thumbgate 1.1.0 → 1.3.0

package/plugins/cursor-marketplace/README.md

@@ -11,7 +11,7 @@ How it works:
 
 - 👎 **Thumbs down** a mistake — it gets blocked from happening again
 - 👍 **Thumbs up** good work — the agent remembers the pattern
-- **History-aware lesson distillation** — vague thumbs signals can be grounded in the last ~10 messages and failed tool call
+- **History-aware lesson distillation** — vague thumbs-down signals can be grounded in up to 8 prior recorded entries and the failed tool call
 - **Cross-session memory** — lessons persist between conversations
 - **Automatic enforcement** — repeated failures become prevention rules
 - **Zero config** — install and start giving feedback
@@ -108,7 +108,7 @@ Or copy the plugin MCP config into `.cursor/mcp.json`:
 
 Use the `/capture-feedback` command or the `capture_feedback` MCP tool to send structured feedback directly to the ThumbGate memory system. Feedback drives prevention rule generation — repeated failure patterns are auto-promoted into enforceable gates.
 
-When the user only gives a quick `thumbs_down`, `wrong`, or `correct`, the Cursor plugin should include the last ~10 messages and the failed tool call in `chatHistory` so ThumbGate can propose the lesson automatically. If the explanation comes later, reuse the earlier event with `relatedFeedbackId` instead of creating an isolated duplicate.
+When the user only gives a quick `thumbs_down`, `wrong`, or `correct`, the Cursor plugin should include up to 8 prior recorded entries and the failed tool call in `chatHistory` so ThumbGate can propose the lesson automatically. If the explanation comes later, reuse the earlier event with `relatedFeedbackId` so the linked 60-second follow-up session refines the same feedback record instead of creating an isolated duplicate.
 
 ## What makes this useful in Cursor
 

package/plugins/cursor-marketplace/commands/capture-feedback.md

@@ -15,7 +15,7 @@ Invoke this command to record a feedback signal with context and tags. If the us
 
 1. Specify signal: `thumbs_up` or `thumbs_down`.
 2. If the user already gave a clear explanation, pass it as `context`.
-3. If the explanation is vague or omitted, gather the last ~10 messages plus the failed tool call into `chatHistory`.
+3. If the explanation is vague or omitted, gather up to 8 prior recorded entries plus the failed tool call into `chatHistory`.
 4. Add tags for categorization.
 5. If this is a later clarification for an existing feedback event, include `relatedFeedbackId`.
 6. Call the `capture_feedback` MCP tool and show the proposed lesson or corrective rule back to the user.
@@ -30,4 +30,4 @@ Invoke this command to record a feedback signal with context and tags. If the us
 /capture-feedback thumbs_down
 ```
 
-In the second example, the command should rely on history-aware distillation instead of refusing the signal outright.
+In the second example, the command should rely on history-aware distillation instead of refusing the signal outright. If the user adds more detail later, reuse `relatedFeedbackId` so the linked 60-second follow-up session updates the same record.

package/plugins/cursor-marketplace/rules/feedback-capture.mdc

@@ -20,15 +20,15 @@ After any mistake, unexpected behavior, or completed task, prompt to capture str
    - `signal`: "thumbs_up" or "thumbs_down"
    - `context`: Brief description of what happened when the explanation is already explicit
    - `tags`: Relevant tags (e.g., "test-failure", "wrong-approach", "deployment-error")
-   - `chatHistory`: the last ~10 messages plus the failed tool call when the signal is vague and the lesson should be inferred from recent conversation
+   - `chatHistory`: up to 8 prior recorded entries plus the failed tool call when the thumbs-down signal is vague and the lesson should be inferred from recent conversation
    - `relatedFeedbackId`: use for a later clarification that should refine an existing feedback event
    - `rubric_scores`: Optional structured scores if a rubric applies
 
 2. Include enough context so the feedback is useful for future prevention rule generation.
-3. If the user only gives a bare thumbs signal or short correction, still capture it and let ThumbGate propose `whatWentWrong` / `whatToChange` from recent history.
