agent-knowledge 1.9.8 → 3.0.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (179) hide show
  1. package/CHANGELOG.md +82 -0
  2. package/README.md +53 -52
  3. package/dist/events.d.ts +16 -0
  4. package/dist/events.d.ts.map +1 -0
  5. package/dist/events.js +31 -0
  6. package/dist/events.js.map +1 -0
  7. package/dist/index.js +31 -13
  8. package/dist/index.js.map +1 -1
  9. package/dist/knowledge/analyze.d.ts.map +1 -1
  10. package/dist/knowledge/analyze.js +18 -11
  11. package/dist/knowledge/analyze.js.map +1 -1
  12. package/dist/knowledge/consolidate.d.ts +2 -2
  13. package/dist/knowledge/consolidate.js +2 -2
  14. package/dist/knowledge/distill.d.ts.map +1 -1
  15. package/dist/knowledge/distill.js +28 -16
  16. package/dist/knowledge/distill.js.map +1 -1
  17. package/dist/knowledge/freshness.js +1 -1
  18. package/dist/knowledge/freshness.js.map +1 -1
  19. package/dist/knowledge/git.d.ts +3 -4
  20. package/dist/knowledge/git.d.ts.map +1 -1
  21. package/dist/knowledge/git.js +62 -49
  22. package/dist/knowledge/git.js.map +1 -1
  23. package/dist/knowledge/graph.d.ts +2 -0
  24. package/dist/knowledge/graph.d.ts.map +1 -1
  25. package/dist/knowledge/graph.js +8 -0
  26. package/dist/knowledge/graph.js.map +1 -1
  27. package/dist/knowledge/ingest-freshness.d.ts +22 -0
  28. package/dist/knowledge/ingest-freshness.d.ts.map +1 -0
  29. package/dist/knowledge/ingest-freshness.js +93 -0
  30. package/dist/knowledge/ingest-freshness.js.map +1 -0
  31. package/dist/knowledge/promote.js +1 -1
  32. package/dist/knowledge/promote.js.map +1 -1
  33. package/dist/knowledge/reflect.js +2 -2
  34. package/dist/knowledge/reflect.js.map +1 -1
  35. package/dist/knowledge/remove.d.ts +11 -0
  36. package/dist/knowledge/remove.d.ts.map +1 -0
  37. package/dist/knowledge/remove.js +24 -0
  38. package/dist/knowledge/remove.js.map +1 -0
  39. package/dist/knowledge/search.d.ts +76 -4
  40. package/dist/knowledge/search.d.ts.map +1 -1
  41. package/dist/knowledge/search.js +364 -37
  42. package/dist/knowledge/search.js.map +1 -1
  43. package/dist/knowledge/similar.d.ts +7 -0
  44. package/dist/knowledge/similar.d.ts.map +1 -0
  45. package/dist/knowledge/similar.js +51 -0
  46. package/dist/knowledge/similar.js.map +1 -0
  47. package/dist/knowledge/store.d.ts +20 -1
  48. package/dist/knowledge/store.d.ts.map +1 -1
  49. package/dist/knowledge/store.js +119 -84
  50. package/dist/knowledge/store.js.map +1 -1
  51. package/dist/lib.d.ts +1 -1
  52. package/dist/lib.d.ts.map +1 -1
  53. package/dist/lib.js.map +1 -1
  54. package/dist/package-meta.d.ts +4 -1
  55. package/dist/package-meta.d.ts.map +1 -1
  56. package/dist/package-meta.js +18 -9
  57. package/dist/package-meta.js.map +1 -1
  58. package/dist/rest-server.d.ts +31 -0
  59. package/dist/rest-server.d.ts.map +1 -0
  60. package/dist/rest-server.js +142 -0
  61. package/dist/rest-server.js.map +1 -0
  62. package/dist/search/bm25.d.ts +8 -0
  63. package/dist/search/bm25.d.ts.map +1 -1
  64. package/dist/search/bm25.js +32 -20
  65. package/dist/search/bm25.js.map +1 -1
  66. package/dist/search/excerpt.d.ts +1 -1
  67. package/dist/search/excerpt.d.ts.map +1 -1
  68. package/dist/search/excerpt.js +19 -1
  69. package/dist/search/excerpt.js.map +1 -1
  70. package/dist/search/lexicon.d.ts +37 -0
  71. package/dist/search/lexicon.d.ts.map +1 -0
  72. package/dist/search/lexicon.js +83 -0
  73. package/dist/search/lexicon.js.map +1 -0
  74. package/dist/search/mmr.d.ts +9 -3
  75. package/dist/search/mmr.d.ts.map +1 -1
  76. package/dist/search/mmr.js +21 -7
  77. package/dist/search/mmr.js.map +1 -1
  78. package/dist/search/types.d.ts +2 -0
  79. package/dist/search/types.d.ts.map +1 -1
  80. package/dist/server.d.ts +1 -5
  81. package/dist/server.d.ts.map +1 -1
  82. package/dist/server.js +1 -13
  83. package/dist/server.js.map +1 -1
  84. package/dist/sessions/human.d.ts +12 -0
  85. package/dist/sessions/human.d.ts.map +1 -0
  86. package/dist/sessions/human.js +52 -0
  87. package/dist/sessions/human.js.map +1 -0
  88. package/dist/sessions/indexer.d.ts +4 -0
  89. package/dist/sessions/indexer.d.ts.map +1 -1
  90. package/dist/sessions/indexer.js +6 -1
  91. package/dist/sessions/indexer.js.map +1 -1
  92. package/dist/sessions/parser.d.ts +39 -0
  93. package/dist/sessions/parser.d.ts.map +1 -1
  94. package/dist/sessions/parser.js +169 -15
  95. package/dist/sessions/parser.js.map +1 -1
  96. package/dist/sessions/search.d.ts +2 -0
  97. package/dist/sessions/search.d.ts.map +1 -1
  98. package/dist/sessions/search.js +147 -67
  99. package/dist/sessions/search.js.map +1 -1
  100. package/dist/sessions/snapshot.d.ts +19 -0
  101. package/dist/sessions/snapshot.d.ts.map +1 -0
  102. package/dist/sessions/snapshot.js +118 -0
