procedure_callees and procedure_callers returned a bare address list, so an
inferred no-return call that truncates Ghidra's function body silently dropped
real edges (fixes#2326). Wrap both views in the function-observation envelope
with noReturnCallLimitations, matching procedure_info, procedure_assembly,
procedure_pseudo_code, read_function_instructions, and analyze_function
(precedent: 69a8b3a95). Version the boundary-observation contract to
ghidra-terminal-call-limitations-v2 so older snapshots cannot replay
unqualified results.
Validation: red->green unit test for the envelope requirement, MCP boundary
evidence test extended for both views (limitations retained through Evidence
and snapshot replay), 8 harness test files + ghidra behavior tests green,
check:fast, docs:check, verify:test-discovery green. Real Ghidra host lane
unavailable here (no Ghidra installed); Java change mirrors the existing
procedureInfo call pattern.
* fix(pwntools): let callers widen the artifact input bound
The pwntools adapters rejected every artifact over 32 MiB before any
analysis ran, and nothing in configuration, environment, CLI or MCP
could raise that bound, so real recorded core dumps stayed unanalyzable.
Resolve the bound from REA_PWNTOOLS_MAX_INPUT_BYTES in front of the
shipped default. An absent value keeps 32 MiB, and a value outside
1..1073741824 bytes is refused as unusable provider configuration before
the artifact is read instead of falling back silently.
* fix(pwntools): name the bound variable in its own refusal
An unusable REA_PWNTOOLS_MAX_INPUT_BYTES travelled through the shared
engine-absence helper, which always stamped configuration_key
REA_PWNTOOLS_PYTHON, so a caller following that field was pointed at the
Python executable instead of the bound it had just set. The guide also
described the refusal as an invalid request while the decoder reports
provider_unavailable.
Let the failure context carry the key under review and pass the bound's
own name at that call site, then state the code callers actually receive
in the guide. Assert the key on the projected rejection so the engine
default and the bound override each stay attributed to the right setting.
---------
Co-authored-by: morluto <76467478+morluto@users.noreply.github.com>
* feat(ghidra): admit raw NMOS 6502 analysis
* test(ghidra): require verified POSIX verifier cleanup
* test(ghidra): verify 6502 low-memory raw imports
* docs(skill): document raw NMOS 6502 support
* fix(ghidra): preserve requested Windows raw source paths
* docs(ghidra): include raw 6502 in Windows installation scope
* test(ghidra): verify Windows raw junction rejection through CLI and MCP
* fix(validation): make fixtures portable and preserve explicit Codex home
* fix(ghidra): attest raw mapping state and address space
* fix(ghidra): align diagnostics and regression groups
* fix(ghidra): bind raw attestation to language address space
* chore(ghidra): avoid verifier insertion conflict
* docs(ghidra): align linked raw admission guides
* fix(ghidra): retain 6502 verifier workspace on failure
* test(ghidra): isolate Windows authority doctor fixture
* fix(binary): re-admit Windows raw paths on session reopen
* test(binary): require fresh provider execution after path change
* Constrain 6502 to raw lifecycle and snapshot targets
* Bypass saved reads for unadmitted Windows paths
* Guard direct CLI snapshot replay on Windows paths
* fix(snapshot): gate Windows Ghidra replay on native admission
* fix(snapshot): constrain nested 6502 subjects to raw images
* test: make dyld slice fixtures portable
* feat(managed): name the declaring type of call and field targets
Call edges named external targets by member name alone (Join, ToJson,
WriteAllText), and method-spec edges had no name at all, so an analyst could
not tell which library a method calls or find every use of a type without a
decompiler. Field accesses through MemberRefs had no name either.
Name each TypeDef, TypeRef, and generic-instantiation TypeSpec once, in the
namespace-qualified form TypeDef full_name uses, and add member_refs
parent_name, call_edges target_declaring_type, and field_accesses
field_declaring_type. A method-spec target reports the generic method it
instantiates. Fields, member references, methods, and method specs now
resolve through one token-keyed MemberCore lookup instead of three maps.
Evidence recorded before these fields reads them as null.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(managed): validate spec blobs and member kinds before naming targets
Address review findings on declaring-type naming:
- A generic instantiation must name a TypeDef or TypeRef, so a TypeSpec
base is reported as malformed instead of followed. Naming no longer
recurses.
