pi-review-loop 0.3.1 → 0.4.0

package/README.md

@@ -30,13 +30,15 @@ Agents make mistakes. They miss edge cases, introduce typos, forget error handli
 
 **Smart Exit Detection** - Won't be fooled by "Fixed 3 issues. No further issues found." Detects when issues were fixed and keeps looping.
 
+**Fresh Context** - Optional mode that strips prior review iterations from context each pass. The agent is prompted to re-read any relevant plan, spec, or PRD documents, so it truly reviews with fresh eyes instead of through the lens of its previous passes.
+
 **Fully Configurable** - Every pattern is customizable. Change what triggers the loop, what exits it, and what prompt gets sent. Extend the defaults or replace them entirely.
 
 ## Typical Workflow
 
 The loop shines in two scenarios:
 
-**Before implementing** — You've got a plan doc and want to sanity-check it against the actual codebase. Run `/review-start` and let the agent compare the plan to what exists. It'll catch things like outdated assumptions, conflicting patterns, or unnecessary complexity. The funny thing is, it rarely finds everything on the first pass. Second pass catches different issues. Third pass, more still. That's the whole point of the loop.
+**Before implementing** — You've got a plan doc and want to sanity-check it against the actual codebase. Run `/review-plan` and let the agent compare the plan to what exists. It'll catch things like outdated assumptions, conflicting patterns, or unnecessary complexity. The funny thing is, it rarely finds everything on the first pass. Second pass catches different issues. Third pass, more still. That's the whole point of the loop.
 
 **After implementing** — You just finished building a feature and want to catch bugs before calling it done. Run `/review-start` and the agent reviews its own work with fresh eyes. Typos, missed edge cases, forgotten error handling — it finds stuff you'd miss staring at the same code. Again, multiple passes tend to surface different issues each time.
 
@@ -66,7 +68,7 @@ The package includes two prompt templates that are automatically installed to `~
 These prompts are designed to work with the review loop:
 - They instruct the agent to respond with "No issues found." when done (triggering exit)
 - They tell the agent to end with "Fixed [N] issue(s). Ready for another review." when issues are fixed (continuing the loop)
-- They include the "fresh eyes" phrase that triggers the loop when `autoTrigger` is enabled
+- `double-check.md` includes the "fresh eyes" phrase that triggers the loop when `autoTrigger` is enabled
 
 **Recommended workflow:** Use `/review-start` to activate review mode, which sends the review prompt automatically. Alternatively, enable auto-trigger (`/review-auto on`) and the `/double-check` template will activate the loop.
 
@@ -140,6 +142,47 @@ Or just type something else. Any non-trigger input exits review mode.
 
 Changes max iterations for current session.
 
+## Fresh Context
+
+By default, each review iteration sees the full conversation history including all prior iterations. This means by pass 3, the agent is reviewing code through the lens of its two prior reviews -- not truly fresh eyes. Fresh context mode fixes this.
+
+When enabled, prior review iterations are stripped from context before each LLM call. The agent only sees: the original pre-review conversation, a brief pass note instructing it to re-read any relevant plan/spec/PRD documents, and the current iteration's review prompt and tool usage.
+
+### Enable Fresh Context
+
+Per-session:
+```
+/review-fresh on
+```
+
+Or via the tool:
+```typescript
+review_loop({ start: true, freshContext: true })
+```
+
+Or permanently in settings:
+```json
+{
+  "reviewerLoop": {
+    "freshContext": true
+  }
+}
+```
+
+### How It Works
+
+```
+iteration 1: [pre-review context] [review prompt]                     ← full context
+iteration 2: [pre-review context] [pass note] [review prompt]         ← iter 1 stripped
+iteration 3: [pre-review context] [pass note] [review prompt]         ← iters 1-2 stripped
+```
+
+The pre-review context is everything from before review mode was activated. Within a review iteration, multi-turn tool usage (read, bash, edit) is preserved -- only completed prior iterations are stripped.
+
+The pass note tells the agent which pass it's on and instructs it to re-read any relevant plan, spec, PRD, or progress documents before reviewing. This way the agent re-grounds itself in the source of truth each pass using its own tool calls rather than programmatic injection.
+
+If auto-compaction fires during a review loop (large sessions), the fresh context handler gracefully degrades to full context for that session.
+
 ## Configuration
 
