* feat(diagnostics): report a byte entropy profile for one local file
The diff adds the MCP tool inspect_byte_entropy and the CLI command
inspect-byte-entropy, both implemented by ByteEntropyService. One
absolute regular file is read offline into observed Evidence: file-wide
SHA-256 and Shannon entropy, then fixed blocks and adjacent runs for an
optional window, using a power-of-two block size from 256 to 65536
(default 4096), with high at 7.2 or above and low at 1.0 or below. A
window past EOF, a missing path, or a non-file is an input error; a size
or identity change during the read is an artifact-changed error; empty
input reports entropy 0; cancellation is honored before and during the
read. Summary detail retains the complete series and returns file
entropy, counts, and the ten longest runs for inspect_analysis_view,
while the registered effect remains read-only Evidence and the
diagnostics docs record the operation and its errors.
* fix(diagnostics): flag zero-size files as not shown empty
Generated files such as Linux /proc entries report size 0 while holding
content, so inspect_byte_entropy read nothing and returned an empty-looking
report. A file that reports size 0 now carries a limitation that the result
does not show the file is empty. The result stays a success because the
tool reads the declared size, which the file-stability checks depend on.
* fix(mcp): reject unsupported diagnostic summaries before inspection
* fix(ci): repair integrated checks and Hopper test baseline
* test(entropy): use shared MCP schema verification after merge
* fix(catalog): list inspect_byte_entropy under its own provider
Adding inspect_byte_entropy to BINARY_DIAGNOSTICS_TOOL_CONTRACTS put it
under pwntools-elf in the generated product catalog, because the
pwntools group took every binary diagnostics tool except
inspect_pe_resources. The service reports provider byte-entropy.
Add a BYTE_ENTROPY_PROVIDER identity beside PE_RESOURCES_PROVIDER, use
it in ByteEntropyService, give it its own catalog group, and name the
pwntools tools (inspect_binary_layout, inspect_recorded_crash)
explicitly so a new diagnostics tool is not attributed to pwntools by
default. The catalog test pins both groups.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KFTE4BHbGYzgf8Uhzcppdh
* fix(cli): resolve the byte-entropy path with absoluteCallerPath
main replaced the node:path resolve import in binaryDiagnosticsCommands.ts
with absoluteCallerPath (#4067), so the byte-entropy command's
resolve(args.path) no longer type-checks once the PR is merged with main.
Use absoluteCallerPath like the other commands, and add
src/filesystem/callerPath.ts matching main so the branch builds on its own.
* fix(diagnostics): retain entropy read cleanup ownership
* docs(diagnostics): explain entropy output capacity
* fix(diagnostics): preserve entropy cancellation outcomes
* fix(diagnostics): close the entropy input and warn on large reports
Public inspect_byte_entropy input now uses publicInputStrictObject before
detail is added, so unknown keys stay rejected on the advertised schema.
The tool description warns that complete output can become very large and
points callers to a shorter length or a larger block_size.
Co-authored-by: morluto <morluto@users.noreply.github.com>
* fix(diagnostics): keep byte-entropy registration mergeable with main
Register inspect_byte_entropy beside layout inspection and reject
unretained summaries before the file read, without editing workflows.
Co-authored-by: morluto <morluto@users.noreply.github.com>
* fix(diagnostics): register byte entropy with the shared tool options
Use toolRegistrationOptions so the handler typechecks on this branch and
on main, and still refuse unretained summaries before the file read.
Co-authored-by: morluto <morluto@users.noreply.github.com>
---------
Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>
Co-authored-by: morluto <76467478+morluto@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: morluto <morluto@users.noreply.github.com>
* feat(environment): render more read tools from the state and describe responses
Keep main's environment serving and fold in coverage for non-entity reads,
response summaries, and MCP tool annotations.
Squashed rebase of #5352 onto main.
* fix(environment): keep rendered documents and quoted samples inside their budgets
A recorded body of exponential numbers can render past 256 KiB and turn a
working read into a state size error, so that document stays a recorded replay.
