Fork Analysis Results

Fork Network — Cross-Referenced Analysis & Results

Synthesized analysis of all findings from the fork network sweep. Cross-references related items, deduplicates overlapping work, and provides actionable recommendations.

Generated: 2026-06-12 Source data: FORKS.md (tracking), FORKS_REPORT.md (raw findings)

---

Table of Contents

1. Deduplicated Fix Clusters
2. Deduplicated Feature Clusters
3. Cross-Fork Dependency Graph
4. Port Priority Matrix
5. JPEG/MJPEG Specific Analysis
6. Upstream Issue Resolution Map
7. Risk Assessment

---

1. Deduplicated Fix Clusters

Many forks independently discovered or ported the same fixes. This section groups related fixes to identify the best source for each.

FC-1: Decoder Memory Safety (cgutman — 4 fixes)

Root cause: Original decoder has multiple memory safety issues in lifecycle management.

Fix	Commit	Impact
Mutex deadlock on EOS	58a977d4	Decoder hangs at stream end
delete vs delete[]	ff8fd95a	Heap corruption (decoder + encoder)
Uninitialized pointer deletion	49d55d10	Crash if < MAX_BUFFERS used
Hang on close without data	00455615	Indefinite hang if no frames sent

Sources: cgutman (original author) → cybernhl (inherits). No other fork has these. Best source: cybernhl/jetson-ffmpeg moonlight_fixes branch — all 4 in sequence. Cross-refs: jocover#10, #104 (thread lifecycle); Moonlight game streaming use case. Status in our repo: FIXED — all four commits are ancestors of main; they entered the lineage via jocover's merges of cgutman's PR#22/#23/#28. Verified 2026-06-12 (gjrtimmer#2, closed).

FC-2: Missing Return Statement UB (2 locations)

Root cause: V4L2 capture callback functions declared void* but have no return statement. Newer GCC on L4T 5.1+ treats this as UB → segfault.

Location	Commit	Author
Encoder: enc_capture_loop_fcn	0229e5ba	Extend-Robotics
Decoder: dec_capture_loop_fcn	6d327624	Extend-Robotics

Sources: Extend-Robotics only fork with explicit fix. Cross-refs: jocover#127 (decoder segfault reports), jocover#98. Status in our repo: FIXED — 184a67a ("add missing return statements") is an ancestor of main, and the lineage refactor removed the UB class entirely (void capture fcn via std::thread, bool DQ callback). Verified 2026-06-12 (gjrtimmer#3, closed).

FC-3: Key Frame Flag Bug (encoder)

Root cause: nvmpi_enc.c line ~92 does pkt->flags |= 0x0001 (AV_PKT_FLAG_KEY) on EVERY packet. Result: every frame marked as keyframe. Breaks seeking, GOP structure, rate control, HLS/DASH segmentation.

Fix	Commit	Author
Use enc_metadata.KeyFrame from V4L2	1324d7d6	spotai

Sources: spotai only. Filed as Keylost#26 (UNFIXED upstream). Cross-refs: Keylost#26. Status in our repo: FIXED — spotai's 1324d7d is an ancestor of main (flag from V4L2 KeyFrame metadata, conditional in wrapper). Upstream Keylost#26 remains unfixed. Verified 2026-06-12 (gjrtimmer#4, closed).

FC-4: H.265 Encoder Header / IDR Scan

Root cause: GLOBAL_HEADER scan for SPS/PPS NALs was (a) not applied to H.265, (b) had unbounded scan loop that could read past buffer end.

Fix	Commit	Author
Extended to H.265 + bounded scan	2b715e81	spotai

Sources: spotai only. Status in our repo: FIXED — spotai's 2b715e8 is an ancestor of main (H.265 included, scan bounded by payload_size); a defensive tail-read bound on the start-code probe landed with gjrtimmer#5.

FC-5: glibc 2.34+ pthread Segfaults

Root cause: NvV4l2ElementPlane has uninitialized pthread_t dq_thread. On glibc 2.34+ (Ubuntu 22.04+), pthread_join() on uninitialized pthread_t segfaults instead of returning an error.

