HaloKeymind/docs/ota_protocol.md

MeshCore OTA - .mota container & LoRa protocol

This is the single source of truth for MeshCore's over-the-air firmware update system ("mOTA"). It is written for developers who want to implement an interoperable peer (server, fetcher, relay, or host tool) in another codebase or project. Everything below is implemented in this repository and covered by host, simulation, build, or hardware tests as noted in the relevant section. Hardware qualification is target- and chain-specific; do not infer it from implementation alone. Where a section names a source file, that file is the authoritative reference for byte-level details.

Just want to update your node? See the plain-language OTA user guide - this document is the technical/wire specification.

Design goals

• Distribute firmware over LoRa as a self-verifying, resumable, single-source block transfer that survives reboots and never auto-applies without explicit consent.
• Trustless mesh relay: repeaters may forward packets while the source alone serves firmware data; integrity is content-addressed against a signed merkle root, so a relay need not be trusted and never needs the signing keys.
• Primary while transferring: periodic discovery stays at background priority, but manifest, block, data, and proof packets for an active fetch use primary queue priority at every relay hop.
• Portable: the engine (src/helpers/ota/OtaManager) is Arduino/radio/crypto-free and host-testable, so the same logic drives a device, a simulation, or a third-party implementation.

Source map (all under src/helpers/ota/ unless noted)

Concern	File
Constants, enums, flags	OtaFormat.h
Container/manifest parse	MotaContainer.{h,cpp}
Merkle tree + proofs	MerkleTree.{h,cpp}
EndF self-identity	FirmwareInfo.{h,cpp}
Wire message codec	OtaProtocol.{h,cpp}
Session engine (serve+fetch+discovery)	OtaManager.{h,cpp}
Multi-mota / folder relay	OtaSource.h, MotaSourceSerial.{h,cpp}, MotaSeederProto.h
Staging stores	OtaStore.h, OtaStoreFlashNrf52.*, OtaStoreFlashEsp32.*
Apply	OtaApply.*, bootloader Adafruit_nRF52_Bootloader_OTAFIX
Device glue (CLI/context)	OtaCli.cpp, OtaContext.h
Host tooling	motatool (standalone Rust CLI: build/verify/inspect/serve); tools/mota/ (Python reference lib motalib.py + build/test glue)

---

1. Conventions

• Endianness: all multi-byte integers are little-endian unless stated.
• Hashes (multihash): the hash family is declared once per manifest via hash_algo = 0x12 = SHA-256 (the multihash code for sha2-256). Truncations used:
  • sha2-256:4 - first 4 bytes of the SHA-256 digest. Merkle leaves, internal nodes, root, proofs, manifest_id, and the discovery set_digest.
  • sha2-256:8 - first 8 bytes. Base-firmware identity (base_hash, EndF.body_hash).
  • sha2-256:32 - full digest. The image security anchor (image_hash). Digests are stored bare (just the truncated bytes); the family is implied by hash_algo.
• Signatures: Ed25519 (RFC 8032), 64-byte detached signature, 32-byte public key.

Reference constants (OtaFormat.h):

Name	Value	ASCII / note
Container MAGIC	6D 4F 54 41	mOTA
Container TRAILER	76 6B 34 39 36	vk496
EndF marker	45 6E 64 46	EndF
hash_algo (sha2-256)	0x12	multihash code
format_ver	0x02	this spec
approval = not approved	FF FF FF FF	erased NOR word
approval = approved	41 50 52 56	APRV
MFLAG_FULL	0x01	flags bit0
MFLAG_SIGNED	0x02	flags bit1
CODEC_FULL / _SEQUENTIAL / _INPLACE	0 / 1 / 2	Section 5
PAYLOAD_TYPE_OTA	0x0C	MeshCore packet type (src/Packet.h)
MAX_PACKET_PAYLOAD	184	usable bytes per packet (src/MeshCore.h)
Default block size	1024	block_size_log2 = 0x0A
OTA discovery TX priority	250	background (OTA_TX_PRIORITY, src/Mesh.h)
OTA active-transfer TX priority	0	primary (OTA_TRANSFER_TX_PRIORITY, src/Mesh.h)

---

2. Firmware image & the EndF trailer

Every OTA-capable build appends a fixed 56-byte EndF trailer to its flashed image so a running node can discover its own size and self-describing identity on any MCU (no linker symbols needed). Every field is always present at a constant offset. Implemented by FirmwareInfo.cpp; appended at build time by tools/mota/pio_endf.py (post-build hook).

flashed image = BODY (image bytes) || EndF trailer
EndF trailer (fixed 56 bytes):
  off 0   4   "EndF"        45 6E 64 46
  off 4   4   body_len      uint32 LE - length of BODY (excludes the whole trailer)
  off 8   8   body_hash     sha2-256:8 of BODY
  off 16  4   fw_version    uint32 LE, packed MAJOR<<24|MINOR<<16|PATCH<<8|pre  (0 = unknown)
  off 20  4   target_id     uint32 LE - sha2-256:4(pio_env): hardware + role + partition (fetch routing)
  off 24  32  hw_id         NUL-padded ASCII hardware tag (brick-safety), e.g. "RAK4631" ("" = unknown)

• Self-describing identity. pio_endf.py uses build.sh's MOTA_TARGET_ID when present (required for virtual LoRa-OTA build names), otherwise it computes target_id from the PlatformIO env name. It reads hw_id from MOTA_HW_ID and fw_version from FIRMWARE_VERSION. The device reads them back (ota_self_firmware()), so a node's advertised identity is correct regardless of how it was built - and the packaging tool reads them straight from a raw .bin (no --target-env/--fw-version flags, no reliance on filenames; Section 9, Section 13). A dev build with no dotted version simply carries fw_version = 0 / empty hw_id (= unknown) - still a full 56-byte trailer.
• Size discovery: scan flash from the partition top downward for the EndF marker; the byte before it is the last BODY byte (the trailer is always 56 bytes). See ota_self_firmware().
• Delta base matching: a node's body_hash is read directly from its own EndF; a delta's base_hash (Section 5) must equal it. body_hash is over BODY only.
• No circularity: EndF hashes only the BODY, never itself.

The "reconstructed image" referenced by the manifest is the full BODY || EndF (what gets flashed).

