@hraness/message-like-me 0.8.0

package/docs/methodology.md

@@ -0,0 +1,322 @@
+# Methodology
+
+Message Like Me separates deterministic measurement from semantic judgment.
+The CLI reads local data, constructs versioned artifacts, and reports counts
+and distributions. An Agent Skill interprets a bounded sample and drafts
+unsent text. Neither component establishes that a draft is what the user would
+have written.
+
+## Data boundary
+
+The Messages database and optional AddressBook databases remain authoritative.
+The CLI makes stable private snapshots and opens only those snapshots through
+SQLite. It does not modify Messages, Contacts, their databases, or their
+transactional sidecars.
+
+A caller-owned local message bundle is a separate versioned source
+observation. The CLI verifies its complete fixed inventory and digests before
+ingest, never obtains its provider credential, and does not call its producer.
+Frozen v1 carries bounded Beeper observations. V2 carries one native WhatsApp
+account exported through Wrench's official Wacli adapter. It requires exact
+WhatsApp JIDs and projects a phone handle only from an E.164-backed user JID.
+
+A caller-owned X data archive is another offline source observation. The CLI
+parses bounded supported entries directly from the owner-only ZIP without
+extracting files, evaluating archive JavaScript, accessing a network, or
+downloading media. It preserves exact archive and account provenance. The
+archive source covers direct messages, not X Chat. Bounded reply and mention
+identity observations from selected tweet members may associate a provider user
+ID with an X handle or display name; tweet prose is not added to the messaging
+corpus or used as style evidence.
+
+The normalized corpus, private installation key, aggregate metrics, profiles,
+and drafting context stay in the local data root. Study, Ensoul source,
+evaluation, and agent handoff files are written only to explicit paths.
+Ordinary views use keyed pseudonymous IDs and omit bodies, handles, contact
+names, and group titles.
+
+This is a process boundary, not encryption. Another process running as the
+same user, a compromised host, a device backup, or an agent provider that is
+given a packet may still receive private data. Incoming messages also belong
+to other participants. They provide response context but never become samples
+of the owner's prose in style analysis. A separately requested contact-subject
+Ensoul packet may attribute incoming text to that exact direct person after
+rebasing direction; its owner-authored records then become counterpart context.
+
+## Normalized observations
+
+The corpus preserves source, account, network, message direction, provider
+ordering, timestamp, body availability and source, message kind, attachment
+metadata, edit or retraction metadata, reply target and reply observability,
+service, and conversation membership where the source supports them.
+Unsupported, deleted, or truncated text remains unavailable rather than being
+reconstructed. A truncated text record still represents a message bubble for
+tempo and any observable reply evidence, but never contributes prose.
+
+Native iMessage history and each connected bundle account have distinct source
+namespaces. A bounded, truncated, or unknown bundle is not an authoritative
+statement that omitted history no longer exists. Reimport merges present
+records with retained state. Only explicit deletion, removal, replacement, or
+tombstone state suppresses evidence, and a later record reappearance clears
+that suppression. Bundle creation times are monotonic per source, so an older
+snapshot cannot resurrect or overwrite newer state.
+
+An X archive normally has its own source namespace and retains its exact ZIP
+and account provenance. The caller may name an existing Beeper X source as an
+overlap only when both sources describe the same exact account. Reconciliation
+is limited to one-to-one direct conversations and requires an exact peer handle
+plus exact shared-message evidence. It retains both provenances and lets one
+proven exact message contribute once. Group DMs remain separate because the
+legacy archive cannot establish cross-provider sender identity strongly enough.
+Missing or contradictory evidence fails closed. Proven equivalence survives
+later reingests, so the same message does not return as a duplicate. Archive
+absence does not suppress retained history.