 Configure in `~/.pi/agent/settings.json`. Works out of the box, but everything is customizable:
@@ -149,6 +192,7 @@ Configure in `~/.pi/agent/settings.json`. Works out of the box, but everything i
   "reviewerLoop": {
     "maxIterations": 7,
     "autoTrigger": true,
+    "freshContext": true,
     "reviewPrompt": "template:double-check",
     "triggerPatterns": {
       "mode": "extend",
@@ -172,6 +216,7 @@ Configure in `~/.pi/agent/settings.json`. Works out of the box, but everything i
 |--------|-------------|
 | `maxIterations` | Max review prompts before auto-exit (default: 7) |
 | `autoTrigger` | Enable keyword-based auto-trigger (default: false) |
+| `freshContext` | Strip prior iterations from context each pass (default: false) |
 | `reviewPrompt` | The prompt to send each iteration |
 | `triggerPatterns` | What activates review mode (requires autoTrigger: true) |
 | `exitPatterns` | What indicates "review complete" |
@@ -250,6 +295,7 @@ The loop exits when:
 | `/review-max <n>` | Set max iterations (session only) |
 | `/review-auto [on\|off]` | Toggle auto-trigger from keywords (session only) |
 | `/review-auto <focus>` | Enable auto-trigger AND start review with custom focus |
+| `/review-fresh [on\|off]` | Toggle fresh context mode (session only) |
 | `/review-status` | Show current state |
 
 ## Tool API
@@ -278,15 +324,19 @@ review_loop({ maxIterations: 10 })
 // Enable/disable auto-trigger
 review_loop({ autoTrigger: true })
 review_loop({ autoTrigger: false })
+
+// Enable fresh context
+review_loop({ start: true, freshContext: true })
 ```
 
 **Returns:**
 ```json
 {
   "active": true,
-  "currentIteration": 2,
+  "currentIteration": 1,
   "maxIterations": 7,
   "autoTrigger": false,
+  "freshContext": true,
   "focus": "focus on error handling",
   "message": "Review mode active: iteration 2/7"
 }
@@ -316,6 +366,7 @@ otherwise → exit (max reached)
 - `session_start` - Reload settings
 - `input` - Detect triggers (if autoTrigger enabled), handle interrupts
 - `before_agent_start` - Check expanded prompts for triggers (if autoTrigger enabled)
+- `context` - Strip prior iterations and inject pass note (if freshContext enabled)
 - `agent_end` - Analyze response, decide to loop or exit
 
 ## Limitations

package/index.ts

@@ -7,13 +7,15 @@ export default function (pi: ExtensionAPI) {
   let reviewModeActive = false;
   let currentIteration = 0;
   let customPromptSuffix = "";
+  let freshContext = settings.freshContext;
+  let reviewBoundaryCount = -1;
+  let boundaryNeedsCapture = false;
 
   function updateStatus(ctx: ExtensionContext) {
     if (reviewModeActive) {
-      ctx.ui.setStatus(
-        "review-loop",
-        `Review mode (${currentIteration}/${settings.maxIterations})`
-      );
+      const parts = [`Review mode (${currentIteration + 1}/${settings.maxIterations})`];
+      if (freshContext) parts.push("fresh");
+      ctx.ui.setStatus("review-loop", parts.join(" | "));
     } else {
       ctx.ui.setStatus("review-loop", undefined);
     }
@@ -30,12 +32,8 @@ export default function (pi: ExtensionAPI) {
   function parseCustomText(args: string): string {
     const trimmed = args.trim();
     if (!trimmed) return "";
-    
-    // Try quoted strings first: "text" or 'text'
-    const match = trimmed.match(/^"(.+)"$/s) || trimmed.match(/^'(.+)'$/s) ||
-                  trimmed.match(/"(.+?)"/s) || trimmed.match(/'(.+?)'/s);
-    
-    // If no quotes found, use the entire text as focus
+
+    const match = trimmed.match(/^"(.+)"$/s) || trimmed.match(/^'(.+)'$/s);
     return match ? match[1].trim() : trimmed;
   }
 