Description samples are measured after JSON quoting, and the replay-fallback
catalog check uses a tool count the longer response descriptions still separate.
Co-authored-by: morluto <morluto@users.noreply.github.com>
* feat(environment): apply evidence-gated write effects to the served state
Serve write tools whose later reads confirm the effect, and keep main's
state-file ownership and part-wise reply measurement.
Squashed rebase of #5600 onto the rebased #5352 tip.
* fix(environment): bound and validate write state history
* feat(environment): render more read tools from the state and describe responses
Keep main's environment serving and fold in coverage for non-entity reads,
response summaries, and MCP tool annotations.
Squashed rebase of #5352 onto main.
* fix(environment): keep rendered documents and quoted samples inside their budgets
A recorded body of exponential numbers can render past 256 KiB and turn a
working read into a state size error, so that document stays a recorded replay.
Description samples are measured after JSON quoting, and the replay-fallback
catalog check uses a tool count the longer response descriptions still separate.
Co-authored-by: morluto <morluto@users.noreply.github.com>
* feat(environment): apply evidence-gated write effects to the served state
Serve write tools whose later reads confirm the effect, and keep main's
state-file ownership and part-wise reply measurement.
Squashed rebase of #5600 onto the rebased #5352 tip.
* fix(environment): bound and validate write state history
* fix(environment): bound additional collection history
* fix(environment): credit sent fields, negative sequences, and encoded targets
A later overwrite copied surviving derived fields into the sent-member
claim, so a members-only check rejected an update whose sent field still
matched. Serving floored sequences at zero and stored missing history as
zero, so a collection sequence learned from negative ranks skipped the
next value. An absence after a delete forgot the resolved collection, and
an encoded read could not observe an update spelled with another
collection encoding.
Regression tests cover each case.
Co-authored-by: morluto <morluto@users.noreply.github.com>
* test(environment): cover resolved encoded update identities
* fix: validate timed update observations
* fix(environment): share write overlap and check clocks on earlier calls
Observation search skips a read begun before the write ended. That
comparison now uses the same overlap as the check path.
A clock learned on a later call was not checked against earlier calls
that kept the field. Serving would stamp those calls too, so the plan
stays unserved. A clock learned earlier is still refused when a later
call does not advance it.
Regression tests cover an overlapping empty update and a published_at
clock that one recorded call leaves unchanged.
Co-authored-by: morluto <morluto@users.noreply.github.com>
* fix: preserve partial update evidence
* fix(environment): let an unmodelled write explain only the list members that differ
A failed whole-list read now reports which members alone account for the
difference. An unmodelled write excuses the read's writes only when it came
after each such member's last check and pending writes, so an older change
elsewhere no longer hides a newer write the list shows undone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 7b6aab9d7191e600051693f2c192afcfc743d083)
* fix(environment): explain added fields, tie creates to serving's collection, and bound write descriptions
An update explains a field the entity lacked before when the observing
view had shown the entity, so a server-filled field is not dropped. A
create is tied to the held collection serving picks for its path. Write
descriptions and body refusals quote member names and sample them within
a byte bound. The request body is indexed once per explanation, so copied
and slugged fields no longer rescan it.
Replays 07aca54e onto the restacked branch. Its rule that a read begun
before a write ended teaches nothing is already here (fb6785dc), so only
its test is kept.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(environment): state how list reads weigh unmodelled writes, added update fields and quoted members
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* refactor(environment): keep the observing view helpers with the observations
shownBefore and the item-or-list fields check move to the observe module,
keeping the write effects within the line limit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(environment): resolve unmodelled writes onto the held collection
An unmodelled path kept the request's collection spelling. An archive under
the plain spelling of a collection held encoded marked only that plain
collection unknown, so a later update was blamed for the archive's fields
and dropped. Name that entity the way a modelled write does.
