@jhizzard/termdeck 0.10.0 → 0.10.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@jhizzard/termdeck",
-  "version": "0.10.0",
+  "version": "0.10.2",
   "description": "Browser-based terminal multiplexer with metadata overlays, panel flashback memory recall, and AI-aware session management",
   "bin": {
     "termdeck": "./packages/cli/src/index.js"

package/packages/client/public/app.js

@@ -237,6 +237,9 @@
             case 'meta':
               updatePanelMeta(id, msg.session.meta);
               break;
+            case 'proactive_memory':
+              showProactiveToast(id, msg.hit);
+              break;
             case 'exit':
               updatePanelMeta(id, {
                 status: 'exited',
@@ -1245,6 +1248,9 @@
             case 'meta':
               updatePanelMeta(id, msg.session.meta);
               break;
+            case 'proactive_memory':
+              showProactiveToast(id, msg.hit);
+              break;
             case 'exit':
               updatePanelMeta(id, { status: 'exited', statusDetail: `Exited (${msg.exitCode})` });
               const p = document.getElementById(`panel-${id}`);
@@ -1253,6 +1259,17 @@
             case 'status_broadcast':
               updateGlobalStats(msg.sessions);
               break;
+            case 'config_changed':
+              // Sprint 40 T1: parity with the main panel WS handler. The
+              // server broadcasts config_changed to ALL ws clients, including
+              // reconnected sessions; previously the reconnect path silently
+              // dropped these. Idempotent — safe to re-receive.
+              if (msg.config) {
+                state.config = { ...state.config, ...msg.config };
+                if (typeof renderSettingsPanel === 'function') renderSettingsPanel();
+                if (typeof updateRagIndicator === 'function') updateRagIndicator();
+              }
+              break;
           }
         } catch (err) { console.error('[client] reconnect ws message failed:', err); }
       };
@@ -2509,12 +2526,29 @@
       {
         targets: ['#btn-status', '#btn-config'],
         title: 'Status and config',
-        body: `<strong>status</strong> opens a global-metrics modal (session counts by state, RAG mode, memory bridge). <strong>config</strong> shows your loaded project list and theme defaults. Both are in the polish queue for Sprint 3 — buttons are visible but unwired right now.`,
+        body: `<strong>status</strong> opens a global-metrics modal (session counts by state, RAG mode, memory bridge). <strong>config</strong> shows your loaded project list and theme defaults — plus a live RAG-mode toggle (Sprint 36) that flips Flashback on/off without a server restart.`,
+      },
+      {
+        targets: ['#btn-sprint', '#btn-graph'],
+        title: 'Sprint runner and knowledge graph',
+        body: `<strong>sprint</strong> opens the in-dashboard 4+1 sprint runner (Sprint 37): name the sprint, define T1–T4 lane goals, click kick off — TermDeck spawns four panels and injects boot prompts via the two-stage submit pattern automatically. Optional <strong>--isolation=worktree</strong> creates a git worktree per lane so concurrent edits can't stomp. <strong>graph</strong> opens the D3.js force-directed knowledge graph (Sprint 38) of your memory_items + memory_relationships in a new tab — click any node to open its memory in a drawer, filter by relationship type, search, zoom/pan.`,
       },
       {
         targets: ['#btn-how', '#btn-help'],
         title: 'How this works and help',
-        body: `Click <strong>how this works</strong> any time to replay this tour. <strong>help</strong> opens the full TermDeck documentation in a new tab.`,
+        body: `Click <strong>how this works</strong> any time to replay this tour. <strong>help</strong> opens the full TermDeck documentation in a new tab. The <strong>📖 Guide</strong> tab on the right edge of the screen — also opens with the <kbd>g</kbd> keyboard shortcut — is the always-on Orchestrator Guide (Sprint 37): nine sections covering the 4+1 sprint pattern, inject mandate, CLAUDE.md hierarchy, memory-first discipline, sprint discipline, restart-prompt rituals, scaffolding files, channel inject patterns. Search built in.`,
+      },
+      {
+        target: '#guideRail',
+        title: 'Right-rail Orchestrator Guide',
+        body: `The <strong>📖 Guide</strong> rail is your orchestration cheat-sheet — collapsed by default, one click (or <kbd>g</kbd>) to expand. It auto-scrolls to the relevant section based on what you're focused on: clicking a terminal panel jumps the Guide to the 4+1 pattern; opening the project drawer jumps to CLAUDE.md hierarchy. Useful when you forget exactly how the two-stage submit pattern works at 2 AM in the middle of a sprint inject.`,
+        fallback: '#btn-how',
+      },
+      {
+        target: '#btnPreviewProject',
+        title: 'Orchestration preview',
+        body: `The <strong>preview</strong> button next to the project + button (Sprint 37) shows you exactly what <code>termdeck init --project &lt;name&gt;</code> would create for the selected project — file tree, contents per file, expand-on-click. Read-only by default; optional generate button writes the scaffolding (CLAUDE.md, CONTRADICTIONS.md, project_facts.md, .claude/settings.json, docs/orchestration/, RESTART-PROMPT.md template). Lets you see-before-commit instead of running the CLI blind.`,
+        fallback: '#btn-how',
       },
       {
         target: '.panel-header',
@@ -2553,10 +2587,15 @@
         title: 'Prompt bar',
         body: `Type any command here to launch it as a new terminal — <kbd>claude code ~/myproject</kbd>, <kbd>python3 manage.py runserver</kbd>, <kbd>npm run dev</kbd>. Pick a project from the dropdown to auto-cd into its path and apply its default theme. <kbd>Ctrl+Shift+N</kbd> focuses this bar from anywhere.`,
       },
+      {
+        target: null,
+        title: 'Knowledge graph + memory inference',
+        body: `Sprint 38 brought your <strong>memory_relationships</strong> table to life. The <strong>graph</strong> button (top toolbar) renders your memories as a force-directed network — supersedes / relates_to / contradicts / elaborates / caused_by / blocks / inspired_by / cross_project_link edges, color-coded, filterable. The Mnestra MCP server now exposes four new tools: <code>memory_link</code>, <code>memory_unlink</code>, <code>memory_related</code>, and <code>memory_recall_graph</code> — Claude Code can connect related memories explicitly, traverse N-hop neighborhoods, and recall via graph-aware re-ranking (vector_score × edge_weight × recency). Edges populate automatically from Joshua's private rag-system classifier; a nightly cron in Sprint 39+ will surface cross-project connections.`,
+      },
       {
         target: null,
         title: 'You are ready.',
-        body: `That's every major surface. Click <strong>how this works</strong> in the top toolbar to replay this walkthrough. <strong>help</strong> opens the full docs. Questions, bugs, feedback: <a href="https://github.com/jhizzard/termdeck/issues" target="_blank" style="color:var(--tg-accent)">github.com/jhizzard/termdeck/issues</a>. Now launch something.`,
+        body: `That's every major surface. Click <strong>how this works</strong> in the top toolbar to replay this walkthrough. <strong>help</strong> opens the full docs. Press <kbd>g</kbd> any time to crack open the Orchestrator Guide. Questions, bugs, feedback: <a href="https://github.com/jhizzard/termdeck/issues" target="_blank" style="color:var(--tg-accent)">github.com/jhizzard/termdeck/issues</a>. Now launch something.`,
       },
     ];
 

