open-agents-ai 0.187.474 → 0.187.476

package/dist/index.js

@@ -518229,6 +518229,52 @@ function getSystemPromptForTier(tier) {
       return SYSTEM_PROMPT;
   }
 }
+function detectTaskMode(task) {
+  if (!task)
+    return false;
+  const head = task.slice(0, 4e3).toLowerCase();
+  if (task.length > 2e3)
+    return true;
+  if (/(\/[\w.-]+){2,}/.test(task.slice(0, 2e3)))
+    return true;
+  if (/\b(implement|build|create|refactor|write|fix|migrate|deploy|generate|setup|set up|develop|design|integrate)\b/.test(head)) {
+    if (/\b(spec|file|module|component|api|endpoint|database|schema|test|build|next\.js|typescript|react|prisma|tailwind|sql|python|rust|go)\b/.test(head)) {
+      return true;
+    }
+  }
+  return false;
+}
+function slimSystemPromptForTaskMode(prompt) {
+  const SECTION_HEADERS_TO_REMOVE = [
+    /^##\s*Interactive\s*\/\s*Long-?Running Sessions\s*$/im,
+    /^##\s*Document Generation Strategy\s*$/im,
+    /^##\s*Calculations\s*[—-]\s*Always Execute, Never Guess\s*$/im,
+    /^##\s*Knowledge Gaps\s*[—-]\s*Search, Don't Hallucinate\s*$/im,
+    /^##\s*Self-Awareness( & Introspection)?\s*$/im,
+    /^##\s*Debugging\s*[—-]\s*Observe Before Reasoning\s*$/im
+  ];
+  const TOOL_LINES_TO_REMOVE = [
+    /^- nexus:.*$/im,
+    /^- background_run.*task_status.*task_output.*task_stop:.*$/im,
+    /^- (asr_listen|audio_capture|audio_playback|audio_analyze|camera_capture|desktop_click|bluetooth_scan|browser_action):.*$/im,
+    /^Voice\/TTS:.*$/im,
+    /^- Voice\/TTS:.*$/im,
+    /^- Desktop\/Vision:.*$/im,
+    /^- P2P:.*$/im
+  ];
+  const CHAT_MODE_BLOCK = /^\*\*CHAT MODE\*\*[\s\S]*?(?=\*\*TASK MODE\*\*)/im;
+  let out = prompt;
+  for (const re of SECTION_HEADERS_TO_REMOVE) {
+    out = out.replace(new RegExp(re.source + "[\\s\\S]*?(?=^##\\s|\\Z)", "im"), "");
+  }
+  for (const re of TOOL_LINES_TO_REMOVE) {
+    out = out.replace(re, "");
+  }
+  out = out.replace(CHAT_MODE_BLOCK, "");
+  out = out.replace(/^\*\*TASK MODE\*\*[^\n]*\n/im, "");
+  out = out.replace(/\n{3,}/g, "\n\n");
+  return out.trim() + "\n";
+}
 function computeTodoReminder(input) {
   const turnsSinceWriteThreshold = input.turnsSinceWriteThreshold ?? 10;
   const turnsBetweenReminders = input.turnsBetweenReminders ?? 10;
@@ -518638,7 +518684,17 @@ var init_agenticRunner = __esm({
       async assembleContext(task, context2) {
         const sections = [];
         const pressureCue = pressureCheck(task);
-        const basePrompt = getSystemPromptForTier(this.options.modelTier) + pressureCue;
+        const rawPrompt = getSystemPromptForTier(this.options.modelTier);
+        const taskModeOn = detectTaskMode(task);
+        const slimmedPrompt = taskModeOn ? slimSystemPromptForTaskMode(rawPrompt) : rawPrompt;
+        const basePrompt = slimmedPrompt + pressureCue;
+        if (taskModeOn) {
+          this.emit({
+            type: "status",
+            content: `REG-19: TASK MODE detected — system prompt slimmed ${rawPrompt.length}→${slimmedPrompt.length} bytes`,
+            timestamp: (/* @__PURE__ */ new Date()).toISOString()
+          });
+        }
         const _BATCH_GUIDANCE = {
           small: "\n\n## Response batching\n\nEmit AT MOST 2 tool calls per response. After observing their results, plan the next 2 in your following response. Smaller batches let the orchestrator deliver cache/failure/progress signals to you between actions. Tool calls beyond the cap are dropped. Use todo_write between batches to mark progress.",
           medium: "\n\n## Response batching\n\nEmit AT MOST 4 tool calls per response. After observing their results, plan the next batch in your following response. Smaller batches let the orchestrator deliver cache/failure/progress signals to you between actions. Tool calls beyond the cap are dropped. Use todo_write between batches to mark progress.",
@@ -520548,6 +520604,28 @@ TASK: ${task}` : task;
         for (const [tool, budget] of Object.entries(toolBudgets)) {
           toolCallBudget.set(tool, budget);
