omnius 1.0.211 → 1.0.212

package/dist/index.js

@@ -559029,6 +559029,7 @@ var init_agenticRunner = __esm({
           bruteForce: options2?.bruteForce ?? true,
           bruteForceMaxCycles: options2?.bruteForceMaxCycles ?? 100,
           allowTurnExtension: options2?.allowTurnExtension ?? true,
+          completionProvenanceGuard: options2?.completionProvenanceGuard ?? true,
           modelTier: options2?.modelTier ?? "large",
           contextWindowSize: options2?.contextWindowSize ?? 0,
           personality: options2?.personality ?? PERSONALITY_PRESETS.balanced,
@@ -559898,6 +559899,198 @@ Do NOT call task_complete until all items are marked completed via todo_write.`;
           `Continue the work loop: inspect the failed evidence, make the smallest targeted fix, then rerun the relevant verification. Use the full available verification spectrum for the artifact you changed: static syntax, build/typecheck, tests, service startup, runtime logs, browser/page errors, console output, network failures, screenshots, accessibility/DOM state, and end-to-end user flow checks where applicable. The exact tools are stack-dependent; the standard is objective runtime evidence, not self-report.`
         ].join("\n");
       }
+      buildMissionCompletionContract(task, context2) {
+        if (process.env["OMNIUS_DISABLE_MISSION_COMPLETION_CONTRACT"] === "1")
+          return "";
+        const profile = this._inferCompletionProfile(`${task}
+${context2 ?? ""}`);
+        const requirements = [];
+        if (profile.browser)
+          requirements.push("browser/UI state must be proven by a post-action screenshot/DOM/observe_bundle pass");
+        if (profile.desktop)
+          requirements.push("desktop state must be proven by vision_action_loop observe or desktop_describe after the final action");
+        if (profile.code)
+          requirements.push("code/file changes must be proven by a relevant build/test/typecheck/runtime command after the last edit");
+        if (profile.research)
+          requirements.push("research/root-cause claims must cite concrete inspected files, commands, logs, or source artifacts");
+        if (requirements.length === 0)
+          requirements.push("final claims must name the concrete evidence used or state that the task required no external action");
+        return [
+          `[MISSION COMPLETION CONTRACT]`,
+          `Current ask: ${task.slice(0, 500)}`,
+          ``,
+          `Before claiming success or calling task_complete, satisfy the mission-specific evidence requirements:`,
+          ...requirements.map((line) => `- ${line}.`),
+          ``,
+          `The final task_complete summary for any action-heavy task must include a compact Provenance/Evidence note naming the validating tool output, command, screenshot, DOM state, file path, or blocker. Self-confidence is not evidence.`,
+          `For browser/form/account/send flows: after the last click/type/navigate/submit action, capture a fresh browser observation and verify the visible final state before completion.`,
+          `If completion is impossible, use a summary beginning BLOCKED: and name the exact blocker plus the evidence already collected.`
+        ].join("\n");
+      }
+      _inferCompletionProfile(text) {
+        const t2 = text.toLowerCase();
+        const browser3 = /\b(browser|web\s*page|website|page|playwright|selenium|chromedriver|chrome|headless|gui|proton|login|captcha|form|account|compose|mail|submit|click|type|fill)\b/.test(t2);
+        const desktop = /\b(desktop|screen|application|app\b|window|file manager|open a file|laptop|screenshot|vision_action_loop|desktop_describe|desktop_click)\b/.test(t2);
+        const code8 = /\b(implement|fix|patch|refactor|rewrite|build|compile|typecheck|test suite|unit test|integration test|source file|codebase|package|typescript|javascript|python|rust|golang)\b/.test(t2);
+        const research = /\b(discover|root cause|triage|deep dive|review|audit|investigate|prove|validate|forensics|diagnostic|failure mode)\b/.test(t2);
+        const formLike = /\b(form|fill|submit|signup|sign up|login|log in|account|compose|send|sent|mail|captcha|checkout|payment|upload)\b/.test(t2);
+        return { browser: browser3, desktop, code: code8, research, formLike };
+      }
+      _completionSummaryHasProvenance(summary) {
