Are you the author? Sign in to claim
A deterministic crash diagnosis engine for .ips reports (CLI + MCP server)
A crash diagnosis engine for Apple platforms: parses .ips crash reports, symbolicates
them against dSYMs, and produces an evidence-cited, ranked diagnosis, not just a
symbolicated stack trace.
Ships as a dependency-free Swift library (CrashDXCore imports only Foundation), a CLI
(crashdx), and an MCP server (crashdx-mcp), so agents and humans use the same engine.
A symbolicated stack trace tells you where a process died, not why. crashdx adds a second
stage on top of symbolication: deterministic, rule-based evidence extraction (watchdog
budgets, jetsam tables, register and memory state, lastExceptionBacktrace/asi) feeding
a ranked set of competing hypotheses, each citing the specific facts that support it and
pointing back into the raw report.
Two properties it holds onto deliberately:
There are no LLM calls in the engine: it produces verifiable facts and ranked interpretations, and leaves the narrative to whatever consumes them. See docs/DESIGN.md for the full architecture.
Everything below is real, unedited output from fixtures in this repo (elisions are marked).
Paths are relative to the repo root, and the --dsym flags point at dSYMs the repo ships,
so these reproduce as-is; use swift run crashdx ... if you have not installed the binary.
$ crashdx analyze Tests/CrashDXCoreTests/Fixtures/nsexcrash.ips \
--dsym Tests/CrashDXCoreTests/Fixtures/nsexcrash.dSYM
process: nsexcrash
bug_type: 309
os: macOS 26.3.1 (25D2128)
exception: EXC_CRASH (SIGABRT)
terminated: Abort trap: 6
faulting thread (15 frames):
libsystem_kernel.dylib __pthread_kill
libsystem_pthread.dylib pthread_kill
libsystem_c.dylib abort
libc++abi.dylib __abort_message
libc++abi.dylib demangling_terminate_handler()
libobjc.A.dylib _objc_terminate()
libc++abi.dylib std::__terminate(void (*)())
libc++abi.dylib __cxxabiv1::failed_throw(__cxxabiv1::__cxa_exception*)
libc++abi.dylib __cxa_throw
libobjc.A.dylib objc_exception_throw
CoreFoundation -[NSException raise]
nsexcrash throwingHelper() (main.swift:8)
nsexcrash doWork() (main.swift:12)
nsexcrash main (main.swift:15)
dyld start
last exception backtrace (7 frames):
CoreFoundation __exceptionPreprocess
libobjc.A.dylib objc_exception_throw
CoreFoundation -[NSException raise]
nsexcrash throwingHelper() (main.swift:8)
nsexcrash doWork() (main.swift:12)
nsexcrash main (main.swift:15)
dyld start
symbolication (engine: crashSymbolicator):
nsexcrash symbolicated
libsystem_kernel.dylib no_dsym
libsystem_pthread.dylib no_dsym
...
DIAGNOSIS: Uncaught Objective-C exception (NSException) (strong, score 5)
A lastExceptionBacktrace is present and an objc_exception_throw frame appears in it (or on the
faulting thread) — together these are pathognomonic for an uncaught NSException: `-raise` walked up
through objc_exception_throw and was never caught, so the runtime called terminate and aborted. The
asi "Terminating app due to uncaught exception" message is only corroboration when present; plain
CLI/Foundation processes never write it, so its absence is not evidence against this hypothesis.
evidence: lastExceptionBacktrace is present with 7 frame(s); lastExceptionBacktrace frame 1 matches
sentinel 'objc-exception-throw': objc_exception_throw; Faulting thread frame 9 matches
sentinel 'objc-exception-throw': objc_exception_throw
inspect: nsexcrash throwingHelper() (main.swift:8)
confirm: Symbolicate the app frames in lastExceptionBacktrace to find the throw site; Check the
asi/console log for the exception name and reason, if available
also considered: cxx-terminate (moderate, 2), abort-generic (weak, 1)
The evidence line cites the supporting facts the rule scored on, up to four, with a
running total when there are more; each fact carries a path back into the original report,
which --json --tier standard exposes. also considered is the full remainder of the
ranked list, so you can see what the verdict beat and by how much.
