatic/docs checks and real Linux CLI and fresh MCP inspection of two unchanged DLLs. All 118 selected CIL hashes matched original PE-mapped bytes; result/caller-argument parity and unmatched controls passed. Full artifact/native-boundary metadata is complete; whole-assembly members retain the unchanged 64 MiB resource constraint. No target code executes.
Refs #1878
Co-authored-by: skared <[email protected]>
Co-authored-by: Claude Opus 5.5 <[email protected]>
Co-authored-by: Ysopwnable <[email protected]>
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Co-authored-by: Kaoru0822 <[email protected]>
Group repeated session tool availability and provider capability facts, preserve complete named contracts and ordered limitations, and remove duplicate catalog descriptors. Update consumers and migration guidance.
Validated by all 31 CI jobs, focused source checks, and actual Claude Code/Linux Hopper/native Windows Ghidra small-target workflows on the current integration tree.
Refs #723.
BREAKING CHANGE: binary_session no longer includes server_identity.catalog.tools; tool_availability and provider capabilities use grouped facts documented in docs/mcp-contracts.md.
Co-authored-by: Kaoru0822 <[email protected]>
Give atomic Evidence and snapshot writes an explicit resource owner, preserve primary failures and completed publication receipts, and retry retained staging cleanup at the owning lifecycle. Verify inode identity before publishing or deleting; unknown path disappearance remains a cleanup failure. Keep snapshot admission reusable when a session closes and reopens.
Consolidate JavaScript source-scope classification so cancellation cannot fabricate failures for unselected files. Classify configuration failures as invalid input. Remove stale assertions and reuse shared test cleanup ownership.
Validation: 266 focused tests before final ownership tightening, 108 affected tests afterward, static and documentation checks, actual invalid-configuration CLI behavior, and isolated test temporary-directory hygiene.
Refs #1968
Remove repeated MCP input parsing after the SDK has validated named contracts. Reject unmodeled fixture operations and locate prompt contracts by name so unrelated catalog ordering cannot select a contract.
Validated focused browser, Electron, runtime-observation and MCP workflow tests, advertised schemas, packaged MCP and static checks. Application-boundary validation remains in place.
Separate observed relationship coverage from general provider caveats. Authenticate and replay coverage without changing legacy Evidence IDs; expose unknown and capped relationship sets through CLI/MCP graphs and traces.
Validated focused domain, snapshot, IDA/Ghidra boundary, MCP, schema, package and static checks. Provider fixtures establish contract behavior; licensed engine behavior was not reverified.
Serialize resolution with shutdown, retain the owned browser after close failures, and retry cleanup before another launch. Wire cleanup through the shared service and CLI/MCP lifecycles, preserving primary failures and completed native reports alongside cleanup diagnostics.
Separate the browser resolution workflow from lifecycle ownership. Validation: seven resolver source tests, eleven application tests, MCP shutdown retry, public CLI parity, and full static/architecture checks passed. Real Chromium behavior remains unverified.
Supply schema-validated inspect_analysis_view arguments after completed native function Evidence is recorded and exceeds MCP delivery capacity. Preserve the full canonical result and its provenance. Verify exact source paging, export and following-call usability through real Ghidra stdio MCP.
Fixes#1849
Co-authored-by: Minghang Chamling <[email protected]>
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Co-authored-by: Kaoru0822 <[email protected]>
* fix: order remaining outputs independently of host locale
String.prototype.localeCompare without a pinned collation follows the
host's ICU default locale, so the same input could be ordered differently
on different machines. 614fb768 moved canonical domain projections to
compareUnicodeCodePoints; provider, application, catalog, and setup
orderings still used locale collation.
Order every remaining source site by Unicode code point, including Mach-O
symbols, firmware extraction paths, ASAR entries (previously pinned to
"en" collation), evidence ID sets, browser and Electron target, script,
and resource inventories, reconstruction obligations, capability and
provider listings, and setup environment details.
The current tool and prompt catalog names order identically under both
comparators, so the catalog identity digest is unchanged.
