getcompanion-ai/co-mono
1
0
Fork 0
mirror of https://github.com/getcompanion-ai/co-mono.git synced 2026-04-15 07:04:45 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Simplify compaction: remove proactive abort, use Agent.continue() for retry

Browse source 
- Add agentLoopContinue() to pi-ai for resuming from existing context
- Add Agent.continue() method and transport.continue() interface
- Simplify AgentSession compaction to two cases: overflow (auto-retry) and threshold (no retry)
- Remove proactive mid-turn compaction abort
- Merge turn prefix summary into main summary
- Add isCompacting property to AgentSession and RPC state
- Block input during compaction in interactive mode
- Show compaction count on session resume
- Rename RPC.md to rpc.md for consistency

Related to #128
This commit is contained in:
Mario Zechner 2025-12-09 21:43:49 +01:00
parent d67c69c6e9
commit 5a9d844f9a
27 changed files with 1261 additions and 1011 deletions
Download patch file Download diff file Expand all files Collapse all files

725
package-lock.json generated
View file

@ -11,6 +11,9 @@
"packages/*",
"packages/web-ui/example"
],
"dependencies": {
"get-east-asian-width": "^1.4.0"
},
"devDependencies": {
"@biomejs/biome": "2.3.5",
"@types/node": "^22.10.5",
@ -758,6 +761,29 @@
"node": ">=12"
}
},
"node_modules/@isaacs/cliui/node_modules/emoji-regex": {
"version": "9.2.2",
"resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-9.2.2.tgz",
"integrity": "sha512-L18DaJsXSUk2+42pv8mLs5jJT2hqFkFE4j21wOmgbUqsZ2hL72NsUU785g9RXgo3s0ZNgVl42TiHp3ZtOv/Vyg==",
"license": "MIT"
},
"node_modules/@isaacs/cliui/node_modules/string-width": {
"version": "5.1.2",
"resolved": "https://registry.npmjs.org/string-width/-/string-width-5.1.2.tgz",
"integrity": "sha512-HnLOCR3vjcY8beoNLtcjZ5/nxn2afmME6lhrDrebokqMap+XbeW8n9TXpPDOqdGK5qcI3oT0GKTW6wC7EMiVqA==",
"license": "MIT",
"dependencies": {
"eastasianwidth": "^0.2.0",
"emoji-regex": "^9.2.2",
"strip-ansi": "^7.0.1"
},
"engines": {
"node": ">=12"
},
"funding": {
"url": "https://github.com/sponsors/sindresorhus"
}
},
"node_modules/@jridgewell/gen-mapping": {
"version": "0.3.13",
"resolved": "https://registry.npmjs.org/@jridgewell/gen-mapping/-/gen-mapping-0.3.13.tgz",
@ -1451,6 +1477,19 @@
"url": "https://opencollective.com/parcel"
}
},
"node_modules/@parcel/watcher/node_modules/detect-libc": {
"version": "1.0.3",
"resolved": "https://registry.npmjs.org/detect-libc/-/detect-libc-1.0.3.tgz",
"integrity": "sha512-pGjwhsmsp4kL2RTz08wcOlGN83otlqHeD/Z5T8GXZB+/YcpQ/dgo+lbU8ZsGxV0HIvqqxo9l7mqYwyYMD9bKDg==",
"dev": true,
"license": "Apache-2.0",
"bin": {
"detect-libc": "bin/detect-libc.js"
},
"engines": {
"node": ">=0.10"
}
},
"node_modules/@pkgjs/parseargs": {
"version": "0.11.0",
"resolved": "https://registry.npmjs.org/@pkgjs/parseargs/-/parseargs-0.11.0.tgz",
@ -2593,21 +2632,6 @@
"readable-stream": "^3.4.0"
}
},
"node_modules/bl/node_modules/readable-stream": {
"version": "3.6.2",
"resolved": "https://registry.npmjs.org/readable-stream/-/readable-stream-3.6.2.tgz",
"integrity": "sha512-9u/sniCrY3D5WdsERHzHE4G2YCXqoG5FTHUiCC4SIbr6XcLZBY05ya9EKjYek9O5xOAwjGq+1JdGBAS7Q9ScoA==",
"dev": true,
"license": "MIT",
"dependencies": {
"inherits": "^2.0.3",
"string_decoder": "^1.1.1",
"util-deprecate": "^1.0.1"
},
"engines": {
"node": ">= 6"
}
},
"node_modules/brace-expansion": {
"version": "2.0.2",
"resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-2.0.2.tgz",
@ -2798,13 +2822,6 @@
"url": "https://github.com/chalk/ansi-styles?sponsor=1"
}
},
"node_modules/cliui/node_modules/emoji-regex": {
"version": "8.0.0",
"resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
"integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
"dev": true,
"license": "MIT"
},
"node_modules/cliui/node_modules/string-width": {
"version": "4.2.3",
"resolved": "https://registry.npmjs.org/string-width/-/string-width-4.2.3.tgz",
@ -3062,16 +3079,12 @@
}
},
"node_modules/detect-libc": {
"version": "1.0.3",
"resolved": "https://registry.npmjs.org/detect-libc/-/detect-libc-1.0.3.tgz",
"integrity": "sha512-pGjwhsmsp4kL2RTz08wcOlGN83otlqHeD/Z5T8GXZB+/YcpQ/dgo+lbU8ZsGxV0HIvqqxo9l7mqYwyYMD9bKDg==",
"dev": true,
"version": "2.1.2",
"resolved": "https://registry.npmjs.org/detect-libc/-/detect-libc-2.1.2.tgz",
"integrity": "sha512-Btj2BOOO83o3WyH59e8MgXsxEQVcarkUOpEYrubB0urwnN10yQ364rsiByU11nZlqWYZm05i/of7io4mzihBtQ==",
"license": "Apache-2.0",
"bin": {
"detect-libc": "bin/detect-libc.js"
},
"engines": {
"node": ">=0.10"
"node": ">=8"
}
},
"node_modules/diff": {
@ -3122,9 +3135,9 @@
}
},
"node_modules/emoji-regex": {
"version": "9.2.2",
"resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-9.2.2.tgz",
"integrity": "sha512-L18DaJsXSUk2+42pv8mLs5jJT2hqFkFE4j21wOmgbUqsZ2hL72NsUU785g9RXgo3s0ZNgVl42TiHp3ZtOv/Vyg==",
"version": "8.0.0",
"resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
"integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
"license": "MIT"
},
"node_modules/end-of-stream": {
@ -3947,6 +3960,36 @@
"setimmediate": "^1.0.5"
}
},
"node_modules/jszip/node_modules/readable-stream": {
"version": "2.3.8",
"resolved": "https://registry.npmjs.org/readable-stream/-/readable-stream-2.3.8.tgz",
"integrity": "sha512-8p0AUk4XODgIewSi0l8Epjs+EVnWiK7NoDIEGU0HhE7+ZyY8D1IMY7odu5lRrFXGg71L15KG8QrPmum45RTtdA==",
"license": "MIT",
"dependencies": {
"core-util-is": "~1.0.0",
"inherits": "~2.0.3",
"isarray": "~1.0.0",
"process-nextick-args": "~2.0.0",
"safe-buffer": "~5.1.1",
"string_decoder": "~1.1.1",
"util-deprecate": "~1.0.1"
}
},
"node_modules/jszip/node_modules/safe-buffer": {
"version": "5.1.2",
"resolved": "https://registry.npmjs.org/safe-buffer/-/safe-buffer-5.1.2.tgz",
"integrity": "sha512-Gd2UZBJDkXlY7GbJxfsE8/nvKkUEU1G38c1siN6QP6a9PT9MmHB8GnpscSmMJSoF8LOIrt8ud/wPtojys4G6+g==",
"license": "MIT"
},
"node_modules/jszip/node_modules/string_decoder": {
"version": "1.1.1",
"resolved": "https://registry.npmjs.org/string_decoder/-/string_decoder-1.1.1.tgz",
"integrity": "sha512-n/ShnvDi6FHbbVfviro+WojiFzv+s8MPMHBczVePfUpDJLwoLT0ht1l4YwBCbi8pJAveEEdnkHyPyTP/mzRfwg==",
"license": "MIT",
"dependencies": {
"safe-buffer": "~5.1.0"
}
},
"node_modules/jwa": {
"version": "2.0.1",
"resolved": "https://registry.npmjs.org/jwa/-/jwa-2.0.1.tgz",
@ -4251,15 +4294,6 @@
"url": "https://opencollective.com/parcel"
}
},
"node_modules/lightningcss/node_modules/detect-libc": {
"version": "2.1.2",
"resolved": "https://registry.npmjs.org/detect-libc/-/detect-libc-2.1.2.tgz",
"integrity": "sha512-Btj2BOOO83o3WyH59e8MgXsxEQVcarkUOpEYrubB0urwnN10yQ364rsiByU11nZlqWYZm05i/of7io4mzihBtQ==",
"license": "Apache-2.0",
"engines": {
"node": ">=8"
}
},
"node_modules/lit": {
"version": "3.3.1",
"resolved": "https://registry.npmjs.org/lit/-/lit-3.3.1.tgz",
@ -4795,16 +4829,6 @@
"node": ">=10"
}
},
"node_modules/prebuild-install/node_modules/detect-libc": {
"version": "2.1.2",
"resolved": "https://registry.npmjs.org/detect-libc/-/detect-libc-2.1.2.tgz",
"integrity": "sha512-Btj2BOOO83o3WyH59e8MgXsxEQVcarkUOpEYrubB0urwnN10yQ364rsiByU11nZlqWYZm05i/of7io4mzihBtQ==",
"dev": true,
"license": "Apache-2.0",
"engines": {
"node": ">=8"
}
},
"node_modules/process-nextick-args": {
"version": "2.0.1",
"resolved": "https://registry.npmjs.org/process-nextick-args/-/process-nextick-args-2.0.1.tgz",
@ -4845,25 +4869,19 @@
}
},
"node_modules/readable-stream": {
"version": "2.3.8",
"resolved": "https://registry.npmjs.org/readable-stream/-/readable-stream-2.3.8.tgz",
"integrity": "sha512-8p0AUk4XODgIewSi0l8Epjs+EVnWiK7NoDIEGU0HhE7+ZyY8D1IMY7odu5lRrFXGg71L15KG8QrPmum45RTtdA==",
"version": "3.6.2",
"resolved": "https://registry.npmjs.org/readable-stream/-/readable-stream-3.6.2.tgz",
"integrity": "sha512-9u/sniCrY3D5WdsERHzHE4G2YCXqoG5FTHUiCC4SIbr6XcLZBY05ya9EKjYek9O5xOAwjGq+1JdGBAS7Q9ScoA==",
"dev": true,
"license": "MIT",
"dependencies": {
"core-util-is": "~1.0.0",
"inherits": "~2.0.3",
"isarray": "~1.0.0",
"process-nextick-args": "~2.0.0",
"safe-buffer": "~5.1.1",
"string_decoder": "~1.1.1",
"util-deprecate": "~1.0.1"
}
"inherits": "^2.0.3",
"string_decoder": "^1.1.1",
"util-deprecate": "^1.0.1"
},
"node_modules/readable-stream/node_modules/safe-buffer": {
"version": "5.1.2",
"resolved": "https://registry.npmjs.org/safe-buffer/-/safe-buffer-5.1.2.tgz",
"integrity": "sha512-Gd2UZBJDkXlY7GbJxfsE8/nvKkUEU1G38c1siN6QP6a9PT9MmHB8GnpscSmMJSoF8LOIrt8ud/wPtojys4G6+g==",
"license": "MIT"
"engines": {
"node": ">= 6"
}
},
"node_modules/require-directory": {
"version": "2.1.1",
@ -5203,32 +5221,26 @@
"license": "MIT"
},
"node_modules/string_decoder": {
"version": "1.1.1",
"resolved": "https://registry.npmjs.org/string_decoder/-/string_decoder-1.1.1.tgz",
"integrity": "sha512-n/ShnvDi6FHbbVfviro+WojiFzv+s8MPMHBczVePfUpDJLwoLT0ht1l4YwBCbi8pJAveEEdnkHyPyTP/mzRfwg==",
"version": "1.3.0",
"resolved": "https://registry.npmjs.org/string_decoder/-/string_decoder-1.3.0.tgz",
"integrity": "sha512-hkRX8U1WjJFd8LsDJ2yQ/wWWxaopEsABU1XfkM8A+j0+85JAGppt16cr1Whg6KIbb4okU6Mql6BOj+uup/wKeA==",
"dev": true,
"license": "MIT",
"dependencies": {
"safe-buffer": "~5.1.0"
"safe-buffer": "~5.2.0"
}
},
"node_modules/string_decoder/node_modules/safe-buffer": {
"version": "5.1.2",
"resolved": "https://registry.npmjs.org/safe-buffer/-/safe-buffer-5.1.2.tgz",
"integrity": "sha512-Gd2UZBJDkXlY7GbJxfsE8/nvKkUEU1G38c1siN6QP6a9PT9MmHB8GnpscSmMJSoF8LOIrt8ud/wPtojys4G6+g==",
"license": "MIT"
},
"node_modules/string-width": {
"version": "5.1.2",
"resolved": "https://registry.npmjs.org/string-width/-/string-width-5.1.2.tgz",
"integrity": "sha512-HnLOCR3vjcY8beoNLtcjZ5/nxn2afmME6lhrDrebokqMap+XbeW8n9TXpPDOqdGK5qcI3oT0GKTW6wC7EMiVqA==",
"version": "8.1.0",
"resolved": "https://registry.npmjs.org/string-width/-/string-width-8.1.0.tgz",
"integrity": "sha512-Kxl3KJGb/gxkaUMOjRsQ8IrXiGW75O4E3RPjFIINOVH8AMl2SQ/yWdTzWwF3FevIX9LcMAjJW+GRwAlAbTSXdg==",
"license": "MIT",
"dependencies": {
"eastasianwidth": "^0.2.0",
"emoji-regex": "^9.2.2",
"strip-ansi": "^7.0.1"
"get-east-asian-width": "^1.3.0",
"strip-ansi": "^7.1.0"
},
"engines": {
"node": ">=12"
"node": ">=20"
},
"funding": {
"url": "https://github.com/sponsors/sindresorhus"
@ -5258,12 +5270,6 @@
"node": ">=8"
}
},
"node_modules/string-width-cjs/node_modules/emoji-regex": {
"version": "8.0.0",
"resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
"integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
"license": "MIT"
},
"node_modules/string-width-cjs/node_modules/strip-ansi": {
"version": "6.0.1",
"resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-6.0.1.tgz",
@ -5430,21 +5436,6 @@
"node": ">=6"
}
},
"node_modules/tar-stream/node_modules/readable-stream": {
"version": "3.6.2",
"resolved": "https://registry.npmjs.org/readable-stream/-/readable-stream-3.6.2.tgz",
"integrity": "sha512-9u/sniCrY3D5WdsERHzHE4G2YCXqoG5FTHUiCC4SIbr6XcLZBY05ya9EKjYek9O5xOAwjGq+1JdGBAS7Q9ScoA==",
"dev": true,
"license": "MIT",
"dependencies": {
"inherits": "^2.0.3",
"string_decoder": "^1.1.1",
"util-deprecate": "^1.0.1"
},
"engines": {
"node": ">= 6"
}
},
"node_modules/tinybench": {
"version": "2.9.0",
"resolved": "https://registry.npmjs.org/tinybench/-/tinybench-2.9.0.tgz",
@ -6342,12 +6333,6 @@
"url": "https://github.com/chalk/ansi-styles?sponsor=1"
}
},
"node_modules/wrap-ansi-cjs/node_modules/emoji-regex": {
"version": "8.0.0",
"resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
"integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
"license": "MIT"
},
"node_modules/wrap-ansi-cjs/node_modules/string-width": {
"version": "4.2.3",
"resolved": "https://registry.npmjs.org/string-width/-/string-width-4.2.3.tgz",
@ -6374,6 +6359,29 @@
"node": ">=8"
}
},
"node_modules/wrap-ansi/node_modules/emoji-regex": {
"version": "9.2.2",
"resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-9.2.2.tgz",
"integrity": "sha512-L18DaJsXSUk2+42pv8mLs5jJT2hqFkFE4j21wOmgbUqsZ2hL72NsUU785g9RXgo3s0ZNgVl42TiHp3ZtOv/Vyg==",
"license": "MIT"
},
"node_modules/wrap-ansi/node_modules/string-width": {
"version": "5.1.2",
"resolved": "https://registry.npmjs.org/string-width/-/string-width-5.1.2.tgz",
"integrity": "sha512-HnLOCR3vjcY8beoNLtcjZ5/nxn2afmME6lhrDrebokqMap+XbeW8n9TXpPDOqdGK5qcI3oT0GKTW6wC7EMiVqA==",
"license": "MIT",
"dependencies": {
"eastasianwidth": "^0.2.0",
"emoji-regex": "^9.2.2",
"strip-ansi": "^7.0.1"
},
"engines": {
"node": ">=12"
},
"funding": {
"url": "https://github.com/sponsors/sindresorhus"
}
},
"node_modules/wrappy": {
"version": "1.0.2",
"resolved": "https://registry.npmjs.org/wrappy/-/wrappy-1.0.2.tgz",
@ -6463,13 +6471,6 @@
"node": ">=8"
}
},
"node_modules/yargs/node_modules/emoji-regex": {
"version": "8.0.0",
"resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
"integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
"dev": true,
"license": "MIT"
},
"node_modules/yargs/node_modules/string-width": {
"version": "4.2.3",
"resolved": "https://registry.npmjs.org/string-width/-/string-width-4.2.3.tgz",
@ -6522,8 +6523,8 @@
"version": "0.16.0",
"license": "MIT",
"dependencies": {
"@mariozechner/pi-ai": "^0.15.0",
"@mariozechner/pi-tui": "^0.15.0"
"@mariozechner/pi-ai": "^0.16.0",
"@mariozechner/pi-tui": "^0.16.0"
},
"devDependencies": {
"@types/node": "^24.3.0",
@ -6534,93 +6535,20 @@
"node": ">=20.0.0"
}
},
"packages/agent/node_modules/@mariozechner/pi-ai": {
"version": "0.15.0",
"resolved": "https://registry.npmjs.org/@mariozechner/pi-ai/-/pi-ai-0.15.0.tgz",
"integrity": "sha512-vCwDfL4DtIZ73+gnWngFDpi6e7yTzHnX21sWTkpWHOT86BRLVu9gWhUo9lEvQbwA8R15qO5rjRR5J6MRu34Tjw==",
"license": "MIT",
"dependencies": {
"@anthropic-ai/sdk": "0.71.2",
"@google/genai": "1.31.0",
"@sinclair/typebox": "^0.34.41",
"ajv": "^8.17.1",
"ajv-formats": "^3.0.1",
"chalk": "^5.6.2",
"openai": "6.10.0",
"partial-json": "^0.1.7",
"zod-to-json-schema": "^3.24.6"
},
"engines": {
"node": ">=20.0.0"
}
},
"packages/agent/node_modules/@mariozechner/pi-tui": {
"version": "0.15.0",
"resolved": "https://registry.npmjs.org/@mariozechner/pi-tui/-/pi-tui-0.15.0.tgz",
"integrity": "sha512-4kSCO1fbMmpsyro4F+v4hqkkkPBlM1t5LF8UE51AgcA9i/OwVX3Nq5tPgMJ5vvPgexhd8N9hgzFwlQfTzYOetg==",
"license": "MIT",
"dependencies": {
"@types/mime-types": "^2.1.4",
"chalk": "^5.5.0",
"marked": "^15.0.12",
"mime-types": "^3.0.1",
"string-width": "^8.1.0"
},
"engines": {
"node": ">=20.0.0"
}
},
"packages/agent/node_modules/@types/node": {
"version": "24.10.1",
"version": "24.10.2",
"resolved": "https://registry.npmjs.org/@types/node/-/node-24.10.2.tgz",
"integrity": "sha512-WOhQTZ4G8xZ1tjJTvKOpyEVSGgOTvJAfDK3FNFgELyaTpzhdgHVHeqW8V+UJvzF5BT+/B54T/1S2K6gd9c7bbA==",
"dev": true,
"license": "MIT",
"dependencies": {
"undici-types": "~7.16.0"
}
},
"packages/agent/node_modules/mime-db": {
"version": "1.54.0",
"resolved": "https://registry.npmjs.org/mime-db/-/mime-db-1.54.0.tgz",
"integrity": "sha512-aU5EJuIN2WDemCcAp2vFBfp/m4EAhWJnUNSSw0ixs7/kXbd6Pg64EmwJkNdFhB8aWt1sH2CTXrLxo/iAGV3oPQ==",
"license": "MIT",
"engines": {
"node": ">= 0.6"
}
},
"packages/agent/node_modules/mime-types": {
"version": "3.0.2",
"resolved": "https://registry.npmjs.org/mime-types/-/mime-types-3.0.2.tgz",
"integrity": "sha512-Lbgzdk0h4juoQ9fCKXW4by0UJqj+nOOrI9MJ1sSj4nI8aI2eo1qmvQEie4VD1glsS250n15LsWsYtCugiStS5A==",
"license": "MIT",
"dependencies": {
"mime-db": "^1.54.0"
},
"engines": {
"node": ">=18"
},
"funding": {
"type": "opencollective",
"url": "https://opencollective.com/express"
}
},
"packages/agent/node_modules/string-width": {
"version": "8.1.0",
"resolved": "https://registry.npmjs.org/string-width/-/string-width-8.1.0.tgz",
"integrity": "sha512-Kxl3KJGb/gxkaUMOjRsQ8IrXiGW75O4E3RPjFIINOVH8AMl2SQ/yWdTzWwF3FevIX9LcMAjJW+GRwAlAbTSXdg==",
"license": "MIT",
"dependencies": {
"get-east-asian-width": "^1.3.0",
"strip-ansi": "^7.1.0"
},
"engines": {
"node": ">=20"
},
"funding": {
"url": "https://github.com/sponsors/sindresorhus"
}
},
"packages/agent/node_modules/undici-types": {
"version": "7.16.0",
"resolved": "https://registry.npmjs.org/undici-types/-/undici-types-7.16.0.tgz",
"integrity": "sha512-Zz+aZWSj8LE6zoxD+xrjh4VfkIG8Ya6LvYkZqtUQGJPZjYl53ypCaUwWqo7eI0x66KBGeRo+mlBEkMSeSZ38Nw==",
"dev": true,
"license": "MIT"
},
@ -6649,7 +6577,9 @@
}
},
"packages/ai/node_modules/@types/node": {
"version": "24.10.1",
"version": "24.10.2",
"resolved": "https://registry.npmjs.org/@types/node/-/node-24.10.2.tgz",
"integrity": "sha512-WOhQTZ4G8xZ1tjJTvKOpyEVSGgOTvJAfDK3FNFgELyaTpzhdgHVHeqW8V+UJvzF5BT+/B54T/1S2K6gd9c7bbA==",
"dev": true,
"license": "MIT",
"dependencies": {
@ -6658,6 +6588,8 @@
},
"packages/ai/node_modules/undici-types": {
"version": "7.16.0",
"resolved": "https://registry.npmjs.org/undici-types/-/undici-types-7.16.0.tgz",
"integrity": "sha512-Zz+aZWSj8LE6zoxD+xrjh4VfkIG8Ya6LvYkZqtUQGJPZjYl53ypCaUwWqo7eI0x66KBGeRo+mlBEkMSeSZ38Nw==",
"dev": true,
"license": "MIT"
},
@ -6666,9 +6598,9 @@
"version": "0.16.0",
"license": "MIT",
"dependencies": {
"@mariozechner/pi-agent-core": "^0.15.0",
"@mariozechner/pi-ai": "^0.15.0",
"@mariozechner/pi-tui": "^0.15.0",
"@mariozechner/pi-agent-core": "^0.16.0",
"@mariozechner/pi-ai": "^0.16.0",
"@mariozechner/pi-tui": "^0.16.0",
"chalk": "^5.5.0",
"diff": "^8.0.2",
"glob": "^11.0.3"
@ -6686,55 +6618,6 @@
"node": ">=20.0.0"
}
},
"packages/coding-agent/node_modules/@mariozechner/pi-agent-core": {
"version": "0.15.0",
"resolved": "https://registry.npmjs.org/@mariozechner/pi-agent-core/-/pi-agent-core-0.15.0.tgz",
"integrity": "sha512-p1SosZkdYR2fGG/2/2RzUyl4kVkrjrJQ2VGgc0lQklBf/zlKBtkiMDKHGdEVDq3y3J8JnDUleqTZZEAlzMcdXQ==",
"license": "MIT",
"dependencies": {
"@mariozechner/pi-ai": "^0.15.0",
"@mariozechner/pi-tui": "^0.15.0"
},
"engines": {
"node": ">=20.0.0"
}
},
"packages/coding-agent/node_modules/@mariozechner/pi-ai": {
"version": "0.15.0",
"resolved": "https://registry.npmjs.org/@mariozechner/pi-ai/-/pi-ai-0.15.0.tgz",
"integrity": "sha512-vCwDfL4DtIZ73+gnWngFDpi6e7yTzHnX21sWTkpWHOT86BRLVu9gWhUo9lEvQbwA8R15qO5rjRR5J6MRu34Tjw==",
"license": "MIT",
"dependencies": {
"@anthropic-ai/sdk": "0.71.2",
"@google/genai": "1.31.0",
"@sinclair/typebox": "^0.34.41",
"ajv": "^8.17.1",
"ajv-formats": "^3.0.1",
"chalk": "^5.6.2",
"openai": "6.10.0",
"partial-json": "^0.1.7",
"zod-to-json-schema": "^3.24.6"
},
"engines": {
"node": ">=20.0.0"
}
},
"packages/coding-agent/node_modules/@mariozechner/pi-tui": {
"version": "0.15.0",
"resolved": "https://registry.npmjs.org/@mariozechner/pi-tui/-/pi-tui-0.15.0.tgz",
"integrity": "sha512-4kSCO1fbMmpsyro4F+v4hqkkkPBlM1t5LF8UE51AgcA9i/OwVX3Nq5tPgMJ5vvPgexhd8N9hgzFwlQfTzYOetg==",
"license": "MIT",
"dependencies": {
"@types/mime-types": "^2.1.4",
"chalk": "^5.5.0",
"marked": "^15.0.12",
"mime-types": "^3.0.1",
"string-width": "^8.1.0"
},
"engines": {
"node": ">=20.0.0"
}
},
"packages/coding-agent/node_modules/@types/node": {
"version": "24.10.2",
"resolved": "https://registry.npmjs.org/@types/node/-/node-24.10.2.tgz",
@ -6745,47 +6628,6 @@
"undici-types": "~7.16.0"
}
},
"packages/coding-agent/node_modules/mime-db": {
"version": "1.54.0",
"resolved": "https://registry.npmjs.org/mime-db/-/mime-db-1.54.0.tgz",
"integrity": "sha512-aU5EJuIN2WDemCcAp2vFBfp/m4EAhWJnUNSSw0ixs7/kXbd6Pg64EmwJkNdFhB8aWt1sH2CTXrLxo/iAGV3oPQ==",
"license": "MIT",
"engines": {
"node": ">= 0.6"
}
},
"packages/coding-agent/node_modules/mime-types": {
"version": "3.0.2",
"resolved": "https://registry.npmjs.org/mime-types/-/mime-types-3.0.2.tgz",
"integrity": "sha512-Lbgzdk0h4juoQ9fCKXW4by0UJqj+nOOrI9MJ1sSj4nI8aI2eo1qmvQEie4VD1glsS250n15LsWsYtCugiStS5A==",
"license": "MIT",
"dependencies": {
"mime-db": "^1.54.0"
},
"engines": {
"node": ">=18"
},
"funding": {
"type": "opencollective",
"url": "https://opencollective.com/express"
}
},
"packages/coding-agent/node_modules/string-width": {
"version": "8.1.0",
"resolved": "https://registry.npmjs.org/string-width/-/string-width-8.1.0.tgz",
"integrity": "sha512-Kxl3KJGb/gxkaUMOjRsQ8IrXiGW75O4E3RPjFIINOVH8AMl2SQ/yWdTzWwF3FevIX9LcMAjJW+GRwAlAbTSXdg==",
"license": "MIT",
"dependencies": {
"get-east-asian-width": "^1.3.0",
"strip-ansi": "^7.1.0"
},
"engines": {
"node": ">=20"
},
"funding": {
"url": "https://github.com/sponsors/sindresorhus"
}
},
"packages/coding-agent/node_modules/undici-types": {
"version": "7.16.0",
"resolved": "https://registry.npmjs.org/undici-types/-/undici-types-7.16.0.tgz",
@ -6799,8 +6641,8 @@
"license": "MIT",
"dependencies": {
"@anthropic-ai/sandbox-runtime": "^0.0.16",
"@mariozechner/pi-agent-core": "^0.15.0",
"@mariozechner/pi-ai": "^0.15.0",
"@mariozechner/pi-agent-core": "^0.16.0",
"@mariozechner/pi-ai": "^0.16.0",
"@sinclair/typebox": "^0.34.0",
"@slack/socket-mode": "^2.0.0",
