open-agents-ai 0.187.573 → 0.187.575

package/dist/index.js

@@ -252600,7 +252600,9 @@ async function probeService() {
   try {
     const controller = new AbortController();
     const timeout2 = setTimeout(() => controller.abort(), 3e3);
-    const res = await fetch(`${BASE_URL}/health`, { signal: controller.signal });
+    const res = await fetch(`${BASE_URL}/health`, {
+      signal: controller.signal
+    });
     clearTimeout(timeout2);
     return res.ok;
   } catch {
@@ -252610,7 +252612,10 @@ async function probeService() {
 function findPython3() {
   for (const cmd of ["python3", "python"]) {
     try {
-      const ver = execSync19(`${cmd} --version 2>&1`, { stdio: "pipe", timeout: 5e3 }).toString().trim();
+      const ver = execSync19(`${cmd} --version 2>&1`, {
+        stdio: "pipe",
+        timeout: 5e3
+      }).toString().trim();
       if (ver.includes("Python 3"))
         return cmd;
     } catch {
@@ -252682,7 +252687,10 @@ async function ensureSession() {
   });
   const data = await res.json();
   if (!data.ok)
-    return { error: String(data.message ?? "Failed to start browser session"), sessionId: "" };
+    return {
+      error: String(data.message ?? "Failed to start browser session"),
+      sessionId: ""
+    };
   activeSessionId = data.session_id;
   return { sessionId: activeSessionId };
 }
@@ -252828,7 +252836,22 @@ var init_browser_action = __esm({
         properties: {
           action: {
             type: "string",
-            enum: ["navigate", "click", "click_xy", "type", "screenshot", "dom", "dom_summary", "vision_click", "scroll", "scroll_up", "scroll_down", "back", "forward", "close"],
+            enum: [
+              "navigate",
+              "click",
+              "click_xy",
+              "type",
+              "screenshot",
+              "dom",
+              "dom_summary",
+              "vision_click",
+              "scroll",
+              "scroll_up",
+              "scroll_down",
+              "back",
+              "forward",
+              "close"
+            ],
             description: "Browser action to perform. Key actions:\n- 'dom_summary': compact view of interactive elements (~1KB vs 200KB raw DOM)\n- 'vision_click': screenshot the page, use Moondream vision to find an element by description, then click it. Pass the element description in 'text' parameter (e.g. text='the login button'). This is the visual grounding loop from SeeAct.\n- 'click': click by CSS selector (fastest when you know the selector)\n- 'click_xy': click at pixel coordinates (when you have exact coords)"
           },
           url: {
@@ -252858,12 +252881,44 @@ var init_browser_action = __esm({
         },
         required: ["action"]
       };
+      /** TASK-CLEANUP: gracefully close the browser session when the task completes. */
+      async cleanup() {
+        if (activeSessionId) {
+          try {
+            const res = await fetch(`${BASE_URL}/session/close`, {
+              method: "POST",
+              headers: { "Content-Type": "application/json" },
+              body: JSON.stringify({ sid: activeSessionId }),
+              signal: AbortSignal.timeout(5e3)
+            });
+            await res.json();
+          } catch {
+          }
+          activeSessionId = null;
+        }
+        if (serviceProcess && serviceProcess.pid && !serviceProcess.killed) {
+          try {
+            process.kill(-serviceProcess.pid, "SIGKILL");
+          } catch {
+          }
+          try {
+            serviceProcess.kill("SIGKILL");
+          } catch {
+          }
+          serviceProcess = null;
+        }
+      }
       async execute(args) {
         const start2 = Date.now();
         const action = args.action;
         const launchErr = await launchService();
         if (launchErr) {
-          return { success: false, output: "", error: launchErr, durationMs: Date.now() - start2 };
+          return {
+            success: false,
+            output: "",
+            error: launchErr,
+            durationMs: Date.now() - start2
+          };
         }
         if (action === "close") {
           if (activeSessionId) {
@@ -252873,21 +252928,41 @@ var init_browser_action = __esm({
             }
             activeSessionId = null;
           }
-          return { success: true, output: "Browser session closed.", durationMs: Date.now() - start2 };
+          return {
+            success: true,
+            output: "Browser session closed.",
+            durationMs: Date.now() - start2
+          };
         }
         const session = await ensureSession();
         if (session.error) {
-          return { success: false, output: "", error: session.error, durationMs: Date.now() - start2 };
+          return {
+            success: false,
+            output: "",
+            error: session.error,
+            durationMs: Date.now() - start2
