feat(guards): lesson guards — a recorded mistake warns before it repeats (G1)

AGENTS.md

@@ -64,6 +64,10 @@ Under Claude Code with the MeMesh plugin, hooks capture automatically:
 - **Stop** captures session knowledge and turns failures into lessons.
 - **PreCompact** saves important knowledge before history is compressed.
 - **UserPromptSubmit** detects "remember this" intent in the prompt.
+- **PreToolUse (Bash)** fires accepted lesson-guards: a fenced warning
+  citing the source lesson (`[mem:id]`) when a command matches a mistake
+  the graph has recorded. Heed it, and cite the lesson if it changes what
+  you do.
 
 So under Claude Code: skip step 1 of the loop (the topology is already
 injected), and do not `remember` commits or session summaries by hand. Manual

CHANGELOG.md

@@ -4,6 +4,32 @@ All notable changes to MeMesh are documented here.
 
 ## [Unreleased]
 
+### Added
+
+- **Lesson guards: a recorded mistake can now warn at the moment it is
+  about to repeat.** The dreamer gained a guard stage: for each
+  failure-shaped lesson (the Error / Root cause / Fix structure) it
+  proposes a `{tool, pattern, message}` trigger — a regex over the Bash
+  command or the Edit/Write content about to run — and every proposal is
+  verified mechanically before staging: the pattern must compile, must
+  not match a benign-input probe list (`git status`, `npm test`, …), and
+  must pass its own attached evidence (at least two inputs that trigger,
+  two that stay silent — executed, not trusted). A human accepts or
+  rejects in the same review queue as every other proposal; acceptance
+  writes `metadata.guard` onto the source lesson (nothing is created or
+  archived, and the spec is re-verified with no model in the loop). At
+  runtime a new `guard-check` hook (PreToolUse Bash — MeMesh's seventh
+  hook) and the existing pre-edit-recall hook (Edit/Write, deliberately
+  outside its recall throttle: a dangerous edit is dangerous every time)
+  evaluate accepted guards as plain regex tests — no LLM, no network —
+  and a hit injects the lesson's warning as a fenced reference block
+  carrying the lesson's `[mem:id]` citation handle. v1 guards only WARN;
+  the schema carries `action` so per-guard blocking can arrive once
+  measured fire accuracy justifies it, and every fire is counted on the
+  guard (`fires`, `last_fired_at`) so a guard that never fires or fires
+  constantly surfaces for review. Guard failure of any kind degrades to
+  silence — the guard system can never block or break the user's work.
+
 ### Changed
 
 - **Dashboard design system: VIVARIUM supersedes Precision Engineer.** The

README.de.md

@@ -342,7 +342,7 @@ Reproduktionsbefehle, Datensatz-SHA256, rohe Ergebnisse pro Frage und Analyse be
 
 ## Was läuft in Claude Code automatisch ab
 
-Sie müssen nicht manuell alles speichern. MeMesh verfügt über **6 Hooks**, die Wissen während der Arbeit erfassen und injizieren:
+Sie müssen nicht manuell alles speichern. MeMesh verfügt über **7 Hooks**, die Wissen während der Arbeit erfassen und injizieren:
 
 | Wenn | Was MeMesh tut |
 |------|------------------|
@@ -351,6 +351,7 @@ Sie müssen nicht manuell alles speichern. MeMesh verfügt über **6 Hooks**, di
 | **Nach jedem `git commit`** | Erfasst Ihre Änderungen mit Diff-Statistiken |
 | **Wenn Claude stoppt** | Erfasst bearbeitete Dateien und behobene Fehler; generiert automatisch strukturierte Lektionen aus Fehlern |
 | **Vor Context-Verdichtung** | Speichert Wissen, bevor es durch Context-Limits verloren geht |
+| **Vor riskanten Befehlen und Edits** | Löst die von Ihnen akzeptierten Lektions-Guards aus — eine Warnung genau in dem Moment, in dem sich ein erfasster Fehler wiederholen würde |
 
 > **Jederzeit abschalten:** `export MEMESH_AUTO_CAPTURE=false`
 

README.md

@@ -354,7 +354,7 @@ Re-runnable in ~10 seconds. Full instructions, dataset SHA256, raw per-question
 
 ## What Happens Automatically In Claude Code
 
-You don't need to manually remember everything. MeMesh has **6 hooks** that capture and inject knowledge while you work:
+You don't need to manually remember everything. MeMesh has **7 hooks** that capture and inject knowledge while you work:
 
 | When | What MeMesh does |
 |------|------------------|
@@ -364,6 +364,7 @@ You don't need to manually remember everything. MeMesh has **6 hooks** that capt
 | **After every `git commit`** | Records what you changed, with diff stats |
 | **When Claude stops** | Captures files edited, errors fixed, and auto-generates structured lessons from failures |
 | **Before context compaction** | Saves knowledge before it's lost to context limits |
+| **Before risky commands and edits** | Fires the lesson-guards you accepted — a warning at the exact moment a recorded mistake is about to repeat |
 
