amethyst/quartz/plans/2026-05-08-local-headers-explorer.md

Local Bitcoin headers explorer for NIP-03 OTS

Date: 2026-05-08 (revised 2026-05-19) Branch: claude/review-bitcoin-blockchain-plan-6NGZ5 (originally claude/review-ots-blockchain-deps-bKns7) Module: quartz/ (with thin Android wiring in amethyst/) Status: Plan — not yet implemented

Revision log

2026-05-19 — decisions locked, scope tightened. After review against the intervening 2026-05-14 onchain-zaps work:

• Scope stays OTS-only. Trustless verification of NIP-BC onchain zaps (kind 8333) needs more than headers — at minimum BIP-37 merkleblock or full-block fetch on top of this stack. That becomes a follow-up plan (see §20).
• Module moves to quartz/.../bitcoin/ — sibling of nip03Timestamp/ and nipBCOnchainZaps/, not nested under NIP-03. The future merkle-block phase will reuse BitcoinPeer / HeaderStore / HeaderValidator.
• Storage is SQLite via androidx.sqlite + BundledSQLiteDriver, matching the existing SQLiteEventStore (nip01Core/store/sqlite/). No flat append-only file; no hand-rolled height index. See §9.
• No bundled headers file. Ship a single hardcoded Checkpoint(height, blockHash, chainWork, time) constant, bumped at release. First-run sync starts at the checkpoint and pulls forward to the tip over P2P. See §7.
• Pre-checkpoint OTS proofs fall through to HTTP (OkHttpBitcoinExplorer). No backward-from-checkpoint sync. Strict-mode users get a clear error instead of opening an HTTP connection.
• BitcoinExplorerEndpoint (amethyst/.../model/nip03Timestamp/) is now the shared Esplora-URL resolver for OTS and onchain zaps. The composite resolver builder must read from it for the HTTP fallback path.

1. Motivation

NIP-03 (OpenTimestamps attestations, kind 1040) is currently verified by hitting a third-party block explorer (blockstream.info or mempool.space) over HTTPS. That explorer is a trusted oracle: it can lie about a block's merkleRoot and silently invalidate or fabricate a timestamp.

OTS verification only needs one piece of chain data per attestation: the 80-byte block header at a given height. From that header we use:

• merkleRoot — to compare against the digest produced by the proof's ops,
• time — to report as the proven timestamp.

Nothing else. No transactions, no UTXOs, no scripts, no witness data, no compact filters. The OTS proof file itself carries every other byte of witness (calendar tree siblings, the prefix/suffix of the on-chain transaction, and the merkle path from that tx up to the block root) inline as operands of APPEND/PREPEND ops.

The Bitcoin P2P protocol exposes a first-class getheaders/headers message pair specifically for downloading just headers; Bitcoin Core has used it for "headers-first" sync since v0.10 (2014). This plan implements a headers-only client that lets Amethyst verify NIP-03 attestations against the actual proof-of-work chain instead of a trusted explorer.

2. Goals & non-goals

Goals

• Verify any NIP-03 OTS attestation without trusting any single party other than Bitcoin proof-of-work (and the app build itself, via a checkpoint).
• Run on Android phones; budget ≤100 MB disk, ≤1 % battery/day, ≤10 min first-run sync on WiFi.
• Plug into existing BitcoinExplorer interface — zero changes to OTS verification logic.
• Work over Tor when the user has Tor enabled.
• Keep existing HTTP explorers as a fallback for proofs older than the pinned checkpoint, or when local sync hasn't completed yet.
• Live in quartz/src/commonMain/.../bitcoin/ so cli/, desktopApp/ and any future merkle-block / onchain-zap work inherit the same primitives.

Non-goals

• We are not building a full Bitcoin node. No mempool, no UTXO set, no block bodies, no script verification.
• We are not implementing wallet features (BIP-32 keys, BIP-37 bloom, BIP-157/158 compact filters). NIP-03 doesn't need them.
• We are not replacing OkHttp explorers entirely — they remain a fallback path and the way unsynced proofs get verified.
• No support for testnet/signet (mainnet only — NIP-03 attestations are on mainnet).