- TypeSpec and MethodSpec instantiation blobs are decoded with the
member-signature grammar. Malformed blobs and out-of-range TypeDef rows
are reported against their owning row and make coverage partial.
- A MethodSpec reads both columns. An invalid Method coded index is
reported; a readable Method keeps its name beside a malformed
instantiation.
- MemberCore carries its signature kind, so a field operand is named only
by a field or field-signature MemberRef. This replaces the token-table
guard.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(managed): validate MethodSpec arity and report unsupported spec blobs
- A MethodSpec now checks that its method signature is generic and that
the instantiation supplies as many type arguments as it declares. A
non-generic target, a field reference, or a count mismatch is reported
as invalid-row; the readable method keeps its name.
- Unsupported TypeSpec and MethodSpec forms raise an owner-scoped
unsupported-signature issue instead of reading as an intentional
unknown.
- MethodSpec reading moves to ManagedMethodSpecs.ts, and its tests get
their own describe block.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(managed): reject non-method signature headers and invalid TypeRef scopes
The method signature header reader now owns the calling-convention rule
that only FNPTR applied: FIELD, LOCAL_SIG, PROPERTY, and GENERICINST
headers, the reserved bit, and EXPLICITTHIS without HASTHIS are malformed.
A MethodSpec whose method has a PROPERTY header is no longer accepted as a
generic method, and member signatures with such headers no longer decode
as methods.
The type-name lookup validates a TypeRef's ResolutionScope coded index and
reports an out-of-range scope as invalid-row, keeping the readable name.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(managed): hold member rows to MethodDefSig and FieldSig headers
Method signature headers are now checked in context. MethodDef and
MemberRef rows accept only the MethodDefSig/MethodRefSig headers: DEFAULT,
VARARG, or GENERIC with DEFAULT. FNPTR keeps the unmanaged conventions
compilers emit for function pointers. A field signature must be exactly
FIELD, so a flagged FIELD header is malformed and never names a field.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(managed): accept MethodSpec arity only from a fully decoded signature
The arity check read only the target's signature header, so a truncated
generic method signature passed. Under type selection a MethodDef target
outside the selection is never decoded elsewhere, so coverage stayed
complete. The arity now comes from a transient full decode, and only a
decoded method signature supplies it. signature() keeps its raw length
and digest around the shared decodeSignature.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* perf(managed): decode each MethodSpec target signature once
edgeTargets caches the generic arity, including a failed decode, by
target method token, so many instantiations of one generic method share
a single signature decode. Unity.InputSystem has 472 distinct MethodSpecs
over about 133 target methods.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(managed): validate spec type operands and transitive reference scopes
* fix(managed): retain readable names when a scope row is unreadable
* test(managed): pin owner-scoped spec operand and scope-chain checks
e59217c83 validates spec type operands inside specSignature, so a
malformed generic TypeSpec base is now invalid-row under its owning
TypeSpec row; the three stale expectations follow that. New cases cover
a TypeSpec argument and a MethodSpec argument that name a missing TypeRef,
and a TypeRef whose ResolutionScope names itself. The scope cases move to
their own describe block.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(managed): keep field targets from naming MethodSpecs
A MethodSpec over a field-signature MemberRef inherited kind "field", so
an ldsfld on that MethodSpec published a field name and declaring type
for malformed IL. A MethodSpec now borrows its target's name only when
the target is not a field; the arity check still reports it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(managed): reuse owner-decoded signatures for MethodSpec arity
MemberRef signatures, and MethodDef signatures in the type selection, are
already fully decoded, but the MethodSpec arity check decoded them again
and admitted a second transient allocation, which could exhaust the
budget. MemberCore now carries the owner's genericArity, and only a target
nothing decoded, such as an unselected MethodDef, falls back to the raw
signature. signatureGenericArity is the one rule for both paths.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: morluto <76467478+morluto@users.noreply.github.com>
* test(mcp): expect trace_feature to report references the provider truncated
A provider that reports a capped xrefs read makes the trace truncated with a residual unknown for that target. An explicit complete report and a provider that reports nothing leave the trace as before.
* fix(mcp): report references a provider truncated in trace_feature
Executions gain an optional flag for a capped xrefs read, set by the IDA adapter from the upstream cap. trace_feature marks the trace truncated with a residual unknown for that target only when the provider reports a truncation; a provider that reports nothing keeps its previous result. The trace_feature workflow profile is revised so earlier results are not replayed.
