feat: ingest source-aware messaging bundles

AGENTS.md

@@ -1,8 +1,8 @@
 # Contents
 
-- `src/` – the deterministic local iMessage and Contacts readers, normalized
-  corpus and metrics, private SQLite store, profile parser, Agent Skill
-  installer, and `messagelikeme` CLI.
+- `src/` – the deterministic local iMessage, Contacts, and private source-bundle
+  readers, normalized corpus and metrics, private SQLite store, profile parser,
+  Agent Skill installer, and `messagelikeme` CLI.
 - `schema/` – public versioned JSON Schemas for deterministic artifacts and
   agent-authored profiles.
 - `docs/` – public methodology, evidence limits, research review, and prior-art
@@ -25,14 +25,20 @@
 - Use Bun 1.3.14 and run `bun run check` before handing off a change. Do not add
   another package manager or lockfile.
 - Keep the public description exact: “A local-first CLI and Agent Skill for
-  studying your private iMessage history and drafting messages that sound like
+  studying private messaging history and drafting messages that sound like
   you.”
 - Keep the public repository independently buildable. Do not reference another
   source repository, private packages, sibling paths, private fixtures, or
   publication mechanics.
 - Keep `chat.db` authoritative and ingestion read-only, query-only,
   ownership-checked, schema-validated, and bounded. Never modify Messages,
   contacts, attachments, or SQLite sidecars.
+- Treat a `message-like-me.local-message-bundle` as an untrusted, private,
+  versioned directory boundary. Require its fixed inventory, canonical UTF-8,
+  owner-only modes, bounded records, artifact digests, and manifest digest.
+  Never let bundle absence erase retained history unless a future contract
+  explicitly declares authoritative coverage; apply explicit deletions and
+  tombstones separately.
 - Treat AddressBook databases as optional label-enrichment sources. Isolate
   every database plus WAL or journal before SQLite opens it, validate contact
   entities and property owners dynamically, read only names and exact
@@ -54,7 +60,8 @@
 - Keep the command name `messagelikeme`, the repository and package name
   `message-like-me`, and the Agent Skill name `message-like-me`. Treat
   `messagelikeme.com` as an informational project page, never as a data plane.
-- Keep CLI commands namespaced as `ingest imessage|contacts`,
+- Keep CLI commands namespaced as `ingest imessage|contacts|bundle`,
+  `sources list|show`,
   `contacts list|show|resolve`,
   `inspect tempo|sessions`, `study prepare`, `profile apply|show|export`, plus
   `init`, `context`, `skill`, and `doctor`. Machine-readable commands support

CHANGELOG.md

@@ -1,5 +1,25 @@
 # Changelog
 
+## 0.3.0
+
+- Add strict private version-one source-bundle ingestion for local Beeper
+  exports produced through Wrench, including account and network provenance,
+  replies, edits, deletions, attachments, reactions, and tombstones.
+- Namespace corpus ownership by source so native iMessage and multiple provider
+  accounts coexist. Bounded snapshot absence retains prior history, explicit
+  terminal state suppresses evidence, and later reappearance restores it.
+- Add `sources list` and `sources show` with active message, conversation,
+  reaction, undated-reaction, completeness, and warning health.
+- Require Wrench 0.13.0 or newer for its direct official Beeper CLI path,
+  sequential per-account progress, retained-shard validation, and atomic
+  seven-file publication.
+- Partition sessions, bursts, and response episodes by conversation, preserve
+  truncated text bubbles as tempo evidence, and count undated reactions without
+  inventing timestamps or exposing raw provider reaction values in aggregate
+  output.
+- Upgrade existing version-two stores in place while retaining conversations,
+  profiles, study packets, and evidence provenance.
+
 ## 0.2.0
 
 - Aggregate conservatively matched direct threads into one AddressBook person

README.md

@@ -1,12 +1,14 @@
 # Message Like Me
 
-**A local-first CLI and Agent Skill for studying your private iMessage history
-and drafting messages that sound like you.**
+**A local-first CLI and Agent Skill for studying private messaging history and
+drafting messages that sound like you.**
 
-Message Like Me turns a local Messages database into deterministic conversation
-metrics, bounded study packets, and reusable style profiles. Its Agent Skill
-teaches Codex, Claude, and other coding agents how to interpret those local
-artifacts and draft unsent replies in your voice.
+Message Like Me turns private local messaging history into deterministic
+conversation metrics, bounded study packets, and reusable style profiles. It
+reads native iMessage history and strict local source bundles, including
+multi-account Beeper exports produced through Wrench. Its Agent Skill teaches
+Codex, Claude, and other coding agents how to interpret those local artifacts
+and draft unsent replies in your voice.
 
