refactor: drive tool prompt shaping through TOOL_INSTRUCTIONS_BROWSER…

.env.example

@@ -68,7 +68,7 @@ DEEPSEEK_CONTEXT_ROLLOVER=0
 # RATE_LIMIT_PER_MINUTE=30               # per-client-IP limit on /v1 endpoints
 
 # --- TOOL / TOOLCALL ---
-# Optional raw per-chat traces (prompts, tool calls/results and replies) in ./llog/.
+# Optional raw per-chat traces (prompts, tool calls/results and replies) in ./logs/.
 # Files can contain private project data. Disabled by default.
 TOOLCALL_LOG=0
 

.github/workflows/tests.yml

@@ -10,7 +10,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ["3.9", "3.12"]
+        python-version: ["3.10", "3.12"]
     env:
       SERVER_INTERACTIVE_LOGIN: "0"
       SESSION_REFRESH_ENABLED: "0"

README.en.md

@@ -577,13 +577,13 @@ With `TOOLCALL_FORMAT_REPAIR=1` (off by default), an erroneous DeepSeek reply ge
 one format correction in the same conversation. A second correction is never
 requested. Verified offline.
 
-In compact browser mode, an agent's tool result (for example the option you picked)
-is sent as `Tool result (name id="..."): <result>`. IDs retain result binding;
-multiple calls to the same tool also receive an exact call mapping. Complete schemas
-are normally omitted, but the current `tool_choice` follows the result even in
-`none` mode. `auto` explicitly replaces an earlier per-reply selection; task
-constraints and schema validation remain enforced. An authorized catalog upgrade
-sends complete schemas once.
+A tool-result tail into an existing chat uses mode `none` (current `tool_choice`
+policy only) and is sent as `Tool result (name): <result>`; `id` is attached only
+when several results share a tool name. Complete schemas are omitted on that
+tail; an authorized catalog upgrade sends complete schemas once
+(`tool_instructions=full`). Automatic `auto` policy text is included in
+`full`/`medium` only; `light`/`none` continuations omit it. Task constraints and
+schema validation remain enforced.
 Verified offline.
 
 For DeepSeek, disabling `TOOLCALL_FORMAT_REPAIR` prevents an extra format-correction generation;

README.md

@@ -574,14 +574,13 @@ DeepSeek без угадывания аргументов, невалидный
 получает одну коррекцию формата в той же беседе. Вторая коррекция не
 запрашивается. Проверено офлайн.
 
-Результат инструмента, который возвращает агент (например выбранный вами вариант),
-в компактном браузерном режиме уходит как `Tool result (имя id="..."): <результат>`.
-ID сохраняется для привязки; несколько вызовов одного инструмента дополнительно
-получают точное сопоставление вызовов. Полные схемы обычно не повторяются, но
-текущий `tool_choice` передаётся после результата даже при режиме `none`.
-`auto` явно снимает прежний выбор для отдельного ответа; ограничения задачи
-и проверки схем сохраняются. При разрешённой смене каталога один раз отправляются
-полные схемы.
+Хвост с результатом инструмента в существующий чат уходит в режиме `none`
+(только политика `tool_choice`) как `Tool result (имя): <результат>`; `id`
+добавляется лишь когда несколько результатов делят одно имя инструмента.
+Полные схемы на таком хвосте не повторяются; при разрешённой смене каталога
+один раз отправляются полные схемы (`tool_instructions=full`). Текст про
+автоматический `auto` идёт в режимах `full`/`medium`; в `light`/`none` на
+продолжениях его нет. Ограничения задачи и проверки схем сохраняются.
 Проверено офлайн.
 