When the leading hypothesis is not strongly supported and clearly ahead, you get the candidates instead of a label:
$ crashdx analyze Tests/CrashDXCoreTests/Fixtures/synthetic/wild-address.ips \
--dsym corpus/fixtures/nullderef/nullderef.dSYM
process: synthetic-wild-address
bug_type: 309
os: macOS 26.3.1 (25D2128)
exception: EXC_BAD_ACCESS (SIGSEGV)
terminated: Segmentation fault: 11
faulting thread (4 frames):
nullderef readThroughNullPointer() (main.swift:9)
nullderef run() (main.swift:13)
nullderef main (main.swift:16)
dyld start
symbolication (engine: crashSymbolicator):
nullderef symbolicated
dyld no_dsym
DIAGNOSIS: INCONCLUSIVE — competing hypotheses:
1. Wild pointer or use-after-free [wild-or-uaf-address] (moderate, score 3)
The faulting address is outside the null page, and vmregioninfo reports it is not inside any known
VM region at all — consistent with either a WILD pointer (an uninitialized or garbage value used as
an address) or a USE-AFTER-FREE into memory the OS has since unmapped. This is deliberately a
lower-confidence hypothesis than null-dereference: absence of region information doesn't distinguish
between these two causes.
evidence: Faulting address: 0x1deadbeef; Exception type: EXC_BAD_ACCESS
inspect: nullderef readThroughNullPointer() (main.swift:9)
confirm: Re-run with Address Sanitizer enabled to catch the use-after-free at the free/access site;
Enable NSZombies (Malloc Scribble/Guard Malloc) to turn this into an immediate,
informative crash; Profile with Instruments' Allocations tool, recording reference counts
The same shape applies across the rule set. Verdict and evidence lines from four more fixtures, with the explanation and follow-up steps elided:
# synthetic/watchdog-scene-create.ips
DIAGNOSIS: Watchdog timeout (main thread stall) (strong, score 6)
evidence: Termination code: 2343432205 (0x8badf00d); Termination namespace: FRONTBOARD; Watchdog
event: scene-create, allowance 19.97s
# synthetic/jetsam-per-process-limit.ips
DIAGNOSIS: Jetsam memory kill (strong, score 5)
evidence: Exception subtype: RESOURCE_TYPE_MEMORY; Jetsam per-process-limit indicator:
per-process-limit 350808KB exceeds task limit 335872KB
# synthetic/stack-overflow.ips
DIAGNOSIS: Stack overflow (strong, score 5)
evidence: Faulting address 0x3000 is near the stack region: vmregioninfo names a STACK GUARD region;
within 500 bytes of sp (0x31f4); Faulting thread has 5 consecutive frames with identical
symbol 'recurse(_:)' starting at frame 0 — a recursion signature
# crashspike-unsymbolicated.ips, with --dsym Tests/CrashDXCoreTests/Fixtures/crashspike.dSYM
DIAGNOSIS: Swift runtime fatal trap (strong, score 6)
evidence: Exception type: EXC_BREAKPOINT; Signal: SIGTRAP; Faulting thread frame 0 matches sentinel
'assertion-failure': _assertionFailure(_:_:file:line:flags:)
Each of those also prints the full explanation, inspect point, and confirm steps shown
in the first example.