+        return /\b(provenance|evidence|verified|validated|confirmed|observed|screenshot|dom|console|network|log|test|typecheck|build|passed|opened|sent|created|submitted|blocked)\b/i.test(summary);
+      }
+      _isBlockedCompletionSummary(summary) {
+        return /^\s*(?:BLOCKED|PARTIAL|NO FILE CHANGES REQUIRED)\b/i.test(summary);
+      }
+      _browserActionKind(entry) {
+        if (!/^(browser_action|playwright_browser|carbonyl_browser)$/.test(entry.name))
+          return "other";
+        const args = this._parseExactArgsKey(entry.argsKey);
+        const action = String(args.get("action") ?? args.get("tool") ?? args.get("command") ?? "").toLowerCase();
+        if (/^(screenshot|dom|dom_summary|observe|observe_bundle|page_errors|console_logs|network_log|accessibility|snapshot|state|url|title)$/.test(action)) {
+          return "observe";
+        }
+        if (/^(navigate|click|click_xy|vision_click|visual_click|type|fill|press|select|submit|evaluate|scroll|scroll_up|scroll_down|back|forward)$/.test(action)) {
+          return "state";
+        }
+        return entry.name === "carbonyl_browser" ? "state" : "other";
+      }
+      _desktopActionKind(entry) {
+        if (/^(desktop_describe|screenshot)$/.test(entry.name))
+          return "observe";
+        if (entry.name === "desktop_click")
+          return "state";
+        if (entry.name !== "vision_action_loop")
+          return "other";
+        const args = this._parseExactArgsKey(entry.argsKey);
+        const action = String(args.get("action") ?? "").toLowerCase();
+        if (/^(observe|screenshot|describe|ocr|state)$/.test(action))
+          return "observe";
+        if (/^(click|type|key|press|open|focus|drag|scroll)$/.test(action))
+          return "state";
+        return "other";
+      }
+      _isVerificationShell(entry) {
+        if (entry.name !== "shell" && entry.name !== "background_run")
+          return false;
+        if (entry.success !== true)
+          return false;
+        const args = this._parseExactArgsKey(entry.argsKey);
+        const command = String(args.get("command") ?? args.get("cmd") ?? "").toLowerCase();
+        return /\b(test|typecheck|check|build|compile|verify|lint|pytest|vitest|jest|playwright|cypress|tsc|cargo\s+test|go\s+test|npm\s+run|pnpm\s+run|yarn\s+run)\b/.test(command);
+      }
+      _evaluateCompletionProvenanceGate(input) {
+        if (this.options.completionProvenanceGuard === false)
+          return { proceed: true };
+        if (process.env["OMNIUS_DISABLE_COMPLETION_PROVENANCE_GUARD"] === "1")
+          return { proceed: true };
+        const summary = input.summary || "";
+        const blockedSummary = this._isBlockedCompletionSummary(summary);
+        const profile = this._inferCompletionProfile(input.taskGoal);
+        const log22 = input.toolCallLog.filter((entry) => entry.name !== "task_complete");
+        const browserUsed = log22.some((entry) => /^(browser_action|playwright_browser|carbonyl_browser)$/.test(entry.name));
+        const desktopUsed = log22.some((entry) => /^(desktop_describe|desktop_click|vision_action_loop|screenshot)$/.test(entry.name));
+        const mutated = log22.some((entry) => entry.mutated === true);
+        const issues = [];
+        const actionHeavy = profile.browser || profile.desktop || profile.code || profile.research || browserUsed || desktopUsed || mutated || this._fileWritesThisRun > 0;
+        if (!actionHeavy)
+          return { proceed: true };
+        if (blockedSummary)
+          return { proceed: true };
+        const successfulNonCompletion = log22.filter((entry) => entry.success === true);
+        if (successfulNonCompletion.length === 0) {
+          issues.push("No successful objective tool result is recorded for this action-oriented task.");
+        }
+        const requiresBrowser = profile.browser || browserUsed;
+        const requiresDesktop = profile.desktop || desktopUsed;
+        if (requiresBrowser) {
+          let lastStateIdx = -1;
+          let lastObserveIdx = -1;
+          log22.forEach((entry, idx) => {
+            if (entry.success !== true)
+              return;
+            const kind = this._browserActionKind(entry);
+            if (kind === "state")
+              lastStateIdx = idx;
+            if (kind === "observe")
+              lastObserveIdx = idx;
+          });
+          if (lastStateIdx >= 0 && lastObserveIdx <= lastStateIdx) {