 Для DeepSeek отключение `TOOLCALL_FORMAT_REPAIR` запрещает дополнительную генерацию для

browser/opencode-controller.user.js

@@ -1,7 +1,7 @@
 // ==UserScript==
 // @name         OpenCode signed-in tab controller
 // @namespace    opencode-local-bridge
-// @version      0.3.35
+// @version      0.3.37
 // @description  Opt-in local jobs in your existing signed-in browser tab.
 // @match        https://chat.z.ai/*
 // @match        https://grok.com/*
@@ -195,15 +195,98 @@
     label();
   }
 
+  // Attached to the next finish so TOOLCALL_LOG sees what happened after emit.
+  let pendingLateRewrite = null;
+  let pendingDomCheck = null;
+  let domCheckToken = null;
   async function finish(result) {
     const job = pending;
     retire();
     if (!job) return;
     finishing = true;
-    try { await rpc("/browser/result", "POST", { ...identity(job), result }); }
+    try {
+      const payload = { ...result };
+      if (pendingLateRewrite && typeof pendingLateRewrite === "object") {
+        payload.late_rewrite = pendingLateRewrite;
+        pendingLateRewrite = null;
+      }
+      if (pendingDomCheck && typeof pendingDomCheck === "object") {
+        payload.dom_check = pendingDomCheck;
+        pendingDomCheck = null;
+      }
+      await rpc("/browser/result", "POST", { ...identity(job), result: payload });
+    }
     finally { finishing = false; }
   }
 
+  // ---- DOM check (docs/BRIDGE_TRACE_STATS.ru.md): any site with answerSnapshot ----
+  // Rendering turns Markdown into DOM, so the comparison ignores everything but
+  // letters and digits; similarity is the word-multiset overlap in percent.
+  // Markdown round-trips (tables, links, list markers) leave a few percent of
+  // noise: >= 97% overlap with lengths within 5% still counts as the same answer.
+  // coverage is the share of emitted words present in the bubble: a rendered
+  // widget (a mermaid diagram) adds DOM text but keeps coverage near 100%.
+  const domWords = text => String(text || "").toLowerCase().match(/[\p{L}\p{N}]+/gu) || [];
+  const domCompare = (wire, dom) => {
+    const result = domOverlap(wire, dom);
+    const a = String(wire || "").length, b = String(dom || "").length;
+    const close = Math.abs(a - b) <= Math.max(a, b) * 0.05;
+    return { exact: result.match, match: result.match || (result.similarity >= 97 && close),
+      similarity: result.similarity, coverage: result.coverage };
+  };
+  const domOverlap = (wire, dom) => {
+    const a = domWords(wire), b = domWords(dom);
+    const counts = new Map();
+    for (const word of a) counts.set(word, (counts.get(word) || 0) + 1);
+    let common = 0;
+    for (const word of b) {
+      const left = counts.get(word) || 0;
+      if (left > 0) { common += 1; counts.set(word, left - 1); }
+    }
+    return {
+      match: a.join("") === b.join(""),
+      similarity: a.length + b.length ? Math.round(200 * common / (a.length + b.length)) : 100,
+      coverage: a.length ? Math.round(100 * common / a.length) : 100,
+    };
+  };
+  // After an emit, watch the newest answer bubble (snapshot() returns the
+  // rendered answers, oldest first) until the next job starts (or 15 s) and
+  // record metadata only: whether the final rendered text matches what the
+  // agent received, and how often the bubble changed meanwhile.
+  async function domCheck(wireText, snapshot) {
+    const token = {};
+    domCheckToken = token;
+    const startedAt = performance.now();
+    let last = null, changes = -1, changedAt = startedAt;
+    let ended = "timeout";
+    while (performance.now() - startedAt < 15000) {
+      if (domCheckToken !== token) { ended = "next_job"; break; }
+      let texts = [];
+      try { texts = snapshot() || []; } catch (_) { texts = []; }
+      const dom = texts.length ? String(texts[texts.length - 1] || "") : "";
+      if (dom !== last) { last = dom; changes += 1; changedAt = performance.now(); }
+      await new Promise(resolve => setTimeout(resolve, 250));
+    }
+    if (domCheckToken === token) domCheckToken = null;
+    const dom = last || "";
+    const result = {
+      found: !!dom,
+      ...domCompare(wireText, dom),
+      changes: Math.max(0, changes),
+      stable_ms: Math.max(0, Math.min(10000000, Math.round(performance.now() - changedAt))),
+      wire_len: String(wireText || "").length,
+      dom_len: dom.length,
+      tool_wire: String(wireText || "").includes("```tool_calls"),
+      tool_dom: dom.includes("```tool_calls"),
+      ended,
+    };
+    pendingDomCheck = result;
+    if (!result.match) {
+      try { console.warn("[opencode] rendered answer differs from emitted wire", result); } catch (_) { /* ignore */ }
+    }
+    return result;
+  }
+
   button.addEventListener("click", async () => {
     if (TOKEN === "__BRIDGE_TOKEN__") { label("Сначала экспортируй приватные скрипты моста"); return; }
     button.title = "";
@@ -216,7 +299,18 @@
 