package/packages/server/src/flashback-diag.js

@@ -0,0 +1,51 @@
+// Flashback diagnostic ring buffer (Sprint 39 T1).
+//
+// Six decision points along the Flashback pipeline write structured events
+// here so production-flow regressions surface as a readable timeline instead
+// of a silent gate failure. The ring is in-memory and lost on restart by
+// design — persistence is a Sprint-40+ concern. Public surface:
+//
+//   log({ sessionId, event, ...fields })  — append one event
+//   snapshot({ sessionId?, eventType?, limit? }) — read back filtered tail
+//   _resetForTest()                                — test-only ring clear
+//
+// Event shape (all events): { ts, sessionId, event, ...event-specific fields }.
+//
+// Event types and their producers:
+//   pattern_match           — session.js _detectErrors (PATTERNS.error /
+//                             errorLineStart / shellError matched)
+//   error_detected          — session.js _detectErrors at onErrorDetected
+//                             entry, before rate-limit check
+//   rate_limit_blocked      — session.js _detectErrors when 30s limiter rejects
+//   bridge_query            — mnestra-bridge queryMnestra at call return
+//   bridge_result           — mnestra-bridge queryMnestra at call return
+//   proactive_memory_emit   — index.js onErrorDetected WS send block
+//
+// The route GET /api/flashback/diag (registered in index.js) returns
+// snapshot() output as JSON for ad-hoc inspection by Joshua and consumption
+// by T4's production-flow e2e test.
+
+const RING_SIZE = 200;
+
+let ring = [];
+
+function log(event) {
+  ring.push({ ts: new Date().toISOString(), ...event });
+  if (ring.length > RING_SIZE) {
+    ring = ring.slice(-RING_SIZE);
+  }
+}
+
+function snapshot({ sessionId, eventType, limit = RING_SIZE } = {}) {
+  let out = ring;
+  if (sessionId) out = out.filter((e) => e.sessionId === sessionId);
+  if (eventType) out = out.filter((e) => e.event === eventType);
+  const cap = Math.max(1, Math.min(RING_SIZE, Number(limit) || RING_SIZE));
+  return out.slice(-cap);
+}
+
+function _resetForTest() {
+  ring = [];
+}
+
+module.exports = { log, snapshot, _resetForTest, RING_SIZE };