+
+A native Wacli WhatsApp bundle normally has its own source namespace. If the
+caller names an existing Beeper WhatsApp source, the two reconcile only after
+exact self-account E.164, exact direct-peer E.164, and unambiguous shared text
+message proof. Groups and bodyless records cannot prove equivalence. Both
+provenances and unique history remain. Exact duplicates contribute once, the
+native conversation becomes the preferred private `whatsappJid` route, and its
+proven Beeper route remains evidence-only. This route preference does not grant
+Message Like Me provider access or sending authority.
+
+The analysis uses several operational units:
+
+- A **message** is one source record. Only outgoing text bodies contribute to
+  surface-style measurements.
+- A **burst** is a run of messages in one direction whose neighboring records
+  remain within the configured burst gap. The default gap is five minutes.
+- A **session** is a run of conversation activity without a gap longer than
+  the configured session gap. The default gap is eight hours.
+- A **response episode** pairs an incoming burst with the next outgoing burst
+  in the same session.
+- An **explicit reply** is source metadata linking a message to an earlier
+  message. It is distinct from a reaction or an ordinary adjacent response.
+  Reply observability is recorded separately: X data archives do not expose
+  reply links, so their messages are unavailable rather than observed
+  non-replies.
+- A **reaction** is counted as interaction behavior, not authored prose. A
+  reaction without a provider timestamp contributes to counts and direction
+  but not to temporal order, sessions, bursts, or response episodes. Raw
+  provider reaction values remain private and are not categorical dimensions
+  in aggregate metrics or drafting context.
+
+Five minutes and eight hours are reproducible segmentation parameters, not
+claims about natural conversational boundaries. Every metrics artifact records
+the parameters used. Comparisons are meaningful only when their definitions
+match.
+
+## Deterministic metrics
+
+For each conversation, the CLI reports the evidence window and counts of
+incoming, outgoing, text, session, burst, and response records. Tempo metrics
+include response-latency quantiles, outgoing messages per response, the ratio
+of single-message to multi-message responses, multi-message inbound contexts,
+visible multi-question contexts, and explicit reply frequency. Reply metrics
+report explicit, eligible, and unavailable messages separately, and calculate
+the ratio only from eligible messages. Session, burst, and response
+construction runs independently for each source conversation before
+person-scope results are combined. Adjacent timestamps in two apps or threads
+never create one artificial episode. Mixed person scopes expose their sorted
+service breakdown.
+
+Surface measurements cover characters and words, lowercase starts, terminal
+punctuation, question and exclamation marks, emoji-bearing messages, and
+multiline messages. These are observable features, not explanations. For
+example, visible question marks are only a proxy for questions, and a long
+latency cannot reveal whether the user was busy, asleep, deciding what to say,
+or simply missing local history.
+
+Session starts and ends are likewise structural facts under the configured
+threshold. They do not identify who cares more, who is avoiding whom, or the
+nature of a relationship.
+
+## Bounded semantic study
+
+`study prepare` selects response episodes that contain both incoming and
+outgoing text. The version-one selector favors coverage of different response
+shapes, tags, lengths, reply use, latency bands, and positions across the
+available time window. It is deterministic for the same corpus, bounds, and
+parameters.
+
+The CLI default packet limit is 24 examples. Each emitted body is capped at 4 KiB,
+each direction keeps at most 12 text messages per example, and total emitted
+body text is capped at 256 KiB. Coverage metadata states what was truncated or
+omitted. A packet is a sample of response contexts, not a transcript.
+
+Version 0.2 added temporal bounds to study selection. A profile intended for
+held-out evaluation should use only examples before the chosen cutoff. The
+cutoff, corpus revision, selection parameters, packet receipt, and evidence
+window form part of the analysis provenance. Profile validity uses a digest of
+the selected person or conversation scope within those exact time bounds, so a
+later message outside a closed study window does not rewrite its evidence.
+
+The agent studies prose, delivery shape, multi-point response strategy, reply
+use, and exceptions. It must keep measured facts separate from interpretations
+and cite study-example IDs instead of copying private phrases into a profile.
