diff --git a/AGENTS.md b/AGENTS.md
new file mode 100644
index 00000000..dfad8535
--- /dev/null
+++ b/AGENTS.md
@@ -0,0 +1,122 @@
+# SnapRAID Development Guide
+
+This document provides guidance for agentic coding agents working on the SnapRAID codebase.
+
+## Project Structure
+
+### Core Source Files (`cmdline/`)
+
+The CLI is organized into focused C modules:
+
+| Modules | Purpose |
+|---|---|
+| `main.c`, `snapraid.h` / `snapraid.c` | Program entry points, option parsing, and CLI command dispatch |
+| `state.h` / `state.c` | Global state (`snapraid_state`), config file parsing, and content file read/write |
+| `elem.h` / `elem.c` | Core data model (files, links, dirs, disks, parity, filters) and filter logic |
+| `scan.c`, `import.h` / `import.c` | Multi-threaded directory scanner and file-move hash importer |
+| `search.h` / `search.c`, `locate.h` / `locate.c` | Path pattern matching and file lookup / path location commands |
+| `io.h` / `io.c` | Parallel read-ahead and write-behind I/O queue management |
+| `parity.h` / `parity.c` | Parity file operations supporting split parity files |
+| `handle.h` / `handle.c` | File handles abstraction for data disk I/O with POSIX fadvise settings |
+| `bw.h` / `bw.c` | Bandwidth token-bucket rate limiter |
+| `stream.h` / `stream.c` | Buffered stream abstraction with CRC32 integrity checks |
+| `sync.c`, `scrub.c`, `check.c`, `status.c` | CLI commands for core operations (synchronize, integrity scrub, check, health status) |
+| `pool.c`, `dup.c`, `list.c`, `touch.c`, `rehash.c` | CLI commands for pooling layouts, finding duplicates, listing files, updating tracking/hashes |
+| `device.c`, `dry.c` | Disk device geometry/UUID querying, and dry-run execution simulation |
+| `murmur3.c`, `spooky2.c`, `metro.c`, `museair.c` (and tests) | Non-cryptographic hash implementations (MurmurHash3, SpookyHash, MetroHash, MuseAir) and tests |
+| `support.h` / `support.c`, `app.h` | Locks, logging, thread wrappers, global constants |
+| `str.h` / `str.c` | Path/string utilities |
+| `memory.h` / `memory.c` | Safe memory allocation and tracking helpers |
+| `util.h` / `util.c` | Low-level utilities (CRC32 calculations, numeric formatting, hash helpers) |
+| `unixapp.c`, `mingwapp.c` | Platform-specific implementations of system info, process management, and services |
+| `thermal.h` / `thermal.c` | Temperature monitoring to prevent hardware overheating |
+| `selftest.c`, `speed.c` | Self-test command (`test`) and CPU RAID parity performance benchmark (`-T`) |
+| `mktest.c`, `mkstream.c` | Standalone test utility tools for recovery and stream validation |
+
+### OS Abstraction Files (`os/`)
+
+The platform abstraction layer is separated into `os/`. Note that `os/` acts as a library layer, so having unused functions (e.g., platform helper functions not currently invoked by the main CLI application) is expected.
+
+| Module | Purpose |
+|--------|---------|
+| `os.h` | Core OS-independent interface declarations (syslog, signals, execution, threads, mutexes) |
+| `portable.h` | Platform detection macros and compatibility definitions |
+| `unix.c/h` | UNIX implementation of OS abstraction layer (fork, exec, signals, thread primitives) |
+| `mingw.c/h` | Windows implementation of OS abstraction layer (process creation, registry, thread primitives) |
+
+### Key Data Structures
+
+| Struct | Defined in | Description |
+|---|---|---|
+| `snapraid_state` | `state.h` | Global program state: all disks, parity levels, filter list, options, hash kind, progress |
+| `snapraid_option` | `state.h` | Runtime options (force flags, sort order, IO advise mode, bandwidth limit, etc.) |
+| `snapraid_disk` | `elem.h` | One data disk: name, mount dir, device id, UUID, file/link/dir hash tables and lists, parity extent tree |
+| `snapraid_file` | `elem.h` | A tracked file: sub-path, size, mtime, inode, physical offset, block vector |
+| `snapraid_block` | `elem.h` | One parity block: state (`BLK/CHG/REP/DELETED/EMPTY`) + hash |
+| `snapraid_extent` | `elem.h` | Contiguous run of file blocks mapped to contiguous parity positions |
+| `snapraid_filter` | `elem.h` | One include/exclude rule: pattern, root scope, `is_disk`, `is_abs`, `is_dir`, direction |
+| `snapraid_parity` | `elem.h` | One parity level: up to `SPLIT_MAX=8` split files, total/free blocks |
+| `snapraid_io` | `io.h` | Parallel I/O controller: buffer pool, worker threads for each disk |
+| `snapraid_worker` | `io.h` | One I/O thread: reads or writes one disk at a time using a task ring buffer |
+
+### Block States
+
+| Constant | Value | Meaning |
+|---|---|---|
+| `BLOCK_STATE_EMPTY` | 0 | No file mapped here; parity position is zero-filled |
+| `BLOCK_STATE_BLK` | 1 | File block with valid hash and up-to-date parity |
+| `BLOCK_STATE_CHG` | 2 | File block changed; hash may be old (pre-sync); parity not yet updated |
+| `BLOCK_STATE_REP` | 3 | File block hashed (from copy heuristic); parity not yet updated |
+| `BLOCK_STATE_DELETED` | 4 | File deleted; block retains old hash for parity recovery; parity not yet zeroed |
+
+### Crash-Recovery Invariants
+
+SnapRAID is designed to recover from an interruption, including `SIGKILL`, at any point. It achieves this by reading recovery data from multiple snapshots, checking file metadata when reading, and validating data with hashes. Account for this design before treating a non-atomic transition or mixed snapshot lifecycle as an integrity defect.