@@ -43,6 +41,8 @@ export default function (pi: ExtensionAPI) {
     reviewModeActive = false;
     currentIteration = 0;
     customPromptSuffix = "";
+    reviewBoundaryCount = -1;
+    boundaryNeedsCapture = false;
     updateStatus(ctx);
     ctx.ui.notify(`Review mode ended: ${reason}`, "info");
   }
@@ -50,12 +50,15 @@ export default function (pi: ExtensionAPI) {
   function enterReviewMode(ctx: ExtensionContext) {
     reviewModeActive = true;
     currentIteration = 0;
+    reviewBoundaryCount = -1;
+    boundaryNeedsCapture = false;
     updateStatus(ctx);
     ctx.ui.notify("Review mode activated", "info");
   }
 
   pi.on("session_start", async () => {
     settings = loadSettings();
+    freshContext = settings.freshContext;
   });
 
   pi.on("input", async (event, ctx) => {
@@ -86,6 +89,54 @@ export default function (pi: ExtensionAPI) {
     }
   });
 
+  pi.on("context", async (event) => {
+    if (!reviewModeActive || !freshContext) return;
+
+    const messages = event.messages;
+    if (messages.length === 0) return;
+
+    if (boundaryNeedsCapture) {
+      for (let i = messages.length - 1; i >= 0; i--) {
+        if (messages[i].role === "user") {
+          reviewBoundaryCount = i;
+          break;
+        }
+      }
+      boundaryNeedsCapture = false;
+    }
+
+    if (currentIteration === 0) return;
+    if (reviewBoundaryCount < 0) return;
+
+    let lastUserIdx = -1;
+    for (let i = messages.length - 1; i >= 0; i--) {
+      if (messages[i].role === "user") {
+        lastUserIdx = i;
+        break;
+      }
+    }
+    if (lastUserIdx < 0) return;
+    if (reviewBoundaryCount >= lastUserIdx) return;
+
+    const preReview = messages.slice(0, reviewBoundaryCount);
+    const currentIterationMsgs = messages.slice(lastUserIdx);
+
+    const assembled: typeof messages = [...preReview];
+
+    assembled.push({
+      role: "user",
+      content: [{
+        type: "text",
+        text: `[Review pass ${currentIteration + 1}. ${currentIteration} prior pass(es) completed, fixes applied to code. Re-read any relevant plan, spec, PRD, or progress documents before reviewing. Review with fresh eyes.]`,
+      }],
+      timestamp: Date.now(),
+    } as any);
+
+    assembled.push(...currentIterationMsgs);
+
+    return { messages: assembled };
+  });
+
   pi.on("agent_end", async (event, ctx) => {
     if (!ctx.hasUI) return;
     if (!reviewModeActive) return;
@@ -119,11 +170,12 @@ export default function (pi: ExtensionAPI) {
     }
 
     currentIteration++;
-    if (currentIteration > settings.maxIterations) {
+    if (currentIteration >= settings.maxIterations) {
       exitReviewMode(ctx, `max iterations (${settings.maxIterations}) reached`);
       return;
     }
 
+    if (freshContext && reviewBoundaryCount < 0) boundaryNeedsCapture = true;
     updateStatus(ctx);
     pi.sendUserMessage(buildReviewPrompt(settings.reviewPromptConfig), {
       deliverAs: "followUp",
@@ -131,26 +183,28 @@ export default function (pi: ExtensionAPI) {
   });
 
   pi.registerCommand("review-start", {
-    description: "Activate review loop and send review prompt. Optional: add custom focus text.",
+    description: "Activate review loop with optional custom focus text.",
     handler: async (args, ctx) => {
       if (reviewModeActive) {
         ctx.ui.notify("Review mode is already active", "info");
       } else {
         customPromptSuffix = parseCustomText(args);
         enterReviewMode(ctx);
+        if (freshContext && reviewBoundaryCount < 0) boundaryNeedsCapture = true;
         pi.sendUserMessage(buildReviewPrompt(settings.reviewPromptConfig));
       }
     },
   });
 
