Files
amethyst/cli/DEVELOPMENT.md
T
Claude 4efb2cca98 fix(quartz,cli): audit fixes — bounded drains, no event loss, honest verdicts, false-reject traps
Adversarial audit of the PR's own changes (8 finder angles, verified
before fixing). Quartz core:

- fetchAll-family drains get a wall-clock ceiling (maxTotalMs, default
  10x the idle window, delay()-watchdog: cancellable and virtual-time
  testable). The pure idle window was unbounded when a relay trickled
  events forever — sandboxed napplet queries, set -e fetches, and
  marmot await stuck inside one drain. Streaming relays still finish.
- The suspending onEvent hook no longer runs inside a cancellable
  timeout scope (an expiring window could cancel verifyAndStore
  mid-write and silently drop a received event); the timeout is armed
  only when the channels are dry (no per-message timeout-job churn).
- fetchAll is a projection over fetchAllWithHooks: fixes its
  unsynchronized events/seenIds mutation from concurrent socket
  threads and deletes the duplicate loop + per-event activity channel.
- publishAndConfirmDetailed regains its only-responders contract
  (synthetic no-response entries no longer render as 'relay rejected
  your message' in app callers); results built by pure associateWith;
  shared failure-reason constants + PublishResult.isTransportFailure.
- NIP-65 mutations: split read+write r-tags for the same URL now merge
  to BOTH instead of last-wins dropping a facet (+ test).
- TcpProber's 128-thread pool drains after 60s idle.

CLI:

- publishGuard: all-transport failure exits 124 as timeout; rejected/1
  is reserved for an actual OK-false answer.
- --help anywhere in argv is hoisted centrally; 'amy notes post "x"
  --help' prints usage instead of publishing.
- rejectUnknown false-reject traps fixed: geochat --no-fetch behind an
  early return, and 13 elvis-alias short-circuit sites read eagerly.
- Aliases load once per Context and only match name-shaped inputs (no
  shadowing a real npub/NIP-05/hex); stderr color requires a
  positively-known terminal (TERM sniff polluted captured logs).
- Relay-CSV strictness unified on RawEventSupport.relayFlag (post,
  graperank publish/followers/register no longer silently drop
  malformed URLs); Args.timeoutMs(+OrNull) replaces 27 hand-rolled
  conversions, all strict; offer/debit --timeout > 3600 rejected with
  a 'looks like milliseconds' hint; NPub.create idiom; stale jq .id in
  the marmot reactions harness; printUsage drift (offer pay --with,
  profile --clink-offer, search --kind).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CP4kfLCa3wWtE8Khy21Pkj
2026-07-19 00:17:43 +00:00

27 KiB
Raw Blame History

Developing Amy

How to touch the cli/ module without breaking its public contract.

  • What Amy is + how to use it: README.md.
  • What to build next and in what order: ROADMAP.md.
  • Plans for cross-cutting work: see this module's plans/ folder and commons/plans/ for shared-code work consumed by Amy.

The rule this doc defends: cli/ is a thin assembly layer. If you're writing Nostr protocol logic, filter building, state machines, or encryption in here, stop — that code belongs in quartz/ or commons/.


Public contract

What every caller — user, script, agent, CI — can rely on:

  • Default stdout is human-readable text. A YAML-ish render of the underlying result map. Friendly at a terminal; no shape promises.
  • --json is the machine contract. One line. One object. Stable snake_case keys. Pipe it into jq, parse it from Python, hand it to an agent. Pass --json anywhere before the subcommand.
  • stderr is for humans. Progress, warnings, per-relay ACK traces. Safe to discard. Errors land here too: error: <code>: <detail> by default, or JSON {"error":"…","detail":"…"} under --json.
  • Exit codes are the real signal. The exit code is derived from the error-code stringOutput.error(code, …) returns the exit code, so return Output.error(…) always honours the contract with no per-site bookkeeping:
    • 0 — success
    • 2 — the code is bad_args (every bad argument, including unknown flags and malformed numeric/relay/--author/--id values)
    • 124 — the code is timeout (every timeout: await verbs, pow mine, offer/debit round-trips)
    • 1 — every other code (runtime errors, rejected, not_member, …)
  • No interactive prompts — passwords, names, keys are all flags — with exactly two deliberate, opt-in exceptions: the ncryptsec secret-backend passphrase falls back to a TTY prompt when neither --passphrase-file nor $AMY_PASSPHRASE is set, and bunker --interactive prompts y/N per signing request (it errors with no_tty without a terminal). Neither can trigger in a correctly-configured script.
  • ~/.amy/ is the whole world. Per-account dirs hold identity, cursors, MLS state, and aliases at ~/.amy/<account>/; every observed Nostr event lands in the shared store under ~/.amy/shared/ (a SQLite events.db by default; the events-store/ tree when AMY_STORE=fs). Delete to reset; copy to move. Shell tests isolate by overriding $HOME for the amy subprocess (HOME=/tmp/run.123 amy --account alice …) — same convention git, gpg, and npm use; the in-process JVM tests use the amy.home system-property seam instead.
  • An account is only required to sign. Read-only verbs (relay queries, the shared store, offer/debit info, and the stateless primitives) run against an empty ~/.amy/DataDir.resolveOptional hands them an accountless dir (its hasAccount = false) pointing only at the shared event store, and Context.openOrAnonymous gives them an ephemeral key-less identity (they read fine, they just can't authenticate). Signing verbs go through Context.open, which re-asserts the account requirement — init/create/login/logoff/whoami resolve strictly, since they operate on the account dir itself.

Only the --json shape and the exit codes are public API. The default text format is allowed to change between releases. The five design principles below are how we keep that promise.


Design principles

  1. Non-interactive. One verb = one result on stdout = one exit code. No REPL, no daemon, no prompts. Any network wait is an explicit await verb with a --timeout.
  2. Thin command layer. Each file in commands/ parses args, calls into commons/ or quartz/, and emits a result map via Output.emit(...). A file longer than ~200 lines is a code smell — the logic is living in the wrong module.
  3. Everything persistent is on-disk. No in-memory caches that survive between invocations. Every run reloads cursors, MLS state, identity, and relay config. This is what makes Amy safe to run from CI and from 100 parallel interop scenarios.
  4. Shared defaults. When Amethyst picks a default relay, kind, or tag — Amy calls the same helper. No hand-rolled duplicates. If the helper doesn't exist yet, extract it to commons/ first.
  5. The --json output shape is the public API. Default stdout is human-readable text (a YAML-ish render of the result map by way of Output.kt's default formatter) — that text shape can change freely without warning. The --json shape cannot: changes to keys, types, or nesting are breaking changes. Version them explicitly in commit messages; update interop fixtures.

Architecture

cli/src/main/kotlin/com/vitorpamplona/amethyst/cli/
├── Main.kt                    # argv → subcommand dispatch (runCli is the
│                              #   testable seam; main() adds exitProcess)
├── Args.kt                    # tiny flag parser: --key value / --key=value,
│                              #   `--` ends flag parsing, rejectUnknown()
├── Output.kt                  # text/json emitter; error(code) → exit code
├── Aliases.kt                 # per-account aliases.json read/write
├── Config.kt                  # Identity, RunState, DataDir (~/.amy layout,
│                              #   `amy.home` system-property override)
├── Context.kt                 # per-run wiring: signer + NostrClient +
│                              #   MarmotManager + publish/drain helpers;
│                              #   syncIncoming delegates to the shared
│                              #   commons/marmot/MarmotSyncPolicy
├── StoreFactory.kt            # AMY_STORE backend pick: sqlite | fs
├── StoreStats.kt              # shared store-size reporting (status, store stat)
├── OperatorKeys.kt            # ~/.amy/operator/ GrapeRank operator keys
├── RelayDiagnostics.kt        # per-relay OK/REJECT stderr traces
├── SecureFileIO.kt            # 0600/0700 atomic writes, perm tighten
├── stores/                    # File-backed MLS/KP/message stores, ConcordStore,
│                              #   cashu keyset counters
├── secrets/                   # SecretStore backends (keychain / ncryptsec / plaintext)
└── commands/                  # ~70 command files + commands/cashu/ (8) — one
    ├── Router.kt              #   file per verb family, named <Verb>Command(s).kt.
    └── …                      # Router.kt's `route(name, tail, usage, routes,
                               #   help=USAGE)` is the shared sub-verb dispatcher;
                               #   most families expose a `val USAGE` that both
                               #   `--help` and the README are kept in sync with.

