Rebase the change from PR #3776 (21ad0bc2f1) while preserving current document-identity and host-version tests.
Review-support branch for the original contributor: the author-disabled maintainer-edit setting prevents updating the original PR directly.
Preserve invalid V8 script events as unavailable protocol data instead of merging them into target-policy exclusions. Keep independent producer counts, original Evidence, and existing target authority.
Validated with 90 focused Linux tests, real Node Inspector CLI/MCP captures, complete static/architecture checks, and current-head automated review/CI.
Fixes#4376.
Include request and response endpoint carriers in the shared semantic query seed matcher so traces retain resolved response consumers, source identities, and incomplete coverage.
Validated with 44 domain tests, compiled CLI/MCP parity, full static/architecture checks, and current-head automated review/CI.
Fixes#4213.
Refs #4502, #4506, #4508, #4518, #4522, #4526.
Retain snapshot, endpoint, launch rollback and compiler cleanup owners. Join request finalization and cleanup cancellation. Preserve bounded compiler output and original byte counts through capture and Ghidra failures. Normalize execFile cancellation once at its producer.
Validation includes real macOS Ghidra TCP MCP success workflows, controlled Node compiler/process fault fixtures, focused regressions and static checks. Native Swift compiler faults, Windows execution and stdio transport were not verified. Broader startup primary classification in #4517 and native adapter propagation remain unfinished.
Refs #4502, #4504.
Keep the original snapshot failure alongside disposal uncertainty and retain exact cancellation reasons at the TypeScript runtime boundary.
Native Windows execution and addon fault behavior were not verified on this macOS host; checked handle-release ownership in #4505 remains open.
Refs #4495.
Normalize dylib file access once at the admitted descriptor boundary. Preserve full file identity, physical read outcomes, cancellation provenance and cleanup retry through artifact analysis.
Refs #4471, #3946.
Read only admitted regular descriptors without blocking on pipes. Retain failed closes, join discovery workers, and propagate cancellation through CLI and MCP. Preserve command and descriptor cleanup as independent owners.
Refs #4467.
Retain failed socket owners across translated failures and resolve aggregate cleanup only after every required transport closes. Preserve concurrently active peer sessions.
Refs #4373, #4455, #4457, #4458.
Use canonical artifact ownership in Go, Rust, SQLite and PE readers and their failure consumers. Admit PE read budgets before allocation and ignore zero-extent sections in interval ownership. Close Rust resources through CLI and MCP lifecycles.
Refs #4458, #4483.
Own admitted descriptors through final close, retain exact failed owners, and preserve physical read failures and cancellation reasons. Track stream coverage independently from delivered bytes and share the boundary across archive, Mach-O, hashing, and output readback consumers.
Refs #4378, #4384, #4432, #4496, #4501.
Keep inaccessible MSIX targets unknown, preserve literal AppData roots, and distinguish absent skill destinations from retained files. Isolate registration tests from host provider discovery.
Windows package behavior was exercised through filesystem fixtures; the installed Windows Claude application was not verified.
Correct mistranslated technical terms and missing details across the Korean
website catalog. Improve sentence flow and spacing, and remove trailing
periods from Korean headings while keeping comparable overall text length.
Preserve English source keys, placeholders, comments, metadata, and obsolete
translation memory.
Refs #3798
Co-authored-by: morluto <76467478+morluto@users.noreply.github.com>
* fix(reference): resolve extensionless imports to .mts and .cts files
Include .mts and .cts extension probe slots in RESOLUTION_SUFFIXES so that relative imports and requires targeting ESM and CJS TypeScript files without an explicit extension resolve to internal relationships instead of remaining unresolved.
* test(reference): restore regression coverage for extensionless .mts and .cts imports
Restore tests covering extensionless imports, requires, and directory indexes to .mts and .cts files so that probe suffix omissions are caught by automated test regression.
* fix(native): report PE input byte budgets as file-size limits
PeResourcesProvider translated every ArtifactReaderFailure with reason
"limit" into a memory resource constraint, so an image above the
caller's max_file_bytes was reported as a memory problem with memory
remediation advice. The bound that was exceeded is the input file's
byte length, which the caller raises with a flag.
Record the constrained bound on the failure the way ElectronFuseReader
already does, declare it at the reader's two input-byte refusals and at
the parser's metadata and entry budgets, and report that bound instead
of assuming memory. Readers that do not declare a boundary are
unchanged.
