flatbread-cli 0.3.1 → 0.3.2

package/dist/flatbread.mjs

@@ -4307,6 +4307,73 @@ function tallyFailures(loops) {
   };
 }
 
+// ../schema/src/harness.ts
+var HARNESS_INFO = {
+  "claude-code": {
+    id: "claude-code",
+    label: "Claude Code",
+    formatLabel: "session JSONL (~/.claude/projects)",
+    confidence: "observed",
+    implemented: true,
+    humanTest: "A1",
+    toolResults: "measured",
+    note: "Built from real observed logs. Anthropic publishes no format guarantee, so watch the unknown-type counter."
+  },
+  cursor: {
+    id: "cursor",
+    label: "Cursor",
+    formatLabel: "state.vscdb SQLite (cursorDiskKV + ItemTable)",
+    confidence: "documented",
+    implemented: true,
+    humanTest: "A5",
+    toolResults: "unavailable",
+    note: "Reads composerData:/bubbleId: from cursorDiskKV (modern) and aiService.prompts from ItemTable (legacy). Cursor publishes no schema and has moved these keys between releases."
+  },
+  opencode: {
+    id: "opencode",
+    label: "OpenCode",
+    formatLabel: "SQLite opencode.db (1.3+) or storage/ JSON tree (1.0.x)",
+    confidence: "documented",
+    implemented: true,
+    humanTest: "A4",
+    toolResults: "measured",
+    note: "Two layouts. 1.3+ keeps sessions in an SQLite opencode.db (schema read off a real 1.3.17 install); 1.0.x kept a session/message/part JSON tree. OpenCode publishes no schema and its migrations have moved this once already, so a version bump can move it again."
+  },
+  antigravity: {
+    id: "antigravity",
+    label: "Google Antigravity",
+    formatLabel: "app brain/<id>/.system_generated/logs/transcript.jsonl, or IDE conversations/<id>.pb",
+    confidence: "documented",
+    implemented: true,
+    humanTest: "A6",
+    toolResults: "unavailable",
+    note: "Both stores measured on one machine. The app's transcript is read the way deja-vu reads it: prompts, replies and the planner's tool calls, not tool output. The IDE's .pb is read the way deja-vu reads Windsurf: prompts and replies only."
+  },
+  terminal: {
+    id: "terminal",
+    label: "Terminal / PTY recording",
+    formatLabel: "asciinema cast v2, or a plain text transcript",
+    confidence: "spec",
+    implemented: true,
+    humanTest: "A3",
+    toolResults: "unavailable",
+    note: "Cast format is specified. Segmenting a raw terminal stream into prompts and replies is heuristic and will misread unfamiliar prompts."
+  },
+  otel: {
+    id: "otel",
+    label: "OpenTelemetry",
+    formatLabel: "OTLP/JSON spans with GenAI semantic conventions",
+    confidence: "spec",
+    implemented: true,
+    humanTest: "A2",
+    toolResults: "measured",
+    note: "Both the wire format and the gen_ai.* attribute names are specified. Unproven: that real instrumentation populates them."
+  }
+};
+function failuresRecorded(harness) {
+  return HARNESS_INFO[harness].toolResults === "measured";
+}
+
 // ../schema/src/brief.ts
 var MUTATING_TOOL_NAMES = /* @__PURE__ */ new Set([
   "write",
@@ -4642,14 +4709,15 @@ function runDiagnostics(trace) {
     if (hit) waste.push({ kind: "repeated-search", label, steps: hit, windowMinutes: WASTE_RULES.repeatedSearch.windowMinutes });
   }
   const stats = trace.stats;
+  const recorded = failuresRecorded(trace.source.harness);
   return {
     counts: {
       steps: steps.length,
       agentSteps: steps.length - prompts.length,
       humanPrompts: prompts.length,
       toolCalls: stats.toolCallCount,
-      failures: steps.filter((step) => step.type === "error_recovery").length,
-      recoveries: loops.filter((loop) => loop.outcome === "recovered").length
+      failures: recorded ? steps.filter((step) => step.type === "error_recovery").length : null,
+      recoveries: recorded ? loops.filter((loop) => loop.outcome === "recovered").length : null
     },
     tokens: {
       input: stats.inputTokens,
@@ -4778,64 +4846,6 @@ function buildExecutionBrief(trace) {
   };
 }
 