+          };
         }
         try {
           let result;
           switch (action) {
             case "navigate": {
               if (!args.url)
-                return { success: false, output: "", error: "url is required for navigate action", durationMs: Date.now() - start2 };
-              result = await apiCall("/navigate", "POST", { url: args.url });
+                return {
+                  success: false,
+                  output: "",
+                  error: "url is required for navigate action",
+                  durationMs: Date.now() - start2
+                };
+              result = await apiCall("/navigate", "POST", {
+                url: args.url
+              });
               if (result.ok) {
-                return { success: true, output: `Navigated to ${args.url}`, durationMs: Date.now() - start2 };
+                return {
+                  success: true,
+                  output: `Navigated to ${args.url}`,
+                  durationMs: Date.now() - start2
+                };
               }
               const navMsg = String(result.message ?? "Navigation failed");
               const navHint = navMsg.toLowerCase().includes("connection") || navMsg.toLowerCase().includes("refused") || navMsg.toLowerCase().includes("err_connection") ? " (the URL appears unreachable — check if the target server is running and accepting connections)" : navMsg.toLowerCase().includes("timeout") ? " (page load timed out — try again or use a different URL)" : "";
@@ -252900,10 +252975,21 @@ var init_browser_action = __esm({
             }
             case "click": {
               if (!args.selector)
-                return { success: false, output: "", error: "selector is required for click action", durationMs: Date.now() - start2 };
-              result = await apiCall("/click", "POST", { selector: args.selector });
+                return {
+                  success: false,
+                  output: "",
+                  error: "selector is required for click action",
+                  durationMs: Date.now() - start2
+                };
+              result = await apiCall("/click", "POST", {
+                selector: args.selector
+              });
               if (result.ok) {
-                return { success: true, output: `Clicked element: ${args.selector}`, durationMs: Date.now() - start2 };
+                return {
+                  success: true,
+                  output: `Clicked element: ${args.selector}`,
+                  durationMs: Date.now() - start2
+                };
               }
               const clickMsg = String(result.message ?? "Click failed");
               return {
@@ -252915,10 +253001,19 @@ var init_browser_action = __esm({
             }
             case "click_xy": {
               if (args.x == null || args.y == null)
-                return { success: false, output: "", error: "x and y are required for click_xy action", durationMs: Date.now() - start2 };
+                return {
+                  success: false,
+                  output: "",
+                  error: "x and y are required for click_xy action",
+                  durationMs: Date.now() - start2
+                };
               result = await apiCall("/click_xy", "POST", { x: args.x, y: args.y });
               if (result.ok) {
-                return { success: true, output: `Clicked at (${args.x}, ${args.y})`, durationMs: Date.now() - start2 };
+                return {
+                  success: true,
+                  output: `Clicked at (${args.x}, ${args.y})`,
+                  durationMs: Date.now() - start2
+                };
               }
               const xyMsg = String(result.message ?? "Click failed");
               return {
@@ -252930,10 +253025,22 @@ var init_browser_action = __esm({
             }
             case "type": {
               if (!args.selector || !args.text)
-                return { success: false, output: "", error: "selector and text are required for type action", durationMs: Date.now() - start2 };
-              result = await apiCall("/type", "POST", { selector: args.selector, text: args.text });
+                return {
+                  success: false,
+                  output: "",
+                  error: "selector and text are required for type action",
+                  durationMs: Date.now() - start2
+                };
+              result = await apiCall("/type", "POST", {
+                selector: args.selector,
+                text: args.text
+              });
               if (result.ok) {
-                return { success: true, output: `Typed "${args.text.slice(0, 50)}" into ${args.selector}`, durationMs: Date.now() - start2 };
+                return {
+                  success: true,
+                  output: `Typed "${args.text.slice(0, 50)}" into ${args.selector}`,
