Merge pull request #3576 from jeremymcs/feat/llm-safety-harness

feat(gsd): LLM safety harness for auto-mode damage control

packages/pi-coding-agent/src/core/extensions/provider-registration.test.ts

@@ -0,0 +1,81 @@
+// GSD2 — Regression test: pendingProviderRegistrations must be flushed exactly once (#3576)
+// Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
+
+import { describe, it } from "node:test";
+import assert from "node:assert/strict";
+
+/**
+ * This test validates that the provider preflush pattern in sdk.ts clears
+ * pendingProviderRegistrations after iterating, so bindCore() doesn't
+ * re-register the same providers.
+ *
+ * The bug: createAgentSession() iterated pendingProviderRegistrations but
+ * did not clear the array. Later, bindCore() replayed and registered the
+ * same providers again, stacking wrappers.
+ */
+
+interface ProviderEntry {
+  name: string;
+  config: Record<string, unknown>;
+}
+
+interface MockRuntime {
+  pendingProviderRegistrations: ProviderEntry[];
+}
+
+describe("provider registration preflush", () => {
+  it("clears pending registrations after preflush so bindCore does not replay", () => {
+    const registered: string[] = [];
+    const runtime: MockRuntime = {
+      pendingProviderRegistrations: [
+        { name: "ollama", config: { type: "ollama" } },
+        { name: "custom-provider", config: { type: "custom" } },
+      ],
+    };
+
+    // Simulate sdk.ts preflush (lines 220-223)
+    for (const { name } of runtime.pendingProviderRegistrations) {
+      registered.push(name);
+    }
+    // The fix: clear after preflush
+    runtime.pendingProviderRegistrations = [];
+
+    // Simulate bindCore() flush (runner.ts lines 268-271)
+    for (const { name } of runtime.pendingProviderRegistrations) {
+      registered.push(name);
+    }
+    runtime.pendingProviderRegistrations = [];
+
+    assert.deepEqual(
+      registered,
+      ["ollama", "custom-provider"],
+      "each provider should be registered exactly once",
+    );
+  });
+
+  it("without the fix, providers are registered twice", () => {
+    const registered: string[] = [];
+    const runtime: MockRuntime = {
+      pendingProviderRegistrations: [
+        { name: "ollama", config: { type: "ollama" } },
+      ],
+    };
+
+    // Old behavior: preflush without clearing
+    for (const { name } of runtime.pendingProviderRegistrations) {
+      registered.push(name);
+    }
+    // NOT clearing — simulating the old bug
+
+    // bindCore() replays the same queue
+    for (const { name } of runtime.pendingProviderRegistrations) {
+      registered.push(name);
+    }
+
+    assert.deepEqual(
+      registered,
+      ["ollama", "ollama"],
+      "without clearing, providers are registered twice (demonstrating the bug)",
+    );
+  });
+});

packages/pi-coding-agent/src/core/sdk.ts

@@ -221,9 +221,8 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 	for (const { name, config } of extensionsForModelResolution.runtime.pendingProviderRegistrations) {
 		modelRegistry.registerProvider(name, config);
 	}
-	// Note: we do NOT clear pendingProviderRegistrations here — bindCore() will iterate
-	// an empty array harmlessly, and clearing here would require the runtime to track
-	// whether the flush already happened.
+	// Clear the queue so bindCore() doesn't re-register the same providers.
+	extensionsForModelResolution.runtime.pendingProviderRegistrations = [];
 
 	// If still no model, use findInitialModel (checks settings default, then provider defaults)
 	if (!model) {

src/resources/extensions/gsd/auto-post-unit.ts

@@ -52,6 +52,13 @@ import { hasPendingCaptures, loadPendingCaptures, revertExecutorResolvedCaptures
 import { debugLog } from "./debug-logger.js";
 import { runSafely } from "./auto-utils.js";
 import type { AutoSession, SidecarItem } from "./auto/session.js";
+import { getEvidence } from "./safety/evidence-collector.js";
+import { validateFileChanges } from "./safety/file-change-validator.js";
+// crossReferenceEvidence available for future use when verification_evidence is stored in DB
+// import { crossReferenceEvidence, type ClaimedEvidence } from "./safety/evidence-cross-ref.js";
+import { validateContent } from "./safety/content-validator.js";
+import { resolveSafetyHarnessConfig } from "./safety/safety-harness.js";
+import { resolveExpectedArtifactPath as resolveArtifactForContent } from "./auto-artifact-paths.js";
 