package/packages/server/src/index.js

@@ -56,6 +56,7 @@ const { SessionManager } = require('./session');
 const { initDatabase, logCommand, getSessionHistory, getProjectSessions } = require('./database');
 const { RAGIntegration } = require('./rag');
 const { createBridge } = require('./mnestra-bridge');
+const flashbackDiag = require('./flashback-diag');
 const { writeSessionLog } = require('./session-logger');
 const { TranscriptWriter } = require('./transcripts');
 const { createHealthHandler, runPreflight } = require('./preflight');
@@ -853,30 +854,69 @@ function createServer(config) {
             question,
             project: sess.meta.project,
             searchAll: false,
+            cwd: sess.meta.cwd,
+            sessionId: sess.id,
             sessionContext: {
               type: sess.meta.type,
               project: sess.meta.project,
+              cwd: sess.meta.cwd,
               lastCommands: sess.meta.lastCommands.slice(-5),
               status: 'errored'
             }
           }).then((result) => {
-            const count = (result.memories || []).length;
+            const memories = (result && result.memories) || [];
+            const count = memories.length;
             console.log(`[flashback] query returned ${count} matches for session ${sess.id}`);
-            const hit = (result.memories || [])[0];
+            const hit = memories[0];
+            const wsReadyState = sess.ws ? sess.ws.readyState : null;
             if (!hit) {
               console.log(`[flashback] no matches — skipping proactive_memory send for session ${sess.id}`);
+              flashbackDiag.log({
+                sessionId: sess.id,
+                event: 'proactive_memory_emit',
+                ws_ready_state: wsReadyState,
+                frame_size_bytes: 0,
+                result_count_in_frame: 0,
+                outcome: 'dropped_empty',
+              });
               return;
             }
             if (sess.ws && sess.ws.readyState === 1) {
+              const frame = JSON.stringify({ type: 'proactive_memory', hit });
               try {
-                sess.ws.send(JSON.stringify({ type: 'proactive_memory', hit }));
+                sess.ws.send(frame);
                 console.log(`[flashback] proactive_memory sent to session ${sess.id} (source_type=${hit.source_type}, project=${hit.project})`);
+                flashbackDiag.log({
+                  sessionId: sess.id,
+                  event: 'proactive_memory_emit',
+                  ws_ready_state: 1,
+                  frame_size_bytes: Buffer.byteLength(frame, 'utf8'),
+                  result_count_in_frame: 1,
+                  outcome: 'emitted',
+                });
               } catch (err) {
                 console.error('[flashback] proactive_memory send failed:', err);
                 console.error('[ws] proactive_memory send failed:', err);
+                flashbackDiag.log({
+                  sessionId: sess.id,
+                  event: 'proactive_memory_emit',
+                  ws_ready_state: 1,
+                  frame_size_bytes: Buffer.byteLength(frame, 'utf8'),
+                  result_count_in_frame: 1,
+                  outcome: 'error',
+                  error_message: err && err.message ? err.message : String(err),
+                });
               }
             } else {
               console.log(`[flashback] ws not open for session ${sess.id} (readyState=${sess.ws ? sess.ws.readyState : 'null'}) — dropped hit`);
+              flashbackDiag.log({
+                sessionId: sess.id,
+                event: 'proactive_memory_emit',
+                ws_ready_state: wsReadyState,
+                frame_size_bytes: 0,
+                result_count_in_frame: count,
+                outcome: 'dropped_no_ws',
+              });
             }
           }).catch((err) => {
             console.error(`[flashback] query failed for session ${sess.id}: ${err.message}`);
@@ -1347,6 +1387,23 @@ function createServer(config) {
     });
   });
 
+  // GET /api/flashback/diag - Sprint 39 T1 diagnostic ring buffer.
+  // Returns the last N Flashback decision-point events so Joshua can trigger
+  // a real-shell error and read the timeline of which gate dropped the toast.
+  // Optional filters: ?sessionId=<uuid>, ?eventType=pattern_match, ?limit=N
+  // (capped at 200, the ring size).
+  app.get('/api/flashback/diag', (req, res) => {
+    const { sessionId, eventType } = req.query || {};
+    const rawLimit = req.query && req.query.limit;
+    const limit = rawLimit != null ? parseInt(rawLimit, 10) : undefined;
+    const events = flashbackDiag.snapshot({
+      sessionId: typeof sessionId === 'string' && sessionId.length ? sessionId : undefined,
+      eventType: typeof eventType === 'string' && eventType.length ? eventType : undefined,
+      limit: Number.isFinite(limit) && limit > 0 ? Math.min(limit, flashbackDiag.RING_SIZE) : undefined,
+    });
+    res.json({ count: events.length, events });
+  });
+
   // ==================== Transcript endpoints (Sprint 6 T3) ====================
 
   // GET /api/transcripts/search - FTS across all sessions
@@ -1568,6 +1625,7 @@ function createServer(config) {
     const sessionContext = session ? {
       type: session.meta.type,
       project: session.meta.project,
+      cwd: session.meta.cwd,
       lastCommands: session.meta.lastCommands.slice(-5),
       status: session.meta.status
     } : null;
@@ -1577,6 +1635,7 @@ function createServer(config) {
         question,
         project,
         searchAll,
+        cwd: session ? session.meta.cwd : undefined,
         sessionContext
       });
 

package/packages/server/src/mnestra-bridge/index.js

@@ -10,6 +10,7 @@
 
 const { spawn } = require('child_process');
 const { resolveProjectName } = require('../rag');
+const flashbackDiag = require('../flashback-diag');
 
 function createBridge(config) {
   const mode = config.rag?.mnestraMode || 'direct';
@@ -225,7 +226,7 @@ function createBridge(config) {
     }
   }
 