Dependencies: :quartz + :commons (+ :geode for amy serve) + kotlinx-coroutines + OkHttp + Jackson. No Android, no Compose. Amy compiles on any JDK 21 host. Never add a Gradle dependency on :amethyst or :desktopApp.

Context.kt is the backbone. Most commands follow this template:

Context.open(dataDir).use { ctx ->   // .use closes the Context on exit
    ctx.prepare()               // restore MLS state + connect relays
    ctx.syncIncoming()          // pull new gift-wraps + group events
    // ...call into commons/ or quartz/ to build an event...
    val ack = ctx.publish(event, targets)
    Output.emit(mapOf(...))
    return 0
}                               // close() flushes RunState + disconnects

Context is AutoCloseable, so wrap it in use { } rather than a hand-rolled try { } finally { ctx.close() }.


How to add a command

Rule of thumb: no new logic in cli/. Every command is an assembly of things that already work elsewhere.

1. Audit (mandatory)

Before writing anything, answer three questions:

  1. Is the Nostr-protocol piece (event kind, tags, encryption) already in quartz/? If not, add it there first.
  2. Is the business logic (state, default values, ordering, filter assembly) already in commons/? If not, extract it from amethyst/ into commons/ in a preceding commit. See the extraction recipe below.
  3. What is the smallest signed event or query this command has to produce? That shape is the JSON your command will echo.

2. Extract from Android

The single most important recurring task for Amy's growth. Most Amethyst features today live in amethyst/src/main/java/…/model/ or …/service/ with Android-only imports (Context, SharedPreferences, WorkManager, Log, Bitmap). Amy cannot call those directly — they have to move.

  1. Identify the class in amethyst/ (e.g. ReactionPost.kt).
  2. List its Android dependencies.
  3. For each dependency, choose:
    • Inline-able (one call, trivial): delete.
    • Platform-abstractable: add expect/actual in commons/commonMain/… + commons/androidMain/… + commons/jvmMain/…. (See kotlin-multiplatform skill.)
    • Inversion-of-control: take it as a constructor arg. Amy supplies a JVM flavour.
  4. Move the file to commons/commonMain/….
  5. Update the Android caller to use the new location. Add a JVM test.
  6. Only now, add the cli/commands/… file that calls it.

What to keep in amethyst/: screens, navigation, Android-specific side-effects (notifications, background services, camera, Intents). Everything else is a candidate to move.

3. Command file template

package com.vitorpamplona.amethyst.cli.commands

import com.vitorpamplona.amethyst.cli.Args
import com.vitorpamplona.amethyst.cli.Context
import com.vitorpamplona.amethyst.cli.DataDir
import com.vitorpamplona.amethyst.cli.Output

object NoteCommands {
    // The per-group usage text. `amy note --help` prints it; keep it in
    // lock-step with printUsage() in Main.kt and the README table.
    val USAGE: String =
        """
        |Notes:
        |  note publish TEXT [--relay URL[,URL…]]   publish a kind:1 note
        """.trimMargin()