ESP32 application-slot profiles

ESP32 companion firmware is exempt from the portable-slot limit. USB and WiFi companion artifacts retain LoRa OTA and carry -ota- in their filenames so they can seed a host folder over serial or TCP; they keep their target partition table rather than using the FULL profile. A small set of high-capacity, non-PSRAM classic ESP32 companions cannot combine their configured contact, group-channel, and offline-queue capacities with LoRa OTA in internal DRAM. Their normal artifacts remain unchanged, and option 3 also emits -full-ota- and -full-logging-ota- variants with 100 contacts, 8 group channels, and a 16-frame offline queue. MQTT observers and ESP-NOW bridges always use FULL builds because fitting them into the legacy slot would require removing CLI and role features. Except for those FULL roles and the ESP32-C6 case below, non-companion ESP32 artifacts, including room, sensor, and repeater roles, must fit the legacy slot from 0x10000 up to 0x150000 (0x140000, 1,310,720 bytes), including the 56-byte EndF trailer. The build checks both that limit and the target's actual app partition. The ESP32-C6 no_external_sensors OTA siblings are the narrow exception: the Arduino 3.x WiFi runtime cannot fit that cross-family ceiling, so those images retain their established target-specific 1920 KiB or larger A/B app layout and are checked against the actual app partition. For every standalone ESP32 and nRF52 repeater, build.sh also exposes an explicit *_lora_ota_no_external_sensors artifact: the ordinary repeater remains sensor-enabled, while that sibling disables optional external environmental-sensor drivers for LoRa distribution. Integrated GPS and other board-native telemetry remain enabled where the target selects the GPS-preserving lean profile. RAK3401 is the explicit exception: RAK_3401_repeater_lora_ota_no_external_sensors undefines ENV_INCLUDE_GPS, so it does not detect or use a RAK12501. RAK12501 GPS requires the ordinary full-sensor RAK_3401_repeater build and sensor slot A; slot D conflicts with the RAK13302 radio's BUSY/DIO1 lines. ESP32 siblings retain the compact browser WiFi updater and use the full 254-entry neighbor table. RP2040 and STM32 targets are not offered because those platforms do not yet have a safe bootloader/apply path.

nRF52 LoRa-OTA siblings use size optimization rather than the Adafruit platform's default -Ofast. This keeps the runtime software Ed25519 fallback from being expanded into tens of kilobytes of repeated curve arithmetic while retaining CC310 hardware crypto, hardware RNG mixing, telemetry history, and board-native features.

WiFi-heavy non-companion roles are not reduced to fit the legacy application slot. build.sh automatically promotes every ESP32 MQTT observer and ESP-NOW bridge to the expanded FULL partition profile. These artifacts retain the complete role CLI, WebConfig where supported, display and optional sensor support, full timezone and TLS behavior, and the board's normal power-management implementation. The compact CLI is not compiled into any build. Ordinary repeater builds remain sensor-enabled; only explicitly named *_lora_ota_no_external_sensors siblings omit sensors for LoRa distribution, and those siblings retain the complete CLI.

MQTT observer radio and bridge preferences use verified temporary files plus a recoverable backup. A reset during a settings save restores the last committed common preference image or publishes the completed new image; it does not leave a partially written /com_prefs file to fail on the next boot. A truncated legacy image is rejected before any partial radio or string fields are applied, then rewritten from safe defaults.

Option 3 in build.sh emits *-full-ota-* and *-full-logging-ota-* ESP32 artifacts for FULL-capable non-companion roles and for the constrained companion fallbacks described above. MQTT observers and ESP-NOW bridges are emitted only with expanded FULL partitions. Menu option 8, or build-full-esp32-firmwares, builds the logging-off FULL artifacts from matching MQTT targets. Menu option 9, or build-full-esp32-logging-firmwares, builds the FULL logging artifacts from matching non-MQTT targets. FULL builds restore WebConfig, display support, optional external sensors, and the full role CLI and feature set, full ElegantOTA where that target declares the required library, and LoRa OTA for every included role, including room servers, sensors, observers, and bridges. They use expanded A/B partition tables: 1984 KiB application slots on 4 MiB boards and the framework's larger dual-OTA tables on 8 MiB and 16 MiB boards. Explicit *_lora_ota_no_external_sensors targets are not duplicated; their ordinary repeater build is the FULL, sensor-enabled counterpart. The *-full-logging-ota-* profile enables USB debug and packet logging and explicitly disables MQTT. Install a matching *-full-ota-*-merged.bin or *-full-logging-ota-*-merged.bin over USB once to write the expanded partition table. After that, its matching non-merged FULL application image can be installed through USB, WiFi OTA, or LoRa OTA. Do not install a non-merged FULL image onto a node that still has its old partition table.

Implementer note: the bootloader (and any non-Arduino consumer) MUST locate the body extent by scanning for EndF, never by trusting a stored size - see the bootloader contract in Section 12.

---

3. The .mota container

The distributed form (host-built, wire-transferred). Parsed by mota_parse() in MotaContainer.cpp.

off            size   field
0              4      MAGIC = 6D 4F 54 41
4              4      MOTA_TOTAL_SIZE  uint32 LE - total container bytes (incl. manifest, leaves[],
                                       payload, trailer). Lets a node pre-reserve staging and compute
                                       write_start = staging_region_end - MOTA_TOTAL_SIZE.
8              M      MANIFEST         (Section 4; M = 197 fixed + leaves[], 4*BC; no length field - BC from payload_size)
8 + M          P      PAYLOAD          (payload_size bytes; delta or full image)
8 + M + P      5      TRAILER = 76 6B 34 39 36

MOTA_TOTAL_SIZE = 4 + 4 + M + P + 5. The manifest M includes leaves[]; the manifest-minus-leaves prefix (mfl, sent over the wire as OTA_MANIFEST) is [8, leaves_off).

Staged (in-flash) form. Written bottom-aligned so TRAILER ends at staging_region_end. Identical bytes, except the device mutates two regions in place (both NOR-safe, no re-erase): the leaves[] slots (filled as blocks arrive - Section 7) and the 4-byte approval field (on owner consent - Section 4.2). Everything else is immutable.