3. Architecture overview

┌────────────────────────────────────────────────────────────────────┐
│                         Existing OTS code                          │
│  OtsEvent.verify(resolver) ──► OpenTimestamps.verify ──► explorer  │
│                                              │                     │
│                                              ▼                     │
│                                    BitcoinExplorer  (interface)    │
└────────────────────────────────────────────────────────────────────┘
                                      ▲
                       ┌──────────────┼──────────────┐
                       │              │              │
              OkHttpBitcoinExplorer   │   LocalHeadersBitcoinExplorer  ◄── NEW
              (existing, fallback,    │              │
               URL from              │              ▼
               BitcoinExplorer-      │   ┌─────────────────────────┐
               Endpoint)             │   │   HeaderStore           │
                                      │   │   (SQLite, androidx-    │
                                      │   │    sqlite + Bundled-    │
                                      │   │    SQLiteDriver)        │
                                      │   └─────────────────────────┘
                                      │              ▲
                                      │              │
                                      │   ┌─────────────────────────┐
                                      │   │   HeadersSyncEngine     │
                                      │   │ (validates PoW + chain) │
                                      │   └─────────────────────────┘
                                      │              ▲
                                      │              │
                                      │   ┌─────────────────────────┐
                                      │   │   PeerPool              │
                                      │   │ (multiple BitcoinPeer)  │
                                      │   └─────────────────────────┘
                                      │              ▲
                                      │              │
                                      │   ┌─────────────────────────┐
                                      │   │   BitcoinPeer           │
                                      │   │ (TCP socket + framing + │
                                      │   │  version/getheaders)    │
                                      │   └─────────────────────────┘
                                      │              ▲
                                      │              │
                                      │   ┌─────────────────────────┐
                                      │   │   PeerDiscovery         │
                                      │   │ (DNS seeds, fixed seeds)│
                                      │   └─────────────────────────┘
                                      │
                            CompositeBitcoinExplorer
                            (try local first; if height
                             below pinned checkpoint OR not yet
                             reached by sync, fall back to HTTP
                             via BitcoinExplorerEndpoint)

4. Module layout

All new code in quartz/src/commonMain/.../bitcoin/, sibling of nip03Timestamp/ and nipBCOnchainZaps/. Protocol code is platform-independent; socket I/O is the only expect/actual boundary. Storage uses androidx.sqlite which is already KMP — no platform split.

quartz/src/commonMain/kotlin/com/vitorpamplona/quartz/bitcoin/
├── proto/
│   ├── BitcoinNetMagic.kt          # 0xf9beb4d9 mainnet
│   ├── MessageHeader.kt            # 24-byte envelope codec
│   ├── VarInt.kt                   # CompactSize encode/decode
│   ├── NetAddr.kt                  # net_addr (26B / 30B with timestamp)
│   ├── ServiceFlags.kt             # NODE_NETWORK, NODE_NETWORK_LIMITED, NODE_WITNESS
│   ├── messages/
│   │   ├── VersionMessage.kt
│   │   ├── VerackMessage.kt
│   │   ├── PingMessage.kt
│   │   ├── PongMessage.kt
│   │   ├── GetHeadersMessage.kt    # locator + stop hash
│   │   ├── HeadersMessage.kt       # up to 2,000 × (80B header + varint=0)
│   │   └── SendHeadersMessage.kt   # BIP-130
│   └── BitcoinMessage.kt           # sealed interface
├── header/
│   ├── BlockHeader80.kt            # 80-byte parsed header (ver, prev, merkle, time, bits, nonce)
│   ├── BlockHash.kt                # @JvmInline value class wrapping ByteArray(32)
│   ├── HeaderHasher.kt             # dsha256 over 80B → blockHash
│   ├── DifficultyTarget.kt         # bits ↔ target (256-bit) compact form
│   ├── ChainWork.kt                # cumulative work accumulator (uint256 BE)
│   └── PowValidator.kt             # checks hash ≤ target
├── consensus/
│   ├── RetargetCalculator.kt       # every 2016 blocks, BIP-? rules
│   ├── MedianTimePast.kt           # last-11 median, for time sanity
│   └── HeaderValidator.kt          # composes the above + linkage check
├── store/
│   ├── HeaderStore.kt              # public API: append/getByHeight/getByHash/tip
│   ├── SqliteHeaderStore.kt        # androidx.sqlite impl (commonMain — driver-agnostic)
│   ├── HeaderStoreSchema.kt        # IModule for schema + migrations (see §9)
│   ├── Checkpoint.kt               # (height, blockHash, chainWork, time)
│   └── PinnedCheckpoint.kt         # hardcoded release-time checkpoint constant
├── sync/
│   ├── LocatorBuilder.kt           # exponential locator
│   ├── HeadersSyncEngine.kt        # drives getheaders loop, applies validator, persists
│   └── ReorgHandler.kt             # rare; switches to higher-chainwork tip
├── peer/
│   ├── PeerSocket.kt               # expect (open / send / recv / close)
│   ├── PeerPool.kt                 # holds N BitcoinPeer connections
│   ├── BitcoinPeer.kt              # high-level: handshake, send msg, receive msg
│   ├── PeerScoring.kt              # ban misbehaving peers
│   └── PeerDiscovery.kt            # expect — DNS seeds + fixed seeds
└── LocalHeadersBitcoinExplorer.kt  # implements nip03Timestamp.ots.BitcoinExplorer
                                    # (with small LRU on top, like OtsBlockHeightCache)

quartz/src/jvmAndroid/kotlin/com/vitorpamplona/quartz/bitcoin/
└── peer/
    ├── TcpPeerSocket.jvmAndroid.kt     # actual TCP socket (java.net.Socket / SOCKS)
    └── DnsSeedDiscovery.jvmAndroid.kt  # InetAddress.getAllByName on seed hostnames

Two expect/actual seams only: PeerSocket (TCP plus optional SOCKS5 for Tor) and PeerDiscovery (DNS-seed resolution). Header validation, message codecs, store, sync engine, peer pool — all compile in commonMain, all testable cross-platform without a single platform shim. Storage compiles in commonMain too because androidx.sqlite + BundledSQLiteDriver are KMP.