  103. package/dist/sessions/snapshot.js.map +1 -0
  104. package/dist/sessions/summary.d.ts +24 -5
  105. package/dist/sessions/summary.d.ts.map +1 -1
  106. package/dist/sessions/summary.js +122 -146
  107. package/dist/sessions/summary.js.map +1 -1
  108. package/dist/slice.d.ts +5 -0
  109. package/dist/slice.d.ts.map +1 -0
  110. package/dist/slice.js +28 -0
  111. package/dist/slice.js.map +1 -0
  112. package/dist/text.d.ts +9 -0
  113. package/dist/text.d.ts.map +1 -0
  114. package/dist/text.js +16 -0
  115. package/dist/text.js.map +1 -0
  116. package/dist/tool-handlers.d.ts.map +1 -1
  117. package/dist/tool-handlers.js +6 -10
  118. package/dist/tool-handlers.js.map +1 -1
  119. package/dist/transport/guard.d.ts +5 -0
  120. package/dist/transport/guard.d.ts.map +1 -0
  121. package/dist/transport/guard.js +51 -0
  122. package/dist/transport/guard.js.map +1 -0
  123. package/dist/transport/http.d.ts +17 -0
  124. package/dist/transport/http.d.ts.map +1 -0
  125. package/dist/transport/http.js +97 -0
  126. package/dist/transport/http.js.map +1 -0
  127. package/dist/transport/rest.d.ts +4 -0
  128. package/dist/transport/rest.d.ts.map +1 -0
  129. package/dist/transport/rest.js +424 -0
  130. package/dist/transport/rest.js.map +1 -0
  131. package/dist/vectorstore/store.d.ts +9 -0
  132. package/dist/vectorstore/store.d.ts.map +1 -1
  133. package/dist/vectorstore/store.js +23 -0
  134. package/dist/vectorstore/store.js.map +1 -1
  135. package/dist/wakeup.d.ts +5 -0
  136. package/dist/wakeup.d.ts.map +1 -1
  137. package/dist/wakeup.js +10 -1
  138. package/dist/wakeup.js.map +1 -1
  139. package/docs/ARCHITECTURE.md +25 -39
  140. package/docs/HOOKS.md +138 -237
  141. package/docs/INGEST.md +1 -1
  142. package/docs/SETUP.md +87 -143
  143. package/docs/USER-MANUAL.md +169 -100
  144. package/package.json +20 -17
  145. package/{skills → plugin/skills}/knowledge-ingest/SKILL.md +1 -1
  146. package/{skills → plugin/skills}/knowledge-ingest/scripts/validate.mjs +2 -2
  147. package/scripts/plugins/opencode/agent-knowledge.ts +37 -86
  148. package/scripts/setup.js +80 -146
  149. package/scripts/source-walk.d.mts +12 -0
  150. package/scripts/source-walk.mjs +94 -0
  151. package/scripts/tree-sitter-extract.mjs +10 -76
  152. package/.claude-plugin/marketplace.json +0 -19
  153. package/.claude-plugin/plugin.json +0 -12
  154. package/.mcp.json +0 -8
  155. package/agent-desk-plugin.json +0 -20
  156. package/dist/dashboard.d.ts +0 -5
  157. package/dist/dashboard.d.ts.map +0 -1
  158. package/dist/dashboard.js +0 -666
  159. package/dist/dashboard.js.map +0 -1
  160. package/dist/ui/app.js +0 -1013
  161. package/dist/ui/index.html +0 -460
  162. package/dist/ui/morphdom.min.js +0 -1
  163. package/dist/ui/render-graph.js +0 -482
  164. package/dist/ui/render-knowledge.js +0 -288
  165. package/dist/ui/render-panel.js +0 -475
  166. package/dist/ui/render-search.js +0 -105
  167. package/dist/ui/render-sessions.js +0 -237
  168. package/dist/ui/styles.css +0 -2322
  169. package/dist/ui/template.js +0 -173
  170. package/dist/ui/ui-utils.js +0 -217
  171. package/dist/ui/vis-network.min.js +0 -34
  172. package/docs/DASHBOARD.md +0 -191
  173. package/scripts/copy-ui.js +0 -14
  174. package/scripts/hooks/first-prompt-inject.mjs +0 -242
  175. package/scripts/hooks/precompact-distill.mjs +0 -105
  176. package/scripts/hooks/precompact-flush.mjs +0 -192
  177. package/scripts/hooks/session-start-ingest.mjs +0 -238
  178. package/scripts/hooks/session-start.js +0 -212
  179. package/scripts/hooks/sessionend-distill.mjs +0 -198
@@ -5,7 +5,7 @@
5
5
  1. [Overview](#1-overview)
6
6
  2. [Installation](#2-installation)
7
7
  3. [Configuration](#3-configuration)
8
- 4. [Dashboard Guide](#4-dashboard-guide)
8
+ 4. [The /knowledge Pane](#4-the-knowledge-pane)
9
9
  5. [MCP Tools Reference](#5-mcp-tools-reference)
10
10
  6. [REST API Reference](#6-rest-api-reference)
11
11
  7. [Knowledge Base Management](#7-knowledge-base-management)
@@ -35,16 +35,17 @@ agent-knowledge is an MCP (Model Context Protocol) server that provides cross-se
35
35
  - **Knowledge analysis** -- find most-connected concepts (god nodes), cross-category bridges, and isolated entries (gaps).
36
36
  - **Knowledge brief** -- compact ~200 token summary of the knowledge base state for session-start orientation.
37
37
  - **Pre-extracted session insights** -- distillation extracts git commits, error patterns, URLs, and package changes from session transcripts via regex (no LLM cost).
38
- - **Real-time dashboard** -- web UI showing knowledge entries, sessions, and search results.
38
+ - **`/knowledge` pane** -- the knowledge base beside the conversation in Claude Code (agent-knowledge plugin).
39
39
 