"@slack/web-api": "^7.0.0",
@ -6819,106 +6661,20 @@
"node": ">=20.0.0"
}
},
"packages/mom/node_modules/@mariozechner/pi-agent-core": {
"version": "0.15.0",
"resolved": "https://registry.npmjs.org/@mariozechner/pi-agent-core/-/pi-agent-core-0.15.0.tgz",
"integrity": "sha512-p1SosZkdYR2fGG/2/2RzUyl4kVkrjrJQ2VGgc0lQklBf/zlKBtkiMDKHGdEVDq3y3J8JnDUleqTZZEAlzMcdXQ==",
"license": "MIT",
"dependencies": {
"@mariozechner/pi-ai": "^0.15.0",
"@mariozechner/pi-tui": "^0.15.0"
},
"engines": {
"node": ">=20.0.0"
}
},
"packages/mom/node_modules/@mariozechner/pi-ai": {
"version": "0.15.0",
"resolved": "https://registry.npmjs.org/@mariozechner/pi-ai/-/pi-ai-0.15.0.tgz",
"integrity": "sha512-vCwDfL4DtIZ73+gnWngFDpi6e7yTzHnX21sWTkpWHOT86BRLVu9gWhUo9lEvQbwA8R15qO5rjRR5J6MRu34Tjw==",
"license": "MIT",
"dependencies": {
"@anthropic-ai/sdk": "0.71.2",
"@google/genai": "1.31.0",
"@sinclair/typebox": "^0.34.41",
"ajv": "^8.17.1",
"ajv-formats": "^3.0.1",
"chalk": "^5.6.2",
"openai": "6.10.0",
"partial-json": "^0.1.7",
"zod-to-json-schema": "^3.24.6"
},
"engines": {
"node": ">=20.0.0"
}
},
"packages/mom/node_modules/@mariozechner/pi-tui": {
"version": "0.15.0",
"resolved": "https://registry.npmjs.org/@mariozechner/pi-tui/-/pi-tui-0.15.0.tgz",
"integrity": "sha512-4kSCO1fbMmpsyro4F+v4hqkkkPBlM1t5LF8UE51AgcA9i/OwVX3Nq5tPgMJ5vvPgexhd8N9hgzFwlQfTzYOetg==",
"license": "MIT",
"dependencies": {
"@types/mime-types": "^2.1.4",
"chalk": "^5.5.0",
"marked": "^15.0.12",
"mime-types": "^3.0.1",
"string-width": "^8.1.0"
},
"engines": {
"node": ">=20.0.0"
}
},
"packages/mom/node_modules/@types/node": {
"version": "24.10.1",
"version": "24.10.2",
"resolved": "https://registry.npmjs.org/@types/node/-/node-24.10.2.tgz",
"integrity": "sha512-WOhQTZ4G8xZ1tjJTvKOpyEVSGgOTvJAfDK3FNFgELyaTpzhdgHVHeqW8V+UJvzF5BT+/B54T/1S2K6gd9c7bbA==",
"dev": true,
"license": "MIT",
"dependencies": {
"undici-types": "~7.16.0"
}
},
"packages/mom/node_modules/mime-db": {
"version": "1.54.0",
"resolved": "https://registry.npmjs.org/mime-db/-/mime-db-1.54.0.tgz",
"integrity": "sha512-aU5EJuIN2WDemCcAp2vFBfp/m4EAhWJnUNSSw0ixs7/kXbd6Pg64EmwJkNdFhB8aWt1sH2CTXrLxo/iAGV3oPQ==",
"license": "MIT",
"engines": {
"node": ">= 0.6"
}
},
"packages/mom/node_modules/mime-types": {
"version": "3.0.2",
"resolved": "https://registry.npmjs.org/mime-types/-/mime-types-3.0.2.tgz",
"integrity": "sha512-Lbgzdk0h4juoQ9fCKXW4by0UJqj+nOOrI9MJ1sSj4nI8aI2eo1qmvQEie4VD1glsS250n15LsWsYtCugiStS5A==",
"license": "MIT",
"dependencies": {
"mime-db": "^1.54.0"
},
"engines": {
"node": ">=18"
},
"funding": {
"type": "opencollective",
"url": "https://opencollective.com/express"
}
},
"packages/mom/node_modules/string-width": {
"version": "8.1.0",
"resolved": "https://registry.npmjs.org/string-width/-/string-width-8.1.0.tgz",
"integrity": "sha512-Kxl3KJGb/gxkaUMOjRsQ8IrXiGW75O4E3RPjFIINOVH8AMl2SQ/yWdTzWwF3FevIX9LcMAjJW+GRwAlAbTSXdg==",
"license": "MIT",
"dependencies": {
"get-east-asian-width": "^1.3.0",
"strip-ansi": "^7.1.0"
},
"engines": {
"node": ">=20"
},
"funding": {
"url": "https://github.com/sponsors/sindresorhus"
}
},
"packages/mom/node_modules/undici-types": {
"version": "7.16.0",
"resolved": "https://registry.npmjs.org/undici-types/-/undici-types-7.16.0.tgz",
"integrity": "sha512-Zz+aZWSj8LE6zoxD+xrjh4VfkIG8Ya6LvYkZqtUQGJPZjYl53ypCaUwWqo7eI0x66KBGeRo+mlBEkMSeSZ38Nw==",
"dev": true,
"license": "MIT"
},
@ -6927,7 +6683,7 @@
"version": "0.16.0",
"license": "MIT",
"dependencies": {
"@mariozechner/pi-agent-core": "^0.15.0",
"@mariozechner/pi-agent-core": "^0.16.0",
"chalk": "^5.5.0"
},
"bin": {
@ -6938,96 +6694,6 @@
"node": ">=20.0.0"
}
},
"packages/pods/node_modules/@mariozechner/pi-agent-core": {
"version": "0.15.0",
"resolved": "https://registry.npmjs.org/@mariozechner/pi-agent-core/-/pi-agent-core-0.15.0.tgz",
"integrity": "sha512-p1SosZkdYR2fGG/2/2RzUyl4kVkrjrJQ2VGgc0lQklBf/zlKBtkiMDKHGdEVDq3y3J8JnDUleqTZZEAlzMcdXQ==",
"license": "MIT",
"dependencies": {
"@mariozechner/pi-ai": "^0.15.0",
"@mariozechner/pi-tui": "^0.15.0"
},
"engines": {
"node": ">=20.0.0"
}
},
"packages/pods/node_modules/@mariozechner/pi-ai": {
"version": "0.15.0",
"resolved": "https://registry.npmjs.org/@mariozechner/pi-ai/-/pi-ai-0.15.0.tgz",
"integrity": "sha512-vCwDfL4DtIZ73+gnWngFDpi6e7yTzHnX21sWTkpWHOT86BRLVu9gWhUo9lEvQbwA8R15qO5rjRR5J6MRu34Tjw==",
"license": "MIT",
"dependencies": {
"@anthropic-ai/sdk": "0.71.2",
"@google/genai": "1.31.0",
"@sinclair/typebox": "^0.34.41",
"ajv": "^8.17.1",
"ajv-formats": "^3.0.1",
"chalk": "^5.6.2",
"openai": "6.10.0",
"partial-json": "^0.1.7",
"zod-to-json-schema": "^3.24.6"
},
"engines": {
"node": ">=20.0.0"
}
},
"packages/pods/node_modules/@mariozechner/pi-tui": {
"version": "0.15.0",
"resolved": "https://registry.npmjs.org/@mariozechner/pi-tui/-/pi-tui-0.15.0.tgz",
"integrity": "sha512-4kSCO1fbMmpsyro4F+v4hqkkkPBlM1t5LF8UE51AgcA9i/OwVX3Nq5tPgMJ5vvPgexhd8N9hgzFwlQfTzYOetg==",
"license": "MIT",
"dependencies": {
"@types/mime-types": "^2.1.4",
"chalk": "^5.5.0",
"marked": "^15.0.12",
"mime-types": "^3.0.1",
"string-width": "^8.1.0"
},
"engines": {
"node": ">=20.0.0"
}
},
"packages/pods/node_modules/mime-db": {
"version": "1.54.0",
"resolved": "https://registry.npmjs.org/mime-db/-/mime-db-1.54.0.tgz",
"integrity": "sha512-aU5EJuIN2WDemCcAp2vFBfp/m4EAhWJnUNSSw0ixs7/kXbd6Pg64EmwJkNdFhB8aWt1sH2CTXrLxo/iAGV3oPQ==",
"license": "MIT",
"engines": {
"node": ">= 0.6"
}
},
"packages/pods/node_modules/mime-types": {
"version": "3.0.2",
"resolved": "https://registry.npmjs.org/mime-types/-/mime-types-3.0.2.tgz",
"integrity": "sha512-Lbgzdk0h4juoQ9fCKXW4by0UJqj+nOOrI9MJ1sSj4nI8aI2eo1qmvQEie4VD1glsS250n15LsWsYtCugiStS5A==",
"license": "MIT",
"dependencies": {
"mime-db": "^1.54.0"
},
"engines": {
"node": ">=18"
},
"funding": {
"type": "opencollective",
"url": "https://opencollective.com/express"
}
},
"packages/pods/node_modules/string-width": {
"version": "8.1.0",
"resolved": "https://registry.npmjs.org/string-width/-/string-width-8.1.0.tgz",
"integrity": "sha512-Kxl3KJGb/gxkaUMOjRsQ8IrXiGW75O4E3RPjFIINOVH8AMl2SQ/yWdTzWwF3FevIX9LcMAjJW+GRwAlAbTSXdg==",
"license": "MIT",
"dependencies": {
"get-east-asian-width": "^1.3.0",
"strip-ansi": "^7.1.0"
},
"engines": {
"node": ">=20"
},
"funding": {
"url": "https://github.com/sponsors/sindresorhus"
}
},
"packages/proxy": {
"name": "@mariozechner/pi-proxy",
"version": "0.16.0",
@ -7065,6 +6731,8 @@
},
"packages/tui/node_modules/mime-db": {
"version": "1.54.0",
"resolved": "https://registry.npmjs.org/mime-db/-/mime-db-1.54.0.tgz",
"integrity": "sha512-aU5EJuIN2WDemCcAp2vFBfp/m4EAhWJnUNSSw0ixs7/kXbd6Pg64EmwJkNdFhB8aWt1sH2CTXrLxo/iAGV3oPQ==",
"license": "MIT",
"engines": {
"node": ">= 0.6"
@ -7072,6 +6740,8 @@
},
"packages/tui/node_modules/mime-types": {
"version": "3.0.2",
"resolved": "https://registry.npmjs.org/mime-types/-/mime-types-3.0.2.tgz",
"integrity": "sha512-Lbgzdk0h4juoQ9fCKXW4by0UJqj+nOOrI9MJ1sSj4nI8aI2eo1qmvQEie4VD1glsS250n15LsWsYtCugiStS5A==",
"license": "MIT",
"dependencies": {
"mime-db": "^1.54.0"
@ -7084,30 +6754,14 @@
"url": "https://opencollective.com/express"
}
},
"packages/tui/node_modules/string-width": {
"version": "8.1.0",
"resolved": "https://registry.npmjs.org/string-width/-/string-width-8.1.0.tgz",
"integrity": "sha512-Kxl3KJGb/gxkaUMOjRsQ8IrXiGW75O4E3RPjFIINOVH8AMl2SQ/yWdTzWwF3FevIX9LcMAjJW+GRwAlAbTSXdg==",
"license": "MIT",
"dependencies": {
"get-east-asian-width": "^1.3.0",
"strip-ansi": "^7.1.0"
},
"engines": {
"node": ">=20"
},
"funding": {
"url": "https://github.com/sponsors/sindresorhus"
}
},
"packages/web-ui": {
"name": "@mariozechner/pi-web-ui",
"version": "0.16.0",
"license": "MIT",
"dependencies": {
"@lmstudio/sdk": "^1.5.0",
"@mariozechner/pi-ai": "^0.15.0",
"@mariozechner/pi-tui": "^0.15.0",
"@mariozechner/pi-ai": "^0.16.0",
"@mariozechner/pi-tui": "^0.16.0",
"docx-preview": "^0.3.7",
"jszip": "^3.10.1",
"lucide": "^0.544.0",
@ -7141,87 +6795,6 @@
"typescript": "^5.7.3",
"vite": "^7.1.6"
}
},
"packages/web-ui/example/node_modules/@mariozechner/pi-ai": {
"resolved": "packages/ai",
"link": true
},
"packages/web-ui/node_modules/@mariozechner/pi-ai": {
"version": "0.15.0",
"resolved": "https://registry.npmjs.org/@mariozechner/pi-ai/-/pi-ai-0.15.0.tgz",
"integrity": "sha512-vCwDfL4DtIZ73+gnWngFDpi6e7yTzHnX21sWTkpWHOT86BRLVu9gWhUo9lEvQbwA8R15qO5rjRR5J6MRu34Tjw==",
"license": "MIT",
"dependencies": {
"@anthropic-ai/sdk": "0.71.2",
"@google/genai": "1.31.0",
"@sinclair/typebox": "^0.34.41",
"ajv": "^8.17.1",
"ajv-formats": "^3.0.1",
"chalk": "^5.6.2",
"openai": "6.10.0",
"partial-json": "^0.1.7",
"zod-to-json-schema": "^3.24.6"
},
"engines": {
"node": ">=20.0.0"
}
},
"packages/web-ui/node_modules/@mariozechner/pi-tui": {
"version": "0.15.0",
"resolved": "https://registry.npmjs.org/@mariozechner/pi-tui/-/pi-tui-0.15.0.tgz",
"integrity": "sha512-4kSCO1fbMmpsyro4F+v4hqkkkPBlM1t5LF8UE51AgcA9i/OwVX3Nq5tPgMJ5vvPgexhd8N9hgzFwlQfTzYOetg==",
"license": "MIT",
"dependencies": {
"@types/mime-types": "^2.1.4",
"chalk": "^5.5.0",
"marked": "^15.0.12",
"mime-types": "^3.0.1",
"string-width": "^8.1.0"
},
"engines": {
"node": ">=20.0.0"
}
},
"packages/web-ui/node_modules/mime-db": {
"version": "1.54.0",
"resolved": "https://registry.npmjs.org/mime-db/-/mime-db-1.54.0.tgz",
"integrity": "sha512-aU5EJuIN2WDemCcAp2vFBfp/m4EAhWJnUNSSw0ixs7/kXbd6Pg64EmwJkNdFhB8aWt1sH2CTXrLxo/iAGV3oPQ==",
"license": "MIT",
"engines": {
"node": ">= 0.6"
}
},
"packages/web-ui/node_modules/mime-types": {
"version": "3.0.2",
"resolved": "https://registry.npmjs.org/mime-types/-/mime-types-3.0.2.tgz",
"integrity": "sha512-Lbgzdk0h4juoQ9fCKXW4by0UJqj+nOOrI9MJ1sSj4nI8aI2eo1qmvQEie4VD1glsS250n15LsWsYtCugiStS5A==",
"license": "MIT",
"dependencies": {
"mime-db": "^1.54.0"
},
"engines": {
"node": ">=18"
},
"funding": {
"type": "opencollective",
"url": "https://opencollective.com/express"
}
},
"packages/web-ui/node_modules/string-width": {
"version": "8.1.0",
"resolved": "https://registry.npmjs.org/string-width/-/string-width-8.1.0.tgz",
"integrity": "sha512-Kxl3KJGb/gxkaUMOjRsQ8IrXiGW75O4E3RPjFIINOVH8AMl2SQ/yWdTzWwF3FevIX9LcMAjJW+GRwAlAbTSXdg==",
"license": "MIT",
"dependencies": {
"get-east-asian-width": "^1.3.0",
"strip-ansi": "^7.1.0"
},
"engines": {
"node": ">=20"
},
"funding": {
"url": "https://github.com/sponsors/sindresorhus"
}
}
}
}