Wiring on Android

amethyst/src/main/java/com/vitorpamplona/amethyst/model/nip03Timestamp/
├── HeaderSyncWorker.kt              # WorkManager job (CONNECTED + optional UNMETERED)
├── HeaderSyncForegroundService.kt   # FGS for first-run / large catch-up
├── LocalHeadersOtsResolverBuilder.kt # plugs LocalHeadersBitcoinExplorer in
└── (modify) TorAwareOkHttpOtsResolverBuilder.kt → CompositeOtsResolverBuilder

AppModules.kt switches from TorAwareOkHttpOtsResolverBuilder to a composite that prefers the local explorer and falls back to HTTP only when local is missing the height. The HTTP path resolves its base URL via the existing BitcoinExplorerEndpoint (shared with the onchain-zap EsploraBackend), so a user-configured custom Esplora is honoured on both the OTS and the zap side.

AppModules.kt:433 already passes rootFilesDir = { appContext.filesDir } into Quartz; the headers SQLite file goes under $filesDir/bitcoin/headers.db via a new entry in that wiring.

On desktop and CLI, the equivalent data-dir paths are already plumbed for SQLiteEventStore — same hook reused here.

5. Bitcoin P2P protocol — minimum we implement

We implement a strict subset. Anything not in this list is unimplemented and silently dropped on receive.

Outbound messages we send

Command	When
version	Right after TCP connect
verack	After receiving peer's version
ping	Every 60 s if idle, to keep connection alive
pong	In response to peer's ping
getheaders	The actual workhorse: locator + zero stop hash
sendheaders	(optional) tell peer we want unsolicited new tips via headers

Inbound messages we handle

Command	Action
version	Validate services flag, store user-agent + start_height; reply verack
verack	Mark handshake complete
ping	Reply pong with same nonce
pong	Update RTT estimate
headers	Hand off to HeadersSyncEngine
addr/addrv2	(later) feed into peer pool for fallback discovery
anything else	ignore, do not disconnect

version payload we send

• version: 70016 (supports wtxidrelay, but we don't use it)
• services: 0 (we serve nothing — pure leech)
• timestamp: current unix time
• addr_recv/addr_from: zeroed
• nonce: random 64-bit
• user_agent: /Amethyst-Quartz:<quartz-version>/
• start_height: 0
• relay: false (we don't want unsolicited tx invs)

getheaders request rules

• Block locator built with exponential back-off: [tip, tip-1, tip-2, tip-4, tip-8, ..., tip-2^k, ..., genesisHash] ~33 entries even at height 1 M.
• hash_stop = 0x00..00 → "give me as many as you have, up to 2000".
• Pipeline tip-extension: after a headers of length 2000 arrives, immediately fire the next getheaders with the new tip. (This is the standard optimization Core itself uses.)

Magic bytes & ports

• Mainnet magic: 0xf9 0xbe 0xb4 0xd9
• Default port: 8333 (allow override for users running their own node on a non-standard port)

6. Validation rules — what makes a header chain "valid"

Per-header (HeaderValidator):

1. Hash ≤ target. Compute dsha256(header80) (interpreted as little-endian 256-bit int) and compare to compactToTarget(bits).
2. Linkage. header.prevHash == storedTip.hash.
3. MTP sanity. header.time > medianOf(last 11 headers' time). (time is allowed to be up to 2 h ahead of our clock, but we don't enforce that on sync — we'd reject reorgs where we're skewed; just log.)
4. Difficulty retarget at multiples of 2016. Reproduce Core's CalculateNextWorkRequired:
   • actualTimespan = clamp(prev2016.time - first2016.time, target/4, target*4)
   • newTarget = oldTarget * actualTimespan / (14 * 24 * 3600)
   • Cap at pow_limit (max target).
5. Version: not validated — soft-fork bits don't matter for headers-only.
6. Cumulative chainwork updated as prevWork + (2^256 / (target+1)).

Chain selection

• We track the highest-cumulative-chainwork chain we've seen, not the longest. Standard Bitcoin rule.
• On receiving a fork: keep both branches in memory while their tips are within MAX_REORG_DEPTH = 100 of each other; once one accumulates clearly more work, drop the loser.
• A reorg deeper than 100 blocks is treated as adversarial and refused (operator must bump the checkpoint).

