Compare commits

..

24 Commits
1.2.4 ... 1.4.2

Author SHA1 Message Date
077f75cdd9 Fixed delegate agent history... again
All checks were successful
Publish Library / Build NPM Project (push) Successful in 48s
Publish Library / Tag Version (push) Successful in 13s
2026-08-04 13:58:47 -04:00
566d84fd7a Added memory graph traversal helpers
All checks were successful
Publish Library / Build NPM Project (push) Successful in 43s
Publish Library / Tag Version (push) Successful in 14s
2026-08-04 12:58:39 -04:00
4230b534fc bump 1.4.0
All checks were successful
Publish Library / Build NPM Project (push) Successful in 1m17s
Publish Library / Tag Version (push) Successful in 14s
2026-08-04 12:44:41 -04:00
119f8472f2 token pools
Some checks failed
Publish Library / Tag Version (push) Has been cancelled
Publish Library / Build NPM Project (push) Has been cancelled
2026-08-04 12:44:21 -04:00
9c04e58c63 Pass deligate subagents full history, improved memory managment 2026-08-04 12:24:23 -04:00
7fbb42c26a improved subagent instructions 2026-08-04 12:03:31 -04:00
be08db8e2c Attach tps to response promise
All checks were successful
Publish Library / Build NPM Project (push) Successful in 42s
Publish Library / Tag Version (push) Successful in 9s
2026-08-04 09:48:20 -04:00
497f051c62 bump 1.3.5
All checks were successful
Publish Library / Build NPM Project (push) Successful in 52s
Publish Library / Tag Version (push) Successful in 7s
2026-08-04 09:30:45 -04:00
62fbe73b22 Added tps + duration to AI history
All checks were successful
Publish Library / Build NPM Project (push) Successful in 52s
Publish Library / Tag Version (push) Successful in 11s
2026-08-04 09:26:57 -04:00
d53b1c6328 Removed <tool> blocks from responses
All checks were successful
Publish Library / Build NPM Project (push) Successful in 39s
Publish Library / Tag Version (push) Successful in 11s
2026-08-03 20:23:22 -04:00
89619e211e Fixed message history and response
All checks were successful
Publish Library / Build NPM Project (push) Successful in 59s
Publish Library / Tag Version (push) Successful in 22s
2026-08-03 19:30:39 -04:00
afc6653364 fixed openai system calls in history breaking anthropic calls
All checks were successful
Publish Library / Build NPM Project (push) Successful in 55s
Publish Library / Tag Version (push) Successful in 21s
2026-08-02 22:35:17 -04:00
68e72445a2 Keep recent memories in context
All checks were successful
Publish Library / Build NPM Project (push) Successful in 53s
Publish Library / Tag Version (push) Successful in 17s
2026-08-01 21:42:05 -04:00
1aa6cdf329 Agent/subagent support
All checks were successful
Publish Library / Build NPM Project (push) Successful in 45s
Publish Library / Tag Version (push) Successful in 15s
2026-08-01 18:28:16 -04:00
d022a5ef4d Improved levenshtein fuzzy match
All checks were successful
Publish Library / Build NPM Project (push) Successful in 52s
Publish Library / Tag Version (push) Successful in 17s
2026-08-01 12:00:26 -04:00
a1d438a20a Tools can now emit "done" event and end chat early gracefully
All checks were successful
Publish Library / Build NPM Project (push) Successful in 1m0s
Publish Library / Tag Version (push) Successful in 9s
2026-07-31 17:49:06 -04:00
52a9e3aaa4 Fixed history poisoning on empty tool response
All checks were successful
Publish Library / Build NPM Project (push) Successful in 51s
Publish Library / Tag Version (push) Successful in 13s
2026-07-30 22:12:49 -04:00
a7aec4ee29 Improved memory prompt slightly
All checks were successful
Publish Library / Build NPM Project (push) Successful in 1m9s
Publish Library / Tag Version (push) Successful in 19s
2026-07-30 16:00:03 -04:00
dda2d4c2a3 Bump 1.2.8
All checks were successful
Publish Library / Build NPM Project (push) Successful in 49s
Publish Library / Tag Version (push) Successful in 7s
2026-07-29 22:35:29 -04:00
58e0e488e4 Added Geo, FS and flarescraperr tools
Some checks failed
Publish Library / Tag Version (push) Has been cancelled
Publish Library / Build NPM Project (push) Has been cancelled
2026-07-29 22:34:51 -04:00
8dfcd06752 More memory fixes
All checks were successful
Publish Library / Build NPM Project (push) Successful in 43s
Publish Library / Tag Version (push) Successful in 14s
2026-07-29 22:11:09 -04:00
14f6cdd313 Personal file memory organization instructions
All checks were successful
Publish Library / Build NPM Project (push) Successful in 33s
Publish Library / Tag Version (push) Successful in 12s
2026-07-27 22:47:48 -04:00
73d6ee0f2a Personal file memory organization instructions
All checks were successful
Publish Library / Build NPM Project (push) Successful in 45s
Publish Library / Tag Version (push) Successful in 12s
2026-07-27 22:39:06 -04:00
bee4085666 updatememory awaits full result
Some checks failed
Publish Library / Tag Version (push) Has been cancelled
Publish Library / Build NPM Project (push) Has been cancelled
2026-07-27 22:34:36 -04:00
13 changed files with 1965 additions and 918 deletions