   pi.registerCommand("review-plan", {
-    description: "Activate review loop for plans/specs/PRDs. Optional: add custom focus text.",
+    description: "Activate review loop for plans/specs/PRDs with optional custom focus text.",
     handler: async (args, ctx) => {
       if (reviewModeActive) {
         ctx.ui.notify("Review mode is already active", "info");
       } else {
         customPromptSuffix = parseCustomText(args);
         enterReviewMode(ctx);
+        if (freshContext && reviewBoundaryCount < 0) boundaryNeedsCapture = true;
         pi.sendUserMessage(buildReviewPrompt({ type: "template", value: "double-check-plan" }));
       }
     },
@@ -184,13 +238,12 @@ export default function (pi: ExtensionAPI) {
     description: "Show review mode status",
     handler: async (_args, ctx) => {
       if (reviewModeActive) {
-        ctx.ui.notify(
-          `Review mode active: iteration ${currentIteration}/${settings.maxIterations}`,
-          "info"
-        );
+        const parts = [`iteration ${currentIteration + 1}/${settings.maxIterations}`];
+        if (freshContext) parts.push("fresh context");
+        ctx.ui.notify(`Review mode active: ${parts.join(", ")}`, "info");
       } else {
         ctx.ui.notify(
-          `Review mode inactive (max: ${settings.maxIterations}, auto-trigger: ${settings.autoTrigger ? "on" : "off"})`,
+          `Review mode inactive (max: ${settings.maxIterations}, auto-trigger: ${settings.autoTrigger ? "on" : "off"}, fresh: ${freshContext ? "on" : "off"})`,
           "info"
         );
       }
@@ -202,8 +255,7 @@ export default function (pi: ExtensionAPI) {
     handler: async (args, ctx) => {
       const arg = args.trim();
       const argLower = arg.toLowerCase();
-      
-      // Check for standard on/off/toggle keywords first
+
       if (argLower === "on" || argLower === "true" || argLower === "1") {
         settings.autoTrigger = true;
         ctx.ui.notify(`Auto-trigger enabled`, "info");
@@ -222,21 +274,36 @@ export default function (pi: ExtensionAPI) {
         );
         return;
       }
-      
-      // Anything else is treated as custom focus text
-      const customText = parseCustomText(arg);
+
       settings.autoTrigger = true;
-      customPromptSuffix = customText;
+      customPromptSuffix = parseCustomText(arg);
       if (reviewModeActive) {
         ctx.ui.notify(`Auto-trigger enabled, focus updated for next iteration`, "info");
       } else {
         enterReviewMode(ctx);
+        if (freshContext && reviewBoundaryCount < 0) boundaryNeedsCapture = true;
         pi.sendUserMessage(buildReviewPrompt(settings.reviewPromptConfig));
         ctx.ui.notify(`Auto-trigger enabled, review started with custom focus`, "info");
       }
     },
   });
 
+  pi.registerCommand("review-fresh", {
+    description: "Toggle fresh context mode for review iterations",
+    handler: async (args, ctx) => {
+      const arg = args.trim().toLowerCase();
+
+      if (arg === "on" || arg === "true" || arg === "1") {
+        freshContext = true;
+      } else if (arg === "off" || arg === "false" || arg === "0") {
+        freshContext = false;
+      } else {
+        freshContext = !freshContext;
+      }
+      ctx.ui.notify(`Fresh context ${freshContext ? "enabled" : "disabled"}`, "info");
+    },
+  });
+
   pi.registerTool({
     name: "review_loop",
     description:
@@ -268,15 +335,18 @@ export default function (pi: ExtensionAPI) {
           description: "Custom focus/instructions to append to the review prompt (e.g., \"focus on error handling\")",
         })
       ),
+      freshContext: Type.Optional(
+        Type.Boolean({
+          description: "Enable/disable fresh context mode (strips prior review iterations from context)",
+        })
+      ),
     }),
 
     async execute(_toolCallId, params, _onUpdate, ctx) {
-      // Update maxIterations if provided
       if (typeof params.maxIterations === "number" && params.maxIterations >= 1) {
         settings.maxIterations = params.maxIterations;
       }
 
-      // Update autoTrigger if provided
       if (typeof params.autoTrigger === "boolean") {
         settings.autoTrigger = params.autoTrigger;
         ctx.ui.notify(
@@ -285,12 +355,18 @@ export default function (pi: ExtensionAPI) {
         );
       }
 
-      // Update custom focus if provided
       if (typeof params.focus === "string") {
         customPromptSuffix = params.focus.trim();
       }
 
-      // Mode: start > stop > status
+      if (typeof params.freshContext === "boolean") {
+        freshContext = params.freshContext;
+        ctx.ui.notify(
+          `Fresh context ${freshContext ? "enabled" : "disabled"}`,
+          "info"
+        );
+      }
+
       if (params.start) {
         if (reviewModeActive) {
           return {
@@ -302,6 +378,7 @@ export default function (pi: ExtensionAPI) {
                   currentIteration,
                   maxIterations: settings.maxIterations,
                   autoTrigger: settings.autoTrigger,
+                  freshContext,
                   focus: customPromptSuffix || undefined,
                   message: "Review mode is already active",
                 }),
@@ -311,6 +388,7 @@ export default function (pi: ExtensionAPI) {
         }
 