Fixes#4192
* fix(native): keep caller-selected PE budget refusals actionable
Address review findings on the budget classification:
- withCleanup now forwards the constrained bound, so a file that grows past
its byte limit while its handle also fails to close is still reported as
a file-size refusal rather than a memory one.
- Both caller-selected bounds carry their own remediation, pointing at
max_file_bytes and max_entries instead of the file-size fallback's worker
write-limit advice.
- AnalysisResourceConstraintError can declare that retrying cannot succeed.
The entry budget is deterministic for the same file and limit, so it no
longer inherits admission capacity's retryable behaviour.
The projection falls back to the previous resource-kind rule when a caller
does not declare retryability, so other capacity refusals are unchanged.
* fix(native): preserve budget retryability and name the schema ceiling
Address the remaining review findings on the budget classification:
- Failure redaction and the caller-copy path rebuilt resource constraint
errors without forwarding retryable, so a deterministic capacity refusal
projected as retryable again and automated callers could loop on it.
Both constructors now forward the declared boolean.
- Both remediations recommended raising a bound the caller had already
set to the schema maximum. The maxima are now named constants shared
with the schema, and the remediation names the operation's hard ceiling
when the caller is already at it.
* Fix formatting in Korean README file
Corrected formatting in the README_ko.md file.
* Fix formatting of REA website link in README_ko.md
Update the REA website link format in Korean README.
* fix(docs): prevent misleading Korean website instructions
Correct the reviewed Korean catalog translations and repeated errors so
readers retain the English source's technical meaning and evidence.
Align the Touhou blog card and article bylines with the existing Korean
author/date format. Restore separators, identifiers, prerequisites,
license names and runtime status descriptions without rewriting source
messages or obsolete translation memory.
Constraint: Preserve msgid, msgctxt, named placeholders and recorded source facts
Confidence: high
Scope-risk: moderate
Tested: Strict Korean catalog validation; website checks and 18 regressions; npm run check; git diff --check; local blog and article browser checks
Not-tested: Every translated page across mobile widths and themes; production deployment
build_reconstruction_obligation_ledger reported a bundle with duplicate
record IDs, or a reviewed obligation citing Evidence absent from the
bundle, as unreadable_output and pointed callers at rea doctor. Both are
caller input mistakes.
Throw dedicated TypeError subclasses at those checks and map them to
workflow input errors, following the NonCanonicalEvidenceBundleError
approach from #3219 so unrelated TypeErrors stay protocol failures.
Fixes#4444
* fix(domain): classify TypeScript parameter properties as declarations rather than identifier reads
Recognize TSParameterProperty in isPatternOwnerDeclaration, isDeclarationIdentifier, and patternContains so parameter property definitions are not emitted as identifier reads, while retaining valid initializer reads. Fixes#4371
* test(domain): assert constructor body parameter property reads survive
* fix(native): reject empty Swift class vtable owner before slot insertion
Throw an explicit RangeError within the descriptor recovery block when a Swift class vtable descriptor class name is empty. This prevents emitting malformed slots that fail final schema validation, allowing partial coverage and valid sibling descriptors to be preserved. Fixes#4369
* test(native): verify decoding continues for valid Swift vtable siblings after a malformed descriptor
Preserve caller-selected max_capture_age_ms in compare_process_captures Evidence parameters so returned and session-retained records distinguish freshness-constrained comparisons and derive distinct semantic Evidence identities.
Add boundary MCP regression coverage asserting that max_capture_age_ms is omitted when unconstrained, retained in parameters and session evidence when specified, and differentiates comparison Evidence IDs.
Closes#4000
Older Android builds print `wm density` as a bare number. The parser's
documentation named that form, but only the labeled `Physical density:`
and `Override density:` lines were matched, so a bare `420` left
`physical_density` null in successful display Evidence with no limitation
saying the declaration was unread (#3942).
The parser now reads the first bare numeric line as the physical density
when no labeled physical declaration is present; a labeled declaration
still wins so a mixed dump cannot downgrade the value. Override density
and physical size are unchanged.
Validation: parser unit tests for the labeled, bare, CRLF, mixed and
unrecognized forms, plus a provider-level regression through the
inspect_adb_display operation; two of them fail on main and pass here.