5
package.json
View file

@ -34,5 +34,8 @@
"engines": {
"node": ">=20.0.0"
},
"version": "0.0.3"
"version": "0.0.3",
"dependencies": {
"get-east-asian-width": "^1.4.0"
}
}

101
packages/agent/src/agent.ts
View file

@ -176,11 +176,6 @@ export class Agent {
throw new Error("No model configured");
}
// Set up running prompt tracking
this.runningPrompt = new Promise<void>((resolve) => {
this.resolveRunningPrompt = resolve;
});
// Build user message with attachments
const content: Array<TextContent | ImageContent> = [{ type: "text", text: input }];
if (attachments?.length) {
@ -204,6 +199,62 @@ export class Agent {
timestamp: Date.now(),
};
await this._runAgentLoop(userMessage);
}
/**
* Continue from the current context without adding a new user message.
* Used for retry after overflow recovery when context already has user message or tool results.
*/
async continue() {
const messages = this._state.messages;
if (messages.length === 0) {
throw new Error("No messages to continue from");
}
const lastMessage = messages[messages.length - 1];
if (lastMessage.role !== "user" && lastMessage.role !== "toolResult") {
throw new Error(`Cannot continue from message role: ${lastMessage.role}`);
}
await this._runAgentLoopContinue();
}
/**
* Internal: Run the agent loop with a new user message.
*/
private async _runAgentLoop(userMessage: AppMessage) {
const { llmMessages, cfg } = await this._prepareRun();
const events = this.transport.run(llmMessages, userMessage as Message, cfg, this.abortController!.signal);
await this._processEvents(events);
}
/**
* Internal: Continue the agent loop from current context.
*/
private async _runAgentLoopContinue() {
const { llmMessages, cfg } = await this._prepareRun();
const events = this.transport.continue(llmMessages, cfg, this.abortController!.signal);
await this._processEvents(events);
}
/**
* Prepare for running the agent loop.
*/
private async _prepareRun() {
const model = this._state.model;
if (!model) {
throw new Error("No model configured");
}
this.runningPrompt = new Promise<void>((resolve) => {
this.resolveRunningPrompt = resolve;
});
this.abortController = new AbortController();
this._state.isStreaming = true;
this._state.streamMessage = null;
@ -222,9 +273,7 @@ export class Agent {
model,
reasoning,
getQueuedMessages: async <T>() => {
// Return queued messages based on queue mode
if (this.queueMode === "one-at-a-time") {
// Return only first message
if (this.messageQueue.length > 0) {
const first = this.messageQueue[0];
this.messageQueue = this.messageQueue.slice(1);
@ -232,7 +281,6 @@ export class Agent {
}
return [];
} else {
// Return all queued messages at once
const queued = this.messageQueue.slice();
this.messageQueue = [];
return queued as QueuedMessage<T>[];
@ -240,32 +288,30 @@ export class Agent {
},
};
// Track all messages generated in this prompt
const generatedMessages: AppMessage[] = [];
try {
let partial: Message | null = null;
// Transform app messages to LLM-compatible messages (initial set)
const llmMessages = await this.messageTransformer(this._state.messages);
for await (const ev of this.transport.run(
llmMessages,
userMessage as Message,
cfg,
this.abortController.signal,
)) {
// Update internal state BEFORE emitting events
// so handlers see consistent state
return { llmMessages, cfg, model };
}
/**
* Process events from the transport.
*/
private async _processEvents(events: AsyncIterable<AgentEvent>) {
const model = this._state.model!;
const generatedMessages: AppMessage[] = [];
let partial: AppMessage | null = null;
try {
for await (const ev of events) {
switch (ev.type) {
case "message_start": {
partial = ev.message;
this._state.streamMessage = ev.message;
partial = ev.message as AppMessage;
this._state.streamMessage = ev.message as Message;
break;
}
case "message_update": {
partial = ev.message;
this._state.streamMessage = ev.message;
partial = ev.message as AppMessage;
this._state.streamMessage = ev.message as Message;
break;
}
case "message_end": {
@ -299,7 +345,6 @@ export class Agent {
}
}
// Emit after state is updated
this.emit(ev as AgentEvent);
}

49
packages/agent/src/transports/AppTransport.ts
View file

@ -11,7 +11,7 @@ import type {
ToolCall,
UserMessage,
} from "@mariozechner/pi-ai";
import { agentLoop } from "@mariozechner/pi-ai";
import { agentLoop, agentLoopContinue } from "@mariozechner/pi-ai";
import { AssistantMessageEventStream } from "@mariozechner/pi-ai/dist/utils/event-stream.js";
import { parseStreamingJson } from "@mariozechner/pi-ai/dist/utils/json-parse.js";
import type { ProxyAssistantMessageEvent } from "./proxy-types.js";
@ -335,14 +335,8 @@ export class AppTransport implements AgentTransport {
this.options = options;
}
async *run(messages: Message[], userMessage: Message, cfg: AgentRunConfig, signal?: AbortSignal) {
const authToken = await this.options.getAuthToken();
if (!authToken) {
throw new Error("Auth token is required for AppTransport");
}
// Use proxy - no local API key needed
const streamFn = <TApi extends Api>(model: Model<TApi>, context: Context, options?: SimpleStreamOptions) => {
private async getStreamFn(authToken: string) {
return <TApi extends Api>(model: Model<TApi>, context: Context, options?: SimpleStreamOptions) => {
return streamSimpleProxy(
model,
context,
@ -353,24 +347,51 @@ export class AppTransport implements AgentTransport {
this.options.proxyUrl,
);
};
}
// Messages are already LLM-compatible (filtered by Agent)
const context: AgentContext = {
private buildContext(messages: Message[], cfg: AgentRunConfig): AgentContext {
return {
systemPrompt: cfg.systemPrompt,
messages,
tools: cfg.tools,
};
}
const pc: AgentLoopConfig = {
private buildLoopConfig(cfg: AgentRunConfig): AgentLoopConfig {
return {
model: cfg.model,
reasoning: cfg.reasoning,
getQueuedMessages: cfg.getQueuedMessages,
};
}
async *run(messages: Message[], userMessage: Message, cfg: AgentRunConfig, signal?: AbortSignal) {
const authToken = await this.options.getAuthToken();
if (!authToken) {
throw new Error("Auth token is required for AppTransport");
}
const streamFn = await this.getStreamFn(authToken);
const context = this.buildContext(messages, cfg);
const pc = this.buildLoopConfig(cfg);
// Yield events from the upstream agentLoop iterator
// Pass streamFn as the 5th parameter to use proxy
for await (const ev of agentLoop(userMessage as unknown as UserMessage, context, pc, signal, streamFn as any)) {
yield ev;
}
}
async *continue(messages: Message[], cfg: AgentRunConfig, signal?: AbortSignal) {
const authToken = await this.options.getAuthToken();
if (!authToken) {
throw new Error("Auth token is required for AppTransport");
}
const streamFn = await this.getStreamFn(authToken);
const context = this.buildContext(messages, cfg);
const pc = this.buildLoopConfig(cfg);
for await (const ev of agentLoopContinue(context, pc, signal, streamFn as any)) {
yield ev;
}
}
}

34
packages/agent/src/transports/ProviderTransport.ts
View file

@ -2,6 +2,7 @@ import {
type AgentContext,
type AgentLoopConfig,
agentLoop,
agentLoopContinue,
type Message,
type UserMessage,
} from "@mariozechner/pi-ai";
@ -33,18 +34,15 @@ export class ProviderTransport implements AgentTransport {
this.options = options;
}
async *run(messages: Message[], userMessage: Message, cfg: AgentRunConfig, signal?: AbortSignal) {
// Get API key
private async getModelAndKey(cfg: AgentRunConfig) {
let apiKey: string | undefined;
if (this.options.getApiKey) {
apiKey = await this.options.getApiKey(cfg.model.provider);
}
if (!apiKey) {
throw new Error(`No API key found for provider: ${cfg.model.provider}`);
}
// Clone model and modify baseUrl if CORS proxy is enabled
let model = cfg.model;
if (this.options.corsProxyUrl && cfg.model.baseUrl) {
model = {
@ -53,23 +51,43 @@ export class ProviderTransport implements AgentTransport {
};
}
// Messages are already LLM-compatible (filtered by Agent)
const context: AgentContext = {
return { model, apiKey };
}
private buildContext(messages: Message[], cfg: AgentRunConfig): AgentContext {
return {
systemPrompt: cfg.systemPrompt,
messages,
tools: cfg.tools,
};
}
const pc: AgentLoopConfig = {
private buildLoopConfig(model: typeof cfg.model, apiKey: string, cfg: AgentRunConfig): AgentLoopConfig {
return {
model,
reasoning: cfg.reasoning,
apiKey,
getQueuedMessages: cfg.getQueuedMessages,
};
}
async *run(messages: Message[], userMessage: Message, cfg: AgentRunConfig, signal?: AbortSignal) {
const { model, apiKey } = await this.getModelAndKey(cfg);
const context = this.buildContext(messages, cfg);
const pc = this.buildLoopConfig(model, apiKey, cfg);
// Yield events from agentLoop
for await (const ev of agentLoop(userMessage as unknown as UserMessage, context, pc, signal)) {
yield ev;
}
}
async *continue(messages: Message[], cfg: AgentRunConfig, signal?: AbortSignal) {
const { model, apiKey } = await this.getModelAndKey(cfg);
const context = this.buildContext(messages, cfg);
const pc = this.buildLoopConfig(model, apiKey, cfg);
for await (const ev of agentLoopContinue(context, pc, signal)) {
yield ev;
}
}
}

4
packages/agent/src/transports/types.ts
View file

@ -19,10 +19,14 @@ export interface AgentRunConfig {
* Events yielded must match the @mariozechner/pi-ai AgentEvent types.
*/
export interface AgentTransport {
/** Run with a new user message */
run(
messages: Message[],
userMessage: Message,
config: AgentRunConfig,
signal?: AbortSignal,
): AsyncIterable<AgentEvent>;
/** Continue from current context (no new user message) */
continue(messages: Message[], config: AgentRunConfig, signal?: AbortSignal): AsyncIterable<AgentEvent>;
}