 The CLI does not call an AI service, authenticate with a product account, send
 messages, or operate Messages. The agent already running the skill supplies the
@@ -23,7 +25,7 @@ Message Like Me requires Bun 1.3.14 or newer. Install the immutable public
 release from GitHub, then install the Agent Skill:
 
 ```sh
-bun add --global github:hraness/message-like-me#v0.2.0
+bun add --global github:hraness/message-like-me#v0.3.0
 messagelikeme skill install
 ```
 
@@ -40,7 +42,7 @@ messagelikeme skill path
 Message Like Me is distributed directly through GitHub and is not published to
 npm.
 
-## Start with your local history
+## Start with private local history
 
 Initialize the private data store and inspect its location:
 
@@ -79,6 +81,58 @@ copy of the database and its transactional sidecars, and opens only that copy
 with SQLite. It does not change Messages, `chat.db`, or its sidecars. macOS may
 require permission for the terminal or agent host to read Messages data.
 
+To study accounts connected through Beeper, install or update to
+[Wrench 0.13.0 or newer](https://github.com/hraness/wrench/releases), then ask
+it to create a new private Message Like Me bundle:
+
+```sh
+wrench beeper export-message-like-me \
+  --auth <beeper-auth-id> \
+  --output /absolute/private/path/beeper-bundle \
+  --json
+```
+
+The optional `--limit-chats`, `--limit-messages`, and `--max-participants`
+flags lower the export bounds. The output path must be a normalized absolute
+path to a directory that does not already exist. Wrench calls the pinned
+[official Beeper CLI](https://github.com/beeper/cli) directly. It enumerates
+the connected account realm, invokes `export --no-attachments` once per
+account in deterministic order, and reports the account ordinal, elapsed-time
+heartbeats, and cumulative validated chat and message counts on stderr. It
+retains each private raw shard until it can atomically publish the complete
+mode-`0700` seven-file bundle with mode-`0600` files.
+
+The export does not use the separate
+[Beeper Desktop API MCP project](https://github.com/beeper/desktop-api-mcp).
+The CLI path supplies the bounded account snapshots and local files needed for
+hash validation, deterministic conversion, crash recovery, and atomic
+publication. Provider URLs and credentials are excluded. Message Like Me does
+not receive the Beeper credential and does not call Beeper or Wrench itself.
+
+Ingest the finished directory, then inspect its redacted source health:
+
+```sh
+messagelikeme ingest bundle --input /absolute/private/path/beeper-bundle --json
+messagelikeme sources list --json
+messagelikeme sources show <source-id> --json
+```
+
+The importer verifies the fixed version-one inventory, canonical UTF-8 NDJSON,
+record and byte bounds, owner-only permissions, artifact digests, and manifest
+digest before changing the store. One bundle may contain several connected
+accounts and networks; each becomes a separate source namespace. Native
+iMessage and prior bundle sources remain alongside it.
+
+The complete interchange, integrity, identity, and reimport laws are in the
+[version-one local message bundle contract](docs/local-message-bundle-v1.md).
+
+Beeper exports describe bounded local observations. A later bounded export
+that omits an older record does not delete retained history. Explicit deletion,
+removal, replacement, and tombstone records suppress their target, and a later
+reappearance restores it. Older snapshots cannot overwrite newer state. Use
+`sources show <source-id> --private --json` only when you deliberately need the
+private provider account and source metadata.
+
 Optionally enrich and join direct conversations with private identities from
 macOS Contacts:
 