Hardcoded sanity checkpoints

In Checkpoints.kt we ship a small list of (height, expectedBlockHash) pairs mirroring Bitcoin Core's own historical checkpoints, plus one release-pinned checkpoint updated at app build time near the current tip. Sync must agree with all checkpoints — disagreement aborts and bans the peer.

7. Bundled data — none. Just a pinned checkpoint constant.

The point of this work is "verify against the Bitcoin network, not a trusted party," so we ship no headers in the APK. Instead:

• One hardcoded PinnedCheckpoint constant in bitcoin/store/PinnedCheckpoint.kt: (height, blockHash, chainWork, time). ~80 bytes of source.
• The checkpoint is pinned ~1 month before each release and bumped via a one-line PR per release. This is the only thing trusted at build time — and it's a single block hash any user can independently verify against any Bitcoin node before installing.
• First-run sync starts at the checkpoint and pulls forward to the current tip over P2P. At a ~1-month checkpoint that's ~4,300 headers (~340 KB on the wire, ~15 minutes' worth of getheaders round-trips at worst), seconds over WiFi.
• Pre-checkpoint heights are not synced locally at all. OTS proofs whose height precedes the checkpoint fall through to OkHttpBitcoinExplorer via the composite. Users in strict-mode (localHeadersTrustOnly = true) get a clear "this OTS proof predates the bundled checkpoint" error and no network call.
• Catch-up after long absence. A device that hasn't run the app for months will, on next sync, fast-forward from tip_local (its stored best_height) to tip_network. Each headers message brings 2,000 headers; even 6 months of absence is ~26,000 headers (~2 MB) in 13 round trips.
• Header growth on disk is ~4 MB raw per year (+ SQLite overhead, see §9).

APK growth: 0 bytes. Trust delta vs the old HTTP-explorer model: one block hash, code-reviewable in git.

8. Multi-peer strategy & eclipse mitigation

A single peer can lie. We protect against that with:

1. Peer count. Maintain ≥4 simultaneous peer connections during sync; accept a header batch only after at least 2 distinct peers (different /16 IPv4 prefixes, different ASNs where we can tell) have served the same batch.
2. Chainwork comparison. After convergence we should land on the chain with highest cumulative work. If peers disagree, prefer the higher-work chain.
3. Checkpoint cross-check. Bundled checkpoints (above) detect any peer that disagrees with the consensus history. Disagreement → ban that IP for the session.
4. DNS seed diversity. Rotate across all 9 of Bitcoin Core's DNS seeds, not a single one.
5. Hardcoded fallback IPs. Ship the chainparams.cpp fallback IP list (about 1,000 v4 + v6 entries) so first-run works even if every DNS seed is blocked or poisoned.
6. Optional user override. A user can configure their own trusted node (bitcoin.example.com:8333) — same path as the existing custom explorer URL setting in OtsSettings.kt.

9. Storage — SQLite via androidx.sqlite + BundledSQLiteDriver

We reuse the same storage primitive Quartz already uses for SQLiteEventStore (quartz/.../nip01Core/store/sqlite/): androidx.sqlite with the bundled driver. This is KMP, ships in androidx-sqlite-bundled / …-bundled-jvm (already a Quartz dependency), and is the project's standard storage choice.

The complete SQLiteEventStore infrastructure — BundledSQLiteDriver, SQLiteConnectionPool, transaction { }, IModule for migrations, QueryBuilder helpers — is reused by SqliteHeaderStore and HeaderStoreSchema.

Schema

PRAGMA journal_mode = WAL;
PRAGMA synchronous  = NORMAL;
PRAGMA foreign_keys = OFF;

CREATE TABLE block_header (
    height       INTEGER PRIMARY KEY,        -- 0..tip, dense
    hash         BLOB NOT NULL,              -- 32 B, dsha256(header), big-endian display order
    header       BLOB NOT NULL,              -- 80 B raw header
    chainwork    BLOB NOT NULL               -- 32 B cumulative work, big-endian
);

CREATE UNIQUE INDEX idx_block_header_hash ON block_header(hash);

CREATE TABLE chain_tip (
    id           INTEGER PRIMARY KEY CHECK (id = 1),  -- singleton row
    best_height  INTEGER NOT NULL,
    best_hash    BLOB    NOT NULL,
    best_work    BLOB    NOT NULL
);

CREATE TABLE checkpoint_seen (
    height       INTEGER PRIMARY KEY,         -- which checkpoints we've passed and confirmed
    hash         BLOB    NOT NULL,
    source       TEXT    NOT NULL             -- 'pinned' | 'historical'
);

Migrations live in HeaderStoreSchema.kt as an IModule (same pattern as the event store's modules). v1 is the schema above; future additions (prev-hash cache column, merkle-block annotations) ALTER from there.