260
packages/agent/test/e2e.test.ts
View file

@ -1,19 +1,11 @@
import type { Model } from "@mariozechner/pi-ai";
import type { AssistantMessage, Model, ToolResultMessage, UserMessage } from "@mariozechner/pi-ai";
import { calculateTool, getModel } from "@mariozechner/pi-ai";
import { describe, expect, it } from "vitest";
import { Agent, ProviderTransport } from "../src/index.js";
async function basicPrompt(model: Model<any>) {
const agent = new Agent({
initialState: {
systemPrompt: "You are a helpful assistant. Keep your responses concise.",
model,
thinkingLevel: "off",
tools: [],
},
transport: new ProviderTransport({
function createTransport() {
return new ProviderTransport({
getApiKey: async (provider) => {
// Map provider names to env var names
const envVarMap: Record<string, string> = {
google: "GEMINI_API_KEY",
openai: "OPENAI_API_KEY",
@ -26,7 +18,18 @@ async function basicPrompt(model: Model<any>) {
const envVar = envVarMap[provider] || `${provider.toUpperCase()}_API_KEY`;
return process.env[envVar];
},
}),
});
}
async function basicPrompt(model: Model<any>) {
const agent = new Agent({
initialState: {
systemPrompt: "You are a helpful assistant. Keep your responses concise.",
model,
thinkingLevel: "off",
tools: [],
},
transport: createTransport(),
});
await agent.prompt("What is 2+2? Answer with just the number.");
@ -54,22 +57,7 @@ async function toolExecution(model: Model<any>) {
thinkingLevel: "off",
tools: [calculateTool],
},
transport: new ProviderTransport({
getApiKey: async (provider) => {
// Map provider names to env var names
const envVarMap: Record<string, string> = {
google: "GEMINI_API_KEY",
openai: "OPENAI_API_KEY",
anthropic: "ANTHROPIC_API_KEY",
xai: "XAI_API_KEY",
groq: "GROQ_API_KEY",
cerebras: "CEREBRAS_API_KEY",
zai: "ZAI_API_KEY",
};
const envVar = envVarMap[provider] || `${provider.toUpperCase()}_API_KEY`;
return process.env[envVar];
},
}),
transport: createTransport(),
});
await agent.prompt("Calculate 123 * 456 using the calculator tool.");
@ -111,22 +99,7 @@ async function abortExecution(model: Model<any>) {
thinkingLevel: "off",
tools: [calculateTool],
},
transport: new ProviderTransport({
getApiKey: async (provider) => {
// Map provider names to env var names
const envVarMap: Record<string, string> = {
google: "GEMINI_API_KEY",
openai: "OPENAI_API_KEY",
anthropic: "ANTHROPIC_API_KEY",
xai: "XAI_API_KEY",
groq: "GROQ_API_KEY",
cerebras: "CEREBRAS_API_KEY",
zai: "ZAI_API_KEY",
};
const envVar = envVarMap[provider] || `${provider.toUpperCase()}_API_KEY`;
return process.env[envVar];
},
}),
transport: createTransport(),
});
const promptPromise = agent.prompt("Calculate 100 * 200, then 300 * 400, then sum the results.");
@ -156,22 +129,7 @@ async function stateUpdates(model: Model<any>) {
thinkingLevel: "off",
tools: [],
},
transport: new ProviderTransport({
getApiKey: async (provider) => {
// Map provider names to env var names
const envVarMap: Record<string, string> = {
google: "GEMINI_API_KEY",
openai: "OPENAI_API_KEY",
anthropic: "ANTHROPIC_API_KEY",
xai: "XAI_API_KEY",
groq: "GROQ_API_KEY",
cerebras: "CEREBRAS_API_KEY",
zai: "ZAI_API_KEY",
};
const envVar = envVarMap[provider] || `${provider.toUpperCase()}_API_KEY`;
return process.env[envVar];
},
}),
transport: createTransport(),
});
const events: Array<string> = [];
@ -204,22 +162,7 @@ async function multiTurnConversation(model: Model<any>) {
thinkingLevel: "off",
tools: [],
},
transport: new ProviderTransport({
getApiKey: async (provider) => {
// Map provider names to env var names
const envVarMap: Record<string, string> = {
google: "GEMINI_API_KEY",
openai: "OPENAI_API_KEY",
anthropic: "ANTHROPIC_API_KEY",
xai: "XAI_API_KEY",
groq: "GROQ_API_KEY",
cerebras: "CEREBRAS_API_KEY",
zai: "ZAI_API_KEY",
};
const envVar = envVarMap[provider] || `${provider.toUpperCase()}_API_KEY`;
return process.env[envVar];
},
}),
transport: createTransport(),
});
await agent.prompt("My name is Alice.");
@ -284,8 +227,8 @@ describe("Agent E2E Tests", () => {
});
});
describe.skipIf(!process.env.ANTHROPIC_API_KEY)("Anthropic Provider (claude-3-5-haiku-20241022)", () => {
const model = getModel("anthropic", "claude-3-5-haiku-20241022");
describe.skipIf(!process.env.ANTHROPIC_API_KEY)("Anthropic Provider (claude-haiku-4-5)", () => {
const model = getModel("anthropic", "claude-haiku-4-5");
it("should handle basic text prompt", async () => {
await basicPrompt(model);
@ -404,3 +347,164 @@ describe("Agent E2E Tests", () => {
});
});
});
describe("Agent.continue()", () => {
describe("validation", () => {
it("should throw when no messages in context", async () => {
const agent = new Agent({
initialState: {
systemPrompt: "Test",
model: getModel("anthropic", "claude-haiku-4-5"),
},
transport: createTransport(),
});
await expect(agent.continue()).rejects.toThrow("No messages to continue from");
});
it("should throw when last message is assistant", async () => {
const agent = new Agent({
initialState: {
systemPrompt: "Test",
model: getModel("anthropic", "claude-haiku-4-5"),
},
transport: createTransport(),
});
const assistantMessage: AssistantMessage = {
role: "assistant",
content: [{ type: "text", text: "Hello" }],
api: "anthropic-messages",
provider: "anthropic",
model: "claude-haiku-4-5",
usage: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
totalTokens: 0,
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
},
stopReason: "stop",
timestamp: Date.now(),
};
agent.replaceMessages([assistantMessage]);
await expect(agent.continue()).rejects.toThrow("Cannot continue from message role: assistant");
});
});
describe.skipIf(!process.env.ANTHROPIC_API_KEY)("continue from user message", () => {
const model = getModel("anthropic", "claude-haiku-4-5");
it("should continue and get response when last message is user", async () => {
const agent = new Agent({
initialState: {
systemPrompt: "You are a helpful assistant. Follow instructions exactly.",
model,
thinkingLevel: "off",
tools: [],
},
transport: createTransport(),
});
// Manually add a user message without calling prompt()
const userMessage: UserMessage = {
role: "user",
content: [{ type: "text", text: "Say exactly: HELLO WORLD" }],
timestamp: Date.now(),
};
agent.replaceMessages([userMessage]);
// Continue from the user message
await agent.continue();
expect(agent.state.isStreaming).toBe(false);
expect(agent.state.messages.length).toBe(2);
expect(agent.state.messages[0].role).toBe("user");
expect(agent.state.messages[1].role).toBe("assistant");
const assistantMsg = agent.state.messages[1] as AssistantMessage;
const textContent = assistantMsg.content.find((c) => c.type === "text");
expect(textContent).toBeDefined();
if (textContent?.type === "text") {
expect(textContent.text.toUpperCase()).toContain("HELLO WORLD");
}
});
});
describe.skipIf(!process.env.ANTHROPIC_API_KEY)("continue from tool result", () => {
const model = getModel("anthropic", "claude-haiku-4-5");
it("should continue and process tool results", async () => {
const agent = new Agent({
initialState: {
systemPrompt:
"You are a helpful assistant. After getting a calculation result, state the answer clearly.",
model,
thinkingLevel: "off",
tools: [calculateTool],
},
transport: createTransport(),
});
// Set up a conversation state as if tool was just executed
const userMessage: UserMessage = {
role: "user",
content: [{ type: "text", text: "What is 5 + 3?" }],
timestamp: Date.now(),
};
const assistantMessage: AssistantMessage = {
role: "assistant",
content: [
{ type: "text", text: "Let me calculate that." },
{ type: "toolCall", id: "calc-1", name: "calculate", arguments: { expression: "5 + 3" } },
],
api: "anthropic-messages",
provider: "anthropic",
model: "claude-haiku-4-5",
usage: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
totalTokens: 0,
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
},
stopReason: "toolUse",
timestamp: Date.now(),
};
const toolResult: ToolResultMessage = {
role: "toolResult",
toolCallId: "calc-1",
toolName: "calculate",
content: [{ type: "text", text: "5 + 3 = 8" }],
isError: false,
timestamp: Date.now(),
};
agent.replaceMessages([userMessage, assistantMessage, toolResult]);
// Continue from the tool result
await agent.continue();
expect(agent.state.isStreaming).toBe(false);
// Should have added an assistant response
expect(agent.state.messages.length).toBeGreaterThanOrEqual(4);
const lastMessage = agent.state.messages[agent.state.messages.length - 1];
expect(lastMessage.role).toBe("assistant");
if (lastMessage.role === "assistant") {
const textContent = lastMessage.content
.filter((c) => c.type === "text")
.map((c) => (c as { type: "text"; text: string }).text)
.join(" ");
// Should mention 8 in the response
expect(textContent).toMatch(/8/);
}
});
});
});

2
packages/agent/vitest.config.ts
View file

@ -4,6 +4,6 @@ export default defineConfig({
test: {
globals: true,
environment: "node",
testTimeout: 10000, // 10 seconds
testTimeout: 30000, // 30 seconds for API calls
},
});

4
packages/ai/CHANGELOG.md
View file

@ -2,6 +2,10 @@
## [Unreleased]
### Added
- **`agentLoopContinue` function**: Continue an agent loop from existing context without adding a new user message. Validates that the last message is `user` or `toolResult`. Useful for retry after context overflow or resuming from manually-added tool results.
### Breaking Changes
- Removed provider-level tool argument validation. Validation now happens in `agentLoop` via `executeToolCalls`, allowing models to retry on validation errors. For manual tool execution, use `validateToolCall(tools, toolCall)` or `validateToolArguments(tool, toolCall)`.