@@ -96,16 +150,17 @@ messagelikeme ingest contacts \
   --json
 ```
 
-Contacts ingest may run before or after iMessage ingest. It reads only bounded
-name, email, and phone fields from a stable private copy. Exact normalized
-email or phone handles can join several one-to-one iMessage, SMS, and email
-threads for the same AddressBook person into one analysis scope. Existing
-conversation IDs remain aliases for that person scope. Shared handles remain
-ambiguous, local phone numbers never gain a guessed country code, unmatched
-threads stay separate, and groups are never collapsed to one person. Contact
-labels have their own revision, so a rename does not stale a messaging-style
-profile. `messagelikeme doctor` reports local aggregate state without asking
-for an account or credential.
+Contacts ingest may run before or after any message source. It reads only
+bounded name, email, and phone fields from a stable private copy. Exact
+normalized email or E.164 phone handles can join several one-to-one threads
+for the same AddressBook person into one analysis scope. A bundle conversation
+is eligible only when the producer positively marks its direct participant
+roster complete. Existing conversation IDs remain aliases for that person
+scope. Shared handles remain ambiguous, local phone numbers never gain a
+guessed country code, unmatched threads stay separate, and groups are never
+collapsed to one person. Contact labels have their own revision, so a rename
+does not stale a messaging-style profile. `messagelikeme doctor` reports local
+aggregate state without asking for an account or credential.
 
 ## Inspect behavior without exposing prose
 
@@ -124,8 +179,15 @@ outgoing turns, within-session response latency, single-message versus
 multi-message replies, surface prose features, multi-point response contexts,
 reactions, and explicit reply use. Incoming messages establish what you were
 responding to; they are never counted as examples of your writing style.
-Session and burst gaps are configurable seconds and are recorded with each
-result. They are segmentation choices, not universal facts about conversation.
+Sessions, bursts, and response episodes never cross a source conversation
+boundary. Person scopes spanning several apps expose a sorted `services`
+breakdown instead of hiding the mixed-channel evidence behind a null service.
+Reactions with no provider timestamp still contribute to reaction counts and
+direction, but never to temporal metrics. Raw provider reaction values remain
+private; ordinary metrics and drafting context expose only fixed-size counts,
+direction, datedness, and the outgoing reaction ratio. Session and burst gaps
+are configurable seconds and are recorded with each result. They are
+segmentation choices, not universal facts about conversation.
 
 Pass `--private` to `contacts list` or `contacts show` only when you need to
 resolve a pseudonymous contact to its local private label or participants.
@@ -244,6 +306,9 @@ Run `messagelikeme --help` for the checked grammar. The public surfaces are:
 messagelikeme init [--json]
 messagelikeme ingest imessage [--database PATH] [--json]
 messagelikeme ingest contacts [--addressbook PATH] [--json]
+messagelikeme ingest bundle --input ABS_PATH [--json]
+messagelikeme sources list [--private] [--json]
+messagelikeme sources show SOURCE_ID [--private] [--json]
 messagelikeme contacts list [--min-outgoing N] [--limit N] [--private] [--json]
 messagelikeme contacts show CONTACT_ID [--private] [--json]
 messagelikeme contacts resolve QUERY --private [--limit N] [--json]
@@ -271,10 +336,13 @@ Place global `--data-dir PATH` before the command.
 
 - The original `chat.db` and AddressBook databases remain authoritative.
   SQLite opens only stable private copies, never the source files or sidecars.
+- Source bundles remain private caller-owned inputs. Import verifies their
+  fixed inventory, canonical bytes, digests, bounds, and owner-only modes.
 - The normalized corpus, profiles, and installation key stay in a private local
   store with owner-only permissions.
-- Stable contact, conversation, and message IDs are derived with a private
-  per-install HMAC key. Pseudonymous IDs are not encryption.
+- Stable source, contact, participant, conversation, message, and reaction IDs
+  are derived with a private per-install HMAC key. Pseudonymous IDs are not
+  encryption.
 - Aggregate commands omit bodies and private labels. Study and evaluation
   packets are bounded, explicit body-bearing exports.
 - Message text never goes to a Message Like Me server. There is no service,
@@ -312,9 +380,9 @@ bun install --frozen-lockfile --ignore-scripts
 bun run check
 ```
 
-Tests use synthetic Messages and AddressBook databases plus synthetic
-conversations. Never add a real message, handle, group title, attachment,
-contact record, private path, or derived profile to a fixture.
+Tests use synthetic Messages and AddressBook databases plus synthetic source
+bundles and conversations. Never add a real message, handle, group title,
+attachment, contact record, private path, or derived profile to a fixture.
 
 The canonical repository is
 [`hraness/message-like-me`](https://github.com/hraness/message-like-me).

SECURITY.md

@@ -11,10 +11,12 @@ boundary, observed result, and a reproduction built from synthetic data.
 
 ## Private-data boundary
 
-Message Like Me reads private iMessage history to derive local analysis. The
+Message Like Me reads private messaging history to derive local analysis. The
 following values are sensitive even when they do not contain an obvious name:
 
 - the source Messages and AddressBook databases and their SQLite sidecars;
+- local message bundles, manifests, connected-account metadata, and provider
+  provenance;
 - contact names, email addresses, and phone numbers;
 - message bodies, timestamps, reply links, tapbacks, and attachment metadata;
 - contact, participant, conversation, and group metadata;
@@ -54,6 +56,36 @@ provenance. Missing or unsupported text remains unavailable rather than being
 guessed. Reply targets and tapbacks remain separate from prose so they cannot
 silently become authored style evidence.
 