* refactor(ida): report reference truncation whenever the runner observed an xrefs cap
* test(mcp): cover session forwarding, per-target state, native reports and failures of the trace coverage
* test(mcp): format the session forwarding test
* test(ida): expect headless IDA to resolve containing procedures from lookups and segment facts
Cover positive and negative containment, segment boundaries including addresses above 2^53, failures of the lookup and of the segment survey, cancellation, input validation and the admitted program as the only document.
* test(ida): expect feature tracing and address context to work on headless IDA
Run trace_feature, including a capped reference read, and inspect_address_context with and without an explicit document through the IDA session client.
* test(mcp): expect trace_feature availability to follow the search operations it calls
* fix(mcp): derive trace_feature availability from the search operations it calls
* feat(ida): resolve containing procedures and report the admitted program as the document
The modern headless profile resolves an address through the function lookup and, when no function contains it, tells a mapped address from an unmapped one using the database-scoped minimal survey. A survey that is missing, truncated or inconsistent fails the read. list_documents reports the single admitted program. The attached profiles do not declare either operation.
* docs(ida): describe containing procedures, the admitted document and truncated feature traces
* test(ida): verify containing procedures, the admitted document and feature tracing on real IDA
* docs(testing): list the headless IDA workflows the real lane verifies
* test(ida): make each survey failure case isolate a single defect
* test(ida): pin large addresses, segment order, document identity, attached refusals and each trace_feature prerequisite
* test(mcp): expect a repeated reference source to be resolved once in trace_feature
* fix(mcp): resolve the containing procedure of a repeated reference source once
* test(ida): trace the fixture procedure and derive probe addresses from its facts in the real lane
* test(process): format processCaptureLifecycle.test.ts after the merge
* test(ida): declare the alignment each synthetic universal slice offset satisfies
* refactor(ida): move the real lane's feature trace assertions into a helper
* test(ida): expect the real lane's feature trace check to accept references to every matched address
* fix(ida): check each trace reference against its own matched address in the real lane
* test(ida): pin that the new operations reject a document other than the admitted program
* refactor(ida): move the real lane's mapped non-function probe into a helper
* test(ida): expect the real lane's mapped data probe to skip strings inside functions
* fix(ida): try the target's strings in order for the mapped non-function probe
* test(ida): declare the real lane helper types for the verifier tests
* test(ida): expect the real lane to reject a listed string that resolves as unmapped
* fix(ida): fail the real lane when a listed string resolves as unmapped
Limit classification to five minutes, static/build/docs/report jobs to ten minutes, and each portable shard to thirty minutes. The saved successful-job sample peaked at 392 seconds per shard and 61 seconds for other core jobs, leaving generous setup and coverage headroom. Runner queue waiting is unaffected.
Select SQLite compatibility by explicit source, fixture, lifecycle, and script ownership while retaining every version in full qualification. Share the canonical scope inventory between planning and the required gate.
Reuse recent exact-SHA full qualification only for hourly schedules, with trusted successful-run metadata and receipts. Force a daily refresh, reject stale original runs even after retries, and never renew proof from reused validation.
Report attempt-specific queue, execution, and matched step timing in a separate completion workflow using trusted default-branch code and read-only API access. Preserve raw metadata and unknown timings without delaying validation or publication.
Validation: 177 focused tests; typecheck/lint; knip; actionlint 1.7.7; all 1046 tests discovered; live 37-job API report smoke. Remote workflow scheduling and real-provider matrices remain CI verification.
Remove redundant ref-based concurrency from reusable provider and website workflows. Their old groups could cancel qualification from another CI tier even though the callers used separate groups. Disposable hosted jobs now inherit caller scheduling; direct manual dispatches remain independent.
Validate every called CI workflow has no independent cancellation group. Validation: 100 focused scope/gate tests and actionlint 1.7.7.
Scope main pushes from a successful ancestor so canceled runs cannot hide changes. Keep the complete portable suite in fast feedback and move coverage plus Intel macOS to hourly, manual, explicit-full, and release qualification. Let qualification finish independently of superseded feedback.
Bind publication to full qualification at its exact revision and publish the verified tarball after checking workflow identity, version, and SHA-256. Keep trusted cache writers and reject incomplete merged test inventories.