         }
+        const stagnationWindow = [];
+        let stagnationCooldownUntilTurn = -1;
+        const STAG_WINDOW_TURNS = 40;
+        const STAG_WINDOW_MS = 10 * 60 * 1e3;
+        const STAG_MIN_SAMPLES = 30;
+        const STAG_FAILURE_THRESHOLD = 5;
+        const STAG_VARIANT_THRESHOLD = 4;
+        const STAG_FILES_DELTA_MIN = 3;
+        let injectionsThisTurn = 0;
+        const INJECTION_BUDGET_SOFT = 2;
+        const deferredSoftInjections = [];
+        const pushSoftInjection = (role, content) => {
+          if (injectionsThisTurn < INJECTION_BUDGET_SOFT) {
+            messages2.push({ role, content });
+            injectionsThisTurn++;
+            return true;
+          }
+          if (deferredSoftInjections.length < 6) {
+            deferredSoftInjections.push({ role, content });
+          }
+          return false;
+        };
         for (let turn = 0; turn < this.options.maxTurns; turn++) {
           clearTurnState(this._appState);
           this._maybeApplyThinkGuard();
@@ -520562,6 +520640,76 @@ TASK: ${task}` : task;
             this.emit({ type: "error", content: "Task aborted by user", timestamp: (/* @__PURE__ */ new Date()).toISOString() });
             break;
           }
+          injectionsThisTurn = 0;
+          while (deferredSoftInjections.length > 0 && injectionsThisTurn < INJECTION_BUDGET_SOFT) {
+            const next = deferredSoftInjections.shift();
+            messages2.push({ role: next.role, content: next.content });
+            injectionsThisTurn++;
+          }
+          if (turn > stagnationCooldownUntilTurn && stagnationWindow.length >= STAG_MIN_SAMPLES) {
+            const cutoffTurn = turn - STAG_WINDOW_TURNS;
+            const cutoffTs = Date.now() - STAG_WINDOW_MS;
+            while (stagnationWindow.length && (stagnationWindow[0].turn < cutoffTurn || stagnationWindow[0].ts < cutoffTs)) {
+              stagnationWindow.shift();
+            }
+            if (stagnationWindow.length >= STAG_MIN_SAMPLES) {
+              const completedDelta = stagnationWindow[stagnationWindow.length - 1].completedTodos - stagnationWindow[0].completedTodos;
+              const fileSet = /* @__PURE__ */ new Set();
+              for (const s2 of stagnationWindow)
+                for (const p2 of s2.filesTouchedThisTurn)
+                  fileSet.add(p2);
+              const filesDelta = fileSet.size;
+              const failureSum = stagnationWindow.reduce((a2, s2) => a2 + s2.failuresThisTurn, 0);
+              const variantSet = /* @__PURE__ */ new Set();
+              for (const s2 of stagnationWindow)
+                for (const p2 of s2.shellPrefixesThisTurn)
+                  variantSet.add(p2);
+              const variantCount = variantSet.size;
+              if (completedDelta === 0 && filesDelta < STAG_FILES_DELTA_MIN && failureSum >= STAG_FAILURE_THRESHOLD && variantCount >= STAG_VARIANT_THRESHOLD) {
+                const variantList = [...variantSet].slice(0, 8).map((v) => `  • ${v}`).join("\n");
+                const stagMsg = [
+                  `[STAGNATION DETECTED — DIAGNOSTIC MODE REQUIRED]`,
+                  ``,
+                  `Over the last ${stagnationWindow.length} turns you have:`,
+                  `  • Completed 0 new todos`,
+                  `  • Written/edited only ${filesDelta} unique file(s) (need ≥${STAG_FILES_DELTA_MIN} for healthy progress)`,
+                  `  • Accumulated ${failureSum} failures`,
+                  `  • Tried ${variantCount} different shell-command variants:`,
+                  variantList,
+                  ``,
+                  `You are not making progress — you are trying surface-level variants of the same approach without diagnosing root cause. This is the failure mode that prevents real completion.`,
+                  ``,
+                  `MANDATORY NEXT ACTIONS (do NOT call task_complete; do NOT try another variant):`,
+                  ``,