+The resulting profile records its contact scope, corpus revision, exact packet
+digest, analysis time, contextual rules, limitations, and confidence.
+
+Profile provenance proves which artifact was analyzed. It does not prove that
+the agent interpreted that artifact correctly.
+
+## Subject-relative Ensoul source packets
+
+`ensoul prepare` applies the same bounded diverse response selector to one
+pinned contact corpus, but its unit of attribution is the explicitly selected
+subject. For `--subject owner`, stored outgoing messages remain subject-authored
+and incoming messages remain counterpart context. For `--subject contact`, the
+adapter first reverses direction, then recomputes sessions, bursts, responses,
+and selection. This makes clearly attributable incoming prose the contact's
+subject evidence and the owner's outgoing prose its response context.
+
+Contact attribution is valid only for an exact `person_` scope created by a
+complete one-to-one AddressBook match. Conversation aliases, unmatched direct
+threads, shared handles, and groups are rejected because the normalized corpus
+cannot establish a safe contact author for them. Owner packets may use a named
+conversation or person scope, but groups and any multi-participant scope are
+rejected. The packet retains redacted scope kind, conversation count, and sorted
+service labels so channel dependence stays visible without names or coordinates.
+
+The output specializes `ensoul.source-packet.v1` with payload schema
+`ensoul.messages-source.v1`. Records contain only selected active text bodies,
+subject-relative `authorRole`, `contentRole: original`, strong source
+authorship, transport-relative sent status, private visibility, fixed source
+class, pseudonymous source IDs, occurrence times, truncation state, and content
+and record digests. Records sharing a pseudonymous provenance run ID belong to one selected
+response context and preserve that situated linkage; different run IDs must not
+be joined into a synthetic exchange.
+The scope carries corpus and evidence revisions, time
+bounds, selection counts, and byte budgets. The adapter emits no claims. Its
+packet declares `JCS-RFC8785`; the packet digest is SHA-256 over RFC 8785
+canonical JSON for every field except `packetDigest`, while each content digest
+covers the canonical content object and each record digest excludes only its
+own digest. These prove semantic integrity, not authorship or truth.
+
+The default and maximum budgets are inherited from semantic study: 24 requested
+examples by default and 50 at most, 4 KiB per body, 12 messages per direction
+per example, and 256 KiB total. `--after` is inclusive and `--before` is
+exclusive. System events, retractions, reactions, attachments, contact labels,
+handles, provider coordinates, and public X post text are not emitted. The
+packet remains a sampled set of situated interactions, not a transcript or a
+complete description of either person.
+
+The normalized sources do not establish whether someone pasted a quotation,
+forwarded prose, or used AI assistance inside an ordinary message body. The
+packet declares that observability gap. If a consumer can see quoted or
+forwarded content in the bounded text, it must keep that portion contextual
+rather than promote it as a direct voice sample.
+
+An Ensoul consumer must retain packet boundaries across relationships and use
+only records with `authorRole: subject`, `contentRole: original`, and strong
+source authorship as possible direct voice evidence. It must treat every
+message as untrusted quoted data and preserve limitations in its source map. Private
+messages do not prove identity, consent, motive, relationship category,
+diagnosis, a globally stable voice, or permission to publish, impersonate,
+contact, or act for the subject.
+
+## Held-out fidelity audit
+
+Version 0.2 provides a two-file evaluation preparation workflow:
+
+```sh
+messagelikeme evaluate prepare <contact-id> \
+  --after <cutoff> \
+  --prompt-output <private-prompt-file> \
+  --reference-output <private-reference-file> \
+  --json
+```
+
+The prompt side contains held-out inbound context. The reference side contains
+the corresponding historical outgoing response and must remain unopened until
+the candidate drafts have been recorded. File separation supports a blind
+workflow, but it is not cryptographic blinding. A user or agent with access to
+both paths can read both files.