Validation: 151 focused CI tests; typecheck/lint; actionlint 1.7.7; knip; all 1044 files discovered; bounded blob/merge smoke; macOS package verifier with retention and cleanup; npm pack dry run. Native Linux/ARM/Windows and remote workflow execution remain CI verification.
Compare import and require outcomes for static TypeScript references when compiler options are unknown. Preserve branch facts through worker IR and both graph projections; retain declaration fallbacks as candidates without changing explicitly selected paths.
Validation: 210 focused resolver and semantic tests; 21 production-boundary and source-extension tests. A bounded 2,000-reference probe adds about 16–17 ms wall time for dual resolution; full task token costs and real compiled application behavior remain unmeasured.
* perf(native): add opt-in Mach-O summary delivery and retained pages
inspect_macho returns every import and export inline. For VS Code's
Electron Framework one MCP call carries 2.5 MB of text (about 828k o200k
tokens), and the same again as structuredContent; the MCP response
budget admits it. Agents asking for slice identity, UUID or dependency
counts paid that cost on every first call.
Add the established opt-in detail selection: omitted or complete keeps
the existing complete result and Evidence identity, while summary
retains the complete Evidence and returns slice identity with
per-collection returned/total/exhaustive counts (956 tokens).
inspect_analysis_view kind macho pages the original rows of any
collection or the command provenance, repeating the producer's own
total and exhaustive flag so page exhaustion is never read as absence.
Delivery detail never reaches the native tools or the source Evidence
ID: registerEvidenceTools gains a parametersForEvidence hook, so the
summary's parent is the same Evidence as a complete call.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(mcp): refuse unretainable summary delivery before analysis
A server without Evidence retention (a fixed analysis source) ran the
complete inspect_macho or inspect_artifact analysis for detail summary
and only then refused it in summarizeRetainedAnalysis, paying the very
cost summary delivery exists to avoid. registerEvidenceTools gains a
preflight hook; both registrations check retention there, as the WASM,
EVM, managed, pcap and capture handlers already do before dispatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(native): keep whole rows in Mach-O architecture and segment pages
The page item union listed plain-object row arrays, and zod returns the
first union member that parses while dropping unknown keys. Entrypoint
rows declare only file_offset, so every architecture and segment row
(which also carry a numeric file_offset) matched it and lost its name,
CPU fields, permissions and sections. Rows now use exact key sets, and
imports/exports share one member. The MCP test pages every facet of a
real /bin/ls and compares it with the unparsed complete rows.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(windows): preserve caller affinity only within one processor group
`spawnProcess` read the caller's affinity mask and applied it to the
suspended child unconditionally. A mask names processors relative to one
processor group, and Windows assigns each new process to a group of its
own choosing, so on hosts with more than 64 logical processors the child
is often placed in another group. `SetProcessAffinityMask` then rejects
the mask with ERROR_INVALID_PARAMETER (87) and the spawn fails, making
every owned provider process start intermittently on such hosts.
Affinity is a nicety, not an authority or isolation control, so it is now
applied only when the caller and the suspended child both occupy the same
single processor group. A child placed in another group keeps that group's
default affinity and the spawn proceeds. Job Object membership,
kill-on-owner-close policy, owner normalization and priority preservation
are unchanged, and the mask failure remains fatal within a shared group,
where it is not expected.
Single-group hosts, including every host with 64 or fewer logical
processors, are unaffected: `GetProcessGroupAffinity` reports one group
for both processes and the same mask is applied as before.
The optional process-boundary fixture gains an `affinity` observation,
and the native verification lane asserts that an owned spawn succeeds and
that the child's mask matches the caller's whenever both share a single
group.
Fixes#2181
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(windows): refuse to widen a pinned caller's affinity across groups
Addresses review feedback on the first commit.
An unconstrained caller loses nothing when Windows places an owned child
in another processor group, so that spawn proceeds with the group's
default. A caller deliberately pinned to fewer processors than its group
holds is a different case: silently accepting the default would hand the
provider tree the whole group, which `docs/installation.md` tells
operators to rely on affinity to prevent. Such a spawn now fails naming
the constraint instead. `GetProcessAffinityMask` already reports the
group's active set, which distinguishes the two without another query.