Sizing

Item	Per record	At height 1,000,000
block_header row (80+32+32 B payload + SQLite overhead)	~180–220 B	~180–220 MB
idx_block_header_hash B-tree (32 B key + rowid)	~40 B	~40 MB
chain_tip	1 row	negligible
Total		~220–260 MB at height 1 M

In practice users will start from a recent checkpoint and grow ~4 MB raw (~10 MB stored) per year, so most installs sit at 10–40 MB. The "1 M height" line is the long-tail worst case if someone with a multi-year-old device's local store retroactively backfills (which we explicitly don't do in v1 — see §15 Q1).

Atomicity & crash safety

• HeadersSyncEngine writes each headers batch (up to 2,000) inside a single transaction { }. WAL gives us atomic-or-nothing per batch.
• chain_tip is updated in the same transaction as the headers it points to, so the tip can never reference rows that didn't commit.
• A crash mid-batch leaves the WAL inconsistent; SQLite rolls it back on next open. No custom recovery code.
• On open, SqliteHeaderStore.verifyTipIntegrity() re-reads the tip row and confirms the referenced header exists. If not (unexpected DB corruption, ransomware-like sandbox damage), we delete the db file and fall back to re-sync from the pinned checkpoint.

Concurrency

Same model as SQLiteEventStore: one connection pool, all writes go through the sync engine on a dedicated dispatcher; reads use the pool's read slots. OTS verifier reads are short and indexed — no contention with sync writes.

10. Resolving a height query (the actual BitcoinExplorer impl)

class LocalHeadersBitcoinExplorer(
    private val store: HeaderStore,
    private val sync: HeadersSyncEngine,
    private val fallback: BitcoinExplorer? = null,  // OkHttpBitcoinExplorer
) : BitcoinExplorer {
    override suspend fun blockHash(height: Int): String {
        store.getByHeight(height)?.let { return it.blockHash.toHex() }
        // height not yet synced; ask sync engine to extend, with timeout
        sync.ensureHeight(height, timeoutMs = 10_000)
        store.getByHeight(height)?.let { return it.blockHash.toHex() }
        return fallback?.blockHash(height) ?: throw NotSyncedException(height)
    }

    override suspend fun block(hash: String): BlockHeader {
        store.getByHash(hash)?.let { rec ->
            return BlockHeader(
                merkleRoot = rec.header.merkleRoot.toHex(),
                blockHash  = hash,
                time       = rec.header.time.toString(),
            )
        }
        return fallback?.block(hash) ?: throw NotSyncedException(hash)
    }
}

The fallback path is gated by a setting; a privacy-conscious user can disable it ("local-only verification").

11. Mobile (Android) integration

Lifecycle

• First launch after install/upgrade. Show a one-time "Verify Bitcoin timestamps locally?" prompt. If accepted, kick off a foreground-service sync (FGS so OS doesn't kill us mid-download). Default WiFi-only; user can override.
• Subsequent launches. Background WorkManager job, periodic, ~6 h cadence. Constraints: NetworkType.UNMETERED by default, BatteryNotLow.
• On-demand. When a NIP-03 attestation needs a height we don't yet have, request a targeted sync up to that height (capped 10 s, otherwise fall back to HTTP).

Permissions

• No new runtime permissions needed. We use the foreground-service permission declared for nests audio (already present) plus INTERNET (already granted). Add FOREGROUND_SERVICE_DATA_SYNC to manifest if not present.

Settings UI

Extend OtsSettings.kt:

data class OtsSettings(
    val customExplorerUrl: String? = null,        // existing — shared via BitcoinExplorerEndpoint
    // NEW:
    val localHeadersEnabled: Boolean = true,
    val localHeadersTrustOnly: Boolean = false,   // disables HTTP fallback; pre-checkpoint proofs error
    val localHeadersWifiOnly: Boolean = true,
    val customP2pNode: String? = null,            // e.g. "myriad.example.com:8333"
)

A short "Bitcoin verification" settings screen lets the user pick:

• Local headers + HTTP fallback (default) — best UX, trustless for post-checkpoint heights, HTTP for older proofs.
• Local headers only (strict) — refuses to verify proofs older than the pinned checkpoint. Most paranoid setting.
• HTTP explorer only — pre-local-headers behaviour. Lightest disk/CPU.

The customExplorerUrl field continues to feed BitcoinExplorerEndpoint, so a custom HTTP Esplora is honoured both by the fallback in this feature and by the onchain-zaps EsploraBackend. No new Esplora setting.