+3. If the user only gives a bare thumbs signal or short correction, still capture it and let ThumbGate propose `whatWentWrong` / `whatToChange` from recent history. If the explanation arrives later, reuse `relatedFeedbackId` so the linked 60-second follow-up session refines the same record.
 
 ## Examples
 
 - Test failure after refactor: signal=thumbs_down, tags=["test-failure", "refactor"], context="Tests broke after extracting helper function — missed internal dependency"
 - Successful deployment: signal=thumbs_up, tags=["deployment", "success"], context="Clean deploy after running gate checks"
-- Vague correction after failed tool use: signal=thumbs_down, chatHistory=[last 10 messages + failed tool call], tags=["wrong-approach", "tool-failure"]
+- Vague correction after failed tool use: signal=thumbs_down, chatHistory=[up to 8 prior recorded entries + failed tool call], tags=["wrong-approach", "tool-failure"]

package/plugins/cursor-marketplace/skills/capture-feedback/SKILL.md

@@ -21,7 +21,7 @@ Use the `capture_feedback` MCP tool with:
 - **signal** — `"thumbs_up"` or `"thumbs_down"`
 - **context** — Description of what happened and why when the user already said it clearly
 - **tags** — Array of relevant tags for categorization (e.g., `["test-failure", "refactor"]`)
-- **chatHistory** — The last ~10 messages plus the failed tool call when the thumbs signal is vague and the lesson must be distilled from recent context
+- **chatHistory** — Up to 8 prior recorded entries plus the failed tool call when the thumbs-down signal is vague and the lesson must be distilled from recent context
 - **relatedFeedbackId** — Use when the user adds clarifying detail later and it should refine the existing feedback event
 - **rubric_scores** — Optional object with structured quality scores
 
@@ -39,8 +39,9 @@ If the user only says `thumbs_down`, `wrong`, `correct`, or `this failed`, do no
 
 - the signal
 - any minimal context the user already gave
-- `chatHistory` containing the recent conversation window
+- `chatHistory` containing up to 8 prior recorded entries from the current correction thread
 - the failed tool call or command when available
+- `relatedFeedbackId` if the user is clarifying an already-open 60-second follow-up session
 
 That lets ThumbGate propose `whatWentWrong`, `whatToChange`, and a candidate rule automatically.
 

package/plugins/opencode-profile/INSTALL.md

@@ -25,7 +25,7 @@ The portable profile adds this MCP server entry:
   "mcp": {
     "thumbgate": {
       "type": "local",
-      "command": ["npx", "--yes", "--package", "thumbgate@1.1.0", "thumbgate", "serve"],
+      "command": ["npx", "--yes", "--package", "thumbgate@1.3.0", "thumbgate", "serve"],
       "enabled": true
     }
   }

package/public/compare.html

@@ -0,0 +1,302 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="UTF-8">
+<meta name="viewport" content="width=device-width, initial-scale=1.0">
+<title>Best Pre-Action Gate Tools for AI Coding Agents (2026 Comparison)</title>
+<!-- Privacy-friendly analytics by Plausible -->
+<script defer data-domain="rlhf-feedback-loop-production.up.railway.app" src="https://plausible.io/js/script.js"></script>
+<meta name="description" content="Compare pre-action gate tools that prevent AI coding agents from making costly mistakes. ThumbGate vs manual review vs post-hoc fixes.">
+<meta name="keywords" content="AI agent safety, pre-action gates, AI coding agent comparison, ThumbGate vs manual review, AI agent guardrails, PreToolUse hooks, Claude Code safety, Codex safety, Gemini safety, Cursor rules alternative">
+<meta property="og:title" content="Best Pre-Action Gate Tools for AI Coding Agents (2026 Comparison)">