+            issues.push("Browser state changed after the last browser observation. Capture a fresh screenshot/DOM/observe_bundle after the final click/type/navigate/submit before completion.");
+          } else if (profile.formLike && lastObserveIdx < 0) {
+            issues.push("This looks like a form/account/send flow, but no successful post-action browser observation is recorded.");
+          }
+        }
+        if (requiresDesktop) {
+          let lastStateIdx = -1;
+          let lastObserveIdx = -1;
+          log22.forEach((entry, idx) => {
+            if (entry.success !== true)
+              return;
+            const kind = this._desktopActionKind(entry);
+            if (kind === "state")
+              lastStateIdx = idx;
+            if (kind === "observe")
+              lastObserveIdx = idx;
+          });
+          if (lastStateIdx >= 0 && lastObserveIdx <= lastStateIdx) {
+            issues.push("Desktop state changed after the last visual observation. Run vision_action_loop observe or desktop_describe after the final desktop action before completion.");
+          }
+        }
+        const mutatedEntries = log22.map((entry, idx) => ({ entry, idx })).filter(({ entry }) => entry.mutated === true);
+        if (profile.code || mutatedEntries.length > 0 || this._fileWritesThisRun > 0) {
+          const lastMutationIdx = mutatedEntries.length > 0 ? Math.max(...mutatedEntries.map(({ idx }) => idx)) : -1;
+          const verifiedAfterMutation = log22.some((entry, idx) => {
+            if (lastMutationIdx >= 0 && idx <= lastMutationIdx)
+              return false;
+            if (this._isVerificationShell(entry))
+              return true;
+            if (entry.success === true && this._browserActionKind(entry) === "observe")
+              return true;
+            if (entry.success === true && this._desktopActionKind(entry) === "observe")
+              return true;
+            return false;
+          });
+          if (lastMutationIdx >= 0 && !verifiedAfterMutation) {
+            issues.push("Files were changed but no successful verification command or runtime observation appears after the last mutation.");
+          }
+        }
+        const lastNonCompletion = [...log22].reverse().find(Boolean);
+        if (lastNonCompletion?.success === false) {
+          issues.push(`The most recent non-completion tool result failed (${lastNonCompletion.name}); resolve or explicitly report BLOCKED before completing.`);
+        }
+        if (!this._completionSummaryHasProvenance(summary)) {
+          issues.push("The completion summary does not include an explicit Evidence/Provenance note.");
+        }
+        if (issues.length === 0)
+          return { proceed: true };
+        const recentEvidence = successfulNonCompletion.slice(-6).map((entry) => {
+          const preview = (entry.outputPreview ?? "").replace(/\s+/g, " ").slice(0, 160);
+          return `- ${entry.name}(${entry.argsKey.slice(0, 120)})${preview ? ` -> ${preview}` : ""}`;
+        }).join("\n");
+        return {
+          proceed: false,
+          reason: issues[0] ?? "missing provenance",
+          feedback: [
+            `[COMPLETION PROVENANCE REQUIRED]`,
+            ``,
+            `You attempted to finish, but the completion claim is not yet proven against the current mission.`,
+            ``,
+            `Blocking issues:`,
+            ...issues.map((issue, index) => `${index + 1}. ${issue}`),
+            ``,
+            recentEvidence ? `Recent successful evidence already available:
+${recentEvidence}` : `Recent successful evidence already available: none recorded.`,
+            ``,
+            `Do the smallest missing verification step now. For browser/UI work, take a fresh screenshot/DOM/observe_bundle after the final action. For desktop work, run vision_action_loop observe or desktop_describe after the final action. For code/file changes, run the relevant build/test/typecheck/runtime check after the last edit.`,
+            ``,
+            `Only then call task_complete with this shape:`,
+            `Summary: <what changed or what final state was reached>`,
+            `Provenance: <tool/command/screenshot/DOM/log/file evidence proving it>`,
+            `If impossible, call task_complete with summary starting BLOCKED: and name the blocker plus evidence.`
+          ].join("\n")
+        };
+      }
       /**
        * REG-47: post-implementation backward-pass review.
