gsd-pi 2.63.0 → 2.64.0-dev.1a85e85

package/src/resources/extensions/gsd/preferences.ts

@@ -224,9 +224,13 @@ export function parsePreferencesMarkdown(content: string): GSDPreferences | null
     return parseHeadingListFormat(content);
   }
 
-  if (!_warnedUnrecognizedFormat) {
+  // Warn when a non-empty file exists but lacks frontmatter delimiters (#2036).
+  if (content.trim().length > 0 && !_warnedUnrecognizedFormat) {
     _warnedUnrecognizedFormat = true;
-    console.warn("[parsePreferencesMarkdown] preferences.md exists but uses an unrecognized format — skipping.");
+    console.warn(
+      "[GSD] Warning: preferences file has unrecognized format — content does not use YAML frontmatter delimiters (---). " +
+      "Wrap your preferences in --- fences. See https://github.com/gsd-build/gsd-2/issues/2036",
+    );
   }
   return null;
 }
@@ -363,6 +367,10 @@ function mergePreferences(base: GSDPreferences, override: GSDPreferences): GSDPr
     verification_commands: mergeStringLists(base.verification_commands, override.verification_commands),
     verification_auto_fix: override.verification_auto_fix ?? base.verification_auto_fix,
     verification_max_retries: override.verification_max_retries ?? base.verification_max_retries,
+    enhanced_verification: override.enhanced_verification ?? base.enhanced_verification,
+    enhanced_verification_pre: override.enhanced_verification_pre ?? base.enhanced_verification_pre,
+    enhanced_verification_post: override.enhanced_verification_post ?? base.enhanced_verification_post,
+    enhanced_verification_strict: override.enhanced_verification_strict ?? base.enhanced_verification_strict,
     search_provider: override.search_provider ?? base.search_provider,
     context_selection: override.context_selection ?? base.context_selection,
     auto_visualize: override.auto_visualize ?? base.auto_visualize,
@@ -384,6 +392,9 @@ function mergePreferences(base: GSDPreferences, override: GSDPreferences): GSDPr
           ].filter(Boolean),
         }
       : undefined,
+    slice_parallel: (base.slice_parallel || override.slice_parallel)
+      ? { ...(base.slice_parallel ?? {}), ...(override.slice_parallel ?? {}) }
+      : undefined,
   };
 }
 

package/src/resources/extensions/gsd/prompt-loader.ts

@@ -51,6 +51,14 @@ const __extensionDir = resolveExtensionDir();
 const promptsDir = join(__extensionDir, "prompts");
 const templatesDir = join(__extensionDir, "templates");
 
+/**
+ * Return the resolved templates directory path for use in prompts.
+ * Avoids hardcoding `~/.gsd/agent/extensions/gsd/templates/` in templates. (#3575)
+ */
+export function getTemplatesDir(): string {
+  return templatesDir;
+}
+
 // Cache all templates eagerly at module load — a running session uses the
 // template versions that were on disk at startup, immune to later overwrites.
 const templateCache = new Map<string, string>();

package/src/resources/extensions/gsd/prompts/complete-milestone.md

@@ -24,6 +24,8 @@ Then:
 7. Fill the **Decision Re-evaluation** table in the milestone summary. For each key decision from `.gsd/DECISIONS.md` made during this milestone, evaluate whether it is still valid given what was actually built. Flag decisions that should be revisited next milestone.
 8. Validate **requirement status transitions**. For each requirement that changed status during this milestone, confirm the transition is supported by evidence. Requirements can move between Active, Validated, Deferred, Blocked, or Out of Scope — but only with proof.
 
+**DB access safety:** Do NOT query `.gsd/gsd.db` directly via `sqlite3` or `node -e require('better-sqlite3')` — the engine owns the WAL connection. Use `gsd_milestone_status` to read milestone and slice state. All data you need is already inlined in the context above or accessible via the `gsd_*` tools — never via direct SQL.
+
 ### Verification Gate — STOP if verification failed
 
 **If ANY verification failure was recorded in steps 3, 4, or 5, you MUST follow the failure path below. Do NOT proceed to step 9.**

package/src/resources/extensions/gsd/prompts/complete-slice.md

@@ -35,6 +35,8 @@ Then:
 
 **Autonomous execution:** Do not call `ask_user_questions` or `secure_env_collect`. You are running in auto-mode — there is no human available to answer questions. Make reasonable assumptions and document them in the slice summary. If a decision genuinely requires human input, note it in the summary and proceed with the best available option.
 