-  async function queryMnestra({ question, project, searchAll, sessionContext, cwd }) {
+  async function queryMnestra({ question, project, searchAll, sessionContext, cwd, sessionId }) {
     // Flashback callers pass the session's project (from config.yaml). If that
     // slot is empty — e.g. a session created without an explicit project — fall
     // back to resolving the session's cwd against config.projects so queries
@@ -246,15 +247,68 @@ function createBridge(config) {
     // out-of-repo session-end hook), the mismatch surfaces here at query time.
     console.log(`[mnestra-bridge] query project=${effectiveProject ?? 'ALL'} source=${searchAll ? 'searchAll' : projectSource} mode=${mode}`);
 
-    switch (mode) {
-      case 'webhook':
-        return queryWebhook({ question, project: effectiveProject, searchAll });
-      case 'mcp':
-        return queryMcp({ question, project: effectiveProject, searchAll });
-      case 'direct':
-      default:
-        return queryDirect({ question, project: effectiveProject, searchAll });
+    const projectTagInFilter = searchAll ? null : (effectiveProject || null);
+    const t0 = Date.now();
+    let result;
+    let callError;
+    try {
+      switch (mode) {
+        case 'webhook':
+          result = await queryWebhook({ question, project: effectiveProject, searchAll });
+          break;
+        case 'mcp':
+          result = await queryMcp({ question, project: effectiveProject, searchAll });
+          break;
+        case 'direct':
+        default:
+          result = await queryDirect({ question, project: effectiveProject, searchAll });
+          break;
+      }
+    } catch (err) {
+      callError = err;
     }
+    const durationMs = Date.now() - t0;
+
+    // Sprint 39 T1 — bridge_query / bridge_result diag events. Emitted at
+    // queryMnestra's outer boundary so all three backends (direct, webhook,
+    // mcp) flow through one observability point. T3 reads project_tag_in_filter
+    // (the tag the bridge SENT to the RPC) and top_3_project_tags (the tags
+    // it GOT BACK) to confirm or refute the project-mismatch hypothesis.
+    flashbackDiag.log({
+      sessionId,
+      event: 'bridge_query',
+      project_tag_in_filter: projectTagInFilter,
+      query_text: typeof question === 'string' ? question.slice(0, 200) : '',
+      mode,
+      rpc_args: {
+        project: projectTagInFilter,
+        searchAll: !!searchAll,
+        project_source: searchAll ? 'searchAll' : projectSource,
+      },
+      duration_ms: durationMs,
+    });
+
+    const memories = (result && Array.isArray(result.memories)) ? result.memories : [];
+    const tagCounts = {};
+    for (const m of memories) {
+      const tag = m && m.project != null ? String(m.project) : '(null)';
+      tagCounts[tag] = (tagCounts[tag] || 0) + 1;
+    }
+    const top3 = Object.entries(tagCounts)
+      .sort((a, b) => b[1] - a[1])
+      .slice(0, 3)
+      .map(([tag, count]) => ({ tag, count }));
+
+    flashbackDiag.log({
+      sessionId,
+      event: 'bridge_result',
+      result_count: memories.length,
+      error_message: callError ? (callError.message || String(callError)) : null,
+      top_3_project_tags: top3,
+    });
+
+    if (callError) throw callError;
+    return result;
   }
 
   return { mode, queryMnestra };

package/packages/server/src/session.js

@@ -14,6 +14,7 @@ const { v4: uuidv4 } = require('uuid');
 const os = require('os');
 const path = require('path');
 const { resolveTheme } = require('./theme-resolver');
+const flashbackDiag = require('./flashback-diag');
 