+                  durationMs: Date.now() - start2
+                };
               }
               const typeMsg = String(result.message ?? "Type failed");
               return {
@@ -252981,16 +253088,30 @@ var init_browser_action = __esm({
                   durationMs: Date.now() - start2
                 };
               }
-              return { success: false, output: "", error: "Screenshot failed", durationMs: Date.now() - start2 };
+              return {
+                success: false,
+                output: "",
+                error: "Screenshot failed",
+                durationMs: Date.now() - start2
+              };
             }
             case "dom": {
               result = await apiCall("/dom", "GET");
               const dom = result.dom;
               if (dom) {
                 const truncated = dom.length > 5e4 ? dom.slice(0, 5e4) + "\n... (truncated)" : dom;
-                return { success: true, output: truncated, durationMs: Date.now() - start2 };
+                return {
+                  success: true,
+                  output: truncated,
+                  durationMs: Date.now() - start2
+                };
               }
-              return { success: false, output: "", error: "DOM capture failed", durationMs: Date.now() - start2 };
+              return {
+                success: false,
+                output: "",
+                error: "DOM capture failed",
+                durationMs: Date.now() - start2
+              };
             }
             // dom_summary: Research-grounded DOM downsampling
             // Paper: AgentOccam (arXiv:2410.13825, ICLR 2025) — pivotal node extraction
@@ -253004,9 +253125,18 @@ var init_browser_action = __esm({
               result = await apiCall("/dom", "GET");
               const rawDom = result.dom;
               if (!rawDom)
-                return { success: false, output: "", error: "DOM capture failed", durationMs: Date.now() - start2 };
+                return {
+                  success: false,
+                  output: "",
+                  error: "DOM capture failed",
+                  durationMs: Date.now() - start2
+                };
               const summary = downsampleDom(rawDom);
-              return { success: true, output: summary, durationMs: Date.now() - start2 };
+              return {
+                success: true,
+                output: summary,
+                durationMs: Date.now() - start2
+              };
             }
             // vision_click: Screenshot → Moondream point detection → Click
             // Paper: SeeAct (arXiv:2401.01614) — visual grounding for web agents
@@ -253019,14 +253149,24 @@ var init_browser_action = __esm({
             case "vision_click": {
               const target = args.text;
               if (!target)
-                return { success: false, output: "", error: "text parameter is required for vision_click — describe what to click (e.g. 'the login button')", durationMs: Date.now() - start2 };
+                return {
+                  success: false,
+                  output: "",
+                  error: "text parameter is required for vision_click — describe what to click (e.g. 'the login button')",
+                  durationMs: Date.now() - start2
+                };
               const ssResult = await apiCall("/screenshot", "GET");
               const ssB64 = ssResult.b64;
               const ssWidth = ssResult.width || 1280;
               const ssHeight = ssResult.height || 720;
               const ssFile = ssResult.file;
               if (!ssB64 && !ssFile) {
-                return { success: false, output: "", error: "Screenshot failed — cannot perform vision click", durationMs: Date.now() - start2 };
+                return {
+                  success: false,
+                  output: "",
+                  error: "Screenshot failed — cannot perform vision click",
+                  durationMs: Date.now() - start2
+                };
               }
               let imagePath = "";
               if (ssFile) {
@@ -253038,7 +253178,12 @@ var init_browser_action = __esm({
                   wfs(tmpPath, fileBuffer);
                   imagePath = tmpPath;
                 } catch (e2) {
-                  return { success: false, output: "", error: `Failed to save screenshot: ${e2}`, durationMs: Date.now() - start2 };
+                  return {
+                    success: false,
+                    output: "",
+                    error: `Failed to save screenshot: ${e2}`,
+                    durationMs: Date.now() - start2
+                  };
                 }
               } else if (ssB64) {
                 const tmpPath = join41(process.env["TMPDIR"] || "/tmp", `oa-vision-click-${Date.now()}.png`);
@@ -253056,7 +253201,12 @@ var init_browser_action = __esm({
                   prompt: target