234
package-lock.json generated
View File

@@ -1,12 +1,12 @@
{
"name": "@ztimson/ai-utils",
"version": "1.0.6",
"version": "1.2.6",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "@ztimson/ai-utils",
"version": "1.0.6",
"version": "1.2.6",
"license": "MIT",
"dependencies": {
"@anthropic-ai/sdk": "^0.102.0",
@@ -57,34 +57,38 @@
}
},
"node_modules/@emnapi/core": {
"version": "1.11.1",
"resolved": "https://registry.npmjs.org/@emnapi/core/-/core-1.11.1.tgz",
"integrity": "sha512-RSvbQmHzdKzNsLYa/wHrbc3KN4sYLKAdPZxqiM2HATqv/SBk2/ENSHpvXGaLOMcsAyz0poEGqkmmKYG3OWiJEQ==",
"version": "2.0.0-alpha.3",
"resolved": "https://registry.npmjs.org/@emnapi/core/-/core-2.0.0-alpha.3.tgz",
"integrity": "sha512-AZypUeJ/yByuxyS7BlSNRDOMLMlROYtjYdIAuBmJssVz1UJDSeYxLrdizhXCFYhedC5bqd/ASy8EuNXbVVXp9g==",
"dev": true,
"license": "MIT",
"optional": true,
"peer": true,
"dependencies": {
"@emnapi/wasi-threads": "1.2.2",
"@emnapi/wasi-threads": "2.0.1",
"tslib": "^2.4.0"
}
},
"node_modules/@emnapi/runtime": {
"version": "1.11.2",
"resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-1.11.2.tgz",
"integrity": "sha512-kyOl3X0DuTiT1h2ft8r2fYO8JYtU9a9Xis/zBSiGArNaagCOWx90N1k2wxp18czFDH+OgcWGb5ZP/XMt3dcyPA==",
"version": "2.0.0-alpha.3",
"resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-2.0.0-alpha.3.tgz",
"integrity": "sha512-hFPAhMUjJD9BSyCANEISPOogeXC9Zo9ZQl7L6vKnaVsMkCtzznaW/naYypeyl0Gv5rYfWYsZbpixTMpjDJzQeA==",
"dev": true,
"license": "MIT",
"optional": true,
"peer": true,
"dependencies": {
"tslib": "^2.4.0"
}
},
"node_modules/@emnapi/wasi-threads": {
"version": "1.2.2",
"resolved": "https://registry.npmjs.org/@emnapi/wasi-threads/-/wasi-threads-1.2.2.tgz",
"integrity": "sha512-c95qOXkHdydNKhscBTebqEC1CVAZpyqOfVfBzQ1qgzyl3gfeldUjIggDbIZgDKsHLgnsM+igH7TJ/eAasaVuMA==",
"version": "2.0.1",
"resolved": "https://registry.npmjs.org/@emnapi/wasi-threads/-/wasi-threads-2.0.1.tgz",
"integrity": "sha512-9DsSk+o5NBX0CCJT8s0EROGSGxjR/tKu6aBTaVyq+SjAEQH4XcdcRxPBRzsBLizTTJ49MJjF+jgu3qnO9GLQcQ==",
"dev": true,
"license": "MIT",
"optional": true,
"peer": true,
"dependencies": {
"tslib": "^2.4.0"
}
@@ -573,6 +577,16 @@
"url": "https://opencollective.com/libvips"
}
},
"node_modules/@img/sharp-wasm32/node_modules/@emnapi/runtime": {
"version": "1.11.3",
"resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-1.11.3.tgz",
"integrity": "sha512-Xz4Tpyki7XyrpbUK1jR1AhdAdaXyhhY4lZ3neLodmhpuWfy2PAQN5B46sAiU4liOXGLkHypn/qU+jvfWSCYYLA==",
"license": "MIT",
"optional": true,
"dependencies": {
"tslib": "^2.4.0"
}
},
"node_modules/@img/sharp-win32-arm64": {
"version": "0.34.5",
"resolved": "https://registry.npmjs.org/@img/sharp-win32-arm64/-/sharp-win32-arm64-0.34.5.tgz",
@@ -681,22 +695,25 @@
}
},
"node_modules/@napi-rs/wasm-runtime": {
"version": "1.1.6",
"resolved": "https://registry.npmjs.org/@napi-rs/wasm-runtime/-/wasm-runtime-1.1.6.tgz",
"integrity": "sha512-ZLv/JdUfkvOy9eCnnBaGfiO+XimbjebAeO+MRQqD/B+FR1tnRN0tpKSJHRbE8sFfS6aqsXZ67TQjfwfsxULVbg==",
"version": "1.2.0",
"resolved": "https://registry.npmjs.org/@napi-rs/wasm-runtime/-/wasm-runtime-1.2.0.tgz",
"integrity": "sha512-kDoONqMa+VnZ4vvvu/ZUurpJ4gkZU57e7g69qpNgWhYcZFPUHZM2CEMKm+cG6ufDVALbjMvfmMjFVqaK7uEMnA==",
"dev": true,
"license": "MIT",
"optional": true,
"dependencies": {
"@tybys/wasm-util": "^0.10.3"
},
"engines": {
"node": "^20.19.0 || ^22.13.0 || >=23.5.0"
},
"funding": {
"type": "github",
"url": "https://github.com/sponsors/Brooooooklyn"
},
"peerDependencies": {
"@emnapi/core": "^1.7.1",
"@emnapi/runtime": "^1.7.1"
"@emnapi/core": "^2.0.0-alpha.3",
"@emnapi/runtime": "^2.0.0-alpha.3"
}
},
"node_modules/@oxc-project/types": {
@@ -1007,6 +1024,18 @@
"node": "^20.19.0 || >=22.12.0"
}
},
"node_modules/@rolldown/binding-wasm32-wasi/node_modules/@emnapi/core": {
"version": "1.11.1",
"resolved": "https://registry.npmjs.org/@emnapi/core/-/core-1.11.1.tgz",
"integrity": "sha512-RSvbQmHzdKzNsLYa/wHrbc3KN4sYLKAdPZxqiM2HATqv/SBk2/ENSHpvXGaLOMcsAyz0poEGqkmmKYG3OWiJEQ==",
"dev": true,
"license": "MIT",
"optional": true,
"dependencies": {
"@emnapi/wasi-threads": "1.2.2",
"tslib": "^2.4.0"
}
},
"node_modules/@rolldown/binding-wasm32-wasi/node_modules/@emnapi/runtime": {
"version": "1.11.1",
"resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-1.11.1.tgz",
@@ -1018,6 +1047,17 @@
"tslib": "^2.4.0"
}
},
"node_modules/@rolldown/binding-wasm32-wasi/node_modules/@emnapi/wasi-threads": {
"version": "1.2.2",
"resolved": "https://registry.npmjs.org/@emnapi/wasi-threads/-/wasi-threads-1.2.2.tgz",
"integrity": "sha512-c95qOXkHdydNKhscBTebqEC1CVAZpyqOfVfBzQ1qgzyl3gfeldUjIggDbIZgDKsHLgnsM+igH7TJ/eAasaVuMA==",
"dev": true,
"license": "MIT",
"optional": true,
"dependencies": {
"tslib": "^2.4.0"
}
},
"node_modules/@rolldown/binding-win32-arm64-msvc": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-win32-arm64-msvc/-/binding-win32-arm64-msvc-1.1.5.tgz",
@@ -1408,18 +1448,18 @@
"license": "MIT"
},
"node_modules/@ztimson/utils": {
"version": "0.29.5",
"resolved": "https://registry.npmjs.org/@ztimson/utils/-/utils-0.29.5.tgz",
"integrity": "sha512-8mUuhi//3agwrueR006emOvJu1JXxVEryJmD3nkEmK4yQ9qS24oZijdoaiLRid/6xc75j3Fk6YjmnYmjq61iPQ==",
"version": "0.29.7",
"resolved": "https://registry.npmjs.org/@ztimson/utils/-/utils-0.29.7.tgz",
"integrity": "sha512-cjQ9+RjC5X7gKNA/hJHDf7OtyYCa+5E0PDc76lIaATwNAxXCSx2IO9r2wHiHtZGV5bldjnFmw7aOV8Jmq7SgKQ==",
"license": "MIT",
"dependencies": {
"var-persist": "^1.0.1"
}
},
"node_modules/acorn": {
"version": "8.17.0",
"resolved": "https://registry.npmjs.org/acorn/-/acorn-8.17.0.tgz",
"integrity": "sha512-xRQbDb9BnwDafYNn6Vwl839DYVjqXYb1XVGtWAZ1kcDc6iwAL4hg3B1dZlRiuENFeO2H53gFG3in621AdERVAg==",
"version": "8.18.0",
"resolved": "https://registry.npmjs.org/acorn/-/acorn-8.18.0.tgz",
"integrity": "sha512-lGq+9yr1/GuAWaVYIHRjvvySG5/4VfKIvC8EWxStPdcDh/Ka7FG3twP6v4d5BkravUilhIAsG4Qj83t02LWUPQ==",
"dev": true,
"license": "MIT",
"bin": {
@@ -1504,9 +1544,9 @@
"license": "MIT"
},
"node_modules/brace-expansion": {
"version": "2.1.2",
"resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-2.1.2.tgz",
"integrity": "sha512-w5JZcKgdhDOgOwm8H+KgbosopHMuGcl6qbulwjtz3SM7I7P3yW1eAjzMPLrIE+NQ9vjgANKHWeMHnrT0OXW1oA==",
"version": "2.1.3",
"resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-2.1.3.tgz",
"integrity": "sha512-DRdx5neNsG/QXbniLFWi2YmC/68oeOOmKz6zOjVk6ZS1ZLXgLIKqVEc6hWsmkjBbgii0SwaBTcJ5XKj5gzY/4A==",
"dev": true,
"license": "MIT",
"dependencies": {
@@ -2009,9 +2049,9 @@
"license": "MIT"
},
"node_modules/exsolve": {
"version": "1.1.0",
"resolved": "https://registry.npmjs.org/exsolve/-/exsolve-1.1.0.tgz",
"integrity": "sha512-D+42+T12DdIlJM3uepa55qGiL3sYdLBOxIl2ifQCzCHz4c7eiolaHsi3BIqEr7JxBzxv2pYZQX9kw16ziMcEmw==",
"version": "1.1.1",
"resolved": "https://registry.npmjs.org/exsolve/-/exsolve-1.1.1.tgz",
"integrity": "sha512-9U/jZUgjnSGyntRr6y5Muu1MJcwFl6kPu7k8qLF0IMNfLqvw0NZ4nnVDq0RVoZ0RvCyumib4Ez3KYrVfilrw+g==",
"dev": true,
"license": "MIT"
},
@@ -2382,9 +2422,9 @@
"license": "MIT"
},
"node_modules/lightningcss": {
"version": "1.32.0",
"resolved": "https://registry.npmjs.org/lightningcss/-/lightningcss-1.32.0.tgz",
"integrity": "sha512-NXYBzinNrblfraPGyrbPoD19C1h9lfI/1mzgWYvXUTe414Gz/X1FD2XBZSZM7rRTrMA8JL3OtAaGifrIKhQ5yQ==",
"version": "1.33.0",
"resolved": "https://registry.npmjs.org/lightningcss/-/lightningcss-1.33.0.tgz",
"integrity": "sha512-WkUDrojuJs0xkgGf2udWxa3yGBRxPtxUkB79i6aCZLRgc7PM8fZe9TosfPDcvEpQZbuFASnHYmRLBLUbmLOIIA==",
"dev": true,
"license": "MPL-2.0",
"dependencies": {
@@ -2398,23 +2438,23 @@
"url": "https://opencollective.com/parcel"
},
"optionalDependencies": {
"lightningcss-android-arm64": "1.32.0",
"lightningcss-darwin-arm64": "1.32.0",
"lightningcss-darwin-x64": "1.32.0",
"lightningcss-freebsd-x64": "1.32.0",
"lightningcss-linux-arm-gnueabihf": "1.32.0",
"lightningcss-linux-arm64-gnu": "1.32.0",
"lightningcss-linux-arm64-musl": "1.32.0",
"lightningcss-linux-x64-gnu": "1.32.0",
"lightningcss-linux-x64-musl": "1.32.0",
"lightningcss-win32-arm64-msvc": "1.32.0",
"lightningcss-win32-x64-msvc": "1.32.0"
"lightningcss-android-arm64": "1.33.0",
"lightningcss-darwin-arm64": "1.33.0",
"lightningcss-darwin-x64": "1.33.0",
"lightningcss-freebsd-x64": "1.33.0",
"lightningcss-linux-arm-gnueabihf": "1.33.0",
"lightningcss-linux-arm64-gnu": "1.33.0",
"lightningcss-linux-arm64-musl": "1.33.0",
"lightningcss-linux-x64-gnu": "1.33.0",
"lightningcss-linux-x64-musl": "1.33.0",
"lightningcss-win32-arm64-msvc": "1.33.0",
"lightningcss-win32-x64-msvc": "1.33.0"
}
},
"node_modules/lightningcss-android-arm64": {
"version": "1.32.0",
"resolved": "https://registry.npmjs.org/lightningcss-android-arm64/-/lightningcss-android-arm64-1.32.0.tgz",
"integrity": "sha512-YK7/ClTt4kAK0vo6w3X+Pnm0D2cf2vPHbhOXdoNti1Ga0al1P4TBZhwjATvjNwLEBCnKvjJc2jQgHXH0NEwlAg==",
"version": "1.33.0",
"resolved": "https://registry.npmjs.org/lightningcss-android-arm64/-/lightningcss-android-arm64-1.33.0.tgz",
"integrity": "sha512-gEpRTalKdosp4Bb8qWtc2iOgE5SeIHlpS1up9bFq2wAyYhl1UdTObYiHe98zEM9SQvSoqQZ1IQD0JNpg3Ml5pg==",
"cpu": [
"arm64"
],
@@ -2433,9 +2473,9 @@
}
},
"node_modules/lightningcss-darwin-arm64": {
"version": "1.32.0",
"resolved": "https://registry.npmjs.org/lightningcss-darwin-arm64/-/lightningcss-darwin-arm64-1.32.0.tgz",
"integrity": "sha512-RzeG9Ju5bag2Bv1/lwlVJvBE3q6TtXskdZLLCyfg5pt+HLz9BqlICO7LZM7VHNTTn/5PRhHFBSjk5lc4cmscPQ==",
"version": "1.33.0",
"resolved": "https://registry.npmjs.org/lightningcss-darwin-arm64/-/lightningcss-darwin-arm64-1.33.0.tgz",
"integrity": "sha512-Sciaz8eenNTKn9b3t7+xr0ipTp9YxKQY4npwQ3mrRuL0BAVHBLyZxofhaKBAVtzmtRZ/zTyo0/to4B1uWG/Djg==",
"cpu": [
"arm64"
],
@@ -2454,9 +2494,9 @@
}
},
"node_modules/lightningcss-darwin-x64": {
"version": "1.32.0",
"resolved": "https://registry.npmjs.org/lightningcss-darwin-x64/-/lightningcss-darwin-x64-1.32.0.tgz",
"integrity": "sha512-U+QsBp2m/s2wqpUYT/6wnlagdZbtZdndSmut/NJqlCcMLTWp5muCrID+K5UJ6jqD2BFshejCYXniPDbNh73V8w==",
"version": "1.33.0",
"resolved": "https://registry.npmjs.org/lightningcss-darwin-x64/-/lightningcss-darwin-x64-1.33.0.tgz",
"integrity": "sha512-Z5UPAxzrjlWNNyGy6i65cJzzvgJ5D3T6wMvs+gWpY9d7qRhANrxqAp6LhxIgZhWEw18RfJTGcRxjuLIBr+m8XQ==",
"cpu": [
"x64"
],
@@ -2475,9 +2515,9 @@
}
},
"node_modules/lightningcss-freebsd-x64": {
"version": "1.32.0",
"resolved": "https://registry.npmjs.org/lightningcss-freebsd-x64/-/lightningcss-freebsd-x64-1.32.0.tgz",
"integrity": "sha512-JCTigedEksZk3tHTTthnMdVfGf61Fky8Ji2E4YjUTEQX14xiy/lTzXnu1vwiZe3bYe0q+SpsSH/CTeDXK6WHig==",
"version": "1.33.0",
"resolved": "https://registry.npmjs.org/lightningcss-freebsd-x64/-/lightningcss-freebsd-x64-1.33.0.tgz",
"integrity": "sha512-QQM/Ti/hQajJwCY+RiWuCZ9sdtI/XQk7nDK5vC8kkdwixezOlDgvDx7+RT+QjK6FcFT4MpsuoBnHIo/O3StRRg==",
"cpu": [
"x64"
],
@@ -2496,9 +2536,9 @@
}
},
"node_modules/lightningcss-linux-arm-gnueabihf": {
"version": "1.32.0",
"resolved": "https://registry.npmjs.org/lightningcss-linux-arm-gnueabihf/-/lightningcss-linux-arm-gnueabihf-1.32.0.tgz",
"integrity": "sha512-x6rnnpRa2GL0zQOkt6rts3YDPzduLpWvwAF6EMhXFVZXD4tPrBkEFqzGowzCsIWsPjqSK+tyNEODUBXeeVHSkw==",
"version": "1.33.0",
"resolved": "https://registry.npmjs.org/lightningcss-linux-arm-gnueabihf/-/lightningcss-linux-arm-gnueabihf-1.33.0.tgz",
"integrity": "sha512-N7FVBe6iS24MlM6R/4RBTxGhQheZGs7tiQ9U32UtF75NzP5Q7xWPRqLBCKxlRQRk3rY1jCIPLzx7WzOhuUIRLQ==",
"cpu": [
"arm"
],
@@ -2517,9 +2557,9 @@
}
},
"node_modules/lightningcss-linux-arm64-gnu": {
"version": "1.32.0",
"resolved": "https://registry.npmjs.org/lightningcss-linux-arm64-gnu/-/lightningcss-linux-arm64-gnu-1.32.0.tgz",
"integrity": "sha512-0nnMyoyOLRJXfbMOilaSRcLH3Jw5z9HDNGfT/gwCPgaDjnx0i8w7vBzFLFR1f6CMLKF8gVbebmkUN3fa/kQJpQ==",
"version": "1.33.0",
"resolved": "https://registry.npmjs.org/lightningcss-linux-arm64-gnu/-/lightningcss-linux-arm64-gnu-1.33.0.tgz",
"integrity": "sha512-j2v/itmy4HlNxlc6voKXYgBqNi0Ng2LShg4z7GufpEgs05P+2suBVyi9I6YHq5uoVFx9ETin3eCEhLVyXGQnKg==",
"cpu": [
"arm64"
],
@@ -2541,9 +2581,9 @@
}
},
"node_modules/lightningcss-linux-arm64-musl": {
"version": "1.32.0",
"resolved": "https://registry.npmjs.org/lightningcss-linux-arm64-musl/-/lightningcss-linux-arm64-musl-1.32.0.tgz",
"integrity": "sha512-UpQkoenr4UJEzgVIYpI80lDFvRmPVg6oqboNHfoH4CQIfNA+HOrZ7Mo7KZP02dC6LjghPQJeBsvXhJod/wnIBg==",
"version": "1.33.0",
"resolved": "https://registry.npmjs.org/lightningcss-linux-arm64-musl/-/lightningcss-linux-arm64-musl-1.33.0.tgz",
"integrity": "sha512-yiO5ROMuYQgXbC60yjZU5CYSFZGKXL0HFATXt9mHJn1+zW55oCtMI9NfcVhYLMFDL7gV7oBPon/EmMMGg2OvtQ==",
"cpu": [
"arm64"
],
@@ -2565,9 +2605,9 @@
}
},
"node_modules/lightningcss-linux-x64-gnu": {
"version": "1.32.0",
"resolved": "https://registry.npmjs.org/lightningcss-linux-x64-gnu/-/lightningcss-linux-x64-gnu-1.32.0.tgz",
"integrity": "sha512-V7Qr52IhZmdKPVr+Vtw8o+WLsQJYCTd8loIfpDaMRWGUZfBOYEJeyJIkqGIDMZPwPx24pUMfwSxxI8phr/MbOA==",
"version": "1.33.0",
"resolved": "https://registry.npmjs.org/lightningcss-linux-x64-gnu/-/lightningcss-linux-x64-gnu-1.33.0.tgz",
"integrity": "sha512-ar+Ju7LmcN0Jo4FpL4hpFybwNG9/3A/Br5KW2n2jyODg3MEZXaDYADdemoNS+BDNfMgKvylJLj4S5tyRActuAg==",
"cpu": [
"x64"
],
@@ -2589,9 +2629,9 @@
}
},
"node_modules/lightningcss-linux-x64-musl": {
"version": "1.32.0",
"resolved": "https://registry.npmjs.org/lightningcss-linux-x64-musl/-/lightningcss-linux-x64-musl-1.32.0.tgz",
"integrity": "sha512-bYcLp+Vb0awsiXg/80uCRezCYHNg1/l3mt0gzHnWV9XP1W5sKa5/TCdGWaR/zBM2PeF/HbsQv/j2URNOiVuxWg==",
"version": "1.33.0",
"resolved": "https://registry.npmjs.org/lightningcss-linux-x64-musl/-/lightningcss-linux-x64-musl-1.33.0.tgz",
"integrity": "sha512-RYiYbkokw0trfKqqzfF55lginwEPrD3OJDfTuJzFs1MK6iFnDenaz1fqLLtX4ITG3OktJQXOeTaw1awrBAlZPw==",
"cpu": [
"x64"
],
@@ -2613,9 +2653,9 @@
}
},
"node_modules/lightningcss-win32-arm64-msvc": {
"version": "1.32.0",
"resolved": "https://registry.npmjs.org/lightningcss-win32-arm64-msvc/-/lightningcss-win32-arm64-msvc-1.32.0.tgz",
"integrity": "sha512-8SbC8BR40pS6baCM8sbtYDSwEVQd4JlFTOlaD3gWGHfThTcABnNDBda6eTZeqbofalIJhFx0qKzgHJmcPTnGdw==",
"version": "1.33.0",
"resolved": "https://registry.npmjs.org/lightningcss-win32-arm64-msvc/-/lightningcss-win32-arm64-msvc-1.33.0.tgz",
"integrity": "sha512-1K+MPfLSFVpphzpdbfkhlWk6wBrTObBzS2T6db10PNOZgR9GoVsAWzwNyuhUYYbTp23j+4RrncfujZ4uAzXvwA==",
"cpu": [
"arm64"
],
@@ -2634,9 +2674,9 @@
}
},
"node_modules/lightningcss-win32-x64-msvc": {
"version": "1.32.0",
"resolved": "https://registry.npmjs.org/lightningcss-win32-x64-msvc/-/lightningcss-win32-x64-msvc-1.32.0.tgz",
"integrity": "sha512-Amq9B/SoZYdDi1kFrojnoqPLxYhQ4Wo5XiL8EVJrVsB8ARoC1PWW6VGtT0WKCemjy8aC+louJnjS7U18x3b06Q==",
"version": "1.33.0",
"resolved": "https://registry.npmjs.org/lightningcss-win32-x64-msvc/-/lightningcss-win32-x64-msvc-1.33.0.tgz",
"integrity": "sha512-OlEICDx/Xl0FqSp4bry8zFnCvGpig3Gl4gCquvYwHuqJKEC1+n9NgDniFvqHGmMv1ZkqDJrDqKKSykTDX+ehuA==",
"cpu": [
"x64"
],
@@ -2794,9 +2834,9 @@
}
},
"node_modules/mdurl": {
"version": "2.0.0",
"resolved": "https://registry.npmjs.org/mdurl/-/mdurl-2.0.0.tgz",
"integrity": "sha512-Lf+9+2r+Tdp5wXDXC4PcIBjTDtq4UKjCPMQhKIuzpJNW0b96kVqSwW0bT7FhRSfmAiFYgP+SCRvdrDozfh0U5w==",
"version": "2.1.0",
"resolved": "https://registry.npmjs.org/mdurl/-/mdurl-2.1.0.tgz",
"integrity": "sha512-1+HBaOx0zi/dQWht8rNv9MYf9qqpqL/kxI0hXImU6Y547zM6Sni8BQibt7ifgMcYtQg41ao3Ivd6cnSM86inpg==",
"dev": true,
"license": "MIT"
},
@@ -2971,9 +3011,9 @@
"license": "MIT"
},
"node_modules/nanoid": {
"version": "3.3.15",
"resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.15.tgz",
"integrity": "sha512-y7Wygv/7mEOvxTuEQDB8StXdMRBWf1kR/tlhAzBRUFkB2jfcLOAxO/SHmOO2zgz1pVgK29/kyupn059/bCHdjA==",
"version": "3.3.16",
"resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.16.tgz",
"integrity": "sha512-bzlKTyNJ7+LdGIIwy8ijFpIqEQIvafahV7eYykJ8Cvh42EdJeODoJ6gUJXpQJvej1BddH8OqTXZNE/KfbWAu8Q==",
"dev": true,
"funding": [
{
@@ -3092,9 +3132,9 @@
"license": "MIT"
},
"node_modules/openai": {
"version": "6.46.0",
"resolved": "https://registry.npmjs.org/openai/-/openai-6.46.0.tgz",
"integrity": "sha512-DFg6jEPT2RO+oAyXtddeUJU8zkGy1OQ1AjGzNIJUMQG03TTqvCpy9tBpQ+2VVVnvrl3E56F8GEin2JYtWpITtA==",
"version": "6.49.0",
"resolved": "https://registry.npmjs.org/openai/-/openai-6.49.0.tgz",
"integrity": "sha512-aYCc0C6L864eR6WSYIwQGyXriw/nIyZx0ObvhzOEVuk0zoBDpynjSbrionWI7q65B5H8jJX0DXR9snEzM6bfPg==",
"license": "Apache-2.0",
"peerDependencies": {
"@aws-sdk/credential-provider-node": ">=3.972.0 <4",
@@ -3232,9 +3272,9 @@
"license": "MIT"
},
"node_modules/postcss": {
"version": "8.5.17",
"resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.17.tgz",
"integrity": "sha512-J7EF+8X+CzRPaJPOv9Ck2wNWJvGnnl3PcNPAdGg6GTLjyVpyQ0yATMSXRFRV01BviT/9Gwuc3rjEyJbDJG9a4w==",
"version": "8.5.25",
"resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.25.tgz",
"integrity": "sha512-DTPx3RWSSnWyzLxQnlH0rJP+EW5ekl16ZU4/psbIhA0e53kJfdgaN5vKM+xP7yJtXVu+nfdVFmlgFDEKAe4Pyw==",
"dev": true,
"funding": [
{
@@ -3252,7 +3292,7 @@
],
"license": "MIT",
"dependencies": {
"nanoid": "^3.3.12",
"nanoid": "^3.3.16",
"picocolors": "^1.1.1",
"source-map-js": "^1.2.1"
},
@@ -3787,9 +3827,9 @@
"license": "MIT"
},
"node_modules/undici": {
"version": "7.28.0",
"resolved": "https://registry.npmjs.org/undici/-/undici-7.28.0.tgz",
"integrity": "sha512-cRZYrTDwWznlnRiPjggAGxZXanty6M8RV1ff8Wm4LWXBp7/IG8v5DnOm74DtUBp9OONpK75YlPnIjQqX0dBDtA==",
"version": "7.29.0",
"resolved": "https://registry.npmjs.org/undici/-/undici-7.29.0.tgz",
"integrity": "sha512-IDxfleLmmbSskfWSUATiN1nfn2rDuvnMOqb5CWR92iIfojA0Ud+ulOAAEQ57LPr9rWmsreUyf5lwyao+7GNNVw==",
"license": "MIT",
"engines": {
"node": ">=20.18.1"
@@ -3981,16 +4021,16 @@
}
},
"node_modules/vite": {
"version": "8.1.4",
"resolved": "https://registry.npmjs.org/vite/-/vite-8.1.4.tgz",
"integrity": "sha512-bTT9PsdWO+MQMNG9ZXIP/qM9wGh37DFxTV/sPq9cFpHr3w4jkgef032PkAL9jAqhk3Nz8NQw3O8n6/xFkqO4QQ==",
"version": "8.1.5",
"resolved": "https://registry.npmjs.org/vite/-/vite-8.1.5.tgz",
"integrity": "sha512-7ULLwsCdYx/nRyrpiEwvqb5TFHrMVZyBt+rg/OAXT7rgj/z+DtTDyKFeLAdDkubDVDKD8jOsndmy7m55XcfUsw==",
"dev": true,
"license": "MIT",
"dependencies": {
"lightningcss": "^1.32.0",
"picomatch": "^4.0.5",
"postcss": "^8.5.16",
"rolldown": "~1.1.4",
"postcss": "^8.5.17",
"rolldown": "~1.1.5",
"tinyglobby": "^0.2.17"
},
"bin": {

View File

@@ -1,6 +1,6 @@
{
"name": "@ztimson/ai-utils",
"version": "1.2.4",
"version": "1.4.2",
"description": "AI Utility library",
"author": "Zak Timson",
"license": "MIT",