Co-authored-by: morluto <morluto@users.noreply.github.com>
* fix(environment): confirm no write through volatile fields or overwritten replacements, and bound the served clock
A members-only check no longer compares volatile fields, which render
from the read itself. A document replacement credits the members it
wrote or removed at its object, so a later merge that overwrites them
takes its confirmation away. A recorded time past the range of a date
gives no served clock instead of throwing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(environment): leave unknown entities unchecked by failed reads, volatile fields and unchecked calls
A failed item read that is no absence leaves an unknown entity unknown, so
a later usable read still checks its writes. A read showing an unknown
entity again checks no write on a volatile field. An update of an unknown
entity takes its recorded outcome and adds nothing to a built plan. A later
update that does not answer with its entity is observed only by a read
whose view shows every learned clock; without one it is unmodelled.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(environment): fill a create's path fields from the collection serving selects
The gate built a later create's path fields from the recorded request path,
while serving fills them from the held collection it selects for that path,
so a read showing the other spelling's parent could confirm a create serving
cannot reproduce. Rules still learn from the recorded path.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(environment): name unmodelled writes as modelled ones, count fields updates write, and keep later members' types
An unmodelled write below an item path names its entity through itemTarget,
and one at or above a collection resolves the held collection, as modelled
writes do, so an encoded spelling no longer hides the entity it changed.
An update's answer counts the fields its rules write, other than increments,
as available, so a field only the update writes no longer makes it
incomplete-entities. A member a later call first stores keeps that call's
JSON type in the body's kinds, which the body check falls back to.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(environment): let a document replacement overwrite everything under its object
A replacement claims only the members it wrote or removed, but it now
overwrites every earlier effect under its object, so a merge that wrote
the object it replaces loses its confirmation.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(environment): start the served clock from the records whose times a date can hold
A record timed past the range of a date no longer leaves the served clock
without a start, which the store read as zero and stamped writes in 1970;
the clock starts after the latest record a date can hold.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: morluto <morluto@users.noreply.github.com>
Co-authored-by: morluto <76467478+morluto@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Authenticated capture Evidence whose operation is supported but whose
normalized_result fails its capture schema fell through to a generic
evidence.operation refusal, discarding the schema issues. Each recognized
operation now refuses with the projected schema issues beneath
evidence.normalized_result and retains the schema failure as the cause;
only an operation no capture schema owns blames evidence.operation.
The view Evidence parser had the same conflation: a wrong operation, a
wrong kind and a malformed payload of the requested kind all blamed
normalized_result.kind. It now reports role.operation, the kind, or the
payload's own field paths respectively.
Fixes#6119
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Electron and process scenario path admission caught every realpath or
stat failure as an invalid_value input issue, so a correct path beneath a
directory without search permission returned invalid_request with
schema-correction recovery instead of a host access denial.
Admission now classifies EACCES/EPERM from path resolution and metadata
reads as the existing AnalysisAccessDeniedError (selected path,
system_code, filesystem-read boundary). Missing, looping or otherwise
unresolvable paths and reachable files without execute permission keep
their input-error semantics. Electron reports invalid selections before a
denial on another field. Process PATH search still continues past a
denied directory; only an explicitly selected executable reports denial.
Fixes#6117
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Construct integer formats directly while preserving bounds, refinements, metadata and current shared schema owners. Keep Incur-dependent CLI numeric coercion declarations unchanged.
Verify canonical advertisements, 126 numeric boundary cases, 974 actual CLI parser cases, broad source tests and real public-artifact SDK workflows. Record 72 normal tasks and 12 separate owned-inspector heap diagnostics with whole-family resource bounds; distinguish SDK spans and asymmetric audit overhead. Generate numerical report sections from receipts and reject stale tables through documentation conformance checks.
Closes#6236. Current-head CI and Greptile pass; actionable review findings are addressed. Codex review was unavailable due repository usage limits.
* fix(ci): run the lockfile-pinned Skills CLI in package verification
The repository skill stage fetched skills@1.7.2 and its unlocked
dependencies from the npm registry on every run, taking 5-120 s and
intermittently exceeding the 120 s stage timeout on main and PRs.
