@aisystemresources/emdee 0.2.2 → 0.3.2

package/.claude/commands/describe-image.md

@@ -0,0 +1,17 @@
+---
+name: describe-image
+description: Describe an EMDEE vault image doc and rename it from timestamp to semantic title.
+argument-hint: <vault-path>
+---
+
+Take the vault doc path from `$ARGUMENTS`. Run the emdee-describe-image skill workflow:
+
+1. `emdee get-image --doc-path $ARGUMENTS --remote --out /tmp/emdee-describe.jpg` — save the image locally
+2. Read the image visually and compose:
+   - A semantic title, 3-6 words, UPPERCASE-with-HYPHENS (e.g. `HANDSTAND-BALANCE-DRILL`)
+   - A one-line summary, 15-30 words
+3. `emdee rename-doc --old-path $ARGUMENTS --new-title <TITLE> --remote`
+4. Get the fresh preamble hash: `emdee get-doc --path <new-path> --remote --json`
+5. `emdee patch-preamble --path <new-path> --body "> <summary>" --expected-hash <hash> --remote`
+
+Report the before/after path + title + summary as a punch list at the end.

package/.claude/commands/summarise.md

@@ -0,0 +1,24 @@
+---
+name: summarise
+description: Refresh drifting doc summaries. Optional path prefix argument narrows the batch.
+argument-hint: [path-prefix]
+---
+
+Run the emdee-summariser skill's batch flow:
+
+1. `emdee list-summary-drift --remote --limit 20 --format text` (append `--prefix $ARGUMENTS` if the user provided one) → get up to 20 candidate paths
+2. For each path, `emdee get-doc --path <path> --remote --full --format text` → read the body
+3. Compose a fresh summary (15-40 words) that answers "what is this and why does it exist"
+4. Present all proposals as a punch list — do NOT patch yet:
+
+```
+<path>
+  Current: <old blockquote>
+  Proposed: <new blockquote>
+```
+
+5. Wait for the user's approval (per-doc or bulk). On approval:
+   - `emdee get-doc --path <path> --remote --json` → fetch preamble.content_hash
+   - `emdee patch-preamble --path <path> --body "> <new>" --expected-hash <hash> --remote`
+
+Never batch-patch without human review — image content is high-signal but the model can misread ambiguous shots.

package/.claude/commands/vault-status.md

@@ -0,0 +1,21 @@
+---
+name: vault-status
+description: Show a mini dashboard of the user's EMDEE vault — total docs + top 3 drifting summaries.
+---
+
+Run these commands via the Bash tool and present the results as a compact dashboard:
+
+1. `emdee list --remote | wc -l` → total doc count
+2. `emdee list-summary-drift --remote --limit 3 --format text` → top 3 drifting paths
+
+Format the output as:
+
+```
+Vault: <N> docs total, <M> with drifting summaries
+Top drift candidates:
+  - <path 1>
+  - <path 2>
+  - <path 3>
+```
+
+If `emdee list --remote` errors with "run login", tell the user to run `emdee login` first and stop.

package/.claude-plugin/plugin.json

@@ -0,0 +1,21 @@
+{
+  "name": "emdee",
+  "displayName": "Emdee — Local-first knowledge graph",
+  "description": "Local-first knowledge graph backed by plain markdown. Bundles the emdee.tech MCP server, the emdee CLI, and 4 Claude Code skills that teach vault conventions + workflows.",
+  "version": "0.2.2",
+  "author": {
+    "name": "AI System Resources",
+    "url": "https://emdee.tech"
+  },
+  "homepage": "https://emdee.tech",
+  "repository": "https://github.com/AISystemResources/emdee",
+  "license": "MIT",
+  "keywords": ["knowledge-graph", "markdown", "vault", "mcp", "claude"],
+  "mcpServers": {
+    "emdee": {
+      "type": "http",
+      "url": "https://emdee.tech/api/mcp"
+    }
+  },
+  "skills": "./skills"
+}

package/README.md

@@ -1,13 +1,30 @@
 # Emdee
 
-Local-first knowledge graph backed by plain markdown. Humans browse it through a Next.js renderer; agents (Claude Code, Claude.ai, Cursor, Codex) read and write the same files through an MCP server. The vault is the source of truth — anything an LLM says traces back to a file you wrote.
+A markdown knowledge graph shared by humans and their AI agents. Files you can edit like Obsidian, a vault you can share like Google Drive, and a live MCP endpoint every agent (Claude Code, Claude.ai, Cursor, Codex) reads and writes to natively. One vault, three surfaces — with the plain markdown you wrote as the only source of truth.
 