+                  `1. READ THE FULL ERROR — re-read your most recent failure output ENTIRELY. If it's in a log packet, call log_explore({op:"errors"}) then log_explore({op:"lines", start:..., end:...}) for context. Do not skim.`,
+                  ``,
+                  `2. STATE A HYPOTHESIS in writing — what specifically is wrong? "I think X is failing because Y." Be concrete. Do NOT propose a fix yet.`,
+                  ``,
+                  `3. VERIFY ONE ASSUMPTION — pick the ONE thing you most BELIEVE to be true and test it with the smallest possible command:`,
+                  `     • If you think a package is installed: ls node_modules/<name>/package.json`,
+                  `     • If you think an env var is set: printenv <NAME>`,
+                  `     • If you think a file imports correctly: head -5 <file>`,
+                  `     • If you don't know what an error means: web_search("<exact error string>")`,
+                  ``,
+                  `4. CHECK SILENT FAILURES — npm install reporting "added N packages" does NOT mean ALL declared deps installed; npm sometimes drops packages with peer-dep conflicts without erroring. Verify each expected dep individually.`,
+                  ``,
+                  `DO NOT in your next response:`,
+                  `  • Try another version, flag, or variant of any command in the list above`,
+                  `  • Wipe node_modules / re-install — that hides the original error`,
+                  `  • Call task_complete — being stuck on a debug problem is NEVER grounds for task_complete`,
+                  ``,
+                  `task_complete is ONLY for actual completion or unrecoverable hardware/permission errors. You are stuck on a fixable problem; diagnose it.`
+                ].join("\n");
+                messages2.push({ role: "system", content: stagMsg });
+                stagnationCooldownUntilTurn = turn + 5;
+                this.emit({
+                  type: "status",
+                  content: `STAGNATION DETECTED — injected diagnostic mode at turn ${turn} (${variantCount} variants, ${failureSum} failures, ${filesDelta} files in window)`,
+                  timestamp: (/* @__PURE__ */ new Date()).toISOString()
+                });
+              }
+            }
+          }
           if (pendingConstraintWarnings.length > 0) {
             const warningMsg = "<constraint-recall>\n" + pendingConstraintWarnings.join("\n") + "\n</constraint-recall>";
             messages2.push({ role: "system", content: warningMsg });
@@ -520745,11 +520893,8 @@ Now call file_write with YOUR skeleton for this task.`
             if (toolHints.length > 0) {
               toolHints.sort((a2, b) => b.score - a2.score);
               const top = toolHints.slice(0, 5);
-              messages2.push({
-                role: "system",
-                content: `[Relevant tools for this task]
-${top.map((t2) => `- ${t2.name}: ${t2.desc}`).join("\n")}`
-              });
+              pushSoftInjection("system", `[Relevant tools for this task]
+${top.map((t2) => `- ${t2.name}: ${t2.desc}`).join("\n")}`);
             }
           }
           if (turn === 0 && (turnTier === "small" || turnTier === "medium")) {
@@ -520773,11 +520918,8 @@ ${top.map((t2) => `- ${t2.name}: ${t2.desc}`).join("\n")}`
             }
             hints.push("EFFICIENCY: Aim for 3-5 tool calls total. Each call should make measurable progress. Do not repeat a tool call with the same arguments.");
             if (hints.length > 0) {
-              messages2.push({
-                role: "system",
-                content: `[Efficiency Guide]
-${hints.join("\n")}`
-              });
+              pushSoftInjection("system", `[Efficiency Guide]
+${hints.join("\n")}`);
             }
           }
           if (turn === 0 && (turnTier === "small" || turnTier === "medium")) {
@@ -520787,21 +520929,18 @@ ${hints.join("\n")}`
             const hasMultiStepRequirement = taskGoal.length > 200 && (taskGoal.match(/\d\./g) || []).length >= 2;