40
40
  ### Architecture
41
41
 
42
42
  agent-knowledge has two entry points:
43
43
 
44
- | Entry Point | File | Purpose |
45
- | ---------------- | ---------------- | ----------------------------------------------------------------------------------------- |
46
- | MCP stdio server | `dist/index.js` | Communicates with the AI agent via JSON-RPC over stdin/stdout. Auto-starts the dashboard. |
47
- | HTTP server | `dist/server.js` | Standalone dashboard + REST API + WebSocket. |
44
+ | Entry Point | File | Purpose |
45
+ | ---------------- | --------------------- | ---------------------------------------------------------------------------------------------------- |
46
+ | MCP stdio server | `dist/index.js` | Communicates with the AI agent via JSON-RPC over stdin/stdout. |
47
+ | REST server | `dist/rest-server.js` | Started by the MCP process that holds the port (127.0.0.1:3423); the others take over when it exits. |
48
+ | Claude Code mod | `plugin/hooks/` | The `/knowledge` pane, reading the REST API. |
48
49
 
49
50
  Internally:
50
51
 
@@ -97,7 +98,7 @@ npx agent-knowledge
97
98
 
98
99
  | Variable | Default | Description |
99
100
  | ------------------------------------- | ------------------- | ----------------------------------------------- |
100
- | `AGENT_KNOWLEDGE_PORT` | `3423` | Dashboard HTTP/WebSocket port |
101
+ | `AGENT_KNOWLEDGE_PORT` | `3423` | Local REST API port |
101
102
  | `AGENT_KNOWLEDGE_DATA_DIR` | (platform config) | Override primary host data root (auto-detected) |
102
103
  | `AGENT_KNOWLEDGE_GIT_URL` | (none) | Git remote URL for knowledge base sync |
103
104
  | `AGENT_KNOWLEDGE_MEMORY_DIR` | `~/agent-knowledge` | Local knowledge base directory |
@@ -119,7 +120,7 @@ Add to `~/.claude.json`:
119
120
  }
120
121
  ```
121
122
 
122
- The dashboard auto-starts at http://localhost:3423.
123
+ The first MCP process also serves the local REST API on `127.0.0.1:3423`; in Claude Code, `/knowledge` (agent-knowledge plugin) shows the knowledge base in a pane.
123
124
 
124
125
  ### Permissions (settings.json)
125
126
 
@@ -141,20 +142,18 @@ knowledge_admin with action "config", memory_dir "/custom/path/knowledge"
141
142
  knowledge_admin with action "config", auto_distill false
142
143
  ```
143
144
 
144
- ### Lifecycle Hooks
145
+ ### The session lifecycle
145
146
 
146
- agent-knowledge ships six lifecycle hook scripts that integrate with Claude Code's event system. Running `node scripts/setup.js` installs all six into `~/.claude/settings.json`; on other hosts that don't support lifecycle hooks (Cursor, Windsurf), the MCP tools still work — you just lose the automatic context injection.
147
+ With the Claude Code plugin installed, each session works with the knowledge base on its own; there is nothing to configure:
147
148
 
148
- | Script | Event | Purpose |
149
- | -------------------------- | ---------------- | ------------------------------------------------------------------------- |
150
- | `session-start.js` | SessionStart | Dashboard URL + auto-loads token-budgeted wakeup payload into context |
151
- | `session-start-ingest.mjs` | SessionStart | Detects project + reports knowledge-ingest cache drift via SHA256 diff |
152
- | `first-prompt-inject.mjs` | UserPromptSubmit | Query-targeted knowledge hits injected on the session's first real prompt |
153
- | `precompact-flush.mjs` | PreCompact | Rich session summary on disk + save-unsaved-context nudge into context |
154
- | `precompact-distill.mjs` | PreCompact | Lightweight text snapshot of recent user prompts |
155
- | `sessionend-distill.mjs` | SessionEnd | Final summary (turn counts, tool uses, first 20 prompts) |
149
+ | When | What happens |
150
+ | -------------------------------- | --------------------------------------------------------------------------------------------------- |
151
+ | The conversation's first message | The model knows what the knowledge base holds, who you are (or asks you once), and the ingest state |
152
+ | Your first real prompt | The entries that match it ride beside it; one line in the transcript names them |
153
+ | A compaction | The session is snapshotted into `sessions/<project>/`; the summary keeps what is still to be saved |
154
+ | The session ending or `/clear` | The session's last snapshot; open `/knowledge` panes toast it |
156
155
 