Fix	Commits	Author
Init dq_thread, fix teardown, vendor patched NvV4l2ElementPlane.cpp	3 commits	LanderN

Sources: LanderN only (optimize/jetson branch). Cross-refs: Keylost#21, PR#48 (Ubuntu 22.04 compilation issues). Note: This patches NVIDIA's sample code, not our code. Need to vendor or upstream the fix. Status in our repo: NOT FIXED. CRITICAL for Ubuntu 22.04+.

FC-6: 4K I-Frame Buffer Overflow

Root cause: CHUNK_SIZE (input buffer for encoded data) hardcoded at 4MB. 4K I-frames can exceed this → truncation or crash.

Fix	Source	Author
CHUNK_SIZE 4→10MB + bounds check + configurable chunk_size	cuda_buffers branch	GlassBil

Sources: GlassBil only. Status in our repo: FIXED (gjrtimmer#7) — decoder input default raised to 10 MB with a chunk_size AVOption and an entry bounds check; encoder packet buffers raised to 10 MiB with an overflow guard in the DQ callback.

FC-7: RTSP / Out-of-Band SPS/PPS Decoder Hang

Root cause: Decoder hangs when SPS/PPS is only in SDP (not in-band). Decoder needs extradata fed before first frame.

Fix	Source	Approach
Prime decoder with extradata at init	w3sip 0d320fe5	Original implementation
Same fix, different approach	hsuanguo PR#3	Co-authored with w3sip
Same fix	gjrtimmer 08e32c15	Our implementation
extradata-at-init	danhawker, matthewj-t6	Same concept
Partial (camera-specific)	LanderN	Cites w3sip

This is the most-duplicated fix in the entire network. 5+ independent implementations. Status in our repo: FIXED (gjrtimmer@08e32c15). Not merged upstream to Keylost.

FC-8: Decoder Crash on Exit

Root cause: Queued DMA buffers not cleaned up in destructor → crash on rapid open/close or error paths.

Fix	Source	Author
Cleanup queued buffers in destructor	49039f81	w3sip
Related: adaptive poll timeout for resolution event	00455615	cgutman

Two different aspects of the same lifecycle problem. Both should be ported. Status in our repo: NOT FIXED.

FC-9: Child Process Decoder Segfault

Root cause: V4L2 fd inheritance after fork() breaks decoder. NvBufferSession create/destroy around transform fixes it.

Fix	Source	Author
NvBufferSession lifecycle	xia-chu fd0900d7 (via Keylost fork_fix)	Keylost/xia-chu
nvjpeg removal (related)	hsuanguo PR#2	Removes unused dep

Status in our repo: VERIFY — may already be in Keylost lineage we inherited.

---

2. Deduplicated Feature Clusters

FT-1: Zero-Copy / GPU Pipeline

Multiple forks approach GPU-resident decode/encode from different angles:

Approach	Author	Maturity	Files
Full DMA-BUF pipeline (get_frame_fd/put_frame_fd + DRM_PRIME)	YuriiHoliuk	Most complete (26 commits)	nvmpi_dec.cpp, nvmpi_enc.cpp, nvmpi.h, nvmpi_surface.cpp
VIC hardware scale filter (vf_scale_vic.c)	YuriiHoliuk	Working	+468 lines new file
hwcontext_nvmpi (FFmpeg HW device type)	mguzzina	WIP (6 commits)	hwcontext_nvmpi.c/h
CUDA buffer return (frames stay on GPU via FFmpeg API)	GlassBil	Working	cuda_buffers branch
dlopen lazy-loading (prevent EGL crash)	YuriiHoliuk	Working	dynlink_nvmpi.h
dlopen (independent impl)	wangxiaoshuang	Working	different approach

These are complementary, not competing:

• YuriiHoliuk's DMA-BUF pipeline handles the V4L2↔FFmpeg data path
• mguzzina's hwcontext_nvmpi provides the FFmpeg framework integration
• GlassBil's CUDA buffer return provides the CUDA interop
• dlopen prevents startup crashes when libnvmpi is missing

Recommended port order: dlopen → DMA-BUF pipeline → hwcontext → VIC filter → CUDA

FT-2: Encoder Control APIs

Feature	Author	API
Dynamic bitrate	muxable	nvmpi_encoder_set_bitrate(ctx, bitrate)
Force IDR	muxable	nvmpi_encoder_force_idr(ctx)
Per-frame QP, SAR, bit_depth, chroma_format_idc	muxable	Extended nvPacket struct
Lossless encode	muxable	enableLossless (also requested in Keylost#45)
ROI encode	muxable	Region of interest
CABAC toggle, VUI insert, AUD insert	muxable	Additional encoder params

Single source: muxable. Most comprehensive encoder param expansion. Cross-refs: Keylost#45 (lossless), Keylost#50 (setLevel), jocover#50,#52,#7 (quality control).