28
packages/ai/README.md
View file

@ -898,6 +898,34 @@ const messages = await stream.result();
context.messages.push(...messages);
```
### Continuing from Existing Context
Use `agentLoopContinue` to resume an agent loop without adding a new user message. This is useful for:
- Retrying after context overflow (after compaction reduces context size)
- Resuming from tool results that were added manually to the context
```typescript
import { agentLoopContinue, AgentContext } from '@mariozechner/pi-ai';
// Context already has messages - last must be 'user' or 'toolResult'
const context: AgentContext = {
systemPrompt: 'You are helpful.',
messages: [userMessage, assistantMessage, toolResult],
tools: [myTool]
};
// Continue processing from the tool result
const stream = agentLoopContinue(context, { model });
for await (const event of stream) {
// Same events as agentLoop, but no user message events emitted
}
const newMessages = await stream.result();
```
**Validation**: Throws if context has no messages or if the last message is an assistant message.
### Defining Tools with TypeBox
Tools use TypeBox schemas for runtime validation and type inference:

92
packages/ai/src/agent/agent-loop.ts
View file

@ -4,7 +4,10 @@ import { EventStream } from "../utils/event-stream.js";
import { validateToolArguments } from "../utils/validation.js";
import type { AgentContext, AgentEvent, AgentLoopConfig, AgentTool, AgentToolResult, QueuedMessage } from "./types.js";
// Main prompt function - returns a stream of events
/**
* Start an agent loop with a new user message.
* The prompt is added to the context and events are emitted for it.
*/
export function agentLoop(
prompt: UserMessage,
context: AgentContext,
@ -12,31 +15,80 @@ export function agentLoop(
signal?: AbortSignal,
streamFn?: typeof streamSimple,
): EventStream<AgentEvent, AgentContext["messages"]> {
const stream = new EventStream<AgentEvent, AgentContext["messages"]>(
(event: AgentEvent) => event.type === "agent_end",
(event: AgentEvent) => (event.type === "agent_end" ? event.messages : []),
);
const stream = createAgentStream();
// Run the prompt async
(async () => {
// Track new messages generated during this prompt
const newMessages: AgentContext["messages"] = [];
// Create user message for the prompt
const messages = [...context.messages, prompt];
newMessages.push(prompt);
const newMessages: AgentContext["messages"] = [prompt];
const currentContext: AgentContext = {
...context,
messages: [...context.messages, prompt],
};
stream.push({ type: "agent_start" });
stream.push({ type: "turn_start" });
stream.push({ type: "message_start", message: prompt });
stream.push({ type: "message_end", message: prompt });
// Update context with new messages
const currentContext: AgentContext = {
...context,
messages,
};
await runLoop(currentContext, newMessages, config, signal, stream, streamFn);
})();
// Keep looping while we have tool calls or queued messages
return stream;
}
/**
* Continue an agent loop from the current context without adding a new message.
* Used for retry after overflow - context already has user message or tool results.
* Throws if the last message is not a user message or tool result.
*/
export function agentLoopContinue(
context: AgentContext,
config: AgentLoopConfig,
signal?: AbortSignal,
streamFn?: typeof streamSimple,
): EventStream<AgentEvent, AgentContext["messages"]> {
// Validate that we can continue from this context
const lastMessage = context.messages[context.messages.length - 1];
if (!lastMessage) {
throw new Error("Cannot continue: no messages in context");
}
if (lastMessage.role !== "user" && lastMessage.role !== "toolResult") {
throw new Error(`Cannot continue from message role: ${lastMessage.role}. Expected 'user' or 'toolResult'.`);
}
const stream = createAgentStream();
(async () => {
const newMessages: AgentContext["messages"] = [];
const currentContext: AgentContext = { ...context };
stream.push({ type: "agent_start" });
stream.push({ type: "turn_start" });
// No user message events - we're continuing from existing context
await runLoop(currentContext, newMessages, config, signal, stream, streamFn);
})();
return stream;
}
function createAgentStream(): EventStream<AgentEvent, AgentContext["messages"]> {
return new EventStream<AgentEvent, AgentContext["messages"]>(
(event: AgentEvent) => event.type === "agent_end",
(event: AgentEvent) => (event.type === "agent_end" ? event.messages : []),
);
}
/**
* Shared loop logic for both agentLoop and agentLoopContinue.
*/
async function runLoop(
currentContext: AgentContext,
newMessages: AgentContext["messages"],
config: AgentLoopConfig,
signal: AbortSignal | undefined,
stream: EventStream<AgentEvent, AgentContext["messages"]>,
streamFn?: typeof streamSimple,
): Promise<void> {
let hasMoreToolCalls = true;
let firstTurn = true;
let queuedMessages: QueuedMessage<any>[] = (await config.getQueuedMessages?.()) || [];
@ -61,8 +113,6 @@ export function agentLoop(
queuedMessages = [];
}
// console.log("agent-loop: ", [...currentContext.messages]);
// Stream assistant response
const message = await streamAssistantResponse(currentContext, config, signal, stream, streamFn);
newMessages.push(message);
@ -91,11 +141,9 @@ export function agentLoop(
// Get queued messages after turn completes
queuedMessages = (await config.getQueuedMessages?.()) || [];
}
stream.push({ type: "agent_end", messages: newMessages });
stream.end(newMessages);
})();
return stream;
}
// Helper functions

2
packages/ai/src/agent/index.ts
View file

@ -1,3 +1,3 @@
export { agentLoop } from "./agent-loop.js";
export { agentLoop, agentLoopContinue } from "./agent-loop.js";
export * from "./tools/index.js";
export type { AgentContext, AgentEvent, AgentLoopConfig, AgentTool, QueuedMessage } from "./types.js";

12
packages/ai/src/models.generated.ts
View file

@ -3275,13 +3275,13 @@ export const MODELS = {
reasoning: true,
input: ["text"],
cost: {
input: 0.19999999999999998,
output: 0.7999999999999999,
input: 0.15,
output: 0.75,
cacheRead: 0,
cacheWrite: 0,
},
contextWindow: 163840,
maxTokens: 163840,
contextWindow: 8192,
maxTokens: 7168,
} satisfies Model<"openai-completions">,
"openai/gpt-4o-audio-preview": {
id: "openai/gpt-4o-audio-preview",
@ -4516,8 +4516,8 @@ export const MODELS = {
reasoning: false,
input: ["text", "image"],
cost: {
input: 0.049999999999999996,
output: 0.22,
input: 0.04,
output: 0.15,
cacheRead: 0,
cacheWrite: 0,
},

186
packages/ai/test/agent.test.ts
View file

@ -1,9 +1,17 @@
import { describe, expect, it } from "vitest";
import { agentLoop } from "../src/agent/agent-loop.js";
import { agentLoop, agentLoopContinue } from "../src/agent/agent-loop.js";
import { calculateTool } from "../src/agent/tools/calculate.js";
import type { AgentContext, AgentEvent, AgentLoopConfig } from "../src/agent/types.js";
import { getModel } from "../src/models.js";
import type { Api, Message, Model, OptionsForApi, UserMessage } from "../src/types.js";
import type {
Api,
AssistantMessage,
Message,
Model,
OptionsForApi,
ToolResultMessage,
UserMessage,
} from "../src/types.js";
async function calculateTest<TApi extends Api>(model: Model<TApi>, options: OptionsForApi<TApi> = {}) {
// Create the agent context with the calculator tool
@ -282,7 +290,7 @@ describe("Agent Calculator Tests", () => {
});
describe.skipIf(!process.env.ANTHROPIC_API_KEY)("Anthropic Provider Agent", () => {
const model = getModel("anthropic", "claude-3-5-haiku-20241022");
const model = getModel("anthropic", "claude-haiku-4-5");
it("should calculate multiple expressions and sum the results", async () => {
const result = await calculateTest(model);
@ -351,3 +359,175 @@ describe("Agent Calculator Tests", () => {
}, 30000);
});
});
describe("agentLoopContinue", () => {
describe("validation", () => {
const model = getModel("anthropic", "claude-haiku-4-5");
const baseContext: AgentContext = {
systemPrompt: "You are a helpful assistant.",
messages: [],
tools: [],
};
const config: AgentLoopConfig = { model };
it("should throw when context has no messages", () => {
expect(() => agentLoopContinue(baseContext, config)).toThrow("Cannot continue: no messages in context");
});
it("should throw when last message is an assistant message", () => {
const assistantMessage: AssistantMessage = {
role: "assistant",
content: [{ type: "text", text: "Hello" }],
api: "anthropic-messages",
provider: "anthropic",
model: "claude-haiku-4-5",
usage: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
totalTokens: 0,
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
},
stopReason: "stop",
timestamp: Date.now(),
};
const context: AgentContext = {
...baseContext,
messages: [assistantMessage],
};
expect(() => agentLoopContinue(context, config)).toThrow(
"Cannot continue from message role: assistant. Expected 'user' or 'toolResult'.",
);
});
// Note: "should not throw" tests for valid inputs are covered by the E2E tests below
// which actually consume the stream and verify the output
});
describe.skipIf(!process.env.ANTHROPIC_API_KEY)("continue from user message", () => {
const model = getModel("anthropic", "claude-haiku-4-5");
it("should continue and get assistant response when last message is user", async () => {
const userMessage: UserMessage = {
role: "user",
content: [{ type: "text", text: "Say exactly: HELLO WORLD" }],
timestamp: Date.now(),
};
const context: AgentContext = {
systemPrompt: "You are a helpful assistant. Follow instructions exactly.",
messages: [userMessage],
tools: [],
};
const config: AgentLoopConfig = { model };
const events: AgentEvent[] = [];
const stream = agentLoopContinue(context, config);
for await (const event of stream) {
events.push(event);
}
const messages = await stream.result();
// Should have gotten an assistant response
expect(messages.length).toBe(1);
expect(messages[0].role).toBe("assistant");
// Verify event sequence - no user message events since we're continuing
const eventTypes = events.map((e) => e.type);
expect(eventTypes).toContain("agent_start");
expect(eventTypes).toContain("turn_start");
expect(eventTypes).toContain("message_start");
expect(eventTypes).toContain("message_end");
expect(eventTypes).toContain("turn_end");
expect(eventTypes).toContain("agent_end");
// Should NOT have user message events (that's the difference from agentLoop)
const messageEndEvents = events.filter((e) => e.type === "message_end");
expect(messageEndEvents.length).toBe(1); // Only assistant message
expect((messageEndEvents[0] as any).message.role).toBe("assistant");
}, 30000);
});
describe.skipIf(!process.env.ANTHROPIC_API_KEY)("continue from tool result", () => {
const model = getModel("anthropic", "claude-haiku-4-5");
it("should continue processing after tool results", async () => {
// Simulate a conversation where:
// 1. User asked to calculate something
// 2. Assistant made a tool call
// 3. Tool result is ready
// 4. We continue from here
const userMessage: UserMessage = {
role: "user",
content: [{ type: "text", text: "What is 5 + 3? Use the calculator." }],
timestamp: Date.now(),
};
const assistantMessage: AssistantMessage = {
role: "assistant",
content: [
{ type: "text", text: "Let me calculate that for you." },
{ type: "toolCall", id: "calc-1", name: "calculate", arguments: { expression: "5 + 3" } },
],
api: "anthropic-messages",
provider: "anthropic",
model: "claude-haiku-4-5",
usage: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
totalTokens: 0,
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
},
stopReason: "toolUse",
timestamp: Date.now(),
};
const toolResult: ToolResultMessage = {
role: "toolResult",
toolCallId: "calc-1",
toolName: "calculate",
content: [{ type: "text", text: "5 + 3 = 8" }],
isError: false,
timestamp: Date.now(),
};
const context: AgentContext = {
systemPrompt: "You are a helpful assistant. After getting a calculation result, state the answer clearly.",
messages: [userMessage, assistantMessage, toolResult],
tools: [calculateTool],
};
const config: AgentLoopConfig = { model };
const events: AgentEvent[] = [];
const stream = agentLoopContinue(context, config);
for await (const event of stream) {
events.push(event);
}
const messages = await stream.result();
// Should have gotten an assistant response
expect(messages.length).toBeGreaterThanOrEqual(1);
const lastMessage = messages[messages.length - 1];
expect(lastMessage.role).toBe("assistant");
// The assistant should mention the result (8)
if (lastMessage.role === "assistant") {
const textContent = lastMessage.content
.filter((c) => c.type === "text")
.map((c) => (c as any).text)
.join(" ");
expect(textContent).toMatch(/8/);
}
}, 30000);
});
});

20
packages/coding-agent/CHANGELOG.md
View file

@ -2,6 +2,26 @@
## [Unreleased]
### Changed
- **Simplified compaction flow**: Removed proactive compaction (aborting mid-turn when threshold approached). Compaction now triggers in two cases only: (1) overflow error from LLM, which compacts and auto-retries, or (2) threshold crossed after a successful turn, which compacts without retry.
- **Compaction retry uses `Agent.continue()`**: Auto-retry after overflow now uses the new `continue()` API instead of re-sending the user message, preserving exact context state.
- **Merged turn prefix summary**: When a turn is split during compaction, the turn prefix summary is now merged into the main history summary instead of being stored separately.
### Added
- **`isCompacting` property on AgentSession**: Check if auto-compaction is currently running.
- **Session compaction indicator**: When resuming a compacted session, displays "Session compacted N times" status message.
### Fixed
- **Block input during compaction**: User input is now blocked while auto-compaction is running to prevent race conditions.
- **Skip error messages in usage calculation**: Context size estimation now skips both aborted and error messages, as neither have valid usage data.
## [0.16.0] - 2025-12-09
### Breaking Changes

6
packages/coding-agent/README.md
View file

@ -267,7 +267,11 @@ Long sessions can exhaust context windows. Compaction summarizes older messages
**Manual:** `/compact` or `/compact Focus on the API changes`
**Automatic:** Enable with `/autocompact`. Triggers when context exceeds threshold.
**Automatic:** Enable with `/autocompact`. When enabled, triggers in two cases:
- **Overflow recovery**: LLM returns context overflow error. Compacts and auto-retries.
- **Threshold maintenance**: Context exceeds `contextWindow - reserveTokens` after a successful turn. Compacts without retry.
When disabled, neither case triggers automatic compaction (use `/compact` manually if needed).
**How it works:**
1. Cut point calculated to keep ~20k tokens of recent messages

1
packages/coding-agent/docs/RPC.md → packages/coding-agent/docs/rpc.md
View file

@ -109,6 +109,7 @@ Response:
"model": {...},
"thinkingLevel": "medium",
"isStreaming": false,
"isCompacting": false,
"queueMode": "all",
"sessionFile": "/path/to/session.jsonl",
"sessionId": "abc123",

155
packages/coding-agent/src/core/agent-session.ts
View file

@ -14,11 +14,11 @@
*/
import type { Agent, AgentEvent, AgentState, AppMessage, Attachment, ThinkingLevel } from "@mariozechner/pi-agent-core";