* fix(artifacts): order ASAR entries by their portable path
@electron/asar lists host-spelled paths, so code-point order over the raw
spelling still differed by host: "a\b" sorts after "a0" on Windows while
"a/b" sorts before it on POSIX. Pair each listing with its portable
archive path once, order by that path, and keep the host spelling for
@electron/asar calls and adapter keys.
Focused and zero-match search_android_classes responses embedded every
examined class name: JadxSession records each raw list_classes page into
observed calls, those calls ride into Evidence raw_result on the
registration path, and the complete inline Evidence is serialized into
every tool response (issue #1850: 382,435 bytes for one match).
When the server session retains Evidence, record the complete raw-call
record first, then deliver a summary Evidence that keeps the pagination
audit (totals, offsets, page sizes, elided item counts), the unchanged
normalized matches and truthful coverage metadata, and links the
retained complete inventory by Evidence ID. Sessions without retention
keep the complete inline delivery, so complete CLI JSON output is
unchanged and no provenance is discarded.
🤖 Generated with AI assistance (disclosure voluntary per repo policy)
* feat(native): admit raw binaries with caller-supplied architecture and entry
* fix(native): honor raw-binary preparation and entry identity
* fix(ghidra): stream raw snapshot attestation with cancellation
* fix(native): preserve and validate raw lifecycle entries
---------
Co-authored-by: morluto <[email protected]>
Forward Ghidra startup waiting observations through the existing progress reporter. Count emitted MCP status notifications and omit the unknown total to keep progress strictly increasing.
Verified with focused tests, type/lint/docs checks, packaging and actual installed Windows timeout/recovery workflows. Native first-query speed and large-target memory remain open under #1065.
Co-authored-by: Kaoru0822 <[email protected]>
* feat(harmony): project HarmonyOS package inventory evidence
First HarmonyOS increment from #1431: HAP, HSP, and App Pack roots are
classified through the hardened ZIP reader and projected into an
execution-free application graph with packaging-model hints, exact
component paths and hashes, App Pack children by path, runtime-family
hints, and path-based N-API bridge hypotheses. HarmonyOS .har stays
unclassified by suffix because .har also names HTTP Archive JSON.
Adds the project_harmony_application_graph tool (CLI and MCP), synthetic
composition coverage, a pinned real-package verify lane, and docs.
* refactor(domain): share application graph projection scaffolding
The HarmonyOS projection mirrors the Android and Apple projections by
design, which pushed jscpd duplication over its 2.5 percent budget.
Extract the identical pieces (inventory input schema, projected
component schema and occurrence flattening, bridge candidate schema
factory, coverage assembly, comparator, canonical digest) into
domain/mobileApplicationGraphProjection.ts and route all three
projections plus the composition fixtures through it. Duplication drops
from 2.52 to 2.49 percent with no behavior change: projection ids,
classifications, and limitations are byte-identical.
* fix(harmony): address review, stable app-pack roots, and rotated fixture
- rebased onto main; the shared projection scaffolding now keeps main's
nativeRuntimeConvention loop and the harmony runtime families adopt the
same react-native/flutter/unity path conventions
- the .app App Pack suffix is now a root-input rule only (verified ZIP
magic in classifyRootFormat and BinaryTargetResolver); nested .app ZIP
members inside other packages keep the generic zip classification, with
a regression test, so existing inventories do not change
- the signing limitation now states that an empty signing list does not
mean unsigned because packaging-tool signatures live in the ZIP signing
block, and the app-pack root rule is disclosed as suffix-based
- repinned the real-package fixture to the current VHome 2.6.14-beta asset
after the upstream deleted v5.0.111-beta and replaced the v2.6.14-beta
asset within hours; the pin records both observed digests
---------
Co-authored-by: morluto <[email protected]>
Use the canonical current and supported legacy process-provider rule for authenticated capture Evidence at both CLI and MCP admission boundaries. Preserve successful Evidence and distinguish provider input errors from digest authentication failures.
Closes#1757.
Stream legacy snapshot cache rows independently, authenticate complete Evidence once, and restore exact immutable payloads without assembling duplicate full graphs. Preserve the existing file format, provenance, resource admission, cancellation and cache export isolation.