 /** Maximum verification retry attempts before escalating to blocker placeholder (#2653). */
 const MAX_VERIFICATION_RETRIES = 3;
@@ -437,6 +444,87 @@ export async function postUnitPreVerification(pctx: PostUnitContext, opts?: PreV
       debugLog("postUnit", { phase: "rogue-detection", error: String(e) });
     }
 
+    // ── Safety harness: post-unit validation ──
+    try {
+      const { loadEffectiveGSDPreferences } = await import("./preferences.js");
+      const prefs = loadEffectiveGSDPreferences()?.preferences;
+      const safetyConfig = resolveSafetyHarnessConfig(
+        prefs?.safety_harness as Record<string, unknown> | undefined,
+      );
+
+      if (safetyConfig.enabled) {
+        const { milestone: sMid, slice: sSid, task: sTid } = parseUnitId(s.currentUnit.id);
+
+        // File change validation (execute-task only, after auto-commit)
+        if (safetyConfig.file_change_validation && s.currentUnit.type === "execute-task" && sMid && sSid && sTid && isDbAvailable()) {
+          try {
+            const taskRow = getTask(sMid, sSid, sTid);
+            if (taskRow) {
+              const expectedOutput = taskRow.expected_output ?? [];
+              const plannedFiles = taskRow.files ?? [];
+              const audit = validateFileChanges(s.basePath, expectedOutput, plannedFiles);
+              if (audit && audit.violations.length > 0) {
+                const warnings = audit.violations.filter(v => v.severity === "warning");
+                for (const v of warnings) {
+                  logWarning("safety", `file-change: ${v.file} — ${v.reason}`);
+                }
+                if (warnings.length > 0) {
+                  ctx.ui.notify(
+                    `Safety: ${warnings.length} unexpected file change(s) outside task plan`,
+                    "warning",
+                  );
+                }
+              }
+            }
+          } catch (e) {
+            debugLog("postUnit", { phase: "safety-file-change", error: String(e) });
+          }
+        }
+
+        // Evidence cross-reference (execute-task only)
+        // Verification evidence is passed via the complete-task tool call and
+        // stored in the SUMMARY.md on disk — not available as structured data
+        // in the DB. The evidence collector tracks actual bash tool calls, so
+        // we can still detect units that claimed success but ran no commands.
+        if (safetyConfig.evidence_cross_reference && s.currentUnit.type === "execute-task") {
+          try {
+            const actual = getEvidence();
+            const bashCalls = actual.filter(e => e.kind === "bash");
+            // If the task is marked complete but zero bash commands were run,
+            // it's suspicious — the LLM may have fabricated results.
+            if (sMid && sSid && sTid && isDbAvailable()) {
+              const taskRow = getTask(sMid, sSid, sTid);
+              if (taskRow?.status === "complete" && taskRow.verify && bashCalls.length === 0) {
+                logWarning("safety", "task marked complete with verification commands but no bash calls were executed");
+                ctx.ui.notify(
+                  `Safety: task ${sTid} has verification commands but no bash calls were recorded`,
+                  "warning",
+                );
+              }
+            }
+          } catch (e) {
+            debugLog("postUnit", { phase: "safety-evidence-xref", error: String(e) });
+          }
+        }
+
+        // Content validation (plan-slice, plan-milestone)
+        if (safetyConfig.content_validation) {
+          try {
+            const artifactPath = resolveArtifactForContent(s.currentUnit.type, s.currentUnit.id, s.basePath);
+            const contentViolations = validateContent(s.currentUnit.type, artifactPath);
+            for (const v of contentViolations) {
+              logWarning("safety", `content: ${v.reason}`);
+              ctx.ui.notify(`Content validation: ${v.reason}`, "warning");
+            }
+          } catch (e) {
+            debugLog("postUnit", { phase: "safety-content-validation", error: String(e) });
+          }
+        }
+      }
+    } catch (e) {
+      debugLog("postUnit", { phase: "safety-harness", error: String(e) });
+    }
+
     // Artifact verification
     let triggerArtifactVerified = false;
     if (!s.currentUnit.type.startsWith("hook/")) {

src/resources/extensions/gsd/auto-timers.ts

@@ -106,8 +106,9 @@ export function startUnitSupervision(sctx: SupervisionContext): void {
     }
   }
   const estimateMinutes = taskEstimate ? parseEstimateMinutes(taskEstimate) : null;
+  const MAX_TIMEOUT_SCALE = 6; // Cap at 6x (60min task). Prevents 2h+ tasks from creating 120min+ timeout windows.
   const timeoutScale = estimateMinutes && estimateMinutes > 0
-    ? Math.max(1, estimateMinutes / 10)  // 10min task = 1x, 30min = 3x, 2h = 12x
+    ? Math.min(MAX_TIMEOUT_SCALE, Math.max(1, estimateMinutes / 10))
     : 1;
 