157
- Every hook fails open — if a script errors, it logs to stderr and the session continues. Each hook has env-var toggles (`AGENT_KNOWLEDGE_AUTOWAKE`, `AGENT_KNOWLEDGE_FIRSTPROMPT_INJECT`, `AGENT_KNOWLEDGE_PRECOMPACT_NUDGE`, etc.) — see [`docs/HOOKS.md`](HOOKS.md) for the full reference, test coverage, and budget/threshold knobs.
156
+ Every part fails open: with the server down the session goes on as without the plugin. Each has a switch (`AGENT_KNOWLEDGE_AUTOWAKE`, `AGENT_KNOWLEDGE_FIRSTPROMPT_INJECT`, `AGENT_KNOWLEDGE_PRECOMPACT_NUDGE`, budgets) — see [`docs/HOOKS.md`](HOOKS.md). Hosts without the plugin (Cursor, Windsurf) keep the MCP tools; OpenCode has its own plugin.
158
157
 
159
158
  ### Persistent Memory — agent-knowledge, not host auto-memory
160
159
 
@@ -176,60 +175,148 @@ sessions.
176
175
 
177
176
  ---
178
177
 
179
- ## 4. Dashboard Guide
178
+ ## 4. The /knowledge Pane
180
179
 
181
- ### Accessing the Dashboard
180
+ In Claude Code (2.1.289+), the agent-knowledge plugin adds `/knowledge`: the knowledge base in a pane beside the conversation. Install it with:
182
181
 
183
- The dashboard is available at **http://localhost:3423** (or the port configured via `AGENT_KNOWLEDGE_PORT`).
182
+ ```bash
183
+ claude plugin marketplace add keshrath/agent-knowledge
184
+ claude plugin install agent-knowledge@agent-knowledge
185
+ ```
186
+
187
+ The pane reads the local REST API (section 6). There is no web dashboard: it was removed in 2.0.
188
+
189
+ Every picture below is from a recorded session in the real Claude Code CLI against a scratch knowledge base (`npm run manual:record`).
190
+
191
+ ![The /knowledge pane: open it, open an entry, follow a link](images/manual/hero.gif)
192
+
193
+ ### Opening it
194
+
195
+ | Type | Opens |
196
+ | -------------------- | --------------------------------------------------------------------------- |
197
+ | `/knowledge` | Browse, every entry newest first |
198
+ | `/knowledge <words>` | Search, with the hits for the words (entries and past sessions) |
199
+ | `/knowledge <path>` | that entry's page; `.md` may be left off (`/knowledge decisions/rest-only`) |
200
+
201
+ The command's reply goes to the model too, so it says something only when something is wrong: nothing found for the words, no such entry, nothing answers at the port, or the pane could not be shown (then with the counts). Plain `/knowledge` shows every entry, clearing a Find or category left from before. The pane opens beside the conversation in the fullscreen layout, above the prompt otherwise; an open pane is opened afresh.
202
+
203
+ ![Opening the pane](images/manual/open.gif)
204
+
205
+ ### Keys
206
+
207
+ | Key | Does |
208
+ | ------------------- | ------------------------------------------------------------------- |
209
+ | `1` `2` `3` `4` | Browse, Search, Sessions, Health |
210
+ | `r` | Refresh |
211
+ | `Tab` / `shift+Tab` | move the focus ring |
212
+ | `Enter` | press what holds the ring (open a row, follow a link, fold, delete) |
213
+ | `↑` `↓` | scroll the pane |
214
+ | `Esc` | give the keyboard back to the prompt |
215
+ | `ctrl+x tab` | take it again; the ring is where you left it |
216
+
217
+ The line under the tabs says what the keys do. In a text field (Find, Search) the digits and `r` type into the field; the tab bar drops their numbers and the line says `Tab leaves the field`. With the mouse (fullscreen layout), a click presses a tab, a row or a button.
218
+
219
+ The same line shows what an action is doing (`searching "zebrafish"…`) and how it went (`✓ synced with the remote`, `✗ sync failed: …`), and clears itself after a few seconds.
220
+
221
+ ### Browse
222
+
223
+ ![Browse](images/manual/browse.png)
184
224
 
185
- ### Knowledge Tab
225
+ Every entry, newest first. Under each title: its category, `evergreen` for entries kept from decay, its maturity once it is past `candidate` (`established`, `proven`), its first tags and how long ago it was updated. Files at the top of the knowledge folder are under `other`; the session snapshots are their own category, `snapshots`, last in the bar and not in `all`. An entry another session writes or deletes shows on the next poll, without a key; the ring stays on the row it held, and while the pane has the keyboard the rows keep their order (a new one comes at the end). An empty knowledge base offers the next step: **Ingest &lt;project&gt; ›** puts `/knowledge-ingest <folder>` in the prompt for you to run.
186
226
 
187
- Shows all knowledge base entries organized by category. Each entry card displays:
227
+ ![Browse, the whole screen](images/manual/browse-full.png)
188
228
 
189
- - **Title** extracted from the Markdown content or filename.
190
- - **Category** badge (`projects`, `people`, `decisions`, `workflows`, `notes`).
191
- - **Tags** from YAML frontmatter.
192
- - **Maturity** level (`candidate`, `established`, `proven`).
193
- - **Access count** -- how many times the entry has been read.
194
- - **Last accessed** timestamp.
229
+ The category bar narrows the list; the counts are the knowledge base's.
195
230
 
196
- Click an entry to view its full Markdown content, related entries (from the knowledge graph), and score data.
231
+ ![A category](images/manual/category.gif)
197
232
 
198
- The Knowledge tab header includes analysis buttons:
233
+ **Find** narrows by title, path or tag; a row found by a tag says `tag <name>`, and the list is headed **Matches** (`Matches in notes` with a category picked). When nothing matches, the ring is on **Clear the filter**.
199
234
 