+<meta property="og:description" content="Compare pre-action gate tools that prevent AI coding agents from making costly mistakes. ThumbGate vs manual review vs post-hoc fixes.">
+<meta property="og:type" content="article">
+<meta property="og:url" content="https://thumbgate-production.up.railway.app/compare">
+<link rel="canonical" href="https://thumbgate-production.up.railway.app/compare">
+
+<script type="application/ld+json">
+{
+  "@context": "https://schema.org",
+  "@type": "TechArticle",
+  "headline": "Best Pre-Action Gate Tools for AI Coding Agents (2026 Comparison)",
+  "description": "Compare pre-action gate tools that prevent AI coding agents from making costly mistakes. ThumbGate vs manual review vs post-hoc fixes vs Cursor rules vs linters.",
+  "author": {
+    "@type": "Person",
+    "name": "Igor Ganapolsky",
+    "url": "https://github.com/IgorGanapolsky"
+  },
+  "publisher": {
+    "@type": "Organization",
+    "name": "ThumbGate",
+    "url": "https://thumbgate-production.up.railway.app"
+  },
+  "datePublished": "2026-04-08",
+  "dateModified": "2026-04-08",
+  "mainEntityOfPage": "https://thumbgate-production.up.railway.app/compare"
+}
+</script>
+
+<script type="application/ld+json">
+{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "What is a pre-action gate for AI coding agents?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "A pre-action gate is an enforcement layer that intercepts AI agent tool calls before they execute. Unlike prompt rules that agents can ignore, pre-action gates physically block dangerous actions such as force-pushing to main, deleting production files, or committing code with failing tests. ThumbGate implements pre-action gates via PreToolUse hooks that fire before every tool call."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Does ThumbGate work with Claude Code?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Yes. ThumbGate works with Claude Code, Cursor, Codex, Gemini CLI, Amp, OpenCode, and any MCP-compatible agent. Run npx thumbgate init --agent claude-code to auto-configure PreToolUse hooks in your .claude/settings.json."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Is ThumbGate free?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "ThumbGate has a free tier that includes local enforcement with 3 daily feedback captures, 5 lesson searches, unlimited recall, blocking, and history-aware lesson distillation. Pro ($19/mo or $149/yr) adds a personal local dashboard and DPO export. Team rollout ($12/seat/mo) adds a shared lesson database and org dashboard."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "How is ThumbGate different from .cursorrules or CLAUDE.md rules?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Cursor rules and CLAUDE.md are prompt-level instructions that the AI agent can read, forget, or override. ThumbGate enforces rules at the tool-call level via PreToolUse hooks. When a tool call matches a known failure pattern, it is physically blocked before execution. Additionally, ThumbGate auto-generates prevention rules from feedback, so you never have to write rules manually."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Can ThumbGate learn from mistakes automatically?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Yes. When your AI agent makes a mistake, you give a thumbs-down with context. After repeated failures with the same pattern, ThumbGate auto-generates a prevention rule and wires it as a pre-action gate. Gates adapt their sensitivity over time using Thompson Sampling, a Bayesian multi-armed bandit algorithm."