   const softTimeoutMs = (supervisor.soft_timeout_minutes ?? 0) * 60 * 1000 * timeoutScale;

src/resources/extensions/gsd/auto/phases.ts

@@ -37,6 +37,9 @@ import { withTimeout, FINALIZE_POST_TIMEOUT_MS } from "./finalize-timeout.js";
 import { getEligibleSlices } from "../slice-parallel-eligibility.js";
 import { startSliceParallel } from "../slice-parallel-orchestrator.js";
 import { isDbAvailable, getMilestoneSlices } from "../gsd-db.js";
+import { resetEvidence } from "../safety/evidence-collector.js";
+import { createCheckpoint, cleanupCheckpoint, rollbackToCheckpoint } from "../safety/git-checkpoint.js";
+import { resolveSafetyHarnessConfig } from "../safety/safety-harness.js";
 
 // ─── generateMilestoneReport ──────────────────────────────────────────────────
 
@@ -1079,6 +1082,21 @@ export async function runUnitPhase(
   if (mid)
     deps.updateSliceProgressCache(s.basePath, mid, state.activeSlice?.id);
 
+  // ── Safety harness: reset evidence + create checkpoint ──
+  const safetyConfig = resolveSafetyHarnessConfig(
+    prefs?.safety_harness as Record<string, unknown> | undefined,
+  );
+  if (safetyConfig.enabled && safetyConfig.evidence_collection) {
+    resetEvidence();
+  }
+  // Only checkpoint code-executing units (not lifecycle/planning units)
+  if (safetyConfig.enabled && safetyConfig.checkpoints && unitType === "execute-task") {
+    s.checkpointSha = createCheckpoint(s.basePath, unitId);
+    if (s.checkpointSha) {
+      debugLog("runUnitPhase", { phase: "checkpoint-created", unitId, sha: s.checkpointSha.slice(0, 8) });
+    }
+  }
+
   // Prompt injection
   let finalPrompt = prompt;
 
@@ -1376,6 +1394,27 @@ export async function runUnitPhase(
 
   deps.emitJournalEvent({ ts: new Date().toISOString(), flowId: ic.flowId, seq: ic.nextSeq(), eventType: "unit-end", data: { unitType, unitId, status: unitResult.status, artifactVerified, ...(unitResult.errorContext ? { errorContext: unitResult.errorContext } : {}) }, causedBy: { flowId: ic.flowId, seq: unitStartSeq } });
 
+  // ── Safety harness: checkpoint cleanup or rollback ──
+  if (s.checkpointSha) {
+    if (unitResult.status === "error" && safetyConfig.auto_rollback) {
+      const rolled = rollbackToCheckpoint(s.basePath, unitId, s.checkpointSha);
+      if (rolled) {
+        ctx.ui.notify(`Rolled back to pre-unit checkpoint for ${unitId}`, "info");
+        debugLog("runUnitPhase", { phase: "checkpoint-rollback", unitId });
+      }
+    } else if (unitResult.status === "error") {
+      ctx.ui.notify(
+        `Unit ${unitId} failed. Pre-unit checkpoint available at ${s.checkpointSha.slice(0, 8)}`,
+        "warning",
+      );
+    } else {
+      // Success — clean up checkpoint ref
+      cleanupCheckpoint(s.basePath, unitId);
+      debugLog("runUnitPhase", { phase: "checkpoint-cleaned", unitId });
+    }
+    s.checkpointSha = null;
+  }
+
   return { action: "next", data: { unitStartedAt: s.currentUnit?.startedAt } };
 }
 

src/resources/extensions/gsd/auto/session.ts

@@ -145,6 +145,10 @@ export class AutoSession {
   lastBaselineCharCount: number | undefined;
   pendingQuickTasks: CaptureEntry[] = [];
 
+  // ── Safety harness ───────────────────────────────────────────────────────
+  /** SHA of the pre-unit git checkpoint ref. Cleared on success or rollback. */
+  checkpointSha: string | null = null;
+
   // ── Signal handler ───────────────────────────────────────────────────────
   sigtermHandler: (() => void) | null = null;
 