The issue's three CLI controls through `rea inspect-adb-display` against
a stub adb return physical density 420 for the bare form and keep the
labeled and override controls. Typecheck, lint and module boundaries pass.
Closes#3942
* fix(ghidra): consume inspected configuration directory identity
Carry the parsed release coordinate through installation inspection and provider prerequisites into the launcher. Remove duplicate property reads and basename guesses, preserve unresolved identity, and canonicalize directory versions without changing observed provider versions.
Keep verifier callers, lifecycle regressions, and installation guidance aligned. Extract the Ghidra-only follow-up from #2293 without changing CLI output or unrelated documentation.
* docs(ghidra): clarify startup failure side-effect boundary
Distinguish the launcher's directory and Java startup guard from the client's earlier private runtime allocation and target snapshot. Keep existing failure cleanup and runtime behavior unchanged.
---------
Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Windows ships two tar implementations. `C:\Windows\System32\tar.exe` (bsdtar)
terminates every listed member with CRLF, so `verify-package-pack.mjs`
splitting the listing on a bare "\n" left a trailing "\r" on each non-final
member. `verifyPackedBridge` compares member identities exactly, so it reported
`package omitted package/bridge/hopper_bridge.py` although the tarball ships
that member (#4128).
`parseTarListing` normalizes line endings once, at the listing owner, and
preserves every member string beyond its line ending. `verifyPackagePack` is
the only tar listing producer in the package verifiers, so its consumers now
compare bare member names; the prefix checks in the same file keep their
behavior because they never matched a line ending.
Covered by `tests/conformance/package/packedTarListing.test.ts`: CRLF and LF
listings both yield exact bridge identities, the trailing line ending no longer
produces an empty member, member strings keep embedded spaces, and a regression
guard pins the previous parse's failure.
Real Windows 11 evidence (`package/bridge/hopper_bridge.py` in the listing):
| tar | listing | split("\n") | parseTarListing |
| ---------------------------------- | ------- | ----------- | --------------- |
| Git for Windows GNU tar | LF | true | true |
| C:\Windows\System32\tar.exe | CRLF | false | true |
Fixes#4128
Co-authored-by: Chow Kim <catch@126.com>
Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(android): make the JADX operation deadline configurable
JADX operations used a fixed 120 s provider deadline and a separate fixed
120 s per-call bound in the session, so an APK whose cold load exceeds
120 s could not complete, and raising only the MCP client timeout could
not extend either. Add REA_JADX_OPERATION_TIMEOUT_MS, parsed once by the
JADX configuration owner (decimal milliseconds from 1 to 2147483647,
default 120000). The provider resolves it before acquiring anything and
uses it for the operation deadline and its timeout errors, and the session
uses it for every upstream call. Doctor reports the effective value, and
setup forwards an explicit value like the other JADX settings.
Fixes#3918
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(android): isolate JADX fixtures from inherited shell settings
The shared JADX protocol fixture spread the developer's process.env, so an
exported REA_JADX_OPERATION_TIMEOUT_MS (or REA_JADX_HEAP_MIB) changed every
provider test that never selected one. Keep inheriting the host environment
for process launch, but take REA_JADX_* settings only from the fixture.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(website): correct mistranslated zh-CN catalog entries
The zh-CN gettext catalog had not been reviewed. Terminology was taken from
unrelated dictionary senses and several file names, version numbers and
observed byte values were translated as if they were prose, so rendered pages
showed content that did not match the English source or the recorded target
identity.
Rewrites 236 msgstr values across the catalog:
- Terminology now uses one term per concept: mask -> 掩码, assembly -> 汇编,
helper -> 辅助函数, source -> 源码, native -> 原生, oracle -> 校验基准,
authenticator -> 身份验证器, agent -> 编程助手, and flag stays untranslated.
- Literal values are preserved: MAIN.EXE, main/, Ghidra 12.1.x, 0x5b?,
export.js, preload.js and npx rea-agents@latest setup are no longer
translated.
- Word-salad sentences are rewritten in natural Chinese, including
"剂将它们转化为方程" -> "编程助手把它们转化为方程".
- Step-marker separators damaged from "·" to "*" are restored.
No msgid or msgctxt changes, so translations extract is not required and no
entry becomes fuzzy or empty.
