claude-db 0.2.3 → 0.4.0
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.
- package/README.md +17 -4
- package/dist/capture/flush.d.ts +20 -1
- package/dist/capture/flush.js +45 -5
- package/dist/capture/git.d.ts +2 -0
- package/dist/capture/git.js +96 -0
- package/dist/capture/index.d.ts +4 -3
- package/dist/capture/index.js +4 -3
- package/dist/capture/manual.d.ts +9 -0
- package/dist/capture/manual.js +0 -0
- package/dist/capture/transcript.d.ts +8 -0
- package/dist/capture/transcript.js +30 -0
- package/dist/capture/turn-extractor.d.ts +9 -1
- package/dist/capture/turn-extractor.js +74 -13
- package/dist/cli/index.js +421 -27
- package/dist/cli/install.d.ts +8 -0
- package/dist/cli/install.js +31 -0
- package/dist/mcp/render.d.ts +23 -0
- package/dist/mcp/render.js +78 -0
- package/dist/mcp/server.d.ts +1 -7
- package/dist/mcp/server.js +73 -36
- package/dist/store/adapter.d.ts +10 -0
- package/dist/store/adapter.js +8 -0
- package/dist/store/mongo/index.d.ts +1 -0
- package/dist/store/mongo/index.js +34 -3
- package/dist/store/postgres/index.d.ts +15 -0
- package/dist/store/postgres/index.js +89 -4
- package/dist/store/sqlite/index.d.ts +1 -0
- package/dist/store/sqlite/index.js +32 -2
- package/dist/types.d.ts +16 -0
- package/dist/usages/find.d.ts +29 -0
- package/dist/usages/find.js +115 -0
- package/dist/usages/index.d.ts +2 -0
- package/dist/usages/index.js +1 -0
- package/dist/util/snippet.d.ts +28 -0
- package/dist/util/snippet.js +52 -0
- package/package.json +2 -1
- package/skills/cdb-scan/SKILL.md +72 -0
package/dist/cli/install.js
CHANGED
|
@@ -33,6 +33,28 @@ export function instructionsPathFor(scope, project) {
|
|
|
33
33
|
? join(project, 'CLAUDE.local.md')
|
|
34
34
|
: join(homedir(), '.claude', 'CLAUDE.md');
|
|
35
35
|
}
|
|
36
|
+
/**
|
|
37
|
+
* Where the `/cdb-scan` skill is installed.
|
|
38
|
+
*
|
|
39
|
+
* A skill rather than a CLI command: surveying a codebase means reading it, and
|
|
40
|
+
* the agent is already sitting in the repo with the tools to do that. A command
|
|
41
|
+
* would need its own model, its own key and its own opinion about what matters.
|
|
42
|
+
*/
|
|
43
|
+
export function skillPathFor(scope, project) {
|
|
44
|
+
const root = scope === 'project' ? join(project, '.claude') : join(homedir(), '.claude');
|
|
45
|
+
return join(root, 'skills', 'cdb-scan', 'SKILL.md');
|
|
46
|
+
}
|
|
47
|
+
function writeSkill(distDir, scope, project) {
|
|
48
|
+
const source = resolve(distDir, '..', 'skills', 'cdb-scan', 'SKILL.md');
|
|
49
|
+
const body = readText(source);
|
|
50
|
+
// Absent only in a partial checkout; the rest of the install still stands.
|
|
51
|
+
if (body.length > 0)
|
|
52
|
+
writeAtomic(skillPathFor(scope, project), body);
|
|
53
|
+
}
|
|
54
|
+
function removeSkill(scope, project) {
|
|
55
|
+
const path = skillPathFor(scope, project);
|
|
56
|
+
rmSync(dirname(path), { recursive: true, force: true });
|
|
57
|
+
}
|
|
36
58
|
const BLOCK_START = '<!-- claude-db:start -->';
|
|
37
59
|
const BLOCK_END = '<!-- claude-db:end -->';
|
|
38
60
|
const INSTRUCTIONS = [
|
|
@@ -52,6 +74,13 @@ const INSTRUCTIONS = [
|
|
|
52
74
|
'Verifying against the code afterwards is right — memory records what was',
|
|
53
75
|
'done, the source shows what is. Starting there is what wastes the memory.',
|
|
54
76
|
'',
|
|
77
|
+
'`search` covers *why*: past decisions, bugs, reasoning. It does not know',
|
|
78
|
+
'*where* something is used right now — for that, call `find_usages` instead',
|
|
79
|
+
'of grepping yourself, specifically before editing or removing a shared or',
|
|
80
|
+
'exported name, and when asked what uses something. It re-derives the answer',
|
|
81
|
+
'from the current source on every call, so it is never stale. Plain-text or',
|
|
82
|
+
'log searching is still a job for grep.',
|
|
83
|
+
'',
|
|
55
84
|
'When the user states a standing rule or preference, record it with',
|
|
56
85
|
'`remember` so it survives this session.',
|
|
57
86
|
BLOCK_END,
|
|
@@ -165,6 +194,7 @@ export function install(distDir, scope, project) {
|
|
|
165
194
|
writeJson(path, settings);
|
|
166
195
|
const server = resolve(distDir, 'mcp', 'server.js');
|
|
167
196
|
writeInstructions(instructionsPathFor(scope, project));
|
|
197
|
+
writeSkill(distDir, scope, project);
|
|
168
198
|
// A global install lands in ~/.claude.json, which holds all of Claude Code's
|
|
169
199
|
// own state. Let its CLI own that file rather than round-tripping megabytes
|
|
170
200
|
// of someone else's config through JSON.parse to add one key. The direct
|
|
@@ -225,6 +255,7 @@ export function uninstall(distDir, scope, project) {
|
|
|
225
255
|
delete settings['hooks'];
|
|
226
256
|
writeJson(path, settings);
|
|
227
257
|
removeInstructions(instructionsPathFor(scope, project));
|
|
258
|
+
removeSkill(scope, project);
|
|
228
259
|
// Symmetric with install: if the CLI put it there, the CLI takes it out.