         enterReviewMode(ctx);
+        if (freshContext && reviewBoundaryCount < 0) boundaryNeedsCapture = true;
         pi.sendUserMessage(buildReviewPrompt(settings.reviewPromptConfig));
 
         return {
@@ -322,6 +400,7 @@ export default function (pi: ExtensionAPI) {
                 currentIteration,
                 maxIterations: settings.maxIterations,
                 autoTrigger: settings.autoTrigger,
+                freshContext,
                 focus: customPromptSuffix || undefined,
                 message: customPromptSuffix
                   ? `Review mode started with custom focus. Review prompt sent.`
@@ -343,6 +422,7 @@ export default function (pi: ExtensionAPI) {
                   currentIteration: 0,
                   maxIterations: settings.maxIterations,
                   autoTrigger: settings.autoTrigger,
+                  freshContext,
                   message: "Review mode is not active",
                 }),
               },
@@ -361,6 +441,7 @@ export default function (pi: ExtensionAPI) {
                 currentIteration: 0,
                 maxIterations: settings.maxIterations,
                 autoTrigger: settings.autoTrigger,
+                freshContext,
                 message: "Review mode stopped",
               }),
             },
@@ -368,7 +449,6 @@ export default function (pi: ExtensionAPI) {
         };
       }
 
-      // Default: status
       return {
         content: [
           {
@@ -378,9 +458,10 @@ export default function (pi: ExtensionAPI) {
               currentIteration,
               maxIterations: settings.maxIterations,
               autoTrigger: settings.autoTrigger,
+              freshContext,
               focus: customPromptSuffix || undefined,
               message: reviewModeActive
-                ? `Review mode active: iteration ${currentIteration}/${settings.maxIterations}`
+                ? `Review mode active: iteration ${currentIteration + 1}/${settings.maxIterations}`
                 : "Review mode inactive",
             }),
           },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-review-loop",
-  "version": "0.3.1",
+  "version": "0.4.0",
   "description": "Automated code review loop extension for pi coding agent",
   "scripts": {
     "postinstall": "node scripts/install.js",

package/prompts/double-check-plan.md

@@ -1,10 +1,12 @@
 ---
 description: Double check / critique implementation plan with fresh eyes
 ---
-Great, now I want you to carefully read over the plan/spec and compare against the codebase with "fresh eyes," looking super carefully for any obvious bugs, errors, problems, issues, confusion, etc. Also double check that the patterns and coding styles match existing codebase conventions and preferences. Consider whether there is any unnecessary complexity or duplication, or if the changes can be more minimal without changing scope of requirements and without causing regressions or increasing the risk of bugs. Also keep in mind that this is not an MVP - so don't skip anything. If any issues are found, proceed to fix them without being asked to do so.
+Great, now I want you to carefully read over the plan/spec and compare against the codebase with "fresh eyes," looking super carefully for any obvious bugs, errors, problems, issues, confusion, etc. Think through how best to implement this elegantly given our architecture and goals. Also keep in mind that this is not an MVP - so don't skip anything. If any issues are found, proceed to fix them without being asked to do so.
 
 **Response format:**
 - If you find ANY issues: fix them, then list what you fixed. Do NOT say "no issues found" - instead end with "Fixed [N] issue(s). Ready for another review."
-- If you find ZERO issues: respond with exactly "No issues found."
+- If you find ZERO issues: describe what you examined and verified, then conclude with "No issues found."
+
+Do not rush to a verdict. Read all relevant code first, trace through edge cases, and only then decide. I am pushing you to do a genuinely thorough review and not just lazily rubber-stamp it. Make sure you think deeply, then ultrathink some more.
 
 $@

package/prompts/double-check.md

@@ -3,8 +3,20 @@ description: Double check / review implementation with fresh eyes
 ---
 Great, now I want you to carefully read over all of the new code you just wrote and other existing code with "fresh eyes," looking super carefully for any obvious bugs, errors, problems, issues, confusion, etc. Also, if you notice any pre-existing issues/bugs those should be addressed.
 