Verified through real Linux SDK MCP import of the original 83 MB Ghidra snapshot and full 41 MB bundle hash, actual Windows CLI/MCP PE recovery and the installed Windows large-file decoder boundary, focused tests and static checks.
Fixes#1680. Refs #1065.
Seal file-backed MCP imports at the session boundary while preserving mutable
ordinary file reads. Validate internally constructed ledger candidates directly
with the existing immutable bundle constructor.
Assisted-by: Suika (gensokyo-skills:suika-scatter-gather-planning)
* fix(build): emit no runtime imports for type-only bindings
Under verbatimModuleSyntax, an import whose specifiers all carry inline
`type` modifiers compiles to `import {} from "..."`, a side-effect import
that still resolves and evaluates the module. 78 compiled modules carried
one, including `import {} from "playwright-core"` in the browser scenario
session, which loaded Playwright's module graph at MCP startup.
Use top-level `import type` for these bindings and enable oxlint's
typescript/no-import-type-side-effects rule so the pattern cannot return.
Refs #1685
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* perf(mcp): load Playwright when a scenario launches, not at startup
The browser scenario and Electron active providers are constructed before
the stdio server serves, and both bound Playwright's launchers at module
load. Resolve `chromium` and `_electron` from playwright-core when a
capture opens instead, as the web module resolver already does.
MCP startup now resolves 1,312 modules (4,902 resolutions) instead of
1,598 (6,076) before reading the first request, and no longer loads
playwright-core. Provider construction, the tool catalog, and capability
inventory are unchanged.
A boundary test spawns the stdio server, initializes, lists tools, and
asserts playwright-core and isomorphic-git were not resolved; the CLI
probe now exercises the real launch path to prove detection.
Closes#1685
Co-Authored-By: Claude Opus 5.5 <[email protected]>
---------
Co-authored-by: Claude Opus 5.5 <[email protected]>
* feat(evidence): discover retained records without full payloads
* fix(evidence): guide oversized results to retained discovery
* style(tests): format retained discovery recovery assertion
Require an ArtifactResourceScope for web script exports and share its lifetime with CLI or MCP shutdown. Retain setup-failed output trees in firmware and Wakaru providers. Preserve completed publications when descriptor cleanup fails, returning the full result after a successful retry or as a partial observation while cleanup remains unresolved.
Keep failed reader, root descriptor, and output-tree cleanup with the caller. Drain admitted work before close, retry the same retained owners, and block new resource work while cleanup is incomplete. Preserve primary failures and complete inventory observations.
Thread ownership through JavaScript reconstruction, the CLI's one-shot scope, MCP server shutdown, verifier scripts, and test fixtures. Remove redundant ownership flags and the unused validated test wrapper.
Validate with 850 passing focused tests across 90 suites (two skipped), targeted classification and cleanup retries, static and unused-code checks, and real macOS CLI/MCP directory, ZIP, and DMG workflows. Native Windows and deep-analysis providers were not verified.
BREAKING CHANGE: Direct artifact inventory/extraction and JavaScript analysis APIs require an explicit caller-owned ArtifactResourceScope. CLI and MCP input contracts are unchanged.
* feat(apktool): decode Android resource tables through a caller-supplied launcher
Add an Apktool family with two operations over the caller's apktool
launcher (REA_APKTOOL_COMMAND, absolute, or apktool on PATH):
- inspect_apktool_client: resolved launcher, selection source, and
version; decodes nothing.
- decode_android_resources: apktool d --no-src in a provider-owned
workspace that is removed before the result returns. Projects the
decoded AndroidManifest.xml (bounded at 512 KiB), apktool.yml version
and SDK facts exactly as printed, string resources verbatim (default or
one locale, bounded at 2000 entries, nested-markup entries counted as
unparsed), available string locales, and the decoded file count with a
hard cap. The target APK is digested with SHA-256 before decoding so
Evidence binds the observation to the exact input.
Smali production is skipped by design: Java-source recovery belongs to
the JADX family. Workspace cleanup happens before results return; an
unremovable workspace is named in the limitations instead of abandoned.