             const isAnalysisTask = (taskGoal.match(/\banalyze\b|\baudit\b|\breview\b|\bdiagnose\b|\binvestigate\b|\bcompare\b|\bevaluate\b/gi) || []).length >= 1;
             if (hasMultiplePremises || hasConditionalLogic || hasMultiStepRequirement || isAnalysisTask) {
-              messages2.push({
-                role: "system",
-                content: [
-                  "[Structured Reasoning Guide]",
-                  "This task requires multi-step reasoning. Follow this structure:",
-                  "",
-                  "1. DECOMPOSE: List the sub-questions this task requires, from simplest to most complex.",
-                  "2. For each sub-question:",
-                  "   a. State what you KNOW (verified from evidence/tool output)",
-                  "   b. State what you ASSUME (hypotheses not yet confirmed)",
-                  "   c. Derive your conclusion using ONLY verified facts",
-                  "3. If a tool result contradicts your earlier reasoning, UPDATE your conclusions — don't ignore new evidence.",
-                  "4. Before your final answer, verify: does each conclusion follow from the evidence?"
-                ].join("\n")
-              });
+              pushSoftInjection("system", [
+                "[Structured Reasoning Guide]",
+                "This task requires multi-step reasoning. Follow this structure:",
+                "",
+                "1. DECOMPOSE: List the sub-questions this task requires, from simplest to most complex.",
+                "2. For each sub-question:",
+                "   a. State what you KNOW (verified from evidence/tool output)",
+                "   b. State what you ASSUME (hypotheses not yet confirmed)",
+                "   c. Derive your conclusion using ONLY verified facts",
+                "3. If a tool result contradicts your earlier reasoning, UPDATE your conclusions — don't ignore new evidence.",
+                "4. Before your final answer, verify: does each conclusion follow from the evidence?"
+              ].join("\n"));
             }
           }
           const turnBudget = turnTier === "small" ? 5 : turnTier === "medium" ? 8 : 0;
@@ -522354,6 +522493,39 @@ Your most recent tool calls SUCCEEDED. If the task is complete, call task_comple
               });
             }
           }
+          try {
+            const turnLogTail = toolCallLog.filter((t2) => t2.turn === turn || t2.turn === void 0);
+            const filesTouched = /* @__PURE__ */ new Set();
+            const shellPrefixes = /* @__PURE__ */ new Set();
+            let failuresThisTurn = 0;
+            for (const tc of turnLogTail) {
+              if (tc.success === false)
+                failuresThisTurn++;
+              if (["file_write", "file_edit", "batch_edit", "file_patch"].includes(tc.name)) {
+                const m2 = tc.argsKey?.match(/path=([^,]+)/);
+                if (m2 && m2[1])
+                  filesTouched.add(m2[1]);
+              }
+              if (tc.name === "shell") {
+                const cmdMatch = tc.argsKey?.match(/command=([^,]{0,200})/);
+                const cmd = cmdMatch?.[1] ?? "";
+                const prefix = cmd.replace(/^cd\s+\S+\s*&&\s*/, "").split(/\s+/).slice(0, 3).join(" ");
+                if (prefix)
+                  shellPrefixes.add(prefix);
+              }
+            }
+            const todosNow = this.readSessionTodos() || [];
+            const completedNow = todosNow.filter((t2) => t2.status === "completed").length;
+            stagnationWindow.push({
+              turn,
+              ts: Date.now(),
+              completedTodos: completedNow,
+              filesTouchedThisTurn: filesTouched,
+              failuresThisTurn,
+              shellPrefixesThisTurn: shellPrefixes
+            });
+          } catch {
+          }
         }
         let prevCycleToolCalls = toolCallCount;
         while (!completed && !this.aborted && this.options.bruteForce && bruteForceCycle < this.options.bruteForceMaxCycles) {