import type { AssistantMessage, Model, ToolResultMessage } from "@mariozechner/pi-ai";
import type { AssistantMessage, Model } from "@mariozechner/pi-ai";
import { isContextOverflow } from "@mariozechner/pi-ai";
import { getModelsPath } from "../config.js";
import { type BashResult, executeBash as executeBashCommand } from "./bash-executor.js";
import { calculateContextTokens, compact, estimateTokens, shouldCompact } from "./compaction.js";
import { calculateContextTokens, compact, shouldCompact } from "./compaction.js";
import { exportSessionToHtml } from "./export-html.js";
import type { BashExecutionMessage } from "./messages.js";
import { getApiKeyForModel, getAvailableModels } from "./model-config.js";
@ -112,8 +112,6 @@ export class AgentSession {
// Compaction state
private _compactionAbortController: AbortController | null = null;
private _autoCompactionAbortController: AbortController | null = null;
private _abortingForCompaction = false;
private _lastUserMessageText: string | null = null;
// Bash execution state
private _bashAbortController: AbortController | null = null;
@ -148,48 +146,19 @@ export class AgentSession {
// Handle session persistence
if (event.type === "message_end") {
// Skip saving aborted message if we're aborting for compaction
const isAbortedForCompaction =
this._abortingForCompaction &&
event.message.role === "assistant" &&
(event.message as AssistantMessage).stopReason === "aborted";
if (!isAbortedForCompaction) {
this.sessionManager.saveMessage(event.message);
}
// Initialize session after first user+assistant exchange
if (this.sessionManager.shouldInitializeSession(this.agent.state.messages)) {
this.sessionManager.startSession(this.agent.state);
}
// Track user message text for potential retry after overflow
if (event.message.role === "user") {
const content = (event.message as { content: unknown }).content;
if (typeof content === "string") {
this._lastUserMessageText = content;
} else if (Array.isArray(content)) {
this._lastUserMessageText = content
.filter((c): c is { type: "text"; text: string } => c.type === "text")
.map((c) => c.text)
.join("\n");
}
}
// Track assistant message for auto-compaction (checked on agent_end)
if (event.message.role === "assistant") {
this._lastAssistantMessage = event.message as AssistantMessage;
this._lastAssistantMessage = event.message;
}
}
// Handle turn_end for proactive compaction check
if (event.type === "turn_end") {
await this._checkProactiveCompaction(
event.message as AssistantMessage,
event.toolResults as ToolResultMessage[],
);
}
// Check auto-compaction after agent completes
if (event.type === "agent_end" && this._lastAssistantMessage) {
const msg = this._lastAssistantMessage;
@ -274,6 +243,11 @@ export class AgentSession {
return this.agent.state.isStreaming;
}
/** Whether auto-compaction is currently running */
get isCompacting(): boolean {
return this._autoCompactionAbortController !== null || this._compactionAbortController !== null;
}
/** All messages including custom types like BashExecutionMessage */
get messages(): AppMessage[] {
return this.agent.state.messages;
@ -622,91 +596,41 @@ export class AgentSession {
this._autoCompactionAbortController?.abort();
}
/**
* Check for proactive compaction after turn_end (before next LLM call).
* Estimates context size and aborts if threshold would be crossed.
*/
private async _checkProactiveCompaction(
assistantMessage: AssistantMessage,
toolResults: ToolResultMessage[],
): Promise<void> {
const settings = this.settingsManager.getCompactionSettings();
if (!settings.enabled) return;
// Skip if message was aborted or errored
if (assistantMessage.stopReason === "aborted" || assistantMessage.stopReason === "error") return;
// Only check if there are tool calls (meaning another turn will happen)
const hasToolCalls = assistantMessage.content.some((c) => c.type === "toolCall");
if (!hasToolCalls) return;
// Estimate context size: last usage + tool results
const contextTokens = calculateContextTokens(assistantMessage.usage);
const toolResultTokens = toolResults.reduce((sum, msg) => sum + estimateTokens(msg), 0);
const estimatedTotal = contextTokens + toolResultTokens;
const contextWindow = this.model?.contextWindow ?? 0;
if (!shouldCompact(estimatedTotal, contextWindow, settings)) return;
// Threshold crossed - abort for compaction
this._abortingForCompaction = true;
this.agent.abort();
}
/**
* Handle compaction after agent_end.
* Checks for overflow (reactive) or threshold (proactive after abort).
* Two cases:
* 1. Overflow: LLM returned context overflow error, remove error message from agent state, compact, auto-retry
* 2. Threshold: Turn succeeded but context over threshold, compact, NO auto-retry (user continues manually)
*/
private async _handleAgentEndCompaction(assistantMessage: AssistantMessage): Promise<void> {
const settings = this.settingsManager.getCompactionSettings();
if (!settings.enabled) return;
// Skip if message was aborted (user cancelled)
if (assistantMessage.stopReason === "aborted") return;
const contextWindow = this.model?.contextWindow ?? 0;
// Check 1: Overflow detection (reactive recovery)
const isOverflow = isContextOverflow(assistantMessage, contextWindow);
// Check 2: Aborted for compaction (proactive)
const wasAbortedForCompaction = this._abortingForCompaction;
this._abortingForCompaction = false;
// Check 3: Threshold crossed but turn succeeded (maintenance compaction)
const contextTokens =
assistantMessage.stopReason === "error" ? 0 : calculateContextTokens(assistantMessage.usage);
const thresholdCrossed = settings.enabled && shouldCompact(contextTokens, contextWindow, settings);
// Determine which action to take
let reason: "overflow" | "threshold" | null = null;
let willRetry = false;
if (isOverflow) {
reason = "overflow";
willRetry = true;
// Remove the overflow error message from agent state
// Case 1: Overflow - LLM returned context overflow error
if (isContextOverflow(assistantMessage, contextWindow)) {
// Remove the error message from agent state (it IS saved to session for history,
// but we don't want it in context for the retry)
const messages = this.agent.state.messages;
if (messages.length > 0 && messages[messages.length - 1].role === "assistant") {
this.agent.replaceMessages(messages.slice(0, -1));
}
} else if (wasAbortedForCompaction) {
reason = "threshold";
willRetry = true;
// Remove the aborted message from agent state
const messages = this.agent.state.messages;
if (
messages.length > 0 &&
messages[messages.length - 1].role === "assistant" &&
(messages[messages.length - 1] as AssistantMessage).stopReason === "aborted"
) {
this.agent.replaceMessages(messages.slice(0, -1));
}
} else if (thresholdCrossed) {
reason = "threshold";
willRetry = false; // Turn succeeded, no retry needed
await this._runAutoCompaction("overflow", true);
return;
}
if (!reason) return;
// Case 2: Threshold - turn succeeded but context is getting large
// Skip if this was an error (non-overflow errors don't have usage data)
if (assistantMessage.stopReason === "error") return;
// Run compaction
await this._runAutoCompaction(reason, willRetry);
const contextTokens = calculateContextTokens(assistantMessage.usage);
if (shouldCompact(contextTokens, contextWindow, settings)) {
await this._runAutoCompaction("threshold", false);
}
}
/**
@ -754,11 +678,22 @@ export class AgentSession {
};
this._emit({ type: "auto_compaction_end", result, aborted: false, willRetry });
// Auto-retry if needed
if (willRetry && this._lastUserMessageText) {
// Small delay to let UI update
await new Promise((resolve) => setTimeout(resolve, 100));
await this.prompt(this._lastUserMessageText);
// Auto-retry if needed - use continue() since user message is already in context
if (willRetry) {
// Remove trailing error message from agent state (it's kept in session file for history)
// This is needed because continue() requires last message to be user or toolResult
const messages = this.agent.state.messages;
const lastMsg = messages[messages.length - 1];
if (lastMsg?.role === "assistant" && (lastMsg as AssistantMessage).stopReason === "error") {
this.agent.replaceMessages(messages.slice(0, -1));
}
// Use setTimeout to break out of the event handler chain
setTimeout(() => {
this.agent.continue().catch(() => {
// Retry failed - silently ignore, user can manually retry
});
}, 100);
}
} catch (error) {
// Compaction failed - emit end event without retry

116
packages/coding-agent/src/core/compaction.ts
View file

@ -41,11 +41,12 @@ export function calculateContextTokens(usage: Usage): number {
/**
* Get usage from an assistant message if available.
* Skips aborted and error messages as they don't have valid usage data.
*/
function getAssistantUsage(msg: AppMessage): Usage | null {
if (msg.role === "assistant" && "usage" in msg) {
const assistantMsg = msg as AssistantMessage;
if (assistantMsg.stopReason !== "aborted" && assistantMsg.usage) {
if (assistantMsg.stopReason !== "aborted" && assistantMsg.stopReason !== "error" && assistantMsg.usage) {
return assistantMsg.usage;
}
}
@ -81,36 +82,59 @@ export function shouldCompact(contextTokens: number, contextWindow: number, sett
/**
* Estimate token count for a message using chars/4 heuristic.
* This is conservative (overestimates tokens).
* Accepts any message type (AppMessage, ToolResultMessage, etc.)
*/
export function estimateTokens(message: {
role: string;
content?: unknown;
command?: string;
output?: string;
}): number {
export function estimateTokens(message: AppMessage): number {
let chars = 0;
// Handle custom message types that don't have standard content
// Handle bashExecution messages
if (message.role === "bashExecution") {
chars = (message.command?.length || 0) + (message.output?.length || 0);
const bash = message as unknown as { command: string; output: string };
chars = bash.command.length + bash.output.length;
return Math.ceil(chars / 4);
}
// Standard messages with content
const content = message.content;
// Handle user messages
if (message.role === "user") {
const content = (message as { content: string | Array<{ type: string; text?: string }> }).content;
if (typeof content === "string") {
chars = content.length;
} else if (Array.isArray(content)) {
for (const block of content) {
if (block.type === "text") {
if (block.type === "text" && block.text) {
chars += block.text.length;
} else if (block.type === "thinking") {
chars += block.thinking.length;
}
}
}
return Math.ceil(chars / 4);
}
// Handle assistant messages
if (message.role === "assistant") {
const assistant = message as AssistantMessage;
for (const block of assistant.content) {
if (block.type === "text") {
chars += block.text.length;
} else if (block.type === "thinking") {
chars += block.thinking.length;
} else if (block.type === "toolCall") {
chars += block.name.length + JSON.stringify(block.arguments).length;
}
}
return Math.ceil(chars / 4);
}
// Handle tool results
if (message.role === "toolResult") {
const toolResult = message as { content: Array<{ type: string; text?: string }> };
for (const block of toolResult.content) {
if (block.type === "text" && block.text) {
chars += block.text.length;
}
}
return Math.ceil(chars / 4);
}
return 0;
}
/**
@ -166,6 +190,9 @@ export interface CutPointResult {
/**
* Find the cut point in session entries that keeps approximately `keepRecentTokens`.
*
* Algorithm: Walk backwards from newest, accumulating estimated message sizes.
* Stop when we've accumulated >= keepRecentTokens. Cut at that point.
*
* Can cut at user OR assistant messages (never tool results). When cutting at an
* assistant message with tool calls, its tool results come after and will be kept.
*
@ -188,46 +215,23 @@ export function findCutPoint(
return { firstKeptEntryIndex: startIndex, turnStartIndex: -1, isSplitTurn: false };
}
// Collect assistant usages walking backwards from endIndex
const assistantUsages: Array<{ index: number; tokens: number }> = [];
for (let i = endIndex - 1; i >= startIndex; i--) {
const entry = entries[i];
if (entry.type === "message") {
const usage = getAssistantUsage(entry.message);
if (usage) {
assistantUsages.push({
index: i,
tokens: calculateContextTokens(usage),
});
}
}
}
if (assistantUsages.length === 0) {
// No usage info, keep from last cut point
const lastCutPoint = cutPoints[cutPoints.length - 1];
const entry = entries[lastCutPoint];
const isUser = entry.type === "message" && entry.message.role === "user";
return {
firstKeptEntryIndex: lastCutPoint,
turnStartIndex: isUser ? -1 : findTurnStartIndex(entries, lastCutPoint, startIndex),
isSplitTurn: !isUser,
};
}
// Walk through and find where cumulative token difference exceeds keepRecentTokens
const newestTokens = assistantUsages[0].tokens;
// Walk backwards from newest, accumulating estimated message sizes
let accumulatedTokens = 0;
let cutIndex = startIndex; // Default: keep everything in range
for (let i = 1; i < assistantUsages.length; i++) {
const tokenDiff = newestTokens - assistantUsages[i].tokens;
if (tokenDiff >= keepRecentTokens) {
// Find the valid cut point at or after the assistant we want to keep
const lastKeptAssistantIndex = assistantUsages[i - 1].index;
for (let i = endIndex - 1; i >= startIndex; i--) {
const entry = entries[i];
if (entry.type !== "message") continue;
// Find closest valid cut point at or before lastKeptAssistantIndex
for (let c = cutPoints.length - 1; c >= 0; c--) {
if (cutPoints[c] <= lastKeptAssistantIndex) {
// Estimate this message's size
const messageTokens = estimateTokens(entry.message);
accumulatedTokens += messageTokens;
// Check if we've exceeded the budget
if (accumulatedTokens >= keepRecentTokens) {
// Find the closest valid cut point at or after this entry
for (let c = 0; c < cutPoints.length; c++) {
if (cutPoints[c] >= i) {
cutIndex = cutPoints[c];
break;
}
@ -404,9 +408,8 @@ export async function compact(
}
}
// Generate summaries (can be parallel if both needed)
// Generate summaries (can be parallel if both needed) and merge into one
let summary: string;
let turnPrefixSummary: string | undefined;
if (cutResult.isSplitTurn && turnPrefixMessages.length > 0) {
// Generate both summaries in parallel
@ -416,8 +419,8 @@ export async function compact(
: Promise.resolve("No prior history."),
generateTurnPrefixSummary(turnPrefixMessages, model, settings.reserveTokens, apiKey, signal),
]);
summary = historyResult;
turnPrefixSummary = turnPrefixResult;
// Merge into single summary
summary = historyResult + "\n\n---\n\n**Turn Context (split turn):**\n\n" + turnPrefixResult;
} else {
// Just generate history summary
summary = await generateSummary(
@ -434,7 +437,6 @@ export async function compact(
type: "compaction",
timestamp: new Date().toISOString(),
summary,
turnPrefixSummary,
firstKeptEntryIndex: cutResult.firstKeptEntryIndex,
tokensBefore,
};

13
packages/coding-agent/src/core/session-manager.ts
View file

@ -50,8 +50,6 @@ export interface CompactionEntry {
type: "compaction";
timestamp: string;
summary: string;
/** Summary of turn prefix when a turn was split (user message to first kept message) */
turnPrefixSummary?: string;
firstKeptEntryIndex: number; // Index into session entries where we start keeping