        *
@@ -562728,7 +562921,6 @@ Respond with your assessment, then take action.`;
         const contextComposition = await this.assembleContext(task, context2);
         const systemPrompt = contextComposition.assembled;
         this._contextTree = new ContextTree(`sys-${systemPrompt.length}`, cleanedTask.slice(0, 200));
-        this._phaseMessageStartIdx = 2;
         this.emit({
           type: "status",
           content: `Context assembled: ${contextComposition.sections.map((s2) => `${s2.label}(${s2.tokenEstimate}t)`).join(" + ")} = ~${contextComposition.totalTokenEstimate}t`,
@@ -562772,10 +562964,13 @@ TASK: ${scrubbedTask}` : scrubbedTask;
             }
           });
         }
+        const missionCompletionContract = this.buildMissionCompletionContract(cleanedTask, context2);
         const messages2 = [
           { role: "system", content: systemPrompt },
+          ...missionCompletionContract ? [{ role: "system", content: missionCompletionContract }] : [],
           { role: "user", content: userContent }
         ];
+        this._phaseMessageStartIdx = messages2.length;
         if (process.env["OMNIUS_DISABLE_DECOMP1"] !== "1") {
           try {
             const _taskBodyForDecomp = typeof userContent === "string" ? userContent : "";
@@ -563033,6 +563228,24 @@ TASK: ${scrubbedTask}` : scrubbedTask;
           });
           return true;
         };
+        const holdProvenanceTaskComplete = (args, turn) => {
+          const proposedSummary = extractTaskCompleteSummary(args);
+          const gate = this._evaluateCompletionProvenanceGate({
+            summary: proposedSummary,
+            taskGoal: cleanedTask,
+            toolCallLog
+          });
+          if (gate.proceed)
+            return false;
+          messages2.push({ role: "system", content: gate.feedback });
+          this.emit({
+            type: "status",
+            content: `task_complete held by completion provenance guard: ${gate.reason}`,
+            turn,
+            timestamp: (/* @__PURE__ */ new Date()).toISOString()
+          });
+          return true;
+        };
         const turnCap = this.options.maxTurns && this.options.maxTurns > 0 ? this.options.maxTurns : Number.MAX_SAFE_INTEGER;
         for (let turn = 0; turn < turnCap; turn++) {
           clearTurnState(this._appState);
@@ -566251,6 +566464,13 @@ Respond with EXACTLY this structure before your next tool call:
                 editFeedbackRequiredBeforeMoreEdits = this._buildBatchEditAtomicAbortGuidance(tc.arguments);
                 this.pendingUserMessages.push(editFeedbackRequiredBeforeMoreEdits);
               }
+              const currentLogEntry = toolCallLog[_toolLogTailIdx];
+              if (currentLogEntry) {
+                currentLogEntry.success = result.success;
+                currentLogEntry.mutated = realFileMutation;
+                currentLogEntry.mutatedFiles = realMutationPaths;
+                currentLogEntry.outputPreview = (result.success ? result.llmContent ?? result.output ?? output : result.error ?? result.output ?? output).toString().slice(0, 500);
+              }
               this.emit({
                 type: "tool_result",
                 toolName: tc.name,
@@ -566640,7 +566860,7 @@ ${sr.result.output}`;
                         timestamp: (/* @__PURE__ */ new Date()).toISOString()
                       });
                     } else {
-                      if (holdNoProgressTaskComplete(matchTc.arguments, turn)) {
+                      if (holdNoProgressTaskComplete(matchTc.arguments, turn) || holdProvenanceTaskComplete(matchTc.arguments, turn)) {
                         continue;
                       }
                       const _bp1 = await this._runBackwardPassReview(turn);
@@ -566696,7 +566916,7 @@ ${sr.result.output}`;