+**File system safety:** Task summaries are preloaded in the inlined context above. If you need to re-read any of them, use `find .gsd/milestones/{{milestoneId}}/slices/{{sliceId}}/tasks -name "*-SUMMARY.md"` to list file paths first — never pass `{{slicePath}}` or any other directory path directly to the `read` tool. The `read` tool only accepts file paths, not directories.
+
 **You MUST call `gsd_complete_slice` with the slice summary and UAT content before finishing. The tool persists to both DB and disk and renders `{{sliceSummaryPath}}` and `{{sliceUatPath}}` automatically.**
 
 When done, say: "Slice {{sliceId}} complete."

package/src/resources/extensions/gsd/prompts/doctor-heal.md

@@ -9,6 +9,7 @@ Rules:
 4. For missing summaries or UAT files, generate the real artifact from existing slice/task context when possible — do not leave placeholders if you can reconstruct the real content.
 5. After each repair cluster, verify the relevant invariant directly from disk.
 6. When done, rerun `/gsd doctor {{doctorCommandSuffix}}` mentally by ensuring the remaining issue set for this scope is reduced or cleared.
+7. Do NOT query `.gsd/gsd.db` directly via `sqlite3` or `node -e require('better-sqlite3')` — use `gsd_milestone_status` to inspect DB state. Direct access bypasses the WAL connection owned by the engine and can corrupt in-flight writes.
 
 ## Doctor Summary
 

package/src/resources/extensions/gsd/prompts/forensics.md

@@ -116,6 +116,8 @@ A unit dispatched more than once (`type/id` appears multiple times) indicates a
 
 5. **Read the actual GSD source code** at `{{gsdSourceDir}}` to confirm or deny each hypothesis. Do not guess what code does — read it.
 
+   **DB inspection:** If you need to check DB state as part of investigation, use `gsd_milestone_status` — never run `sqlite3 .gsd/gsd.db` or `node -e require('better-sqlite3')` directly. The engine holds a WAL write lock; direct access will either fail or return stale data.
+
 6. **Trace the code path** from the entry point (usually `auto-loop.ts` dispatch or `auto-dispatch.ts`) through to the failure point. Follow function calls across files.
 
 7. **Identify the specific file and line** where the bug lives. Determine what kind of defect it is:

package/src/resources/extensions/gsd/prompts/reassess-roadmap.md

@@ -63,4 +63,6 @@ If `.gsd/REQUIREMENTS.md` exists and requirement ownership or status changed, up
 
 {{commitInstruction}}
 
+**DB access safety:** Do NOT query `.gsd/gsd.db` directly via `sqlite3` or `node -e require('better-sqlite3')`. Use `gsd_milestone_status` to read current milestone and slice state. All roadmap mutations go through `gsd_reassess_roadmap` — the tool writes to the DB and re-renders ROADMAP.md atomically.
+
 When done, say: "Roadmap reassessed."

package/src/resources/extensions/gsd/prompts/system.md

@@ -24,13 +24,9 @@ Leave the project in a state where the next agent can immediately understand wha
 
 ## Skills
 
-GSD ships with bundled skills. Load the relevant skill file with the `read` tool before starting work when the task matches.
+GSD ships with bundled skills. Load the relevant skill file with the `read` tool before starting work when the task matches. Use bare skill names — GSD resolves them to the correct path automatically.
 