200
- - **Duplicates** — scans entries for near-duplicates (TF-IDF similarity)
201
- - **Reflect** — finds entries with no graph connections, generates a structured prompt for the agent
202
- - **God Nodes** — opens a panel showing the most-connected entries (your core concepts)
203
- - **Bridges** — shows entries that connect different categories with a `why` explanation
204
- - **Gaps** — lists isolated entries (0-1 edges) sorted by maturity
205
- - **Brief** — displays the knowledge base summary suitable for session-start orientation
235
+ ![Find](images/manual/find.png)
206
236
 
207
- ### Sessions Tab
237
+ ![Finding by a tag](images/manual/find.gif)
208
238
 
209
- Lists discovered sessions from all configured sources (Claude Code, Cursor, OpenCode). Each session shows:
239
+ A long list comes 25 rows at a time: `▸ 25 more` adds the next page and puts the ring on its first row, shown whole.
210
240
 
211
- - **Session ID** (UUID).
212
- - **Project** name.
213
- - **Message count** and duration.
214
- - **Last modified** timestamp.
241
+ ### An entry
215
242
 
216
- Click a session to view its messages.
243
+ ![An entry](images/manual/entry.png)
244
+
245
+ Its title, path, category, maturity, how often agents read it and when it was updated; its tags, author and when an agent last read it (reads from the pane do not count). **Links** are its graph edges both ways: `→` it points at another entry, `←` another points at it, with the relation (`depends on`, `related to`, …) and where the link came from when it was not made by hand (`auto-link`). `Enter` on a link opens that entry; `‹` goes back the way you came, with the ring on the link you followed. Dates read as the rows read them up to a week (`2d ago`), then as the date.
246
+
247
+ ![Following a link and back](images/manual/entry.gif)
248
+
249
+ The body is drawn as Markdown, without its frontmatter (those are the facts above) and without a first heading that only repeats the title. Code is drawn apart, a dim `│` before each line: a long command that wraps continues indented, with no bar, so it never reads as a second one. Links in the text are listed with the links (`→ mentions`); a table too wide for the pane reads a row a line, each value with its header (`Owner: alice`). A long body folds after 30 lines: `▸ Show all n lines`. The page follows its file: an edit from another session shows on the next poll, said (`! notes/x.md changed elsewhere`), and an entry deleted elsewhere takes you back to the list (`! notes/x.md was deleted elsewhere`).
250
+
251
+ ![Folding a long body](images/manual/fold.gif)
252
+
253
+ **Delete** is at the end of the page and asks first, with **Keep it** first and the ring on it, and **Delete notes/x.md** beside it. The question says what happens: without a git remote the file goes from this machine for good, with one the deletion is pushed too (its git history keeps it), and how many entries that link to it lose that link. A remote that refuses is said as a warning (`! deleted notes/x.md, not pushed: …`).
254
+
255
+ ![Delete asks first](images/manual/confirm-delete.png)
256
+
257
+ ![Deleting](images/manual/delete.gif)
217
258
 
218
259
  ### Search
219
260
 
220
- The search bar performs hybrid search across both knowledge entries and sessions. Results are ranked by relevance combining TF-IDF and semantic similarity scores.
261
+ ![Search](images/manual/search.png)
262
+
263
+ Search finds words, and close spellings, in every entry and in past sessions (Claude Code, Cursor, OpenCode). Best matches first: **Entries** with their category and a line of the text around the match, then **In past sessions** with the project, when, whether you asked it or the agent answered, and the line. An excerpt is the entry's text without its headings and code blocks, paragraphs, list items and table rows joined with `·`. The ring lands on the best hit; `Enter` opens it. `2` again (or `Enter` on the tab) puts the ring in the field with the last search's words, to edit. With no hit the page says so: a search typed in the field keeps the ring there to try again, and one run as `/knowledge <words>` leaves the ring on a button, so the next command you type runs as a command.
264
+
265
+ ![Searching and opening a hit](images/manual/search.gif)
266
+
267
+ ### Sessions
268
+
269
+ The recent sessions on this machine, newest first: the project's folder name (this session marked `(this session)`), when it started, how long it ran, the git branch and the first thing asked (a slash command is not a question). Sessions where nothing was asked, a `/clear` alone say, are folded behind `▸ n more where nothing was asked`. `▸ 25 more` goes on past the newest 100, loading older sessions from the server. A session opens on its summary: how many prompts, its id and how to resume it (`claude --resume <id>`), what was asked, files changed, commits, errors (in red) and the tools used. A session snapshot's page opens its session when its transcript is on this machine, and back leads to the snapshot.
270
+
271
+ ![A session's summary](images/manual/session.png)
272
+
273
+ ![Sessions](images/manual/sessions.gif)
221
274
 
