fix(ai-plugin): writeMacro family no-ops on web Document Server — nes…

[zhaokaige]

On a web-deployed Document Server, writeMacro (and generateDocx / changeParagraphStyle, same pattern) silently fail: WRITE requests report success but never modify the document; READ requests return no data.

(1) WRITE: the macro code is passed to Asc.Editor.callCommand as a nested eval inside the function body. Asc.plugin.callCommand serializes the function via toString(), and the editor core executes it through _safe_eval_closure (sdk-all-min.js), which wraps the string with Function.apply(... return eval(…)) and shadows the window formal parameter — a deliberate plugin sandbox. A side effect is that a nested eval() inside the function body returns undefined and never runs, so the macro silently no-ops. Verified by four probes: new Function(code) and executeCommand("command", code) mutate; nested eval(code) and nested new Function(code)() do not. The macro code must be the function body, not runtime-eval'd data.

(2) READ: replacing eval with new Function fixes WRITE, but a new Function body discards the last expression's value. The prompts and tool descriptions were built on eval's auto-return semantics, so a READ macro returned nothing.

Fix:

• helpers.js (3 sites): dual-form new Function wrapper. Expression form captures an IIFE / single expression; if new Function throws (multi-statement or has return), fall back to the statement form which captures a top-level return. try/catch + onlyoffice_id_result / onlyoffice_id_error_message contract preserved (safePluginEval swallows thrown errors).
• helperFuncs.js + helpers.js: align prompts and tool descriptions to the new execution model — read examples use explicit return; "last expression" -> "explicit return"; add READ/WRITE/CHAT decision rule (mutate only on explicit modify; reading never mutates), GENERATIVE rule (ask first on ambiguous "write content" requests), and SELF-CHECK rule (verify top-level return on READ macros).

Sandbox untouched; plugin isolation unchanged.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}