FT-3: Performance Optimization

Optimization	Author(s)	Impact
setMaxPerfMode(1)	douo, Extend-Robotics, LanderN, bradcagle	4K HEVC: 28→64fps; 720p H264: 46→290fps
Disable DPB	LanderN	Additional decode perf
condition_variable blocking wait	cgutman	Eliminates polling; true zero-latency
poc-type=2 (display=decode order)	Extend-Robotics	Reduces latency for H.264

setMaxPerfMode is a one-liner that every perf-focused fork independently discovers. We should port it immediately.

FT-4: Pixel Format / Color Space

Fix/Feature	Author	Commit
YUVJ420P→YUV420P remap	bradcagle	67e7bbcd
NV12 format support	xsacha	4 patches
P010 10-bit decode	YuriiHoliuk	4aab1b56
ABGR32→XRGB32 color fix	dsleep	018e6e2e
insert_vui for FPS	vietnx	fdbb815a

Related upstream issues: jocover#20, #60, #66, #68, #73, #83, #116, #75, #80, #11, #133. Pixel format is one of the most-reported issue categories. NV12 + YUVJ420P + 10-bit together would resolve most complaints.

FT-5: Pool Size Configuration

Feature	Author	Why
frame_pool_size option	spotai	Prevents EAGAIN on frame pool exhaustion
packet_pool_size option	spotai	Prevents EAGAIN on packet pool exhaustion
chunk_size option	GlassBil	Configurable input buffer for 4K+

Cross-refs: Keylost#38 (EAGAIN workaround: -packet_pool_size:v 32). Status in our repo: DONE — frame_pool_size/packet_pool_size AVOptions already in the lineage (spotai cf0560b is an ancestor of main); chunk_size added via gjrtimmer#7. Tracked as gjrtimmer#11 (closed).

FT-6: Codec Flush / Lifecycle

Issue	Status	Related Forks
Codec flush not implemented	Keylost#42 UNFIXED	No fork has a fix
Encoder thread race (packet_pools)	jocover#130 UNFIXED	bmegli identified; no fix
Non-monotonic DTS	jocover#76, #110	teplofizik patches (partial)

Gap: No fork in the entire network has implemented proper codec flush. This is an open problem.

FT-7: Modern FFmpeg Encode API (send_frame / receive_packet)

Approach	Author	Maturity
send_frame/receive_packet encoder (coexists with legacy encode2)	bmegli	POC (1 commit, new-ffmpeg-api branch)
POC patch (receive_packet only, 705-line patch file)	Extend-Robotics	Reference only (1c100f1c)

Root cause: Legacy encode2 callback blocks synchronously — 2-5 frame latency, can miss final packets at EOS. bmegli's implementation adds encoder_flushing state for proper EOS drain that the old API lacks. Cross-refs: jocover#131 (encoder latency), jocover#7, Keylost#42 (flush). Note: This is closely related to FT-6 (codec flush) — the new API is the prerequisite for proper flush.

FT-8: Build System / Cross-Compilation

Feature	Author	Unique?
CMake cflags quoting fix	tmm1 428edfec	YES — real bug, breaks strict CMake
Yocto/BITBAKE_SYSROOT cross-compile	kazuki0824 (4 commits)	YES — embedded Linux use case
Cross-compile via TARGET_ROOTFS	muxable	Different approach (Debian sysroot)
CPack .deb packaging	muxable, jkms, Extend-Robotics	Multiple implementations
pkg-config for libv4l2	kazuki0824, D31m05z	Two independent approaches

tmm1's cflags fix is a real bug that affects all builds on strict CMake — should port.