Selecting the caller's group at creation through
PROC_THREAD_ATTRIBUTE_GROUP_AFFINITY was tried first and rejected:
`CreateProcessW` fails with ERROR_INVALID_PARAMETER on a host with a
single processor group, both alongside PROC_THREAD_ATTRIBUTE_JOB_LIST and
on its own, so it cannot carry the ordinary path.
The boundary fixture's affinity observation now reports a process id's
own mask rather than its own, because a separate probe process is
assigned its own group and cannot stand in for the process that spawns
the child. Masks and the group's active set cross as fixed-width
hexadecimal strings: a mask can set bits above the safe-integer range,
where `JSON.parse` would round two different masks to one value.
The verification lane asserts the mask is inherited exactly when caller
and child share a group, and otherwise that the caller was unconstrained.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* docs(windows): record what owned-process affinity does not preserve
A mask covering its group's whole active set does not distinguish a
caller the system placed in that group from one that selected the group
deliberately, so an owned child can still be placed elsewhere with its
processor count bounded and its group placement lost. Treating that case
as constrained instead is not available: on a multi-group host an
ordinary process is assigned a group by the system and its mask is that
group's active set, so refusing it would fail every owned spawn there.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* fix(apple): limit already-loaded reuse to rpath dependencies
dyld reuses an already-loaded image by install name only for @rpath/
dependencies (Loader.cpp:802-816) and evaluates other dependency forms
against their expanded path for each loader. When multiple images in
different directories depend on @loader_path/ libraries sharing an
install name, checking loaded images before candidate expansion caused
subsequent loaders to bind to the first loaded image's copy or appear
resolved when their own copy was absent.
Limit byInstallName registration and lookup to install names that start
with @rpath/, leaving caller-relative and literal dependencies to expand
and evaluate their own candidate paths.
Signed-off-by: Vinícius DS Andrade <95101505+viniciusdsandrade@users.noreply.github.com>
* test(apple): verify loader_path dependencies across isolated directories
Assert that multiple images depending on @loader_path/ libraries with
the same install name do not reuse already-loaded images across loaders.
When the second loader's helper is absent, its candidate path is absent
and dyld reports required-load-unresolved; when both helpers exist, each
loader resolves to its own directory.
Signed-off-by: Vinícius DS Andrade <95101505+viniciusdsandrade@users.noreply.github.com>
---------
Signed-off-by: Vinícius DS Andrade <95101505+viniciusdsandrade@users.noreply.github.com>
Close persisted Evidence, bundle, profile and unknown record shapes so parsing cannot silently remove caller metadata before checking semantic identity. Keep constructor provider projection explicit and JSON payload fields open.
Exercise actual MCP export/import with independent record and manifest mutations, unchanged state after rejection, and a valid round trip. Prune duplicate source assertions. Validation: focused domain/CLI/MCP/filesystem checks, converted JSON Schemas, docs and typecheck/lint.
Bind the owned demo helper to REA's private stdin stream and unwind child
cleanup on EOF during startup or analysis. Verify graceful and SIGKILL owner
exit through real CLI/MCP workflows and preserve unrelated processes.
Assisted-by: Reimu (gensokyo-skills:reimu-incident-triage)
Summarize saved Wireshark4.x captures with enforced native resource limits, serialized admission, stable snapshot verification and retained fact/text pages. Preserve complete inventories within resource budgets and source uncertainty; report cleanup failure with partial Evidence and retained ownership.
Verified six recorded corpora /13026 packets with independent Scapy/Python oracles, real CLI/MCP/schema parity,60 SNI names,132 DNS addresses and1000 URIs. Real TShark CI now exercises summary/paging. Actual Windows/Bun Wireshark remains unverified; macOS capture requires a separate enforceable-bound decision.
Co-authored-by: 3QL <85309073+3QL@users.noreply.github.com>
Co-authored-by: N0zoM1z0 <161784452+N0zoM1z0@users.noreply.github.com>
Reuse the domain ZIP format list for session schemas, verify full Harmony
CLI/MCP lifecycles, and correct native view guidance and audited docs drift.
Refresh locked dependency notices and gate inventory/provenance consistency,
including notice inputs in the cached documentation gate.
Fixes#2454
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Add explicit MCP Dart AOT summary delivery and exact retained string-pool pages while preserving complete defaults, CLI behavior, ABI selection, original Evidence identity, partial coverage and lower bounds.