 // Strip ANSI escape codes for pattern matching
 function stripAnsi(str) {
@@ -43,6 +44,13 @@ const PATTERNS = {
     django: /Starting development server/,
     httpServer: /Serving HTTP on/,
     request: /(?:^|\s|")(GET|POST|PUT|DELETE|PATCH)\s+\S+.*?\s(\d{3})/m,
+    // Sprint 40 T2: HTTP 5xx response in a web-server log line is a real
+    // error condition for the application. Used as a python-server-typed
+    // fallback in _detectErrors when the prose-shape analyzers miss because
+    // the line carries no `Error:` keyword — just `"GET /foo HTTP/1.1" 503`.
+    // 5xx only (not 4xx, which are typically client-caused). The leading
+    // `(?:^|\s|")` mirrors `request` so colon-quoted log shapes still match.
+    serverError: /(?:^|\s|")(?:GET|POST|PUT|DELETE|PATCH)\s+\S+.*?\sHTTP\/\d(?:\.\d)?"?\s+5\d{2}\b/m,
     // Port detection — matches any of:
     //   • "port NNNN" phrase (capture group 1)
     //   • URL with http/https scheme, optionally prefixed with "on " or "at "
@@ -65,11 +73,20 @@ const PATTERNS = {
   // tools (cat, ls, cd, rm, etc.) report filesystem misses in plain English
   // without ever emitting the ENOENT errno code. Flagged as a gap by Rumen's
   // first production kickstart insight on 2026-04-15.
-  error: /(?:^|\n)\s*(?:Error:\s+\S|error:\s+\S|Traceback \(most recent call last\):|npm ERR!|error\[E\d+\]:|Uncaught Exception|Fatal:)/m,
+  // Sprint 40 T2: added uppercase `ERROR:` (mirrors `Error:` / `error:` for
+  // case-symmetry — closes the stripAnsi-ERROR test fixture from Sprint 33)
+  // and Node errno-style colon-prefix shapes (`ENOENT:`, `EACCES:`,
+  // `ECONNREFUSED:`) so `ENOENT: no such file or directory` shapes from
+  // child-process error reporting fire without depending on the line ALSO
+  // containing the `No such file or directory` prose phrase.
+  error: /(?:^|\n)\s*(?:Error:\s+\S|error:\s+\S|ERROR:\s+\S|Traceback \(most recent call last\):|npm ERR!|error\[E\d+\]:|Uncaught Exception|Fatal:|ENOENT:\s+\S|EACCES:\s+\S|ECONNREFUSED:\s+\S)/m,
   // Stricter line-anchored variant for Claude Code, whose tool output (grep
   // results, test logs, file contents) routinely mentions "Error" mid-line
   // without representing an actual failure of the agent itself.
-  errorLineStart: /^\s*(error|Error|ERROR|exception|Exception|Traceback|fatal|FATAL|segmentation fault|panic|EACCES|ECONNREFUSED|ENOENT|command not found|undefined reference|cannot find module|failed with exit code|No such file or directory|Permission denied)\b/m,
+  // Sprint 40 T2: added mixed-case `Fatal` (mirrors `fatal` / `FATAL`) and
+  // the `npm ERR!` shape (special-cased outside the alternation because
+  // `!` is not a word character so `\b` after `npm ERR!` doesn't match).
+  errorLineStart: /^\s*(?:(?:error|Error|ERROR|exception|Exception|Traceback|fatal|Fatal|FATAL|segmentation fault|panic|EACCES|ECONNREFUSED|ENOENT|command not found|undefined reference|cannot find module|failed with exit code|No such file or directory|Permission denied)\b|npm ERR!)/m,
   // Sprint 33: PATTERNS.error misses the most common Unix shell errors —
   // `cat: /foo: No such file or directory`, `bash: foo: command not found`,
   // `rm: cannot remove ...: Permission denied`. These have a colon-prefix
@@ -77,7 +94,27 @@ const PATTERNS = {
   // mentioning the same words. Each branch requires either the colon-prefix
   // structure or a stand-alone anchored keyword. Validated against an
   // adversarial prose suite (see tests/analyzer-error-fixtures.test.js).
-  shellError: /(?:^|\n)(?:[^\n]*:\s+(?:.*?:\s+)?(?:No such file or directory|Permission denied|Is a directory|Not a directory|command not found)\b|[^\n]*?\(\d+\)\s+Could not resolve host\b|\s*ModuleNotFoundError:\s+\S|\s*Segmentation fault\b|\s*fatal:\s+\S)/m
+  //
+  // Sprint 39 T2: separated `command not found` from the other phrases. The
+  // unified branch was matching rcfile-noise lines emitted by version
+  // managers during shell startup — most notably:
+  //   `pyenv: pyenv-virtualenv-init: command not found in path`
+  // …which has the colon-prefix-with-`command not found` shape but with a
+  // descriptive suffix (` in path`) rather than ending the line. The pyenv
+  // case confirms the strong rcfile-noise hypothesis for pyenv users: their
+  // shell startup burns the 30s onErrorDetected rate limit before the user
+  // can type their first command. The dedicated `command not found` branch
+  // below requires the keyword to be either:
+  //   • followed by `:` (the zsh `command not found: <cmd>` form), or
+  //   • at end-of-line (the bash `<sh>: <cmd>: command not found` form).
+  // Suffixes like ` in path`, ` in $PATH`, ` (compinit)` are silenced as
+  // rcfile noise.
+  // Trade-off: custom command_not_found_handler output that adds a comma-
+  // separated "did you mean X" suggestion is silenced — those are cosmetic
+  // suggestions, not the error itself, which the user already saw fire.
+  // See tests/rcfile-noise.test.js and tests/analyzer-error-fixtures.test.js
+  // for the locked corpus.
+  shellError: /(?:^|\n)(?:[^\n]*:\s+(?:.*?:\s+)?(?:No such file or directory|Permission denied|Is a directory|Not a directory)\b|[^\n]*:\s+(?:.*?:\s+)?command not found(?::|\s*(?:[\r\n]|$))|[^\n]*?\(\d+\)\s+Could not resolve host\b|\s*ModuleNotFoundError:\s+\S|\s*Segmentation fault\b|\s*fatal:\s+\S)/m
 };
 