---

4. The manifest

Fixed layout. Every field sits at a constant offset and is always present - base_hash, signer_pubkey and signature are zero-filled when not applicable (a full image / an unsigned container). Only leaves[] is variable (one 4-byte hash per block). So the manifest-minus-leaves (mfl) is always 197 bytes and the parser is plain offset reads - no conditionals. Parsed by mota_parse_manifest().

off  size   field            notes
0    1      format_ver       = 0x02
1    1      flags            bit0 FULL (0=delta/partial, 1=full image); bit1 SIGNED; bits2-7 reserved 0
2    1      hash_algo        0x12 = sha2-256
3    4      target_id        device/arch/role discriminator (Section 9)
7    4      fw_version       MAJOR<<24 | MINOR<<16 | PATCH<<8 | pre   (comparable uint32)
11   4      image_size       size of the reconstructed image (BODY||EndF)
15   4      payload_size     PAYLOAD bytes in this container
19   1      block_size_log2  e.g. 0x0A = 1024
20   4      merkle_root      sha2-256:4 over PAYLOAD blocks (Section 6) - also the manifest_id
24   32     image_hash       sha2-256:32 of the reconstructed image - SECURITY anchor
56   1      codec_id         0=full/raw, 1=detools-sequential, 2=detools-in-place
57   32     hw_id            NUL-padded ASCII hardware tag (e.g. "RAK4631"); same tag => bootable-compatible.
                             SIGNED. Applier refuses a mismatch (brick-safety); empty on either side = skip.
89   8      base_hash        sha2-256:8 of the BASE image's BODY (== that build's EndF.body_hash). 0 if FULL.
97   32     signer_pubkey    Ed25519 public key. 0 if not SIGNED.
129  64     signature        Ed25519 over manifest[0, 129). 0 if not SIGNED.
193  4      approval         FF FF FF FF = not approved; 41 50 52 56 ("APRV") = approved
--- end of manifest-minus-leaves: mfl = 197 (constant); leaves_off = 8 + 197 = 205 in the container ---
197  4*BC   leaves[]         BC = ceil(payload_size / 2^block_size_log2). sha2-256:4 each (the only variable field)

The signature always covers manifest[0, 129) (the head + base_hash + signer_pubkey). approval is outside the signed region so it can be flipped in place on consent without breaking the signature.

Manifest-minus-leaves size (mfl) is a constant 197 bytes for every container (full or delta, signed or unsigned). At 197 bytes the manifest exceeds one packet, so OTA_MANIFEST is always sent multi-fragment (Section 8.4, 2 fragments) and reassembled by the fetcher.

4.1 Signed region

signature covers manifest bytes [0, 129) - the head + base_hash + signer_pubkey. It does not cover approval or leaves[]:

• leaves[] are verified against the signed merkle_root (Section 6), so they need no separate signature.
• approval is device-local consent (Section 4.2), deliberately outside the signature.

4.2 The approval field

• Distributed and forced on ingest to FF FF FF FF (a peer can never pre-approve).
• The local owner's ota applydelta writes 41 50 52 56 ("APRV") - a single NOR-safe write (only clears bits from the erased word). Any partial/other value reads as not-approved (fail-safe).
• Bound to this image (lives in this .mota's manifest, re-erased when a new .mota is staged).
• A consent marker, not a security primitive. Authenticity = signature + image_hash + hw_id.

---

5. Payload, codecs & delta base

PAYLOAD is either the full reconstructed image (FULL) or a delta (!FULL).

codec_id	Meaning	Used by
0	full / raw	PAYLOAD = reconstructed image (`BODY
1	detools sequential	random read of base + sequential write of result -> ESP32 A->B inactive slot.
2	detools in-place	bounded scratch; rewrites the app region in place -> nRF52 single-slot.

For deltas, base_hash = the base build's EndF.body_hash (sha2-256:8 of its BODY). A node applies a delta only if base_hash matches its own EndF.body_hash. After applying, the result MUST hash (sha2-256:32) to image_hash before it is booted - the hard security gate.

A fetcher only requests firmware it can apply. Each node declares the codec(s) it can apply (set_apply_codec/set_apply_codec2): ESP32 accepts full + sequential (+ in-place). Internal-staging nRF52 targets accept only in-place because internal flash cannot hold a second full application image. Matched SD and raw-QSPI nRF52 targets accept full + in-place because external media holds the container. A .mota with an unsupported codec is rejected at discovery time, before any blocks are requested. A manual pull to an external folder may accept other codecs because that path captures bytes and never installs them.

Compression is internal to the detools patch and must be supported by the applier. Patches are produced by detools 0.53.0 (tools/mota -> detools.create_patch) and decoded on-device by detools' embeddable C decoder, vendored verbatim at src/helpers/ota/detools/ (see its README.meshcore.txt). That build enables only the self-contained NONE + CRLE compressions (no malloc/liblzma/heatshrink), so MeshCore deltas use --compression crle. Do not reimplement the codec - use the vendored decoder.

---

6. Merkle tree (sha2-256:4)

Verifies each PAYLOAD block against the signed merkle_root before the whole payload exists, so corruption/forgery is localized to a block. Implemented in MerkleTree.cpp.

• Blocks: PAYLOAD splits into BC = ceil(payload_size / B) blocks, B = 2^block_size_log2 (default 1024). The last block is its real length (no zero padding).
• Leaf: leaves[i] = sha2-256:4( block_i_bytes ).
• Internal node: node = sha2-256:4( left || right ) (4+4 input bytes).
• Odd level: an odd count promotes the last node unchanged to the next level (no duplication).
• Root: reduce until one node remains. BC == 1 -> root = leaves[0]. BC == 0 is invalid.

6.1 Proofs

A proof for block i is the ordered list of sibling digests from leaf to root. Promoted levels contribute no element. Verification (needs BC to know the tree shape):

h = leaf_i ; idx = i ; n = BC ; p = 0
while n > 1:
    if (n is odd) and (idx == n-1):        # this node was promoted
        pass
    else:
        sib, side = proof[p] ; p += 1
        h = sha2-256:4( sib || h ) if side==left else sha2-256:4( h || sib )
    idx //= 2 ; n = (n + 1) // 2
accept iff h == merkle_root and p == len(proof)

Over LoRa, leaves[] are omitted from the manifest transfer. A serving node computes a block's proof on demand from its stored leaves[] and normally sends OTA_PROOF immediately after that block's paced OTA_DATA. OTA_REQ_PROOF remains the fallback for an older source or a lost proactive proof. The fetcher fills its own leaves[i] as each verified block lands.

---

7. Block availability, staging & resume

There is no separate availability structure. Block i is present <=> leaves[i] is non-erased (!= FF FF FF FF). Because leaves[] live in the staged flash region, availability survives reboot.

Commit order per block (crash-safe): (1) verify proof, (2) write block payload to its offset, (3) write leaves[i] last. A power loss before step 3 leaves the slot erased -> the block is simply re-fetched (idempotent). On boot a node rebuilds an in-RAM present-bitmap by scanning leaves[].

Resume (OtaManager::resumeStaged + OtaStore::checkpoint/reopen): an interrupted fetch resumes from the staged container after a reboot - re-parse the stored manifest, recompute geometry, count present blocks, continue fetching the holes (or jump straight to COMPLETE). The checkpoint cadence (persist progress every N committed blocks) is runtime-tunable (ota config checkpoint <N>, 0 = only finalized containers resume). Stores keep leaves[] in RAM until flush and never auto-GC, preserving resumable progress.