|
|
229
260
|
if (scope === 'global' && claudeMcp(['mcp', 'remove', 'memory', '-s', 'user'])) {
|
|
230
261
|
return path;
|
|
@@ -0,0 +1,23 @@
|
|
|
1
|
+
import type { Observation, ObservationIndexEntry } from '../types.js';
|
|
2
|
+
/**
|
|
3
|
+
* How the three layers are shown to the agent.
|
|
4
|
+
*
|
|
5
|
+
* Split out of the server so it can be exercised without opening a stdio
|
|
6
|
+
* transport: this is the format every search result is paid in, and it had no
|
|
7
|
+
* test of any kind while the server module could only be imported by starting
|
|
8
|
+
* a server.
|
|
9
|
+
*/
|
|
10
|
+
/**
|
|
11
|
+
* Layer 1. Every character here is multiplied by the number of results and
|
|
12
|
+
* paid on every search, so the format stays terse: short id, kind, month-day,
|
|
13
|
+
* title, and one indented line of the body around the match.
|
|
14
|
+
*
|
|
15
|
+
* The snippet is the half that was missing. Terseness only pays if the row can
|
|
16
|
+
* be chosen from what is shown, and a title often cannot carry that — fifteen
|
|
17
|
+
* results dated the same day, titled things like "Committed as 30daa92", leave
|
|
18
|
+
* a reader nothing to pick on, so they expand two at random and pay a full body
|
|
19
|
+
* for each guess. Sixty characters here saves several thousand there.
|
|
20
|
+
*/
|
|
21
|
+
export declare function renderIndex(entries: ObservationIndexEntry[]): string;
|
|
22
|
+
/** Layer 3. Full detail for one observation, bounded by `chars`. */
|
|
23
|
+
export declare function renderFull(obs: Observation, chars: number): string;
|
|
@@ -0,0 +1,78 @@
|
|
|
1
|
+
import { toShortId } from '../util/shortid.js';
|
|
2
|
+
import { clipBody } from '../util/snippet.js';
|
|
3
|
+
/**
|
|
4
|
+
* How the three layers are shown to the agent.
|
|
5
|
+
*
|
|
6
|
+
* Split out of the server so it can be exercised without opening a stdio
|
|
7
|
+
* transport: this is the format every search result is paid in, and it had no
|
|
8
|
+
* test of any kind while the server module could only be imported by starting
|
|
9
|
+
* a server.
|
|
10
|
+
*/
|
|
11
|
+
/**
|
|
12
|
+
* Layer 1. Every character here is multiplied by the number of results and
|
|
13
|
+
* paid on every search, so the format stays terse: short id, kind, month-day,
|
|
14
|
+
* title, and one indented line of the body around the match.
|
|
15
|
+
*
|
|
16
|
+
* The snippet is the half that was missing. Terseness only pays if the row can
|
|
17
|
+
* be chosen from what is shown, and a title often cannot carry that — fifteen
|
|
18
|
+
* results dated the same day, titled things like "Committed as 30daa92", leave
|
|
19
|
+
* a reader nothing to pick on, so they expand two at random and pay a full body
|
|
20
|
+
* for each guess. Sixty characters here saves several thousand there.
|
|
21
|
+
*/
|
|
22
|
+
export function renderIndex(entries) {
|
|
23
|
+
if (entries.length === 0)
|
|
24
|
+
return 'No matching observations.';
|
|
25
|
+
const rows = entries.map((entry) => {
|
|
26
|
+
const date = new Date(entry.createdAt).toISOString().slice(5, 10);
|
|
27
|
+
const head = `${toShortId(entry.id)} ${entry.kind} ${date} ${entry.title}`;
|
|
28
|
+
// Absent for a row only vector search found: it holds no query terms to
|
|
29
|
+
// centre a fragment on. Dropped too when it only restates the title.
|
|
30
|
+
const worth = entry.snippet !== undefined && !echoesTitle(entry.snippet, entry.title);
|
|
31
|
+
return worth ? `${head}\n ${entry.snippet}` : head;
|
|
32
|
+
});
|
|
33
|
+
return `${entries.length} result(s):\n${rows.join('\n')}`;
|
|
34
|
+
}
|
|
35
|
+
/** Lowercased alphanumeric runs — the only form the two sides share. */
|
|
36
|
+
function words(text) {
|
|
37
|
+
return text.toLowerCase().match(/[a-z0-9]+/g) ?? [];
|
|
38
|
+
}
|
|
39
|
+
/**
|
|
40
|
+
* True when a fragment says little the title has not already said.
|
|
41
|
+
*
|
|
42
|
+
* A search engine centres its window on the query match, and the title is taken
|
|
43
|
+
* from the body, so on a short observation the two are frequently the same
|
|
44
|
+
* sentence — the row then pays a second time for a line already shown. Observed
|
|
45
|
+
* live: a title of "Added as section 3 of PLAN.md, with the later sections
|
|
46
|
+
* renumbered" under a fragment reading "…Added as section 3 of [PLAN.md](PLAN.md),
|
|
47
|
+
* with the later…".
|
|
48
|
+
*
|
|
49
|
+
* Compared as a bag of words rather than by substring, because that example is
|
|
50
|
+
* exactly the case a substring test misses: the fragment comes from the raw body
|
|
51
|
+
* and still carries the markdown the title had stripped.