 ## Why
 
-LLM agents need a stable, human-readable substrate to read and write their own context over time. Most knowledge-graph tools are either built for humans (Obsidian) or built for agents (vector stores). Emdee is a single substrate for both: the markdown a human edits is the exact bytes an agent reads, with no hidden index, no parallel summaries, no schema gymnastics. Build up a working journal that survives across sessions.
+Every knowledge tool picks two of three: human-friendly, team-shareable, agent-native. Notion is human-friendly and shareable, but its blocks aren't markdown and agents can't natively touch them. Obsidian is human-friendly and agent-touchable (they're just markdown files), but private-by-default — sharing is a paid add-on or a git repo. Google Drive is shareable but not markdown-native. Emdee is all three: the markdown a human edits is the exact bytes an agent reads through MCP and a teammate reads through a share link — no hidden index, no parallel summaries, no schema gymnastics. Build up a working journal that survives across sessions and travels with the people you invite in.
 
 ## Install
 
+**Claude Code plugin (recommended — MCP + CLI + skills in one shot):**
+
+```bash
+claude plugin install @aisystemresources/emdee
+```
+
+This installs the CLI, wires up the emdee.tech MCP server, and drops in 4 auto-loading skills. Then:
+
+```bash
+emdee login                          # PKCE flow, opens browser
+emdee whoami                         # confirms your namespace
+```
+
+Slash commands available after install: `/vault-status`, `/describe-image <path>`, `/summarise [prefix]`.
+
+**CLI-only (no plugin, for scripting or non-Claude-Code use):**
+
 ```bash
 npm install -g @aisystemresources/emdee
 cd ~/my-vault
@@ -16,6 +33,8 @@ emdee list                           # your owner + 5 virtual system nodes
 emdee mcp                            # stdio MCP server — point Claude Code at it
 ```
 
+**claude.ai users:** the plugin doesn't apply (different install model). Use the connect panel on emdee.tech to add the HTTP MCP server directly.
+
 ## The 5-node OS layer
 
 Every Emdee vault has exactly 5 canonical top-level nodes, plus your owner node:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aisystemresources/emdee",
-  "version": "0.2.2",
+  "version": "0.3.2",
   "description": "Local-first document management with markdown rendering, knowledge graph, and MCP server.",
   "type": "module",
   "bin": {
@@ -36,6 +36,8 @@
     "src/mcp/",
     "skills/",
     "templates/",
+    ".claude-plugin/",
+    ".claude/commands/",
     "README.md"
   ],
   "dependencies": {

package/src/cli/remote-client.ts

@@ -37,6 +37,9 @@ export async function callTool(
     method: "POST",
     headers: {
       "Content-Type": "application/json",
+      // MCP streamable-HTTP transport requires BOTH content types in Accept
+      // or it 406s with "Client must accept both application/json and
+      // text/event-stream". We handle whichever shape the server picks below.
       "Accept": "application/json, text/event-stream",
       Authorization: `Bearer ${creds.access_token}`,
     },
@@ -54,12 +57,33 @@ export async function callTool(
     throw new Error(`remote call failed: ${res.status} ${body}`);
   }
 
-  const body = (await res.json()) as JsonRpcResponse;
+  const body = await parseJsonRpc(res);
   if (body.error) throw new Error(`remote tool error: ${body.error.message}`);
   if (!body.result) throw new Error("remote call returned no result");
   return body.result;
 }
 
+// The MCP streamable-HTTP transport content-negotiates on the request Accept:
+// when the client offers both JSON and SSE (as required — see 406 above),
+// the server may reply with either. JSON we parse directly; SSE arrives as
+// one or more `event: message\ndata: <json>\n\n` frames — for a single
+// tools/call response we take the first `data:` payload.
+async function parseJsonRpc(res: Response): Promise<JsonRpcResponse> {
+  const ct = res.headers.get("content-type") ?? "";
+  if (ct.includes("text/event-stream")) {
+    const text = await res.text();
+    const dataLine = text.split(/\r?\n/).find((l) => l.startsWith("data:"));
+    if (!dataLine) throw new Error(`remote call returned SSE with no data frame: ${text.slice(0, 200)}`);
+    const payload = dataLine.slice("data:".length).trim();
+    try {
+      return JSON.parse(payload) as JsonRpcResponse;
+    } catch (e) {
+      throw new Error(`remote call SSE data frame is not valid JSON: ${(e as Error).message}`);
+    }
+  }
+  return (await res.json()) as JsonRpcResponse;
+}
+
 /**
  * Convenience: unwrap the MCP text envelope back to a string. Tools that
  * return JSON encode it inside `content[0].text`; tools that return plain