12. Tor integration

The codebase already has a Tor-aware HTTP path (TorAwareOkHttpOtsResolverBuilder). For P2P:

• Reuse the project's existing SOCKS proxy plumbing.
• When Tor is on, route TCP socket connections through SOCKS5 to Tor's 127.0.0.1:9050.
• Prefer .onion peer addresses (advertised via addrv2) when available — they don't expose the user's clearnet IP to the peer.
• Increase per-peer timeout (Tor RTT is ~1–5 s).

DNS seed lookup over Tor must use SOCKS5 hostname resolution (no plaintext DNS).

13. Phased delivery

Each phase produces a shippable artifact with measurable behavior.

Phase 0 — Foundation (1–2 days)

• Create quartz/plans/2026-05-08-local-headers-explorer.md (this file). ✓
• Add quartz/src/commonMain/.../bitcoin/ skeleton.
• Wire empty LocalHeadersBitcoinExplorer that always defers to fallback.
• No behavior change yet.

Phase 1 — Header parsing & validation, no I/O (3–5 days)

• BlockHeader80 codec + tests against known mainnet headers (genesis, a few retargets, post-segwit).
• DifficultyTarget compact↔target with RFC test vectors.
• PowValidator, RetargetCalculator, MedianTimePast, HeaderValidator.
• Tests run on a local fixture file containing the first 50 000 headers.

Phase 2 — Storage layer (2–3 days)

• HeaderStore interface + SqliteHeaderStore implementation in commonMain (androidx.sqlite is KMP, no platform shim needed).
• HeaderStoreSchema as an IModule matching the SQLiteEventStore migration pattern.
• append(batch), getByHeight(h), getByHash(hash), tip().
• Reuse BundledSQLiteDriver, SQLiteConnectionPool, transaction { } from nip01Core/store/sqlite/ — no duplication.
• Crash-safety test: kill mid-transaction, reopen, confirm WAL rollback; confirm verifyTipIntegrity() recovers cleanly.

Phase 3 — P2P codec, no socket (3–5 days)

• Message envelope (magic, command, length, checksum) encode/decode.
• All seven message payload codecs.
• Tests using captured wire bytes from a real Bitcoin Core peer (record once, replay forever).

Phase 4 — Single-peer sync (5–7 days)

• BitcoinPeer over a real socket.
• HeadersSyncEngine drives getheaders loop against one peer.
• Validate every header through Phase 1 validator before persisting.
• Integration test: sync first 100 K headers from a public Bitcoin Core peer in CI. Skip in normal ./gradlew test; only run nightly.

Phase 5 — Multi-peer & eclipse hardening (3–5 days)

• PeerPool, PeerScoring.
• DNS seed discovery + fallback IP list.
• Header-batch cross-validation across peers.
• Hardcoded checkpoints in Checkpoints.kt.

Phase 6 — Pinned checkpoint constant (½ day)

• bitcoin/store/PinnedCheckpoint.kt — single hardcoded Checkpoint(height, blockHash, chainWork, time). ~80 bytes of source.
• A small Gradle task (./gradlew bumpBitcoinCheckpoint) that takes (height, hash) arguments and writes the constant. Run by the release process; otherwise no automation.
• No backward-sync, no bundled headers blob, no asset resources. Pre-checkpoint heights are routed to the HTTP fallback by the composite resolver (or refused in strict mode).

Phase 7 — Android wiring (3–4 days)

• HeaderSyncWorker + HeaderSyncForegroundService.
• New "Bitcoin verification" settings screen.
• Switch AppModules.kt to composite resolver builder.
• Manual QA: install fresh, open NIP-03 note, watch sync run, confirm verification matches HTTP explorer.

Phase 8 — Tor (2–3 days)

• SOCKS5-tunnelled TCP socket variant.
• addrv2 decoding for .onion addresses.
• QA on Orbot.

Phase 9 — Desktop & CLI integration (1–2 days)

• desktopApp/ inherits the explorer for free (everything is in quartz/commonMain/). Confirm headers.db path is plumbed via the same data-dir hook already used by SQLiteEventStore on JVM. Confirm sync runs end-to-end on the desktop app.
• cli/ — add an amy verify-ots <event-json> command if not present, wired through the same composite resolver. The CLI is a thin assembly layer over Quartz (per amy-expert), so the actual sync logic is reused as-is.

Phase 10 — Rollout & telemetry (ongoing)

• Ship as opt-in for one release (default off, advertised in release notes).
• Collect (locally, no analytics): time-to-first-sync, time-to-verify, fallback rate. Surface in a debug screen.
• Default-on in the next release if metrics look healthy.

Total: ~22–32 engineer-days end to end. Phases 1–4 are the bulk (protocol + validation); phases 5–10 are integration. Phase 6 dropped from 2–3 d to ½ d because we ship no bundled data — just one constant.