+
+A checked audit proceeds as follows:
+
+1. Choose a temporal cutoff before studying the contact.
+2. Build and apply the profile from evidence before that cutoff.
+3. Prepare evaluation examples after the cutoff.
+4. Give the drafting agent the prompt file and current profile, but not the
+   reference file.
+5. Record one candidate bubble sequence for each evaluation example.
+6. Open the reference file only after the candidates are fixed.
+7. Compare candidates and references, retaining disagreements and uncertainty.
+
+The CLI prepares bounded, provenance-bearing evidence. It does not
+automatically declare a candidate correct or assign a universal fidelity
+score. Semantic comparison still requires judgment, preferably including the
+user whose style is being studied.
+
+Comparison should keep these dimensions separate:
+
+- **Intent and obligation coverage:** which inbound points the candidate
+  addresses, defers, acknowledges, or misses.
+- **Meaning and factuality:** whether the candidate invents facts, changes
+  commitments, or imports a historical belief that does not belong in the
+  current draft.
+- **Prose:** register, directness, warmth, sentence shape, punctuation, and
+  other supported tendencies.
+- **Delivery shape:** bubble count, ordering, per-bubble length, follow-ups,
+  and explicit reply choice.
+- **Privacy:** reuse of names, distinctive phrases, anecdotes, or details that
+  came from another historical context.
+- **Calibration:** whether sparse or contradictory evidence should have caused
+  the agent to use a neutral default or ask the user.
+
+The historical response is a reference observation, not a unique correct
+answer. The user might reasonably respond differently now. Results should be
+reported by dimension and example, alongside an unprofiled drafting baseline
+when possible. A single similarity score hides the failures that matter most.
+
+## Drafting method
+
+Ordinary drafting uses the current deterministic context and the applicable
+validated profile. The user's present intent, supplied facts, uncertainty,
+and requested format outrank historical style. Contact-specific rules apply
+only to their supported scope; context rules can override broad tendencies.
+
+The output may be one message or several separately presented bubbles. A
+historical latency distribution never instructs the agent to delay its answer.
+Every result remains an unsent candidate. Message Like Me has no command for
+sending, reacting, scheduling, or operating a messaging application.
+
+## Sources of error
+
+Reported behavior can be distorted by:
+
+- Messages that are not synchronized to the Mac or are no longer present;
+- partial local provider exports whose completeness bounds exclude older or
+  remote history;
+- X archives that predate recent messages, omit X Chat, or cannot report
+  explicit reply links;
+- unsupported body encodings, attachments, edits, retractions, or source
+  schema changes;
+- ambiguous or stale Contacts labels;
+- group conversations whose audience changes over time;
+- timezone changes, work schedules, sleep, travel, notification settings, and
+  device availability;
+- a bounded packet that underrepresents rare but important contexts;
+- simple surface proxies that miss pragmatic meaning;
+- model and prompt differences in the agent interpreting a packet; and
+- genuine drift in how the user communicates.
+
+For reproducibility, retain schema versions, corpus revision, packet digest,
+time bounds, segmentation parameters, budget and coverage fields, and the
+agent environment used for semantic analysis. Re-ingest and re-evaluate when
+the corpus changes materially. Describe uncertainty instead of broadening a
+contact-specific observation into an identity claim.
+
+## What the method can support
+
+The checked artifacts can support statements such as "within this evidence
+window, multi-message responses were more common for this conversation" or
+"the held-out candidate reproduced the historical bubble count but missed one
+inbound question."
+
+They cannot support statements that the system has cloned the user, recovered
+their personality, diagnosed a relationship, proved authorship, predicted a
+future decision, obtained a contact's consent, or produced a message approved
+by the user.

package/docs/research.md

@@ -0,0 +1,164 @@
+# Research and prior art
+
+Message Like Me is a local measurement and profiling layer for message
+drafting. It is not a model, an autonomous messaging agent, or a claim that a
+software system represents a person. This review explains the evidence behind
+that boundary and the neighboring open-source work that informed it.