Real LocalSend SDK replies shrink 92.05% with approximately unchanged SDK time. Two actual Codex pairs preserve all selected facts and reduce total input 48.75% / 42.38%, uncached input 33.63% / 42.52%, and task time 23.95% / 30.30%; all client errors/repeated calls remain counted. Preserve old portable Evidence and the real CLI complete result after rebasing onto SAZ/qlog main. Curated results keep measured and integration runtime pins separate; raw artifacts stay outside Git.
Refs #2016.
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Inspect bounded recorded raw HTTP/XML SAZ archives and page retained capture Evidence. Preserve exact stored bytes and provenance, validate ZIP/XML integrity, and require an enforceable decoder memory bound.
Verified 356 real recorded sessions against independent byte and metadata oracles, public CLI/MCP parity, retained pages, and related interface checks. Current Fiddler Everywhere GUI and non-Linux producer workflows remain unverified.
Co-authored-by: zahhak <1792039+zahhak@users.noreply.github.com>
Co-authored-by: N0zoM1z0 <161784452+N0zoM1z0@users.noreply.github.com>
Add explicit package summary delivery and exact retained manifest pages while preserving complete MCP defaults, one-shot CLI output, provider requests, source identity and session lifetime. Return distinct linked Evidence and preserve partial/null facts; reject unavailable retention before engine work.
Validation: 27 relevant SDK/domain/Android boundary checks, focused test-only correction rerun, full supported compiler/lint and docs checks, real NewPipe SDK pairs, authentic old Evidence continuation, four actual Codex trials with failures/regressions retained, and all 39 exact-head CI checks. Package replies shrink 119,527 to 8,659 bytes; no model-token or time gain is established.
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Reject native callee successes without an explicit truncation marker while preserving provider failure reasons. Verify positive procedure inventory/search and function-result parity in the real-provider lane; support independent string/xref oracles and label their absence as unverified.
Verified with Windows IDA Pro 9.3: eight normal production CLI/MCP checks passed, and six real-provider response-fault cases reproduced old false positives and validated the correction. Windows IDA Free was not used. No personal paths or license registration information are included.
Repair inconsistent complete-file Ghidra test probe sizes after the shared ELF-header validation change, deriving sizes from the fixture buffers. Reproduced the three failures before the correction; 115 related installation/doctor and ELF tests pass afterward. Runtime parsing and installation behavior are unchanged; no real-Ghidra verification is claimed.
207 focused IDA tests and static/documentation checks passed. Automated review completed on d1e0b1095 without findings. CI run 38327862093 passed all 36 jobs, including Windows, all six test shards, coverage/timing, and real Apple CLI/MCP workflows.
Follow-up to #1544.
Keep complete default delivery and exact filters. Select stable ID pages at the ledger owner before cloning, return filtered totals and continuation offsets, and preserve detached records through the shared session port.
Validate public MCP paging, safe-integer edges, snapshot replay and converted JSON Schemas. Document the bounded Flameox workflow comparison and its token/CPU/latency measurement limits; no real model performance claim.
Add caller-selected managed metadata summaries and original collection pages while preserving complete CLI/MCP defaults and previous-runtime Evidence. Real SDK replies shrink substantially; actual Codex token/latency results remain mixed and are documented without a token-saving claim.
Validated focused SDK/domain tests, full static/docs checks, real two-assembly repeated workflows and portable old Evidence, real SQLite WASM, and all exact-head CI checks. Raw experimental artifacts remain private.
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
* feat(server): frame analysis-target content as untrusted data
Strings, decompiled code, page text, scripts, captures and logs returned
by REA originate from the analysed target, which can embed text aimed at
the agent. Nothing marked that content, so a prompt injection inside a
binary or page could steer the agent toward tools that run code.
- Server instructions now state that target-derived tool output is data,
never instructions, and that any request in it to run, install,
download or send something is a prompt-injection attempt to report.
- Tool contracts declare returnsTargetContent. It is on by default; only
seven tools whose results hold REA session state or caller-supplied
values (binary_session, close_binary, export_evidence_bundle,
record_unknown, update_unknown, inspect_adb_client,
inspect_apktool_client) are audited exceptions, so new tools start
marked.
- Successful results of marked tools carry
_meta.provenance = { untrusted_content: true, source: "analysis_target" }.
content and structuredContent are unchanged, and the response budget
accounts for the fixed marker bytes.