--json emits the same analysis as a structured AnalysisReport, which is what the MCP
server returns and what you would parse in CI. Below is the same crash as the first
example, abridged (/* ... */ marks elisions). The real output is a single minified line
with sorted keys, so pipe it through jq or python3 -m json.tool; it is pretty-printed
and reordered here for readability:
{
"schemaVersion": "0.2",
"tier": "summary",
"diagnosis": {
"status": "verdict",
"verdict": {
"id": "uncaught-objc-exception",
"title": "Uncaught Objective-C exception (NSException)",
"category": "objc-exception",
"explanation": "A lastExceptionBacktrace is present and an objc_exception_throw frame ...",
"supporting": [
{ "factID": "leb.present", "weight": 1 },
{ "factID": "leb.sentinel.objc-exception-throw", "weight": 3 },
{ "factID": "frames.sentinel.objc-exception-throw", "weight": 1 }
],
"contradicting": [],
"inspect": [
{ "frameIndex": 3, "leb": true, "symbol": "throwingHelper()",
"sourceFile": "main.swift", "sourceLine": 8 }
],
"confirmFurtherBy": [ /* ... */ ]
},
"hypotheses": [ /* all 3, each with its band + score */ ]
},
"event": {
"bugType": "309", "exceptionType": "EXC_CRASH", "signal": "SIGABRT",
"terminationIndicator": "Abort trap: 6", "terminationNamespace": "SIGNAL"
},
"faultingThread": { "index": 0, "queue": "com.apple.main-thread", "triggered": true,
"frames": [ /* ... */ ] },
"symbolication": { "engine": "crashSymbolicator", "images": [ /* per-image outcome + uuid */ ] },
"process": { "name": "nsexcrash", "osVersion": "macOS 26.3.1 (25D2128)",
"captureTime": "2026-07-23 00:22:01.2793 +0000" }
}
Note "leb": true on the inspect point: the throw site was recovered from
lastExceptionBacktrace rather than the faulting thread, and the report says so instead of
flattening the distinction.
--tier standard (see Report tiers) adds diagnosis.factsConsidered,
which resolves every factID to its human-readable statement and its sourcePath into the
raw report:
{ "id": "leb.sentinel.objc-exception-throw",
"statement": "lastExceptionBacktrace frame 1 matches sentinel 'objc-exception-throw': objc_exception_throw",
"sourcePath": "lastExceptionBacktrace[1]" }
sourcePath is not optional on a Fact, so every fact the diagnosis cites can be walked
back to the part of the report it was read from.
Committed golden snapshots of both
tiers live in Tests/CrashDXCoreTests/Fixtures/ (nsexcrash-summary-golden.json,
nullderef-standard-golden.json) if you want the complete, unabridged shape.
swift-tools-version:6.0, but the MCP
server's pinned dependencies declare up to 6.2, and SwiftPM resolves the whole
package graph, so 6.2 is the real floor for every target, including CrashDXCore.CrashSymbolicator.py from Xcode's CoreSymbolicationDT.framework, located via
xcode-select -p. Without it, crashdx falls back to atos, which resolves fewer
source locations. Parsing and diagnosis work either way./usr/bin/python3 and /usr/bin/dwarfdump (both provided by Xcode's command line
tools), used to run CrashSymbolicator.py and to verify dSYM UUIDs.Honest scope, because a crash diagnoser that quietly does worse on your platform is worse than one that says so:
x86_THREAD_STATE registers and the 4 KB page
size. There is no Intel fixture in the corpus, so this path is reasoned-about rather
than exercised against a real report.bug_type 309/109). Jetsam event reports (JetsamEvent-*.ips,
bug_type 298) and hang/stackshot reports (bug_type 288) use a different payload
shape with no threads or exception object; crashdx parses them without error but has
no facts to work from and will report inconclusive.crashdx runs entirely on your machine and makes no network calls of any kind. Crash
reports, dSYMs, and everything derived from them stay local. This matters because .ips
files contain identifying data (crashReporterKey, boot/sleep-wake UUIDs, device model,
and usernames in paths), so if you attach one to a bug report, scrub it first.
git clone https://github.com/r00tify/crashdx.git
cd crashdx
swift build -c release
The binaries land in .build/release/. To put them on your PATH:
sudo mkdir -p /usr/local/bin
sudo cp .build/release/crashdx .build/release/crashdx-mcp /usr/local/bin/
(Or copy them somewhere you already own, such as ~/.local/bin.)
Crash reports live in ~/Library/Logs/DiagnosticReports/ on macOS (or Console.app →
Crash Reports); on iOS, Settings → Privacy & Security → Analytics & Improvements →
Analytics Data, and Xcode's Organizer for TestFlight/App Store crashes.
crashdx analyze report.ips # human-readable summary + diagnosis
crashdx analyze report.ips --json # structured AnalysisReport JSON
crashdx analyze report.ips --json --tier full # ...with every thread included
crashdx symbolicate report.ips --dsym path/to.dSYM # just symbolicate, print enriched .ips
crashdx --version
--dsym accepts a .dSYM bundle, an .xcarchive, or a directory to search recursively
(repeatable), and --no-spotlight skips Spotlight-based discovery while --no-archives
skips Xcode's archive directory. Run crashdx <subcommand> --help for the full option
list. Note that --json and --tier apply to analyze only; symbolicate always emits
a complete .ips.