View File

@@ -1,58 +1,52 @@
import {Anthropic as anthropic} from '@anthropic-ai/sdk';
import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, makeArray} from '@ztimson/utils';
import {AbortablePromise, Ai} from './ai.ts';
import {LLMMessage, LLMRequest} from './llm.ts';
import {LLMProvider} from './provider.ts';
import {TokenPool} from './token-pool.ts';
import {convertSchema} from './tools.ts';
export class Anthropic extends LLMProvider {
client!: anthropic;
private clients = new Map<string, anthropic>();
tokenPool!: TokenPool;
constructor(public readonly ai: Ai, public readonly apiToken: string, public model: string) {
constructor(public readonly ai: Ai, public readonly apiToken: string | string[], public model: string) {
super();
this.client = new anthropic({apiKey: apiToken});
this.tokenPool = new TokenPool(...makeArray(apiToken).filter(Boolean));
}
private toStandard(history: any[]): LLMMessage[] {
const timestamp = Date.now();
const messages: LLMMessage[] = [];
for(let h of history) {
if(typeof h.content == 'string') {
messages.push(<any>{timestamp, ...h});
} else {
const textContent = h.content?.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
if(textContent) messages.push({timestamp, role: h.role, content: textContent});
h.content.forEach((c: any) => {
if(c.type == 'tool_use') {
messages.push({timestamp, role: 'tool', id: c.id, name: c.name, args: c.input, content: undefined});
} else if(c.type == 'tool_result') {
const m: any = messages.findLast(m => (<any>m).id == c.tool_use_id);
if(m) m[c.is_error ? 'error' : 'content'] = c.content;
private getClient(token: string): anthropic {
let client = this.clients.get(token);
if(!client) {
client = new anthropic({apiKey: token});
this.clients.set(token, client);
}
});
}
}
return messages;
return client;
}
private fromStandard(history: LLMMessage[]): any[] {
for(let i = 0; i < history.length; i++) {
if(history[i].role == 'tool') {
const h: any = history[i];
history.splice(i, 1,
/** Convert standard history -> Anthropic wire format */
private toWire(history: LLMMessage[]): any[] {
const wire: any[] = [];
for(const h of history) {
if(h.role === 'tool') {
wire.push(
{role: 'assistant', content: [{type: 'tool_use', id: h.id, name: h.name, input: h.args}]},
{role: 'user', content: [{type: 'tool_result', tool_use_id: h.id, is_error: !!h.error, content: h.error || h.content}]}
)
i++;
{role: 'user', content: [{type: 'tool_result', tool_use_id: h.id, is_error: !!h.error, content: h.error || h.content || ''}]}
);
} else {
wire.push({role: h.role, content: h.content});
}
}
return history.map(({timestamp, ...h}) => h);
return wire;
}
ask(message: string, options: LLMRequest = {}): AbortablePromise<string | any> {
const controller = new AbortController();
return Object.assign(new Promise<any>(async (res) => {
let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
return Object.assign(new Promise<any>(async (res, rej) => {
if(!options.history) options.history = [];
const history = options.history;
if(message) history.push({role: 'user', content: message, timestamp: Date.now()});
const tools = options.tools || this.ai.options.llm?.tools || [];
const requestParams: any = {
model: options.model || this.model,
@@ -66,90 +60,97 @@ export class Anthropic extends LLMProvider {
type: 'object',
properties: t.args ? objectMap(t.args, (key, value) => ({...value, required: undefined})) : {},
required: t.args ? Object.entries(t.args).filter(t => t[1].required).map(t => t[0]) : []
},
fn: undefined
}
})),
messages: history,
stream: !!options.stream,
};
// Add structured output support
if(options.schema) {
requestParams.output_config = {
format: {
type: 'json_schema',
schema: convertSchema(options.schema)
}
};
requestParams.output_config = {format: {type: 'json_schema', schema: convertSchema(options.schema)}};
}
let resp: any, isFirstMessage = true;
try {
let terminal = false;
do {
resp = await this.client.messages.create(requestParams).catch(err => {
err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
requestParams.messages = this.toWire(history.filter(h => h.role !== 'system'));
const callStart = Date.now();
const resp: any = await this.tokenPool.run(token => this.getClient(token).messages.create(requestParams)).catch(err => {
err.message += `\n\nMessages:\n${JSON.stringify(requestParams.messages, null, 2)}`;
throw err;
});
// Streaming mode
let usage: any, content: any[] = [];
if(options.stream) {
if(!isFirstMessage) options.stream({text: '\n\n'});
else isFirstMessage = false;
resp.content = [];
for await (const chunk of resp) {
if(controller.signal.aborted) break;
if(chunk.type === 'content_block_start') {
if(chunk.content_block.type === 'text') {
resp.content.push({type: 'text', text: ''});
} else if(chunk.content_block.type === 'tool_use') {
resp.content.push({type: 'tool_use', id: chunk.content_block.id, name: chunk.content_block.name, input: <any>''});
}
if(chunk.content_block.type === 'text') content.push({type: 'text', text: ''});
else if(chunk.content_block.type === 'tool_use') content.push({type: 'tool_use', id: chunk.content_block.id, name: chunk.content_block.name, input: ''});
} else if(chunk.type === 'content_block_delta') {
if(chunk.delta.type === 'text_delta') {
const text = chunk.delta.text;
resp.content.at(-1).text += text;
options.stream({text});
content.at(-1).text += chunk.delta.text;
options.stream({text: chunk.delta.text});
} else if(chunk.delta.type === 'input_json_delta') {
resp.content.at(-1).input += chunk.delta.partial_json;
content.at(-1).input += chunk.delta.partial_json;
}
} else if(chunk.type === 'content_block_stop') {
const last = resp.content.at(-1);
if(last.input != null) last.input = last.input ? JSONAttemptParse(last.input, {}) : {};
const last = content.at(-1);
if(last?.type === 'tool_use') last.input = last.input ? JSONAttemptParse(last.input, {}) : {};
} else if(chunk.type === 'message_delta') {
if(chunk.usage) usage = chunk.usage;
} else if(chunk.type === 'message_stop') {
break;
}
}
} else {
usage = resp.usage;
content = resp.content;
}
const duration = Date.now() - callStart;
const tps = usage?.output_tokens && duration > 0 ? usage.output_tokens / (duration / 1000) : 0;
// Run tools
const toolCalls = resp.content.filter((c: any) => c.type === 'tool_use');
const toolCalls = content.filter((c: any) => c.type === 'tool_use');
if(toolCalls.length && !controller.signal.aborted) {
history.push({role: 'assistant', content: resp.content});
const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
const tool = tools.find(findByProp('name', toolCall.name));
if(options.stream) options.stream({tool: toolCall.name});
if(!tool) return {tool_use_id: toolCall.id, is_error: true, content: 'Tool not found'};
const text = content.filter((c: any) => c.type === 'text').map((c: any) => c.text).join('\n\n').trim();
if(text) history.push({role: 'assistant', content: text, timestamp: Date.now(), duration, tps});
const entries = toolCalls.map((tc: any) => {
const entry: any = {role: 'tool', id: tc.id, name: tc.name, args: tc.input, content: undefined, timestamp: Date.now()};
history.push(entry);
return {tc, entry};
});
await Promise.all(entries.map(async ({tc, entry}: any) => {
const tool = tools.find(findByProp('name', tc.name));
if(options.stream) options.stream({tool: tc.name});
if(!tool) { entry.error = 'Tool not found'; return; }
try {
const result = await tool.fn(toolCall.input, options?.stream, this.ai);
return {type: 'tool_result', tool_use_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result};
const toolStream = options.stream && ((chunk: any) => {
if(chunk.done) { terminal = true; return; }
options.stream!(chunk);
});
const result = await tool.fn(entry.args, toolStream, this.ai, tc.id);
entry.content = typeof result === 'object' ? JSONSanitize(result) : result;
} catch(err: any) {
return {type: 'tool_result', tool_use_id: toolCall.id, is_error: true, content: err?.message || err?.toString() || 'Unknown'};
entry.error = err?.message || err?.toString() || 'Unknown';
}
}));
history.push({role: 'user', content: results});
requestParams.messages = history;
} else {
terminal = true;
const text = content.filter((c: any) => c.type === 'text').map((c: any) => c.text).join('\n\n').trim();
if(text) history.push({role: 'assistant', content: text, timestamp: Date.now(), duration, tps});
}
} while (!controller.signal.aborted && resp.content.some((c: any) => c.type === 'tool_use'));
const textContent = resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
history.push({role: 'assistant', content: textContent});
history = this.toStandard(history);
} while(!terminal && !controller.signal.aborted);
if(options.stream) options.stream({done: true});
if(options.history) options.history.splice(0, options.history.length, ...history);
// Return parsed JSON if schema provided
const finalContent = history.at(-1)?.content;
const turnStart = history.map(h => h.role).lastIndexOf('user');
const finalContent = history.slice(turnStart + 1).reduce((str, h) => h.role === 'assistant' ? str + (h.content || '') : str, '').trim();
res(options.schema ? JSONAttemptParse(finalContent, finalContent) : finalContent);
} catch(err) {
rej(err);
}
}), {abort: () => controller.abort()});
}
}

72
src/helpers.ts Normal file
View File

@@ -0,0 +1,72 @@
import {Memory, MemoryCache} from './memory.ts';
export type MemoryNode = {
name: string;
missing: boolean;
links: string[];
backlinks: string[];
}
export function buildMemoryGraph(memories: Memory[] | MemoryCache): MemoryNode[] {
const mems = memories instanceof MemoryCache ? memories.memories : memories;
const nameSet = new Set(mems.map(m => m.name));
const ghosts = new Set<string>();
const nodes: MemoryNode[] = mems.map(m => ({
name: m.name,
missing: false,
links: m.links,
backlinks: m.backlinks,
}));
for (const node of nodes) {
for (const link of node.links) {
if (!nameSet.has(link)) ghosts.add(link);
}
}
return [
...nodes,
...[...ghosts].map(name => ({
name,
missing: true,
links: [],
backlinks: nodes
.filter(n => n.links.includes(name))
.map(n => n.name),
}))
];
}
export function renderMemoryGraph(nodes) {
if (!nodes.length) return 'No memories yet.';
const groups = new Map();
for (const node of nodes) {
const [prefix, ...rest] = node.name.split('/');
const group = rest.length ? prefix : 'Root';
const label = rest.length ? rest.join('/') : node.name;
if (!groups.has(group)) groups.set(group, []);
groups.get(group).push({...node, label});
}
const ghostCount = nodes.filter(n => n.missing).length;
const lines = [`Memory Graph (${nodes.length} nodes, ${ghostCount} ghost${ghostCount === 1 ? '' : 's'})`, ''];
for (const group of [...groups.keys()].sort()) {
const items = groups.get(group).sort((a, b) => a.label.localeCompare(b.label));
lines.push(`${group}/`);
items.forEach((n, i) => {
const last = i === items.length - 1;
const branch = last ? '└─' : '├─';
const pad = last ? ' ' : '│ ';
const tag = n.missing ? ' (ghost)' : '';
lines.push(` ${branch} ${n.label}${tag}`);
if (n.links.length) lines.push(` ${pad}${n.links.join(', ')}`);
if (n.backlinks.length) lines.push(` ${pad}${n.backlinks.join(', ')}`);
});
lines.push('');
}
return lines.join('\n').trimEnd();
}

View File

@@ -1,9 +1,11 @@
export * from './ai';
export * from './antrhopic';
export * from './audio';
export * from './helpers';
export * from './llm';
export * from './memory';
export * from './open-ai';
export * from './provider';
export * from './token-pool'
export * from './tools';
export * from './vision';

View File

@@ -1,3 +1,4 @@
import {snakeCase} from '@ztimson/utils';
import {AbortablePromise, Ai} from './ai.ts';
import {Anthropic} from './antrhopic.ts';
import {OpenAi} from './open-ai.ts';
@@ -6,10 +7,25 @@ import {AiTool, AiToolArg} from './tools.ts';
import {fileURLToPath} from 'url';
import {dirname, join} from 'path';
import {spawn} from 'node:child_process';
import {Memory, MemoryCache, MemoryManager} from './memory.ts';
import {Memory, MemoryCache, MemoryManager, MemoryOptions} from './memory.ts';
export type AnthropicConfig = {proto: 'anthropic', token: string};
export type OpenAiConfig = {proto: 'openai', host?: string, token: string};
const MAX_AGENT_DEPTH = 5;
export type AnthropicConfig = {proto: 'anthropic', token: string | string[]};
export type OpenAiConfig = {proto: 'openai', host?: string, token: string | string[]};
export type Agent = {
name: string;
description?: string;
model?: string | null;
temperature?: number;
system: string;
delegate?: boolean;
skills?: Skill[] | null;
tools?: AiTool[] | null;
mcp?: McpServer[] | null;
agents?: string[] | null;
}
export type LLMMessage = {
/** Message originator */
@@ -18,6 +34,10 @@ export type LLMMessage = {
content: string | any;
/** Timestamp */
timestamp?: number;
/** Response duration in ms */
duration?: number;
/** Tokens per second */
tps?: number;
} | {
/** Tool call */
role: 'tool';
@@ -33,6 +53,10 @@ export type LLMMessage = {
error?: undefined | string;
/** Timestamp */
timestamp?: number;
/** Response duration in ms */
duration?: number;
/** Tokens per second */
tps?: number;
}
export type LLMRequest = {
@@ -55,13 +79,17 @@ export type LLMRequest = {
/** Compress old messages in the chat to free up context */
compress?: {max: number; min: number};
/** User's memory documents - RAG injected automatically each turn */
memory?: Memory[] | MemoryCache;
memory?: Memory[] | MemoryCache | MemoryOptions;
/** Model to use for memory operations */
memoryModel?: string;
/** Skill documents the AI can browse and read on demand */
skills?: Skill[];
/** MCP servers to connect and expose as tools */
mcp?: McpServer[];
/** Subagents exposed as delegatable/wrapped tools */
agents?: Agent[];
/** @internal recursion guard for nested agent delegation */
_agentDepth?: number;
}
export type McpServer = {
@@ -82,7 +110,6 @@ export type Skill = {
content: string;
}
class LLM {
private memoryManager!: MemoryManager;
@@ -99,6 +126,55 @@ class LLM {
this.memoryManager = new MemoryManager(this);
}
private setupAgent(agents: Agent[] = [], allAgents: Agent[], history: LLMMessage[], aborts: (() => void)[], depth = 0, delegateState: {resp: string | null}): AiTool[] {
return agents.map(a => {
const toolName = `${a.delegate ? '' : 'sub'}agent_${snakeCase(a.name)}`;
return {
name: toolName,
description: `${a.delegate ? 'Delegate to ' : ''}Subagent: ${a.description || a.name}`,
args: <any>(a.delegate ? {} : {
context: {type: 'string', description: 'Summary of related messages, samples, files, etc...', required: true},
instructions: {type: 'string', description: 'Detailed instructions for subagent to complete', required: true},
}),
fn: async (args: any, stream: any, ai: any, id?: string) => {
if(depth >= MAX_AGENT_DEPTH) return 'Max agent delegation depth exceeded';
const nested = (a.agents || [])
.map(name => allAgents.find(x => x.name === name))
.filter((x): x is Agent => !!x && x.name !== a.name);
// Delegate continues the SAME live conversation - no new user turn needed,
// `history` is always current (shared, mutated in place) by the time this runs
const q = a.delegate ? '' : `${args.instructions}${args.context ? `\n\n<context>${args.context}</context>` : ''}`;
const request = this.ask(q, {
system: `You are a specialized subagent. ${a.delegate ? 'Your output streams directly to the user for the remainder of this turn. You are mid conversation - dispense with greetings.' : 'You are wrapped in a tool call that will be analysis by an LLM - dispense with conversation'}
As a subagent, focus on executing your task completely using available tools and returning only the final result - no commentary, questions, or dialogue.
${a.system}`,
model: a.model || undefined,
temperature: a.temperature,
stream: a.delegate ? stream : undefined,
history: a.delegate ? history : [],
mcp: a.mcp || undefined,
skills: a.skills || undefined,
tools: a.tools || undefined,
agents: nested,
_agentDepth: depth + 1,
} as any);
aborts.push(request.abort);
const resp = await request;
if(a.delegate) {
delegateState.resp = resp;
return '';
}
return resp;
}
};
});
}
private async setupMcp(servers: McpServer[] = []): Promise<{prompt: string, tools: AiTool[]}> {
if(!servers?.length) return {prompt: '', tools: []};
const allTools: AiTool[] = [];
@@ -143,7 +219,7 @@ class LLM {
return {
prompt: `You have access to the following skill documents, use \`read_skill\` to access them:\n${list}`,
tools: [{
name: 'read_skill',
name: 'skill_read',
description: 'Read the full content of a skill/knowledge document',
args: {
name: {type: 'string', description: 'Exact skill name', required: true}
@@ -157,6 +233,20 @@ class LLM {
}
}
private wrapToolTiming(tools: AiTool[], timings: Map<string, {duration: number, tps: number}>): AiTool[] {
return tools.map(t => ({
...t,
fn: async (args: any, stream: any, ai: any, id?: string) => {
const start = Date.now();
const result = await t.fn(args, stream, ai, id);
const duration = Date.now() - start;
const tps = duration > 0 ? this.estimateTokens(result) / (duration / 1000) : 0;
if(id) timings.set(id, {duration, tps});
return result;
}
}));
}
ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
options = <any>{
system: '',
@@ -167,11 +257,25 @@ class LLM {
}
const m = options.model || this.defaultModel;
if(!this.models[m]) throw new Error(`Model does not exist: ${m}`);
let abort = () => {};
return Object.assign(new Promise<string>(async res => {
let request: AbortablePromise<string> | null = null;
let aborted = false;
const nestedAborts: (() => void)[] = [];
const abort = () => {
aborted = true;
request?.abort?.();
nestedAborts.forEach(a => a());
};
let promise: any;
const requestStart = Date.now();
promise = (async () => {
let tools: AiTool[] = options.tools || this.ai.options.llm?.tools || [];
const prompts: string[] = [];
// `history` is the single source of truth from here on - mutated in place by
// this call AND by any nested/delegated agent calls sharing the same array
let history = options.history || [];
if(message) history.push({role: 'user', content: message, timestamp: Date.now()});
// MCP
const mcp = options.mcp || this.ai.options?.llm?.mcp;
@@ -189,48 +293,95 @@ class LLM {
tools.push(...s.tools);
}
// Agents
const agents = options.agents || this.ai.options?.llm?.agents;
const delegateState: {resp: string | null} = {resp: null};
if(agents?.length) tools.push(...this.setupAgent(agents, agents, history, nestedAborts, options._agentDepth || 0, delegateState));
// Memory
if (options.memory) {
const mems = options.memory instanceof MemoryCache ? options.memory.memories : options.memory;
const relevant = await this.memoryManager.recollect(message, options.memory, 5);
const mem = MemoryManager.normalize(options.memory);
if(mem) {
const mems = mem.memory instanceof MemoryCache ? mem.memory.memories : mem.memory;
if(mems.length) {
if(mem.inject) {
const pool = 15; // candidates considered, cheap since only refs are listed
const budget = mem.maxTokens ?? 2000; // actual content injected
const relevant = await this.memoryManager.recollect(message, mem.memory, pool);
let used = 0;
const preloaded: typeof relevant = [];
const listed: typeof relevant = [];
for(const r of relevant) {
const t = this.estimateTokens(r.content);
if(used + t <= budget || preloaded.length === 0) {
preloaded.push(r);
used += t;
} else listed.push(r);
}
prompts.unshift(`You have access to the following memory files:
${mems.map(m => `- ${m.name}: ${m.description}`).join('\n')}
${relevant.length ? `
${preloaded.length ? `
Relevant memories have been preloaded:
${relevant.map(r => `
${preloaded.map(r => `
**${r.name}**
${r.description}
${r.content}
`).join('\n---\n')}
` : ''}${listed.length ? `
Also relevant but not preloaded (use \`memory_recall\`): ${listed.map(r => r.name).join(', ')}
` : ''}`.trim());
tools.push(this.memoryManager.tools.read(options.memory));
}
if(mem.tool) tools.push(this.memoryManager.tools.read(mem.memory));
}
}
if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
const toolTimings = new Map<string, {duration: number, tps: number}>();
tools = this.wrapToolTiming(tools, toolTimings);
if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
prompts.unshift(options.system || this.ai.options.llm?.system || '');
const resp = await this.models[m].ask(message, {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
// Message already appended to shared `history` above - pass '' so the provider
// doesn't push a duplicate user turn
request = this.models[m].ask('', {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
let resp = await request;
// Trim memory injections from history
if(options.memory) {
history.splice(0, history.length, ...history.filter(h => h.role !== 'tool' || h.name !== 'recall'));
// Capture meta (duration / tps)
for(const h of history) {
if(h.role === 'tool' && toolTimings.has(h.id)) Object.assign(h, toolTimings.get(h.id));
}
// Auto-memorize before compressing
if(typeof resp === 'string' && !resp.trim() && delegateState.resp !== null) resp = delegateState.resp;
if(mem?.tool) history.splice(0, history.length, ...history.filter(h => h.role !== 'tool' || h.name !== 'memory_recall'));
if(options.compress && this.estimateTokens(history) >= options.compress.max) {
if(options.memory) await this.memoryManager.memorize(history, options.memory, {model: options.memoryModel || this.defaultModel, ...options});
if(mem?.update) await this.memoryManager.memorize(history, mem.memory, {model: options.memoryModel || this.defaultModel, ...options});
const compressed = await this.compressHistory(history, options.compress.max, options.compress.min, options);
if(options.history) options.history.splice(0, options.history.length, ...compressed);
}
return res(resp);
}), {abort});
const requestDuration = Date.now() - requestStart;
const totalTokens = history
.filter((h: any) => h.role === 'assistant' && h.duration && h.tps)
.reduce((sum: number, h: any) => sum + h.tps * (h.duration / 1000), 0);
const requestTps = requestDuration > 0 ? totalTokens / (requestDuration / 1000) : 0;
Object.assign(promise, {duration: requestDuration, tps: requestTps});
return resp;
})();
return Object.assign(promise, {abort});
}
/**
* Digest full conversation history into memory documents.
* Call on session end to persist the conversation.
*/
async updateMemory(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest = {}): Promise<void> {
await this.memoryManager.memorize(history, memories, {model: this.defaultModel, ...options});
async updateMemory(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest = {}): Promise<Memory[]> {
return this.memoryManager.memorize(history, memories, {model: this.defaultModel, ...options});
}
/**
@@ -383,15 +534,33 @@ ${r.content}
* @param {string} searchTerms Multiple search terms to check against target
* @returns {{avg: number, max: number, similarities: number[]}} Similarity values 0-1: 0 = unique, 1 = identical
*/
fuzzyMatch(target: string, ...searchTerms: string[]) {
fuzzyMatch(target, ...searchTerms) {
if (searchTerms.length < 2) throw new Error('Requires at least 2 strings to compare');
const vector = (text: string, dimensions: number = 10): number[] => {
return text.toLowerCase().split('').map((char, index) =>
(char.charCodeAt(0) * (index + 1)) % dimensions / dimensions).slice(0, dimensions);
const levenshtein = (a, b) => {
const m = a.length, n = b.length;
if (!m) return n;
if (!n) return m;
const dp = Array.from({length: m + 1}, (_, i) => [i, ...Array(n).fill(0)]);
for (let j = 0; j <= n; j++) dp[0][j] = j;
for (let i = 1; i <= m; i++) {
for (let j = 1; j <= n; j++) {
dp[i][j] = a[i - 1] === b[j - 1]
? dp[i - 1][j - 1]
: 1 + Math.min(dp[i - 1][j - 1], dp[i - 1][j], dp[i][j - 1]);
}
const v = vector(target);
const similarities = searchTerms.map(t => vector(t)).map(refVector => this.cosineSimilarity(v, refVector));
return {avg: similarities.reduce((acc, s) => acc + s, 0) / similarities.length, max: Math.max(...similarities), similarities};
}
return dp[m][n];
};
const similarity = (a, b) => {
a = a.toLowerCase(); b = b.toLowerCase();
return 1 - levenshtein(a, b) / Math.max(a.length, b.length, 1);
};
const similarities = searchTerms.map(t => similarity(target, t));
return {
avg: similarities.reduce((acc, s) => acc + s, 0) / similarities.length,
max: Math.max(...similarities),
similarities
};
}
/**