+
+### Build System
+
+- `configure.ac`: Autoconf script (detects systemd vs BSD init)
+- `Makefile.am`: Source file lists, dependencies, install hooks (including rules for generating documentation)
+- `uncrustify.cfg`: Code formatting rules for main application source files
+- `linux.cfg`: Code formatting rules for `raid/` directory and its subdirectories
+- Run `make doc` to regenerate all manual pages (`*.1`) and text manuals (`*.txt`)
+
+## Code Style Guidelines
+
+#### Code Style
+
+- **Language**: The codebase uses **C99 standard** (not C11/C++)
+- **Format**: Enforce via `uncrustify -c uncrustify.cfg --no-backup *.c *.h` (or `uncrustify -c linux.cfg --no-backup *.c *.h` for `raid/` directory and subdirectories)
+- **Naming**: Snake_case for functions, UPPER_CASE for macros/constants
+- **Indentation**: Tabs for indentation, no alignment (existing codebase style)
+- **Comments**: C-style `/** */` for multiline comments; C `/* first letter lowercase */` for single-line inline notes
+- **Headers**: All `.h` files have include guards (`#ifndef __NAME_H`)
+- **Preferences**: Use 0 instead of NULL and '\0'
+- **Preferences**: Use prefix ++variable and --variable instead of postfix variable++ and variable-- where both are equivalent
+- **Safety Checks**: Avoid adding safety checks for conditions that never happen
+- **Commit Messages**: Every time a change is done, a single line commit description should be provided for that change
+- **Git Commits**: Never commit changes to git.
+
+#### Error Handling & Memory Management
+
+- **Return codes**: `0` means no-error; `< 0` means error; positive for specific results
+- **Logging**: Use `log_fatal()`, `log_tag()`, `msg_info()`, `msg_progress()`, `msg_verbose()` from `support.h`
+- **Allocation**: Always use custom `malloc_nofail()`, `calloc_nofail()`, `nalloc_nofail()`, or `strdup_nofail()` wrappers (which handle OOM by aborting)
+- **Path Manipulation**: Never use standard unsafe string functions (`strcpy`, `strcat`, `strcmp`) on paths. Always use the platform-correct, size-bounded wrappers in `support.h` (e.g., `pathcpy()`, `pathcat()`, `pathcmp()`, `pathslash()`).
+- **TommyDS Iteration**: For `tommy_list`, iterate using:
+  `tommy_node* node = list.head; while (node) { void* obj = node->data; ... node = node->next; }`
+  Or use callback iteration helpers: `tommy_list_foreach()`, `tommy_list_foreach_arg()`, `tommy_hashdyn_foreach()`, etc.
+- **Cleanup**: Use `goto` pattern for cleanup in complex functions when appropriate
+- **Fatal paths**: Use `exit(EXIT_FAILURE)` for unrecoverable errors; `os_abort()` for internal inconsistencies
+
+#### Threading
+
+- **Compile-time guard**: `#if HAVE_THREAD` wraps all thread-specific code
+- **Mutexes**: Always use the custom `thread_mutex_init/destroy/lock/unlock()` wrappers
+- **IO workers**: The `snapraid_io` / `snapraid_worker` infrastructure handles parallel disk I/O via a pool of reader and writer threads
+
+#### Testing Changes
+
+- Never run `make check` because it's too expansive
+- Run `./snapraid test` for the regression test
+- Verify error paths and recovery scenarios (the test suite covers aborted sync, UUID changes, disk failures)