Flash-store note (RX-safe writes): a flash page-erase halts the CPU (~85 ms on nRF52) and starves LoRa RX, so the flash stores (OtaStoreFlashNrf52/OtaStoreFlashEsp32) coalesce writes to the erase unit (4 KB page / sector) and commit each once off the per-packet path - RAM stays O(one page), not O(image). A small delta that fits page 0 does zero flash I/O until COMPLETE.

---

8. LoRa OTA protocol

Carried in MeshCore packets with PAYLOAD_TYPE_OTA = 0x0C. Every OTA packet payload is:

[0]    ota_msg_type      (OtaMsgType, OtaFormat.h)
[1..]  body              (fixed per type; encode/decode in OtaProtocol.cpp)

Message types:

ota_msg_type	val	routing	purpose
OTA_ADV	0x01	discovery	tiny per-node beacon (discovery tier 1)
OTA_QUERY	0x02	discovery	ask a source for its catalog (discovery tier 2)
OTA_HAVE	0x03	discovery	the catalog reply (fragmented, digest-tagged)
OTA_GET_MANIFEST	0x04	transfer	request a manifest's fragments (want_mask) by manifest_id
OTA_MANIFEST	0x05	transfer	the manifest-minus-leaves, fragmented
OTA_REQ	0x06	transfer	request fragments from an adaptive flight of 1-4 blocks (want_mask per block)
OTA_DATA	0x07	transfer	one self-describing fragment of a block's data
OTA_REQ_PROOF	0x08	transfer	request/re-request a missing proof
OTA_PROOF	0x09	transfer	the merkle proof for one block
OTA_GET_LEAVES	0x0A	transfer	request the target's leaves[] fragments (want_mask) - warm-start only
OTA_LEAVES	0x0B	transfer	a fragment of the leaves[] array (for host-side seed leaf-diff)

• manifest_id = the manifest's merkle_root (4 bytes) - a compact content id present in every transfer message, so a multi-mota server dispatches each request to the right image.
• Priority: OTA_ADV, OTA_QUERY, and OTA_HAVE enqueue at background priority 250. Once a fetch is active, manifest, block request, data, and proof messages enqueue at primary priority 0. Relay-only nodes classify the wire message identically, so a transfer stays primary across the complete path.
• Reliability is eventual: the fetcher re-requests only missing fragments after an adaptive deadline derived from packet airtime, outstanding response packets, duty pacing, and path length. The manager may still run a one-second maintenance tick, but that tick is not itself a retry timer. No hard ACKs or global ordering are required.
• Relay envelope: OTA still uses a bounded flood-shaped mesh header so the same packets can cross the configured number of hops without first discovering an addressed return path. During TempRadio each node forwards one copy; active OTA packets do not use the generic flood-retry subsystem. The fetcher verifies every block against the signed root, and a repeater without ENABLE_OTA can transport PAYLOAD_TYPE_OTA opaquely without the manager, staging store, installer, or destination bootloader.
• Hop limit + duty cycle: OTA floods accumulate one path-hash per relay (the mesh's flood routing). A node with the OTA manager accepts a packet only if it arrived within ota config hops hops (default 3; 0 = direct only) and relays it only while still under that limit, appending its own hash. Relay-only repeaters instead use their ordinary flood limits and forwarding filters. Discovery relays remain lowest-priority and may be skipped when the packet pool runs low. Active-transfer relays bypass that background pool gate and use priority 0; operators should therefore treat TempRadio as a dedicated OTA maintenance window because the transfer can delay unrelated mesh traffic.

8.1 Two-tier discovery

Because a node may serve many mOTAs (its own firmware plus an external folder - Section 10), discovery is split so the periodic beacon stays tiny regardless of catalog size:

Tier 1 - OTA_ADV beacon (10 bytes, constant). Flooded as a short burst at boot, then every advert_mins minutes (default 24h; runtime-tunable via ota config advert, 0 disables the periodic re-advertise). It is also emitted immediately whenever the served set changes (e.g. a motatool folder is attached/detached), so peers learn about newly-available firmware without waiting for the next interval:

seeder_id[4]    advertiser node id = pubkey[0:4]; the QUERY address + distinct-source id
n_motas         uint8 - count of complete servable mOTAs (saturates at 255)
set_digest[4]   sha2-256:4 over the SORTED set of served manifest_ids (see below)

set_digest is a content hash of the offering, not a counter: canonical across nodes, and it changes iff the set of served mids changes. A peer that has already catalogued this {seeder, set_digest} ignores the beacon (steady state is query-free). For a single served mota, set_digest = sha2-256:4(mid).

Tier 2 - OTA_QUERY -> OTA_HAVE (on interest only):

OTA_QUERY  (flood):  seeder_id[4]  set_digest[4]  filter_target(uint32) want_fragments(uint32)
                     # filter_target 0 = everything; want_fragments 0 = every fragment
OTA_HAVE   (flood):  seeder_id[4]  set_digest[4]  frag_idx(1) frag_total(1) n_rows(1)  rows[]
  HaveRow (16 bytes, OTA_HAVE_ROW_BYTES): mid[4] target_id(4) fw_version(4) codec_id(1) flags(1) have_count(2)

have_count is the number of blocks the source holds (== block_count for a complete offered image). Receivers do not advertise partial or completed downloads as new sources.

A node interested in a source's offering schedules a QUERY; the source replies with its full catalog as OTA_HAVE rows (fragmented if they exceed one packet - 10 rows per fragment). A receiver marks the catalog complete only after all frag_total fragments arrive. If any are missing after the recovery timeout, it sends another QUERY whose want_fragments bitmap names only the holes. want_fragments is an append-only extension: an original 13-byte QUERY is still accepted and means "send every fragment." The heavy manifest is fetched per-mid only on commit (Section 8.3).

Fragment numbers are canonical pages of the complete catalog sorted by manifest_id. filter_target may remove rows from a requested page (and can therefore produce an empty fragment), but it never renumbers pages or changes frag_total. This keeps missing-fragment recovery unambiguous when filtered and unfiltered queries for the same {seeder, set_digest} are overheard together.

8.2 Anti-storm (mandatory at mesh scale)

If 50 neighbours all queried a new beacon at once, the mesh would collapse. Mitigations (gossip/mDNS pattern), all in OtaManager:

• OTA_HAVE is flooded and digest-tagged. EVERY node that overhears it caches the rows passively (keyed by {seeder, set_digest}) - no query of its own needed.
• Jittered query: a peer needing a catalog schedules its OTA_QUERY after a random delay OTA_QUERY_MIN_MS (300) + rand(OTA_QUERY_SPREAD_MS (4000)), derived from id +/ digest +/ self.
• Overhear suppression: during the jitter window, overhearing another QUERY that covers the same scope, or completing the HAVE fragment set for the same {seeder, set_digest}, cancels the pending query.
• Per-source recovery: each seeder has independent query/retry state. One source cannot overwrite another source's timer, and a partial reply requests only missing fragments after 15 seconds (five bounded retries, then another source ADV or explicit ota ls can start a fresh series).