14. Testing strategy

Unit tests (commonMain)

• Header codec round-trip (10 known headers across history).
• dsha256 against test vectors.
• Compact-bits encoding edge cases (0x1d00ffff, 0x1b0404cb, etc.).
• Retarget at heights 2016, 4032, 32256 (well-known).
• MTP sanity for a fixture chain.
• Locator-builder shape (geometric back-off, includes genesis).
• Reorg handler: synthetic 5-block fork with higher work wins.

Integration tests (jvmAndroid)

• Replay-based P2P codec test using a captured pcap of a real session (committed as quartz/src/jvmAndroid/test/resources/btc-handshake.bin).
• Local-network test against a Bitcoin Core regtest node spun up in CI (Docker). Validates full handshake + getheaders loop end to end.
• SqliteHeaderStore round-trip on a real on-disk file (jvmTest, in a tmpdir): append 5,000 headers in batches, getByHeight, getByHash, reopen, confirm tip survives.
• Crash-and-resume test: kill the writer in the middle of a transaction { }, reopen, confirm SQLite WAL rolls the partial batch back and sync resumes from the last committed tip.

Property tests

• HeaderValidator — fuzz with random 80-byte inputs; confirm invalid PoW always rejected and valid PoW always accepted.

Manual QA checklist

• Verify a known NIP-03 attestation against the local explorer matches Blockstream's answer.
• Disable network mid-sync; resume cleanly when network returns.
• Tor on — sync completes (slowly) over Orbot.
• Disk-full simulation — fail gracefully, don't corrupt store.

15. Risks & open questions

#	Risk	Mitigation
R1	First-run UX cost (sync from pinned checkpoint)	Pinned checkpoint ~1 month before release means ~4 K headers, ~340 KB, seconds over WiFi. Show progress in settings; sync runs in background WorkManager job anyway.
R2	Mobile OS kills socket mid-sync	Foreground service for first run; resumable design (SQLite WAL handles partial batches automatically).
R3	Peer feeds adversarial headers (eclipse)	Multi-peer + checkpoints + chainwork comparison.
R4	Disk corruption	SQLite WAL + verifyTipIntegrity() on open; on detected corruption, delete db file and re-sync from pinned checkpoint.
R5	Tor + P2P is slow	Increased timeouts; fallback to HTTP-over-Tor explorer (via BitcoinExplorerEndpoint) when sync stalls.
R6	APK size growth	Zero — no bundled data ships in the APK.
R7	Privacy: peers see user's IP	Tor mode supported; version user-agent doesn't leak Amethyst install ID.
R8	Battery cost on background sync	Default to UNMETERED + BATTERY_OK constraints; back off when idle.
R9	Maintenance burden — new soft-forks	Soft-forks don't affect headers-only validation (no script eval). Difficulty algorithm hasn't changed; a future change would require a code update.
R10	Test flakiness if CI talks to public Bitcoin nodes	Use Bitcoin Core regtest in Docker for integration tests; public nodes only nightly.
R11	SQLite store size on long-running installs	~10 MB/year typical; ~220 MB at height 1 M which won't happen in any realistic mobile lifetime. Add a "wipe local headers" button in settings for users who want it.

Open questions

Resolved 2026-05-19

• Q1. Resolved — HTTP fallback only. Pre-checkpoint OTS proofs are routed to OkHttpBitcoinExplorer via the composite resolver. Strict-mode users (localHeadersTrustOnly = true) get a clear error. No backward sync code path.
• Q2. Resolved — cache chainwork. SQLite makes the choice cheap (one extra BLOB column). Reading current tip work without a full scan matters for chain-selection logic; the ~30 MB-at-1M cost is negligible on disk.
• Q4. Resolved — no HTTPS headers blob. The point of this feature is trustless verification, and the pinned-checkpoint approach already gives fast first-run. Adding a second ingestion path (signed blob over HTTPS) doubles the test surface for marginal benefit.

Still open

• Q3. Should LocalHeadersBitcoinExplorer (and friends) be a Quartz public API for 3rd-party Quartz consumers, or internal to Amethyst? Recommendation: public — it's a clean reusable primitive. Decide before Phase 0 lands so the package layout doesn't churn. (See quartz-integration skill for the public-API contract Quartz exposes today.)

16. Migration & rollout

1. Ship behind a feature flag, default off, documented in release notes.
2. Internal QA for one full release cycle on the flag.
3. Default on for new installs the release after, leaving existing installs on their previous setting.
4. Default on for everyone in the release after that.
5. Keep HTTP explorer code path for at least 4 releases as a safety net.

17. Success criteria

• A NIP-03 attestation that verifies via Blockstream also verifies via the local explorer (bit-for-bit identical timestamp).
• 95th-percentile cold-launch verification of a recent attestation: ≤2 s.
• 95th-percentile first-run sync (from pinned checkpoint to tip): ≤30 s on WiFi for a checkpoint ≤2 months old.
• No new crash-rate regression in the verification path.
• App size growth: 0 MB (no bundled data).
• A user with HTTP explorer disabled can still verify any post-checkpoint attestation.