+Question everything: Does each line of code need to exist? Unused parameters, dead code, and unnecessary complexity should be removed, not dressed up with underscore prefixes or comments.
+
+This codebase will outlive you. Every shortcut becomes someone else's burden. Every hack compounds into technical debt that slows the whole team down.
+
+You are not just writing code. You are shaping the future of this project. The patterns you establish will be copied. The corners you cut will be cut again.
+
+Fight entropy. Leave the codebase better than you found it.
+
+You MUST read all relevant code and think deeply (ultrathink!!!) first before you make any edits.
+
 **Response format:**
 - If you find ANY issues: fix them, then list what you fixed. Do NOT say "no issues found" - instead end with "Fixed [N] issue(s). Ready for another review."
-- If you find ZERO issues: respond with exactly "No issues found."
+- If you find ZERO issues: describe what you examined and verified, then conclude with "No issues found."
+
+Do not rush to a verdict. Read all relevant code first, trace through edge cases, and only then decide. I am pushing you to do a genuinely thorough review and not just lazily rubber-stamp it. Make sure you think deeply, then ultrathink some more.
 
 $@

package/settings.ts

@@ -6,11 +6,23 @@ export const SETTINGS_PATH = join(homedir(), ".pi", "agent", "settings.json");
 
 export const DEFAULT_MAX_ITERATIONS = 7;
 
-export const DEFAULT_REVIEW_PROMPT = `Great, now I want you to carefully read over all of the new code you just wrote and other existing code you just modified with "fresh eyes," looking super carefully for any obvious bugs, errors, problems, issues, confusion, etc. If any issues are found, proceed to fix them without being asked to do so.
+export const DEFAULT_REVIEW_PROMPT = `Great, now I want you to carefully read over all of the new code you just wrote and other existing code with "fresh eyes," looking super carefully for any obvious bugs, errors, problems, issues, confusion, etc. Also, if you notice any pre-existing issues/bugs those should be addressed.
+
+Question everything: Does each line of code need to exist? Unused parameters, dead code, and unnecessary complexity should be removed, not dressed up with underscore prefixes or comments.
+
+This codebase will outlive you. Every shortcut becomes someone else's burden. Every hack compounds into technical debt that slows the whole team down.
+
+You are not just writing code. You are shaping the future of this project. The patterns you establish will be copied. The corners you cut will be cut again.
+
+Fight entropy. Leave the codebase better than you found it.
+
+You MUST read all relevant code and think deeply (ultrathink!!!) first before you make any edits.
 
 **Response format:**
 - If you find ANY issues: fix them, then list what you fixed. Do NOT say "no issues found" - instead end with "Fixed [N] issue(s). Ready for another review."
-- If you find ZERO issues: respond with exactly "No issues found."`;
+- If you find ZERO issues: describe what you examined and verified, then conclude with "No issues found."
+
+Do not rush to a verdict. Read all relevant code first, trace through edge cases, and only then decide. I am pushing you to do a genuinely thorough review and not just lazily rubber-stamp it. Make sure you think deeply, then ultrathink some more.`;
 
 export const DEFAULT_TRIGGER_PATTERNS: RegExp[] = [
   /\bimplement\s+(the\s+)?plan\b/i,
@@ -61,6 +73,7 @@ export interface ReviewerLoopSettingsRaw {
   maxIterations?: number;
   reviewPrompt?: string;
   autoTrigger?: boolean;
+  freshContext?: boolean;
   triggerPatterns?: PatternConfig;
   exitPatterns?: PatternConfig;
   issuesFixedPatterns?: PatternConfig;
@@ -70,6 +83,7 @@ export interface ReviewerLoopSettings {
   maxIterations: number;
   reviewPromptConfig: ReviewPromptConfig;
   autoTrigger: boolean;
+  freshContext: boolean;
   triggerPatterns: RegExp[];
   exitPatterns: RegExp[];
   issuesFixedPatterns: RegExp[];
@@ -81,7 +95,8 @@ function parsePattern(input: unknown): RegExp | null {
   const match = input.match(/^\/(.+)\/([gimsuy]*)$/);
   if (match) {
     try {
-      return new RegExp(match[1], match[2]);
+      const flags = match[2].replace(/g/g, "");
+      return new RegExp(match[1], flags);
     } catch {
       return null;
     }
@@ -204,6 +219,7 @@ export function loadSettings(): ReviewerLoopSettings {
         : DEFAULT_MAX_ITERATIONS,
     reviewPromptConfig: parseReviewPromptConfig(raw.reviewPrompt),
     autoTrigger: raw.autoTrigger === true,
+    freshContext: raw.freshContext === true,
     triggerPatterns: loadPatterns(raw.triggerPatterns, DEFAULT_TRIGGER_PATTERNS),
     exitPatterns: loadPatterns(raw.exitPatterns, DEFAULT_EXIT_PATTERNS),
     issuesFixedPatterns: loadPatterns(