222
- ### Stats
275
+ ### Health
276
+
277
+ ![Health](images/manual/health.png)
278
+
279
+ How many entries (by category), session snapshots and sessions there are, whether search uses meaning too (`words and meaning: n vectors`) or words only, where the knowledge folder is and whether a git remote syncs it, and whether this project's code is ingested (**Ingest &lt;project&gt; ›** when it was never, or files changed since). **Sync now** (shown with a remote) pulls and pushes; the ring never lands on it by itself. **Recent activity** lists what was reported, such as a session saved or a sync that failed. **Analyze links** lists the most linked entries (two links or more) and, apart from them, the ones with one link or none, fewest first, 30 at a time (`▸ 30 more`): those are the entries nothing much leads to. The folder is shown whole, wrapped when long.
280
+
281
+ ![Analyzing the links](images/manual/health.gif)
282
+
283
+ ### Status line, toasts and what the model sees
284
+
285
+ The status line under the prompt reads `agent-knowledge: 8 entries · 2 sessions`, from the moment a session starts; `offline` when nothing answers, `old server x on the port` when an agent-knowledge from before the pane does. While the pane is shown it says the last outcome for its 10 seconds (it may be scrolled out of view). Outcomes are marked: ✓ done, ✗ failed or refused, ! changed elsewhere, – nothing happened.
286
+
287
+ A live event (a session saved at its end) is toasted once, beside the conversation (`Session saved: demo · 2026-11-13 · fix the sync`; three at most, then `+n more`), and listed under Health. A project never ingested, or files changed since its ingest, is toasted once with the conversation's first message.
288
+
289
+ ![A toast](images/manual/toast.png)
290
+
291
+ The model is told, once per conversation, what the knowledge base holds (`agent-knowledge holds 8 entries · 2 sessions (projects 2, …)`), to search it before saying something is unknown, and that you browse it in `/knowledge`.
292
+
293
+ ### Layouts
294
+
295
+ Docked in the fullscreen layout the pane takes the right side of the window; on a narrow window it is about 40 columns wide and its tab bar wraps between two tabs. Without the fullscreen layout it opens above the prompt.
296
+
297
+ ![Docked narrow](images/manual/narrow.png)
298
+
299
+ ![Above the prompt](images/manual/inline.png)
300
+
301
+ ### When the server is away
302
+
303
+ One agent-knowledge MCP process serves the REST API on `127.0.0.1:3423`; when its session ends another one takes the port over within a few seconds. While nothing answers, the pane says so and **Refresh** tries again (`✗ still nothing answers at http://127.0.0.1:3423 (12:04:31)`); the tabs stay, so `1`–`4` never land in the prompt, and it comes back by itself on the next poll.
223
304
 
224
- Displays vector store statistics: total entries, knowledge entries, session entries, database size, embedding provider, and dimensions.
305
+ ![The server away](images/manual/server-down.png)
225
306
 
226
- ### Theme Toggle
307
+ ![The server away and back](images/manual/server-down.gif)
227
308
 
228
- Dark and light themes available. Preference saved in `localStorage`.
309
+ ### Troubleshooting
229
310
 
230
- ### Real-Time Updates
311
+ | You see | Do |
312
+ | ----------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------- |
313
+ | `The agent-knowledge server is not reachable` | Check in `/mcp` that agent-knowledge is connected; `AGENT_KNOWLEDGE_PORT` must be the same for the pane and the server. |
314
+ | `An older agent-knowledge (1.x) holds the port` | End the sessions (or the process) still running 1.x; a current one takes the port over. Then **Refresh**. |
315
+ | `words only: no embeddings yet` | Search still works; for meaning too, configure an embedding provider (section 3). |
316
+ | `no remote: entries stay on this machine` | Set `AGENT_KNOWLEDGE_GIT_URL` or make the knowledge folder a git clone. |
317
+ | `pane not shown: …` after `/knowledge` | The window is too narrow for a docked pane; widen it or leave the fullscreen layout. |
231
318
 
232
- The dashboard connects via WebSocket. On connect, it receives the full state snapshot. A file watcher monitors the UI directory for hot-reload during development. The state snapshot is cached for 30 seconds to avoid expensive disk and database scans.
319
+ The pictures are recorded with `npm run manual:record` (needs `npm run build` and a logged-in `claude`; it makes no model calls). `npm run manual:record -- health` writes only the pictures whose name contains `health`.
233
320
 
234
321
  ---
235
322
 
@@ -746,65 +833,50 @@ knowledge_analyze with action "search_gaps", since_days 7, min_count 3
746
833
 
747
834
  ## 6. REST API Reference
748
835
 
749
- The REST API is served by the dashboard. All endpoints are loopback only: foreign `Host`/`Origin` headers get 403, request bodies must be `application/json`, no wildcard CORS. Rate limited: 100 requests/minute general, 20 requests/minute for search/analyze endpoints.
836
+ The first MCP process serves a local REST API on `127.0.0.1:3423` (`AGENT_KNOWLEDGE_PORT`); the other sessions retry the port and take it over when that process exits. It is what the Claude Code plugin (the `/knowledge` pane and the session lifecycle), the opencode plugin and agent-tasks call. Loopback only: a foreign `Host` or any `Origin` header gets 403, request bodies must be `application/json`, no CORS. A `:path` with a slash is sent encoded (`projects%2Fbackend.md`).
750
837
 
751
838
  ### Health
752
839
 
753
840
  ```
754
- GET /health Status, version, uptime, knowledge entry count
841
+ GET /health Status, version, uptime, entries (curated) and snapshots
755
842
  ```
756
843
 
757
844
  ### Knowledge Entries
758
845
 
759
846
  ```
760
- GET /api/knowledge List entries (?category=&tag=)
761
- GET /api/knowledge/:path Read entry content (enriched with score data)
762
- GET /api/knowledge/:path/links Get graph edges for an entry
763
- GET /api/knowledge/search Search entries (?q=&category=&max_results=)
764
- GET /api/knowledge/consolidate Find duplicates (?category=&threshold=)
765
- GET /api/knowledge/reflect Find unconnected entries (?category=&max_entries=)
766
- GET /api/knowledge/god-nodes Most-connected entries (?top_n=)
767
- GET /api/knowledge/bridges Cross-category connectors (?top_n=)
768
- GET /api/knowledge/gaps Isolated entries (?max_entries=)
769
- GET /api/knowledge/brief Cached knowledge base brief
847
+ GET /api/overview entries (curated), snapshots, categories (curated), recent events (the pane)
848
+ GET /api/knowledge List entries with scoring (?category=&tag=)
849
+ GET /api/knowledge/search Search entries (?q=&category=&max_results=&mmr=true&skip=sessions&min_terms=2&prompt=true; prompt: q is a prompt, reply-format phrases dropped)
850
+ GET /api/knowledge/:path Read entry content (enriched with score data; a snapshot adds session_on_machine); 404 when missing
851
+ GET /api/knowledge/:path/links Graph edges for an entry, each with the other end's title
852
+ DELETE /api/knowledge/:path Delete an entry (git synced)
853
+ POST /api/knowledge Write an entry: { category, filename, content } -> 201 { path, git, autoLinks?, duplicateWarnings? }
854
+ POST /api/knowledge/sync git pull + push
855
+ GET /api/analysis { hubs, sparse, sparse_total, offset } (?offset=&limit=30): sparse fewest links first, then by title
770
856
  ```
771
857
 