    suspend fun dispatch(dataDir: DataDir, tail: Array<String>): Int =
        route("note", tail, "note <publish|read|…>", help = USAGE, routes = mapOf(
            "publish" to { rest -> publish(dataDir, rest) },
        ))

    private suspend fun publish(dataDir: DataDir, rest: Array<String>): Int {
        val args = Args(rest)
        val text = args.positional(0, "text")
        val relays = RawEventSupport.relayFlag(args)   // malformed URL → bad_args
        args.rejectUnknown()                           // --typo → bad_args, exit 2
        Context.open(dataDir).use { ctx ->
            ctx.prepare()
            val event = com.vitorpamplona.amethyst.commons.note.buildTextNote(ctx.signer, text)
            val ack = ctx.publish(event, relays.ifEmpty { ctx.outboxRelays() })
            RawEventSupport.publishGuard(ack, event.id)?.let { return it }  // all-rejected → `rejected`, exit 1
            Output.emit(
                mapOf(
                    "event_id" to event.id,
                    "kind" to event.kind,
                ) + RawEventSupport.ackFields(ack),  // published_to + rejected_by[{relay, reason}]
            )
            return 0
        }
    }
}

The shared route(name, tail, usage, routes, help) helper (in Router.kt) handles the empty-input and unknown-verb bad_args cases (an unknown verb echoes the expected verb list) and answers --help / -h / help with the group's USAGE text — so each dispatch is just the verb→handler map plus a USAGE constant. This is the pattern for every new command group.

Inside a handler, the Args contract does the policing for you:

  • Read every supported flag through an accessor (flag, bool, intFlag, longFlag, requireFlag), then call args.rejectUnknown() — any leftover --typo becomes bad_args (exit 2) instead of a silent no-op.
  • intFlag/longFlag throw bad_args on non-numeric values; use RawEventSupport.relayFlag / Output.invalidRelayUrl for relay URLs so malformed inputs fail the same way everywhere.
  • After a publish, call RawEventSupport.publishGuard(ack, id) — it returns the rejected (exit 1) error when every targeted relay refused the event, and null otherwise.

Add a top-level branch in Main.kt's dispatch and extend both printUsage() and printVerbList(). Keep the command tour in README.md and the parity matrix in ROADMAP.md in sync.

4. Output-shape conventions

The result map you pass to Output.emit(...) IS the --json shape — treat its keys and types as the public API. The default text render is derived from the same map by Output.kt and intentionally has no contract.

  • Top-level object always.
  • Stable snake_case keys.
  • Event IDs as hex strings (not npub-style).
  • Pubkeys as hex ("pubkey":…) and bech32 when the pubkey is the primary subject ("npub":…).
  • Relay URLs as strings, normalized. The one structured exception is rejected_by: a list of {relay, reason} objects, because the relay's NIP-01 OK reason (blocked: …, rate-limited: …, a connect error, or no response within timeout) is the answer to "why didn't it post".
  • Lists of events under a plural key ("messages", "members"); each item converges on event_id / author / created_at / content (+ domain extras). author is reserved for the key that signed an event; pubkey for an identity being described.
  • Byte counts under *_bytes keys — the text renderer only KiB/MiB-renders that suffix.
  • Errors via Output.error("code","detail") — single lower_snake code, free-form detail. The return value is the exit code (bad_args → 2, timeout → 124, else 1), so handlers return Output.error(…).
  • Publish results use RawEventSupport.ackFields(ack)published_to (accepting relay URLs) + rejected_by ([{relay, reason}]) everywhere — and freshly published/minted events report event_id. Don't invent accepted_by-style variants or hand-roll the projection.

