Repository navigation
Pots runs on pi 1.6.0: GPT-6.2 Astra and GPT-6.2 Sol, with GPT-6.2 Sol as the default model. Images that any tool gives the bot are fitted to the model's limits, as in the pi command line, and each MCP server is a namespace of its own in code mode. After this upgrade, a bot's database cannot go back to 0.8.7.
Added
- GPT-6.2 Astra and GPT-6.2 Sol, on
openai,azureandopenai-codex(the ChatGPT login). With the ChatGPT login, Sol has a 500K context window and Astra 272K; Azure gives both 1.05M. Neither has thinkingofforminimal; a level a model does not have becomes the nearest one it has. - Images from every tool are fitted to the model. Before, only
readshrank its images, so a large image from an MCP tool or from a script'simage()could make the provider refuse the request. Now pi's image fitting fits every image a tool returns to the limits of the conversation's model before it is stored, for the main agent and its children, as the pi command line does. An image that cannot be prepared is left out with a note.
Changed
- After the upgrade, a bot's database cannot go back to 0.8.7. pi 1.6.0 stores the bot's tool calls and model requests in a newer format, which 0.8.7 and older do not read correctly. If you may want to go back, stop Pots and copy each bot's
bot.sqlitebefore you upgrade. - pi 1.6.0 (was 1.5.2). What the bot gains:
- Models. A request that fails with
Connection closed.or a socket error such asECONNRESET,EPIPEorETIMEDOUTis tried again. With the ChatGPT login, a request that reaches the connection's 60-minute limit during the answer, for example during long thinking, is tried again on a new connection; GPT-6.2 now gets prompt cache hits on the system prompt; a change of thinking level no longer loses the prompt cache on GPT-6 models; and a model'smaxTokensis used. OpenAI no longer refuses a request whose tool schema has a part without atype; such a tool is sent as not strict. Mistral Large 4 tool calls no longer stall for minutes. OpenAI-compatible providers other than OpenAI get system messages with thesystemrole, so Groq's Qwen models no longer fail with400 Unexpected message role. The Qwen Token Plan providers have new models and use the current gateways. - llama.cpp. pi's llama.cpp provider connects to
http://127.0.0.1:9931by default, the current default port of llama-server, instead of 8080. A llama.cpp classifier that takes images, such as Clef with itsmmproj, now gets the images ofmodels.classify(). - Tools.
readno longer keeps a note that the model takes no images, which stayed in the conversation after a change to a model that takes them.bashshows the end of its output when the output ends in one very long line (over about 100KB, such as agrepmatch in a minified file); before, it showed nothing. - undici 8.12.4 (was 8.12.3), for six security advisories.
- Models. A request that fails with
findpath patterns start at the folder searched (pi 1.6.0). A pattern with/now matches from the folderfindsearches, as theglobofgrepdoes: searchingsrc,slack/*.tsno longer findssrc/channels/slack/sdk.ts, and**/slack/*.tsfinds it at any depth. A pattern without/still matches file names at any depth, and a pattern that starts with./now works.- Each MCP server is a namespace of its own in code mode. The
codemodesection lists each server asmcp__<server>, with pi's one-line summary of it (the first line of itsdescriptioninmcp.json, or else of the instructions the server sends, at most 250 characters), also before it connects.describeNamespace("mcp__<server>")gives the whole description, the instructions and the server's tools, andsearchTools()searches them. A server whose every tool ishiddenis not named. For a bot with MCP servers, thecodemodesection changes, so its prompt cache is written once more after the upgrade. - Children follow the bot's MCP servers. A child has the main agent's tools without the channel actions and
task, and keeps them as that difference, so a server added tomcp.jsonwhile a child works reaches the child too, and a server removed there goes from it. Children that 0.8.7 started are changed this way when the bot starts. The order of their tools changes, so each such child's prompt cache is written once more. - The default model is
openai/gpt-6.2-sol(wasopenai/gpt-6.1-sol), as pi's default foropenai. A bot whosebot.jsonhas nomodeluses it after the restart.
What's Changed
- pi 1.6.0: GPT-6.2, MCP servers as code-mode namespaces, image fitting for every tool by @Crokily in #66
Full Changelog: v0.8.7...v0.8.8