 class Session {
@@ -350,14 +387,44 @@ class Session {
     // Claude Code's tool output frequently contains "error"/"Error" mid-line
     // (grep matches, test results, log dumps). Use a line-anchored pattern
     // for that session type so we don't flag content as failure.
-    const pattern = this.meta.type === 'claude-code'
+    const primaryPattern = this.meta.type === 'claude-code'
       ? PATTERNS.errorLineStart
       : PATTERNS.error;
+    const primaryName = this.meta.type === 'claude-code' ? 'errorLineStart' : 'error';
     // Sprint 33 fix: the structured patterns above miss `cat: /foo: No such
     // file or directory` and friends — the most common Unix shell error
     // shapes Josh hits day-to-day. Fall through to PATTERNS.shellError so
     // the analyzer flips status='errored' and Flashback can fire.
-    if (!pattern.test(clean) && !PATTERNS.shellError.test(clean)) return;
+    const primaryMatch = clean.match(primaryPattern);
+    const shellMatch = !primaryMatch ? clean.match(PATTERNS.shellError) : null;
+    // Sprint 40 T2: HTTP 5xx fallback for python-server sessions. The prose
+    // analyzers miss `"GET /foo HTTP/1.1" 503 -` because it carries no
+    // `Error:` keyword — but the response IS the error signal for an
+    // HTTP-server session. Gated on session type to avoid flagging 5xx
+    // status codes that legitimately appear in unrelated content (e.g. a
+    // shell that just printed a copy of an HTTP log).
+    const serverMatch = (!primaryMatch && !shellMatch && this.meta.type === 'python-server')
+      ? clean.match(PATTERNS.pythonServer.serverError)
+      : null;
+    if (!primaryMatch && !shellMatch && !serverMatch) return;
+
+    // Sprint 39 T1 — pattern_match diag event. Emitted on every PATTERNS hit,
+    // including ones that get rate-limited downstream. T2 reads these to
+    // measure the rcfile-noise false-positive rate against real shell output.
+    const matchedSrc = primaryMatch || shellMatch || serverMatch;
+    const matchedLine = (matchedSrc && typeof matchedSrc[0] === 'string')
+      ? matchedSrc[0].replace(/^\n+/, '').slice(0, 200)
+      : '';
+    const matchedPattern = primaryMatch
+      ? primaryName
+      : (shellMatch ? 'shellError' : 'serverError');
+    flashbackDiag.log({
+      sessionId: this.id,
+      event: 'pattern_match',
+      pattern: matchedPattern,
+      matched_line: matchedLine,
+      output_chunk_size: clean.length,
+    });
 
     const oldStatus = this.meta.status;
     this.meta.status = 'errored';
@@ -371,7 +438,30 @@ class Session {
 
     // Server-side rate limit: at most one error_detected event every 30s per session
     const now = Date.now();
+    const remainingMs = this._lastErrorFireAt
+      ? Math.max(0, 30000 - (now - this._lastErrorFireAt))
+      : 0;
+
+    // Sprint 39 T1 — error_detected diag event, before the rate-limit gate.
+    // The (error_detected count − rate_limit_blocked count) is the number of
+    // errors that actually got dispatched to onErrorDetected. T2/T3 use this
+    // to spot rcfile noise burning the rate-limit window before real errors.
+    flashbackDiag.log({
+      sessionId: this.id,
+      event: 'error_detected',
+      error_text: matchedLine,
+      rate_limit_remaining_ms: remainingMs,
+      last_emit_at: this._lastErrorFireAt
+        ? new Date(this._lastErrorFireAt).toISOString()
+        : null,
+    });
+
     if (now - this._lastErrorFireAt < 30000) {
+      flashbackDiag.log({
+        sessionId: this.id,
+        event: 'rate_limit_blocked',
+        rate_limit_remaining_ms: remainingMs,
+      });
       console.log(`[flashback] error detected in session ${this.id} but rate-limited (${Math.round((30000 - (now - this._lastErrorFireAt)) / 1000)}s left)`);
       return;
     }

package/packages/server/src/setup/mnestra-migrations/011_project_tag_backfill.sql