   const visible = element => !!element && element.getClientRects().length > 0
     && getComputedStyle(element).visibility === "visible" && !element.closest('[inert], [aria-hidden=true]');
-  const editor = () => Array.from(document.querySelectorAll(site.input)).find(visible);
+  // Some composers mark an ancestor aria-hidden while dialogs/tool chips are open;
+  // still require a laid-out, non-inert, non-display:none control. Opt in with
+  // site.editorIgnoreAriaHidden when the shell itself stays aria-hidden.
+  const editorVisible = element => {
+    if (!element || element.getClientRects().length === 0) return false;
+    const style = getComputedStyle(element);
+    if (style.visibility === "hidden" || style.display === "none") return false;
+    if (element.closest("[inert]")) return false;
+    if (!site.editorIgnoreAriaHidden && element.closest("[aria-hidden=true]")) return false;
+    return true;
+  };
+  const editor = () => Array.from(document.querySelectorAll(site.input)).find(editorVisible);
   const block = node => node.nodeType === Node.ELEMENT_NODE && /^(P|DIV|LI|PRE|BLOCKQUOTE)$/.test(node.tagName);
   function editorText(node) {
     if (node.nodeType === Node.TEXT_NODE) return node.textContent;
@@ -347,17 +441,22 @@
     ".send-button-container, [class*='send-button-container']",
   )).find(visible) || null;
   const sendArmed = button => {
-    if (!button || button.disabled || button.getAttribute("aria-disabled") === "true"
-        || button.classList.contains("disabled")) return false;
+    if (!button || button.disabled || button.getAttribute("aria-disabled") === "true") {
+      return false;
+    }
+    // Class "disabled" can linger after the attribute is removed; the attribute
+    // is the source of truth when site.sendIgnoreDisabledClass is set.
+    if (!site.sendIgnoreDisabledClass && button.classList.contains("disabled")) return false;
     const wrap = button.closest(".chat-prompt-send-button, .send-button-container")
       || button.parentElement;
     if (wrap) {
       const style = getComputedStyle(wrap);
       if (style.pointerEvents === "none" || Number(style.opacity) < 0.5) return false;
-      if (wrap.classList.contains("disabled")) return false;
+      if (!site.sendIgnoreDisabledClass && wrap.classList.contains("disabled")) return false;
     }
     const style = getComputedStyle(button);
-    if (style.pointerEvents === "none" || Number(style.opacity) < 0.5) return false;
+    if (style.pointerEvents === "none") return false;
+    if (!site.sendIgnoreDisabledClass && Number(style.opacity) < 0.5) return false;
     return true;
   };
   const activateClick = element => {
@@ -603,6 +702,18 @@
   document.addEventListener("opencode-local-response-v1", async event => {
     let value;
     try { value = JSON.parse(event.detail); } catch (_) { return; }
+    if (value?.late_rewrite && typeof value.late_rewrite === "object") {
+      const note = value.late_rewrite;
+      pendingLateRewrite = {
+        delay_ms: Number.isFinite(note.delay_ms) ? Math.max(0, Math.min(10000000, note.delay_ms | 0)) : 0,
+        tool_before: !!note.tool_before,
+        tool_after: !!note.tool_after,
+      };
+      try {
+        console.warn("[opencode] late rewrite after emit", pendingLateRewrite);
+      } catch (_) { /* ignore */ }
+      return;
+    }
     if (!pending || value?.nonce !== pending.nonce) return;
     if (value.observing === true) {
       pending.sawCapture = true;
@@ -653,13 +764,19 @@
     if (claudeChatgbotPage() && site.provider === "claude") {
       reported = claudeCanonicalPath(reported);
     }
+    if (typeof site.answerSnapshot === "function" && typeof value.answer_text === "string") {
+      void domCheck(value.answer_text, site.answerSnapshot);
+    }
     await finish({ status: value.status, body: value.body, path: reported,
       ...(value.request_turn ? { request_turn: value.request_turn } : {}),
+      ...(value.settle && typeof value.settle === "object" ? { settle: value.settle } : {}),
     }).catch(() => label("OpenCode: ответ не принят"));
   });
 