Pin skills as a devDependency and run its locked bin directly.
Fixes#6234
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(ci): lock the Skills CLI outside the root dependency tree
skills 1.6+ requires Node >=22.20.0, so pinning it as a root
devDependency made `npm ci` fail under engine-strict on the supported
Node 22.19.x. Keep the root install unchanged and give the verifier its
own package.json and lockfile, installed with `npm ci --prefer-offline`
in the verification workspace. The lockfile still removes all registry
metadata resolution and pins every transitive dependency.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(native): let callers raise the Apple dispatch metadata input budget
inspect_native_dispatch_metadata refused any Mach-O over a fixed 64 MiB
before decoding, and its recovery said there was no override. Most large
macOS binaries exceed it.
max_input_bytes (CLI --max-input-bytes) now selects the budget, default
64 MiB and at most 1 GiB, through the shared input contract, the enhanced
workflow and the native provider. Results record input_admission with the
effective budget, its source and the retained bytes. A refusal names the
observed size, the budget and its source, and the budget to select or, past
1 GiB, a deep analysis provider.
Fixes#3917
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): describe --max-input-bytes in inspect-native-dispatch-metadata help
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(native): cover max_input_bytes forwarding through CLI and MCP
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(pcap): report workspace allocation failures by their system cause
Pcap workspace allocation and snapshot writes mapped every filesystem
error, including ENOTDIR and EACCES from mkdtemp, to a file-size resource
constraint with null limits and worker file-size advice. Flutter's owned
snapshot storage had the same errno-blind mapping, while TShark and sigrok
carried duplicate copies of the correct split.
Move the private-workspace failure projection into one shared owner:
ENOSPC/EDQUOT/EFBIG stay file-size resource constraints with their system
code and path; descriptor exhaustion and every other failure become
provider workspace failures carrying stage, workspace path, failing path,
system code and cause-specific recovery.
Fixes#6123
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(flutter): locate owned storage failures by their workspace
mapOwnedIoFailure always passed a null workspace path, so a descriptor
write or read failure without its own errno path lost every location.
Thread the owned root through: allocation keeps null (no workspace yet),
snapshot failures use the allocated root, and package-reader system
errors propagate to the request owner, which maps them against the
snapshot root instead of mapping early without it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Authorize CDP targets with the caller's operation instead of a fixed
inspect_web_page label, and reuse the authorization owner's derived
origin scope. When a browser observation failure is attributed to a
different operation, keep its producer detail, remediation, cleanup,
diagnostics, partial observation and original cause rather than
rebuilding it from the reason alone. Both the CDP browser and CDP
Electron providers use the same domain-owned relabel.
Fixes#6122
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(process): record withheld unrelated Linux processes as unverified
The token sweep after an owned command (for example the Ghidra Java probe
run by setup and doctor) failed cleanup when an unrelated same-UID process
started during the sweep and Linux denied its /proc/<pid>/environ. Setgid or
setuid programs (mismatched real/saved IDs) and non-dumpable processes are
withheld by the kernel exactly like another account's processes, which were
already recorded as unverified.
Derive the kernel restriction from the same stat-bracketed procfs read that
already supplies the effective UID: all Uid/Gid IDs from status, and the
root ownership procfs reports for a non-dumpable process's environ. Keep that
fact through exit settlement. Related candidates still fail closed.
Fixes#6204
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(process): define caller credentials on hosts without geteuid
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(native): inspect a universal Mach-O slice in Go and Rust metadata readers
inspect_go_binary and inspect_rust_binary refused universal Mach-O files and
offered no slice selection, so hosts without lipo could not inspect them.
Both now accept architecture (x86, x86_64, arm, arm64). One selector reads
REA's own FAT table, validates the slice range, alignment and header CPU,
and returns the thin image with its macho-selected-image-v1 identity. The
universal file stays the artifact; results add selected_slice and keep
every source range a file offset. Missing, absent, duplicate or misplaced
selections fail on architecture and list the slices present. Subject
format mach-o-universal now has one owner shared by the services and the
retained Go/Rust views.