18. Out of scope (explicit)

• BIP-157/158 compact filters
• BIP-37 merkleblock or full-block fetch
• Trustless verification of NIP-BC onchain zaps — needs merkleblock or full-block fetch on top of this stack. See §20 for the follow-up plan.
• Trustless wallet operations for NIP-BC onchain zap sending — needs UTXO-by-address and broadcast paths that headers can't provide. Stays on user-configured Esplora.
• Lightning, BIP-32 wallet, mempool monitoring
• Verifying NIP-03 attestations against altcoin/Liquid timestamps (not in the spec)
• Replacing the calendar-server upgrade path (OtsState.upgrade()) — unrelated, those are simple HTTPS calls and stay as-is

19. Follow-up: trustless NIP-BC onchain zap verification

Not in this plan. Tracked here so the design choices above stay aligned.

The shipped OnchainZapVerifier (quartz/.../nipBCOnchainZaps/verify/) trusts OnchainBackend.getTx(txid) — i.e. Esplora — for both tx-existence-in-a-block and the output values used to compute verifiedSats. A malicious Esplora can fabricate or suppress zaps.

Closing that gap is a separate plan that builds on this one:

• Foundation reused: BitcoinPeer, PeerPool, HeaderStore, HeaderValidator, LocalHeadersBitcoinExplorer.
• New work (rough ~10–15 days on top of S1): BIP-37 filterload / merkleblock / matched-tx handling, or getdata MSG_BLOCK + stream parse; partial-merkle-tree verification; new OnchainBackend variant that asks the local stack for the inclusion proof and the matched tx while keeping Esplora for UTXO listing / broadcast / fee estimates on the send side.
• Net win: zap-receive verification stops trusting Esplora. Zap-send still needs an address-indexed server (no headers-only equivalent exists), so that part of the trust story is unchanged.
• Filename: quartz/plans/YYYY-MM-DD-onchain-zap-merkle-proofs.md, written once S1 has shipped and we have real numbers on first-run sync, store size, and battery cost.

20. References

• NIP-03: https://github.com/nostr-protocol/nips/blob/master/03.md
• OpenTimestamps spec: https://github.com/opentimestamps/python-opentimestamps
• Bitcoin Core P2P protocol: src/protocol.h, src/net_processing.cpp, src/headerssync.cpp
• BIP-130 (sendheaders): https://github.com/bitcoin/bips/blob/master/bip-0130.mediawiki
• Headers-first sync rationale: Bitcoin Core 0.10 release notes (2015)
• Existing OTS surface in this repo:
  • quartz/src/commonMain/kotlin/com/vitorpamplona/quartz/nip03Timestamp/ots/BitcoinExplorer.kt
  • quartz/src/commonMain/kotlin/com/vitorpamplona/quartz/nip03Timestamp/ots/BlockHeader.kt
  • quartz/src/commonMain/kotlin/com/vitorpamplona/quartz/nip03Timestamp/ots/attestation/BitcoinBlockHeaderAttestation.kt
  • quartz/src/commonMain/kotlin/com/vitorpamplona/quartz/nip03Timestamp/ots/OtsBlockHeightCache.kt
  • quartz/src/jvmAndroid/kotlin/com/vitorpamplona/quartz/nip03Timestamp/okhttp/OkHttpBitcoinExplorer.kt
  • amethyst/src/main/java/com/vitorpamplona/amethyst/model/nip03Timestamp/TorAwareOkHttpOtsResolverBuilder.kt
  • amethyst/src/main/java/com/vitorpamplona/amethyst/model/nip03Timestamp/OtsSettings.kt
  • amethyst/src/main/java/com/vitorpamplona/amethyst/model/nip03Timestamp/BitcoinExplorerEndpoint.kt
• Storage precedent (reused as-is):
  • quartz/src/commonMain/kotlin/com/vitorpamplona/quartz/nip01Core/store/sqlite/SQLiteEventStore.kt
  • quartz/src/commonMain/kotlin/com/vitorpamplona/quartz/nip01Core/store/sqlite/IModule.kt
  • quartz/src/commonMain/kotlin/com/vitorpamplona/quartz/nip01Core/store/sqlite/SQLiteConnectionExt.kt
  • quartz/src/commonMain/kotlin/com/vitorpamplona/quartz/nip01Core/store/sqlite/SQLiteConnectionPool.kt
• Follow-up area:
  • quartz/src/commonMain/kotlin/com/vitorpamplona/quartz/nipBCOnchainZaps/verify/OnchainZapVerifier.kt
  • quartz/src/commonMain/kotlin/com/vitorpamplona/quartz/nipBCOnchainZaps/chain/OnchainBackend.kt