|
|
52
|
+
*/
|
|
53
|
+
function echoesTitle(snippet, title) {
|
|
54
|
+
const tokens = words(snippet);
|
|
55
|
+
if (tokens.length === 0)
|
|
56
|
+
return true;
|
|
57
|
+
const inTitle = new Set(words(title));
|
|
58
|
+
const shared = tokens.filter((word) => inTitle.has(word)).length;
|
|
59
|
+
// Not 1.0: the window carries the ellipses and a spill word or two from the
|
|
60
|
+
// neighbouring sentence, so an exact echo never scores as one.
|
|
61
|
+
return shared / tokens.length >= 0.7;
|
|
62
|
+
}
|
|
63
|
+
/** Layer 3. Full detail for one observation, bounded by `chars`. */
|
|
64
|
+
export function renderFull(obs, chars) {
|
|
65
|
+
return [
|
|
66
|
+
`id: ${toShortId(obs.id)}`,
|
|
67
|
+
`kind: ${obs.kind}`,
|
|
68
|
+
`when: ${new Date(obs.createdAt).toISOString()}`,
|
|
69
|
+
obs.author ? `who: ${obs.author}` : null,
|
|
70
|
+
obs.files.length > 0 ? `files: ${obs.files.join(', ')}` : null,
|
|
71
|
+
'',
|
|
72
|
+
obs.title,
|
|
73
|
+
'',
|
|
74
|
+
clipBody(obs.body, chars),
|
|
75
|
+
]
|
|
76
|
+
.filter((line) => line !== null)
|
|
77
|
+
.join('\n');
|
|
78
|
+
}
|
package/dist/mcp/server.d.ts
CHANGED
|
@@ -1,8 +1,2 @@
|
|
|
1
1
|
#!/usr/bin/env node
|
|
2
|
-
|
|
3
|
-
/**
|
|
4
|
-
* Layer 1 rendering. Every character here is multiplied by the number of
|
|
5
|
-
* results and paid on every search, so the format is deliberately terse:
|
|
6
|
-
* short id, kind, month-day, title. No padding, no separators, no year.
|
|
7
|
-
*/
|
|
8
|
-
export declare function renderIndex(entries: ObservationIndexEntry[]): string;
|
|
2
|
+
export {};
|
package/dist/mcp/server.js
CHANGED
|
@@ -8,7 +8,9 @@ import { z } from 'zod';
|
|
|
8
8
|
import { remember } from '../capture/index.js';
|
|
9
9
|
import { createContext } from '../context.js';
|
|
10
10
|
import { toShortId } from '../util/shortid.js';
|
|
11
|
+
import { renderFull, renderIndex } from './render.js';
|
|
11
12
|
import { resolveProject } from '../util/project.js';
|
|
13
|
+
import { findUsages, formatUsages } from '../usages/index.js';
|
|
12
14
|
import { silenceSqliteWarning } from '../util/warnings.js';
|
|
13
15
|
silenceSqliteWarning();
|
|
14
16
|
const KINDS = [
|
|
@@ -36,21 +38,28 @@ function packageVersion() {
|
|
|
36
38
|
const ctx = await createContext();
|
|
37
39
|
const server = new McpServer({ name: 'claude-db', version: packageVersion() });
|
|
38
40
|
server.tool('search', 'Layer 1. Search project memory and get a compact index of matches ' +
|
|
39
|
-
'(id, kind, title,
|
|
40
|
-
'
|
|
41
|
+
'(id, kind, date, title, and a line of the matching body) without full ' +
|
|
42
|
+
'bodies. Start here, and use the snippet to decide which ids are worth ' +
|
|
43
|
+
'passing to get_observations rather than expanding on the title alone.', {
|
|
41
44
|
query: z.string().describe('Natural language description of what you need'),
|
|
42
45
|
project: z
|
|
43
46
|
.string()
|
|
44
47
|
.optional()
|
|
45
48
|
.describe('Absolute project path; defaults to cwd. Pass "*" to search every project'),
|
|
46
49
|
kind: z.enum(KINDS).optional(),
|
|
50
|
+
tag: z
|
|
51
|
+
.string()
|
|
52
|
+
.optional()
|
|
53
|
+
.describe('Limit to one repository or top-level directory, e.g. "backend". Useful ' +
|
|
54
|
+
'when a workspace pools several repos under one project'),
|
|
47
55
|
limit: z.number().int().min(1).max(50).default(10),
|
|
48
|
-
}, async ({ query, project, kind, limit }) => {
|
|
56
|
+
}, async ({ query, project, kind, tag, limit }) => {
|
|
49
57
|
const entries = await ctx.search.search({
|
|
50
58
|
text: query,
|
|
51
59
|
limit,
|
|
52
60
|
...(project === '*' ? {} : { project: resolveProject(project) }),
|
|
53
61
|
...(kind ? { kind } : {}),
|
|
62
|
+
...(tag ? { tag } : {}),
|
|
54
63
|
});
|
|
55
64
|
return { content: [{ type: 'text', text: renderIndex(entries) }] };
|
|
56
65
|
});
|
|
@@ -61,11 +70,19 @@ server.tool('remember', 'Record something the user stated outright that no trans
|
|
|
61
70
|
text: z.string().describe('What to remember, in full. First line becomes the title'),
|
|
62
71
|
kind: z.enum(KINDS).default('preference'),