                 });
                 if (!visionResult.success) {
-                  return { success: false, output: `Vision could not find "${target}" on the page. Try using dom_summary to find the CSS selector instead.`, error: visionResult.error, durationMs: Date.now() - start2 };
+                  return {
+                    success: false,
+                    output: `Vision could not find "${target}" on the page. Try using dom_summary to find the CSS selector instead.`,
+                    error: visionResult.error,
+                    durationMs: Date.now() - start2
+                  };
                 }
                 const coordMatch = visionResult.output.match(/\((\d+\.?\d*),\s*(\d+\.?\d*)\)/);
                 if (coordMatch) {
@@ -253066,10 +253216,19 @@ var init_browser_action = __esm({
                   pointY = Math.round(normY * ssHeight);
                 }
               } catch (e2) {
-                return { success: false, output: "", error: `Vision detection failed: ${e2}`, durationMs: Date.now() - start2 };
+                return {
+                  success: false,
+                  output: "",
+                  error: `Vision detection failed: ${e2}`,
+                  durationMs: Date.now() - start2
+                };
               }
               if (pointX < 0 || pointY < 0) {
-                return { success: false, output: `Could not determine click coordinates for "${target}". Vision returned no valid points.`, durationMs: Date.now() - start2 };
+                return {
+                  success: false,
+                  output: `Could not determine click coordinates for "${target}". Vision returned no valid points.`,
+                  durationMs: Date.now() - start2
+                };
               }
               const clickResult = await apiCall("/click_xy", "POST", {
                 x: pointX,
@@ -253092,22 +253251,49 @@ var init_browser_action = __esm({
               };
             }
             case "scroll":
-              result = await apiCall("/scroll", "POST", { amount: args.amount ?? 600 });
-              return { success: !!result.ok, output: `Scrolled ${args.amount ?? 600}px`, durationMs: Date.now() - start2 };
+              result = await apiCall("/scroll", "POST", {
+                amount: args.amount ?? 600
+              });
+              return {
+                success: !!result.ok,
+                output: `Scrolled ${args.amount ?? 600}px`,
+                durationMs: Date.now() - start2
+              };
             case "scroll_up":
               result = await apiCall("/scroll/up", "POST");
-              return { success: !!result.ok, output: "Scrolled up", durationMs: Date.now() - start2 };
+              return {
+                success: !!result.ok,
+                output: "Scrolled up",
+                durationMs: Date.now() - start2
+              };
             case "scroll_down":
               result = await apiCall("/scroll/down", "POST");
-              return { success: !!result.ok, output: "Scrolled down", durationMs: Date.now() - start2 };
+              return {
+                success: !!result.ok,
+                output: "Scrolled down",
+                durationMs: Date.now() - start2
+              };
             case "back":
               result = await apiCall("/history/back", "POST");
-              return { success: !!result.ok, output: "Navigated back", durationMs: Date.now() - start2 };
+              return {
+                success: !!result.ok,
+                output: "Navigated back",
+                durationMs: Date.now() - start2
+              };
             case "forward":
               result = await apiCall("/history/forward", "POST");
-              return { success: !!result.ok, output: "Navigated forward", durationMs: Date.now() - start2 };
+              return {
+                success: !!result.ok,
+                output: "Navigated forward",
+                durationMs: Date.now() - start2
+              };
             default:
-              return { success: false, output: "", error: `Unknown action: ${action}. Available: navigate, click, click_xy, type, screenshot, dom, scroll, scroll_up, scroll_down, back, forward, close`, durationMs: Date.now() - start2 };
+              return {
+                success: false,
+                output: "",
+                error: `Unknown action: ${action}. Available: navigate, click, click_xy, type, screenshot, dom, scroll, scroll_up, scroll_down, back, forward, close`,
+                durationMs: Date.now() - start2
+              };
           }
         } catch (err) {
           return {
@@ -527925,6 +528111,10 @@ var init_agenticRunner = __esm({
       // can re-emit the same plan a second time (plan-replay) and execute
       // duplicate work because PROGRESS NUDGE alone is informational.