Net effect: a digest change costs ~1 query + ~1 HAVE flood mesh-wide; a stable mesh is query-free.

8.3 Fetch handshake

fetcher                                   server (any node that has the mid)
  OTA_GET_MANIFEST(mid, want_mask)  >     (want_mask=0xFFFF first; only missing fragments on retry)
                               <-------   OTA_MANIFEST(mid, frag_idx, frag_total, bytes)   x requested frags
  (reassemble manifest, verify, compute geometry: BC, block_size, payload_size)
  for each adaptive flight of missing blocks (starts at 1, grows on clean flights):
    OTA_REQ(mid, {block_idx, want_mask}[]) >  (one packet; all fragments first, only holes on recovery)
                               <-------   OTA_DATA(mid, block_idx, frag_off, data) x requested frags/block
                               <-------   OTA_PROOF(mid, block_idx, n_proof, proof) x requested blocks
    (independently reassemble + verify each block, but remain RX-silent until the flight drains)
    [after adaptive deadline: recover one block's holes, or OTA_REQ_PROOF for a missing proof]
    (clean flight grows by one block; recovered flight halves the next width)
  when all blocks present: verify full merkle_root + image_hash -> COMPLETE

8.4 Message bodies (transfer)

All offsets after the 1-byte type. Encoders/decoders in OtaProtocol.cpp; constants in OtaManager.h.

OTA_GET_MANIFEST:  manifest_id[4]  want_mask(uint16)   # bit k = send manifest fragment k; 0xFFFF = all
OTA_MANIFEST:      manifest_id[4]  frag_idx(1)  frag_total(1)  bytes[]     # up to OTA_MF_FRAG=176 B/frag
OTA_REQ:           manifest_id[4]  { block_idx(uint16)  want_mask(uint16) }[1..4]
                   # one or more rows; bit k = send fragment k of that block
OTA_DATA:          manifest_id[4]  block_idx(uint16)  frag_off(uint16)  data[]   # up to OTA_FRAG_DATA=160 B
OTA_REQ_PROOF:     manifest_id[4]  block_idx(uint16)
OTA_PROOF:         manifest_id[4]  block_idx(uint16)  n_proof(1)  proof[]   # n_proof x 4 bytes
OTA_GET_LEAVES:    manifest_id[4]  want_mask(uint16)   # bit k = send leaves fragment k; 0xFFFF = all
OTA_LEAVES:        manifest_id[4]  frag_idx(1)  frag_total(1)  bytes[]      # up to OTA_LEAVES_FRAG=176 leaf bytes

• Warm-start / leaf-diff (OTA_GET_LEAVES/OTA_LEAVES) - motatool folder-capture only. Capturing a device's firmware into a motatool serve folder is slow (a full image is hundreds of blocks). Because builds here are non-deterministic, you cannot reproduce the exact target on the host - but a similar build (e.g. a fresh recompile) is ~99% identical. So motatool serve --seed <similar.mota> stages that build's payload into the destination .part, and ota pull <mid8> folder validate makes the fetcher (1) bulk- fetch the target's leaves[] via OTA_GET_LEAVES/OTA_LEAVES (bitmap-fragmented with a want_mask, same anti-burst rule as OTA_MANIFEST), (2) recompute the merkle root from them and check it equals the manifest root (authenticate), then (3) keep every seeded block whose leaf matches and pull full OTA_DATA only for the blocks that differ. The want_mask is a fixed uint16, so leaves[] is capped at OTA_LEAVES_MAXFRAG=16 fragments (OTA_DIFF_MAX_BLOCKS=704 blocks); larger images just fall back to a full fetch. Normal P2P nodes never use this - they target only the blocks they want; the only always-on part is answering OTA_GET_LEAVES with leaves the node already holds, so any node's firmware can be captured.

• Block <-> fragments: a 1 KB block remains split into self-describing OTA_DATA fragments. frag_off is the byte offset of data within the block, so the global position is block_idx*block_size + frag_off - a fragment is self-placing when returned by the source. The fetcher tracks a per-block slice bitmap and reassembles before requesting the proof.

• Adaptive flight size is not signed block size. The container continues to use 1 KiB Merkle leaves and each slot is one existing 1 KiB block. A clean link changes how many of those blocks one OTA_REQ names: 1, then 2, then 3, then the compiled cap. The manifest stores block_size_log2, so 3 KiB is not a valid logical geometry; enabling 2 KiB would require larger device reassembly buffers and new-package/bootloader validation while saving only one request/proof pair per 2 KiB. It does not address premature retries, which were the dominant packet multiplier.

• Append-only request-window compatibility: the first four-byte request row is exactly the original block_idx + want_mask body. Old sources decode that row and ignore appended bytes. New sources queue all rows. If a new fetcher meets an old source, the unserved tail rows eventually time out and are recovered as ordinary single-row requests; the dirty flight then contracts. Old fetchers continue sending nine-byte single-row requests, which new sources accept normally.

• Fragment-level requests (anti-deadlock + anti-congestion): OTA_REQ, OTA_GET_MANIFEST, and OTA_QUERY carry fragment masks. For catalog discovery, want_fragments is a 32-bit bitmap and covers the protocol maximum 255-row catalog (26 fragments at the current packet size). For block and manifest transfer, the want_mask is 16 bits. A fetcher requests the full set on the first ask ((1<<nf)-1, or 0xFFFF before frag_total is known) and only the still-missing bits on any retry, so recovering one lost fragment re-sends one fragment, not the whole block/manifest. This is essential on half-duplex radios: re-requesting a whole multi-fragment burst let the periodic retry (a transmit) collide with the tail of the in-flight burst and drop the same fragment forever - a hang. Requesting only the hole removes the burst, so there is nothing to collide with. The block/manifest mask matches the 16-bit reassembly bitmap (<=16 fragments/block; 1 KB blocks = 7). OTA_PROOF is a single packet and needs no mask.

• Data and proof remain separate packets, without a normal extra round trip. A server retains requested blocks in a bounded descriptor queue, admits at most one response per main-loop pass, and sends one OTA_PROOF after each block's requested fragments. Before admitting that proactive proof, the source leaves an airtime/duty-aware 100-3000 ms RX turnaround gap. It accounts for one active transmission plus the two paced-response queue credits. A legacy receiver uses the gap to send its immediate OTA_REQ_PROOF; receiving that explicit request bypasses the remaining gap, avoiding a proof/request collision and an otherwise multi-second legacy retry. A newer receiver uses an airtime/path-aware proof grace (never less than 500 ms) and defers OTA_REQ_PROOF while any flight slot still expects DATA. Thus a missing early proof cannot make the receiver transmit into the rest of a legitimate half-duplex response train. After the complete-flight deadline, only one slot's missing fragments/proof is requested at a time.