Verified against apktool 2.7.0-dirty (Debian packaging) on Linux with
OpenJDK 25 against a signed aapt2-built probe APK with two resource
locales: launcher version, on-disk digest agreement, metadata projection,
manifest package agreement, and both string tables (default and de).
Parser goldens come from the same decode. READMEs carry the new
'Android resources' row in every supported language.
* fix(apktool): own snapshots and bound decoded resource reads
* docs(apktool): describe owned resources and bounded projection
---------
Co-authored-by: REA Contributor <[email protected]>
Co-authored-by: morluto <[email protected]>
* feat(jeb): add JEB engine provider over its MCP interface
Add a bring-your-own JEB provider that serves project, unit, and
decompilation inspection from a caller-started JEB client speaking MCP
(streamable HTTP, selected with REA_JEB_MCP_URL). Expose four provider-neutral
operations through shared CLI/MCP workflows: inspect_jeb_client,
open_jeb_project, list_jeb_units, and decompile_jeb_item, each returning
Evidence with engine identity, endpoint, and preserved refusal reasons.
The adapter keeps JEB's envelope semantics: success fields are parsed
format-aware at the boundary, message-only refusals become unsupported-target
errors with the engine's own reason, transport failures become availability
failures with recovery guidance, and unit pagination derives coverage from the
engine's 100-item page cap. Explicit unknowns (implicitly selected unit,
omitted filters) are reported instead of guessed.
Verified against a real JEB 5.48.0 licensed build serving jeb-mcp-server 1.3.0
headless: client inspection, unit listing, and method decompilation were
exercised through the CLI, MCP boundary, and scripts/verify-jeb.mjs.
JEB 5.48.0 headless instances do not advertise open_project; the refusal is
preserved and documented.
* fix(jeb): verify against a real signed APK and fix CLI option surface
Verify the provider against a locally built, v1/v2/v3-signed probe APK
(binary manifest, classes.dex with a launcher activity) on JEB 5.48.0: the
full unit tree (apk, manifest, three certificate units, dex bytecode,
decompiler), wildcard-filtered and paginated unit listing with derived
coverage, parent-restricted listing, and launcher-activity method
decompilation through both CLI and MCP, including dexdec's inline
constant folding of the probe's obfuscated tag.
Move list-jeb-units filter/pagination and decompile-jeb-item unit selection
from positional arguments to options; the MCP schemas are unchanged.
* fix(jeb): preserve engine refusals through the real MCP boundary
Testing against the published Signal 8.30.3 universal APK exposed a boundary
bug the scripted provider tests missed: decodeJebToolResult raised a protocol
error for isError results before the provider could map JEB's
{success:false, message} envelope, so real engine refusals surfaced as
unreadable output instead of the engine's own reason. Decode refusal
envelopes like any structured result and let the provider preserve the
reason; cover the decoder directly, including contentless and unparseable
results.
Signal verification: full 109 MB R8/Kotlin/Compose APK with four signature
schemes and native arm64 ELF units parsed; MainActivity.onCreate and R8
synthetic lambdas decompiled through CLI and MCP; paginated, filtered, and
parent-restricted unit listing; refusal path confirmed on the live engine.
* docs(jeb): record Flutter Gallery verification on the real-engine lane
Verify the provider against the official Flutter Gallery 2.9.2 release APK
(112 MB): per-ABI libapp.so Dart AOT snapshot images analyzed as native code
with 75k methods on arm64, snapshot internals and wrapper methods decompiled
through both CLI and MCP with explicit unit selection, decoded manifest, and
filtered native-library unit discovery. Note that JEB 5.48.0 headless does
not expose the dedicated Dart snapshot processor or run_script; both remain
GUI-only engine surfaces.
* docs(jeb): translate the README JEB row for every supported language
The JEB row was inserted in English across all translated READMEs; each now
carries its own label, capability summary, requirements, and guide link
following that language's Android-row conventions.