       _progressGateActive = false;
+      // Consecutive gate blocks count. When the model ignores the gate and
+      // retries a blocked tool, this counter increments. ≥2 triggers a system
+      // message escalation to break pattern-lock loops.
+      _consecutiveGateBlocks = 0;
       // REG-5: Rolling buffer of recent tool failures with their error output.
       // Surfaced before every LLM call so the agent can't ignore "I just ran this
       // and it errored". Detects same-fingerprint failure repetition and escalates
@@ -528174,15 +528364,19 @@ var init_agenticRunner = __esm({
       // DECOMP-2 (root-cause from batch531-midi-decomp, 2026-05-03): compelling
       // sub_agent delegation. DECOMP-1's informational directive was ignored
       // (0 sub_agent calls in 466 tool-call run despite directive at turn 1).
-      // Mirrors the BFC-61.G escalation arc: when the agent has edited
-      // ≥THRESHOLD distinct files in main context WITHOUT successful sub_agent,
+      // Mirrors the BFC-61.G escalation arc, but must not deadlock delivery:
+      // when the agent has edited ≥adaptive-threshold distinct files in main context
+      // WITHOUT successful sub_agent,
       // the dispatcher BLOCKS edits to NEW files (paths not yet edited) until
       // sub_agent succeeds. Edits to already-touched files are still allowed
-      // (current-module finishing work). Failed or malformed delegation does
-      // not clear the gate.
+      // (current-module finishing work). Repeated failed delegation attempts
+      // unlock a main-context fallback so the guardrail cannot become a hard
+      // write-deadlock when sub_agent itself is broken or unavailable.
       // Kill switch: OA_DISABLE_DECOMP2=1.
       _decomp2MainContextFiles = /* @__PURE__ */ new Set();
       _decomp2SubAgentCalls = 0;
+      _decomp2FailedDelegationCalls = 0;
+      _decomp2FallbackAllowed = false;
       _decomp2GateActive = false;
       // MEM_PATH item #9: adaptive retrieval cache. When the (goalHash, recent-tool-sig)
       // hasn't changed since last retrieval, skip the PPR call entirely and reuse
@@ -528727,6 +528921,8 @@ Pick the SMALLEST concrete deliverable from the spec — typically the project e
       _maybeDecomp2Block(tc, turn) {
         if (!this._decomp2GateActive)
           return null;
+        if (this._decomp2FallbackAllowed)
+          return null;
         if (process.env["OA_DISABLE_DECOMP2"] === "1")
           return null;
         const _editTools = /* @__PURE__ */ new Set([
@@ -528747,7 +528943,7 @@ Pick the SMALLEST concrete deliverable from the spec — typically the project e
         const decomp2Msg = [
           `[BLOCKED — DECOMP-2 main-context exhaustion]`,
           ``,
-          `You have already edited ${this._decomp2MainContextFiles.size} distinct files in main context without invoking sub_agent. Continuing to edit ANOTHER new file ('${_editPath}') will keep your context window saturated and trigger compaction thrashing.`,
+          `You have already edited ${this._decomp2MainContextFiles.size} distinct files in main context without a successful sub_agent. Continuing to edit another new file ('${_editPath}') may keep your context window saturated and trigger compaction thrashing.`,
           ``,
           `Files you've already edited (will accept further edits to these):`,
           _filesList,
@@ -528763,13 +528959,15 @@ Pick the SMALLEST concrete deliverable from the spec — typically the project e
           `     })`,
           `  3. After sub_agent returns, mark the todo completed.`,
           ``,
+          `If sub_agent keeps failing for reasons outside the module work, retry it once with corrected arguments. After repeated failed delegation attempts, OA will downgrade this from a hard block to an advisory fallback so file writes can continue.`,
+          ``,
           `Why this matters: spreading edits across N files in main context burns ~N × file_size tokens. sub_agent gives the next module a focused context window.`,