+      }
+    }
+  ]
+}
+</script>
+
+<style>
+  *, *::before, *::after { margin: 0; padding: 0; box-sizing: border-box; }
+  :root {
+    --bg: #0a0a0b;
+    --bg-card: #161618;
+    --border: #222225;
+    --text: #e8e8ec;
+    --muted: #8b8b94;
+    --cyan: #22d3ee;
+    --green: #34d399;
+    --red: #f87171;
+  }
+  body { font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif; background: var(--bg); color: var(--text); line-height: 1.7; }
+  .container { max-width: 780px; margin: 0 auto; padding: 2rem 1.5rem 4rem; }
+  nav { padding: 1rem 2rem; border-bottom: 1px solid var(--border); display: flex; gap: 1.5rem; align-items: center; }
+  nav a { color: var(--muted); text-decoration: none; font-size: 0.9rem; }
+  nav a:hover { color: var(--cyan); }
+  nav .brand { color: var(--text); font-weight: 700; font-size: 1.1rem; }
+  h1 { font-size: 2.2rem; line-height: 1.2; margin: 2rem 0 1rem; }
+  h2 { font-size: 1.5rem; margin: 2.5rem 0 1rem; color: var(--cyan); }
+  h3 { font-size: 1.15rem; margin: 1.5rem 0 0.5rem; }
+  p, li { color: var(--text); margin-bottom: 0.75rem; }
+  ul { padding-left: 1.5rem; }
+  li { margin-bottom: 0.5rem; }
+  code { background: #1a1a1e; padding: 0.15em 0.4em; border-radius: 4px; font-size: 0.9em; color: var(--cyan); }
+  pre { background: #111113; border: 1px solid var(--border); border-radius: 8px; padding: 1rem; overflow-x: auto; margin: 1rem 0; }
+  pre code { background: none; padding: 0; color: var(--text); }
+  .highlight { color: var(--cyan); }
+  .card { background: var(--bg-card); border: 1px solid var(--border); border-radius: 12px; padding: 1.5rem; margin: 1.5rem 0; }
+  .comparison-table { width: 100%; border-collapse: collapse; margin: 1.5rem 0; font-size: 0.9rem; }
+  .comparison-table th, .comparison-table td { padding: 0.75rem; border: 1px solid var(--border); text-align: left; }
+  .comparison-table th { background: #111113; color: var(--cyan); }
+  .comparison-table td:first-child { font-weight: 600; }
+  .yes { color: var(--green); }
+  .no { color: var(--red); }
+  .partial { color: #fbbf24; }
+  .cta { display: inline-block; background: var(--cyan); color: #000; padding: 0.75rem 1.5rem; border-radius: 8px; text-decoration: none; font-weight: 600; margin: 1rem 0; }
+  .cta:hover { opacity: 0.9; }
+  .breadcrumb { color: var(--muted); font-size: 0.85rem; margin-bottom: 0.5rem; }
+  .breadcrumb a { color: var(--muted); }
+  .step-grid { display: grid; grid-template-columns: repeat(auto-fit, minmax(200px, 1fr)); gap: 1rem; margin: 1.5rem 0; }
+  .step-card { background: var(--bg-card); border: 1px solid var(--border); border-radius: 12px; padding: 1.25rem; }
+  .step-card .step-number { color: var(--cyan); font-weight: 700; font-size: 1.5rem; margin-bottom: 0.5rem; }
+  .step-card h3 { margin-top: 0; }
+  footer { border-top: 1px solid var(--border); padding: 2rem; text-align: center; color: var(--muted); font-size: 0.85rem; }
+  footer a { color: var(--muted); text-decoration: underline; }
+  @media (max-width: 600px) { h1 { font-size: 1.6rem; } .container { padding: 1rem; } .comparison-table { font-size: 0.8rem; } .comparison-table th, .comparison-table td { padding: 0.5rem; } }
+</style>
+</head>
+<body>
+<nav>
+  <a href="/" class="brand">ThumbGate</a>
+  <a href="/guide">Guide</a>
+  <a href="/compare">Compare</a>
+  <a href="/dashboard">Dashboard</a>
+  <a href="https://github.com/IgorGanapolsky/ThumbGate">GitHub</a>
+</nav>
+
+<div class="container">
+<p class="breadcrumb"><a href="/">Home</a> / Compare</p>
+
+<h1>Best Pre-Action Gate Tools for AI Coding Agents (2026 Comparison)</h1>
+<p style="color:var(--muted);">Your AI coding agent force-pushed to main again. Which tool actually prevents that? We compared the five most common approaches to AI agent safety.</p>
+
+<h2>Comparison Table</h2>
+<div style="overflow-x:auto;">
+<table class="comparison-table">
+  <tr>
+    <th>Feature</th>