+## Local message bundle ingestion
+
+`messagelikeme ingest bundle` accepts only a normalized absolute path to a
+current-user-owned physical mode-`0700` directory. The version-one directory
+contains exactly `manifest.json` and six mode-`0600` canonical UTF-8 NDJSON
+artifacts. Files must be regular, singly linked, owner-controlled, stable while
+read, and free of symbolic-link traversal.
+
+The importer validates the manifest before allocating for its artifacts. It
+caps one line at 2 MiB, the complete bundle at 500,000 records and 512 MiB, and
+connected accounts at 128. It streams each artifact, rejects invalid UTF-8,
+requires canonical JSON plus final newlines, and verifies exact record counts,
+bytes, SHA-256 artifact digests, and the canonical manifest projection digest.
+These checks detect malformed or changed local input. They do not establish
+that the provider data is truthful or complete.
+
+The accepted privacy declaration permits attachment metadata only and requires
+provider URLs and credentials to be excluded. The bundle may still contain
+message bodies, names, handles, timestamps, account identifiers, and graph
+coordinates. Keep it under the same controls as the normalized store, and do
+not place it in Git, logs, issues, packages, or ordinary agent context.
+
+Each connected account is stored in its own per-install HMAC namespace.
+Bounded, truncated, and unknown source absence never deletes retained history.
+Explicit tombstones and terminal message or reaction state suppress their
+validated targets. A later matching record can clear suppression, while an
+older or conflicting equal-time snapshot is rejected. `sources list` is
+redacted. `sources show --private` deliberately reveals provider account and
+source metadata.
+
 ## Contacts enrichment
 
 Contacts enrichment is optional. The reader discovers populated
@@ -83,7 +115,8 @@ a prose profile.
 
 ## Local identifiers
 
-Contact, conversation, and message identifiers are derived with an HMAC key
+Source, contact, participant, conversation, message, and reaction identifiers
+are derived with an HMAC key
 created for one local installation. They reduce accidental disclosure and keep
 stable local references without storing handles in ordinary views. They are
 not anonymization against an attacker who can read the local corpus or key.
@@ -95,8 +128,11 @@ it.
 ## Inspection and study packets
 
 Aggregate contact, session, tempo, and surface-style views omit message bodies
-and private labels by default. `--private` deliberately reveals local private
-identity fields. Use it only when the current task needs that mapping.
+and private labels by default. Raw provider reaction values also remain private;
+aggregate and drafting-context views expose only fixed-size reaction counts,
+direction, datedness, and the outgoing reaction ratio. `--private` deliberately
+reveals local private identity fields. Use it only when the current task needs
+that mapping.
 
 `contacts resolve QUERY --private` performs bounded exact matching against
 private labels. It does not do prefix, substring, phonetic, or fuzzy matching,

dist/cli-xby0v0et.js → dist/cli-mxxakdqk.js

@@ -19,11 +19,12 @@ function prettyJson(value) {
 
 // src/types.ts
 var CORPUS_SCHEMA_VERSION = 1;
-var METRICS_SCHEMA_VERSION = 1;
+var METRICS_SCHEMA_VERSION = 2;
 var PROFILE_SCHEMA_VERSION = 2;
 var LEGACY_PROFILE_SCHEMA_VERSION = 1;
 var STUDY_PACKET_SCHEMA_VERSION = 2;
 var EVALUATION_PACKET_SCHEMA_VERSION = 1;
 var CONTACTS_SCHEMA_VERSION = 1;
+var MESSAGE_BUNDLE_SCHEMA_VERSION = 1;
 
-export { canonicalJson, sha256, prettyJson, CORPUS_SCHEMA_VERSION, METRICS_SCHEMA_VERSION, PROFILE_SCHEMA_VERSION, LEGACY_PROFILE_SCHEMA_VERSION, STUDY_PACKET_SCHEMA_VERSION, EVALUATION_PACKET_SCHEMA_VERSION, CONTACTS_SCHEMA_VERSION };
+export { canonicalJson, sha256, prettyJson, CORPUS_SCHEMA_VERSION, METRICS_SCHEMA_VERSION, PROFILE_SCHEMA_VERSION, LEGACY_PROFILE_SCHEMA_VERSION, STUDY_PACKET_SCHEMA_VERSION, EVALUATION_PACKET_SCHEMA_VERSION, CONTACTS_SCHEMA_VERSION, MESSAGE_BUNDLE_SCHEMA_VERSION };