---

3. Cross-Fork Dependency Graph

jocover/jetson-ffmpeg (abandoned ~2022)
├── Keylost/jetson-ffmpeg (active maintainer, 74 ahead)
│   ├── YuriiHoliuk (+26): DMA-BUF, VIC, P010, dlopen
│   ├── mguzzina (+6): hwcontext_nvmpi
│   ├── wangxiaoshuang (+9): dlopen, decoder rewrite
│   ├── hsuanguo (+8): RTSP fix, JP5.1.2, FFmpeg 6.1
│   ├── fingul (+7): JP6.2 install
│   ├── v-faeez-kadiri (+5) → spotai (+20 on branch): KEY FRAME FIX, H265, FFmpeg 8.0
│   ├── izdeveloper (+3) → Ledrego (behind)
│   ├── madsciencetist (+3): resize, stderr → AndBobsYourUncle (cmake) → traversaro (6.1)
│   ├── w3sip (branches only): RTSP fix origin, crash-on-exit, refactor
│   ├── nhahn (+1): JP6 path detect
│   ├── LanderN (branch): glibc fix, setMaxPerfMode
│   ├── GlassBil (branch): CUDA buffers, 4K overflow fix
│   ├── peters (branch): FFmpeg 6.1
│   └── [mirrors: 30+ forks with 0 unique work]
│
├── cgutman (Moonlight) → cybernhl: 4 safety fixes + low-latency
├── dsleep (+32): vendor NVIDIA classes, color fix, encoder fix, JP4/5
├── muxable (+28): dynamic bitrate, force IDR, expanded params
├── xia-chu (+16): child process fix
├── vietnx (+4): leak fix, insert_vui, linesize fix (PR#79 merged)
├── bradcagle (+7): JPEG prototype, YUVJ fix, NvUtils migration ref
├── douo (+2) → Extend-Robotics (+8): UB fixes, POC new API
│            → ReAlexLiu (+4): JP5 compat
├── bmegli (branch): new send_frame/receive_packet encoder API POC
├── xsacha (branches): pixel format fixes
├── dthk-cogmatix (+1): NvBufferSession lifecycle
├── tmm1 (branch): CMake cflags quoting fix
├── kazuki0824 (branch): Yocto cross-compile + pkg-config
├── koenvandesande (branch): YUVJ420P acceptance (= bradcagle)
├── uschen-thirdparty (branch): setMaxPerfMode (= douo/LanderN)
└── [164 verified mirrors with 0 unique work on any branch]

Total coverage: 237 forks enumerated → 70 with unique work analyzed → 164 verified mirrors → 3 trivial (README/docs only)

---

4. Port Priority Matrix

Scoring: Impact (1-5) x Confidence (1-5) x Effort (inverse, 5=easy, 1=hard)

#	Item	Impact	Confidence	Effort	Score	Category
1	Key frame flag fix	5	5	5	125	FC-3
2	cgutman 4 safety fixes	5	5	4	100	FC-1
3	setMaxPerfMode(1)	4	5	5	100	FT-3
4	Missing return UB fixes	5	4	5	100	FC-2
5	H265 header/IDR scan fix	5	4	5	100	FC-4
6	glibc 2.34+ pthread fix	5	4	3	60	FC-5
7	4K I-frame overflow fix	4	4	4	64	FC-6
8	Pool size configuration	3	5	4	60	FT-5
9	condition_variable wait	4	4	3	48	FT-3
10	YUVJ420P remap	3	5	5	75	FT-4
11	Dynamic bitrate + IDR	4	4	3	48	FT-2
12	Crash-on-exit fix	3	4	4	48	FC-8
13	JetPack 6 path detect	3	5	5	75	—
14	CMake cflags quoting fix	2	5	5	50	FT-8
15	New encode API (send_frame/receive_packet)	4	3	2	24	FT-7
16	DMA-BUF pipeline	5	3	1	15	FT-1
17	hwcontext_nvmpi	5	2	1	10	FT-1
18	Yocto cross-compile	2	4	3	24	FT-8

Recommended Port Waves

Wave 1 — Safety fixes (1 PR, should be fast): Items 1-5, 7, 14. All are isolated fixes with clear commits to cherry-pick. Includes CMake cflags fix.

Wave 2 — Performance + usability (1-2 PRs): Items 3, 8, 9, 10, 12, 13. setMaxPerfMode is one line; pool sizes are config additions; YUVJ remap is small.

Wave 3 — Encoder features (1 PR): Items 11, plus expanded encoder params from muxable.

Wave 4 — glibc fix (1 PR, needs care): Item 6. Requires vendoring patched NVIDIA code. May need upstream engagement.

