v1.16.0

A minor release with one experimental headline feature and one long-session fix: run_script ("code mode") lets the agent batch many tool calls into a single sandboxed JavaScript step, and "Add Project" context is now sent once, in the system prompt, instead of being copied into every user message.

Added

• feat(agent): run_script batches tool calls in one JavaScript step (experimental, opt-in) — In agent mode every tool call costs a model round trip, and every raw result stays in chat memory. With Settings → DevoxxGenie → Agent → Built-in Tools → run_script enabled, the model can instead write a short async JavaScript snippet that calls other tools as await tools.read_file({ path }), loops and filters over their results, and returns only what it prints with text() / console.log. The tool description lists every callable tool as a TypeScript-style signature, and ALL_TOOLS exposes the catalog at runtime. Scripts run in a fresh quickjs4j engine (QuickJS compiled to WebAssembly and then to plain Java bytecode, so no native code): no Node.js, require, timers, file system or network; output capped at 20,000 characters; memory capped at 128 MiB; a 30 s JavaScript time budget that excludes time spent inside tool calls (such as waiting on an approval); cancelling the prompt stops the script. Every call inside a script goes through the same tool toggles, approval dialogs and activity rows as a direct call. run_script itself counts as one call toward Max tool calls per prompt, while the calls inside one script are limited by a new Max tool calls per script setting (default 100). Its description is editable like the other built-in tools. Off by default; when disabled the agent tool chain is unchanged (#1352, #1353)

Fixed

• fix(memory): "Add Project" context is sent once, via the system prompt, instead of being copied into every user message — The full-project snapshot used to be appended to the text of each prompt and stored in chat memory with it, so after n prompts the history carried n stale copies of the project. Context windows filled far too early, every copy was re-sent (and billed) on each request, and with a large project the second prompt could fail outright with "prompt too long". The snapshot now lives in a single <ProjectFilesContext> section of the system prompt, which is updated in place when the project is added or removed mid-conversation. prepareMemory now keeps the system message in sync before every prompt, which also fixes restored conversations running without a system prompt (persona, DEVOXXGENIE.md, CLAUDE.md/AGENTS.md and tool instructions were silently lost after restoring from history). OpenAI o1 models, which take no system message, get the section in their per-message <SystemPrompt> block; the agent/MCP tool loop passes it through on every request; and CLI/ACP runners receive it once per request in a <project_context> block (#1363, #1364)

Dependencies

• build: add quickjs4j 0.1.1 (~1.2 MB including its WebAssembly runtime); jackson-databind/core move 2.22.1 → 2.22.3 as a result (#1353)

Documentation

• docs: align the plugin reference guides with the current UI and fill feature coverage gaps (#1367)
• docs: refresh token-cost and context-window examples, and update the pricing comparison (#1368, #1369)
• docs(blog): compare the Matt Pocock and Addy Osmani agent skills, and add settings discovery links (#1366)
• docs(tips): add tips for the v1.15.0 and v1.15.1 features and fixes (#1362)
• docs: document run_script in Agent Mode, and announce v1.16.0 with a dedicated run_script deep-dive post

Contributors

• @omarmahamid
• @SamyNassiem
• @stephanj

Read more: run_script deep dive · v1.16.0 announcement