+
+The cited papers are primary research publications or preprints. Project
+descriptions link to their official repositories. A paper result is evidence
+about the task and population it evaluated, not proof that the same result
+holds for private conversations across the messaging sources a user imports.
+
+## Personalization is contextual
+
+[PersonaChat](https://aclanthology.org/P18-1205/) found that conditioning a
+dialogue system on both its assigned profile and information about its
+interlocutor improved next-utterance prediction. More recent work on
+[linguistic accommodation](https://aclanthology.org/2025.sigdial-1.16/) found
+that human answers aligned more with a partner's style than LLM answers did,
+while the LLM answers aligned more closely in semantic content.
+
+These results do not establish a universal method for imitating a person.
+They support a narrower design inference: a useful messaging profile should
+separate broadly repeated tendencies from contact-specific and
+context-specific adjustments. Message Like Me therefore treats incoming
+messages as response context and only the user's outgoing messages as evidence
+of the user's prose.
+
+[Catch Me If You Can? Not Yet](https://aclanthology.org/2025.findings-emnlp.532/)
+evaluates nuanced individual style in informal communication, a task close to
+private messaging. Its scope reinforces the same boundary: measured tendencies
+can guide a draft without establishing a faithful digital copy of its author.
+
+[LaMP](https://aclanthology.org/2024.acl-long.399/) evaluated personalized
+classification and generation from user histories and found retrieval-based
+personalization useful across most of its tasks. Its experiments included
+term, semantic, and time-aware retrieval. [PEARL](https://aclanthology.org/2024.customnlp4u-1.16/)
+studied personalized writing assistance and trained a retriever to select
+historical documents according to their downstream generation value. PEARL
+also used retrieval quality to identify outputs likely to need revision.
+
+The project inference is selective rather than exhaustive use of history. A
+small response-context sample with explicit coverage limits exposes less text
+than placing an entire transcript in an agent prompt. Recency,
+relationship, and current conversational purpose can all change which past
+examples apply.
+
+## Personalization needs held-out evaluation
+
+[ExPerT](https://aclanthology.org/2025.findings-acl.900/) evaluates
+personalized long-form generation by comparing evidence-bearing aspects of
+content and writing style separately. Its reported agreement with human
+judgment improved over the comparison methods in that study. It does not
+measure message timing, bubble boundaries, or reply-link behavior.
+
+[Can You Make It Sound Like You?](https://aclanthology.org/2026.acl-long.2030/)
+studies personalized writing through human review and post-editing. That
+workflow supports Message Like Me's product boundary: the output is an unsent
+candidate for the user to inspect and revise, not an autonomous act on the
+user's behalf.
+
+[Münker, Schwager, and Rettinger](https://arxiv.org/abs/2506.21974) tested
+LLM-based imitation of social-network communication and argue that a
+simulation must be validated for empirical realism in the setting where it
+was fitted. This supports holding later conversations out of profile creation,
+drafting from their inbound context without seeing the historical response,
+and only then comparing the candidate with the reference. A match on surface
+features alone does not establish semantic equivalence, authorship, or
+identity.
+
+## Digital-agent results are task-bounded
+
+[Generative Agents](https://arxiv.org/abs/2304.03442) showed that stored
+experiences, retrieval, reflection, and planning can produce believable agent
+behavior in a simulated town. Believability in that environment is not the
+same as fidelity to a real individual.
+
+[Generative Agent Simulations of 1,000 People](https://arxiv.org/abs/2411.10109)
+built agents from two-hour interviews with 1,052 participants and evaluated
+them on surveys, personality measures, and experimental replications. The
+reported survey result is relative to how consistently participants repeated
+their own answers two weeks later. It is not evidence that the agents could
+write private messages like those participants or act on their behalf.