8.5 Sizing against MAX_PACKET_PAYLOAD = 184

message	fixed overhead	payload/packet
OTA_DATA	9 B (type+mid4+idx2+off2)	OTA_FRAG_DATA = 160 -> 7 frags per 1 KB block
OTA_MANIFEST	7 B	OTA_MF_FRAG = 176 -> signed manifest ~ 2 frags
OTA_HAVE	12 B	10 rows x 16 B per fragment
OTA_PROOF	8 B	up to ~44 sibling digests (>> any real tree)

A served mota supports up to OTA_MAX_BLOCK/4 leaves in the default 4 KB proof scratch (<=1024 blocks ~ 1 MB payload); larger self-images pass a bigger scratch buffer.

8.6 Temporary-radio and transfer boundary

OTA packets may cross normal mesh relay hops, but each participating node processes or relays them only while its tempradio window is actually running. A receiver selects missing blocks in serial order into a bounded request flight. Every session starts with one block. A clean completed flight increases the next request by one block; a flight requiring fragment/proof recovery halves the next width (4 -> 2, 3 -> 2, 2 -> 1). The default compiled cap is two blocks; the RAK3401 LoRa-OTA target caps at four, so it probes 1 -> 2 -> 3 -> 4. All rows are sent in one backward-compatible OTA_REQ, and no freed slot is refilled until the current flight is finished. It never serves partial blocks. A normal install receiver never re-advertises its completed download. An SD archive node is the deliberate exception: after a fully proof-verified container is published to its persistent archive, it registers that complete file as a MotaSource and advertises it as a new seeder. This keeps each active transfer as one transmitter and one receiver while still allowing active temporary-radio repeaters between them and persistent archive nodes to improve future availability.

TempRadio is treated as a private maintenance network. Active transfer packets use priority 0, bypass the public-flood receive holdoff, use the full transmit budget without overwriting the saved normal-radio airtime factor, retain the relay role's airtime-scaled transmit collision window, and do not schedule generic flood retries. Deployed firmware predating that TempRadio budget override can be accelerated manually with a saved get af / temporary set af 0 / restore sequence. The bounded serving queue admits at most two DATA/PROOF packets ahead of the radio while preserving at least four free packet-pool entries. CAD remains enabled to arbitrate the half-duplex channel, but its busy retry is scaled to one-quarter of a packet airtime and clamped to 5-50 ms instead of the ordinary 120-360 ms cadence. Discovery traffic keeps collision jitter and background priority. The fetch deadline uses the active radio's measured maximum-packet airtime, remaining DATA/PROOF packet count, dispatcher airtime factor, and longest observed path (falling back to the configured hop horizon before one is observed), with bounded guard time. Faster SF/BW settings therefore recover loss sooner; slower or multi-hop settings do not spuriously re-request a response still on air. These changes remove software waits and duplicate bursts; they do not remove the one required forwarding transmission per hop.

---

9. Identity, trust & versioning

• target_id (4 B): sha2-256:4(pio_env_name) (little-endian uint32). The env name uniquely captures hardware and role/partition, so a node auto-fetches only matching firmware (a companion image is not fetched onto a repeater even though it shares hw_id). It is self-described in the firmware's EndF (Section 2, written by pio_endf.py) and read via ota_self_firmware(), so it is correct on any build; -D MOTA_TARGET_ID / MainBoard::getOtaTargetId() is the fallback when no EndF identity is present. tools/mota reads it from the firmware's EndF (or --target-env). A manual ota pull/want can override target (deliberate role switch); the hw_id brick-safety gate (Section 4) still applies at apply time.
• target_id vs hw_id - complementary, not redundant: target_id is the fetch-routing key (hw + role + partition); hw_id is the human-readable brick-safety key (hardware only). Same board, two roles => same hw_id, different target_id.
• Naming a target_id locally: only the 4-byte target_id ever travels on the wire. To show which board/role a target is, a node (and motatool) reverse-looks-it-up in src/helpers/ota/OtaTargets.h - a generated target_id -> env-name table covering every ENABLE_OTA env (tools/mota/gen_targets.py, resolved from pio project config). So ota ls can render [Heltec_v3_repeater] for a neighbour's beacon without the string being transmitted. Unknown IDs show as raw hw XXXXXXXX / N/A values.
• fw_version: packed comparable uint32 (MAJOR<<24 | MINOR<<16 | PATCH<<8 | pre); also self-described in EndF. ota ls prints the stable eight-hex manifest ID and uses [same target], [unsupported], or [rescue] after combining target equality with the local codec, bootloader, and EndF preflight. A known different target is rendered by environment name; an unknown/unset target remains raw or ?. Target equality is routing information, not by itself an assertion that an image is safe to install.
• hw_id: 32-byte NUL-padded ASCII hardware tag inside the signed head. The applier refuses a .mota whose hw_id differs from the device's own tag (empty on either side = permissive). Brick-safety independent of signature.
• Signing & allowlist: a node keeps a runtime allowlist of trusted Ed25519 signer pubkeys (none embedded in firmware; ota key add/list/rm). A .mota is eligible for auto-install only if signed by an allowlisted key, the signature verifies, and image_hash matches. Manual install permits unsigned packages, but a package that claims to be signed must have a valid signature from an allowlisted key or it is rejected. Transfer needs no trust - blocks are content-addressed against the manifest's merkle root.
• Policies (persisted): autofetch in {off, any, signed} (default off) gates automatic block fetching of own-target adverts; autoinstall in {off, trusted} (default off) gates auto-apply of a COMPLETE signed + allowlisted fetch. Conservative defaults: a fresh node discovers + announces but never fetches/installs without operator intent.
• Supersession: a newer version announced mid-download does not abort the in-progress transfer (finish-current).

---

10. Multi-mota serve & the external "folder" relay

A node serves a set of mOTAs: its own firmware plus, optionally, an external folder of .mota files it relays without holding them in flash. To peers it simply "has N mOTAs"; the relay is trustless (fetchers verify everything). The serve side (OtaManager) keeps a lightweight registry of what it advertises and two resident "views": view0 (its own firmware) and one on-demand view loaded from a source when a request targets an external mota. Every fetch message carries manifest_id, so dispatch is a registry lookup.