* fix(jeb): enforce local endpoints and preserve cancellation
* fix(jeb): refuse redirects across the local MCP boundary
---------
Co-authored-by: REA Contributor <[email protected]>
Co-authored-by: morluto <[email protected]>
* feat(adb): acquire installed APK sets from connected Android devices
Add a read-only ADB device-acquisition family over the caller's adb binary:
inspect_adb_client, list_adb_devices, inspect_adb_device, list_adb_packages,
and pull_adb_package, exposed through the CLI and MCP with shared admission,
Evidence provenance (binary, version, serial, transport), and fixed argv
vectors that never reach a shell.
pull_adb_package acquires the complete split set reported by pm path into a
package-named directory that must not exist, digesting every file with
SHA-256 and keeping per-file failure reasons and partial coverage. Device
kind is reported as an inference with its basis; getprop whitelisting keeps
build identity bounded. REA installs nothing and never manages devices or
emulators; the first device query may start the caller's local adb server,
recorded when observed.
Verified against adb 34.0.5 on Linux with an Android 14 (API 34) x86_64
emulator: full 224-package inventory, a real two-APK split set pulled with
byte-exact digests, and a 73.9 MB system-package pull. npm run verify:adb
records the lane; parser goldens come from the same device. READMEs carry
the new row in every supported language.
* feat(adb): complete the adb surface with observation, transfer, and lifecycle tools
Extend the ADB family from five to twenty-four operations, all through
fixed adb argument vectors with no shell exposure:
Observation: read_adb_logcat (bounded dump, count/buffer/pid), inspect_adb_package
(dumpsys package projection), list_adb_processes (ps -A), list_adb_directory
(ls -la with kinds and symlink targets), list_adb_features (hex GL versions),
list_adb_services (AIDL slash-suffixed names), inspect_adb_display (wm size and
density), inspect_adb_window (focus declarations), read_adb_setting (device-side
nulls reported as such), capture_adb_screen (exec-out screencap with PNG
dimensions), collect_adb_bugreport (digested device archive).
Transfer and search: pull_adb_file and push_adb_file (device-side sha256sum
verification, explicit overwrite authority), resolve_adb_packages (substring
package search with exact-match flag — the find step of an easy get).
Lifecycle for dynamic analysis, each declaring mutatesTarget:
install_adb_package and uninstall_adb_package preserving the device's
Success/Failure verdict, start_adb_app (resolve-activity then am start -W),
start_adb_activity (typed extras as --es/--ez/--ei/--el/--ef), and stop_adb_app
(am force-stop).
Verified end-to-end against adb 34.0.5 on Linux with an Android 14 (API 34)
x86_64 emulator: 287 services, 92 features, split-set pulls with byte-exact
digests, verified push/pull roundtrip, sized screen capture, and the full
install → resolve → start → observe-in-ps → force-stop → uninstall lifecycle.
Fixed a v-mode regex incompatibility in the settings key pattern and added
argument descriptions required by the CLI catalog audit.
* fix(adb): quote remote shell operands and protect captured output
* test(adb): cover bugreport destinations containing spaces
---------
Co-authored-by: REA Contributor <[email protected]>
Co-authored-by: morluto <[email protected]>
inspect_binary_layout accepts the same optional detail input as
analyze_javascript_application. The default complete delivery is
unchanged; detail "summary" retains the complete layout Evidence in the
session and returns its inspect_analysis_view summary, whose
parent_evidence_id selects sections, symbols, mitigations or linkage
later. Both tools share one detail schema and output projection.
Co-authored-by: Claude Opus 5.5 <[email protected]>
analyze_javascript_application returns the complete application graph
inline, which repeats tens of kilobytes even for a few source lines and
exceeds the default 10 MiB MCP frame for a 300-function source file.
The new optional detail input keeps complete delivery by default. With
detail "summary" the server records the complete analysis Evidence in
the session and returns its inspect_analysis_view summary, whose
parent_evidence_id lets callers page modules, select items, or pass the
retained reference to application workflows. A server without Evidence
retention refuses summary detail with capability_unavailable rather
than returning a summary whose analysis cannot be reached.
Co-authored-by: Claude Opus 5.5 <[email protected]>
Long captures emitted only a final result, so a quiet child looked idle.
Pass the shared rate-limited progress reporter through the capture lifecycle
and report the observed phase, elapsed time, and collected-unit counts.