Error codes — the canonical set. The code string is part of the --json contract; reuse an existing code before minting a new one:

  • Contract-wide: bad_args (exit 2), timeout (exit 124), rejected (every targeted relay refused a publish AND at least one actually answered OK false; payload carries event_id + rejected_by — when every failure is transport-level the code is timeout/124 instead), runtime (uncaught exception), invalid_event (event/template fails id or signature checks), http_error, bad_response, not_found, exists, read_only, no_identity, bad_account, signer_error, decrypt_failed.
  • Relay routing: no_relays, no_dm_relays, no_inbox_relays, no_servers, servers_unreachable, relay_error, sync_error.
  • Groups: not_member (you aren't in the group / group unknown), target_not_member (the other user isn't).
  • Cashu: no_wallet, no_mint, insufficient_funds, mint_unreachable, mint_http_<status>, mint_proofs_spent, mint_quote_gone, nutzap_locked_to_wrong_key, invoice_failed.
  • CLINK / zaps: bad_pointer, offer_error, debit_error, no_lightning.
  • Blossom / nsite: upload_failed, hash_mismatch, aggregate_mismatch.
  • Misc: no_follows (fof), invalid_identifier (namecoin), no_tty (bunker --interactive without a terminal).

Old spellings are gone — bad_event/bad_template collapsed into invalid_event, fetch_failed into http_error, not_in_group into target_not_member, and pow_timeout into timeout.

Command-family schemas:

  • Cashu (amy cashu …, NIP-60/61): the full per-verb --json key table and the cashu.json NUT-13 counter layout are pinned in plans/2026-05-28-cashu-cli.md. Common keys: wallet_event_id, mint_url, amount_sats (Long), proofs_count, token_event_id, history_event_id, quote_id, p2pk_pubkey. Error codes include no_wallet, no_mint, insufficient_funds, mint_unreachable, mint_http_<status>, mint_proofs_spent, mint_quote_gone.
  • Admin (amy admin …, NIP-86): {relay, method, result} where result is the relay's raw JSON-RPC result (list/boolean/null). Relay-side failures surface as error: relay_error.
  • Serve (amy serve): a single startup object {listening, host, port, path, persistent, admin_pubkeys[]}, then the process blocks (it embeds geode; teardown is on SIGINT).

Testing

Most of what Amy does is already exercised by tests in quartz and commons — the protocol, the builders, the state machines. The thin Amy-specific layer still needs its own coverage:

Layer Test approach
Argument parsing (Args, flag forms, -- terminator, unknown-flag rejection) ArgsTest in cli/src/test/kotlin/ — plain JVM unit tests.
Error / exit-code contract (bad args → 2, timeout → 124, rejected → 1) ExitCodeContractTest — table-driven tests invoking runCli(argv) with captured stdout/stderr.
JSON output shape (keys and types under --json) JsonContractTest — runs commands under --json and asserts on the parsed object. The default text render has no shape contract and isn't asserted on.
File layout on disk (identity.json, shared/events.db, marmot/groups/*.mls, …) Structural assertions after a command sequence.
Round-trip between two accounts on a local relay End-to-end shell harnesses under cli/tests/: each spins up a local nostr-rs-relay and a fresh $HOME=$STATE_DIR so amy sees a virgin ~/.amy/, then bootstraps multiple accounts sharing one store and drives a scenario through them. Nine suites today — see cli/tests/README.md.

The JVM suite drives runCli in-process through the shared amy(vararg argv) harness in CliResult.kt: it captures stdout/stderr, resets the global Output.mode around every run, and isolates ~/.amy via the amy.home system-property seam in DataDir (a temp dir per invocation, deleted afterwards) — no subprocess, no $HOME games, so the contract tests run in milliseconds with plain ./gradlew :cli:test. runCli exists precisely for this: it is main() minus exitProcess.

What not to test here: event signing, filter assembly, MLS correctness, NIP-44 encryption. Those belong in quartz/commons. If an Amy bug can only be caught here, it's a contract violation (wrong key name, wrong exit code), not a protocol bug.