-| Trigger | Skill to load |
-|---|---|
-| Frontend UI - web components, pages, landing pages, dashboards, React/HTML/CSS, styling | `~/.gsd/agent/skills/frontend-design/SKILL.md` |
-| macOS or iOS apps - SwiftUI, Xcode, App Store | `~/.gsd/agent/skills/swiftui/SKILL.md` |
-| Debugging - complex bugs, failing tests, root-cause investigation after standard approaches fail | `~/.gsd/agent/skills/debug-like-expert/SKILL.md` |
+{{bundledSkillsTable}}
 
 ## Hard Rules
 
@@ -119,7 +115,7 @@ In all modes, slices commit sequentially on the active branch; there are no per-
 ### Artifact Templates
 
 Templates showing the expected format for each artifact type are in:
-`~/.gsd/agent/extensions/gsd/templates/`
+`{{templatesDir}}`
 
 **Always read the relevant template before writing an artifact** to match the expected structure exactly. The parsers that read these files depend on specific formatting:
 
@@ -175,6 +171,7 @@ Templates showing the expected format for each artifact type are in:
 - Never guess at library APIs from training data — use `get_library_docs`.
 - Never ask the user to run a command, set a variable, or check something you can check yourself.
 - Never await stale async jobs after editing source — `cancel_job` them first, then re-run.
+- Never query `.gsd/gsd.db` directly via `sqlite3`, `better-sqlite3`, or `node -e require('better-sqlite3')` — the database uses a single-writer WAL connection managed by the engine. Direct access causes reader/writer conflicts and bypasses validation logic. Use `gsd_milestone_status`, `gsd_journal_query`, or other `gsd_*` tools exclusively for all DB reads and writes.
 
 ### Ask vs infer
 

package/src/resources/extensions/gsd/prompts/validate-milestone.md

@@ -38,6 +38,8 @@ All relevant context has been preloaded below — the roadmap, all slice summari
 
 **Persist validation results through `gsd_validate_milestone`.** Call it with: `milestoneId`, `verdict`, `remediationRound`, `successCriteriaChecklist`, `sliceDeliveryAudit`, `crossSliceIntegration`, `requirementCoverage`, `verificationClasses` (when non-empty), `verdictRationale`, and `remediationPlan` (if verdict is `needs-remediation`). The tool writes the validation to the DB and renders VALIDATION.md to disk.
 
+**DB access safety:** Do NOT query `.gsd/gsd.db` directly via `sqlite3` or `node -e require('better-sqlite3')` — the engine owns the WAL connection. Use `gsd_milestone_status` to read milestone and slice state. All data you need is already inlined in the context above or accessible via the `gsd_*` tools. Direct DB access corrupts the WAL and bypasses tool-level validation.
+
 If verdict is `needs-remediation`:
 - After calling `gsd_validate_milestone`, use `gsd_reassess_roadmap` to add remediation slices. Pass `milestoneId`, a synthetic `completedSliceId` (e.g. "VALIDATION"), `verdict: "roadmap-adjusted"`, `assessment` text, and `sliceChanges` with the new slices in the `added` array. The tool persists the changes to the DB and re-renders ROADMAP.md.
 - These remediation slices will be planned and executed before validation re-runs.

package/src/resources/extensions/gsd/roadmap-mutations.ts

@@ -39,7 +39,7 @@ export function markSliceDoneInRoadmap(basePath: string, mid: string, sid: strin
       new RegExp(`^(#{1,4}\\s+(?:\\*{0,2})(?:Slice\\s+)?${sid}\\*{0,2}[:\\s.\\u2014\\u2013-]+\\s*)(.+)`, "m"),
       (match, prefix, title) => {
         // Already marked done — no-op
-        if (/^\u2713/.test(title) || /\(Complete\)\s*$/i.test(title)) return match;
+        if (/^[\u2713\u2705]/.test(title) || /[\u2705]\s*$/.test(title) || /\(Complete\)\s*$/i.test(title)) return match;
         return `${prefix}\u2713 ${title}`;
       },
     );

package/src/resources/extensions/gsd/roadmap-slices.ts

@@ -82,7 +82,7 @@ function parseTableSlices(section: string): RoadmapSliceEntry[] {
     const fullRow = line.toLowerCase();
     const done =
       /\[x\]/i.test(line) ||
-      /[✅☑✓]/.test(line) ||
+      /[✅☑✓✔]/.test(line) ||
       /\bdone\b/.test(fullRow) ||
       /\bcomplete(?:d)?\b/.test(fullRow);
 