+
+The accompanying [official repository](https://github.com/StanfordHCI/genagents)
+does not publish the interview-derived individual agent bank. It describes
+aggregated access for fixed tasks and reviewed access for individual outputs
+because of participant privacy. That is useful precedent: an open-source
+method can remain public while real person-level evidence stays private.
+
+Message Like Me consequently uses the terms *style profile*, *historical
+tendency*, and *draft candidate*. It does not use a messaging profile to infer
+beliefs, personality, relationship quality, future behavior, or authority to
+represent the user.
+
+## Private text can remain revealing
+
+[Quantifying Memorization Across Neural Language Models](https://arxiv.org/abs/2202.07646)
+found that extractable memorization increased with model capacity, repeated
+examples, and longer prompting context in the evaluated model families.
+[Beyond Memorization](https://arxiv.org/abs/2310.07298) showed that LLMs could
+infer personal attributes from text even when the task was not extraction of a
+memorized training example. Removing obvious names is therefore not a complete
+privacy defense.
+
+[When Personalization Misleads](https://aclanthology.org/2026.findings-acl.395/)
+found that personalization could steer factual answers toward a user's prior
+history rather than objective truth in its evaluated settings. The project
+inference is that meaning, current intent, and factual correctness must outrank
+style fidelity.
+
+[NIST's digital identity risk guidance](https://pages.nist.gov/800-63-4/sp800-63/dirm/)
+lists impersonation, privacy loss, and reputational damage among relevant
+harms. Message Like Me reduces those risks through local storage, bounded
+exports, pseudonymous ordinary views, explicit provenance, and the absence of
+message-sending commands. Those controls do not create consent from a
+conversation partner or make a hosted agent local.
+
+## Tempo is relational and descriptive
+
+A 2026 preprint on [response times in donated WhatsApp and Instagram chats](https://arxiv.org/abs/2605.03687)
+reported persistent response-speed similarity between chat partners in its
+sample. This is preliminary evidence from different platforms and cannot set a
+norm for users of any supported messaging source. It does support comparing
+tempo within a dyad instead of treating one global latency distribution as a
+personal rule.
+
+Historical latency is affected by sleep, work, travel, notifications, device
+availability, urgency, and missing data. Message Like Me reports it as
+descriptive metadata. It does not tell an agent to wait before returning a
+draft or portray a historical delay as a preference or promise.
+
+## Open-source landscape
+
+| Project | Public scope | Relevant distinction |
+| --- | --- | --- |
+| [OpenSelf](https://github.com/Open-Self/Open-Self) | Self-hosted profile and memory system that can automatically reply through WhatsApp, Telegram, and Discord | Message Like Me deliberately ends at an inspectable, unsent draft and does not simulate typing, delay delivery, or operate a messaging account. |
+| [Second-Me](https://github.com/mindverse/Second-Me) | Locally trained and hosted "AI self" with memory, model alignment, and a network | Message Like Me does not train a model, construct an identity, or join an agent network. |
+| [Doppelganger](https://github.com/NotYuSheng/Doppelganger) | LoRA fine-tuning from chat exports | Its documentation warns that trained models can reproduce private data and other participants' text. Message Like Me keeps analysis in inspectable profiles instead of weights. |
+| [Write Like Me](https://github.com/Hiro-Inagawa/write-like-me) | Stylometric profiles for several writing registers with held-out verification | It is useful prior art for measured profiles and verification. Message Like Me focuses on dyadic response context, bubble shape, tempo, and reply behavior. |
+| [imessage-exporter](https://github.com/ReagentX/imessage-exporter) | Broad read-only parsing and export of modern Messages features | It is a valuable compatibility reference. Its GPL-3.0 implementation is not copied into this MIT project. |
+| [iMessageAnalyzer](https://github.com/dsouzarc/imessageanalyzer) | Conversation statistics, starters, and successive-message analysis | Its fixed time thresholds are prior art, not universal conversational facts. Message Like Me records its own versioned thresholds with results. |
+| [iMCP](https://github.com/mattt/iMCP) | Sandboxed native access to Messages and the Contacts framework | It demonstrates a possible future native Contacts boundary. The current CLI uses bounded read-only database snapshots. |
+| [imessage-rag](https://github.com/sapochat/imessage-rag) | Local retrieval and question answering over message history | Retrieval of facts from conversations is a different task from measuring the user's outgoing style. |
+
+## Explicit limitations
+
+The research above does not establish that:
+
+- one stable profile captures how a person writes to every contact;
+- linguistic similarity implies the same intent, judgment, or factual answer;
+- a historical response is the response the user would choose now;
+- contact frequency, latency, or warmth reveals relationship quality;
+- a model-generated draft was authored, approved, or sent by the user;
+- pseudonymous identifiers anonymize a corpus against someone with access to
+  the store or installation key;
+- local CLI processing controls the data practices of the agent environment
+  that opens a study packet; or
+- possession of a conversation database grants permission to publish,
+  fine-tune on, or impersonate its participants.