Interop-test script template:

The canonical examples live under cli/tests/ — read cli/tests/README.md for the layout, then crib from cli/tests/dm/tests-dm.sh or cli/tests/marmot/tests-create.sh. At the byte-banging level, a minimal round-trip looks like:

set -euo pipefail
export HOME=$(mktemp -d)   # virgin ~/.amy/ for the duration of this script

amy --account alice create
amy --account bob   create

# ... the scenario under test ...

amy --account bob marmot await message "$GID" --match "hello" --timeout 60

If an Amethyst scenario cannot be scripted through Amy yet, that's a gap — add it to ROADMAP.md.


Local event store — the source of truth

Every Nostr event amy observes is verified (NIP-01 id + signature check) and persisted to a shared store under ~/.amy/shared/ (one store per machine, shared across every account in ~/.amy/). That includes:

  • events received from any relay subscription (amy notes feed, amy dm list, amy marmot key-package publish, group sync, …),
  • events amy generates and publishes itself,
  • inner events unwrapped from NIP-59 gift wraps.

Malformed events are dropped before reaching command code. Persistence is best-effort — if the store fails (full disk, permissions), the relay subscription still works, but the event is not cached.

The store is the authoritative cache of everything amy has seen: profile metadata, relay lists (NIP-65 and NIP-02), gift wraps, group events, follow lists, etc. Commands that need any of these read from the store first and only fall back to a relay fetch on miss. Three convenience helpers exist on Context:

ctx.profileOf(pubKey)            // latest kind:0    (NIP-01)
ctx.relaysOf(pubKey)             // latest kind:10002 (NIP-65)
ctx.contactsOf(pubKey)           // latest kind:3    (NIP-02)
ctx.dmInboxOf(pubKey)            // latest kind:10050 (NIP-17 DM inbox)
ctx.keyPackageRelaysOf(pubKey)   // latest kind:10051 (MIP-00 KP relays)
ctx.cachedRelayListsOf(pubKey)   // RecipientRelayFetcher.Lists from cache

Two backends, one interface. StoreFactory.kt picks the backend from the AMY_STORE environment variable; both implement quartz's IEventStore, so every command works unchanged either way:

  • sqlite (the default) — quartz's SQLite store (quartz/.../store/sqlite/EventStore.kt) in a single database file at ~/.amy/shared/events.db. Index postings live in shared B-tree pages, so an event's kind/author/tag indexes cost a handful of rows — for crawl-scale corpora (hundreds of thousands of follow lists) this is several times smaller on disk than the FS tree.
  • fs (AMY_STORE=fs) — the file tree at ~/.amy/shared/events-store/ (quartz/.../store/fs/FsEventStore.kt): one pretty-printed JSON file per event plus one file per index posting, under shard directories. Intentionally inspectable with ls, cat, jq, grep, find, rsync, and git — but every posting rounds up to a filesystem block, so a large corpus balloons. Deleting an event file is treated as a deliberate "I never saw this"; dangling indexes are skipped at query time and cleaned up with amy store scrub / amy store compact (on sqlite those verbs are a no-op / VACUUM respectively).

Both backends implement the full feature set — NIP-01 replaceable / addressable uniqueness, NIP-09 deletion tombstones, NIP-40 expiration, NIP-50 full-text search (amy store reindex-fts rebuilds the index), NIP-62 right-to-vanish, NIP-91 multi-tag AND. See cli/plans/2026-04-24-file-event-store-*.md for the FS design.


Relay routing

amy follows the Marmot protocol's per-event routing rules so two users with completely disjoint relay configurations can still marmot each other. No event ever ships blindly to "our configured relays" — amy looks up the right relay set per event per recipient.