Wave 5 — Modern encode API (1 PR): Item 15. bmegli's send_frame/receive_packet POC as starting point. Prerequisite for proper codec flush.

Wave 6 — GPU pipeline (multi-PR, major feature): Items 16, 17. DMA-BUF + hwcontext. Months of work but highest long-term value.

---

5. JPEG/MJPEG Specific Analysis

Upstream Demand Signal

Ref	What
Keylost#37	"High cpu usage when encoding" — JPG sequence → h264_nvmpi. Software JPEG decode burns >1 CPU core. jtop: NVENC active, NVJPG OFF
Keylost#40	"Accelerated JPEG decoding and encoding" — direct feature request. Keylost: "I'll consider it when I have time" (2025-08)
Keylost#45	Lossless-HEVC bug, but workflow is -f image2 -i %d.jpg with 4K JPEGs — third JPEG-input user
Keylost#47	"support direct access to nvbuffer" — zero-copy follow-up that pairs with JPEG
Keylost comment on #37	Confirms root cause: "load could be reduced by using NvJpegDecoder … passing the hardware buffers … directly to the encoder without copying"

Prior Art in Fork Network

Source	What	Status
bradcagle cd9bf72e	NvJPEGDecoder C API prototype: nvmpi_create_jpeg_decoder, nvmpi_decode_jpeg, nvmpi_close_jpeg_decoder	Incomplete — no FFmpeg codec wrapper, no encoder
Keylost#37, #40	Feature requests for JPEG HW decode/encode	UNFIXED
hsuanguo PR#2	Removed nvjpeg (unused dep)	Merged — confirmed JPEG was never functional
Keylost PR#9	Added nvjpeg stubs	Merged — stubs only, no implementation
NVIDIA forum /t/189290	nvjpeg causes segfault in child processes	Root cause documented

No fork in the entire 237-fork network has a working JPEG implementation. (Confirmed by exhaustive sweep.)

Hardware Support Matrix

Module	NVJPG decode	NVJPG encode
Nano (T210), TX2, Xavier NX, AGX Xavier	600 MP/s	Yes
AGX Orin, Orin NX (T234)	2x engines (2x600 MP/s)	Yes
Orin Nano	Decode-only (499.2 MHz max)	NO (NVIDIA confirmed)

• Engine: NVIDIA Tegra-accelerated libjpeg-8b fork (libnvjpeg.so), NOT CUDA nvJPEG
• Limits: baseline JPEG only (no progressive), 8 bpc, max 16384x16384
• Decode: 420/422H/V/444/400; Encode: YUV420/NV12 input only

Proposed Design

Backend: NvJPEGDecoder::decodeToFd / NvJPEGEncoder::encodeFromFd (documented MMAPI classes, present in devcontainer's mounted MMAPI). Identical signatures JP4→JP6.

Shape:

• include/nvmpi.h: add NV_VIDEO_CodingMJPEG to nvCodingType
• src/nvmpi_jpegdec.cpp (new): synchronous decode — put_packet → decodeToFd → VIC transform → frame pool → get_frame copy-out. No capture thread, no resolution events.
• src/nvmpi_jpegenc.cpp (new, phase 2): copy raw YUV420 into DMA buffer → encodeFromFd(quality) → nvPacket. Map FFmpeg -q:v → libjpeg quality 1-100.
• FFmpeg layer: NVMPI_DEC(mjpeg, AV_CODEC_ID_MJPEG, NULL) + encoder registration
• Output yuv420p + color_range = AVCOL_RANGE_JPEG (not deprecated yuvj420p)
• scripts/ffpatch.sh: anchor mjpeg_cuvid_decoder_deps="cuvid" (present in all 7 FFmpeg versions)
• CMake: libnvjpeg already linked; add NvJpegDecoder.cpp/NvJpegEncoder.cpp to sources

Phasing:

1. v1: mjpeg_nvmpi decoder only — resolves #37's CPU on ALL modules incl. Orin Nano
2. v2: mjpeg_nvmpi encoder — everywhere except Orin Nano
3. Follow-up: zero-copy decode→encode (DMA fd passthrough; Keylost#47 territory)