|
|
63
72
|
project: z.string().optional().describe('Absolute project path; defaults to cwd'),
|
|
64
|
-
|
|
73
|
+
key: z
|
|
74
|
+
.string()
|
|
75
|
+
.optional()
|
|
76
|
+
.describe('Stable name for a note meant to be kept current, e.g. "profile:stack". ' +
|
|
77
|
+
'Remembering the same key again replaces it instead of adding a duplicate'),
|
|
78
|
+
tags: z.array(z.string()).max(5).optional().describe('Extra tags for filtering'),
|
|
79
|
+
}, async ({ text, kind, project, key, tags }) => {
|
|
65
80
|
const observation = await remember(ctx, {
|
|
66
81
|
project: resolveProject(project),
|
|
67
82
|
text,
|
|
68
83
|
kind,
|
|
84
|
+
...(key ? { key } : {}),
|
|
85
|
+
...(tags ? { tags } : {}),
|
|
69
86
|
});
|
|
70
87
|
return {
|
|
71
88
|
content: [
|
|
@@ -94,40 +111,60 @@ server.tool('timeline', 'Layer 2. Show what happened immediately before and afte
|
|
|
94
111
|
});
|
|
95
112
|
server.tool('get_observations', 'Layer 3. Fetch full bodies for specific observation ids, using the short ' +
|
|
96
113
|
'ids returned by search. Always batch every id you need into one call ' +
|
|
97
|
-
'rather than calling repeatedly.', {
|
|
114
|
+
'rather than calling repeatedly.', {
|
|
115
|
+
ids: z.array(z.string()).min(1).max(25),
|
|
116
|
+
chars: z
|
|
117
|
+
.number()
|
|
118
|
+
.int()
|
|
119
|
+
.min(200)
|
|
120
|
+
.max(20000)
|
|
121
|
+
.default(2000)
|
|
122
|
+
.describe('Characters of each body to return. Bodies run to 4000 and this call ' +
|
|
123
|
+
'takes 25 ids, so the default keeps a batch readable; raise it when a ' +
|
|
124
|
+
'truncated answer says there is more'),
|
|
125
|
+
}, async ({ ids, chars }) => {
|
|
98
126
|
const observations = await ctx.search.getObservations(ids);
|
|
99
|
-
return {
|
|
127
|
+
return {
|
|
128
|
+
content: [
|
|
129
|
+
{ type: 'text', text: observations.map((obs) => renderFull(obs, chars)).join('\n\n---\n\n') },
|
|
130
|
+
],
|
|
131
|
+
};
|
|
132
|
+
});
|
|
133
|
+
server.tool('find_usages', 'Find real usages of a symbol or component name via `git grep` — a live read ' +
|
|
134
|
+
'of the current source, not a stored index, so it is never stale. Returns ' +
|
|
135
|
+
'file:line and the matching line, with a best-effort [definition?] marker on ' +
|
|
136
|
+
'lines that look like a declaration. Use this before editing or removing a ' +
|
|
137
|
+
'shared or exported name, or to answer "what uses this" (structure). Use ' +
|
|
138
|
+
'`search` instead for "why is this the way it is" (history).', {
|
|
139
|
+
symbol: z.string().min(1).max(200).describe('Exact name to search for, e.g. "useAuth" or "CartButton"'),
|
|
140
|
+
path: z
|
|
141
|
+
.string()
|
|
142
|
+
.optional()
|
|
143
|
+
.describe('Directory inside the repo to search from; defaults to cwd. Must be inside ' +
|
|
144
|
+
'a git working tree. Different from the memory tools’ project param, which ' +
|
|
145
|
+
'can point at a folder pooling several repos rather than being one itself'),
|
|
146
|
+
regex: z
|
|
147
|
+
.boolean()
|
|
148
|
+
.default(false)
|
|
149
|
+
.describe('Treat symbol as an extended regular expression instead of a literal name'),
|
|
150
|
+
context: z
|
|
151
|
+
.number()
|
|
152
|
+
.int()
|
|
153
|
+
.min(0)
|
|
154
|
+
.max(10)
|
|
155
|
+
.default(0)
|
|
156
|
+
.describe('Lines of surrounding context before/after each match'),
|
|
157
|
+
limit: z
|
|
158
|
+
.number()
|
|
159
|
+
.int()
|
|
160
|
+
.min(1)
|
|
161
|
+
.max(500)
|
|
162
|
+
.default(100)
|
|
163
|
+
.describe('Cap on matching lines returned; the result says how many more exist'),
|
|
164
|
+
}, async ({ symbol, path, regex, context, limit }) => {
|
|
165
|
+
const result = findUsages({ symbol, regex, context, limit, ...(path ? { path } : {}) });
|
|
166
|
+
return { content: [{ type: 'text', text: formatUsages(result) }] };
|
|
100
167
|
});
|
|
101
|
-
/**
|
|
102
|
-
* Layer 1 rendering. Every character here is multiplied by the number of
|
|
103
|
-
* results and paid on every search, so the format is deliberately terse:
|
|
104
|
-
* short id, kind, month-day, title. No padding, no separators, no year.