@@ -222,11 +222,11 @@ function parseProseSliceHeaders(content: string): RoadmapSliceEntry[] {
   // numeric prefixes (e.g., "1.", "(1)"), bracketed IDs (e.g., "[S01]"),
   // optional checkmark completion marker, and optional leading indentation.
   // Separator after the ID is flexible: colon, dash, em/en dash, dot, or just whitespace.
-  const headerPattern = /^\s*#{1,4}\s+\*{0,2}(?:\u2713\s+)?(?:\d+[.)]\s+)?(?:\(\d+\)\s+)?(?:Slice\s+)?\[?(S\d+)\]?\*{0,2}[:\s.\u2014\u2013-]*\s*(.+)/gm;
+  const headerPattern = /^\s*#{1,4}\s+\*{0,2}(?:[\u2713\u2705]\s+)?(?:\d+[.)]\s+)?(?:\(\d+\)\s+)?(?:Slice\s+)?\[?(S\d+)\]?\*{0,2}[:\s.\u2014\u2013-]*\s*(.+)/gm;
   let match: RegExpExecArray | null;
 
   // Check for checkmark before the slice ID (e.g., "## checkmark S01: Title")
-  const prefixCheckPattern = /^\s*#{1,4}\s+\*{0,2}\u2713\s+/;
+  const prefixCheckPattern = /^\s*#{1,4}\s+\*{0,2}[\u2713\u2705]\s+/;
 
   while ((match = headerPattern.exec(content)) !== null) {
     const id = match[1]!;
@@ -240,9 +240,14 @@ function parseProseSliceHeaders(content: string): RoadmapSliceEntry[] {
     const line = match[0];
     let done = prefixCheckPattern.test(line);
 
-    if (!done && title.startsWith("\u2713")) {
+    if (!done && /^[\u2713\u2705]/.test(title)) {
       done = true;
-      title = title.replace(/^\u2713\s*/, "");
+      title = title.replace(/^[\u2713\u2705]\s*/, "");
+    }
+
+    if (!done && /[\u2705]\s*$/.test(title)) {
+      done = true;
+      title = title.replace(/\s*[\u2705]\s*$/, "");
     }
 
     if (!done && /\(Complete\)\s*$/i.test(title)) {

package/src/resources/extensions/gsd/safety/content-validator.ts

@@ -0,0 +1,98 @@
+/**
+ * Lightweight content validator for auto-mode safety harness.
+ * Validates that high-value unit outputs contain minimum expected content.
+ *
+ * Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
+ */
+
+import { existsSync, readFileSync } from "node:fs";
+import { logWarning } from "../workflow-logger.js";
+
+// ─── Types ──────────────────────────────────────────────────────────────────
+
+export interface ContentViolation {
+  severity: "warning";
+  reason: string;
+}
+
+// ─── Public API ─────────────────────────────────────────────────────────────
+
+/**
+ * Validate content quality for a completed unit.
+ * Returns an array of violations. Empty array = content looks acceptable.
+ *
+ * @param unitType - The type of unit that completed (e.g. "plan-slice")
+ * @param artifactPath - Absolute path to the primary artifact file
+ */
+export function validateContent(
+  unitType: string,
+  artifactPath: string | null,
+): ContentViolation[] {
+  if (!artifactPath || !existsSync(artifactPath)) return [];
+
+  const validator = VALIDATORS[unitType];
+  if (!validator) return [];
+
+  try {
+    const content = readFileSync(artifactPath, "utf-8");
+    return validator(content);
+  } catch (e) {
+    logWarning("safety", `content validation read failed: ${(e as Error).message}`);
+    return [];
+  }
+}
+
+// ─── Validators ─────────────────────────────────────────────────────────────
+
+type ContentValidatorFn = (content: string) => ContentViolation[];
+
+const VALIDATORS: Record<string, ContentValidatorFn> = {
+  "plan-slice": validatePlanSlice,
+  "plan-milestone": validatePlanMilestone,
+};
+
+function validatePlanSlice(content: string): ContentViolation[] {
+  const violations: ContentViolation[] = [];
+
+  // Must have at least 2 task entries (checkbox pattern)
+  const taskCount = (content.match(/- \[[ x]\] \*\*T\d+/g) || []).length;
+  if (taskCount < 2) {
+    violations.push({
+      severity: "warning",
+      reason: `Slice plan has only ${taskCount} task(s) — expected at least 2`,
+    });
+  }
+
+  // Should have a Files Likely Touched section
+  if (!content.includes("## Files Likely Touched") && !content.includes("## Files")) {
+    violations.push({
+      severity: "warning",
+      reason: "Slice plan missing 'Files Likely Touched' section",
+    });
+  }
+
+  // Should have a verification section
+  if (!content.includes("Verify") && !content.includes("verify")) {
+    violations.push({
+      severity: "warning",
+      reason: "Slice plan has no verification instructions",
+    });
+  }
+
+  return violations;
+}
+
+function validatePlanMilestone(content: string): ContentViolation[] {
+  const violations: ContentViolation[] = [];
+
+  // Must have at least 1 slice entry
+  const sliceCount = (content.match(/##\s+S\d+/g) || []).length;
+  if (sliceCount < 1) {
+    violations.push({
+      severity: "warning",
+      reason: `Milestone roadmap has ${sliceCount} slice(s) — expected at least 1`,
+    });
+  }
+
+  return violations;
+}