-// ../schema/src/harness.ts
-var HARNESS_INFO = {
-  "claude-code": {
-    id: "claude-code",
-    label: "Claude Code",
-    formatLabel: "session JSONL (~/.claude/projects)",
-    confidence: "observed",
-    implemented: true,
-    humanTest: "A1",
-    note: "Built from real observed logs. Anthropic publishes no format guarantee, so watch the unknown-type counter."
-  },
-  cursor: {
-    id: "cursor",
-    label: "Cursor",
-    formatLabel: "state.vscdb SQLite (cursorDiskKV + ItemTable)",
-    confidence: "documented",
-    implemented: true,
-    humanTest: "A5",
-    note: "Reads composerData:/bubbleId: from cursorDiskKV (modern) and aiService.prompts from ItemTable (legacy). Cursor publishes no schema and has moved these keys between releases."
-  },
-  opencode: {
-    id: "opencode",
-    label: "OpenCode",
-    formatLabel: "SQLite opencode.db (1.3+) or storage/ JSON tree (1.0.x)",
-    confidence: "documented",
-    implemented: true,
-    humanTest: "A4",
-    note: "Two layouts. 1.3+ keeps sessions in an SQLite opencode.db (schema read off a real 1.3.17 install); 1.0.x kept a session/message/part JSON tree. OpenCode publishes no schema and its migrations have moved this once already, so a version bump can move it again."
-  },
-  antigravity: {
-    id: "antigravity",
-    label: "Google Antigravity",
-    formatLabel: "app brain/<id>/.system_generated/logs/transcript.jsonl, or IDE conversations/<id>.pb",
-    confidence: "documented",
-    implemented: true,
-    humanTest: "A6",
-    note: "Both stores measured on one machine. The app's transcript is read the way deja-vu reads it: prompts, replies and the planner's tool calls, not tool output. The IDE's .pb is read the way deja-vu reads Windsurf: prompts and replies only."
-  },
-  terminal: {
-    id: "terminal",
-    label: "Terminal / PTY recording",
-    formatLabel: "asciinema cast v2, or a plain text transcript",
-    confidence: "spec",
-    implemented: true,
-    humanTest: "A3",
-    note: "Cast format is specified. Segmenting a raw terminal stream into prompts and replies is heuristic and will misread unfamiliar prompts."
-  },
-  otel: {
-    id: "otel",
-    label: "OpenTelemetry",
-    formatLabel: "OTLP/JSON spans with GenAI semantic conventions",
-    confidence: "spec",
-    implemented: true,
-    humanTest: "A2",
-    note: "Both the wire format and the gen_ai.* attribute names are specified. Unproven: that real instrumentation populates them."
-  }
-};
-
 // ../schema/src/capsule.ts
 var CAPSULE_LIMITS = {
   /** Diff lines across every patch. Past it the rest is counted, never silently dropped. */
@@ -6000,7 +6010,7 @@ function buildTrace(input) {
   const title = titleFromPrompt(firstPrompt?.content ?? "");
   const filesChanged = input.capabilities.filesChanged === "measured" ? changedFiles.size : null;
   const filesPhrase = filesChanged === null ? "an unrecorded number of files changed" : `${plural2(filesChanged, "file")} changed`;
-  const summary = `${plural2(steps.length, "step")}, ${plural2(toolCallCount, "tool call")}, ${filesPhrase}, ${plural2(errorCount, "tool error")} over ${Math.round(durationSeconds)}s.`;
+  const summary = `${plural2(steps.length, "step")}, ${plural2(toolCallCount, "tool call")}, ${filesPhrase}, ${failuresRecorded(input.harness) ? plural2(errorCount, "tool error") : "tool errors not recorded"} over ${Math.round(durationSeconds)}s.`;
   const tokensMeasured = input.capabilities.tokens === "measured";
   const provenance = {
     id: input.harness,
@@ -10204,7 +10214,7 @@ function renderReviewPage(input) {
   <h1>${marked("title", trace.title)}</h1>
   <div class="meta">
     <b>${escapeHtml(trace.model)}</b> &middot; ${plural2(trace.steps.length, "step")} &middot; ${plural2(toolCalls, "tool call")} &middot;
-    ${trace.stats.filesChanged === null ? `${formatCount(null)} files` : plural2(trace.stats.filesChanged, "file")} changed &middot; ${plural2(trace.stats.errorCount, "error")} &middot;
+    ${trace.stats.filesChanged === null ? `${formatCount(null)} files` : plural2(trace.stats.filesChanged, "file")} changed &middot; ${failuresRecorded(trace.source.harness) ? plural2(trace.stats.errorCount, "error") : `${formatCount(null)} errors`} &middot;
     ${Math.round(trace.stats.durationSeconds)}s &middot; <b>${matches.length}</b> redaction match${matches.length === 1 ? "" : "es"}
   </div>
   <div class="meta path">${escapeHtml(sessionPath)}</div>
@@ -10786,7 +10796,7 @@ function diagnosticsLines(diagnostics) {
   const { counts, tokens } = diagnostics;
   const out = [
     `${plural3(counts.steps, "step")}: ${counts.agentSteps.toLocaleString("en-US")} by the agent, ${plural3(counts.humanPrompts, "prompt")} from a person`,
-    `${plural3(counts.toolCalls, "tool call")}, ${plural3(counts.failures, "failure")}, ${plural3(counts.recoveries, "recovery", "recoveries")} by the agent`
+    counts.failures === null || counts.recoveries === null ? `${plural3(counts.toolCalls, "tool call")}; failures not recorded by this harness` : `${plural3(counts.toolCalls, "tool call")}, ${plural3(counts.failures, "failure")}, ${plural3(counts.recoveries, "recovery", "recoveries")} by the agent`
   ];
   const recorded = [];
   if (tokens.input !== null) recorded.push(["in", tokens.input]);
@@ -11069,7 +11079,7 @@ function logCapture(dir, line, now = (/* @__PURE__ */ new Date()).toISOString())
 