                           timestamp: (/* @__PURE__ */ new Date()).toISOString()
                         });
                       } else {
-                        if (holdNoProgressTaskComplete(r2.tc.arguments, turn)) {
+                        if (holdNoProgressTaskComplete(r2.tc.arguments, turn) || holdProvenanceTaskComplete(r2.tc.arguments, turn)) {
                           continue;
                         }
                         const _bp2 = await this._runBackwardPassReview(turn);
@@ -566788,7 +567008,7 @@ ${sr.result.output}`;
                           timestamp: (/* @__PURE__ */ new Date()).toISOString()
                         });
                       } else {
-                        if (holdNoProgressTaskComplete(r2.tc.arguments, turn)) {
+                        if (holdNoProgressTaskComplete(r2.tc.arguments, turn) || holdProvenanceTaskComplete(r2.tc.arguments, turn)) {
                           continue;
                         }
                         const _bp3 = await this._runBackwardPassReview(turn);
@@ -567006,6 +567226,10 @@ Call task_complete(summary="...") NOW with whatever you have.`
               });
             }
             if (/task.?complete|all tests pass/i.test(content)) {
+              const completionArgs = { summary: content };
+              if (holdNoProgressTaskComplete(completionArgs, turn) || holdProvenanceTaskComplete(completionArgs, turn)) {
+                continue;
+              }
               completed = true;
               summary = content;
               break;
@@ -567568,7 +567792,7 @@ Full content available via: repl_exec(code="data = retrieve('${handleId}')")  or
                       timestamp: (/* @__PURE__ */ new Date()).toISOString()
                     });
                   } else {
-                    if (holdNoProgressTaskComplete(tc.arguments, turn)) {
+                    if (holdNoProgressTaskComplete(tc.arguments, turn) || holdProvenanceTaskComplete(tc.arguments, turn)) {
                       continue;
                     }
                     const _bp4 = await this._runBackwardPassReview(turn);
@@ -567621,22 +567845,9 @@ Full content available via: repl_exec(code="data = retrieve('${handleId}')")  or
                     timestamp: (/* @__PURE__ */ new Date()).toISOString()
                   });
                 } else {
-                  if (process.env["OMNIUS_DISABLE_PROGRESS_GATES"] !== "1") {
-                    const gate = computeNoProgressCompletionGate({
-                      summary: content,
-                      toolCallLog,
-                      taskState: this._taskState
-                    });
-                    if (gate.shouldInject && gate.content) {
-                      messages2.push({ role: "system", content: gate.content });
-                      this.emit({
-                        type: "status",
-                        content: "text completion held: discovery happened but no deliverable or explicit blocker is recorded",
-                        turn,
-                        timestamp: (/* @__PURE__ */ new Date()).toISOString()
-                      });
-                      continue;
-                    }
+                  const completionArgs = { summary: content };
+                  if (holdNoProgressTaskComplete(completionArgs, turn) || holdProvenanceTaskComplete(completionArgs, turn)) {
+                    continue;
                   }
                   completed = true;
                   summary = content;

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.211",
+  "version": "1.0.212",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.211",
+      "version": "1.0.212",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.211",
+  "version": "1.0.212",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",