package/src/resources/extensions/gsd/safety/destructive-guard.ts

@@ -0,0 +1,49 @@
+/**
+ * Destructive command classifier for auto-mode safety harness.
+ * Classifies bash commands and warns on potentially destructive operations.
+ * Does NOT block — only classifies for logging/notification.
+ *
+ * Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
+ */
+
+// ─── Pattern Definitions ────────────────────────────────────────────────────
+
+interface DestructivePattern {
+  pattern: RegExp;
+  label: string;
+}
+
+const DESTRUCTIVE_PATTERNS: readonly DestructivePattern[] = [
+  { pattern: /\brm\s+(-[^\s]*[rfRF][^\s]*\s+|.*\s+-[^\s]*[rfRF])/, label: "recursive delete" },
+  { pattern: /\bgit\s+push\s+.*--force/, label: "force push" },
+  { pattern: /\bgit\s+push\s+-f\b/, label: "force push" },
+  { pattern: /\bgit\s+reset\s+--hard/, label: "hard reset" },
+  { pattern: /\bgit\s+clean\s+-[^\s]*[fdxFDX]/, label: "git clean" },
+  { pattern: /\bgit\s+checkout\s+--\s+\./, label: "discard all changes" },
+  { pattern: /\bdrop\s+(database|table|index)\b/i, label: "SQL drop" },
+  { pattern: /\btruncate\s+table\b/i, label: "SQL truncate" },
+  { pattern: /\bchmod\s+777\b/, label: "world-writable permissions" },
+  { pattern: /\bcurl\s.*\|\s*(bash|sh|zsh)\b/, label: "pipe to shell" },
+];
+
+// ─── Public API ─────────────────────────────────────────────────────────────
+
+export interface CommandClassification {
+  destructive: boolean;
+  labels: string[];
+}
+
+/**
+ * Classify a bash command for destructive operations.
+ * Returns the list of matched destructive pattern labels.
+ */
+export function classifyCommand(command: string): CommandClassification {
+  const labels: string[] = [];
+  for (const { pattern, label } of DESTRUCTIVE_PATTERNS) {
+    if (pattern.test(command)) {
+      // Deduplicate labels (e.g., two force-push patterns)
+      if (!labels.includes(label)) labels.push(label);
+    }
+  }
+  return { destructive: labels.length > 0, labels };
+}

