Closes a 6-item field report from an RE engineer driving LDB against a 503 MB iOS arm64 Mach-O, plus the phase-3/phase-4 hardening cycle that surfaced from deep code review of each fix.

Field-report items closed

Highlights

Performance

• target.open on the canonical fixture: 21,913 → 1,712 tokens (12.8× shrink).
• Full agent workflow: 51,280 → 32,627 tokens (−36%).

New endpoints / surface

• ldbd --listen unix:PATH + ldb --socket PATH (single-client phase 1 + multi-client phase 2).
• daemon.shutdown RPC.
• --listen-idle-timeout N opt-in idle exit.
• xref.address + string.xref now return provenance.warnings documenting heuristic skips.
• view.include_sections on target.open.

Correctness

• ARM64e chained-fixup parser (formats 1, 6, 9, 12, auth-rebase) — phase 1.
• Mach-O loader (thin + FAT + fat_arch_64 triple-aware) — phases 1, 2, 4.
• ADRP-pair resolver with full clobber semantics: function-boundary reset, AAPCS64 + PAC call clobber, ADD/SUB/MOV/CSEL/LDP destination clobbering, pre/post-indexed LDR writeback — phases 2, 3, 4.
• TSan-confirmed UAF in NonStopRuntime fixed via shared_ptr-owned notification sinks.
• Dispatcher::dispatch_mu_ recursive mutex serialising multi-client dispatch.

Security

• Socket inode 0600; parent dir 0700; O_NOFOLLOW lockfile; lstat parent-dir symlink check; getpeereid / SO_PEERCRED peer-uid rejection; SO_RCVTIMEO + SO_SNDTIMEO; absolute --listen paths required.

Process notes

Five branches were developed and reviewed via a deliberate fan-out pattern:

• Each item got an implementation agent (worktree-isolated).
• Each branch went through linus-code-reviewer (and security-auditor for the socket exposure; opus-model linus with explicit max-effort framing on the deepest changes).
• Each review produced a precise punch list addressed by a cleanup agent.

The opus reviewer on phase 2 caught a TSan-confirmed use-after-free in the notification fan-out path. The opus reviewer on phase 3 caught the SUB / PAC / writeback / STR family of false-positives the original whitelist of clobber-source mnemonics missed. The opus reviewer on phase 4 caught the CSEL / LDP regression class and demanded the architectural shift from whitelist to clobber-by-default destination tracking — which is what shipped.

Phase-5 deferrals (enhancement, not fix)

Documented in docs/35-field-report-followups.md:

• §2 phase 3: server-side target_id-aware notification routing, per-connection dispatch parallelism, workers list reaping.
• §6 phase 5: full bind imports-table walk, function_starts backward boundary detection, complete clobber-by-default audit across remaining ARM64 instructions, indirect-dispatch entry points, real iOS .ipa CI smoke.

Commits

Per-PR shortlog:

• #20 (merged 2026-05-16): field-report follow-ups — target.open lazy load + §1 CLI + §2 socket daemon + §6 chained-fixups phases 1-3
• #21 (merged 2026-05-16): §2 multi-client socket daemon + §6 ARM64e xref coverage extension + CI fix

🤖 Generated with Claude Code

Patch release: re-analyze ships as a proper Claude Code Skill

v1.6.0 shipped the re-analyze skill as a flat skills/re-analyze.md file
without YAML frontmatter. Claude Code's skill catalog reads the frontmatter to
populate name / description / argument-hint / allowed-tools, so without
it the skill description rendered as a truncated slice of the prose body
("You are an expert reverse engineer…") and the skill couldn't auto-discover
when installed under ~/.claude/skills/.

Changes

• Moved skills/re-analyze.md → skills/re-analyze/SKILL.md (standard
  Claude Code Skill folder layout).
• Added YAML frontmatter: name, description, argument-hint,
  allowed-tools: Bash, Read, Write, Grep, Glob, Agent, effort: high.
• Rewrote skills/README.md for the new install path (clone+symlink into
  ~/.claude/skills/<name>/ is now the recommended install).
• Bumped CMakeLists.txt to 1.6.1.

Install (new path)

git clone https://github.com/zachgenius/LDB ~/ldb
mkdir -p ~/.claude/skills
ln -s ~/ldb/skills/re-analyze ~/.claude/skills/re-analyze

Claude Code will auto-discover the skill on next session start. Body content of
the skill is unchanged from v1.6.0 — full 5-phase RE workflow (static
orientation → dynamic probes + tracepoints + predicates → network / BPF / perf
observers → artifact capture → session record / replay / export).

Verified

• ctest — 68/68 passing on macOS arm64.
• ldbd --version reports 1.6.1.

Upgrading from v1.6.0

If you previously copied the skill into ~/.claude/commands/re-analyze.md,
remove it and follow the install snippet above. The slash-command path no
longer carries the latest skill body.

🤖 Generated with Claude Code

v1.6.0 closes the V1 planned arc (v1.3 → v1.4 → v1.5 → v1.6 per docs/17-version-plan.md). The protocol contract is stable; everything in the wire surface remains additive.