|
|
105
|
-
*/
|
|
106
|
-
export function renderIndex(entries) {
|
|
107
|
-
if (entries.length === 0)
|
|
108
|
-
return 'No matching observations.';
|
|
109
|
-
const rows = entries.map((entry) => {
|
|
110
|
-
const date = new Date(entry.createdAt).toISOString().slice(5, 10);
|
|
111
|
-
return `${toShortId(entry.id)} ${entry.kind} ${date} ${entry.title}`;
|
|
112
|
-
});
|
|
113
|
-
return `${entries.length} result(s):\n${rows.join('\n')}`;
|
|
114
|
-
}
|
|
115
|
-
function renderFull(obs) {
|
|
116
|
-
const date = new Date(obs.createdAt).toISOString();
|
|
117
|
-
return [
|
|
118
|
-
`id: ${toShortId(obs.id)}`,
|
|
119
|
-
`kind: ${obs.kind}`,
|
|
120
|
-
`when: ${date}`,
|
|
121
|
-
obs.author ? `who: ${obs.author}` : null,
|
|
122
|
-
obs.files.length > 0 ? `files: ${obs.files.join(', ')}` : null,
|
|
123
|
-
'',
|
|
124
|
-
obs.title,
|
|
125
|
-
'',
|
|
126
|
-
obs.body,
|
|
127
|
-
]
|
|
128
|
-
.filter((line) => line !== null)
|
|
129
|
-
.join('\n');
|
|
130
|
-
}
|
|
131
168
|
const transport = new StdioServerTransport();
|
|
132
169
|
await server.connect(transport);
|
|
133
170
|
for (const signal of ['SIGINT', 'SIGTERM']) {
|
package/dist/store/adapter.d.ts
CHANGED
|
@@ -42,6 +42,15 @@ export interface MemoryStore {
|
|
|
42
42
|
list(filter: ListFilter): Promise<Observation[]>;
|
|
43
43
|
/** Every project with memory in this database, most recently active first. */
|
|
44
44
|
listProjects(): Promise<ProjectSummary[]>;
|
|
45
|
+
/**
|
|
46
|
+
* Tables or collections in this database that belong to somebody else.
|
|
47
|
+
*
|
|
48
|
+
* Called by `use` before `init()` creates ours, because afterwards there is
|
|
49
|
+
* no way to tell an empty memory database from an application database we
|
|
50
|
+
* just added two tables to. One `use` in anger pointed at a live e-commerce
|
|
51
|
+
* database and nothing said a word.
|
|
52
|
+
*/
|
|
53
|
+
inventory(): Promise<string[]>;
|
|
45
54
|
/** Lexical match. Returns scores on an adapter-defined scale. */
|
|
46
55
|
searchKeyword(query: SearchQuery): Promise<ObservationIndexEntry[]>;
|
|
47
56
|
/** Vector match. Return [] when the backend has no vector support. */
|
|
@@ -56,6 +65,7 @@ export interface MemoryStore {
|
|
|
56
65
|
* SessionStart loses sessions whose work is still there.
|
|
57
66
|
*/
|
|
58
67
|
export declare function isWholeScope(filter: RemoveFilter): boolean;
|
|
68
|
+
export declare function foreignNames(names: string[]): string[];
|
|
59
69
|
export interface StoreFactory {
|
|
60
70
|
/** Lowercase URI schemes this factory claims, e.g. ['mongodb','mongodb+srv']. */
|
|
61
71
|
schemes: string[];
|
package/dist/store/adapter.js
CHANGED
|
@@ -8,3 +8,11 @@
|
|
|
8
8
|
export function isWholeScope(filter) {
|
|
9
9
|
return !filter.ids && !filter.kind && filter.before === undefined;
|
|
10
10
|
}
|
|
11
|
+
/**
|
|
12
|
+
* Ours, on any backend: the two tables, SQLite's FTS shadow tables, and the
|
|
13
|
+
* Postgres schema-version row.
|
|
14
|
+
*/
|
|
15
|
+
const OURS = /^(claude_db_meta|sessions|observations)(_fts(_\w+)?)?$/;
|
|
16
|
+
export function foreignNames(names) {
|
|
17
|
+
return names.filter((name) => name.length > 0 && !OURS.test(name));
|
|
18
|
+
}
|
|
@@ -29,6 +29,7 @@ export declare class MongoStore implements MemoryStore {
|
|
|
29
29
|
remove(filter: RemoveFilter): Promise<number>;
|
|
30
30
|
list(filter: ListFilter): Promise<Observation[]>;
|
|
31
31
|
listProjects(): Promise<ProjectSummary[]>;
|
|
32
|
+
inventory(): Promise<string[]>;
|
|
32
33
|
searchKeyword(query: SearchQuery): Promise<ObservationIndexEntry[]>;
|
|
33
34
|
searchVector(vector: number[], query: SearchQuery): Promise<ObservationIndexEntry[]>;
|
|
34
35
|
timeline(query: TimelineQuery): Promise<ObservationIndexEntry[]>;
|
|
@@ -1,7 +1,8 @@
|
|
|
1
1
|
import { importMongo } from './driver.js';
|
|
2
|
-
import { isWholeScope } from '../adapter.js';
|
|
2
|
+
import { foreignNames, isWholeScope } from '../adapter.js';
|
|
3
3
|
import { cosine } from '../../util/vector.js';
|
|
4
4
|
import { partitionIds } from '../../util/shortid.js';
|
|
5
|
+
import { toSnippet } from '../../util/snippet.js';
|
|
5
6
|
/**
|
|
6
7
|
* Works against any MongoDB: Atlas, self-hosted, or a local container.
|
|
7
8
|
*
|
|
@@ -150,13 +151,37 @@ export class MongoStore {
|
|
|
150
151
|
lastActive: row.last,
|
|
151
152
|
}));
|
|
152
153
|
}
|
|
154
|
+
async inventory() {
|
|
155
|
+
// Through `command` rather than `listCollections`, so the structural driver
|
|
156
|
+
// types this adapter compiles against stay as small as they are.