Risks

• JP5.1.x decodeToFd stale-frame regression (NVIDIA-acknowledged, forum /t/264120)
• Progressive JPEG: HW unsupported — detect SOF2 and fail with clear error
• 422/444 JPEGs: decode OK, VIC converts to YUV420 → chroma loss in v1
• Orin Nano NVJPG decode clock low by default pre-JP5.1.2 (115 MHz)
• nvjpeg causes fork() segfaults (FC-9) — must handle in implementation

Relation to Other Findings

• FC-9 (child process segfault): nvjpeg was specifically identified as causing fork() segfaults
• FT-1 (zero-copy pipeline): JPEG decode should support DMA-BUF output for VIC scaling
• FT-4 (pixel format): JPEG needs YUVJ420P/YUV422P/YUV444P support
• FT-7 (new encode API): JPEG encoder benefits from send_frame/receive_packet pattern
• bradcagle's prototype: Starting point for C API, needs FFmpeg wrapper + encoder

---

6. Upstream Issue Resolution Map

How fork findings map to upstream issues:

Upstream Issue	Fork Resolution	Best Source
jocover#127 (decoder segfault)	Extend-Robotics 6d327624	Extend-Robotics
jocover#131 (encoder latency)	Extend-Robotics POC 1c100f1c	Extend-Robotics (reference)
jocover#130 (encoder thread race)	STILL UNFIXED	—
jocover#120 (setMaxPerfMode)	douo, LanderN, bradcagle	Any (one-liner)
jocover#41,#48,#100 (PTS truncation)	teplofizik patches.zip	teplofizik
jocover#20,#60+ (YUVJ420P)	bradcagle 67e7bbcd	bradcagle
jocover#64,#80 (linesize crash)	vietnx PR#79 (merged)	Already upstream
jocover#11,#73 (NV12 support)	xsacha 4 patches	xsacha
jocover#75,#80 (10-bit)	YuriiHoliuk P010	YuriiHoliuk
jocover#62,#67+ (CUDA/zero-copy)	YuriiHoliuk, GlassBil, mguzzina	Combined approach
jocover#117,#118 (DMA encode)	YuriiHoliuk put_frame_fd	YuriiHoliuk
Keylost#26 (keyframe flag)	spotai 1324d7d6	spotai
Keylost#42 (codec flush)	STILL UNFIXED	—
Keylost#37,#40 (JPEG)	bradcagle prototype (incomplete)	bradcagle (reference)
Keylost#45 (lossless H265)	muxable enableLossless param	muxable
Keylost#49 (AV1)	STILL UNFIXED	Hardware capable on AGX Orin
Keylost#50 (H265 setLevel)	STILL UNFIXED	—
Keylost#31,#47 (hwcontext)	mguzzina WIP	mguzzina

Permanently unfixed across entire network (no fork has a solution):

• Encoder thread race condition (jocover#130)
• Codec flush (Keylost#42)
• AV1 support (Keylost#49)
• H265 setLevel (Keylost#50)

---

7. Risk Assessment

High-Confidence Ports (well-tested, clear commits)

• cgutman safety fixes: Battle-tested in Moonlight game streaming (latency-sensitive)
• spotai key frame fix: Isolated, obvious bug with clear before/after behavior
• setMaxPerfMode: One-liner, no risk, massive perf gain
• YUVJ420P remap: Simple conditional, well-understood
• Pool size config: Additive (new AVOptions), no behavioral change to defaults

Medium-Risk Ports (need adaptation)

• glibc pthread fix: Vendors patched NVIDIA code — needs maintenance strategy
• 4K overflow fix: Changes buffer size constant — verify no regressions at lower resolutions
• condition_variable wait: Changes decoder threading model — needs careful testing
• Extend-Robotics UB fixes: Simple return NULL; additions but verify our code structure

High-Risk Ports (major refactoring)

• DMA-BUF pipeline: 26 commits touching decoder, encoder, and FFmpeg codec layer
• hwcontext_nvmpi: WIP; modifies FFmpeg's hardware device framework
• Dynamic bitrate: Adds new API surface — must design for stability
• Decoder resize: Unmerged even in madsciencetist's own repo; incomplete

Items Requiring Our Own Implementation (no complete fork solution)

• JPEG/MJPEG hardware decode+encode (bradcagle prototype is incomplete)
• Codec flush (no fork has solved this)
• AV1 decode/encode (hardware-capable, zero software work anywhere)
• Full hwcontext integration (mguzzina's WIP is starting point only)