+    <th>ThumbGate</th>
+    <th>Manual Code Review</th>
+    <th>Post-hoc Git Revert</th>
+    <th>Cursor Rules / .cursorrules</th>
+    <th>ESLint / Linters</th>
+  </tr>
+  <tr>
+    <td>Blocks bad actions before execution</td>
+    <td class="yes">Yes -- PreToolUse hooks</td>
+    <td class="partial">Partially -- after PR only</td>
+    <td class="no">No -- damage already done</td>
+    <td class="partial">Partially -- suggestions only</td>
+    <td class="partial">Partially -- static checks only</td>
+  </tr>
+  <tr>
+    <td>Learns from past mistakes</td>
+    <td class="yes">Yes -- auto-generates rules from feedback</td>
+    <td class="no">No -- relies on reviewer memory</td>
+    <td class="no">No</td>
+    <td class="no">No -- rules are hand-written</td>
+    <td class="no">No -- rules are hand-written</td>
+  </tr>
+  <tr>
+    <td>Works with Claude / Codex / Gemini / Forge</td>
+    <td class="yes">Yes -- all MCP-compatible agents</td>
+    <td class="yes">Yes -- agent-agnostic</td>
+    <td class="yes">Yes -- agent-agnostic</td>
+    <td class="partial">Cursor only</td>
+    <td class="partial">Partial -- code-level only</td>
+  </tr>
+  <tr>
+    <td>Zero config setup</td>
+    <td class="yes">Yes -- <code>npx thumbgate init</code></td>
+    <td class="no">No -- requires process + people</td>
+    <td class="no">No -- manual intervention each time</td>
+    <td class="partial">Partial -- requires writing rules</td>
+    <td class="partial">Partial -- requires config + plugins</td>
+  </tr>
+  <tr>
+    <td>Captures feedback for improvement</td>
+    <td class="yes">Yes -- thumbs up/down with context</td>
+    <td class="partial">Partial -- PR comments only</td>
+    <td class="no">No</td>
+    <td class="no">No</td>
+    <td class="no">No</td>
+  </tr>
+  <tr>
+    <td>Generates prevention rules automatically</td>
+    <td class="yes">Yes -- from repeated failure patterns</td>
+    <td class="no">No</td>
+    <td class="no">No</td>
+    <td class="no">No</td>
+    <td class="no">No</td>
+  </tr>
+  <tr>
+    <td>Real-time pre-action gates</td>
+    <td class="yes">Yes -- fires before every tool call</td>
+    <td class="no">No -- asynchronous review</td>
+    <td class="no">No -- reactive only</td>
+    <td class="no">No -- prompt-time only</td>
+    <td class="no">No -- build-time only</td>
+  </tr>
+  <tr>
+    <td>Cost</td>
+    <td>Free tier + Pro $19/mo</td>
+    <td>Engineer time per review</td>
+    <td>Time lost + recovery cost</td>
+    <td>Free</td>
+    <td>Free (open source)</td>
+  </tr>
+</table>
+</div>
+
+<h2>Why ThumbGate Wins</h2>
+<ul>
+  <li><strong>Enforcement, not suggestions.</strong> Prompt rules in CLAUDE.md or .cursorrules are instructions the agent can ignore. ThumbGate intercepts tool calls at the PreToolUse hook level and physically blocks dangerous actions before they execute.</li>
+  <li><strong>Learns and adapts automatically.</strong> Every thumbs-down becomes a data point. After repeated failures with the same pattern, ThumbGate auto-generates a prevention rule. Gates adapt their sensitivity over time using Thompson Sampling -- aggressive gates that over-block get tuned down, effective gates get reinforced.</li>
+  <li><strong>Works across all major AI coding agents.</strong> One install command covers Claude Code, Cursor, Codex, Gemini CLI, Amp, and OpenCode. No per-agent configuration needed.</li>
+  <li><strong>Prevention is cheaper than recovery.</strong> Manual code review catches mistakes after the PR. Git revert catches them after the push. ThumbGate catches them before the tool call executes -- saving the time, context switches, and cleanup cost of undoing damage.</li>
+</ul>
+
+<h2>How It Works</h2>
+<div class="step-grid">
+  <div class="step-card">
+    <div class="step-number">1</div>
+    <h3>Install</h3>
+    <p>Run <code>npx thumbgate init</code>. ThumbGate auto-detects your AI coding agent and configures PreToolUse hooks. No manual setup needed.</p>