Highlights

Own GDB Remote Serial Protocol transport

• target.connect_remote_rsp opens a native TCP connection to any gdb-remote server (lldb-server gdbserver, gdbserver, QEMU, OpenOCD, rr) — parallel path to LLDB's gdb-remote plugin
• Async reader thread with two-mutex split (write_mu_ / byte_mu_) — see docs/25-own-rsp-client.md
• vCont write path for thread.continue / thread.suspend against RSP-backed targets

Non-stop runtime

• Per-thread state machine + monotonic stop_event_seq per target — docs/26-nonstop-runtime.md
• Listener thread polls RspChannels, fires thread.event{kind:"stopped"} push notifications — docs/27-nonstop-listener.md
• thread.list_state query, thread.suspend honoured for both RSP-backed (via vCont;t) and LLDB-backed (via SBThread::Suspend()) targets
• hello.capabilities.non_stop_runtime: true for agent negotiation
• JSON-RPC 2.0 §4.1 notification frames over the same stdio channel (stream-locked OutputChannel)

Agent-expression VM + predicates

• Stack-based bytecode VM, signed int64, ~25 opcodes including the new if_goto / goto (post-#25 phase-3) — docs/28-agent-expressions.md
• S-expression DSL ((eq (reg "rax") (const 42))) compiles in a single pass — docs/29-predicate-compiler.md
• New predicate.compile endpoint returns {bytecode_b64, mnemonics, reg_table} for agent pre-flight
• Anti-DoS caps: 4 KiB program / 64 stack depth / 10K instruction count

Tracepoints (no-stop probes)

• New tracepoint.* family (create / list / enable / disable / delete / frames) — docs/30-tracepoints.md
• Rate-limit grammar: <int>/{s,ms,us,total} enforced at the orchestrator (fixed-pivot window; total = lifetime cap)
• Predicate filtering reuses the agent-expression VM; per-probe predicate_dropped / predicate_errored / rate_limited counters
• QTDP / QTStart / QTStop / QTFrame wire vocabulary added to src/transport/rsp/packets.{h,cpp} — docs/34-tracepoints-in-target.md. Orchestrator integration is phase-2.5.

Other refinements

• Typed backend::NotImplementedError subclass; dispatcher catches by type rather than substring-matching error messages
• SOCK_CLOEXEC Darwin portability fix in the RSP socket open path
• Design notes for deferred items: own ptrace (docs/31), hardware tracing (docs/32), criu (docs/33)
• Skill skills/re-analyze.md rewritten to match the V1 endpoint surface (fixed 8 stale names)
• README simplified (488 → 289 lines)

V1 surface at a glance

Validation

• CI green on Linux x86-64, Linux arm64, macOS arm64, plus the opt-in Capstone leg (gh run 25732148368)
• ctest: 70/70 passing with SKIP-gated live-attach tests when kernel.yama.ptrace_scope=0 or with CAP_SYS_PTRACE
• Build warning-clean under -Wall -Wextra -Wpedantic -Wshadow -Wnon-virtual-dtor -Wold-style-cast -Wcast-align -Wunused -Woverloaded-virtual -Wconversion -Wsign-conversion -Wnull-dereference -Wdouble-promotion -Wformat=2 -Wmisleading-indentation

Post-V1 / deferred

These are explicitly out of v1.6 and on the "demand-driven / opt-in" track:

• LldbBackend SBDebugger::SetAsync(true) flip — would drop the final sync-mode constraint on LLDB-side per-thread observation. Defer because of the test cascade.
• QTDP phase-2.5 — orchestrator integration on top of the wire vocabulary now in place. Tracepoints currently run daemon-side.
• #20 Own Linux ptrace driver, #22 hardware tracing (Intel PT / ARM ETM), #24 criu snapshot/fork — design notes shipped in docs/31-33. None scheduled.

Future v1.x releases will be additive patches; v2.0 is reserved for breaking wire changes per the V1 contract.

Install

git clone https://github.com/zachgenius/LDB ~/ldb
cd ~/ldb
cmake -B build -G Ninja -DLDB_LLDB_ROOT=/path/to/llvm-prefix
cmake --build build
ctest --test-dir build --output-on-failure
build/bin/ldbd --version   # → 1.6.0

See README.md and docs/00-README.md.

🤖 Generated with Claude Code

v1.1.0 — agent-RE dogfood fixes

Five fixes that landed after closing the V1 gates, surfaced by a
real reverse-engineering pass driven through the /re-analyze skill
(see https://github.com/zachgenius/lf_brokers_RE — CFFEX cffex_server
binary RE round).

Changes:

• fix(symbol.find): resolve qualified C++ names and bare method names
  via a 3-pass cascade (exact -> FindFunctions(eFunctionNameTypeAuto)
  -> demangled endsWith("::"+name)).
• fix(symbol.find): also try the arg-stripped form when the query is
  Class::Method(args). Portable across upstream LLDB and Apple
  libLLDB on arm64.
• fix(type.layout): emit a warnings array when DWARF
  DW_AT_byte_size is smaller than the last field's end. Surfaces
  toolchain-emitted DWARF inconsistencies as data, not error.
• fix(disasm): emit address alongside addr on every disassembled
  instruction so --view fields=address,... projections work.
• docs(skill): clarify value.read (path/frame-relative, requires a
  stopped thread) vs mem.read (address-based, works against
  binaries / cores) in the re-analyze skill's Phase 2.

No breaking changes. Adds two response fields (warnings on
type.layout, address on disasm). 591 unit-test cases (up from 588),
all platforms green.

Source-only release; no prebuilt binaries.