The same host-folder link is also a pull destination (the reverse direction): ota pull <mid8> folder fetches a .mota off the mesh and streams it onto the host as <mid>.mota via the seeder STORAGE ops (OP_STAT/BEGIN/WRITE/SREAD/FIN, see MotaSeederProto.h), using a FolderMotaStore as the fetch's OtaStore instead of RAM/flash. This captures an exact copy of a device's firmware - e.g. to build a delta against firmware you don't have. Resume is bookkeeping-free: BEGIN 0xFF-fills the file and, on reconnect after a link drop (the fetch PAUSES, holding progress on the host - no RAM/flash fallback), STAT+SREAD let the fetcher recompute and refill only the missing blocks.

10.1 The MotaSource abstraction (OtaSource.h)

Transport-agnostic provider of one or more complete .mota as random-access bytes. The same serve code drives USB-serial, BLE, a WiFi URL list, an NFS/samba mount, etc. - only read() differs.

struct MotaDesc {                      // catalog metadata + region offsets (no whole image in RAM)
  uint8_t mid[4]; uint32_t target_id, fw_version; uint8_t codec_id, flags;
  uint32_t total_size, leaves_off, block_count, payload_off, payload_size;
};
class MotaSource {
  virtual uint8_t count();                                  // # mOTAs offered
  virtual bool    describe(uint8_t idx, MotaDesc& out);     // metadata + offsets
  virtual bool    read(uint8_t idx, uint32_t off, uint8_t* buf, uint32_t len);   // random-access bytes
};

To serve an external mota the node reads its manifest-minus-leaves + leaves[] into RAM (<=4 KB for <=1024 blocks) and streams payload blocks from the source on demand; proofs are generated from the read leaves.

10.2 The mota-seeder transport (MotaSeederProto.h)

A MotaSource is fed by a host that serves a folder over the device's USB serial (the same console the CLI uses - no extra hardware) or, on an ESP32 WiFi companion or FULL ESP32 role, over WiFi (TCP). The host is the standalone Rust tool motatool (motatool serve --serial <port> / --tcp <host[:port]>, which also builds + verifies + inspects .mota). The device only emits request frames while actively serving a fetch, and reads the reply synchronously, so over the shared USB console binary frames coexist with the text CLI/logs (resync on magic + checksum). Little-endian, XOR-checksummed:

request  (device -> host):  'M' 'S'  op(1)  args...                 xsum(1 = XOR of op+args)
response (host -> device):  'm' 's'  op(1)  status(1)  payload...    xsum(1 = XOR of all prior)

OP_COUNT     0x01   args: -            -> payload: count(1)
OP_DESCRIBE  0x02   args: idx(1)       -> payload: MotaDesc wire (38 B)
OP_READ      0x03   args: idx(1) off(4) len(2)  -> payload: len bytes
MotaDesc wire (38 B): mid[4] target_id(4) fw_version(4) codec(1) flags(1)
                      total_size(4) leaves_off(4) block_count(4) payload_off(4) payload_size(4)
                      block_size_log2(1) reserved(3)
status: 0 = OK, non-zero = error (out of range / past EOF).

SerialMotaSource splits logical reads into replies of at most 192 payload bytes. A manifest's leaf table can exceed 256 bytes and payload blocks are normally 1 KiB; requesting either in one transaction can overrun common USB CDC/UART receive rings even though the host successfully wrote the complete reply. Chunking is internal to the transport and does not change OP_READ or the MotaSource random-access contract.

Manifest fragments are retained as bounded response jobs and admitted one at a time. Their source-side gap follows the active maximum packet airtime and dispatcher duty spacing, clamped to 100-1000 ms. The 100 ms floor protects fast radios' TX-to-RX turnaround; the cap keeps the receiver's one-second manifest progress/retry observation responsive. The source uses the same radio-aware 100-3000 ms drain/turnaround gap before an unsolicited block proof, but immediately serves a legacy receiver's explicit OTA_REQ_PROOF. Relay collision delay is a separate setting: active OTA floods honor the relay role's configured txdelay, and the deployment runner temporarily uses 0.3 on managed relays.

What to plug into --serial. Use the USB serial console of an OTA-enabled MeshCore node built with OTA_FOLDER_SERIAL. The node must have a working LoRa radio plus an ota folder on command; that command confirms it can host and advertise the folder. A KISS modem will not work: KISS firmware exposes a TNC/KISS frame interface, not the MeshCore CLI and mota-seeder request/response transport. An ESP32 WiFi companion or FULL ESP32 role with active WiFi is the alternative source connection: use its dedicated seeder port with motatool serve --tcp <host>:5001.

Device CLI: ota folder on (attach + announce), ota folder (list), ota folder off. Build flag OTA_FOLDER_SERIAL (default stream = console Serial; override OTA_FOLDER_SERIAL_STREAM + define OTA_FOLDER_SERIAL_BEGIN for a dedicated UART). ESP32 WiFi companions and FULL ESP32 roles run a WiFiServer on the dedicated seeder port (OTA_SEEDER_TCP_PORT, default 5001) while WiFi is usable. On a companion it is separate from the app port (TCP_PORT, default 5000); on infrastructure roles it is separate from WebConfig and browser OTA on port 80. The node auto-attaches the source when a seeder client connects and detaches when it closes (no ota folder on needed over TCP). An already-active serial folder causes a TCP client to be rejected instead of silently replacing it. Verified on hardware: a RAK4631 relays a host folder to a Heltec V3 over one USB cable, and a host feeds a Heltec V3 over WiFi (:5001) while the companion serves a phone on :5000 - every block merkle-checked.

The attach reply and bare ota folder report host=advertised/offered. The registry is RAM-bounded (OTA_MAX_SERVE, with the node's own firmware consuming one slot), so a host may correctly index more valid files than this particular firmware can advertise. Omitted entries are now reported instead of silently disappearing. Operators should split a large chain or use a higher-capacity/SD seeder when the two counts differ.

Transport-agnostic by design. The request/response semantics (COUNT / DESCRIBE(idx) / READ(idx, off, len) over a folder catalog) are independent of the link. The 2-byte magic + XOR checksum + resync framing above exists for the shared USB-UART (an unframed byte stream); it is harmless over a reliable stream and the WiFi (TCP) transport reuses it as-is - both ends just treat the socket as a byte stream (on-device, SerialMotaSource runs verbatim over an Arduino Stream-compatible WiFiClient; motatool's TcpTransport mirrors its SerialTransport). A future framed link such as BLE GATT (an Android phone relaying a folder) could carry the same ops with no magic/checksum at all - a request characteristic write delivers op + args, the reply notifies status + payload. motatool reflects this split: a transport-free SeederCore (the catalog logic) under a swappable framing/transport layer.