* Update website/locales/zh-CN/messages.po
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* Update website/locales/zh-CN/messages.po
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
---------
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Normalize client identities once in the shared workflow while preserving first-selected order. Keep planning, readiness, default skill selection and configuration application on the same unique list.
Cover duplicate selections in direct workflow tests and an isolated compiled CLI preview/apply/rerun journey. Refs morluto/rea#3847.
* feat(ghidra): support raw big-endian m68k analysis
Admit `m68k` for headerless raw-binary targets. Admission requires a
2-byte aligned entry and a mapping within 0x0..0xffffffff, and Ghidra
imports with BinaryLoader using 68000:BE:32:default, the default compiler
specification, and the `ram` default address space.
`6502` and `m68k` now share one RAW_ONLY_ARCHITECTURES owner, so detected
targets, Evidence subjects, snapshots, and lifecycle output contracts
exclude both instead of repeating a 6502 literal.
Ghidra's default 68000 language decodes the 68040 superset; separate
68020/68030/CPU32/ColdFire variants, Amiga/Atari containers, and reset
vector discovery remain unsupported. Windows Ghidra P0 still admits raw
6502 only.
Verified with Ghidra 12.1.4 on macOS against an 8-byte self-built fixture
(nop; nop; moveq #1,d0; rts at 0x1000): load image verified in `ram`,
entry decompiles to `return 1`.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* test(ghidra): verify raw m68k analysis with real Ghidra
Add npm run verify:ghidra:m68k. The verifier loads a source-owned 22-byte
68000 fixture (annotated hex, no assembler) at 0x1000 and checks, through
the compiled CLI and stdio MCP server: the 68000:BE:32:default BinaryLoader
import, load-image parity, exact mapped bytes, both functions, the bsr.b
call reference, decompilation by address and file offset, rejection of a
misaligned entry and of a mapping past 0xffffffff without replacing the
active session, unchanged source bytes, and runtime cleanup.
Run with Ghidra 12.1.4 and JDK 21 on macOS arm64. Linux has not been
exercised; Windows Ghidra P0 does not admit m68k.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Identify the Cutter host through Qt applicationVersion rather than the embedded Rizin module-version report. Preserve unknown host metadata without backend fallback and reuse the existing per-plugin lifetime observation.
Validated 13 Python protocol tests, 66 related CLI/MCP/provider tests, full static/docs gates, and official Linux Cutter 2.5.0 Qt5 public workflows. Native discovery and both Evidence version fields match the independently observed host version; source paths and cleanup remain preserved. The named-tool workflow serialization proxy drops from 8767 to 6912 o200k tokens; this is not model billing.
Fixes#4498.
Read file.path through CutterCore's scalar configuration API instead of invoking the debug-only currentlyOpenFile backup field. Preserve exact string paths and unknowns so existing discovery, generation admission and Evidence consumers receive the owning facts.
Validated 11 Python protocol tests, 66 relevant CLI/MCP/provider tests, full static/docs gates, and official Linux Cutter 2.5.0 Qt5 workflows with a Unicode/#? ELF path. Same-path reload identity remains partial and native document switching is unverified.
Fixes#4497.
Both resource-interval overlap predicates in PeResourceLayout used strict
endpoint inequalities without an extent or point ownership rule, so a
section with VirtualSize=0 and SizeOfRawData=0 inside a resource interval
was rejected as overlapping virtual sections, and the same empty section
inside a header-backed payload was rejected as ambiguous header mapping.
An empty payload at an RVA that starts a positive-extent section below
SizeOfHeaders silently resolved to the header offset even though that
section owns the point.
Normalize intersection once at the owning mapper: a section without
virtual extent owns no byte and never conflicts; an empty resource owns
exactly its start point, so a positive-extent section starting there
conflicts with header-backed ownership and keeps the existing
ambiguous-mapping failure. Positive-extent overlap rejection is
unchanged.
Fixes#4455Fixes#4373
Co-authored-by: maxazure <maxazure@gmail.com>
Propagate CLI/MCP cancellation through shared Cutter discovery, stop queued probes, and drain admitted workers. Deliver terminal bridge results only after socket closure. Preserve confirmed command Evidence and caller reasons when cancellation arrives during finalization.