Fixes#4102
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(native): reject damaged universal files and self-contradicting slice reports
A FAT table whose slice could not be read was treated as a non-universal
file, so callers were told to omit architecture. Its reason now fails on
path. Go and Rust reports with a selected slice must be Mach-O and name the
slice's CPU family, through one shared agreement check. The CLI
architecture flag is described in help.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(artifacts): make ASAR header admission configurable
ASAR header/index admission was a fixed 16 MiB with no override, so
operators with large Electron indexes could only extract the archive.
REA_ASAR_MAX_HEADER_BYTES now sets it (default 16777216, maximum
268435456). One owner resolves the effective admission for the
classification probe, inventory, ASAR reader and Electron file scope,
including the Electron header cache budget. Startup and doctor validate
the setting; admission refuses an invalid value instead of falling back.
Recovery and error details name the effective value and its source.
Fixes#6176
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(artifacts): resolve ASAR header admission from the caller's environment
The admission read process.env, so composition roots that select their own
environment (createDirectAnalysis, createServer, CLI, optional observation
factories) could not raise or lower it, and an ambient value leaked into
callers that left it unset.
Nothing reads process.env now. ArtifactProvider and extraction resolve the
admission from their environment; JavaScript analysis parses it from the
environment its server or CLI caller passes and refuses an invalid value;
the Electron and V8 providers receive it from their factories. Inventory,
nested ASAR readers, classification probes and Electron page authorization
carry that one value. A header cached by Electron under a larger admission
is refused for a smaller one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Summary delivery with --evidence_output wrote the bundle, then let the
command's close/cancellation check replace the result. A SIGINT/SIGTERM
during staging cleanup after commit returned typed cancellation without
the receipt, and a persistent staging cleanup failure after commit was
rewrapped as a provider cleanup error that dropped published_path.
Every --evidence_output command now selects its view first and publishes
through publishEvidenceDelivery, the owner inspect-managed-members already
used. Post-commit cancellation keeps {path, bytes} under
details.diagnostics.evidence_bundle; incomplete cleanup keeps
details.published_path and bytes; a cleanup retry that succeeds returns
the normal receipt; pre-commit cancellation creates nothing and names
the command's operation.
Fixes#6095
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A captured script whose bytes matched no static digest was reported as
high-confidence unmatched even when no automatic or caller-declared
mapping applied to its location in any static layer. The matcher now
computes mapping presence once while locating candidates and shares one
no-candidate classification between the captured-source and source-free
branches, so an unmapped location stays unknown with
no-authorized-location-mapping and is not counted as unmatched.
Fixes#6061.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The language catalog kept failed .ldefs closes in a module-global set that
only a later catalog read retried, and the provider returned a cleanup error
bound to no owner. Registry admission retries provider close before profile
resolution, so the retained handle was never retried and the original error
stayed incomplete forever.
Read definitions through the provider's caller-owned inspection scope with the
shared regular-file owner used by the seed and extension readers. A failed
close is retained as an exact lease in that scope, provider close and the next
catalog admission retry it, and the translated error is bound to the lease so
confirmed release reopens the same registry.
Fixes#6133
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(process): settle owned cleanup with the launcher's own exit proof
When owned process-group verification fails while the supervised
launcher is still mid-exit, the supervisor now waits up to the kill grace
for its own child-exit observation and verifies once more after it
arrives. Verification is not repeated when the exit was already observed
before cleanup started, so there is no new evidence and the original
failure is kept. Nothing is signaled based on the wait.
Fixes#6134
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(process): bound Hopper session cleanup attempts by exit evidence
The fake Hopper launcher exits after its shutdown acknowledgement. When
the supervisor first observes that exit during a failed verification, it
verifies once more, so the first close makes one or two attempts depending
on timing. The test now asserts that bound. Later session retries, made
after the exit was already observed, make exactly one attempt each.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