View File

@@ -1,146 +1,22 @@
import {LLMRequest, LLMMessage} from './llm.ts';
import {AiTool} from './tools.ts';
import {KDTree, KDPoint} from './kd-tree.ts';
import {KDPoint, KDTree} from './kd-tree.ts';
export type Memory = {
name: string;
description: string;
content: string;
embedding: number[];
}
const FACTS_HEADING = '## Facts';
type MemoryRef = {
name: string;
description: string;
}
const GENERIC_TEMPLATE = `# {{Title}}
type FactBucket = {
subject: string;
facts: string[];
isNew: boolean;
}
## Summary
export type MemoryNode = {
name: string;
missing: boolean;
links: string[];
backlinks: string[];
}
## Details
export function buildMemoryGraph(memories: Memory[] | MemoryCache): MemoryNode[] {
const mems = memories instanceof MemoryCache ? memories.memories : memories;
const nameSet = new Set(mems.map(m => m.name));
const ghosts = new Set<string>();
const nodes: MemoryNode[] = mems.map(m => {
const {links, backlinks} = extractMetadata(m.content);
return {
name: m.name,
missing: false,
links,
backlinks,
};
});
for (const node of nodes) {
for (const link of node.links) {
if (!nameSet.has(link)) ghosts.add(link);
}
}
return [
...nodes,
...[...ghosts].map(name => ({
name,
missing: true,
links: [],
backlinks: nodes
.filter(n => n.links.includes(name))
.map(n => n.name),
}))
];
}
function extractLinks(content: string): string[] {
const matches = content.matchAll(/\[\[([^\]]+)\]\]/g);
return [...new Set([...matches].map(m => m[1].trim()))];
}
export function extractMetadata(content: string): {links: string[], backlinks: string[]} {
const match = content.match(/^---\n([\s\S]*?)\n---/);
if (!match) return {links: [], backlinks: []};
const fm = match[1];
const getList = (key: string): string[] => {
const m = fm.match(new RegExp(`^${key}:\\s*\\[(.*)\\]$`, 'm'));
if (!m || !m[1].trim()) return [];
return m[1].split(',').map(s => s.trim().replace(/^"|"$/g, '')).filter(Boolean);
};
return {
links: getList('links'),
backlinks: getList('backlinks'),
};
}
function cosineDistance(a: number[], b: number[]): number {
let dot = 0, normA = 0, normB = 0;
for (let i = 0; i < a.length; i++) {
dot += a[i] * b[i];
normA += a[i] * a[i];
normB += b[i] * b[i];
}
const denom = Math.sqrt(normA) * Math.sqrt(normB);
return denom === 0 ? 1 : 1 - dot / denom;
}
function getWeekMonday(date: Date = new Date()): string {
const d = new Date(Date.UTC(date.getFullYear(), date.getMonth(), date.getDate()));
const day = d.getUTCDay();
const diff = day === 0 ? -6 : 1 - day;
d.setUTCDate(d.getUTCDate() + diff);
return d.toISOString().slice(0, 10);
}
function getWeekSunday(monday: string): string {
const d = new Date(`${monday}T00:00:00Z`);
d.setUTCDate(d.getUTCDate() + 6);
return d.toISOString().slice(0, 10);
}
function tagsFromName(name: string): string[] {
const prefix = name.split('/')[0];
return prefix ? [prefix.toLowerCase()] : [];
}
export function serializeMemory(mem: Memory, week?: {monday: string, sunday: string}): string {
return mem.content;
}
export function deserializeMemory(raw: string, embedding: number[] = []): Memory {
const match = raw.match(/^---\n([\s\S]*?)\n---\n\n?([\s\S]*)$/);
if (!match) {
return {name: '', description: '', content: raw.trim(), embedding};
}
const [, fm] = match;
const get = (key: string): string => {
const m = fm.match(new RegExp(`^${key}:\\s*(.+)$`, 'm'));
return m ? m[1].trim() : '';
};
return {
name: get('name'),
description: get('description'),
content: raw.trim(),
embedding,
};
}
## Related`;
export class MemoryCache {
private tree: KDTree<MemoryRef>;
public memories: Memory[];
private locks = new Map<string, Promise<void>>();
get length() { return this.memories.length; }
constructor(memories: Memory[]) {
this.memories = memories;
@@ -191,49 +67,115 @@ export class MemoryCache {
rebuild(): void {
this.tree = this.buildTree();
}
lock<T>(name: string, fn: () => Promise<T>): Promise<T> {
const prev = this.locks.get(name) ?? Promise.resolve();
let resolveLock!: () => void;
const next = new Promise<void>(r => { resolveLock = r; });
this.locks.set(name, next);
const result = prev.then(fn).finally(resolveLock);
result.finally(() => {
if (this.locks.get(name) === next) this.locks.delete(name);
});
return result;
}
export type MemoryOptions = {
/** Memory object */
memory: Memory[] | MemoryCache;
/** Inject N memories into the system prompt */
inject?: boolean;
/** expose recall tool to LLM */
tool?: boolean;
/** Update memory on compression */
update?: boolean;
/** Max context size of memories to inject to each call (removed immediately after use) */
maxTokens?: number;
}
export type Memory = {
name: string;
description: string;
content: string;
embedding: number[];
links: string[];
backlinks: string[];
}
type MemoryRef = {
name: string;
description: string;
}
type FactBucket = {
subject: string;
facts: string[];
}
function extractLinks(content: string): string[] {
if (!content) return [];
const matches = content.matchAll(/\[\[([^\]]+)\]\]/g);
return [...new Set([...matches].map(m => m[1].trim()))];
}
export function rebuildGraph(memories: Memory[]): void {
for (const m of memories) m.links = extractLinks(m.content).filter(l => l !== m.name);
for (const m of memories) m.backlinks = [];
for (const m of memories) {
for (const link of m.links) {
const target = memories.find(t => t.name === link);
if (target) target.backlinks.push(m.name);
}
}
}
function dedupeFacts(facts: string[]): string[] {
const seen = new Map<string, string>();
for (const f of facts) {
const clean = f.trim();
if (clean) seen.set(clean.toLowerCase(), clean);
}
return [...seen.values()];
}
function cosineDistance(a: number[], b: number[]): number {
let dot = 0, normA = 0, normB = 0;
for (let i = 0; i < a.length; i++) {
dot += a[i] * b[i];
normA += a[i] * a[i];
normB += b[i] * b[i];
}
const denom = Math.sqrt(normA) * Math.sqrt(normB);
return denom === 0 ? 1 : 1 - dot / denom;
}
function getWeekMonday(date: Date = new Date()): string {
const d = new Date(Date.UTC(date.getFullYear(), date.getMonth(), date.getDate()));
const day = d.getUTCDay();
const diff = day === 0 ? -6 : 1 - day;
d.setUTCDate(d.getUTCDate() + diff);
return d.toISOString().slice(0, 10);
}
export class MemoryManager {
private pendingMemorizations = new Map<string, {
memories: Memory[] | MemoryCache,
tempMemoryName: string,
timestamp: number,
private recentlyTouched = new Map<string, number>();
private queues = new Map<string, {
dirty: boolean,
request: {abort?: () => void} | null,
task: Promise<void>,
}>();
tools = {
read: (memories: Memory[] | MemoryCache): AiTool => ({
name: 'read_memory',
name: 'memory_recall',
description: 'Read the full content of a memory document',
args: {
name: {type: 'string', description: 'Exact memory name', required: true},
},
fn: (args: any) => {
const mems = memories instanceof MemoryCache ? memories.memories : memories;
const mems = this.unwrap(memories);
const mem = mems.find(m => m.name === args.name);
if (!mem) return 'Document not found';
this.touch(mem.name);
return mem.content;
},
}),
forget: (memories: Memory[] | MemoryCache): AiTool => ({
name: 'forget_memory',
name: 'memory_forget',
description: 'Permanently delete a memory document and clean up all references to it',
args: {
name: {type: 'string', description: 'Exact memory name to forget', required: true},
reason: {type: 'string', description: 'Why this memory is being deleted', required: true},
name: {type: 'string', description: 'Exact memory name to forget', required: true}
},
fn: (args: any) => {
const result = this.forget(args.name, memories);
@@ -244,81 +186,94 @@ export class MemoryManager {
constructor(private llm: any) {}
private async createTempMemory(conversation: string): Promise<Memory> {
const [e] = await this.llm.embedding(conversation);
const timestamp = Date.now();
return {
name: `_temp_${timestamp}`,
description: 'Temporary memory - processing in background',
content: `---
name: _temp_${timestamp}
description: Temporary memory - processing in background
tags: [_temporary]
links: []
backlinks: []
modified: ${new Date().toISOString()}
---
static normalize(m?: Memory[] | MemoryCache | MemoryOptions) {
if(!m) return null;
const raw = m instanceof MemoryCache || Array.isArray(m);
return raw ? {memory: <Memory[] | MemoryCache>m, inject: true, tool: true, update: true} : {inject: true, tool: true, update: true, ...m};
}
# Recent Conversation (Processing)
private unwrap(memories: Memory[] | MemoryCache): Memory[] {
return memories instanceof MemoryCache ? memories.memories : memories;
}
${conversation}`,
embedding: e?.embedding || [],
};
private sync(memories: Memory[] | MemoryCache): void {
if (memories instanceof MemoryCache) memories.rebuild();
}
private parseFrontmatter(content: string): {fm: Map<string, string>, body: string} {
const match = content.match(/^---\n([\s\S]*?)\n---\n?([\s\S]*)$/);
if (!match) return {fm: new Map(), body: content};
const fm = new Map<string, string>();
for (const line of match[1].split('\n')) {
const i = line.indexOf(':');
if (i === -1) continue;
fm.set(line.slice(0, i).trim(), line.slice(i + 1).trim());
}
return {fm, body: match[2]};
}
private writeFrontmatter(fm: Map<string, string>, body: string): string {
const lines = [...fm.entries()].map(([k, v]) => `${k}: ${v}`);
return `---\n${lines.join('\n')}\n---\n\n${body.trimStart()}`;
}
private stripHeader(content: string): string {
return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
}
private touchHeader(node: Memory, body: string): string {
const {fm} = this.parseFrontmatter(node.content);
fm.set('name', node.name);
fm.set('description', node.description || '');
fm.set('modified', new Date().toISOString());
return this.writeFrontmatter(fm, body);
}
private ensureDoc(node: Memory): void {
if (node.content) return;
const title = node.name.split('/').pop() ?? node.name;
node.content = this.touchHeader(node, `# ${title}\n`);
}
private appendFacts(node: Memory, facts: string[]): void {
this.ensureDoc(node);
const body = this.stripHeader(node.content);
const bullets = facts.map(f => `- ${f}`).join('\n');
const idx = body.indexOf(FACTS_HEADING);
const newBody = idx === -1
? `${body.trimEnd()}\n\n${FACTS_HEADING}\n${bullets}\n`
: `${body.slice(0, idx + FACTS_HEADING.length)}\n${bullets}${body.slice(idx + FACTS_HEADING.length)}`;
node.content = this.touchHeader(node, newBody);
}
decay() {
for(const [name, ttl] of this.recentlyTouched) {
if(ttl <= 1) this.recentlyTouched.delete(name);
else this.recentlyTouched.set(name, ttl - 1);
}
}
touch(name: string, ttl = 2) {
this.recentlyTouched.set(name, ttl);
}
getTouched(): string[] {
return [...this.recentlyTouched.keys()];
}
forget(name: string, memories: Memory[] | MemoryCache): boolean {
const mem = memories instanceof MemoryCache ? memories.memories : memories;
const mem = this.unwrap(memories);
const idx = mem.findIndex(m => m.name === name);
if (idx === -1) return false;
for (const node of mem) {
const {links, backlinks} = extractMetadata(node.content);
const newBacklinks = backlinks.filter(b => b !== name);
const newLinks = links.filter(l => l !== name);
if (newBacklinks.length !== backlinks.length || newLinks.length !== links.length) {
node.content = this.updateFrontmatter(node.content, {
links: newLinks,
backlinks: newBacklinks,
});
}
}
mem.splice(idx, 1);
if (memories instanceof MemoryCache) memories.rebuild();
rebuildGraph(mem);
this.sync(memories);
return true;
}
private cosineSearch(query: number[], memories: Memory[], limit: number): MemoryRef[] {
const scored = memories
.filter(m => m.embedding?.length)
.map(m => ({
ref: {name: m.name, description: m.description},
distance: cosineDistance(query, m.embedding),
}))
.sort((a, b) => a.distance - b.distance)
.slice(0, limit);
return scored.map(s => s.ref);
}
private createNode(name: string, memories: Memory[]): Memory {
const existing = memories.find(m => m.name === name);
if (existing) return existing;
return {
name,
description: '',
content: '',
embedding: [],
};
}
private listNodes(memories: Memory[]): MemoryRef[] {
return memories.map(m => ({name: m.name, description: m.description}));
}
async recollect(query: string, memories: Memory[] | MemoryCache, limit = 5, graphDepth = 1): Promise<Memory[]> {
const mem: Memory[] = memories instanceof MemoryCache ? memories.memories : memories;
const mem = this.unwrap(memories);
if (!mem.length) return [];
const [e] = await this.llm.embedding(query);
@@ -336,8 +291,7 @@ ${conversation}`,
for (const name of frontier) {
const node = mem.find(m => m.name === name);
if (!node) continue;
const {links} = extractMetadata(node.content);
for (const link of links) {
for (const link of node.links) {
if (!found.has(link) && mem.find(m => m.name === link)) {
found.add(link);
next.push(link);
@@ -355,284 +309,194 @@ ${conversation}`,
return ordered.map(n => mem.find(m => m.name === n)!).filter(Boolean);
}
async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest): Promise<void> {
private cosineSearch(query: number[], memories: Memory[], limit: number): MemoryRef[] {
const scored = memories
.filter(m => m.embedding?.length)
.map(m => ({
ref: {name: m.name, description: m.description},
distance: cosineDistance(query, m.embedding),
}))
.sort((a, b) => a.distance - b.distance)
.slice(0, limit);
return scored.map(s => s.ref);
}
private listNodes(memories: Memory[]): MemoryRef[] {
return memories.map(m => ({name: m.name, description: m.description}));
}
async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest): Promise<Memory[]> {
const conversation = history
.filter(h => h.role === 'user' || h.role === 'assistant')
.map(h => `[${h.role}]: ${h.content}`).join('\n\n').trim();
if (!conversation) return [];
if (!conversation) return;
const uid = `${Date.now()}_${Math.random().toString(36).slice(2)}`;
// NOTE: adjust field names below (id/tool_call_id/name) to match your LLMMessage/tool-call schema.
const pending = {role: 'tool', name: 'memory_process', id: uid, content: 'Processing…'} as unknown as LLMMessage;
history.push(pending);
const trackingId = `${Date.now()}_${Math.random()}`;
// Create and insert temp memory immediately
const tempMemory = await this.createTempMemory(conversation);
const mem = memories instanceof MemoryCache ? memories.memories : memories;
mem.push(tempMemory);
if (memories instanceof MemoryCache) {
memories.rebuild();
}
this.pendingMemorizations.set(trackingId, {
memories,
tempMemoryName: tempMemory.name,
timestamp: Date.now(),
});
this._memorizeBackground(conversation, memories, options, trackingId)
.catch(err => {