Validated 66 focused tests, full CI static/docs gates, actual CLI SIGINT, compiled stdio cancellation/recovery, and official Linux Cutter 2.5.0 Qt5 CLI/MCP ELF workflows. Native UI cancellation and descriptor filesystem cancellation remain unverified.
Fixes#3946.
Fixes#4388. Refs #4027.
Observe cancellation during first-window waits and state reads; preserve cancellation and caller reasons before cleanup and retain partial observations when cleanup fails. Keep late cancellation and cleanup failure precedence.
Validation: 49 focused cases, static/docs checks, six controlled verifier-oracle cases and 39 current-head CI checks passed. Codex and Greptile found no remaining blocking issues. Real Joplin cancellation phases verified; complete Linux cleanup/first-signal CLI exit and packaged-runtime startup remain explicitly unverified/failed. Both verifiers now fail incomplete cleanup, including after main-process exit.
Co-authored-by: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
Co-authored-by: Adolanium <94890352+Adolanium@users.noreply.github.com>
Retain received complete and unfinished Rizin command bytes with bounded UTF-8 capture, truthful completion uncertainty and exact launch provenance. Keep captures independent of cancellation, protocol and cleanup precedence; wait for bounded output settlement on exit and preserve active bytes across stdin failures.
Validated353 focused tests (one existing skip), full CI static/docs checks, real official Linux Rizin0.9.1 interruption probes, paired completed public CLI/stdio MCP tasks, and focused mutation checks. Required CI passed; exact-head Codex review reported no findings and Greptile5/5. Fault cleanup uncertainty and partial-marker ambiguity remain explicit.
Fixes#4015. Follow-up to #4227.
Fixes#4381. Follow-up to the Linux lifecycle audit in #4227.
Keep managed source hashing on bounded reads through its existing FileHandle owner. Preserve exact opaque cancellation reasons before filesystem or cleanup interpretation, retain genuine failed closes, block unsafe admission, and recover the same provider after verified cleanup.
Validated with 484 related tests, source/docs checks, seven real Linux provider cancellation/recovery scenarios using an unchanged official Microsoft .NET assembly, and paired actual CLI/stdio MCP Evidence workflows. Normal completed-task timing, CPU, memory and token proxies remained comparable.
Adopts xu91102/rea commit ebeec580c839f2f72144282a44f4e11ea5759843 with independent Linux verification and cancellation-reason hardening.
Co-authored-by: hhang <169027046+xu91102@users.noreply.github.com>
Fixes#4379.
Observe cancellation after file metadata, verified descriptor closure and recovered scope cleanup. Retain completed physical read outcomes and acquired-file digests under cancellation; preserve original failures and exact cleanup owners when close fails.
Bind cancellation markers to their verified owner so raw readers preserve nested callback outcomes and exact caller reasons. Reuse one ADB partial-result mapper for file, package and bugreport acquisition.
Validated with 570 related tests, source/docs checks, real Linux filesystem cancellation/cleanup/nesting probes and actual public CLI/MCP Evidence workflows.
* fix(browser): report cancellation after cleanup and source-map abort
An interrupt during CDP cleanup could still publish the captured page, and a source-map download abort was wrapped as a provider failure. Both now return AnalysisCancelledError for the requested operation.
* fix(setup): keep Devin configuration JSON-formatted
The empty DSH_HOME fix also changed Devin's declared format to "devin", which is not a configuration format and stops the typecheck. Devin still writes mcp_config.json.
* test(browser): keep cleanup cancellation coverage in its own suite
The added interrupt cases made the existing lifecycle suite exceed the function-length lint limit.
* test(ghidra): drop the duplicated path join import
The inventory MCP test imported join twice, which fails lint on current main.
* test(cli): type the retained skill directory pairs
noUncheckedIndexedAccess treats the unlabeled pair list as an open array, so the directory passed to join was string or undefined.
* fix(browser): preserve source-map cancellation results
* style(test): format the partial Frida load cancellation fixture
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: REA Maintenance <rea-maintenance@users.noreply.github.com>
Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Fixes#4011.
Normalize exec-aware worker ownership at the shared pwntools decoder boundary. Preserve parent, process-group and run-token checks, explicit command pins, partial observations, and retryable cleanup.
Validated with real Linux pwntools/GDB/pwndbg CLI/MCP workflows, native-exec cancellation and foreign-token controls.
Mohamed Ayoub Eleuch