CLI writes those lines to stderr; MCP forwards them only when the client
supplied a progress token. Final Evidence on stdout keeps its shape.
* feat(mcp): add compact input schema profile for size-capped providers
Moonshot's function-calling API rejects a request while any single tool
parameters schema exceeds its literal serialized size cap, so clients
such as kimi-cli fail every request while the default catalog is
attached (#1484).
Add an opt-in REA_MCP_INPUT_SCHEMA_PROFILE=compact server presentation.
Each advertised input schema is rendered within a 13,312-byte budget,
measured to sit below the observed provider cap: annotation prose
(nested descriptions, examples, defaults, titles) is dropped, repeated
annotation-free subschemas are shared through schema-local $defs/$ref
references the provider validates without expansion, and the three
structurally oversized schemas are advertised in a documented reduced
form. The canonical catalog, tool set, output schemas, and server-side
canonical validation are unchanged; the default full profile is
byte-identical to today's advertisement.
Boundary tests pin every compact schema to the budget, keep them valid
JSON Schema that still accepts every canonical example, hold reduced
presentations at three or fewer, and verify canonical rejection still
guards a permissive advertisement. Verified against the real Moonshot
API through kimi-cli: prompts succeed with the compact catalog attached
where the default catalog fails every request with 400.
* fix(mcp): preserve compact schemas through Kimi conversion
---------
Co-authored-by: morluto <[email protected]>
Remove forwarding wrappers and shadow aliases, migrate callers and verification scripts to their owners, and share one exhaustive semantic resource-bound projection.
Since #1277 an error result reaches the client without structuredContent,
but toErrorToolResult and the transport recovery still budgeted it as
structured JSON plus escaped text. An error whose text fit between about
half and all of the response budget was replaced by resource_constraint,
and without Evidence retention the original diagnostic was lost.
Count only the escaped text for errors. The boundary, process-global and
real Ghidra large-result probes now use diagnostics that exceed the
text-only budget.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* fix(mcp): deliver tool errors within advertised schema contracts
Retain complete typed JSON error content while omitting non-conforming
structured errors at the public transport boundary. Recover oversized
diagnostics into Evidence before removing their private structured carrier.
Update schema-validated client readers and cover actual stdio delivery.
Closes#1261
Assisted-by: Kogasa (gensokyo-skills:kogasa-surprise-testing)
* test(mcp): read typed errors in canonical Evidence regressions
* test(mcp): validate public error diagnostics in acceptance workflows
Support owned JADX stdio on Windows x64, preserve protocol input errors through cancellation, and verify native stdin and Android CLI/MCP lifecycle on the real Windows host.
Closes#1262
Deliver complete Evidence directly through MCP with required nullable analysis profiles and immutable per-server delivery settings. Bind selected environments at construction and require typed cleanup outcomes with retained retry ownership.
Remove test-only routing, duplicate session and snapshot metadata, producer-field repairs, and ambiguous provider arguments. Separate artifact byte identity from occurrence classification and retain typed source-map and partial native facts across CLI, MCP, and verifier consumers.
BREAKING CHANGE: obsolete Evidence envelopes, session aliases, incomplete snapshots, and missing normalized producer facts are no longer accepted.
Add explicit summary, facet, item and paged views of retained analysis Evidence without repeating analysis. Preserve captured observations, graph validation, artifact identity, coverage and complete default results.
Fixes#1050
Verified with focused source, CLI, pinned-SDK schema and MCP tests, serial static checks, and one complete native Windows Notion ASAR workflow. Real Hopper/Ghidra verification was not performed for this provider-independent projection change.
Co-authored-by: Matt Van Horn <[email protected]>
binary_session read the client's identity, protocol revision, and
capabilities only from the 2026-era per-request envelope. A 2025-era
connection sends none, so every such client was reported as `client: null`
with no negotiated protocol and no features, although it declared them in
`initialize`.
When a request carries no envelope, read the values the SDK keeps from the
connection's initialize handshake. An envelope still takes precedence.
Fixes#1167
Co-Authored-By: Claude Opus 5.5 <[email protected]>
morluto