---

11. CLI surface (OtaCli.cpp)

User-facing OTA data should travel via CMD_OTA_* companion binary frames; the text CLI below is debug/operator oriented and replies are snprintf-bounded into a 160-byte buffer.

Commands take intuitive aliases (matched by the first word; see is_cmd in OtaCli.cpp) so they're easy to type and read - status/neighbors/pull/drop/applydelta are the canonical names, the aliases are the recommended user-facing forms. Output is plain-language (a user-facing guide lives at ota_user_guide.md).

ota help | ? | h                   list the commands
ota status | st  (or bare `ota`)   plain-language: running fw, the one fetch session (state/%/id), serving, keys
ota ls | neighbors | nbrs | updates | n [page]   paged updates (queries sources; rows arrive async via OTA_HAVE)
ota get | pull | download <mid8|#index> flash [rescue] | folder [validate]
                                      fetch by stable mid8 (preferred) or current page index
ota install | apply | applydelta   verify + approve + (ESP32) apply / (nRF52) reboot-to-bootloader
ota rescue install <base_hash16>  internal-flash nRF52 only: recover from failed app-side EndF validation
ota cancel | drop | stop           drop the current fetch session (frees the slot)
ota announce | adv                 serve self + send a beacon now
ota self | id                      print this firmware's EndF (body/image size, base_hash)
ota folder | fold [on|off]         attach/detach an external .mota folder (host daemon) ; bare = list
ota config | cfg | set [autofetch|autoinstall|checkpoint] ...   show/set persisted policy
ota key | keys [add|rm <hex>]      trusted signer allowlist ; bare = list
ota dev ...                        bring-up helpers (stage/recv/serve/verify)

---

12. Apply & bootloader contract

• ESP32 (A/B): applied in-firmware via the detools decoder into the inactive OTA slot (OtaApply.cpp::ota_apply_detools_mota + OtaStoreFlashEsp32), then set-boot + reboot (power-safe, rollback-capable). No bootloader changes. Erase ranges must be sector-aligned (4096).
• nRF52 (single-slot): the running firmware never flashes the app. ota install verifies the container fully (image_hash, codec, signature/allowlist, hw_id, and base_hash for a delta), writes approval = "APRV", then reboots into the modified bootloader (Adafruit_nRF52_Bootloader_OTAFIX). The bootloader:
  1. locates the staged .mota in the approved internal, raw-SD, or raw-QSPI store without trusting an unchecked stored size,
  2. re-checks TRAILER, image_hash, and approval == "APRV"; for a delta it also checks that base_hash equals the running firmware's EndF.body_hash (recomputed by scanning for EndF - never trust bank_0_size),
  3. writes a full external-media payload or applies the in-place codec over the app region, then boots only if the result hashes to image_hash.
• nRF52 internal staging ceiling: an internal-store application derives the ceiling from facts available in every build, not a board-name list. A companion that actually links the internal ExtraFS datastore stays below 0xD4000; a default linker region or a role that does not mount ExtraFS can reclaim the unused 100 KiB through 0xED000. The application uses the larger window only when the installed bootloader advertises the GPREGRET2 ceiling-handoff capability. The bootloader treats every unknown/legacy handoff value as 0xD4000, and accepts a container only at the bottom-aligned position for the selected ceiling.
• nRF52 dynamic apply window: the post-build hook records the resolved app base, linked app end, internal-ExtraFS/SD/QSPI storage flags, and desired staging ceiling immediately before EndF. motatool reads that authenticated firmware record and chooses memory_size from the actual patch size and bottom-aligned stage address; firmware without the record retains the conservative 0x98000 default. Before writing APRV, an internal-store app validates the staged-address bound; an external SD/QSPI app validates the full detools geometry against the application workspace. The bootloader independently parses and validates the same geometry before its first application write. Expanded auto-sized packages require a bootloader with the ceiling-handoff capability; use --inplace-memory 0x98000 when intentionally targeting an older bootloader and the images still fit that window.
• nRF52 EndF rescue: ota rescue install <base_hash16> is a pre-provisioned recovery path for an internal-flash nRF52 application that still runs but cannot validate its own EndF identity. It refuses when normal EndF validation succeeds, requires the operator hash to exactly equal the staged delta's base_hash, requires the package target_id to match and its hw_id to pass the normal hardware gate, and retains the normal payload and signature/allowlist gates. Approval only delegates the base decision: OTAFIX independently locates the physical EndF, hashes the running app, and compares that value with the package before its first app write. A physically absent/corrupt EndF or wrong base therefore returns to the unchanged app; it still requires USB recovery if that app does not already contain this command. A chain intended to cross historical firmware must introduce this command in its first bridge and retain it in every later bridge. Manual pulls still use the build-provided target ID when app-side EndF parsing fails, so a rescue-capable bridge can fetch its exact successor before invoking the guarded command. Such a node must acknowledge the condition up front with ota pull <mid8> flash rescue; an ordinary flash pull refuses before altering staged data. Firmware that predates both rescue commands still requires USB recovery.
• MeshTower V2 SD nRF52: the application stores a contiguous /meshcore-ota.mota on microSD and publishes its raw sector range in a checksummed handoff record outside the MBR partition. The matching bootloader reads the card without mounting FAT, supports either a full image or an in-place delta, verifies the staged/full result hash, and never writes through 0xED000 where InternalFS begins.
• Matched external-QSPI nRF52 repeaters: the application reserves the board's dedicated QSPI NOR as a raw store beginning at offset zero. It obtains a 1-16 MiB capacity from JEDEC, checkpoints payload before leaf metadata, and verifies each erased/programmed page. GPREGRET2 0x51 selects QSPI only when the matching bootloader advertises the QSPI storage bit; legacy markers retain the internal scan path. The bootloader pre-hashes a full payload before invalidating the app, or applies an in-place delta with the complete internal application region as workspace. Companion builds never enable this raw store: some use QSPI as a filesystem, while others simply leave that chip outside OTA ownership. See the nRF52 QSPI guide.

A signature, when present, proves author authenticity and must pass the device allowlist. Unsigned packages remain installable when local policy permits them. The one-shot approval marker records local consent for either form before the bootloader may apply it.

Bootloader testing note: always test apply with a real different image (base != target). A same-image (X->X) "delta" trivially reproduces the target and gives a false positive.

---

13. Versioning of this spec

format_ver = 2. A parser accepts exactly this value and rejects anything else - there is one container format, fixed-layout, and no compatibility shims to carry. If the format ever needs to change, bump format_ver; the multihash hash_algo separately allows swapping the digest family without a format bump. Unknown codec_id / ota_msg_type values are ignored (a node simply won't fetch what it can't apply).