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "open-agents-ai",
-  "version": "0.187.474",
+  "version": "0.187.476",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "open-agents-ai",
-      "version": "0.187.474",
+      "version": "0.187.476",
       "hasInstallScript": true,
       "license": "CC-BY-NC-4.0",
       "dependencies": {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-agents-ai",
-  "version": "0.187.474",
+  "version": "0.187.476",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",

package/prompts/agentic/system-large.md

@@ -161,6 +161,26 @@ When you discover image files (png, jpg, gif, svg, webp, bmp) during codebase ex
 - ALWAYS run validation (tests, build, lint) after making changes
 - If tests fail, read the FULL error output. Fix the exact failing assertion or error.
 - Do NOT give up after a failure. Iterate: fix → test → fix → test until it passes.
+- task_complete is ONLY for actual completion or unrecoverable hardware/permission errors. Being stuck on a code/config problem is NEVER grounds for task_complete — use DIAGNOSTIC MODE below.
+
+### DIAGNOSTIC MODE — When You ARE Stuck, Slow Down and Investigate
+
+If you have tried 2+ approaches to the same blocker and both failed, **STOP attempting fixes** and enter diagnostic mode. Repeating fix-attempts on a misunderstood problem just wastes turns. Diagnose ROOT CAUSE first.
+
+**The diagnostic loop (one cycle per turn, NOT batched):**
+
+1. **READ THE FULL ERROR** — re-read the most recent failure output ENTIRELY. Don't skim the first 200 chars. If output is in a log packet, use `log_explore` with `op="errors"`, then `op="lines"` for context.
+2. **VERIFY ONE ASSUMPTION** — pick ONE thing you BELIEVE to be true and test it with the smallest possible command (e.g. "I think tailwindcss is installed" → `ls node_modules/tailwindcss/package.json`).
+3. **STATE A HYPOTHESIS in writing** before your next action. Then design ONE experiment that would CONFIRM or REFUTE it (not fix it — verify it first).
+4. **WEB SEARCH the exact error message** if you don't know what it means. A 30-second lookup beats 10 retry attempts.
+5. **CHECK THE OBVIOUS** — silent failures are common. `npm install` reporting "added 141 packages" doesn't mean ALL declared deps installed; npm sometimes drops packages with peer-dep conflicts without erroring. Verify each expected dep with `ls node_modules/<name>/package.json`.
+6. Only AFTER root cause is verified, attempt ONE fix targeting that cause. If the fix fails, return to step 1 with the new error.
+
+**What diagnostic mode is NOT:**
+- Trying another version (`tailwindcss@3.4.19` after `tailwindcss@4.0.0`) — that's variant-fatigue, not diagnosis.
+- Adding `--force` or `--legacy-peer-deps` — those mask root causes.
+- Wiping node_modules and re-installing — hides the original error.
+- Calling task_complete to escape — task_complete is NEVER the answer to a stuck debugging session.
 - Use grep_search and find_files for efficient exploration (don't dump entire directories)
 - Use file_edit for small changes instead of rewriting entire files
 - Keep tool calls focused — read only what you need