           ``,
           `If you have ALREADY edited '${_editPath}' (this is a continuation), the orchestrator's set must have missed it — call file_read to verify, then re-edit. Otherwise, dispatch sub_agent now.`
         ].join("\n");
         this.emit({
           type: "status",
-          content: `DECOMP-2 NEW-FILE BLOCK — rejected ${tc.name}('${_editPath}') at turn ${turn}; gate stays active until sub_agent succeeds`,
+          content: `DECOMP-2 NEW-FILE BLOCK — rejected ${tc.name}('${_editPath}') at turn ${turn}; gate stays active until sub_agent succeeds or repeated delegation failure unlocks fallback`,
           timestamp: (/* @__PURE__ */ new Date()).toISOString()
         });
         this._tagSyntheticFailure({
@@ -528778,6 +528976,12 @@ Pick the SMALLEST concrete deliverable from the spec — typically the project e
         });
         return decomp2Msg;
       }
+      _decomp2FileSpreadThreshold() {
+        const ctx3 = this.options.contextWindowSize ?? 0;
+        if (ctx3 <= 0)
+          return 5;
+        return Math.max(5, Math.min(30, Math.round(ctx3 / 6400)));
+      }
       /**
        * DECOMP-2 post-dispatch tracking. Refactored from inline so both the
        * main turn loop AND the brute-force re-engagement inner loop record
@@ -528804,12 +529008,12 @@ Pick the SMALLEST concrete deliverable from the spec — typically the project e
           const _editPaths = this._extractToolTargetPaths(tc.name, tc.arguments, result);
           for (const _editPath of _editPaths) {
             this._decomp2MainContextFiles.add(_editPath);
-            const DECOMP2_FILE_SPREAD_THRESHOLD = 5;
-            if (!this._decomp2GateActive && this._decomp2MainContextFiles.size >= DECOMP2_FILE_SPREAD_THRESHOLD && this._decomp2SubAgentCalls === 0) {
+            const DECOMP2_FILE_SPREAD_THRESHOLD = this._decomp2FileSpreadThreshold();
+            if (!this._decomp2GateActive && !this._decomp2FallbackAllowed && this._decomp2MainContextFiles.size >= DECOMP2_FILE_SPREAD_THRESHOLD && this._decomp2SubAgentCalls === 0) {
               this._decomp2GateActive = true;
               this.emit({
                 type: "status",
-                content: `DECOMP-2 NEW-FILE GATE ACTIVATED — ${this._decomp2MainContextFiles.size} distinct files edited in main context, 0 successful sub_agent calls; further edits to NEW files will be blocked until sub_agent succeeds`,
+                content: `DECOMP-2 NEW-FILE GATE ACTIVATED — ${this._decomp2MainContextFiles.size} distinct files edited in main context, 0 successful sub_agent calls, threshold=${DECOMP2_FILE_SPREAD_THRESHOLD}; further edits to NEW files will be blocked until sub_agent succeeds or repeated delegation failure unlocks fallback`,
                 timestamp: (/* @__PURE__ */ new Date()).toISOString()
               });
             }
@@ -528818,15 +529022,27 @@ Pick the SMALLEST concrete deliverable from the spec — typically the project e
         if (tc.name === "sub_agent" || tc.name === "priority_delegate" || tc.name === "background_run") {
           if (result?.success !== true) {
             if (this._decomp2GateActive) {
+              this._decomp2FailedDelegationCalls++;
               this.emit({
                 type: "status",
                 content: `DECOMP-2 DELEGATION FAILED — '${tc.name}' did not clear gate at turn ${turn}; fix delegation arguments/result before editing another new file`,
                 timestamp: (/* @__PURE__ */ new Date()).toISOString()
               });
+              if (this._decomp2FailedDelegationCalls >= 2) {
+                this._decomp2FallbackAllowed = true;
+                this._decomp2GateActive = false;
+                this.emit({
+                  type: "status",
+                  content: `DECOMP-2 FALLBACK UNLOCKED — ${this._decomp2FailedDelegationCalls} failed delegation attempts while gate was active; allowing main-context new-file edits so work can continue`,