+  </div>
+  <div class="step-card">
+    <div class="step-number">2</div>
+    <h3>Gate</h3>
+    <p>Every tool call your agent makes passes through a pre-action gate. Known-bad patterns are blocked before execution. Good actions pass through instantly.</p>
+  </div>
+  <div class="step-card">
+    <div class="step-number">3</div>
+    <h3>Learn</h3>
+    <p>When your agent makes a mistake, give it a thumbs-down. In the current Claude auto-capture path, vague negative signals can reuse up to 8 prior recorded entries plus the failed tool call, then stay linked to a 60-second follow-up session before repeated failures harden into prevention rules.</p>
+  </div>
+</div>
+
+<h2>Get Started</h2>
+<div class="card">
+  <p>Install ThumbGate in one command:</p>
+  <pre><code>npx thumbgate init</code></pre>
+  <p>Then start giving feedback. ThumbGate handles the rest.</p>
+  <p>
+    <a href="https://www.npmjs.com/package/thumbgate" class="cta">View on npm</a>
+    <a href="https://github.com/IgorGanapolsky/ThumbGate" style="color:var(--cyan); margin-left:1.5rem; text-decoration:underline;">View on GitHub</a>
+  </p>
+</div>
+
+<h2>Frequently Asked Questions</h2>
+
+<div class="card">
+  <h3>What is a pre-action gate?</h3>
+  <p>A pre-action gate is an enforcement layer that intercepts AI agent tool calls before they execute. Unlike prompt rules that agents can ignore, pre-action gates physically block dangerous actions such as force-pushing to main, deleting production files, or committing code with failing tests.</p>
+</div>
+
+<div class="card">
+  <h3>Does ThumbGate work with Claude Code?</h3>
+  <p>Yes. ThumbGate works with Claude Code, Cursor, Codex, Gemini CLI, Amp, OpenCode, and any MCP-compatible agent. Run <code>npx thumbgate init --agent claude-code</code> to auto-configure PreToolUse hooks.</p>
+</div>
+
+<div class="card">
+  <h3>Is ThumbGate free?</h3>
+  <p>ThumbGate has a free tier that includes local enforcement with 3 daily feedback captures, 5 lesson searches, unlimited recall, and pre-action gate blocking. Pro ($19/mo or $149/yr) adds a personal local dashboard and DPO export. Team rollout ($12/seat/mo) adds a shared lesson database and org dashboard.</p>
+</div>
+
+<div class="card">
+  <h3>How is ThumbGate different from .cursorrules or CLAUDE.md rules?</h3>
+  <p>Cursor rules and CLAUDE.md are prompt-level instructions that the AI agent can read, forget, or override. ThumbGate enforces rules at the tool-call level via PreToolUse hooks. When a tool call matches a known failure pattern, it is physically blocked before execution. Additionally, ThumbGate auto-generates prevention rules from feedback -- no manual rule writing.</p>
+</div>
+
+<div class="card">
+  <h3>Can ThumbGate learn from mistakes automatically?</h3>
+  <p>Yes. When your AI agent makes a mistake, give a thumbs-down with context. In the current Claude auto-capture path, vague negative signals can reuse up to 8 prior recorded entries plus the failed tool call, then stay linked to a 60-second follow-up session. After repeated failures with the same pattern, ThumbGate auto-generates a prevention rule and wires it as a pre-action gate. Gates adapt over time using Thompson Sampling, a Bayesian multi-armed bandit algorithm.</p>
+</div>
+
+</div>
+
+<footer>
+  <p>ThumbGate -- Pre-action gates for AI coding agents</p>
+  <p><a href="https://github.com/IgorGanapolsky/ThumbGate">GitHub</a> | <a href="https://www.npmjs.com/package/thumbgate">npm</a> | <a href="/guide">Guide</a> | <a href="/dashboard">Dashboard</a></p>
+</footer>
+</body>
+</html>

package/public/guide.html

@@ -58,7 +58,7 @@
       "@type": "HowToStep",
       "position": 2,
       "name": "Give feedback when your agent makes a mistake",
-      "text": "Say thumbs down with context about what went wrong. ThumbGate captures it as structured feedback."