 > **Opt out anytime:** `export MEMESH_AUTO_CAPTURE=false`
 

README.zh-TW.md

@@ -350,7 +350,7 @@ MeMesh 的檢索引擎**只用 FTS5**（熱路徑上不使用 LLM、不使用嵌
 
 ## Claude Code 自動進行的事情
 
-你不需要手動記住所有事情。MeMesh 有 **6 個 hooks**，會在你工作時自動擷取與注入知識：
+你不需要手動記住所有事情。MeMesh 有 **7 個 hooks**，會在你工作時自動擷取與注入知識：
 
 | 何時 | MeMesh 做什麼 |
 |------|------------------|
@@ -360,6 +360,7 @@ MeMesh 的檢索引擎**只用 FTS5**（熱路徑上不使用 LLM、不使用嵌
 | **每次 `git commit` 之後** | 記錄你的變更，包含 diff 統計 |
 | **Claude 停止時** | 擷取已編輯的檔案、已修復的錯誤，並從失敗自動產生結構化教訓 |
 | **上下文壓縮前** | 在知識被上下文限制丟掉之前先保存 |
+| **危險指令與編輯前** | 觸發你接受過的教訓守衛——在記錄過的錯誤即將重演的那一刻發出警告 |
 
 > **隨時退出：** `export MEMESH_AUTO_CAPTURE=false`
 

dashboard/src/components/InsightsTab.tsx

@@ -471,7 +471,54 @@ export function InsightsTab() {
               );
             })()}
 
-            {detail && detail.proposed_digest && p.kind !== 'relation' && (
+            {/* Guard proposals (G1) carry a GuardSpec payload — the reviewer
+                is approving a regex that will warn on future tool inputs, so
+                the card shows the pattern, the message it will speak, and the
+                evidence examples the validator executed. */}
+            {detail && detail.proposed_digest && p.kind === 'guard' && (() => {
+              const g = detail.proposed_digest as unknown as {
+                guard?: { tool?: string; pattern?: string; message?: string; should_match?: string[]; should_not_match?: string[] };
+                source_lesson?: { name?: string; title?: string | null };
+              };
+              return (
+                <div style={{ marginTop: 12, padding: 12, background: 'var(--bg-1)', borderRadius: 'var(--radius-xs)', fontSize: 13 }}>
+                  <div style={{ marginBottom: 8, color: 'var(--text-3)', fontSize: 11 }}>
+                    {t('insights.generatedBy')}: <code>{detail.llm_model ?? t('common.unknown')}</code> · {t('insights.promptVersion')}: <code>{detail.prompt_version}</code>
+                  </div>
+                  <div style={{ marginBottom: 8 }}>
+                    <span class="tag" style={{ fontSize: 11 }}>{g.guard?.tool}</span>
+                    <span style={{ marginLeft: 8, color: 'var(--text-3)', fontSize: 11 }}>{t('guard.sourceLesson')}:</span>{' '}
+                    <span style={{ fontSize: 12 }}>{g.source_lesson?.title || g.source_lesson?.name}</span>
+                  </div>
+                  <div style={{ marginBottom: 8 }}>
+                    <div style={{ fontSize: 11, color: 'var(--text-3)', marginBottom: 2 }}>{t('guard.pattern')}</div>
+                    <code style={{ fontFamily: 'var(--mono)', fontSize: 12, wordBreak: 'break-all' }}>{g.guard?.pattern}</code>
+                  </div>
+                  <div style={{ marginBottom: 8 }}>
+                    <div style={{ fontSize: 11, color: 'var(--text-3)', marginBottom: 2 }}>{t('guard.message')}</div>
+                    <div style={{ lineHeight: 1.5 }}>{g.guard?.message}</div>
+                  </div>
+                  {(g.guard?.should_match?.length || g.guard?.should_not_match?.length) ? (
+                    <div style={{ fontSize: 12, color: 'var(--text-2)' }}>
+                      {g.guard?.should_match?.length ? (
+                        <div style={{ marginBottom: 4 }}>
+                          <span style={{ color: 'var(--text-3)', fontSize: 11 }}>{t('guard.shouldMatch')}:</span>{' '}
+                          {g.guard.should_match.map((ex, i) => <code key={i} style={{ fontFamily: 'var(--mono)', fontSize: 11, marginRight: 8 }}>{ex}</code>)}
+                        </div>
+                      ) : null}
+                      {g.guard?.should_not_match?.length ? (
+                        <div>
+                          <span style={{ color: 'var(--text-3)', fontSize: 11 }}>{t('guard.shouldNotMatch')}:</span>{' '}
+                          {g.guard.should_not_match.map((ex, i) => <code key={i} style={{ fontFamily: 'var(--mono)', fontSize: 11, marginRight: 8 }}>{ex}</code>)}
+                        </div>
+                      ) : null}
+                    </div>
+                  ) : null}
+                </div>
+              );
+            })()}
+
+            {detail && detail.proposed_digest && p.kind !== 'relation' && p.kind !== 'guard' && (
               <div style={{ marginTop: 12, padding: 12, background: 'var(--bg-1)', borderRadius: 'var(--radius-xs)', fontSize: 13 }}>
                 <div style={{ marginBottom: 8, color: 'var(--text-3)', fontSize: 11 }}>
                   {t('insights.generatedBy')}: <code>{detail.llm_model ?? t('common.unknown')}</code> · {t('insights.promptVersion')}: <code>{detail.prompt_version}</code>