Event Publish to Fetch from
kind:30443 (our own KeyPackage) key_package bucket → NIP-65 outbox → any configured
kind:30443 (someone else's KeyPackage) Their kind:10051 → their kind:10002 write → our bootstrap pool
kind:10051 / 10050 / 10002 (our own lists) All configured relays (broadcast)
kind:10051 / 10050 / 10002 (someone else's) Our bootstrap pool = configured relays Amethyst defaults
kind:1059 Welcome gift wrap (kind:444 inside) Recipient's kind:10050 → their kind:10002 read → DefaultDMRelayList → our outbox
kind:1059 gift wraps addressed to us Our kind:10050
kind:445 Group Event (Commit / Proposal / chat) Group's MIP-01 relays field Same

Bootstrap pool: when amy needs to discover a user it's never talked to, it queries configured relays Amethyst's default NIP-65 set Amethyst's default DM-inbox set. These defaults come from commons.defaults.AmethystDefaults and match what the Android/Desktop UI publishes to on first run, so any fresh Amethyst account is reachable via the bootstrap pool even before amy has seen any of their events.


Full on-disk layout

~/.amy/                                  ← root, follows $HOME (or the amy.home property)
├── current                              # marker file written by `amy use NAME`
├── operator/                            # machine-level GrapeRank operator keys
│                                        #   (independent of any account; OperatorKeys.kt)
├── shared/
│   ├── events.db                        # SQLite event store — the DEFAULT backend
│   ├── dns-cache.bin                    # NIP-05 DNS lookup cache
│   └── events-store/                    # FsEventStore — only when AMY_STORE=fs
│       ├── events/<aa>/<bb>/…           # canonical kind:0 / 3 / 10002 / 10050 / 10051 / 1 / 5 / 1059 / …
│       ├── replaceable/<k>/…            # one slot per (kind, pubkey) for kind:0/3/10000-19999
│       ├── addressable/…                # one slot per (kind, pubkey, d-tag) for kind:30000-39999
│       ├── idx/                         # hardlink indexes (kind / author / owner / tag / fts / expires_at)
│       └── tombstones/                  # NIP-09 / NIP-62 enforcement
├── alice/                               # one dir per account (`amy --account alice init`)
│   ├── identity.json                    # nsec/npub/hex — the account
│   ├── state.json                       # sync cursors (giftWrapSince, groupSince)
│   ├── aliases.json                     # local name → npub map (init writes a self-entry)
│   ├── cashu.json                       # NIP-60 NUT-13 counters: {"keyset_counters":{"<id>":<long>}}
│   ├── concord.json                     # Concord community secrets (ConcordStore)
│   └── marmot/
│       ├── keypackages.bundle           # MLS KeyPackage bundles (NostrSignerInternal)
│       └── groups/
│           ├── <gid>.mls                # MLS group state per group
│           └── <gid>.log                # decrypted inner events (one JSON per line)
└── bob/ ...                             # additional accounts sit alongside

Per-account files are plain JSON or framed binary — human-inspectable, easy to diff across two accounts. Every account on the machine shares the one store under ~/.amy/shared/, so a public event observed once doesn't get re-stored per account. operator/ sits beside the account dirs and is excluded from account auto-selection (it is not an account).

The local relay configuration (kind:10002 / 10050 / 10051) is not a separate file — it lives in the shared events-store/ as signed events owned by the account that wrote them. amy relay add builds + signs + ingests a new relay-list event; amy relay list reads URLs straight out of the latest event for each kind; amy relay publish-lists broadcasts those events to upstream relays. There is no relays.json.


Housekeeping

  • Run ./gradlew spotlessApply before every commit.
  • Keep four things in sync: printUsage() in Main.kt, the per-group USAGE constants (amy <cmd> --help), the command tour in README.md, and the parity matrix in ROADMAP.md. They drift fast.
  • Never add a Gradle dependency on :amethyst or :desktopApp. If you need something from there, move it to commons/ first.
  • Never introduce a blocking prompt (readLine(), interactive password input). Take it as a flag. The only sanctioned exceptions are the two opt-in TTY paths listed under the public contract (ncryptsec passphrase fallback, bunker --interactive).
  • Keep each command file small. Past ~200 lines, split — the Marmot group verbs are already a cautionary tale.