console.error('[memorize] Background memorization failed:', err);
})
.finally(() => {
// Remove temp memory from the exact same memory array/cache
const pending = this.pendingMemorizations.get(trackingId);
if (pending) {
const cleanMem = pending.memories instanceof MemoryCache
? pending.memories.memories
: pending.memories;
const idx = cleanMem.findIndex(m => m.name === pending.tempMemoryName);
if (idx !== -1) {
cleanMem.splice(idx, 1);
console.log(`[memorize] Removed temp memory: ${pending.tempMemoryName}`);
}
if (pending.memories instanceof MemoryCache) {
pending.memories.rebuild();
}
}
this.pendingMemorizations.delete(trackingId);
});
}
private async _memorizeBackground(conversation: string, memories: Memory[] | MemoryCache, options: LLMRequest, trackingId: string): Promise<void> {
const mem = memories instanceof MemoryCache ? memories.memories : memories;
const monday = getWeekMonday();
const sunday = getWeekSunday(monday);
const weekKey = monday;
console.log('[memorize] Starting fact extraction...');
const buckets = await this.factAgent(conversation, mem, options, weekKey);
console.log(`[memorize] Extracted ${buckets.length} buckets:`, buckets);
if (!buckets.length) {
console.log('[memorize] No facts extracted, exiting');
return;
}
const runDocAgent = (node: Memory, bucket: FactBucket, embedding?: number[], week?: {monday: string, sunday: string}) => {
if (memories instanceof MemoryCache) {
return memories.lock(node.name, () => this.docAgent(node, bucket, mem, options, embedding, week));
}
return this.docAgent(node, bucket, mem, options, embedding, week);
};
await Promise.all(buckets.map(async bucket => {
let node = mem.find(m => m.name === bucket.subject && !m.name.startsWith('_temp_'));
let embedding: number[] | undefined;
if (!node || bucket.isNew) {
const [e] = await this.llm.embedding(`${bucket.subject}\n${bucket.facts.join('\n')}`);
embedding = e?.embedding;
const mem = this.unwrap(memories);
const buckets = await this.factAgent(conversation, mem, options, getWeekMonday());
const touched: Memory[] = [];
for (const {subject, facts} of buckets) {
let node = mem.find(m => m.name === subject);
if (!node) {
node = this.createNode(bucket.subject, mem);
node = {name: subject, description: '', content: '', embedding: [], links: [], backlinks: []};
mem.push(node);
}
this.appendFacts(node, facts);
const [e] = await this.llm.embedding(node.content);
if (e) node.embedding = e.embedding;
this.touch(node.name);
touched.push(node);
}
const week = bucket.subject.startsWith('Journal/') ? {monday, sunday} : undefined;
await runDocAgent(node, bucket, embedding, week);
}));
if (memories instanceof MemoryCache) {
memories.rebuild();
if (touched.length) {
rebuildGraph(mem);
this.sync(memories);
(pending as any).content = `Saved to ${touched.map(n => `[[${n.name}]]`).join(', ')}`;
for (const node of touched) this.reconcile(node, memories, options).catch(() => {});
} else {
(pending as any).content = 'Nothing worth remembering.';
}
console.log('[memorize] Completed successfully');
return touched;
}
private buildHeader(node: Memory, week?: {monday: string, sunday: string}, links: string[] = [], backlinks: string[] = []): string {
const tags = node.name.split('/')[0]?.toLowerCase();
const lines = [
'---',
`name: ${node.name}`,
`description: ${node.description || ''}`,
tags ? `tags: [${tags}]` : '',
links.length ? `links: [${links.map(l => `"${l}"`).join(', ')}]` : 'links: []',
backlinks.length ? `backlinks: [${backlinks.map(l => `"${l}"`).join(', ')}]` : 'backlinks: []',
week ? `week: ${week.monday} ${week.sunday}` : '',
`modified: ${new Date().toISOString()}`,
'---',
].filter(Boolean);
return lines.join('\n');
/** Manual/cron entry point. scope 'touched' only reconciles docs with a pending Facts inbox. */
async reconcileVault(memories: Memory[] | MemoryCache, options: LLMRequest, scope: 'touched' | 'all' = 'touched'): Promise<void> {
const mem = this.unwrap(memories);
const targets = scope === 'all' ? mem : mem.filter(m => m.content.includes(FACTS_HEADING));
await Promise.all(targets.map(node => this.reconcile(node, memories, options)));
this.sync(memories);
}
private applyHeader(content: string, header: string): string {
const hasFrontmatter = content.trimStart().startsWith('---');
if (hasFrontmatter) {
return content.replace(/^---[\s\S]*?---\n?/, `${header}\n`);
}
return `${header}\n\n${content}`;
/**
* Coalescing queue: if a doc is already reconciling, mark it dirty and abort the in-flight
* request. The loop below always re-reads node.content fresh, so nothing is ever dropped.
*/
private reconcile(node: Memory, memories: Memory[] | MemoryCache, options: LLMRequest): Promise<void> {
const key = node.name;
const existing = this.queues.get(key);
if (existing) {
existing.dirty = true;
existing.request?.abort?.();
return existing.task;
}
private updateFrontmatter(content: string, updates: {links?: string[], backlinks?: string[]}): string {
const match = content.match(/^---\n([\s\S]*?)\n---\n\n?([\s\S]*)$/);
if (!match) return content;
const [, fm, body] = match;
let newFm = fm;
if (updates.links !== undefined) {
const linksList = updates.links.length ? `[${updates.links.map(l => `"${l}"`).join(', ')}]` : '[]';
newFm = newFm.replace(/^links:.*$/m, `links: ${linksList}`);
const entry = {dirty: false, request: null, task: Promise.resolve()};
this.queues.set(key, entry);
const mem = this.unwrap(memories);
entry.task = (async () => {
do {
entry.dirty = false;
await this.reconcileDoc(node, mem, options, entry);
} while (entry.dirty);
})().finally(() => {
this.queues.delete(key);
rebuildGraph(mem);
this.sync(memories);
});
return entry.task;
}
if (updates.backlinks !== undefined) {
const backlinksList = updates.backlinks.length ? `[${updates.backlinks.map(l => `"${l}"`).join(', ')}]` : '[]';
newFm = newFm.replace(/^backlinks:.*$/m, `backlinks: ${backlinksList}`);
}
newFm = newFm.replace(/^modified:.*$/m, `modified: ${new Date().toISOString()}`);
return `---\n${newFm}\n---\n\n${body}`;
}
private stripHeader(content: string): string {
return content.replace(/^---[\s\S]*?---\n?/, '').trimStart();
}
private async docAgent(node: Memory, bucket: FactBucket, memories: Memory[], options: LLMRequest, precomputedEmbedding?: number[], week?: {monday: string, sunday: string}): Promise<void> {
const {links: oldLinks} = extractMetadata(node.content);
let finalContent = node.content;
await this.llm.ask(
`New facts to integrate:\n${bucket.facts.map(f => `- ${f}`).join('\n')}`,
{
private async reconcileDoc(node: Memory, memories: Memory[], options: LLMRequest, entry: {request: {abort?: () => void} | null}): Promise<void> {
const currentBody = this.stripHeader(node.content);
let update;
try {
for (let i = 0; i < 2 && !update?.content; i++) {
const request = this.llm.ask(currentBody, {
model: options.model,
temperature: 0.3,
system: `You are a knowledge base editor. Integrate the provided facts into the document below.
schema: {
description: {type: 'string', description: 'One-line description of what this document covers, no formatting or emojis', required: true},
content: {type: 'string', description: 'Rewritten document body in markdown, without the frontmatter block', required: true},
},
system: `You are a knowledge base editor maintaining one document in an Obsidian-style vault.
If the document has a "${FACTS_HEADING}" section, integrate every bullet under it into the appropriate part of the document, then remove the "${FACTS_HEADING}" section entirely. If there is no such section, just tidy the document per the rules below.
Structure: follow this generic shape loosely, adapting section names/order to what the content actually needs (e.g. journal-style docs may want a timeline instead of "Details"):
\`\`\`markdown
${GENERIC_TEMPLATE}
\`\`\`
Formatting rules:
- Use Obsidian-style markdown: # headings, **bold** for key terms, bullet lists for facts
- Use Obsidian-style markdown: # headings, **bold** for emphasis, bullet & numbered lists for grouped 1D data, tables for 2D data
- Link related concepts with [[WikiLink]] notation using full paths like [[People/Sarah]] or [[Projects/Website]]
- You may create links to nodes that don't exist yet if the concept is important
- Create links for specific entities (person, place, project, program) and abstract concepts, but skip generics (car, red, dog)
- Keep the document concise, factual, and human-readable
- Resolve any contradictions between old content and new facts (new facts win)
- Do not add filler, preamble, or AI commentary — just clean knowledge documents
- The document begins with a YAML frontmatter block (between --- markers) — do not remove or rewrite it, it is maintained automatically
${week ? '- This is a weekly journal entry. The frontmatter contains the week date range.\n' : ''}
All nodes:
${this.listNodes(memories).map(n => n.name).join(', ') || 'none'}
- Resolve contradictions: newer facts always win — delete the outdated statement entirely, never keep both
- Do not add frontmatter blocks, filler, preamble, or AI commentary
Other nodes in the vault (link to these instead of duplicating their content):
${this.listNodes(memories).filter(n => n.name !== node.name).map(n => n.name).join(', ') || 'none'}
Current document:
\`\`\`markdown
${node.content || '(empty — this is a new document)'}
${currentBody}
\`\`\``,
tools: [{
name: 'update_document',
description: 'Write the complete updated document content. Include everything after the frontmatter block — the frontmatter will be recalculated automatically.',
args: {
description: {type: 'string', description: 'One-line description of what this document covers, no formatting or emojis', required: true},
content: {type: 'string', description: 'Document body in markdown, without the frontmatter block', required: true},
},
fn: (args: any) => {
node.description = args.description;
finalContent = args.content;
return 'Saved';
},
}],
}
);
const newLinks = extractLinks(finalContent).filter(l => l !== node.name);
const newLinkSet = new Set(newLinks);
const oldLinkSet = new Set(oldLinks);
for (const added of newLinkSet) {
if (!oldLinkSet.has(added)) {
const target = memories.find(m => m.name === added);
if (target) {
const {backlinks} = extractMetadata(target.content);
if (!backlinks.includes(node.name)) {
target.content = this.updateFrontmatter(target.content, {
backlinks: [...backlinks, node.name],
});
entry.request = request;
update = await request;
}
}
}
}
for (const removed of oldLinkSet) {
if (!newLinkSet.has(removed)) {
const target = memories.find(m => m.name === removed);
if (target) {
const {backlinks} = extractMetadata(target.content);
target.content = this.updateFrontmatter(target.content, {
backlinks: backlinks.filter(b => b !== node.name),
});
}
}
} catch (err: any) {
if (err?.name === 'AbortError') return;
throw err;
} finally {
entry.request = null;
}
const {backlinks} = extractMetadata(node.content);
const header = this.buildHeader(node, week, newLinks, backlinks);
node.content = this.applyHeader(finalContent, header);
if (precomputedEmbedding) {
node.embedding = precomputedEmbedding;
} else {
const embedInput = `${node.description}\n\n${this.stripHeader(node.content)}`.trim();
const [e] = await this.llm.embedding(embedInput);
if (!update?.content) return;
node.description = node.name !== 'People/User' ? update.description : 'All information about the current user';
node.content = this.touchHeader(node, update.content);
const [e] = await this.llm.embedding(node.content);
if (e) node.embedding = e.embedding;
}
}
private async factAgent(conversation: string, memories: Memory[], options: LLMRequest, weekKey: string): Promise<FactBucket[]> {
const buckets: FactBucket[] = [];
console.log('[factAgent] Starting extraction...');
const buckets = new Map<string, string[]>();
await this.llm.ask(conversation, {
model: options.model,
temperature: 0.2,
system: `You are a fact extractor. Analyze this conversation and extract facts worth remembering long-term.
Rules:
- ONLY extract facts the USER explicitly stated about themselves, their work, or their projects
- ONLY extract current facts the USER explicitly stated about themselves, their work, or their projects
- ONLY extract decisions that were MADE during this conversation
- DO NOT extract anything the AI said, its capabilities, or meta-conversation about the AI
- DO NOT extract greetings, pleasantries, or generic exchanges
- DO NOT extract deltas or changes in facts; ONLY the end fact
- If nothing worth remembering was said, do not call any tools
When extracting facts, you MUST also decide the exact destination path:
- Use an existing node name if the facts clearly belong there
- Create a new path following collection/subject format if needed (e.g., People/Sarah, Projects/Oxide)
- For journal entries, use "journal" (will auto-route to Journal/${weekKey})
- All information primarily about the user should go under "People/User"
- When required, create a new path following collection/subject format (e.g., People/Sarah, Projects/Oxide) — you are not limited to any fixed list of collections, use whatever fits
- For journal entries, use "Journal"
Available nodes:
${this.listNodes(memories).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}`,
- Journal
${this.listNodes(memories).filter(n => !n.name.includes('Journal')).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}`,
tools: [{
name: 'extract_facts',
name: 'facts_extract',
description: 'Submit facts with their destination',
args: {
destination: {type: 'string', description: 'Exact existing node name OR new path (e.g. "People/Sarah", "Projects/Oxide")', required: true},
facts: {type: 'string', description: 'Comma-separated facts', required: true},
create_new: {type: 'boolean', description: 'True if this is a new node that doesn\'t exist yet', required: true},
},
fn: (args: any) => {
console.log('[factAgent] Tool called with:', args);
const subject = args.destination.trim().toLowerCase() === 'journal'
? `Journal/${weekKey}`
: args.destination;
buckets.push({
subject,
facts: args.facts.split(',').map((f: string) => f.trim()).filter(Boolean),
isNew: args.create_new,
});
? `Journal/${weekKey}` : args.destination.trim();
const facts = buckets.get(subject) ?? [];
facts.push(...dedupeFacts(String(args.facts).split(',')));
buckets.set(subject, facts);
return 'Recorded';
},
}],
});
console.log(`[factAgent] Extracted ${buckets.length} buckets:`, buckets);
return buckets;
return buckets.entries().toArray().map(([subject, facts]) => ({subject, facts}));
}
}