Both subcommands search Spotlight and Xcode's archives for a matching dSYM automatically.
For foreign reports (user-submitted, TestFlight, CI artifacts) pass the build's dSYM
explicitly with --dsym.
Strings taken from the report (process name, symbols, paths, exception text) are escaped
before the human-readable summary prints them: C0 controls, DEL, and bidi overrides
render as \x0A / \u{202E} instead of being emitted, so a crafted report cannot forge
a DIAGNOSIS: line or repaint your terminal. --json and symbolicate emit the
report's strings verbatim; treat their output as data, not as terminal text.
A dSYM from a different build is refused rather than used: a stale dSYM produces
confidently wrong symbols, which is worse than none. That case is reported as
uuid_mismatch (with the offending path in reason), kept distinct from no_dsym
because the fix is different: find the archive matching the report's build UUID, rather
than go looking for a file you may already have.
--tier controls how much of the report is emitted; it never changes the diagnosis,
which is always computed from the full report:
| Tier | Contents |
|---|---|
summary (default) | Faulting thread (capped at 15 frames), lastExceptionBacktrace, and the diagnosis verdict + ranked hypotheses |
standard | Lifts the frame cap, and adds the binary images, other threads that contain app frames, and factsConsidered (the evidence behind each hypothesis) |
full | Adds every remaining thread, including those with no app frames |
0 success · 64 usage error · 65 the input could not be parsed as an .ips report,
or symbolication failed · 66 file not found or unreadable.
Diagnostics go to stderr; stdout carries report output (and --help/--version).
crashdx-mcp exposes the same pipeline over stdio (newline-delimited JSON-RPC) as the
crashdx_analyze and crashdx_symbolicate tools. With Claude Code:
claude mcp add crashdx /usr/local/bin/crashdx-mcp
A matching crash-triage Agent Skill ships in .claude/skills/crash-triage/, covering
how to read the diagnosis and the interpretation pitfalls specific to each crash family.
Add https://github.com/r00tify/crashdx.git to your Package.swift, depend on the
CrashDXCore product, and call
AnalyzePipeline.analyze(path:tier:dsymPaths:useSpotlight:searchArchives:), the same
entry point both binaries use.
CrashDXCore imports nothing but Foundation, but two limitations are worth knowing
before you depend on it:
atos/dwarfdump/mdfind through
Foundation.Process, which does not exist on iOS. Adding this package to an iOS target
fails during dependency resolution, before anything compiles.Package.resolved even if you never import them.AnalysisReport and the diagnosis model are Sendable, so analysing a directory of
reports concurrently works as you would expect.
Sources/CrashDXCore/ Parsing, symbolication, and the diagnosis engine (imports only Foundation)
Sources/crashdx/ CLI
Sources/crashdx-mcp/ MCP server (depends on swift-sdk)
Tests/ Unit + integration tests, golden snapshots, crash fixtures
corpus/ Ground-truth crash fixtures and their verified findings
docs/DESIGN.md Diagnosis engine architecture
Scripts/ Fixture scrubbing and the CI privacy guard
swift test
See CONTRIBUTING.md. One rule matters more than the rest: crash
reports and dSYMs must be scrubbed before they are committed. They carry device
identifiers and your username. Scripts/scrub-fixture.py handles .ips files and
Scripts/check-fixtures-scrubbed.sh (also run in CI) will tell you if anything slipped
through.
crashdx makes no network calls; nothing leaves your machine. See SECURITY.md
for vulnerability reporting and for what a .ips file actually contains before you share
one.
MIT. See LICENSE.
Run analytics queries on ClickHouse — explore schemas, execute SQL, fetch results
Run Claude Code as an MCP server so any agent can delegate coding tasks to it
Browser automation using accessibility snapshots instead of screenshots
Google's universal MCP server supporting PostgreSQL, MySQL, MongoDB, Redis, and 10+ databases