package/src/resources/extensions/gsd/safety/evidence-collector.ts

@@ -0,0 +1,151 @@
+/**
+ * Real-time tool call evidence collector for auto-mode safety harness.
+ * Tracks every bash command, file write, and file edit during a unit execution.
+ * Evidence is compared against LLM completion claims in evidence-cross-ref.ts.
+ *
+ * Follows the same module-level Map pattern as auto-tool-tracking.ts.
+ * Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
+ */
+
+// ─── Types ──────────────────────────────────────────────────────────────────
+
+export interface BashEvidence {
+  kind: "bash";
+  toolCallId: string;
+  command: string;
+  exitCode: number;
+  outputSnippet: string;
+  timestamp: number;
+}
+
+export interface FileWriteEvidence {
+  kind: "write";
+  toolCallId: string;
+  path: string;
+  timestamp: number;
+}
+
+export interface FileEditEvidence {
+  kind: "edit";
+  toolCallId: string;
+  path: string;
+  timestamp: number;
+}
+
+export type EvidenceEntry = BashEvidence | FileWriteEvidence | FileEditEvidence;
+
+// ─── Module State ───────────────────────────────────────────────────────────
+
+let unitEvidence: EvidenceEntry[] = [];
+
+// ─── Public API ─────────────────────────────────────────────────────────────
+
+/** Reset all evidence for a new unit. Call at unit start. */
+export function resetEvidence(): void {
+  unitEvidence = [];
+}
+
+/** Get a read-only view of all evidence collected for the current unit. */
+export function getEvidence(): readonly EvidenceEntry[] {
+  return unitEvidence;
+}
+
+/** Get only bash evidence entries. */
+export function getBashEvidence(): readonly BashEvidence[] {
+  return unitEvidence.filter((e): e is BashEvidence => e.kind === "bash");
+}
+
+/** Get all file paths touched (write + edit). */
+export function getFilePaths(): string[] {
+  return unitEvidence
+    .filter((e): e is FileWriteEvidence | FileEditEvidence => e.kind === "write" || e.kind === "edit")
+    .map(e => e.path);
+}
+
+// ─── Recording (called from register-hooks.ts) ─────────────────────────────
+
+/**
+ * Record a tool call at dispatch time (before execution).
+ * Exit codes and output are filled in by recordToolResult after execution.
+ */
+export function recordToolCall(toolName: string, input: Record<string, unknown>): void {
+  if (toolName === "bash" || toolName === "Bash") {
+    unitEvidence.push({
+      kind: "bash",
+      toolCallId: "",
+      command: String(input.command ?? ""),
+      exitCode: -1,
+      outputSnippet: "",
+      timestamp: Date.now(),
+    });
+  } else if (toolName === "write" || toolName === "Write") {
+    unitEvidence.push({
+      kind: "write",
+      toolCallId: "",
+      path: String(input.file_path ?? input.path ?? ""),
+      timestamp: Date.now(),
+    });
+  } else if (toolName === "edit" || toolName === "Edit") {
+    unitEvidence.push({
+      kind: "edit",
+      toolCallId: "",
+      path: String(input.file_path ?? input.path ?? ""),
+      timestamp: Date.now(),
+    });
+  }
+}
+
+/**
+ * Record a tool execution result. Matches the most recent unresolved entry
+ * of the same kind and fills in the toolCallId, exit code, and output.
+ */
+export function recordToolResult(
+  toolCallId: string,
+  toolName: string,
+  result: unknown,
+  isError: boolean,
+): void {
+  const normalizedName = toolName.toLowerCase();
+
+  if (normalizedName === "bash") {
+    const entry = findLastUnresolved("bash") as BashEvidence | undefined;
+    if (entry) {
+      entry.toolCallId = toolCallId;
+      const text = extractResultText(result);
+      entry.outputSnippet = text.slice(0, 500);
+      const exitMatch = text.match(/Command exited with code (\d+)/);
+      entry.exitCode = exitMatch ? Number(exitMatch[1]) : (isError ? 1 : 0);
+    }
+  } else if (normalizedName === "write" || normalizedName === "edit") {
+    const entry = findLastUnresolved(normalizedName as "write" | "edit");
+    if (entry) {
+      entry.toolCallId = toolCallId;
+    }
+  }
+}
+
+// ─── Internals ──────────────────────────────────────────────────────────────
+
+function findLastUnresolved(kind: string): EvidenceEntry | undefined {
+  for (let i = unitEvidence.length - 1; i >= 0; i--) {
+    if (unitEvidence[i].kind === kind && unitEvidence[i].toolCallId === "") {
+      return unitEvidence[i];
+    }
+  }
+  return undefined;
+}
+
+function extractResultText(result: unknown): string {
+  if (typeof result === "string") return result;
+  if (result && typeof result === "object") {
+    const r = result as Record<string, unknown>;
+    if (Array.isArray(r.content)) {
+      const textBlock = r.content.find(
+        (c: unknown) => typeof c === "object" && c !== null && (c as Record<string, unknown>).type === "text",
+      ) as Record<string, unknown> | undefined;
+      if (textBlock && typeof textBlock.text === "string") return textBlock.text;
+    }
+    if (typeof r.text === "string") return r.text;
+  }
+  return String(result ?? "");
+}