+
+The defensible claim is smaller: Message Like Me measures selected historical
+messaging behavior, keeps semantic interpretations tied to bounded evidence,
+and helps an already-running agent produce an unsent candidate for user review.

package/package.json

@@ -0,0 +1,82 @@
+{
+  "name": "@hraness/message-like-me",
+  "version": "0.8.0",
+  "description": "A local-first CLI and Agent Skill for studying private messaging history and drafting messages that sound like you.",
+  "license": "MIT",
+  "type": "module",
+  "packageManager": "bun@1.3.14",
+  "engines": {
+    "bun": ">=1.3.14"
+  },
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/hraness/message-like-me.git"
+  },
+  "homepage": "https://messagelikeme.com",
+  "bugs": {
+    "url": "https://github.com/hraness/message-like-me/issues"
+  },
+  "keywords": [
+    "agent-skill",
+    "beeper",
+    "bun",
+    "cli",
+    "imessage",
+    "local-first",
+    "messaging",
+    "messaging-style",
+    "privacy",
+    "wacli",
+    "whatsapp"
+  ],
+  "bin": {
+    "messagelikeme": "./dist/cli.js"
+  },
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.ts",
+      "import": "./dist/index.js"
+    },
+    "./message-bundle-v1": {
+      "types": "./dist/message-bundle-v1.d.ts",
+      "import": "./dist/message-bundle-v1.js"
+    },
+    "./message-bundle-v2": {
+      "types": "./dist/message-bundle-v2.d.ts",
+      "import": "./dist/message-bundle-v2.js"
+    },
+    "./agentic-messaging-v1": {
+      "types": "./dist/agentic-messaging-v1.d.ts",
+      "import": "./dist/agentic-messaging-v1.js"
+    },
+    "./ensoul-source-v1": {
+      "types": "./dist/ensoul-source-v1.d.ts",
+      "import": "./dist/ensoul-source-v1.js"
+    }
+  },
+  "files": [
+    "dist",
+    "docs",
+    "schema",
+    "skills",
+    "CHANGELOG.md",
+    "README.md",
+    "SECURITY.md",
+    "LICENSE"
+  ],
+  "scripts": {
+    "build": "bun scripts/build-dist.ts",
+    "test": "bun test ./src ./scripts",
+    "typecheck": "tsc --noEmit -p tsconfig.json",
+    "check:skill": "bun scripts/validate-skill.ts",
+    "check:standalone": "bun scripts/check-standalone.ts",
+    "check:dist": "bun scripts/check-dist.ts",
+    "check:package": "bun scripts/package-smoke.ts",
+    "check": "bun run typecheck && bun run test && bun run check:skill && bun run check:standalone && bun run build && bun run check:dist && bun run check:package",
+    "prepack": "bun run check"
+  },
+  "devDependencies": {
+    "@types/bun": "1.3.14",
+    "typescript": "6.0.3"
+  }
+}