+                  timestamp: (/* @__PURE__ */ new Date()).toISOString()
+                });
+              }
             }
             return;
           }
           this._decomp2SubAgentCalls++;
+          this._decomp2FailedDelegationCalls = 0;
+          this._decomp2FallbackAllowed = false;
           if (this._decomp2GateActive) {
             this._decomp2GateActive = false;
             this.emit({
@@ -530914,6 +531130,8 @@ Respond with your assessment, then take action.`;
         this._reg61PerpetualGateActive = false;
         this._decomp2MainContextFiles = /* @__PURE__ */ new Set();
         this._decomp2SubAgentCalls = 0;
+        this._decomp2FailedDelegationCalls = 0;
+        this._decomp2FallbackAllowed = false;
         this._decomp2GateActive = false;
         if (!globalThis.__oa_rca1_sigterm_installed) {
           globalThis.__oa_rca1_sigterm_installed = true;
@@ -533252,22 +533470,25 @@ ${memoryLines.join("\n")}`
                     recentWrites.push({ path: path11, turn: info.lastWriteTurn ?? 0 });
                   }
                 }
+                this._consecutiveGateBlocks++;
                 recentWrites.sort((a2, b) => b.turn - a2.turn);
                 const showWrites = recentWrites.slice(0, 16);
+                const isRepeat = this._consecutiveGateBlocks >= 2;
                 const gateMsg = [
-                  `[PROGRESS GATE — call todo_write FIRST before any other tool]`,
+                  `[BLOCKED — PROGRESS GATE active]`,
                   ``,
-                  `You have completed ${this._writesSinceLastTodoWrite} file modification${this._writesSinceLastTodoWrite === 1 ? "" : "s"} since your last todo_write call.`,
-                  `The next tool call MUST be todo_write to mark progress. This is enforced — non-todo tool calls are intercepted until plan state is updated.`,
+                  `CAUSE: ${this._writesSinceLastTodoWrite} file writes since last todo_write call. Without progress tracking, the next turn re-plans the same work (plan-replay).`,
+                  `EFFECT: All non-todo tool calls are now blocked at the runtime level.`,
+                  `ACTION REQUIRED: Call todo_write with updated progress to release the gate.`,
+                  `CONSEQUENCE OF IGNORING: Retrying blocked tools does NOT work — only todo_write is accepted while the gate is active.`,
+                  isRepeat ? `
+[ESCALATION: This is block #${this._consecutiveGateBlocks}. You keep calling blocked tools instead of todo_write. The gate cannot be bypassed. You MUST call todo_write next.]` : "",
                   ``,
                   `Recent file modifications (use these to decide what's done):`,
                   ...showWrites.map((w) => `  • ${w.path} (turn ${w.turn})`),
                   recentWrites.length > showWrites.length ? `  • ... +${recentWrites.length - showWrites.length} more` : "",
                   ``,
-                  `Required action: call todo_write with the updated todo array — mark anything completed that these writes satisfy, advance the next item to in_progress, keep the rest pending.`,
-                  `After todo_write succeeds, this gate releases and you can continue normal work.`,
-                  ``,
-                  `Why this exists: without the explicit progress update, your next turn will see the same in_progress todo, re-plan the same work, and re-emit identical tool calls (the "plan replay" failure mode that causes byte-identical writes to appear twice).`
+                  `Format: todo_write with todos array — mark items completed that these writes satisfy, advance next to in_progress. After todo_write succeeds, normal tools resume.`
                 ].filter(Boolean).join("\n");
                 this.emit({
                   type: "tool_result",
@@ -533753,6 +533974,7 @@ Respond with EXACTLY this structure before your next tool call:
                   }
                   this._writesSinceLastTodoWrite = 0;
                   this._progressGateActive = false;
+                  this._consecutiveGateBlocks = 0;
                 }
                 if (tc.name === "file_read") {
                   const p2 = String(tc.arguments?.["path"] ?? tc.arguments?.["file"] ?? "");
@@ -534761,6 +534983,12 @@ Then use file_read on individual FILES inside it.`);
                   const output = sr.result.success ? sr.result.output : `Error: ${sr.result.error || "unknown"}
 ${sr.result.output}`;
                   messages2.push(this.buildToolMessage(output, matchTc.id, matchTc.name));
+                  if (this._consecutiveGateBlocks >= 2 && this._progressGateActive) {
+                    messages2.push({
+                      role: "system",
+                      content: `[PROGRESS GATE ESCALATION] You have made ${this._consecutiveGateBlocks} consecutive blocked tool calls without calling todo_write. The gate is enforced at the runtime level — retrying the same blocked tool will never work. Your NEXT call MUST be todo_write(todos=[...]) with updated progress. No other tool will be accepted until the gate is released.`
+                    });
+                  }
                   if (matchTc.name === "task_complete") {
                     const open2 = this.getOpenTodoItems();
                     if (open2.length > 0) {
@@ -534801,6 +535029,12 @@ ${sr.result.output}`;
                   const r2 = await executeSingle(tc);
                   if (r2) {
                     messages2.push(this.buildToolMessage(r2.output, r2.tc.id, r2.tc.name));
+                    if (this._consecutiveGateBlocks >= 2 && this._progressGateActive) {
+                      messages2.push({
+                        role: "system",
+                        content: `[PROGRESS GATE ESCALATION] You have made ${this._consecutiveGateBlocks} consecutive blocked tool calls without calling todo_write. The gate is enforced at the runtime level — retrying the same blocked tool will never work. Your NEXT call MUST be todo_write(todos=[...]) with updated progress. No other tool will be accepted until the gate is released.`
+                      });
+                    }
                     if (r2.tc.name === "task_complete") {
                       const open2 = this.getOpenTodoItems();
                       if (open2.length > 0) {
@@ -534893,6 +535127,14 @@ ${sr.result.output}`;
                         } else {
                           completed = true;
                           summary = extractTaskCompleteSummary(r2.tc.arguments);
+                          for (const tool of this.tools.values()) {
+                            if (tool.cleanup) {
+                              try {
+                                await tool.cleanup();
+                              } catch {
+                              }
+                            }
+                          }
                           if (summary && !this._assistantTextEmitted) {
                             this.emit({
                               type: "assistant_text",
@@ -613853,7 +614095,9 @@ function adaptTool6(tool) {
         output: result.output,
         error: result.error
       };
-    }
+    },
+    // Pass through lifecycle hooks from the underlying Tool implementation
+    cleanup: tool.cleanup
   };
 }
 function scanForSessionSignals(toolOutput) {

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "open-agents-ai",
-  "version": "0.187.573",
+  "version": "0.187.575",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "open-agents-ai",
-      "version": "0.187.573",
+      "version": "0.187.575",
       "hasInstallScript": true,
       "license": "CC-BY-NC-4.0",
       "dependencies": {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-agents-ai",
-  "version": "0.187.573",
+  "version": "0.187.575",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",