772
- ### `GET /api/knowledge/god-nodes`
773
-
774
- Query params: `top_n` (number, default 10).
775
-
776
- Returns: same as `knowledge_analyze(action: "god_nodes")`.
777
-
778
- ### `GET /api/knowledge/bridges`
779
-
780
- Query params: `top_n` (number, default 5).
781
-
782
- Returns: same as `knowledge_analyze(action: "bridges")`.
783
-
784
- ### `GET /api/knowledge/gaps`
785
-
786
- Query params: `max_entries` (number, default 30).
787
-
788
- Returns: same as `knowledge_analyze(action: "gaps")`.
789
-
790
- ### `GET /api/knowledge/brief`
858
+ ### Sessions
791
859
 
792
- No query params. Returns the cached knowledge brief.
860
+ ```
861
+ GET /api/sessions Sessions, paged: { sessions, total } (?project=&limit=&offset=; limit default 50, max 500);
862
+ userMessageCount = real prompts (no /commands), preview '' without one
863
+ GET /api/sessions/search Search sessions (?q=&role=&max_results=&ranked=&project=&semantic=); hits carry name
864
+ GET /api/sessions/:id/summary Session summary (?project=); 404 when unknown
865
+ POST /api/sessions/:id/snapshot { event: compact|end, cwd? } -> { path, prompts, messages, tools }: the session into
866
+ sessions/<project>/<day>-<id>.md (path null: nothing was asked); 404 when unknown
867
+ ```
793
868
 
794
- ### Sessions
869
+ ### Session lifecycle
795
870
 
796
871
  ```
797
- GET /api/sessions List sessions (?project=&limit=&offset=)
798
- GET /api/sessions/:id Get session messages (?project=&include_tools=&tail=)
799
- GET /api/sessions/:id/summary Get session summary (?project=)
800
- GET /api/sessions/search Search sessions (?q=&role=&max_results=&ranked=&project=&semantic=)
801
- GET /api/sessions/recall Scoped recall (?scope=&q=&max_results=&project=)
872
+ GET /api/wakeup The wakeup bundle (?sections=identity,...&budget=) and onboarding: configured|declined|pending
873
+ GET /api/ingest The folder's ingest freshness (?cwd=): state none|never|current|stale, project, counts
802
874
  ```
803
875
 
804
- ### Index Status
876
+ ### Events
805
877
 
806
878
  ```
807
- GET /api/index-status Vector store statistics
879
+ POST /api/events { kind, message } -> 201 { id, ts }: a receipt the /knowledge pane toasts (any host)
808
880
  ```
809
881
 
810
882
  ### Example Requests
@@ -817,7 +889,7 @@ curl http://localhost:3423/health
817
889
  curl http://localhost:3423/api/knowledge
818
890
 
819
891
  # Read a specific entry
820
- curl http://localhost:3423/api/knowledge/projects/backend.md
892
+ curl http://localhost:3423/api/knowledge/projects%2Fbackend.md
821
893
 
822
894
  # Search knowledge
823
895
  curl "http://localhost:3423/api/knowledge/search?q=authentication&max_results=5"
@@ -825,9 +897,6 @@ curl "http://localhost:3423/api/knowledge/search?q=authentication&max_results=5"
825
897
  # Search sessions
826
898
  curl "http://localhost:3423/api/sessions/search?q=deploy+error&role=assistant&max_results=10"
827
899
 