|
|
157
|
+
const res = await this.db.command({ listCollections: 1, nameOnly: true });
|
|
158
|
+
const batch = res['cursor']?.firstBatch ?? [];
|
|
159
|
+
return foreignNames(batch.map((doc) => String(doc['name'] ?? '')));
|
|
160
|
+
}
|
|
153
161
|
async searchKeyword(query) {
|
|
154
162
|
const filter = {
|
|
155
163
|
$text: { $search: query.text },
|
|
156
164
|
...this.scopeFilter(query),
|
|
157
165
|
};
|
|
166
|
+
// $text has no fragment operator, so unlike FTS5 and ts_headline this is
|
|
167
|
+
// the head of the body rather than the text around the match. Still enough
|
|
168
|
+
// to tell two same-titled rows apart, and it costs one projected field
|
|
169
|
+
// instead of shipping whole bodies to slice them here.
|
|
170
|
+
//
|
|
171
|
+
// Inclusion throughout: MongoDB rejects a projection that mixes included
|
|
172
|
+
// and excluded fields, so the previous `body: 0, embedding: 0` form cannot
|
|
173
|
+
// carry a computed one.
|
|
158
174
|
const docs = await this.observations
|
|
159
|
-
.find(filter, {
|
|
175
|
+
.find(filter, {
|
|
176
|
+
projection: {
|
|
177
|
+
kind: 1,
|
|
178
|
+
title: 1,
|
|
179
|
+
project: 1,
|
|
180
|
+
createdAt: 1,
|
|
181
|
+
snippet: { $substrCP: ['$body', 0, 240] },
|
|
182
|
+
score: { $meta: 'textScore' },
|
|
183
|
+
},
|
|
184
|
+
})
|
|
160
185
|
.sort({ score: { $meta: 'textScore' } })
|
|
161
186
|
.limit(query.limit)
|
|
162
187
|
.toArray();
|
|
@@ -232,6 +257,8 @@ export class MongoStore {
|
|
|
232
257
|
filter['project'] = query.project;
|
|
233
258
|
if (query.kind)
|
|
234
259
|
filter['kind'] = query.kind;
|
|
260
|
+
if (query.tag)
|
|
261
|
+
filter['tags'] = query.tag;
|
|
235
262
|
if (query.since !== undefined || query.until !== undefined) {
|
|
236
263
|
const range = {};
|
|
237
264
|
if (query.since !== undefined)
|
|
@@ -314,7 +341,7 @@ function toObservation(doc) {
|
|
|
314
341
|
return obs;
|
|
315
342
|
}
|
|
316
343
|
function toIndexEntry(doc, score) {
|
|
317
|
-
|
|
344
|
+
const entry = {
|
|
318
345
|
id: doc._id,
|
|
319
346
|
kind: doc.kind,
|
|
320
347
|
title: doc.title,
|
|
@@ -322,4 +349,8 @@ function toIndexEntry(doc, score) {
|
|
|
322
349
|
createdAt: doc.createdAt,
|
|
323
350
|
score,
|
|
324
351
|
};
|
|
352
|
+
const snippet = toSnippet(doc.snippet);
|
|
353
|
+
if (snippet)
|
|
354
|
+
entry.snippet = snippet;
|
|
355
|
+
return entry;
|
|
325
356
|
}
|
|
@@ -8,7 +8,21 @@ export declare class PostgresStore implements MemoryStore {
|
|
|
8
8
|
private warnedDims;
|
|
9
9
|
private constructor();
|
|
10
10
|
static create(uri: string): Promise<PostgresStore>;
|
|
11
|
+
/**
|
|
12
|
+
* Every hook is its own process, so nothing here is ever pooled or reused:
|
|
13
|
+
* this ran in full on every single prompt. Against a managed Postgres, where
|
|
14
|
+
* one round trip measured 49ms at best and 1045ms at the median, four of them
|
|
15
|
+
* before the first real query is most of why a remote backend feels broken.
|
|
16
|
+
*
|
|
17
|
+
* The version row collapses that to one read on the steady-state path. It
|
|
18
|
+
* carries the vector state too, because otherwise proving the extension is
|
|
19
|
+
* available costs back the round trip just saved. Anything unexpected — no
|
|
20
|
+
* table, an older version, an unreadable row — falls through to the full DDL,
|
|
21
|
+
* so the worst case is exactly the behaviour this replaces.
|
|
22
|
+
*/
|
|
11
23
|
init(): Promise<void>;
|
|
24
|
+
private readMeta;
|
|
25
|
+
private writeMeta;
|
|
12
26
|
/**
|
|
13
27
|
* CREATE TABLE IF NOT EXISTS leaves an existing generated column alone, so a
|
|
14
28
|
* database created before tags were indexed keeps the old expression. Rebuild
|
|
@@ -36,6 +50,7 @@ export declare class PostgresStore implements MemoryStore {
|
|
|
36
50
|
remove(filter: RemoveFilter): Promise<number>;
|
|
37
51
|
list(filter: ListFilter): Promise<Observation[]>;
|
|
38
52
|
listProjects(): Promise<ProjectSummary[]>;
|
|
53
|
+
inventory(): Promise<string[]>;
|
|
39
54
|
searchKeyword(query: SearchQuery): Promise<ObservationIndexEntry[]>;
|
|
40
55
|
searchVector(vector: number[], query: SearchQuery): Promise<ObservationIndexEntry[]>;
|
|
41
56
|
timeline(query: TimelineQuery): Promise<ObservationIndexEntry[]>;
|
|
@@ -1,5 +1,6 @@
|
|
|
1
|
-
import { isWholeScope } from '../adapter.js';
|
|
1
|
+
import { foreignNames, isWholeScope } from '../adapter.js';
|
|
2
2
|
import { partitionIds } from '../../util/shortid.js';
|
|
3
|
+
import { toSnippet } from '../../util/snippet.js';
|
|
3
4
|
/**
|
|
4
5
|
* Postgres backend. Uses tsvector for keyword search, and pgvector for
|
|
5
6
|
* embeddings when the extension is available. If `CREATE EXTENSION vector`
|
|
@@ -16,6 +17,11 @@ const TSV_EXPRESSION = `
|
|
|
16
17
|
setweight(to_tsvector('english', coalesce(title, '')), 'A') ||
|
|
17
18
|
setweight(to_tsvector('english', coalesce(tags::text, '')), 'B') ||
|
|
18
19
|
setweight(to_tsvector('english', coalesce(body, '')), 'C')`;
|
|
20
|
+
/**
|
|
21
|
+
* Bumped whenever the DDL below changes, which makes every process re-run it
|
|
22
|
+
* once. Only ever compared for equality.