View File

@@ -1,82 +1,62 @@
import {OpenAI as openAI} from 'openai';
import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, clean} from '@ztimson/utils';
import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, clean, makeArray} from '@ztimson/utils';
import {AbortablePromise, Ai} from './ai.ts';
import {LLMMessage, LLMRequest} from './llm.ts';
import {LLMProvider} from './provider.ts';
import {TokenPool} from './token-pool.ts';
import {convertSchema} from './tools.ts';
export class OpenAi extends LLMProvider {
client!: openAI;
tokenPool!: TokenPool;
private clients = new Map<string, openAI>();
constructor(public readonly ai: Ai, public readonly host: string | null, public readonly token: string, public model: string) {
constructor(public readonly ai: Ai, public readonly host: string | null, public readonly token: string | string[], public model: string) {
super();
this.client = new openAI(clean({
baseURL: host,
apiKey: token || (host ? 'ignored' : undefined)
}));
const tokens = makeArray(token).filter(Boolean);
this.tokenPool = new TokenPool(...(tokens.length ? tokens : [host ? 'ignored' : '']));
}
private toStandard(history: any[]): LLMMessage[] {
for(let i = 0; i < history.length; i++) {
const h = history[i];
if(h.role === 'assistant' && h.tool_calls) {
const tools = h.tool_calls.map((tc: any) => ({
role: 'tool',
id: tc.id,
name: tc.function.name,
args: JSONAttemptParse(tc.function.arguments, {}),
timestamp: h.timestamp
}));
history.splice(i, 1, ...tools);
i += tools.length - 1;
} else if(h.role === 'tool' && h.content) {
const record = history.find(h2 => h.tool_call_id == h2.id);
if(record) {
if(h.content.includes('"error":')) record.error = h.content;
else record.content = h.content;
private getClient(token: string): openAI {
let client = this.clients.get(token);
if(!client) {
client = new openAI(clean({baseURL: this.host, apiKey: token || undefined}));
this.clients.set(token, client);
}
history.splice(i, 1);
i--;
}
if(!history[i]?.timestamp) history[i].timestamp = Date.now();
}
return history;
return client;
}
private fromStandard(history: LLMMessage[]): any[] {
return history.reduce((result, h) => {
/** Convert standard history -> OpenAI wire format */
private toWire(history: LLMMessage[], system?: string): any[] {
const wire: any[] = [];
if(system) wire.push({role: 'system', content: system});
for(const h of history) {
if(h.role === 'tool') {
result.push({
wire.push({
role: 'assistant',
content: null,
tool_calls: [{id: h.id, type: 'function', function: {name: h.name, arguments: JSON.stringify(h.args)}}],
refusal: null,
annotations: []
}, {
role: 'tool',
tool_call_id: h.id,
content: h.error || h.content
content: h.error || h.content || '',
});
} else {
const {timestamp, ...rest} = h;
result.push(rest);
wire.push({role: h.role, content: h.content});
}
return result;
}, [] as any[]);
}
return wire;
}
ask(message: string, options: LLMRequest = {}): AbortablePromise<string | any> {
const controller = new AbortController();
return Object.assign(new Promise<any>(async (res, rej) => {
if(options.system) {
if(options.history?.[0]?.role != 'system') options.history?.splice(0, 0, {role: 'system', content: options.system, timestamp: Date.now()});
else options.history[0].content = options.system;
}
let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
if(!options.history) options.history = [];
const history = options.history;
if(message) history.push({role: 'user', content: message, timestamp: Date.now()});
const tools = options.tools || this.ai.options.llm?.tools || [];
const requestParams: any = {
model: options.model || this.model,
messages: history,
stream: !!options.stream,
max_completion_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || undefined,
temperature: options.temperature || this.ai.options.llm?.temperature || undefined,
@@ -96,92 +76,94 @@ export class OpenAi extends LLMProvider {
if(options.schema) {
const schema = convertSchema(options.schema);
requestParams.response_format = {
type: 'json_schema',
json_schema: {
name: 'response',
strict: true,
schema
}
};
requestParams.response_format = {type: 'json_schema', json_schema: {name: 'response', strict: true, schema}};
}
if(options.stream) requestParams.stream_options = {include_usage: true};
let resp: any, isFirstMessage = true;
try {
let terminal = false;
do {
resp = await this.client.chat.completions.create(requestParams).catch(err => {
err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
requestParams.messages = this.toWire(history.filter(h => h.role !== 'system'), options.system);
const callStart = Date.now();
const resp: any = await this.tokenPool.run(token => this.getClient(token).chat.completions.create(requestParams)).catch(err => {
err.message += `\n\nMessages:\n${JSON.stringify(requestParams.messages, null, 2)}`;
throw err;
});
let usage: any, msg: any = {content: '', tool_calls: []};
if(options.stream) {
if(!isFirstMessage) options.stream({text: '\n\n'});
else isFirstMessage = false;
resp.choices = [{message: {role: 'assistant', content: '', tool_calls: []}}];
for await (const chunk of resp) {
if(controller.signal.aborted) break;
if(chunk.choices[0].delta.content) {
resp.choices[0].message.content += chunk.choices[0].delta.content;
if(chunk.usage) usage = chunk.usage;
if(chunk.choices[0]?.delta?.content) {
msg.content += chunk.choices[0].delta.content;
options.stream({text: chunk.choices[0].delta.content});
}
if(chunk.choices[0].delta.tool_calls) {
if(chunk.choices[0]?.delta?.tool_calls) {
for(const deltaTC of chunk.choices[0].delta.tool_calls) {
const existing = resp.choices[0].message.tool_calls.find(tc => tc.index === deltaTC.index);
const existing = msg.tool_calls.find((tc: any) => tc.index === deltaTC.index);
if(existing) {
if(deltaTC.id) existing.id = deltaTC.id;
if(deltaTC.type) existing.type = deltaTC.type;
if(deltaTC.function) {
if(!existing.function) existing.function = {};
if(deltaTC.function.name) existing.function.name = deltaTC.function.name;
if(deltaTC.function.arguments) existing.function.arguments = (existing.function.arguments || '') + deltaTC.function.arguments;
}
if(deltaTC.function?.name) existing.function.name = deltaTC.function.name;
if(deltaTC.function?.arguments) existing.function.arguments += deltaTC.function.arguments;
} else {
resp.choices[0].message.tool_calls.push({
msg.tool_calls.push({
index: deltaTC.index,
id: deltaTC.id || '',
type: deltaTC.type || 'function',
function: {
name: deltaTC.function?.name || '',
arguments: deltaTC.function?.arguments || ''
}
function: {name: deltaTC.function?.name || '', arguments: deltaTC.function?.arguments || ''}
});
}
}
}
}
} else {
usage = resp.usage;
msg = resp.choices[0].message;
}
const duration = Date.now() - callStart;
const tps = usage?.completion_tokens && duration > 0 ? usage.completion_tokens / (duration / 1000) : 0;
if(resp.error) throw new Error(resp.error);
const toolCalls = resp.choices[0].message.tool_calls || [];
const toolCalls = msg.tool_calls || [];
if(toolCalls.length && !controller.signal.aborted) {
history.push(resp.choices[0].message);
const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
const tool = tools?.find(findByProp('name', toolCall.function.name));
if(options.stream) options.stream({tool: toolCall.function.name});
if(!tool) return {role: 'tool', tool_call_id: toolCall.id, content: '{"error": "Tool not found"}'};
if(msg.content?.trim()) history.push({role: 'assistant', content: msg.content.trim(), timestamp: Date.now(), duration, tps});
const entries = toolCalls.map((tc: any) => {
const entry: any = {role: 'tool', id: tc.id, name: tc.function.name, args: JSONAttemptParse(tc.function.arguments, {}), content: undefined, timestamp: Date.now()};
history.push(entry);
return {tc, entry};
});
await Promise.all(entries.map(async ({tc, entry}: any) => {
const tool = tools.find(findByProp('name', tc.function.name));
if(options.stream) options.stream({tool: tc.function.name});
if(!tool) { entry.error = 'Tool not found'; return; }
try {
const args = JSONAttemptParse(toolCall.function.arguments, {});
const result = await tool.fn(args, options.stream, this.ai);
return {role: 'tool', tool_call_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result};
const toolStream = options.stream && ((chunk: any) => {
if(chunk.done) { terminal = true; return; }
options.stream!(chunk);
});
const result = await tool.fn(entry.args, toolStream, this.ai, tc.id);
entry.content = typeof result === 'object' ? JSONSanitize(result) : result;
} catch(err: any) {
return {role: 'tool', tool_call_id: toolCall.id, content: JSONSanitize({error: err?.message || err?.toString() || 'Unknown'})};
entry.error = err?.message || err?.toString() || 'Unknown';
}
}));
history.push(...results);
requestParams.messages = history;
} else {
terminal = true;
const text = (msg.content || '').trim();
if(text) history.push({role: 'assistant', content: text, timestamp: Date.now(), duration, tps});
}
} while (!controller.signal.aborted && resp.choices?.[0]?.message?.tool_calls?.length);
const textContent = resp.choices[0].message.content?.trim() || '';
history.push({role: 'assistant', content: textContent});
history = this.toStandard(history);
} while(!terminal && !controller.signal.aborted);
if(options.stream) options.stream({done: true});
if(options.history) options.history.splice(0, options.history.length, ...history);
// Return parsed JSON if schema provided
const finalContent = history.at(-1)?.content;
const turnStart = history.map(h => h.role).lastIndexOf('user');
const finalContent = history.slice(turnStart + 1).reduce((str, h) => h.role === 'assistant' ? str + (h.content || '') : str, '').trim();
res(options.schema ? JSONAttemptParse(finalContent, finalContent) : finalContent);
} catch(err) {
rej(err);
}
}), {abort: () => controller.abort()});
}
}

View File

@@ -1,5 +1,5 @@
import {AbortablePromise} from './ai.ts';
import {LLMMessage, LLMRequest} from './llm.ts';
import {LLMRequest} from './llm.ts';
export abstract class LLMProvider {
abstract ask(message: string, options: LLMRequest): AbortablePromise<string>;

65
src/token-pool.ts Normal file
View File

@@ -0,0 +1,65 @@
const DEFAULT_COOLDOWN = 15 * 60 * 1000;
type TokenState = {
token: string;
cooldownUntil: number; // 0 = available now
lastError?: {code: number, message: string};
};
export class TokenPoolExhaustedError extends Error {
constructor(public tokens: Record<string, {code: number, message: string}>) {
super(`All tokens exhausted:\n${Object.entries(tokens).map(([t, e]) => `${t}: [${e.code}] ${e.message}`).join('\n')}`);
this.name = 'TokenPoolExhaustedError';
}
}
export class TokenPool {
private states: TokenState[];
constructor(...tokens: string[]) {
this.states = tokens.map(token => ({token, cooldownUntil: 0}));
}
private preview(token: string): string {
return token.length <= 8 ? '****' : `${token.slice(0, 4)}...${token.slice(-4)}`;
}
/** Anthropic & OpenAI SDKs both attach `status` to thrown errors */
private statusCode(err: any): number {
return err?.status ?? err?.response?.status ?? err?.statusCode;
}
private retryAfter(err: any): number {
const headers = err?.headers || err?.response?.headers;
const raw = headers?.get?.('retry-after') ?? headers?.['retry-after'];
if(raw) {
const seconds = Number(raw);
if(!isNaN(seconds)) return Date.now() + seconds * 1000;
const date = new Date(raw).getTime();
if(!isNaN(date)) return date;
}
return Date.now() + DEFAULT_COOLDOWN;
}
async run<T>(fn: (token: string) => Promise<T>): Promise<T> {
const now = Date.now();
for(const state of this.states) {
if(state.cooldownUntil > now) continue;
try {
const result = await fn(state.token);
state.cooldownUntil = 0;
state.lastError = undefined;
return result;
} catch(err: any) {
const code = this.statusCode(err);
if(![401, 403, 429].includes(code)) throw err;
state.cooldownUntil = code === 429 ? this.retryAfter(err) : Date.now() + DEFAULT_COOLDOWN;
state.lastError = {code, message: err?.message || 'Unknown error'};
}
}
const failures: Record<string, {code: number, message: string}> = {};
this.states.forEach(s => { if(s.lastError) failures[this.preview(s.token)] = s.lastError; });
throw new TokenPoolExhaustedError(failures);
}
}