   async function execute(job) {
     if (!Number.isFinite(job.expires_in) || job.expires_in <= 0) throw new BridgeFailure("Invalid browser job deadline");
+    // The previous answer's DOM check ends before this prompt adds a new bubble.
+    domCheckToken = null;
     const deadline = performance.now() + Math.min(job.expires_in, 1800) * 1000;
     let input = editor();
     // Leftover text from a previous failed bridge attempt matches this job's
@@ -693,7 +810,9 @@
       return;
     }
     // The extension can run before the SPA hydrates its signed-in editor.
-    for (let attempt = 0; !input && enabled && attempt < 50; attempt++) {
+    // Some sites also remount the composer between turns / while tool UI is open.
+    const editorWait = Number.isFinite(site.editorWaitAttempts) ? site.editorWaitAttempts : 50;
+    for (let attempt = 0; !input && enabled && attempt < editorWait; attempt++) {
       if (attempt % 5 === 0 && !(await rpc("/browser/check", "POST", identity(job))).active) {
         throw new BridgeFailure("The browser job was cancelled");
       }
@@ -703,7 +822,14 @@
     if (!enabled || !sameRoute(location.pathname, target)) throw new BridgeFailure("The tab changed before submission");
     if (!input) throw new BridgeFailure("The signed-in chat input is unavailable");
     await clearLeftoverDraft(input, job.prompt);
-    if (input.disabled || input.getAttribute("aria-disabled") === "true") throw new BridgeFailure("The chat input is busy");
+    for (let attempt = 0; input && (input.disabled || input.getAttribute("aria-disabled") === "true")
+        && attempt < editorWait; attempt++) {
+      await new Promise(resolve => setTimeout(resolve, 100));
+      input = editor() || input;
+    }
+    if (!input || input.disabled || input.getAttribute("aria-disabled") === "true") {
+      throw new BridgeFailure("The chat input is busy");
+    }
     if (job.submitted) throw new BridgeFailure("The prompt was already submitted");
     pending = { ...job, nonce: crypto.randomUUID(), deadline };
     // Use strings across Safari's isolated/page worlds. Do not send upstream

browser/opencode-observer.user.js

@@ -1,7 +1,7 @@
 // ==UserScript==
 // @name         OpenCode browser response observer
 // @namespace    opencode-local-bridge
-// @version      0.3.18
+// @version      0.3.19
 // @description  Observe only the completion caused by an active local bridge job.
 // @match        https://chat.z.ai/*
 // @match        https://grok.com/*
@@ -253,7 +253,7 @@
     const match = location.pathname.match(/^\/(?:u\/\d+\/)?app\/([\w-]+)$/);
     return match ? "/app/" + match[1] : null;
   };
-  const emitBody = (job, status, binary, conversationPath, requestTurn) => {
+  const emitBody = (job, status, binary, conversationPath, requestTurn, settle, answerText) => {
     if (active?.nonce !== job.nonce) return;
     let encoded = "";
     for (let i = 0; i < binary.length; i += 16384) {
@@ -263,6 +263,9 @@
       nonce: job.nonce, status, body: btoa(encoded),
       ...(typeof conversationPath === "string" ? { path: conversationPath } : {}),
       ...(requestTurn ? { request_turn: requestTurn } : {}),
+      ...(settle && typeof settle === "object" ? { settle } : {}),
+      // Page-local only: the controller compares it with the rendered bubble.
+      ...(typeof answerText === "string" ? { answer_text: answerText } : {}),
     }) }));
   };
   const beginCapture = job => {

chat_protocol.py

@@ -13,6 +13,8 @@ class Reply:
     bridge_notice: Optional[dict] = None
     # Ordered text/tool_call parts for streaming; None keeps the legacy framing.
     parts: Optional[list] = None
+    # Optional tab settle metadata (counts/flags only; never response text).
+    bridge_settle: Optional[dict] = None
 
     def __str__(self) -> str:
         return self.text