@@ -0,0 +1,237 @@
+-- Sprint 39 T3 — chopin-nashville project-tag backfill.
+--
+-- Why this exists:
+--   memory_items rows tagged project='chopin-nashville' are ~96% polluted
+--   with content from other projects (termdeck, mnestra, rumen, podium, pvb,
+--   dor). Root cause is the harness session-end hook
+--   (~/.claude/hooks/memory-session-end.js, OUT OF THIS REPO): its
+--   PROJECT_MAP iteration tests /ChopinNashville/i first and there are no
+--   entries for termdeck/mnestra/rumen/podium/dor — so any session whose
+--   cwd lives under ~/Documents/Graciella/ChopinNashville/... falls into
+--   chopin-nashville, including the entire TermDeck checkout (which lives at
+--   ChopinNashville/SideHustles/TermDeck/termdeck) and Podium (which lives at
+--   ChopinNashville/2026/ChopinInBohemia/podium).
+--
+--   This migration heals the historical rows. The forward-fix to the harness
+--   hook is Joshua's responsibility (out-of-repo file) and is NOT covered
+--   here — without it, new mis-tagged rows will continue to be written until
+--   he extends PROJECT_MAP with the missing project entries.
+--
+-- What this migration does NOT do:
+--   - Does NOT touch mnestra_session_memory / mnestra_project_memory / etc.
+--     (legacy rag-events tables; different write path; separate cleanup).
+--   - Does NOT consolidate duplicate project tags like 'gorgias' vs
+--     'gorgias-ticket-monitor', 'pvb' vs 'PVB', or 'mnestra' vs 'engram'.
+--     Those are visible in `SELECT project, count(*) FROM memory_items GROUP
+--     BY project` but they're a separate cleanup pass.
+--   - Does NOT touch the ~898 "other/uncertain" chopin-nashville rows that
+--     don't carry an unambiguous project keyword. A future sprint can run an
+--     LLM-classification pass; for this migration, conservative wins.
+--
+-- Heuristic — content keyword bucketing:
+--   The migration runs UPDATEs sequentially. Earlier buckets claim ambiguous
+--   multi-project rows first; later buckets only see rows that no earlier
+--   bucket has already re-tagged. Order is by bucket size (largest first):
+--
+--     1. termdeck / mnestra      — keywords: termdeck, mnestra, "4+1 sprint"
+--     2. rumen                   — keyword:  rumen
+--     3. podium                  — keyword:  podium
+--     4. pvb                     — keywords: PVB, petvetbid, pet vet bid
+--     5. dor / openclaw          — TIGHTENED:
+--                                    word-boundary uppercase DOR  (rules out
+--                                    "dormant", "vendored", "indoor", etc.),
+--                                    plus path/identifier markers and
+--                                    openclaw substring.
+--
+-- Spot-check baseline (T3 audit, 2026-04-27):
+--   termdeck/mnestra: 130 rows, all 6 sampled were true positives (TermDeck
+--                     server code, Mnestra wizard, sprint orchestration).
+--   rumen:            92 rows, all 6 sampled were true positives.
+--   podium:           58 rows, all 6 sampled were true positives.
+--   pvb:               7 rows, 1 of those overlaps with mnestra ("Mnestra
+--                     repo … petvetbid project") and gets claimed by bucket 1.
+--   dor (tightened):   3 rows after tightening from 6 — the original
+--                     `%dor%` ILIKE pattern caught false positives like
+--                     "dormant", "vendored". Final 3 rows are all true
+--                     DOR/OpenClaw mentions.
+--   chopin-nashville total: 1,169 rows. Legitimate-signal baseline (rows
+--                     matching Acceptd / NICPC / Bohemia / laureate /
+--                     applicant / competition / repertoire keywords): 71.
+--
+-- Idempotence:
+--   Every UPDATE is gated by `WHERE project = 'chopin-nashville'`. After the
+--   first run, those rows have a different project tag, so re-running this
+--   migration is a no-op (zero rows updated per bucket). RAISE NOTICE on a
+--   re-run will print zeros, which is the expected idempotent signal.
+--
+-- Application:
+--   THIS MIGRATION IS NOT EXECUTED BY THE LANE THAT WROTE IT. Orchestrator
+--   reviews the RAISE NOTICE counts after applying. Apply via the bundled
+--   migration runner at packages/server/src/setup/migration-runner.js (which
+--   uses node-postgres client.query — psql metacommands like \gset are NOT
+--   available, so the count probes use GET DIAGNOSTICS ROW_COUNT inside DO
+--   blocks). Manual fallback: `psql "$DATABASE_URL" -f 011_project_tag_backfill.sql`.
+
+BEGIN;
+
+-- ============================================================
+-- AUDIT BEFORE
+-- ============================================================
+DO $$
+DECLARE
+  before_chopin    int;
+  before_termdeck  int;
+  before_rumen     int;
+  before_podium    int;
+  before_pvb       int;
+  before_dor       int;
+BEGIN
+  SELECT count(*) INTO before_chopin   FROM memory_items WHERE project = 'chopin-nashville';
+  SELECT count(*) INTO before_termdeck FROM memory_items WHERE project = 'termdeck';
+  SELECT count(*) INTO before_rumen    FROM memory_items WHERE project = 'rumen';
+  SELECT count(*) INTO before_podium   FROM memory_items WHERE project = 'podium';
+  SELECT count(*) INTO before_pvb      FROM memory_items WHERE project = 'pvb';
+  SELECT count(*) INTO before_dor      FROM memory_items WHERE project = 'dor';
+  RAISE NOTICE '[011-backfill] BEFORE  chopin-nashville=% termdeck=% rumen=% podium=% pvb=% dor=%',
+    before_chopin, before_termdeck, before_rumen, before_podium, before_pvb, before_dor;
+END $$;
+
+-- ============================================================
+-- BUCKET 1 — TermDeck / Mnestra (claims multi-project mentions first)