|
|
23
|
+
*/
|
|
24
|
+
const SCHEMA_VERSION = 1;
|
|
19
25
|
export class PostgresStore {
|
|
20
26
|
pool;
|
|
21
27
|
kind = 'postgres';
|
|
@@ -30,9 +36,34 @@ export class PostgresStore {
|
|
|
30
36
|
const pool = new pg.Pool({ connectionString: uri });
|
|
31
37
|
return new PostgresStore(pool);
|
|
32
38
|
}
|
|
39
|
+
/**
|
|
40
|
+
* Every hook is its own process, so nothing here is ever pooled or reused:
|
|
41
|
+
* this ran in full on every single prompt. Against a managed Postgres, where
|
|
42
|
+
* one round trip measured 49ms at best and 1045ms at the median, four of them
|
|
43
|
+
* before the first real query is most of why a remote backend feels broken.
|
|
44
|
+
*
|
|
45
|
+
* The version row collapses that to one read on the steady-state path. It
|
|
46
|
+
* carries the vector state too, because otherwise proving the extension is
|
|
47
|
+
* available costs back the round trip just saved. Anything unexpected — no
|
|
48
|
+
* table, an older version, an unreadable row — falls through to the full DDL,
|
|
49
|
+
* so the worst case is exactly the behaviour this replaces.
|
|
50
|
+
*/
|
|
33
51
|
async init() {
|
|
52
|
+
const meta = await this.readMeta();
|
|
53
|
+
if (meta?.version === SCHEMA_VERSION) {
|
|
54
|
+
this.vectorEnabled = meta.vectorEnabled;
|
|
55
|
+
this.vectorDims = meta.vectorDims;
|
|
56
|
+
return;
|
|
57
|
+
}
|
|
34
58
|
this.vectorEnabled = await this.tryEnableVector();
|
|
35
59
|
await this.pool.query(`
|
|
60
|
+
CREATE TABLE IF NOT EXISTS claude_db_meta (
|
|
61
|
+
id INT PRIMARY KEY,
|
|
62
|
+
version INT NOT NULL,
|
|
63
|
+
vector_enabled BOOLEAN NOT NULL DEFAULT FALSE,
|
|
64
|
+
vector_dims INT
|
|
65
|
+
);
|
|
66
|
+
|
|
36
67
|
CREATE TABLE IF NOT EXISTS sessions (
|
|
37
68
|
id TEXT PRIMARY KEY,
|
|
38
69
|
project TEXT NOT NULL,
|
|
@@ -60,6 +91,7 @@ export class PostgresStore {
|
|
|
60
91
|
CREATE INDEX IF NOT EXISTS idx_obs_tsv ON observations USING GIN(tsv);
|
|
61
92
|
CREATE INDEX IF NOT EXISTS idx_obs_project_time
|
|
62
93
|
ON observations(project, created_at DESC);
|
|
94
|
+
CREATE INDEX IF NOT EXISTS idx_obs_tags ON observations USING GIN(tags);
|
|
63
95
|
|
|
64
96
|
ALTER TABLE observations ADD COLUMN IF NOT EXISTS embedder TEXT;
|
|
65
97
|
ALTER TABLE observations ADD COLUMN IF NOT EXISTS author TEXT;
|
|
@@ -67,6 +99,33 @@ export class PostgresStore {
|
|
|
67
99
|
await this.ensureTagsIndexed();
|
|
68
100
|
if (this.vectorEnabled)
|
|
69
101
|
this.vectorDims = await this.readVectorDims();
|
|
102
|
+
await this.writeMeta();
|
|
103
|
+
}
|
|
104
|
+
async readMeta() {
|
|
105
|
+
try {
|
|
106
|
+
const res = await this.pool.query('SELECT version, vector_enabled, vector_dims FROM claude_db_meta WHERE id = 1');
|
|
107
|
+
const row = res.rows[0];
|
|
108
|
+
if (!row)
|
|
109
|
+
return null;
|
|
110
|
+
return {
|
|
111
|
+
version: Number(row['version']),
|
|
112
|
+
vectorEnabled: row['vector_enabled'] === true,
|
|
113
|
+
vectorDims: row['vector_dims'] == null ? null : Number(row['vector_dims']),
|
|
114
|
+
};
|
|
115
|
+
}
|
|
116
|
+
catch {
|
|
117
|
+
// No table yet, or no permission to read it. Either way the caller does
|
|
118
|
+
// the full setup, which is what happened before this existed.