+      "text": "Say thumbs down with context about what went wrong. In the current Claude auto-capture path, vague negative signals can reuse up to 8 prior recorded entries and stay linked to a 60-second feedback session."
     },
     {
       "@type": "HowToStep",
@@ -234,7 +234,7 @@ npx thumbgate init --agent gemini</code></pre>
 <h2>How It Works</h2>
 
 <h3>1. You give feedback</h3>
-<p>When your agent makes a mistake, tell it. ThumbGate captures the feedback as structured data with context, tags, and domain. If the signal is vague, it can reuse the recent conversation window and failed tool call to propose a better lesson instead of discarding the feedback.</p>
+<p>When your agent makes a mistake, tell it. ThumbGate captures the feedback as structured data with context, tags, and domain. In the current Claude auto-capture path, if the thumbs-down is vague, it can reuse up to 8 prior recorded entries and the failed tool call to propose a better lesson instead of discarding the feedback.</p>
 <pre><code># Your agent force-pushed. You say:
 "thumbs down — force-pushed to main, lost commits"
 
@@ -265,7 +265,7 @@ npx thumbgate init --agent gemini</code></pre>
 </ul>
 
 <h2>History-Aware Feedback Sessions</h2>
-<p>ThumbGate supports linked feedback sessions for the messy reality of AI debugging. You can start one correction thread, append more context, and finalize it once the lesson is clear.</p>
+<p>ThumbGate supports linked feedback sessions for the messy reality of AI debugging. In the current Claude flow, accepted feedback opens a 60-second follow-up session. You can append more context, reset the timer, and finalize once the lesson is clear.</p>
 <ul>
   <li><code>open_feedback_session</code> starts a linked correction thread.</li>
   <li><code>append_feedback_context</code> adds later notes, failed tool output, or user corrections to the same thread.</li>
@@ -310,7 +310,7 @@ npx thumbgate init --agent gemini</code></pre>
 <pre><code>npx thumbgate init</code></pre>
 <p>One command. Works with Claude Code, Cursor, Codex, Gemini, Amp, and OpenCode. Claude Code can also call Codex for review, adversarial review, and second-pass handoffs through the repo-local bridge plugin.</p>
 <a href="https://thumbgate-production.up.railway.app/checkout/pro?utm_source=guide&utm_medium=cta_button&utm_campaign=pro_pack" class="cta">Get Pro — $19/mo or $149/yr</a>
-<p style="color:var(--muted); font-size:0.85rem;">Free keeps local enforcement with 5 daily feedback captures, 10 lesson searches, unlimited recall, blocking, and history-aware lesson distillation. Pro adds a personal local dashboard and DPO export. Team rollout adds the hosted shared lesson DB, org dashboard, and generated review views. <a href="https://buy.stripe.com/aFa4gz1M84r419v7mb3sI05" style="color:inherit;text-decoration:underline;">Founder $49 one-time link</a> stays preserved.</p>
+<p style="color:var(--muted); font-size:0.85rem;">Free keeps local enforcement with 3 daily feedback captures, 5 lesson searches, unlimited recall, blocking, and history-aware lesson distillation. Pro adds a personal local dashboard and DPO export. Team rollout adds the hosted shared lesson DB, org dashboard, and generated review views. <a href="https://buy.stripe.com/aFa4gz1M84r419v7mb3sI05" style="color:inherit;text-decoration:underline;">Founder $49 one-time link</a> stays preserved.</p>
 
 </div>
 </body>