+-- ============================================================
+DO $$
+DECLARE
+  rows_updated int;
+BEGIN
+  UPDATE memory_items SET project = 'termdeck'
+   WHERE project = 'chopin-nashville'
+     AND (
+       content ILIKE '%termdeck%'
+       OR content ILIKE '%mnestra%'
+       OR content ILIKE '%4+1 sprint%'
+     );
+  GET DIAGNOSTICS rows_updated = ROW_COUNT;
+  RAISE NOTICE '[011-backfill] bucket 1 (termdeck/mnestra): % rows re-tagged', rows_updated;
+END $$;
+
+-- ============================================================
+-- BUCKET 2 — Rumen
+-- ============================================================
+DO $$
+DECLARE
+  rows_updated int;
+BEGIN
+  UPDATE memory_items SET project = 'rumen'
+   WHERE project = 'chopin-nashville'
+     AND content ILIKE '%rumen%';
+  GET DIAGNOSTICS rows_updated = ROW_COUNT;
+  RAISE NOTICE '[011-backfill] bucket 2 (rumen): % rows re-tagged', rows_updated;
+END $$;
+
+-- ============================================================
+-- BUCKET 3 — Podium
+-- ============================================================
+DO $$
+DECLARE
+  rows_updated int;
+BEGIN
+  UPDATE memory_items SET project = 'podium'
+   WHERE project = 'chopin-nashville'
+     AND content ILIKE '%podium%';
+  GET DIAGNOSTICS rows_updated = ROW_COUNT;
+  RAISE NOTICE '[011-backfill] bucket 3 (podium): % rows re-tagged', rows_updated;
+END $$;
+
+-- ============================================================
+-- BUCKET 4 — PVB (case-insensitive PVB / petvetbid markers)
+-- ============================================================
+DO $$
+DECLARE
+  rows_updated int;
+BEGIN
+  UPDATE memory_items SET project = 'pvb'
+   WHERE project = 'chopin-nashville'
+     AND (
+       content ILIKE '%PVB%'
+       OR content ILIKE '%petvetbid%'
+       OR content ILIKE '%pet vet bid%'
+     );
+  GET DIAGNOSTICS rows_updated = ROW_COUNT;
+  RAISE NOTICE '[011-backfill] bucket 4 (pvb): % rows re-tagged', rows_updated;
+END $$;
+
+-- ============================================================
+-- BUCKET 5 — DOR / OpenClaw (TIGHTENED — word boundary + identifiers)
+--
+-- Original briefing heuristic was `content ILIKE '%dor%'`, which produced a
+-- ~33% false-positive rate (matched "dormant", "vendored", "indoor", etc.).
+-- T3 audit tightened to:
+--   • POSIX word boundary `\mDOR\M` — case-sensitive uppercase only, so
+--     "dormant" / "DormHall" / "vendor" / "indoor" no longer match.
+--   • path/identifier markers: /DOR/, ~/Documents/DOR, dor.config,
+--     "Rust LLM gateway" (DOR's tagline).
+--   • openclaw substring (OpenClaw is the slack-channel automation product
+--     that lives next to DOR in Joshua's stack).
+-- ============================================================
+DO $$
+DECLARE
+  rows_updated int;
+BEGIN
+  UPDATE memory_items SET project = 'dor'
+   WHERE project = 'chopin-nashville'
+     AND (
+       content ~ '\mDOR\M'
+       OR content ILIKE '%/DOR/%'
+       OR content ILIKE '%~/Documents/DOR%'
+       OR content ILIKE '%dor.config%'
+       OR content ILIKE '%Rust LLM gateway%'
+       OR content ILIKE '%openclaw%'
+     );
+  GET DIAGNOSTICS rows_updated = ROW_COUNT;
+  RAISE NOTICE '[011-backfill] bucket 5 (dor): % rows re-tagged', rows_updated;
+END $$;
+
+-- ============================================================
+-- AUDIT AFTER
+-- ============================================================
+DO $$
+DECLARE
+  after_chopin    int;
+  after_termdeck  int;
+  after_rumen     int;
+  after_podium    int;
+  after_pvb       int;
+  after_dor       int;
+BEGIN
+  SELECT count(*) INTO after_chopin   FROM memory_items WHERE project = 'chopin-nashville';
+  SELECT count(*) INTO after_termdeck FROM memory_items WHERE project = 'termdeck';
+  SELECT count(*) INTO after_rumen    FROM memory_items WHERE project = 'rumen';
+  SELECT count(*) INTO after_podium   FROM memory_items WHERE project = 'podium';
+  SELECT count(*) INTO after_pvb      FROM memory_items WHERE project = 'pvb';
+  SELECT count(*) INTO after_dor      FROM memory_items WHERE project = 'dor';
+  RAISE NOTICE '[011-backfill] AFTER   chopin-nashville=% termdeck=% rumen=% podium=% pvb=% dor=%',
+    after_chopin, after_termdeck, after_rumen, after_podium, after_pvb, after_dor;
+  RAISE NOTICE '[011-backfill] If apply succeeds and chopin-nashville count is around the legitimate baseline (~71 rows match competition/laureate/applicant/Acceptd/NICPC/Bohemia/repertoire keywords as of T3 audit), the migration succeeded. The ~898 rows that remain under chopin-nashville without a clear keyword signal are deliberate — a future LLM-classification pass can address them if needed.';
+END $$;
+
+COMMIT;
+
+-- ============================================================
+-- POST-APPLY: optional verification queries (NOT part of the migration).
+-- Run separately to confirm Flashback against project='termdeck' now hits
+-- the re-tagged rows.
+-- ============================================================
+--
+-- 1. Tag distribution after migration:
+--    SELECT project, count(*) FROM memory_items GROUP BY project ORDER BY count(*) DESC LIMIT 20;
+--
+-- 2. Confirm no chopin-nashville rows match obvious termdeck/rumen keywords:
+--    SELECT count(*) FROM memory_items
+--     WHERE project='chopin-nashville'
+--       AND (content ILIKE '%termdeck%' OR content ILIKE '%rumen%' OR content ILIKE '%podium%');
+--    -- Expected: 0
+--
+-- 3. Confirm Flashback project-bound test corpus (>= 5 termdeck-tagged rows
+--    matching the canonical probe question):
+--    SELECT count(*) FROM memory_items
+--     WHERE project='termdeck' AND content ILIKE '%shell error%';