|
|
119
|
+
return null;
|
|
120
|
+
}
|
|
121
|
+
}
|
|
122
|
+
async writeMeta() {
|
|
123
|
+
await this.pool.query(`INSERT INTO claude_db_meta (id, version, vector_enabled, vector_dims)
|
|
124
|
+
VALUES (1, $1, $2, $3)
|
|
125
|
+
ON CONFLICT (id) DO UPDATE SET
|
|
126
|
+
version = EXCLUDED.version,
|
|
127
|
+
vector_enabled = EXCLUDED.vector_enabled,
|
|
128
|
+
vector_dims = EXCLUDED.vector_dims`, [SCHEMA_VERSION, this.vectorEnabled, this.vectorDims]);
|
|
70
129
|
}
|
|
71
130
|
/**
|
|
72
131
|
* CREATE TABLE IF NOT EXISTS leaves an existing generated column alone, so a
|
|
@@ -111,7 +170,13 @@ export class PostgresStore {
|
|
|
111
170
|
await this.pool.query(`ALTER TABLE observations ADD COLUMN IF NOT EXISTS embedding vector(${dims})`);
|
|
112
171
|
await this.pool.query(`CREATE INDEX IF NOT EXISTS idx_obs_embedding
|
|
113
172
|
ON observations USING hnsw (embedding vector_cosine_ops)`);
|
|
114
|
-
|
|
173
|
+
// Read back rather than assumed: another process may have created the
|
|
174
|
+
// column at a different width between this one reading the version row
|
|
175
|
+
// and reaching here, and writing 384d values into a 256d column rolls
|
|
176
|
+
// the whole batch back. One extra round trip, once, and only before the
|
|
177
|
+
// first vector this process stores.
|
|
178
|
+
this.vectorDims = await this.readVectorDims();
|
|
179
|
+
await this.writeMeta();
|
|
115
180
|
}
|
|
116
181
|
if (this.vectorDims !== dims && !this.warnedDims) {
|
|
117
182
|
this.warnedDims = true;
|
|
@@ -290,6 +355,11 @@ export class PostgresStore {
|
|
|
290
355
|
lastActive: Number(row['last']),
|
|
291
356
|
}));
|
|
292
357
|
}
|
|
358
|
+
async inventory() {
|
|
359
|
+
const res = await this.pool.query(`SELECT tablename AS name FROM pg_tables
|
|
360
|
+
WHERE schemaname = ANY(current_schemas(false))`);
|
|
361
|
+
return foreignNames(res.rows.map((row) => String(row['name'] ?? '')));
|
|
362
|
+
}
|
|
293
363
|
async searchKeyword(query) {
|
|
294
364
|
const conditions = ['tsv @@ q.tsq'];
|
|
295
365
|
const values = [query.text];
|
|
@@ -304,7 +374,10 @@ export class PostgresStore {
|
|
|
304
374
|
const res = await this.pool.query(`WITH q AS (
|
|
305
375
|
SELECT replace(plainto_tsquery('english', $1)::text, '&', '|')::tsquery AS tsq
|
|
306
376
|
)
|
|
307
|
-
SELECT id, kind, title, project, created_at,
|
|
377
|
+
SELECT id, kind, title, project, created_at,
|
|
378
|
+
ts_headline('english', body, q.tsq,
|
|
379
|
+
'MaxFragments=1, MaxWords=18, MinWords=5') AS snippet,
|
|
380
|
+
ts_rank(tsv, q.tsq) AS score
|
|
308
381
|
FROM observations, q
|
|
309
382
|
WHERE ${conditions.join(' AND ')}
|
|
310
383
|
ORDER BY score DESC
|
|
@@ -356,6 +429,11 @@ export class PostgresStore {
|
|
|
356
429
|
values.push(query.kind);
|
|
357
430
|
conditions.push(`kind = $${values.length}`);
|
|
358
431
|
}
|
|
432
|
+
if (query.tag) {
|
|
433
|
+
// Containment rather than a text match, so `api` cannot match `api-docs`.
|
|
434
|
+
values.push(JSON.stringify([query.tag]));
|
|
435
|
+
conditions.push(`tags @> $${values.length}::jsonb`);
|
|
436
|
+
}
|
|
359
437
|
if (query.since !== undefined) {
|
|
360
438
|
values.push(query.since);
|
|
361
439
|
conditions.push(`created_at >= $${values.length}`);
|
|
@@ -422,7 +500,7 @@ function toObservation(row) {
|
|
|
422
500
|
return obs;
|
|
423
501
|
}
|
|
424
502
|
function toIndexEntry(row) {
|
|
425
|
-
|
|
503
|
+
const entry = {
|
|
426
504
|
id: row['id'],
|
|
427
505
|
kind: row['kind'],
|
|
428
506
|
title: row['title'],
|
|
@@ -430,4 +508,11 @@ function toIndexEntry(row) {
|
|
|
430
508
|
createdAt: Number(row['created_at']),
|
|
431
509
|
score: Number(row['score'] ?? 0),
|
|
432
510
|
};
|
|
511
|
+
// ts_headline wraps matches in <b> by default. Left at the default rather
|
|
512
|
+
// than configured away, because an empty StartSel is not portable across
|
|
513
|
+
// server versions; toSnippet strips them.
|
|
514
|
+
const snippet = toSnippet(row['snippet']);
|
|
515
|
+
if (snippet)
|
|
516
|
+
entry.snippet = snippet;
|
|
517
|
+
return entry;
|
|
433
518
|
}
|
|
@@ -25,6 +25,7 @@ export declare class SqliteStore implements MemoryStore {
|
|
|
25
25
|
remove(filter: RemoveFilter): Promise<number>;
|
|
26
26
|
list(filter: ListFilter): Promise<Observation[]>;
|
|
27
27
|
listProjects(): Promise<ProjectSummary[]>;
|
|
28
|
+
inventory(): Promise<string[]>;
|
|
28
29
|
searchKeyword(query: SearchQuery): Promise<ObservationIndexEntry[]>;
|
|
29
30
|
searchVector(vector: number[], query: SearchQuery): Promise<ObservationIndexEntry[]>;
|
|
30
31
|
timeline(query: TimelineQuery): Promise<ObservationIndexEntry[]>;
|