package/prompts/agentic/system-medium.md

@@ -94,6 +94,36 @@ NEVER write the entire document in ONE file_write call. DECOMPOSE:
 - Do NOT give up after failure. Iterate until it passes.
 - Use file_edit for small changes, not full file rewrites
 - You MUST call task_complete when done — when you have enough information from web tools, STOP fetching and call task_complete with a summary. Do not keep browsing after you have the answer.
+- task_complete is ONLY for actual completion or unrecoverable hardware/permission errors. Being stuck on a code/config problem is NEVER grounds for task_complete — use DIAGNOSTIC MODE below.
+
+### DIAGNOSTIC MODE — When You ARE Stuck, Slow Down and Investigate
+
+If you have tried 2+ approaches to the same blocker and both failed, **STOP attempting fixes** and enter diagnostic mode. Repeating fix-attempts on a misunderstood problem just wastes turns. Diagnose ROOT CAUSE first.
+
+**The diagnostic loop (one cycle per turn, NOT batched):**
+
+1. **READ THE FULL ERROR** — re-read the most recent failure output ENTIRELY. Don't skim the first 200 chars. If the output is in a log packet, use `log_explore` with `op="errors"` to see every marker, then `op="lines"` for surrounding context.
+
+2. **VERIFY ONE ASSUMPTION** — pick ONE thing you BELIEVE to be true and test it with the smallest possible command:
+   - "I think tailwindcss is installed" → `ls node_modules/tailwindcss/package.json` (one line)
+   - "I think the import path is right" → `cat src/lib/x.ts | head -5`
+   - "I think the env var is set" → `printenv VAR_NAME`
+
+3. **STATE A HYPOTHESIS in writing** before your next action:
+   - "Hypothesis: tailwindcss didn't install because @tailwindcss/postcss has a peer-dep conflict with autoprefixer."
+   - Then design ONE experiment that would CONFIRM or REFUTE it (not fix it — verify it first).
+
+4. **WEB SEARCH the exact error message** if you don't know what it means. `web_search("exact error string from terminal")`. A 30-second lookup beats 10 retry attempts.
+
+5. **CHECK THE OBVIOUS** — silent failures are common. `npm install` saying "added 141 packages" doesn't mean ALL declared deps installed; npm sometimes drops packages with peer-dep conflicts without erroring. Verify each expected dep with `ls node_modules/<name>/package.json`.
+
+6. Only AFTER root cause is verified, attempt ONE fix targeting that cause. If the fix fails, return to step 1 with the new error.
+
+**What diagnostic mode is NOT:**
+- Trying another version (`tailwindcss@3.4.19` after `tailwindcss@4.0.0` failed) — that's variant-fatigue, not diagnosis.
+- Adding `--force` or `--legacy-peer-deps` — those mask root causes, they don't reveal them.
+- Wiping node_modules and re-installing — that just hides the original error.
+- Calling task_complete to escape — task_complete is NEVER the answer to a stuck debugging session.
 - Do NOT output long explanations. Focus on tool calls.
 - If file_read/list_directory returns ENOENT, use list_directory on the project root — do NOT guess parent paths
 - Directory listing entries are RELATIVE to the listed directory. If you list "parent/" and see "child", the full path is "parent/child" — NOT ".child" or just "child"

package/prompts/agentic/system-small.md

@@ -99,10 +99,12 @@ Complex tasks (5+ steps) — DECOMPOSE before acting:
 1. Call todo_write with the checklist. Mark item 1 "in_progress".
 2. Execute ONE STEP AT A TIME. After each, update todo_write status.
 3. After each file edit, VERIFY: file_read or shell test.
-4. If stuck after 2 attempts, try a DIFFERENT approach — do not repeat the same tool call.
+4. If stuck after 2 attempts: STOP. Enter DIAGNOSTIC MODE — read the FULL error output, state a hypothesis in writing, verify ONE assumption with the smallest test command, web_search the exact error string. Only fix AFTER you've confirmed root cause. Do NOT keep trying variants of the same approach.
 5. For multi-file changes: read ALL relevant files first, then edit in dependency order.
 6. Final todo_write marks all items "completed", then call task_complete.
 
+task_complete is ONLY for ACTUAL completion. Being stuck on a code/config problem is NEVER grounds for task_complete — diagnose, do not exit.
+
 CRITICAL — NEVER repeat a tool call with the same arguments. If you already read a file, use the data you have. If you already ran a command, use the output. Calling the same tool twice with identical arguments wastes turns and produces the same result.
 
 Long document generation (reports, SOWs, proposals, contracts):