package/src/resources/extensions/gsd/safety/evidence-cross-ref.ts

@@ -0,0 +1,120 @@
+/**
+ * Evidence cross-reference for auto-mode safety harness.
+ * Compares the LLM's claimed verification evidence (command + exitCode)
+ * against actual bash tool calls recorded by the evidence collector.
+ *
+ * Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
+ */
+
+import type { BashEvidence, EvidenceEntry } from "./evidence-collector.js";
+
+// ─── Types ──────────────────────────────────────────────────────────────────
+
+export interface ClaimedEvidence {
+  command: string;
+  exitCode: number;
+  verdict: string;
+}
+
+export interface EvidenceMismatch {
+  severity: "warning" | "error";
+  claimed: ClaimedEvidence;
+  actual: BashEvidence | null;
+  reason: string;
+}
+
+// ─── Public API ─────────────────────────────────────────────────────────────
+
+/**
+ * Cross-reference claimed verification evidence against actual bash tool calls.
+ *
+ * Returns an array of mismatches. Empty array = all claims verified.
+ * Skips entries that were coerced from strings (already flagged by db-tools.ts).
+ */
+export function crossReferenceEvidence(
+  claimedEvidence: readonly ClaimedEvidence[],
+  actualEvidence: readonly EvidenceEntry[],
+): EvidenceMismatch[] {
+  const bashCalls = actualEvidence.filter(
+    (e): e is BashEvidence => e.kind === "bash",
+  );
+  const mismatches: EvidenceMismatch[] = [];
+
+  for (const claimed of claimedEvidence) {
+    // Skip coerced entries — they're already flagged with exitCode: -1
+    // and verdict: "unknown (coerced from string)" by db-tools.ts
+    if (claimed.verdict?.includes("coerced from string")) continue;
+    if (claimed.exitCode === -1) continue;
+
+    // Skip entries with empty or generic commands
+    if (!claimed.command || claimed.command.length < 3) continue;
+
+    // Find matching bash call by command substring match
+    const match = findBestMatch(claimed.command, bashCalls);
+
+    if (!match) {
+      mismatches.push({
+        severity: "warning",
+        claimed,
+        actual: null,
+        reason: `No bash tool call found matching "${claimed.command.slice(0, 80)}"`,
+      });
+      continue;
+    }
+
+    // Exit code mismatch: LLM claims success but actual command failed
+    if (claimed.exitCode === 0 && match.exitCode !== 0) {
+      mismatches.push({
+        severity: "error",
+        claimed,
+        actual: match,
+        reason: `Claimed exitCode=0 but actual exitCode=${match.exitCode}`,
+      });
+    }
+  }
+
+  return mismatches;
+}
+
+// ─── Internals ──────────────────────────────────────────────────────────────
+
+/**
+ * Find the best matching bash evidence entry for a claimed command.
+ * Uses substring matching — the claimed command may be a shortened version
+ * of the actual command, or vice versa.
+ */
+function findBestMatch(
+  claimedCommand: string,
+  bashCalls: readonly BashEvidence[],
+): BashEvidence | null {
+  const normalized = claimedCommand.trim();
+
+  // Exact match first
+  const exact = bashCalls.find(b => b.command.trim() === normalized);
+  if (exact) return exact;
+
+  // Substring match: claimed is contained in actual or actual in claimed
+  const substring = bashCalls.find(
+    b => b.command.includes(normalized) || normalized.includes(b.command),
+  );
+  if (substring) return substring;
+
+  // Token match: split on whitespace and check significant overlap
+  const claimedTokens = normalized.split(/\s+/).filter(t => t.length > 2);
+  if (claimedTokens.length === 0) return null;
+
+  let bestMatch: BashEvidence | null = null;
+  let bestScore = 0;
+
+  for (const call of bashCalls) {
+    const callTokens = new Set(call.command.split(/\s+/));
+    const matchCount = claimedTokens.filter(t => callTokens.has(t)).length;
+    const score = matchCount / claimedTokens.length;
+    if (score > bestScore && score >= 0.5) {
+      bestScore = score;
+      bestMatch = call;
+    }
+  }
+
+  return bestMatch;
+}