@omgbase/sync 0.5.20 → 0.6.0

package/README.md

@@ -39,14 +39,14 @@ What it does, in order: spawn `--server` (split on whitespace) as a stdio MCP se
 
 All registered in `packages/core/src/mcp/server.ts`; every one takes an optional `repo` slug.
 
-- **`observe_many({files:[{path,content}]})`** → one result per file: `{id, path, rev, commitId, converged, echo, conflicted, dispositions}` (surface 3.0; `docId`/`path` before). Bytes are committed as an **observed**-origin revision (a write *around* the engine, as a human edit is recorded) with block identity threaded by the reconciler. If `sha256(content)` already equals the doc's stored `file_hash`, it is an **echo**: `echo:true`, `rev:null`, no commit. Bytes containing git conflict markers are still ingested and flagged `conflicted:true`. The batch shares one timestamp and one resurrection-pool sweep. (`observe` is the single-file form; the coordinator uses the batch.)
-- **`observe_delete({path})`** → `{id, path, deleted}`. Tombstones the live doc as an observed deletion (blocks pooled for resurrection, no file removed). Idempotent.
-- **`changes_since({cursor?, origin?, limit?})`** → `{digests, cursor, truncated, head}`. The repo-wide commit feed after a commit `seq`; each digest is `{commit, seq, ts, origin, actor, summary, revisions:[{id, path, hash}]}` (sync 2.0; `{doc, path, contentHash}` before) where `hash` is the hex of that revision's rendered file hash — enough for a client to tell "changed vs echo" without a follow-up read. `head` is the repo's current max seq (cursor > head means the cursor came from another repo).
-- **`docs_read({path})`** — used by `readDoc()` in the export direction to get a doc's current byte-exact content.
+- **`observe_many({files:[{path,content}]})`** → one result per file: `{_id, _path, rev, commitId, converged, echo, conflicted, dispositions}` (surface 5.0 — "underscore = engine": a result key that carries a row attribute is spelled as the attribute; `{id, path, …}` through 4.11, `docId`/`path` before 3.0. The ARGUMENT keeps `path`: a file event's path is a role, not a row attribute). Bytes are committed as an **observed**-origin revision (a write *around* the engine, as a human edit is recorded) with block identity threaded by the reconciler. If `sha256(content)` already equals the doc's stored `file_hash`, it is an **echo**: `echo:true`, `rev:null`, no commit. Bytes containing git conflict markers are still ingested and flagged `conflicted:true`. The batch shares one timestamp and one resurrection-pool sweep. (`observe` is the single-file form; the coordinator uses the batch.)
+- **`observe_delete({path})`** → `{_id, _path, deleted}`. Tombstones the live doc as an observed deletion (blocks pooled for resurrection, no file removed). Idempotent.
+- **`changes_since({cursor?, origin?, limit?})`** → `{digests, cursor, truncated, head}`. The repo-wide commit feed after a commit `seq`; each digest is `{commit, seq, ts, origin, actor, summary, revisions:[{_id, _path, _hash}]}` (sync 3.0 / surface 5.0 — a revision is a row, so its keys are the row's engine names; `{id, path, hash}` in sync 2.0, `{doc, path, contentHash}` before; the digest's own keys — `commit`, `seq`, … — describe a commit, not a row, and stay bare) where `_hash` is the hex of that revision's rendered file hash — enough for a client to tell "changed vs echo" without a follow-up read. `head` is the repo's current max seq (cursor > head means the cursor came from another repo).
+- **`docs_read({_path})`** → `{_id, _path, rev, properties, _raw, …}` — used by `readDoc()` in the export direction to get a doc's current byte-exact content (`_raw`). The argument is `_path` because it addresses the row's `_path` (5.0 has no tolerance: a bare `path` is `filter_invalid` with the corrected call in `data.fix`).
 
 ### Loop safety and echo suppression
 
-A round-trip source → engine → source → engine terminates because the engine's echo gate makes ingest idempotent: re-observing bytes the engine already holds produces no commit. `syncOut()` adds a coarse filter on top — it **skips every `observed`-origin digest**, since those came from a source and there is nothing to push back — leaving only `api`/`import` (engine-authored) changes to write out. The `hash` in the feed (and `contentHash` in `readDoc()`'s own `EngineClient` type) is carried for finer per-path suppression but the v1 coordinator does not yet compare it; the durable per-path hash store (`sync_state`) is the deferred piece below.
+A round-trip source → engine → source → engine terminates because the engine's echo gate makes ingest idempotent: re-observing bytes the engine already holds produces no commit. `syncOut()` adds a coarse filter on top — it **skips every `observed`-origin digest**, since those came from a source and there is nothing to push back — leaving only `api`/`import` (engine-authored) changes to write out. The `_hash` in the feed (and `contentHash` in `readDoc()`'s own `EngineClient` type) is carried for finer per-path suppression but the v1 coordinator does not yet compare it; the durable per-path hash store (`sync_state`) is the deferred piece below.
 
 ## The identity rule
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@omgbase/sync",
-  "version": "0.5.20",
+  "version": "0.6.0",
   "description": "Open Markdown Graph Base — standalone store-to-store synchronizer: a coordinator over an EngineClient seam (in-process or MCP-client) and source adapters",
   "license": "MIT",
   "repository": {
@@ -53,7 +53,7 @@
   },
   "dependencies": {
     "@modelcontextprotocol/sdk": "^1.30.0",
-    "@omgbase/core": "^0.27.0",
+    "@omgbase/core": "^0.28.0",
     "@omgbase/fs-adapter": "^0.2.0"
   },
   "scripts": {