Tests cover the encoding and budget accounting, the audited exception
list against the contract catalog, and an in-memory MCP client that sees
the instructions and the marker on a marked tool but not on
binary_session.
* fix(server): mark failed target-content results as untrusted too
The untrusted provenance marker was attached only on the success branch
of ToolResultDelivery, but failures can carry target-derived text as
well: error messages, captured output, and partial observations such as
the terminal frames a failed capture_process_scenario retains in
partial_observation. A host gating later process or device actions on the
marker would have treated an adversarial error payload as trusted.
Every error path in src/server now carries the failing tool's provenance:
- toErrorToolResult and toRecordedPartialErrorToolResult take the tool's
target-content flag explicitly, so each caller is audited by the type
checker; toCallToolResult, toEvidenceToolResult (recording failure), the
response-budget error that replaces an oversized success, and the
Electron partial-evidence path pass contract.returnsTargetContent.
- Admission failures in runAdmittedToolOperation resolve the flag from the
tool name through the new toolReturnsTargetContent, the same predicate
the contract metadata uses; unknown names stay marked.
- The transport's oversized-error replacement keeps the original's marker
and counts it in its budget check, matching the projected error.
- Errors the MCP SDK builds itself (input validation, thrown handlers) are
marked in the transport for target-content tools only.
Error encoding accounts for the marker bytes exactly as success encoding
does. docs/mcp-contracts.md now states that error results carry the
marker. Tests cover marked and unmarked error results, the
capture_process_scenario partial-observation errors in-process and over
the SDK transport, the recorded partial-evidence path, the budget
replacement, the oversized-error replacement, and SDK validation errors
for a marked and a content-free tool.
* fix(contracts): mark session status, close and client probes as target content
The content-free exception set was audited against successful results
only. Now that failures carry the marker too, every exception was
re-audited against its error paths with one question: can the result
include text produced by an external process, provider or device, or
stored content REA did not author?
Removed from TARGET_CONTENT_FREE_TOOLS:
- close_binary: a failed provider cleanup (GhidraClient cleanup error via
createGhidraDiagnostics) projects the launcher's captured stdout and
stderr, which can contain target-derived log text.
- binary_session: capability status reports runtime unavailability as the
provider error message (ProviderAdapterError and similar), and the
provider candidate and Android availability diagnostics include probe
output such as the apktool stderr excerpt.
- inspect_adb_client: the success result carries the version and
installed path parsed from `adb version` stdout, and a failed probe
keeps a stderr_excerpt.
- inspect_apktool_client: the success result carries the first line of
the launcher's `--version` output, and a failed probe keeps a
stderr_excerpt.
- update_unknown: the new revision carries forward the stored question
and domain, which an imported evidence bundle can supply; list_unknowns
and get_evidence_bundle already return that data marked.
The set now holds only export_evidence_bundle (path, byte and record
counts; failures are filesystem, cancellation, or a fixed integrity
message) and record_unknown (the caller's input plus REA-computed ids
and digests; registry errors project only a reason code).
The conformance test pins the reduced set and asserts each removed tool
is marked. Tests that used binary_session as the unmarked example now use
record_unknown. docs/mcp-contracts.md lists the remaining exceptions and
why the removed tools stay marked.
* fix(contracts): include target-content policy in the catalog identity
createToolCatalog() and CATALOG_IDENTITY.tools projected every
client-visible contract field except returnsTargetContent. Because the
untrusted-content framing left schemas, effects, prompts, CLI names and
the package version unchanged, a same-version MCP process started before
the update digested equal to the new server and binary_session reported
it aligned although it never emits _meta.provenance.
Project the flag as returns_target_content in both the digested tool
catalog and CATALOG_IDENTITY.tools, and export createCatalogDigests so a
boundary test proves that toggling the flag on one tool changes the
tools and combined digests.
Danish Rafie Ekaputra
AICodeLion
Saim Shafique
To3akaRin
The one -Niu
记得吃中饭
Nelglor
jhuang-tw
Mykola Dementiev
Guangqi Zhang
morluto
Kaoru0822
Olorin
Stefano Rizzello
Vinícius Andrade
Ysopwnable
Quentin Limo
Luis
Ilian Iliev
Lluís Ariño
nKOxxx