View File

@@ -41,7 +41,7 @@ export type AiTool = {
/** Tool arguments */
args?: AiToolArg,
/** Callback function */
fn: (args: any, stream: LLMRequest['stream'], ai: Ai) => any | Promise<any>,
fn: (args: any, stream: LLMRequest['stream'], ai: Ai, toolId?: string) => any | Promise<any>,
};
export function convertSchema(schema: any): any {
@@ -91,20 +91,33 @@ export function convertSchema(schema: any): any {
};
}
export const CliTool: AiTool = {
export const ExecCliTool: AiTool = {
name: 'cli',
description: 'Use the command line interface, returns any output',
args: {command: {type: 'string', description: 'Command to run', required: true}},
fn: (args: {command: string}) => $Sync`${args.command}`
}
export const DateTimeTool: AiTool = {
name: 'get_datetime',
description: 'Get local/UTC date/time',
export const ExecJSTool: AiTool = {
name: 'exec_javascript',
description: 'Execute commonjs javascript',
args: {
timezone: {type: 'string', description: 'Which timezone to return, defaults to local', enum: ['local', 'utc'], default: 'local'}
code: {type: 'string', description: 'CommonJS javascript', required: true}
},
fn: ({timezone}) => new Date()[timezone === 'local' ? 'toString' : 'toUTCString']()
fn: async (args: {code: string}) => {
const c = consoleInterceptor(null);
const resp = await Fn<any>({console: c}, args.code, true).catch((err: any) => c.output.error.push(err));
return {...c.output, return: resp, stdout: undefined, stderr: undefined};
}
}
export const ExecPythonTool: AiTool = {
name: 'exec_python',
description: 'Execute commonjs javascript',
args: {
code: {type: 'string', description: 'CommonJS javascript', required: true}
},
fn: async (args: {code: string}) => ({result: $Sync`python -c "${args.code}"`})
}
export const ExecTool: AiTool = {
@@ -118,11 +131,11 @@ export const ExecTool: AiTool = {
try {
switch(args.language) {
case 'cli':
return await CliTool.fn({command: args.code}, stream, ai);
return await ExecCliTool.fn({command: args.code}, stream, ai);
case 'node':
return await JSTool.fn({code: args.code}, stream, ai);
return await ExecJSTool.fn({code: args.code}, stream, ai);
case 'python':
return await PythonTool.fn({code: args.code}, stream, ai);
return await ExecPythonTool.fn({code: args.code}, stream, ai);
default:
throw new Error(`Unsupported language: ${args.language}`);
}
@@ -132,8 +145,483 @@ export const ExecTool: AiTool = {
}
}
export const FetchTool: AiTool = {
name: 'fetch',
export const FsDeleteTool = (whitelist: null | string[] = null): AiTool => {
return {
name: 'fs_delete',
description: 'Delete a file or directory',
args: {
path: {type: 'string', description: 'Path to file or directory', required: true},
recursive: {type: 'boolean', description: 'Delete all children', required: false}
},
fn: async ({path, recursive = false}) => {
const {existsSync, rmSync} = await import('fs');
const normalizePath = p => p.replace(/\\/g, '/');
path = normalizePath(path);
if(whitelist && !whitelist.some(p => path.startsWith(p))) return {error: 'Permission denied'};
if(!existsSync(path)) return {error: 'Path does not exist'};
rmSync(path, {recursive, force: true});
return {success: true, path};
}
}
}
export const FsMoveTool = (whitelist: null | string[] = null): AiTool => {
return {
name: 'fs_move',
description: 'Move or rename a file or directory',
args: {
source: {type: 'string', description: 'Path to source file or directory', required: true},
destination: {type: 'string', description: 'Path to destination file or directory', required: true}
},
fn: async ({source, destination}) => {
const {existsSync, renameSync} = await import('fs');
const normalizePath = p => p.replace(/\\/g, '/');
source = normalizePath(source);
destination = normalizePath(destination);
if(whitelist && !whitelist.some(p => source.startsWith(p) && destination.startsWith(p))) return {error: 'Permission denied'};
if(!existsSync(source)) return {error: 'Source path does not exist'};
if(existsSync(destination)) return {error: 'Destination path already exists'};
renameSync(source, destination);
return {success: true, source, destination};
}
}
}
export const FsReadTool = (whitelist: null | string[] = null): AiTool => {
return {
name: 'fs_read',
description: 'Read the contents of a provided path. Works with files and directories',
args: {path: {type: 'string', description: 'Path to file or directory', required: true}},
fn: async ({path}) => {
const {existsSync, lstatSync, readdirSync, readFileSync} = await import('fs');
const {join} = await import('path');
const normalizePath = p => p.replace(/\\/g, '/');
path = normalizePath(path);
if(whitelist && !whitelist.some(p => path.startsWith(p))) return {error: 'Permission denied'};
if(!existsSync(path)) return {error: 'Path does not exist'};
const stats = lstatSync(path);
if(stats.isDirectory()) {
const children = readdirSync(path).map(name => {
const childPath = normalizePath(join(path, name));
const childStats = lstatSync(childPath);
return {name, type: childStats.isDirectory() ? 'directory' : 'file', size: childStats.size};
});
return {type: 'directory', children};
}
const content = readFileSync(path, 'utf-8');
return {type: 'file', content};
}
}
}
export const FsSearchTool = (whitelist: null | string[] = null): AiTool => {
return {
name: 'fs_search',
description: 'Scan a directory for matching glob patterns (e.g. "**/*.js", "src/**/*.test.ts")',
args: {
pattern: {type: 'string', description: 'Glob pattern to match against paths', required: true},
root: {type: 'string', description: 'Directory to search from', required: false, default: '.'}
},
fn: async ({pattern, root = '.'}) => {
const {existsSync, lstatSync, readdirSync} = await import('fs');
const {join, relative} = await import('path');
const normalizePath = p => p.replace(/\\/g, '/');
root = normalizePath(root);
if(!existsSync(root)) return {error: 'Root path does not exist'};
if(!lstatSync(root).isDirectory()) return {error: 'Root path is not a directory'};
if(whitelist && !whitelist.some(p => root.startsWith(p))) return {error: 'Permission denied'};
const globToRegex = (glob) => {
let re = '';
for(let i = 0; i < glob.length; i++) {
const c = glob[i];
if(c === '*') {
if(glob[i + 1] === '*') {
const isSlash = glob[i + 2] === '/';
re += '.*';
i += isSlash ? 2 : 1;
} else {
re += '[^/]*';
}
} else if(c === '?') {
re += '[^/]';
} else if('.+^$(){}|[]\\'.includes(c)) {
re += '\\' + c;
} else {
re += c;
}
}
return new RegExp('^' + re + '$');
};
const regex = globToRegex(pattern);
const results: any = [];
const walk = (dir) => {
for(const name of readdirSync(dir)) {
const fullPath = normalizePath(join(dir, name));
const stats = lstatSync(fullPath);
const relPath = normalizePath(relative(root, fullPath));
if(regex.test(relPath)) {
results.push({path: relPath, type: stats.isDirectory() ? 'directory' : 'file', size: stats.size});
}
if(stats.isDirectory()) walk(fullPath);
}
};
walk(root);
return results;
}
}
}
export const FsWriteTool = (whitelist: null | string[] = null): AiTool => {
return {
name: 'fs_write',
description: 'Create a directory, write content to a file or preform a find & replace',
args: {
path: {type: 'string', description: 'Path to file or directory', required: true},
content: {type: 'string', description: 'Content to write or replace (Omit to create a directory)'},
find: {type: 'string', description: 'Text or regex pattern to match (regex must match pattern: "/pattern/g")'}
},
fn: async ({path, content, find}) => {
const {existsSync, mkdirSync, readFileSync, writeFileSync} = await import('fs');
const {dirname} = await import('path');
const normalizePath = p => p.replace(/\\/g, '/');
path = normalizePath(path);
if(whitelist && !whitelist.some(p => path.startsWith(p))) return {error: 'Permission denied'};
if(content === undefined) {
mkdirSync(path, {recursive: true});
return {success: true, type: 'directory', path};
}
const dir = normalizePath(dirname(path));
if(!existsSync(dir)) mkdirSync(dir, {recursive: true});
if(find && existsSync(path)) {
const existing = readFileSync(path, 'utf-8');
const regexMatch = find.match(/^\/(.+)\/([gimuy]*)$/);
const pattern = regexMatch ? new RegExp(regexMatch[1], regexMatch[2]) : find;
if(!existing.match(pattern)) return {error: 'Find pattern not found in file'};
const updated = existing.replace(pattern, content);
writeFileSync(path, updated, 'utf-8');
return {success: true, type: 'file', path, replaced: true, content: updated};
}
writeFileSync(path, content, 'utf-8');
return {success: true, type: 'file', path, content};
}
}
}
export const GetPathsTool: AiTool = {
name: 'get_paths',
description: 'Get the current working directory, and paths to the users home directory',
fn: async () => {
return {
home: os.homedir(),
cwd: process.cwd()
};
}
}
export const GetDatetimeTool: AiTool = {
name: 'get_datetime',
description: 'Get local/UTC timestamp',
args: {
timezone: {type: 'string', description: 'Which timezone to return, defaults to local', enum: ['local', 'utc'], default: 'local'}
},
fn: ({timezone}) => new Date()[timezone === 'local' ? 'toString' : 'toUTCString']()
}
export const GetDevice: AiTool = {
name: 'get_device',
description: 'Get comprehensive system information including hostname, specs, load, storage, and network status',
args: {},
fn: async () => {
const platform = os.platform();
const hostname = os.hostname();
// CPU Info
const cpus = os.cpus();
const cpuModel = cpus[0].model;
const cpuCores = cpus.length;
// Memory Info
const totalMem: any = (os.totalmem() / 1024 / 1024 / 1024).toFixed(2);
const freeMem: any = (os.freemem() / 1024 / 1024 / 1024).toFixed(2);
const usedMem: any = (totalMem - freeMem).toFixed(2);
const memUsage: any = ((usedMem / totalMem) * 100).toFixed(1);
// Load Average (not available on Windows)
const loadAvg = platform === 'win32' ? ['N/A', 'N/A', 'N/A'] : os.loadavg().map(l => l.toFixed(2));
// Storage Usage
let storage = {};
if(platform === 'win32') {
const ps = $Sync`powershell "Get-PSDrive C | Select-Object Used,Free | ConvertTo-Json"`.trim();
const drive = JSON.parse(ps);
const used: any = (drive.Used / 1024 / 1024 / 1024).toFixed(2);
const free: any = (drive.Free / 1024 / 1024 / 1024).toFixed(2);
const total: any = (parseFloat(used) + parseFloat(free)).toFixed(2);
const usage: any = ((used / total) * 100).toFixed(1);
storage = {
filesystem: 'C:',
size: `${total} GB`,
used: `${used} GB`,
available: `${free} GB`,
usage: `${usage}%`
};
} else {
const df = $Sync`df -h / | tail -1`.trim();
const s = df.split(/\s+/);
storage = {
filesystem: s[0],
size: s[1],
used: s[2],
available: s[3],
usage: s[4]
};
}
// Network Status
const interfaces = os.networkInterfaces();
const activeIfaces = Object.entries(interfaces)
.filter(([name]) => name !== 'lo' && !name.includes('Loopback'))
.map(([name, addrs]) => {
const ipv4 = addrs?.find(a => a.family === 'IPv4');
return ipv4 ? {name, ip: ipv4.address} : null;
})
.filter(Boolean);
// Internet connectivity check
let internet = false;
try {
if(platform === 'win32') {
$Sync`powershell "Test-Connection -ComputerName 8.8.8.8 -Count 1 -Quiet"`;
} else {
$Sync`ping -c 1 -W 2 8.8.8.8 > /dev/null 2>&1`;
}
internet = true;
} catch {}
// Uptime
const uptime = os.uptime();
const days = Math.floor(uptime / 86400);
const hours = Math.floor((uptime % 86400) / 3600);
const minutes = Math.floor((uptime % 3600) / 60);
return {
hostname,
cpu: {
model: cpuModel,
cores: cpuCores
},
memory: {
total: `${totalMem} GB`,
used: `${usedMem} GB`,
free: `${freeMem} GB`,
usage: `${memUsage}%`
},
load: {
'1min': loadAvg[0],
'5min': loadAvg[1],
'15min': loadAvg[2]
},
storage,
network: {
interfaces: activeIfaces,
internet: internet ? 'connected' : 'disconnected'
},
uptime: `${days}d ${hours}h ${minutes}m`,
platform: `${os.type()} ${os.release()}`
};
}
}
export const GetWikipediaTool: AiTool = {
name: 'get_wikipedia',
description: 'Search Wikipedia for matching articles',
args: {
query: {type: 'string', description: 'Search term or article title', required: true},
mode: {type: 'string', description: 'search - look for articles, summary - intro of first found article (default), full - complete first found article', enum: ['search', 'summary', 'full'], default: 'summary'},
ua: {type: 'string', description: 'User Agent'},
},
fn: async ({query, mode, ua}) => {
class WikipediaClient {
useragent = 'Mozilla/5.0 (Windows NT 10.0; Win64; x64)';
constructor(useragent: string) {
this.useragent = useragent;
}
async get(url) {
const resp = await fetch(url, {headers: {'User-Agent': this.useragent}});
return resp.json();
}
api(params) {
const qs = new URLSearchParams({...params, format: 'json', utf8: '1'}).toString();
return this.get(`https://en.wikipedia.org/w/api.php?${qs}`);
}
clean(text) {
const cutoffs = ['== See also ==', '== References ==', '== Bibliography ==', '== External links =='];
for (const marker of cutoffs) {
const idx = text.indexOf(marker);
if (idx !== -1) text = text.slice(0, idx);
}
return text
.replace(/^={4}\s*(.+?)\s*={4}$/gm, '#### $1')
.replace(/^={3}\s*(.+?)\s*={3}$/gm, '### $1')
.replace(/^={2}\s*(.+?)\s*={2}$/gm, '## $1')
.replace(/\n{3,}/g, '\n\n')
.replace(/ {2,}/g, ' ')
.replace(/\[\d+]/g, '')
.trim();
}
async searchTitles(query: string, limit = 6) {
const data = await this.api({action: 'query', list: 'search', srsearch: query, srlimit: limit, srprop: 'snippet'});
return data.query?.search || [];
}
async fetchExtract(title: string, introOnly = false) {
const params: any = {action: 'query', prop: 'extracts', titles: title, explaintext: 1, redirects: 1};
if(introOnly) params.exintro = 1;
const data = await this.api(params);
const page: any = Object.values(data.query?.pages || {})[0];
return this.clean(page?.extract || '');
}
pageUrl(title: string) {
return `https://en.wikipedia.org/wiki/${encodeURIComponent(title.replace(/ /g, '_'))}`;
}
stripHtml(text: string) {
return text.replace(/<[^>]+>/g, '');
}
async lookup(query: string, detail = 'summary') {
const results = await this.searchTitles(query, 6);
if(!results.length) return `❌ No Wikipedia articles found for "${query}"`;
const title = results[0].title;
const url = this.pageUrl(title);
const introOnly = detail !== 'full';
const content = await this.fetchExtract(title, introOnly);
return `## ${title}\n🔗 ${url}\n\n${content}`;
}
async search(query: string) {
const results = await this.searchTitles(query, 8);
if(!results.length) return `❌ No results for "${query}"`;
const lines = [`### Search results for "${query}"\n`];
for(let i = 0; i < results.length; i++) {
const r = results[i];
const snippet = this.stripHtml(r.snippet || '').trim();
lines.push(`**${i + 1}. ${r.title}**\n${snippet}\n${this.pageUrl(r.title)}`);
}
return lines.join('\n\n');
}
}
const wiki = new WikipediaClient(ua);
if(mode === 'search') return wiki.search(query);
return wiki.lookup(query, mode || 'summary');
}
};
export const GeoCodeTool: AiTool = {
name: 'geo_code',
description: 'Converts coordinates to address OR vice versa',
args: {
query: {type: 'string', description: 'Search query - coordinates (lat,lon) or address string', required: true},
},
fn: async ({query}) => {
const coordinates = /(-?\d+(?:\.\d+)?).*?,.*?(-?\d+(?:\.\d+)?)/.exec(query);
if(coordinates) { // Geolocate
const url = `https://nominatim.openstreetmap.org/reverse?format=json&lat=${encodeURIComponent(coordinates[1])}&lon=${encodeURIComponent(coordinates[2])}`;
const response = await fetch(url, {headers: {'User-Agent': 'OpenSight/1.0', 'Accept-Language': 'en'}});
const data = await response.json();
if(data.display_name) return {address: data.display_name, mode: 'geolocate'};
} else { // Geocode
const url = `https://nominatim.openstreetmap.org/search?format=json&q=${encodeURIComponent(query)}`;
const response = await fetch(url, {headers: {'User-Agent': 'OpenSight/1.0'}});
const data = await response.json();
if(data[0]) return {latitude: parseFloat(data[0].lat), longitude: parseFloat(data[0].lon), mode: 'geocode'};
}
return {error: 'Not found'};
},
}
export const GeoWeatherTool: AiTool = {
name: 'geo_weather',
description: 'Gets weather and air quality info for a location and time',
args: {
query: {type: 'string', description: 'Location - address or place name', required: true},
day: {type: 'string', description: 'Date to retrieve (YYYY-MM-DD), defaults to today'},
},
fn: async ({query, day}) => {
day = day || new Date().toISOString().slice(0, 10);
const geoUrl = `https://nominatim.openstreetmap.org/search?format=json&q=${encodeURIComponent(query)}`;
const geoResponse = await fetch(geoUrl, {headers: {'User-Agent': 'OpenSight/1.0'}});
const geoData = await geoResponse.json();
if(!geoData[0]) return {error: 'Location not found'};
const lat = parseFloat(geoData[0].lat);
const lon = parseFloat(geoData[0].lon);
const weatherUrl = `https://api.open-meteo.com/v1/forecast?latitude=${lat}&longitude=${lon}&start_date=${day}&end_date=${day}&daily=weathercode,temperature_2m_max,temperature_2m_min,apparent_temperature_max,apparent_temperature_min,precipitation_sum,precipitation_probability_max,windspeed_10m_max,winddirection_10m_dominant,uv_index_max,sunrise,sunset&timezone=auto`;
const airUrl = `https://air-quality-api.open-meteo.com/v1/air-quality?latitude=${lat}&longitude=${lon}&start_date=${day}&end_date=${day}&hourly=us_aqi,european_aqi,pm10,pm2_5&timezone=auto`;
const [weatherResponse, airResponse] = await Promise.all([fetch(weatherUrl), fetch(airUrl)]);
const weatherData = await weatherResponse.json();
const airData = await airResponse.json();
const avg = arr => (arr && arr.length) ? arr.reduce((a, b) => a + b, 0) / arr.length : null;
return {
location: geoData[0].display_name,
latitude: lat,
longitude: lon,
elevation: weatherData.elevation,
date: day,
weatherCode: weatherData.daily?.weathercode?.[0],
tempMax: weatherData.daily?.temperature_2m_max?.[0],
tempMin: weatherData.daily?.temperature_2m_min?.[0],
feelsLikeMax: weatherData.daily?.apparent_temperature_max?.[0],
feelsLikeMin: weatherData.daily?.apparent_temperature_min?.[0],
precipitation: weatherData.daily?.precipitation_sum?.[0],
precipitationChance: weatherData.daily?.precipitation_probability_max?.[0],
windSpeedMax: weatherData.daily?.windspeed_10m_max?.[0],
windDirection: weatherData.daily?.winddirection_10m_dominant?.[0],
uvIndexMax: weatherData.daily?.uv_index_max?.[0],
sunrise: weatherData.daily?.sunrise?.[0],
sunset: weatherData.daily?.sunset?.[0],
usAqi: avg(airData.hourly?.us_aqi),
europeanAqi: avg(airData.hourly?.european_aqi),
pm10: avg(airData.hourly?.pm10),
pm2_5: avg(airData.hourly?.pm2_5),
};
},
}
export const WebFetchTool: AiTool = {
name: 'web_fetch',
description: 'Make HTTP request to URL',
args: {
url: {type: 'string', description: 'URL to fetch', required: true},
@@ -149,30 +637,59 @@ export const FetchTool: AiTool = {
}) => new Http({url: args.url, headers: args.headers}).request({method: args.method || 'GET', body: args.body})
}
export const JSTool: AiTool = {
name: 'exec_javascript',
description: 'Execute commonjs javascript',
export const WebFlareSolverTool = (host: string) => {
return {
name: 'web_flaresolverr',
description: 'Use a flaresolverr proxy to bypass cloudflare bot detection',
args: {
code: {type: 'string', description: 'CommonJS javascript', required: true}
url: {type: 'string', description: 'URL to fetch', required: true},
cmd: {type: 'string', description: 'Flaresolverr cmd', enum: ['request.get', 'request.post'], default: 'request.get'},
maxTimeout: {type: 'number', description: 'Fetch time limit', default: 60_000},
postData: {type: 'object', description: 'Data to send during request.post requests'},
},
fn: async (args: {code: string}) => {
const c = consoleInterceptor(null);
const resp = await Fn<any>({console: c}, args.code, true).catch((err: any) => c.output.error.push(err));
return {...c.output, return: resp, stdout: undefined, stderr: undefined};
fn: async ({url, cmd, maxTimeout, postData}) => {
function toFormUrlEncoded(obj, prefix = '') {
const pairs: any = [];
for (const key in obj) {
if (!obj.hasOwnProperty(key)) continue;
const value = obj[key];
const encodedKey = prefix
? `${prefix}[${encodeURIComponent(key)}]`
: encodeURIComponent(key);
if (value === null || value === undefined) {
pairs.push(`${encodedKey}=`);
} else if (typeof value === 'object' && !Array.isArray(value)) {
pairs.push(toFormUrlEncoded(value, encodedKey));
} else if (Array.isArray(value)) {
value.forEach(item => {
pairs.push(`${encodedKey}[]=${encodeURIComponent(item)}`);
});
} else {
pairs.push(`${encodedKey}=${encodeURIComponent(value)}`);
}
}
export const PythonTool: AiTool = {
name: 'exec_python',
description: 'Execute commonjs javascript',
args: {
code: {type: 'string', description: 'CommonJS javascript', required: true}
},
fn: async (args: {code: string}) => ({result: $Sync`python -c "${args.code}"`})
return pairs.join('&');
}
export const ReadWebpageTool: AiTool = {
name: 'read_webpage',
const res = await fetch(host + '/v1', {
method: 'POST',
headers: {'Content-Type': 'application/json'},
body: JSON.stringify({cmd, url, maxTimeout, postData: postData ? toFormUrlEncoded(postData) : undefined}),
});
if(!res.ok) throw new Error(`FlareSolverr HTTP error: ${res.status} ${res.statusText}`);
const data = await res.json();
if(data.status !== 'ok') throw new Error(`FlareSolverr error: ${data.message ?? data.status}`);
return data.solution.response;
}
}
}
export const WebReadTool: AiTool = {
name: 'web_read',
description: 'Extract clean content from webpages, or convert media/documents to accessible formats',
args: {
url: {type: 'string', description: 'URL to read', required: true},
@@ -300,91 +817,3 @@ export const WebSearchTool: AiTool = {
return results;
}
}
export const WikipediaTool: AiTool = {
name: 'wikipedia_search',
description: 'Search Wikipedia for matching articles',
args: {
query: {type: 'string', description: 'Search term or article title', required: true},
mode: {type: 'string', description: 'search - look for articles, summary - intro of first found article (default), full - complete first found article', enum: ['search', 'summary', 'full'], default: 'summary'}
},
fn: async (args: {query: string, mode: 'search' | 'summary' | 'full'}) => {
const UA = 'Mozilla/5.0 (Windows NT 10.0; Win64; x64)';
class WikipediaClient {
async get(url: string) {
const resp = await fetch(url, {headers: {'User-Agent': UA}});
return resp.json();
}
api(params: any) {
const qs = new URLSearchParams({...params, format: 'json', utf8: '1'}).toString();
return this.get(`https://en.wikipedia.org/w/api.php?${qs}`);
}
clean(text: string) {
const cutoffs = ['== See also ==', '== References ==', '== Bibliography ==', '== External links =='];
for (const marker of cutoffs) {
const idx = text.indexOf(marker);
if (idx !== -1) text = text.slice(0, idx);
}
return text
.replace(/^={4}\s*(.+?)\s*={4}$/gm, '#### $1')
.replace(/^={3}\s*(.+?)\s*={3}$/gm, '### $1')
.replace(/^={2}\s*(.+?)\s*={2}$/gm, '## $1')
.replace(/\n{3,}/g, '\n\n')
.replace(/ {2,}/g, ' ')
.replace(/\[\d+\]/g, '')
.trim();
}
async searchTitles(query: string, limit = 6) {
const data = await this.api({action: 'query', list: 'search', srsearch: query, srlimit: limit, srprop: 'snippet'});
return data.query?.search || [];
}
async fetchExtract(title: string, introOnly = false) {
const params: any = {action: 'query', prop: 'extracts', titles: title, explaintext: 1, redirects: 1};
if(introOnly) params.exintro = 1;
const data = await this.api(params);
const page: any = Object.values(data.query?.pages || {})[0];
return this.clean(page?.extract || '');
}
pageUrl(title: string) {
return `https://en.wikipedia.org/wiki/${encodeURIComponent(title.replace(/ /g, '_'))}`;
}
stripHtml(text: string) {
return text.replace(/<[^>]+>/g, '');
}
async lookup(query: string, detail = 'summary') {
const results = await this.searchTitles(query, 6);
if(!results.length) return `❌ No Wikipedia articles found for "${query}"`;
const title = results[0].title;
const url = this.pageUrl(title);
const introOnly = detail !== 'full';
const content = await this.fetchExtract(title, introOnly);
return `## ${title}\n🔗 ${url}\n\n${content}`;
}
async search(query: string) {
const results = await this.searchTitles(query, 8);
if(!results.length) return `❌ No results for "${query}"`;
const lines = [`### Search results for "${query}"\n`];
for(let i = 0; i < results.length; i++) {
const r = results[i];
const snippet = this.stripHtml(r.snippet || '').trim();
lines.push(`**${i + 1}. ${r.title}**\n${snippet}\n${this.pageUrl(r.title)}`);
}
return lines.join('\n\n');
}
}
const wiki = new WikipediaClient();
if(args.mode == 'search') return wiki.search(args.query);
return wiki.lookup(args.query, args.mode || 'summary');
}
};