tokensBefore: number;
}
@ -180,18 +178,9 @@ export function loadSessionFromEntries(entries: SessionEntry[]): LoadedSession {
}
}
// Build final messages: summaries + kept messages
// Build final messages: summary + kept messages
const messages: AppMessage[] = [];
// Add history summary
messages.push(createSummaryMessage(compactionEvent.summary));
// Add turn prefix summary if present (when a turn was split)
if (compactionEvent.turnPrefixSummary) {
messages.push(createSummaryMessage(compactionEvent.turnPrefixSummary));
}
// Add kept messages
messages.push(...keptMessages);
return { messages, thinkingLevel, model };

17
packages/coding-agent/src/modes/interactive/interactive-mode.ts
View file

@ -407,6 +407,11 @@ export class InteractiveMode {
}
}
// Block input during compaction (will retry automatically)
if (this.session.isCompacting) {
return;
}
// Queue message if agent is streaming
if (this.session.isStreaming) {
await this.session.queueMessage(text);
@ -604,10 +609,6 @@ export class InteractiveMode {
compactionComponent.setExpanded(this.toolOutputExpanded);
this.chatContainer.addChild(compactionComponent);
this.footer.updateState(this.session.state);
if (event.willRetry) {
this.showStatus("Compacted context, retrying...");
}
}
this.ui.requestRender();
break;
@ -743,6 +744,14 @@ export class InteractiveMode {
renderInitialMessages(state: AgentState): void {
this.renderMessages(state.messages, { updateFooter: true, populateHistory: true });
// Show compaction info if session was compacted
const entries = this.sessionManager.loadEntries();
const compactionCount = entries.filter((e) => e.type === "compaction").length;
if (compactionCount > 0) {
const times = compactionCount === 1 ? "1 time" : `${compactionCount} times`;
this.showStatus(`Session compacted ${times}`);
}
}
async getUserInput(): Promise<string> {

1
packages/coding-agent/src/modes/rpc/rpc-mode.ts
View file

@ -90,6 +90,7 @@ export async function runRpcMode(session: AgentSession): Promise<never> {
model: session.model,
thinkingLevel: session.thinkingLevel,
isStreaming: session.isStreaming,
isCompacting: session.isCompacting,
queueMode: session.queueMode,
sessionFile: session.sessionFile,
sessionId: session.sessionId,

1
packages/coding-agent/src/modes/rpc/rpc-types.ts
View file

@ -63,6 +63,7 @@ export interface RpcSessionState {
model: Model<any> | null;
thinkingLevel: ThinkingLevel;
isStreaming: boolean;
isCompacting: boolean;
queueMode: "all" | "one-at-a-time";
sessionFile: string;
sessionId: string;

233
packages/coding-agent/test/agent-session-compaction.test.ts Normal file
View file

@ -0,0 +1,233 @@
/**
* E2E tests for AgentSession compaction behavior.
*
* These tests use real LLM calls (no mocking) to verify:
* - Manual compaction works correctly
* - Session persistence during compaction
* - Compaction entry is saved to session file
*/
import { existsSync, mkdirSync, rmSync } from "node:fs";
import { tmpdir } from "node:os";
import { join } from "node:path";
import { Agent, ProviderTransport } from "@mariozechner/pi-agent-core";
import { getModel } from "@mariozechner/pi-ai";
import { afterEach, beforeEach, describe, expect, it } from "vitest";
import { AgentSession, type AgentSessionEvent } from "../src/core/agent-session.js";
import { SessionManager } from "../src/core/session-manager.js";
import { SettingsManager } from "../src/core/settings-manager.js";
import { codingTools } from "../src/core/tools/index.js";
const API_KEY = process.env.ANTHROPIC_API_KEY || process.env.ANTHROPIC_OAUTH_TOKEN;
describe.skipIf(!API_KEY)("AgentSession compaction e2e", () => {
let session: AgentSession;
let tempDir: string;
let sessionManager: SessionManager;
let events: AgentSessionEvent[];
beforeEach(() => {
// Create temp directory for session files
tempDir = join(tmpdir(), `pi-compaction-test-${Date.now()}`);
mkdirSync(tempDir, { recursive: true });
// Track events
events = [];
});
afterEach(async () => {
if (session) {
session.dispose();
}
if (tempDir && existsSync(tempDir)) {
rmSync(tempDir, { recursive: true });
}
});
function createSession() {
const model = getModel("anthropic", "claude-sonnet-4-5")!;
const transport = new ProviderTransport({
getApiKey: () => API_KEY,
});
const agent = new Agent({
transport,
initialState: {
model,
systemPrompt: "You are a helpful assistant. Be concise.",
tools: codingTools,
},
});
sessionManager = new SessionManager(false);
const settingsManager = new SettingsManager(tempDir);
session = new AgentSession({
agent,
sessionManager,
settingsManager,
});
// Subscribe to track events
session.subscribe((event) => {
events.push(event);
});
return session;
}
it("should trigger manual compaction via compact()", async () => {
createSession();
// Send a few prompts to build up history
await session.prompt("What is 2+2? Reply with just the number.");
await session.agent.waitForIdle();
await session.prompt("What is 3+3? Reply with just the number.");
await session.agent.waitForIdle();
// Manually compact
const result = await session.compact();
expect(result.summary).toBeDefined();
expect(result.summary.length).toBeGreaterThan(0);
expect(result.tokensBefore).toBeGreaterThan(0);
// Verify messages were compacted (should have summary + recent)
const messages = session.messages;
expect(messages.length).toBeGreaterThan(0);
// First message should be the summary (a user message with summary content)
const firstMsg = messages[0];
expect(firstMsg.role).toBe("user");
}, 120000);
it("should maintain valid session state after compaction", async () => {
createSession();
// Build up history
await session.prompt("What is the capital of France? One word answer.");
await session.agent.waitForIdle();
await session.prompt("What is the capital of Germany? One word answer.");
await session.agent.waitForIdle();
// Compact
await session.compact();
// Session should still be usable
await session.prompt("What is the capital of Italy? One word answer.");
await session.agent.waitForIdle();
// Should have messages after compaction
expect(session.messages.length).toBeGreaterThan(0);
// The agent should have responded
const assistantMessages = session.messages.filter((m) => m.role === "assistant");
expect(assistantMessages.length).toBeGreaterThan(0);
}, 180000);
it("should persist compaction to session file", async () => {
createSession();
await session.prompt("Say hello");
await session.agent.waitForIdle();
await session.prompt("Say goodbye");
await session.agent.waitForIdle();
// Compact
await session.compact();
// Load entries from session manager
const entries = sessionManager.loadEntries();
// Should have a compaction entry
const compactionEntries = entries.filter((e) => e.type === "compaction");
expect(compactionEntries.length).toBe(1);
const compaction = compactionEntries[0];
expect(compaction.type).toBe("compaction");
if (compaction.type === "compaction") {
expect(compaction.summary.length).toBeGreaterThan(0);
// firstKeptEntryIndex can be 0 if all messages fit within keepRecentTokens
// (which is the case for small conversations)
expect(compaction.firstKeptEntryIndex).toBeGreaterThanOrEqual(0);
expect(compaction.tokensBefore).toBeGreaterThan(0);
}
}, 120000);
it("should work with --no-session mode (in-memory only)", async () => {
const model = getModel("anthropic", "claude-sonnet-4-5")!;
const transport = new ProviderTransport({
getApiKey: () => API_KEY,
});
const agent = new Agent({
transport,
initialState: {
model,
systemPrompt: "You are a helpful assistant. Be concise.",
tools: codingTools,
},
});
// Create session manager and disable file persistence
const noSessionManager = new SessionManager(false);
noSessionManager.disable();
const settingsManager = new SettingsManager(tempDir);
const noSessionSession = new AgentSession({
agent,
sessionManager: noSessionManager,
settingsManager,
});
try {
// Send prompts
await noSessionSession.prompt("What is 2+2? Reply with just the number.");
await noSessionSession.agent.waitForIdle();
await noSessionSession.prompt("What is 3+3? Reply with just the number.");
await noSessionSession.agent.waitForIdle();
// Compact should work even without file persistence
const result = await noSessionSession.compact();
expect(result.summary).toBeDefined();
expect(result.summary.length).toBeGreaterThan(0);
// In-memory entries should have the compaction
const entries = noSessionManager.loadEntries();
const compactionEntries = entries.filter((e) => e.type === "compaction");
expect(compactionEntries.length).toBe(1);
} finally {
noSessionSession.dispose();
}
}, 120000);
it("should emit correct events during auto-compaction", async () => {
createSession();
// Build some history
await session.prompt("Say hello");
await session.agent.waitForIdle();
// Manually trigger compaction and check events
await session.compact();
// Check that no auto_compaction events were emitted for manual compaction
const autoCompactionEvents = events.filter(
(e) => e.type === "auto_compaction_start" || e.type === "auto_compaction_end",
);
// Manual compaction doesn't emit auto_compaction events
expect(autoCompactionEvents.length).toBe(0);
// Regular events should have been emitted
const messageEndEvents = events.filter((e) => e.type === "message_end");
expect(messageEndEvents.length).toBeGreaterThan(0);
}, 120000);
});

46
packages/web-ui/src/agent/transports/AppTransport.ts
View file

@ -11,7 +11,7 @@ import type {
ToolCall,
UserMessage,
} from "@mariozechner/pi-ai";
import { agentLoop } from "@mariozechner/pi-ai";
import { agentLoop, agentLoopContinue } from "@mariozechner/pi-ai";
import { AssistantMessageEventStream } from "@mariozechner/pi-ai/dist/utils/event-stream.js";
import { parseStreamingJson } from "@mariozechner/pi-ai/dist/utils/json-parse.js";
import { clearAuthToken, getAuthToken } from "../../utils/auth-token.js";
@ -315,51 +315,57 @@ function streamSimpleProxy(
return stream;
}
// Proxy transport executes the turn using a remote proxy server
/**
* Transport that uses an app server with user authentication tokens.
* The server manages user accounts and proxies requests to LLM providers.
*/
export class AppTransport implements AgentTransport {
// Hardcoded proxy URL for now - will be made configurable later
private readonly proxyUrl = "https://genai.mariozechner.at";
async *run(messages: Message[], userMessage: Message, cfg: AgentRunConfig, signal?: AbortSignal) {
private async getStreamFn() {
const authToken = await getAuthToken();
if (!authToken) {
throw new Error(i18n("Auth token is required for proxy transport"));
}
// Use proxy - no local API key needed
const streamFn = <TApi extends Api>(model: Model<TApi>, context: Context, options?: SimpleStreamOptions) => {
return streamSimpleProxy(
model,
context,
{
...options,
authToken,
},
this.proxyUrl,
);
return <TApi extends Api>(model: Model<TApi>, context: Context, options?: SimpleStreamOptions) => {
return streamSimpleProxy(model, context, { ...options, authToken }, this.proxyUrl);
};
}
// Messages are already LLM-compatible (filtered by Agent)
const context: AgentContext = {
private buildContext(messages: Message[], cfg: AgentRunConfig): AgentContext {
return {
systemPrompt: cfg.systemPrompt,
messages,
tools: cfg.tools,
};
}
const pc: AgentLoopConfig = {
private buildLoopConfig(cfg: AgentRunConfig): AgentLoopConfig {
return {
model: cfg.model,
reasoning: cfg.reasoning,
getQueuedMessages: cfg.getQueuedMessages,
};
}
async *run(messages: Message[], userMessage: Message, cfg: AgentRunConfig, signal?: AbortSignal) {
const streamFn = await this.getStreamFn();
const context = this.buildContext(messages, cfg);
const pc = this.buildLoopConfig(cfg);
// Yield events from the upstream agentLoop iterator
// Pass streamFn as the 5th parameter to use proxy
for await (const ev of agentLoop(userMessage as unknown as UserMessage, context, pc, signal, streamFn as any)) {
yield ev;
}
}
async *continue(messages: Message[], cfg: AgentRunConfig, signal?: AbortSignal) {
const streamFn = await this.getStreamFn();
const context = this.buildContext(messages, cfg);
const pc = this.buildLoopConfig(cfg);
for await (const ev of agentLoopContinue(context, pc, signal, streamFn as any)) {
yield ev;
}
}
}

35
packages/web-ui/src/agent/transports/ProviderTransport.ts
View file

@ -2,6 +2,7 @@ import {
type AgentContext,
type AgentLoopConfig,
agentLoop,
agentLoopContinue,
type Message,
type UserMessage,
} from "@mariozechner/pi-ai";
@ -14,37 +15,53 @@ import type { AgentRunConfig, AgentTransport } from "./types.js";
* Uses CORS proxy only for providers that require it (Anthropic OAuth, Z-AI).
*/
export class ProviderTransport implements AgentTransport {
async *run(messages: Message[], userMessage: Message, cfg: AgentRunConfig, signal?: AbortSignal) {
// Get API key from storage
private async getModelAndKey(cfg: AgentRunConfig) {
const apiKey = await getAppStorage().providerKeys.get(cfg.model.provider);
if (!apiKey) {
throw new Error("no-api-key");
}
// Get proxy URL from settings (if available)
const proxyEnabled = await getAppStorage().settings.get<boolean>("proxy.enabled");
const proxyUrl = await getAppStorage().settings.get<string>("proxy.url");
// Apply proxy only if this provider/key combination requires it
const model = applyProxyIfNeeded(cfg.model, apiKey, proxyEnabled ? proxyUrl || undefined : undefined);
// Messages are already LLM-compatible (filtered by Agent)
const context: AgentContext = {
return { model, apiKey };
}
private buildContext(messages: Message[], cfg: AgentRunConfig): AgentContext {
return {
systemPrompt: cfg.systemPrompt,
messages,
tools: cfg.tools,
};
}
const pc: AgentLoopConfig = {
private buildLoopConfig(model: typeof cfg.model, apiKey: string, cfg: AgentRunConfig): AgentLoopConfig {
return {
model,
reasoning: cfg.reasoning,
apiKey,
getQueuedMessages: cfg.getQueuedMessages,
};
}
async *run(messages: Message[], userMessage: Message, cfg: AgentRunConfig, signal?: AbortSignal) {
const { model, apiKey } = await this.getModelAndKey(cfg);
const context = this.buildContext(messages, cfg);
const pc = this.buildLoopConfig(model, apiKey, cfg);
// Yield events from agentLoop
for await (const ev of agentLoop(userMessage as unknown as UserMessage, context, pc, signal)) {
yield ev;
}
}
async *continue(messages: Message[], cfg: AgentRunConfig, signal?: AbortSignal) {
const { model, apiKey } = await this.getModelAndKey(cfg);
const context = this.buildContext(messages, cfg);
const pc = this.buildLoopConfig(model, apiKey, cfg);
for await (const ev of agentLoopContinue(context, pc, signal)) {
yield ev;
}
}
}

6
packages/web-ui/src/agent/transports/types.ts
View file

@ -13,10 +13,14 @@ export interface AgentRunConfig {
// We re-export the Message type above; consumers should use the upstream AgentEvent type.
export interface AgentTransport {
/** Run with a new user message */
run(
messages: Message[],
userMessage: Message,
config: AgentRunConfig,
signal?: AbortSignal,
): AsyncIterable<AgentEvent>; // passthrough of AgentEvent from upstream
): AsyncIterable<AgentEvent>;
/** Continue from current context (no new user message) */
continue(messages: Message[], config: AgentRunConfig, signal?: AbortSignal): AsyncIterable<AgentEvent>;
}