828
- # Scoped recall
829
- curl "http://localhost:3423/api/sessions/recall?scope=errors&q=ECONNREFUSED"
830
-
831
900
  # List sessions for a project
832
901
  curl "http://localhost:3423/api/sessions?project=backend&limit=20"
833
902
  ```
@@ -983,14 +1052,14 @@ If no provider is available, search falls back to pure TF-IDF.
983
1052
 
984
1053
  ## 9. Troubleshooting
985
1054
 
986
- ### Dashboard Won't Start
1055
+ ### The /knowledge pane says the server is not reachable
987
1056
 
988
- **Symptom:** Port 3423 already in use.
1057
+ **Cause:** no running session serves the REST API, or something other than agent-knowledge holds port 3423.
989
1058
 
990
1059
  **Solutions:**
991
1060
 
992
- 1. Set `AGENT_KNOWLEDGE_PORT=3424` in the MCP config env.
993
- 2. Multiple MCP instances share the same database. Only one serves the dashboard.
1061
+ 1. Check in `/mcp` that agent-knowledge is connected: the first MCP process serves the port, and when it exits another session takes it over within about 5 seconds.
1062
+ 2. If another program holds 3423, set `AGENT_KNOWLEDGE_PORT=3424` for the MCP server and the pane alike.
994
1063
 
995
1064
  ### Search Returns No Results
996
1065
 
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "agent-knowledge",
3
- "version": "1.9.8",
3
+ "version": "3.0.0",
4
4
  "description": "Cross-session memory and recall for AI agents — git-synced knowledge base, knowledge graph, confidence scoring, hybrid semantic+TF-IDF search, auto-distillation with secrets scrubbing",
5
5
  "type": "module",
6
6
  "main": "dist/index.js",
@@ -26,17 +26,14 @@
26
26
  "files": [
27
27
  "dist/",
28
28
  "scripts/",
29
- "skills/",
30
- ".claude-plugin/",
31
- ".mcp.json",
32
- "agent-desk-plugin.json",
29
+ "plugin/skills/",
33
30
  "docs/*.md",
34
31
  "README.md",
35
32
  "CHANGELOG.md",
36
33
  "LICENSE"
37
34
  ],
38
35
  "scripts": {
39
- "build": "tsc && node scripts/copy-ui.js",
36
+ "build": "tsc",
40
37
  "start": "node dist/index.js",
41
38
  "dev": "tsc --watch",
42
39
  "test": "vitest run",
@@ -46,14 +43,17 @@
46
43
  "bench:longmemeval:matrix": "tsx bench/longmemeval-matrix.ts",
47
44
  "bench:promote": "tsx bench/promote-bench.ts",
48
45
  "test:coverage": "vitest run --coverage",
49
- "lint": "eslint src/ tests/",
50
- "lint:fix": "eslint src/ tests/ --fix",
46
+ "lint": "eslint src/ tests/ plugin/",
47
+ "lint:fix": "eslint src/ tests/ plugin/ --fix",
51
48
  "format": "prettier --write .",
52
49
  "format:check": "prettier --check .",
53
50
  "typecheck": "tsc --noEmit",
54
51
  "check": "npm run typecheck && npm run lint && npm run format:check && npm run test",
52
+ "plugin:check": "claude plugin validate plugin && claude plugin test plugin",
55
53
  "clean": "rm -rf dist",
56
- "prepare": "husky"
54
+ "prepare": "husky",
55
+ "e2e:claude": "tsc --outDir dist-e2e && node tests/e2e-claude/run.mjs",
56
+ "manual:record": "tsc --outDir dist-e2e && tsx tests/e2e-claude/manual.ts"
57
57
  },
58
58
  "keywords": [
59
59
  "mcp",
@@ -84,25 +84,28 @@
84
84
  "homepage": "https://github.com/keshrath/agent-knowledge#readme",
85
85
  "dependencies": {
86
86
  "@modelcontextprotocol/sdk": "^1.0.0",
87
- "agent-common": "^1.1.2",
88
87
  "better-sqlite3": "^12.8.0",
89
- "morphdom": "^2.7.8",
90
88
  "sqlite-vec": "^0.1.0",
91
89
  "tree-sitter-wasms": "^0.1.13",
92
- "vis-data": "^8.0.3",
93
- "vis-network": "^10.0.2",
94
- "web-tree-sitter": "^0.24.7",
95
- "ws": "^8.20.0"
90
+ "web-tree-sitter": "^0.24.7"
96
91
  },
97
92
  "devDependencies": {
98
93
  "@eslint/js": "^10.0.1",
94
+ "@playwright/test": "^1.59.1",
99
95
  "@types/better-sqlite3": "^7.6.0",
100
96
  "@types/node": "^25.5.0",
97
+ "@types/pngjs": "^6.0.5",
101
98
  "@types/stream-json": "^1.7.8",
102
- "@types/ws": "^8.18.1",
99
+ "@xterm/addon-serialize": "^0.14.0",
100
+ "@xterm/addon-unicode11": "^0.9.0",
101
+ "@xterm/headless": "^6.0.0",
102
+ "@xterm/xterm": "^6.0.0",
103
103
  "eslint": "^10.1.0",
104
+ "gifenc": "^1.0.3",
104
105
  "husky": "^9.0.0",
105
106
  "lint-staged": "^16.0.0",
107
+ "node-pty": "^1.1.0",
108
+ "pngjs": "^7.0.0",
106
109
  "prettier": "^3.8.1",
107
110
  "stream-json": "^2.1.0",
108
111
  "tsx": "^4.21.0",
@@ -114,7 +117,7 @@
114
117
  "@huggingface/transformers": "3.8.1"
115
118
  },
116
119
  "lint-staged": {
117
- "*.{ts,js}": [
120
+ "*.{ts,tsx,js,mjs}": [
118
121
  "prettier --write",
119
122
  "eslint --fix"
120
123
  ],
@@ -180,7 +180,7 @@ knowledge_graph({ action: "link", source: "decisions/<project>-<decision>.md", t
180
180
  23. Run the validation script to verify the ingest produced correct results:
181
181
 
182
182
  ```bash
183
- node "<agent-knowledge-repo>/skills/knowledge-ingest/scripts/validate.mjs" "<target-path>"
183
+ node "<agent-knowledge-repo>/plugin/skills/knowledge-ingest/scripts/validate.mjs" "<target-path>"
184
184
  ```
185
185
 
186
186
  24. Parse the JSON output. If `status` is `FAIL`:
@@ -134,7 +134,7 @@ function main() {
134
134
  if (cache.files && typeof cache.files === 'object') {
135
135
  let missingHashes = 0;
136
136
  let missingEntryRefs = 0;
137
- for (const [filePath, meta] of Object.entries(cache.files)) {
137
+ for (const meta of Object.values(cache.files)) {
138
138
  if (!meta.sha256 || typeof meta.sha256 !== 'string' || meta.sha256.length !== 64) {
139
139
  missingHashes++;
140
140
  }
@@ -163,7 +163,7 @@ function main() {
163
163
  issues,
164
164
  };
165
165
 
166
- console.log(JSON.stringify(report, null, 2));
166
+ process.stdout.write(`${JSON.stringify(report, null, 2)}\n`);
167
167
  process.exit(issues.length === 0 ? 0 : 1);
168
168
  }
169
169