167
tests/llm.spec.ts Normal file
View File

@@ -0,0 +1,167 @@
import {describe, it, expect, vi, beforeEach} from 'vitest';
import LLM from '../src/llm';
const {FakeProvider, providerLog} = vi.hoisted(() => {
const providerLog: any[] = [];
class FakeProvider {
model: string;
constructor(...args: any[]) { this.model = args[args.length - 1]; }
ask(message: string, opts: any) {
let aborted = false;
const p = (async () => {
const script = (globalThis as any).__scripts?.[this.model];
const plan = script ? script(message, opts) : {text: ''};
providerLog.push({model: this.model, message, system: opts.system, tools: (opts.tools || []).map((t: any) => t.name)});
for (const c of plan.calls || []) {
if (aborted) break;
const tool = (opts.tools || []).find((t: any) => t.name === c.tool);
const id = c.id || `${c.tool}_${Math.random()}`;
const content = await tool.fn(c.args, opts.stream, null, id);
opts.history.push({role: 'tool', id, name: c.tool, args: c.args, content, timestamp: Date.now()});
}
const text = plan.text ?? '';
if (opts.stream && text) opts.stream({text, done: true});
opts.history.push({role: 'assistant', content: text, timestamp: Date.now(), duration: 10, tps: 5});
return text;
})();
return Object.assign(p, {abort: () => { aborted = true; }});
}
}
return {FakeProvider, providerLog};
});
vi.mock('../src/antrhopic.ts', () => ({Anthropic: FakeProvider}));
vi.mock('../src/open-ai.ts', () => ({OpenAi: FakeProvider}));
function makeAi(models: any) {
return {options: {llm: {models}}} as any;
}
beforeEach(() => {
providerLog.length = 0;
(globalThis as any).__scripts = {};
});
describe('LLM cross-provider interchangeability', () => {
it('runs identical tool calls the same way on an anthropic-backed model and an openai-backed model', async () => {
const ai = makeAi({
claude: {proto: 'anthropic', token: 'x'},
gpt: {proto: 'openai', token: 'y', host: 'http://local'},
});
const llm = new LLM(ai);
const calc = {
name: 'calc_add',
description: 'Add two numbers',
args: {a: {type: 'number', required: true}, b: {type: 'number', required: true}},
fn: (args: any) => String(args.a + args.b),
};
(globalThis as any).__scripts.claude = () => ({calls: [{tool: 'calc_add', args: {a: 2, b: 3}}], text: 'Result: 5'});
(globalThis as any).__scripts.gpt = () => ({calls: [{tool: 'calc_add', args: {a: 2, b: 3}}], text: 'Result: 5'});
const historyA: any[] = [], historyB: any[] = [];
const respA = await llm.ask('add 2 and 3', {model: 'claude', tools: [calc], history: historyA});
const respB = await llm.ask('add 2 and 3', {model: 'gpt', tools: [calc], history: historyB});
expect(respA).toBe('Result: 5');
expect(respB).toBe('Result: 5');
expect(providerLog.find(l => l.model === 'claude')!.tools).toContain('calc_add');
expect(providerLog.find(l => l.model === 'gpt')!.tools).toContain('calc_add');
// tool timing gets recomputed from real execution regardless of proto
for (const h of [historyA.find(h => h.name === 'calc_add'), historyB.find(h => h.name === 'calc_add')]) {
expect(h.content).toBe('5');
expect(typeof h.duration).toBe('number');
expect(typeof h.tps).toBe('number');
}
});
it('lets the same shared history flow across model + proto swaps with different system prompts', async () => {
const ai = makeAi({
claude: {proto: 'anthropic', token: 'x'},
gpt: {proto: 'openai', token: 'y', host: 'http://local'},
});
const llm = new LLM(ai);
const history: any[] = [];
(globalThis as any).__scripts.claude = () => ({text: 'Hi from claude'});
(globalThis as any).__scripts.gpt = () => ({text: 'Hi from gpt'});
const r1 = await llm.ask('hello', {model: 'claude', system: 'You are terse.', history});
const r2 = await llm.ask('follow up', {model: 'gpt', system: 'You are verbose.', history});
expect(r1).toBe('Hi from claude');
expect(r2).toBe('Hi from gpt');
expect(history.filter(h => h.role === 'assistant').map(h => h.content)).toEqual(['Hi from claude', 'Hi from gpt']);
expect(providerLog[0].system).toContain('You are terse.');
expect(providerLog[1].system).toContain('You are verbose.');
});
it('exposes MCP tools the same way no matter which proto backs the model', async () => {
const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
const llm = new LLM(ai);
const mcp = [{name: 'weather', host: 'http://mcp.local'}];
global.fetch = vi.fn(async (url: string, opts?: any) => {
if (url.endsWith('/tools')) {
return {json: async () => ({tools: [{name: 'lookup', description: 'Look up weather', inputSchema: {properties: {city: {type: 'string'}}, required: ['city']}}]})} as any;
}
const body = JSON.parse(opts.body);
return {json: async () => ({content: [{text: `Sunny in ${body.arguments.city}`}]})} as any;
}) as any;
for (const model of ['claude', 'gpt']) {
(globalThis as any).__scripts[model] = () => ({calls: [{tool: 'weather_lookup', args: {city: 'Rome'}}], text: 'done'});
const history: any[] = [];
await llm.ask('weather?', {model, mcp, history});
expect(history.find(h => h.name === 'weather_lookup')?.content).toBe('Sunny in Rome');
}
});
it('exposes and resolves skill documents identically across protos', async () => {
const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
const llm = new LLM(ai);
const skills = [{name: 'Onboarding', description: 'How to onboard a user', content: 'Step 1...'}];
for (const model of ['claude', 'gpt']) {
(globalThis as any).__scripts[model] = () => ({calls: [{tool: 'skill_read', args: {name: 'Onboarding'}}], text: 'done'});
const history: any[] = [];
await llm.ask('onboard me', {model, skills, history});
expect(history.find(h => h.name === 'skill_read')?.content).toContain('Step 1...');
}
});
it('delegate agent mutates the shared history directly and backfills the orchestrator response, across protos', async () => {
const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
const llm = new LLM(ai);
const history: any[] = [{role: 'user', content: 'research quantum computing'}];
const researcher = {name: 'researcher', system: 'You research topics.', delegate: true, model: 'gpt'};
(globalThis as any).__scripts.claude = () => ({calls: [{tool: 'agent_researcher', args: {}}], text: ''});
(globalThis as any).__scripts.gpt = () => ({text: 'Quantum computers use qubits.'});
const resp = await llm.ask('go', {model: 'claude', agents: [researcher], history});
expect(resp).toBe('Quantum computers use qubits.');
expect(history.some(h => h.role === 'assistant' && h.content === 'Quantum computers use qubits.')).toBe(true);
expect(history.find(h => h.name === 'agent_researcher')?.content).toBe('');
});
it('regular (non-delegate) subagent keeps its own isolated history separate from the parent, across protos', async () => {
const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
const llm = new LLM(ai);
const history: any[] = [];
const summarizer = {name: 'summarizer', system: 'You summarize text.', model: 'gpt'};
(globalThis as any).__scripts.claude = () => ({calls: [{tool: 'subagent_summarizer', args: {context: 'a long article', instructions: 'summarize it'}}], text: 'Summary: short version'});
(globalThis as any).__scripts.gpt = () => ({text: 'short version'});
const resp = await llm.ask('summarize this', {model: 'claude', agents: [summarizer], history});
expect(resp).toBe('Summary: short version');
expect(history.find(h => h.name === 'subagent_summarizer')?.content).toBe('short version');
// isolated history - subagent's own assistant turn never leaks into the parent
expect(history.some(h => h.role === 'assistant' && h.content === 'short version')).toBe(false);
});
});

256
tests/memory.spec.ts Normal file
View File

@@ -0,0 +1,256 @@
import {describe, it, expect, vi, beforeEach} from 'vitest';
import {MemoryManager, MemoryCache, rebuildGraph, Memory} from '../src/memory';
function makeMemory(overrides: Partial<Memory> = {}): Memory {
return {
name: 'Test/Doc',
description: '',
content: '',
embedding: [],
links: [],
backlinks: [],
...overrides,
};
}
function makeLLM() {
return {
embedding: vi.fn(async (_text: string) => [{embedding: [1, 0, 0]}]),
ask: vi.fn(async () => undefined),
};
}
describe('rebuildGraph', () => {
it('extracts [[WikiLinks]] from content, excluding self-links', () => {
const a = makeMemory({name: 'A', content: '[[B]] and [[A]] and [[C]]'});
const b = makeMemory({name: 'B', content: 'no links here'});
const mem = [a, b];
rebuildGraph(mem);
expect(a.links).toEqual(['B', 'C']);
expect(b.links).toEqual([]);
});
it('computes backlinks only for links that resolve to a real node', () => {
const a = makeMemory({name: 'A', content: '[[B]] [[Missing]]'});
const b = makeMemory({name: 'B', content: ''});
const mem = [a, b];
rebuildGraph(mem);
expect(b.backlinks).toEqual(['A']);
expect(mem.find(m => m.name === 'Missing')).toBeUndefined();
});
it('resets stale backlinks on every rebuild (no leftover from a removed link)', () => {
const a = makeMemory({name: 'A', content: '[[B]]'});
const b = makeMemory({name: 'B', content: ''});
const mem = [a, b];
rebuildGraph(mem);
expect(b.backlinks).toEqual(['A']);
a.content = 'no more links';
rebuildGraph(mem);
expect(b.backlinks).toEqual([]);
});
});
describe('MemoryCache', () => {
it('finds nearest neighbor by embedding via KD-tree search', () => {
const close = makeMemory({name: 'Close', embedding: [1, 0, 0]});
const far = makeMemory({name: 'Far', embedding: [0, 0, 1]});
const cache = new MemoryCache([close, far]);
const results = cache.search([1, 0, 0], 1);
expect(results[0].name).toBe('Close');
});
it('rebuilds the tree on add/update/remove', () => {
const cache = new MemoryCache([makeMemory({name: 'A', embedding: [1, 0, 0]})]);
cache.add(makeMemory({name: 'B', embedding: [0, 1, 0]}));
expect(cache.search([0, 1, 0], 1)[0].name).toBe('B');
cache.remove('B');
expect(cache.search([0, 1, 0], 1)[0]?.name).not.toBe('B');
});
});
describe('MemoryManager.forget', () => {
it('removes the node and recomputes backlinks for the rest of the graph', () => {
const llm = makeLLM();
const mgr = new MemoryManager(llm);
const a = makeMemory({name: 'A', content: '[[B]]'});
const b = makeMemory({name: 'B', content: '[[C]]'});
const c = makeMemory({name: 'C', content: ''});
const mem = [a, b, c];
rebuildGraph(mem);
expect(c.backlinks).toEqual(['B']);
const ok = mgr.forget('B', mem);
expect(ok).toBe(true);
expect(mem.find(m => m.name === 'B')).toBeUndefined();
expect(a.links).toEqual(['B']);
expect(c.backlinks).toEqual([]);
});
it('returns false for an unknown name', () => {
const mgr = new MemoryManager(makeLLM());
expect(mgr.forget('Nope', [makeMemory({name: 'A'})])).toBe(false);
});
});
describe('MemoryManager.recollect', () => {
it('orders vector matches first, then expands one hop via links', async () => {
const llm = makeLLM();
llm.embedding.mockResolvedValue([{embedding: [1, 0, 0]}]);
const mgr = new MemoryManager(llm);
const near = makeMemory({name: 'Near', embedding: [1, 0, 0], content: '[[Linked]]'});
const linked = makeMemory({name: 'Linked', embedding: [0, 0, 1], content: ''});
const far = makeMemory({name: 'Far', embedding: [0, 1, 0], content: ''});
const mem = [near, linked, far];
rebuildGraph(mem);
const result = await mgr.recollect('query', mem, 1, 1);
expect(result.map(r => r.name)).toEqual(['Near', 'Linked']);
});
it('returns [] when there are no memories', async () => {
const mgr = new MemoryManager(makeLLM());
expect(await mgr.recollect('q', [])).toEqual([]);
});
});
describe('MemoryManager.memorize (fast path)', () => {
let llm: ReturnType<typeof makeLLM>;
let mgr: MemoryManager;
beforeEach(() => {
llm = makeLLM();
mgr = new MemoryManager(llm);
});
it('pushes a pending tool message, then resolves it to links once facts land', async () => {
llm.ask.mockImplementation(async (_prompt: string, opts: any) => {
if (opts.tools) {
opts.tools[0].fn({destination: 'Projects/Oxide', facts: 'Uses a hybrid memory system'});
return undefined;
}
return {description: 'd', content: '# doc'};
});
const history: any[] = [{role: 'user', content: 'we use a hybrid memory system'}];
const touched = await mgr.memorize(history, [], {model: 'test'} as any);
const pending = history.find(h => h.name === 'memory_process');
expect(pending).toBeDefined();
expect(pending.content).toContain('[[Projects/Oxide]]');
expect(touched.map(t => t.name)).toEqual(['Projects/Oxide']);
});
it('creates a new node and appends facts under "## Facts" without calling the doc LLM', async () => {
llm.ask.mockImplementation(async (_prompt: string, opts: any) => {
if (opts.tools) opts.tools[0].fn({destination: 'People/Sarah', facts: 'Works at Acme, Likes hiking'});
return undefined;
});
const mem: Memory[] = [];
await mgr.memorize([{role: 'user', content: 'Sarah works at Acme and likes hiking'}] as any, mem, {model: 'test'} as any);
const node = mem.find(m => m.name === 'People/Sarah')!;
expect(node).toBeDefined();
expect(node.content).toContain('## Facts');
expect(node.content).toContain('- Works at Acme');
expect(node.content).toContain('- Likes hiking');
// doc reconciler LLM (schema call) should NOT have been awaited synchronously in this fast path assertion
});
it('routes "journal" destination to Journal/{weekMonday}', async () => {
llm.ask.mockImplementation(async (_prompt: string, opts: any) => {
if (opts.tools) opts.tools[0].fn({destination: 'journal', facts: 'Shipped v1'});
return undefined;
});
const mem: Memory[] = [];
const touched = await mgr.memorize([{role: 'user', content: 'shipped v1 today'}] as any, mem, {model: 'test'} as any);
expect(touched[0].name).toMatch(/^Journal\/\d{4}-\d{2}-\d{2}$/);
});
it('reports nothing to remember when no facts are extracted', async () => {
llm.ask.mockResolvedValue(undefined); // tools present but fn never called
const history: any[] = [{role: 'user', content: 'hey'}];
const touched = await mgr.memorize(history, [], {model: 'test'} as any);
expect(touched).toEqual([]);
expect(history.find(h => h.name === 'memory_process').content).toBe('Nothing worth remembering.');
});
it('returns [] and does nothing for an empty conversation', async () => {
const touched = await mgr.memorize([], [], {model: 'test'} as any);
expect(touched).toEqual([]);
expect(llm.ask).not.toHaveBeenCalled();
});
});
describe('MemoryManager reconcileVault', () => {
it('integrates the "## Facts" section via the doc LLM and removes it', async () => {
const llm = makeLLM();
llm.ask.mockResolvedValue({description: 'Tidy summary', content: '# Doc\n\nIntegrated fact.'});
const mgr = new MemoryManager(llm);
const node = makeMemory({
name: 'Projects/Oxide',
content: '---\nname: Projects/Oxide\n---\n\n# Doc\n\n## Facts\n- some raw fact\n',
});
const mem = [node];
await mgr.reconcileVault(mem, {model: 'test'} as any, 'all');
expect(node.content).not.toContain('## Facts');
expect(node.content).toContain('Integrated fact.');
expect(node.description).toBe('Tidy summary');
});
it('only targets docs with a pending Facts inbox when scope is "touched"', async () => {
const llm = makeLLM();
llm.ask.mockResolvedValue({description: 'd', content: '# clean'});
const mgr = new MemoryManager(llm);
const dirty = makeMemory({name: 'A', content: '## Facts\n- x'});
const clean = makeMemory({name: 'B', content: '# already tidy'});
await mgr.reconcileVault([dirty, clean], {model: 'test'} as any, 'touched');
expect(dirty.content).toContain('# clean'); // rewritten (frontmatter now wraps it)
expect(clean.content).toBe('# already tidy'); // untouched, never queued
});
});
describe('MemoryManager reconcile coalescing', () => {
it('coalesces a second call while one is in-flight: marks dirty, aborts, reuses the same task promise', () => {
const llm = makeLLM();
const abort = vi.fn();
let calls = 0;
llm.ask.mockImplementation(() => {
calls++;
const pending: any = new Promise(() => {}); // never resolves in this test
pending.abort = abort;
return pending;
});
const mgr: any = new MemoryManager(llm);
const node = makeMemory({name: 'Q', content: '# Q\n\n## Facts\n- f'});
const mem = [node];
const p1 = mgr.reconcile(node, mem, {model: 'test'});
const p2 = mgr.reconcile(node, mem, {model: 'test'});
expect(p2).toBe(p1); // same in-flight task, not a new queue entry
expect(abort).toHaveBeenCalledTimes(1); // second call aborted the in-flight request
expect(calls).toBe(1); // no second ask() fired synchronously — it'll rerun via the dirty loop
});
});