 // src/mcp.ts
 var MCP_PROTOCOL_VERSIONS = ["2025-06-18", "2025-03-26", "2024-11-05"];
-var MCP_SERVER_VERSION = "0.3.1";
+var MCP_SERVER_VERSION = "0.3.2";
 function defaultMcpDeps(storeDir2, bucketDir2) {
   return {
     storeDir: storeDir2,
@@ -12732,7 +12742,7 @@ function report(result, outputPath) {
   lines2.push(`  tool calls   ${formatInt(trace.stats.toolCallCount)}`);
   lines2.push(`  tools used   ${result.toolNames.length > 0 ? result.toolNames.join(", ") : "(none)"}`);
   lines2.push(`  files changed ${formatCount(trace.stats.filesChanged)}`);
-  lines2.push(`  tool errors  ${formatInt(trace.stats.errorCount)}`);
+  lines2.push(`  tool errors  ${failuresRecorded(trace.source.harness) ? formatInt(trace.stats.errorCount) : formatCount(null)}`);
   lines2.push("");
   lines2.push(
     `  tokens       in ${formatCount(trace.stats.inputTokens)} / out ${formatCount(trace.stats.outputTokens)} / cache read ${formatCount(trace.stats.cacheReadTokens)} / cache write ${formatCount(trace.stats.cacheCreationTokens)}`

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "flatbread-cli",
-  "version": "0.3.1",
+  "version": "0.3.2",
   "description": "Flatbread: read a coding agent's session off disk, redact it, share it, and pull a run's Context Capsule into another agent.",
   "type": "module",
   "bin": {