Plank scripts intelligence

README, landing and whitepaper now tell the AI story: model, prompt, tool fn, agent, Turing and Joshua Tree.

Structured answers

m.ask_json(question) returns a Json value: asks for only JSON, strips a code fence, retries once with the parse error. Scorecard 60 of 60.

Bounded agent memory

ai.agent takes keep: the saved conversation is capped at about the last n messages, cut only where a turn starts so tool results never lose their call. Scorecard 59 of 59.

Turing as a model

ai.turing() points a Plank script at Turing's own OpenAI-compatible server on port 8127.

Agents on local models

Tools and agents now work with any OpenAI-compatible server (oMLX, Ollama, Turing): the conversation converts to tool_calls and back. Scorecard 58 of 58.

tool fn

tool fn reads a function's signature and the comment above it and writes the ai.Tool, so the schema the model sees cannot drift from the code. Scorecard 57 of 57.

Division by zero stops the program

Integer / and % by zero now stop with the line number (catchable) instead of returning garbage. Found while testing agent tools.

AI-native Plank: model, prompt, tool, agent

lib/ai.pk: ask any model (Anthropic or OpenAI-compatible), prompt templates, tools with schemas, agents with file memory, and a mock model for byte-stable tests. Scorecard 56 of 56.

Plank 4.0

Result with ?, generic enums, packages fetched straight from GitHub, plank doc, on top of modules, sets, dates and static builds. Every scorecard row a yes against Python, Ruby and Swift.

3.3.0: plank doc

plank doc lists every built-in with its signature and a line; plank doc name shows one.