@@ -223,6 +227,7 @@ export class AutoSession {
     this.lastToolInvocationError = null;
     this.isolationDegraded = false;
     this.milestoneMergedInPhases = false;
+    this.checkpointSha = null;
 
     // Signal handler
     this.sigtermHandler = null;

src/resources/extensions/gsd/bootstrap/register-hooks.ts

@@ -18,6 +18,9 @@ import { isParallelActive, shutdownParallel } from "../parallel-orchestrator.js"
 import { checkToolCallLoop, resetToolCallLoopGuard } from "./tool-call-loop-guard.js";
 import { saveActivityLog } from "../activity-log.js";
 import { resetAskUserQuestionsCache } from "../../ask-user-questions.js";
+import { recordToolCall as safetyRecordToolCall, recordToolResult as safetyRecordToolResult } from "../safety/evidence-collector.js";
+import { classifyCommand } from "../safety/destructive-guard.js";
+import { logWarning as safetyLogWarning } from "../workflow-logger.js";
 
 // Skip the welcome screen on the very first session_start — cli.ts already
 // printed it before the TUI launched. Only re-print on /clear (subsequent sessions).
@@ -203,6 +206,26 @@ export function registerHooks(pi: ExtensionAPI): void {
     if (result.block) return result;
   });
 
+  // ── Safety harness: evidence collection + destructive command warnings ──
+  pi.on("tool_call", async (event, ctx) => {
+    if (!isAutoActive()) return;
+    safetyRecordToolCall(event.toolName, event.input as Record<string, unknown>);
+
+    // Destructive command classification (warn only, never block)
+    if (isToolCallEventType("bash", event)) {
+      const classification = classifyCommand(event.input.command);
+      if (classification.destructive) {
+        safetyLogWarning("safety", `destructive command: ${classification.labels.join(", ")}`, {
+          command: String(event.input.command).slice(0, 200),
+        });
+        ctx.ui.notify(
+          `Destructive command detected: ${classification.labels.join(", ")}`,
+          "warning",
+        );
+      }
+    }
+  });
+
   pi.on("tool_result", async (event) => {
     if (event.toolName !== "ask_user_questions") return;
     const milestoneId = getDiscussionMilestoneId();
@@ -268,6 +291,10 @@ export function registerHooks(pi: ExtensionAPI): void {
         : (typeof event.result?.content?.[0]?.text === "string" ? event.result.content[0].text : String(event.result));
       recordToolInvocationError(event.toolName, errorText);
     }
+    // Safety harness: record tool execution results for evidence cross-referencing
+    if (isAutoActive()) {
+      safetyRecordToolResult(event.toolCallId, event.toolName, event.result, event.isError);
+    }
   });
 
   pi.on("model_select", async (_event, ctx) => {

src/resources/extensions/gsd/preferences-types.ts

@@ -105,6 +105,7 @@ export const KNOWN_PREFERENCE_KEYS = new Set<string>([
   "experimental",
   "codebase",
   "slice_parallel",
+  "safety_harness",
 ]);
 
 /** Canonical list of all dispatch unit types. */
@@ -291,6 +292,18 @@ export interface GSDPreferences {
   codebase?: CodebaseMapPreferences;
   /** Slice-level parallelism within a milestone. Disabled by default. */
   slice_parallel?: { enabled?: boolean; max_workers?: number };
+  /** LLM safety harness configuration. Monitors, validates, and constrains LLM behavior during auto-mode. Enabled by default with warn-and-continue policy. */
+  safety_harness?: {
+    enabled?: boolean;
+    evidence_collection?: boolean;
+    file_change_validation?: boolean;
+    evidence_cross_reference?: boolean;
+    destructive_command_warnings?: boolean;
+    content_validation?: boolean;
+    checkpoints?: boolean;
+    auto_rollback?: boolean;
+    timeout_scale_cap?: number;
+  };
 }
 
 export interface LoadedGSDPreferences {

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;
+}

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 };
+}

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 ?? "");
+}

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;
+}

src/resources/extensions/gsd/safety/file-change-validator.ts

@@ -0,0 +1,108 @@
+/**
+ * Post-unit file change validator for auto-mode safety harness.
+ * Compares actual git diff against the task plan's expected output files.
+ *
+ * Uses tasks.expected_output (DB column, populated from per-task ## Expected Output)
+ * and tasks.files (from slice PLAN.md - Files: subline) as the expected set.
+ * Compares against git diff HEAD~1 --name-only after auto-commit.
+ *
+ * Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
+ */
+
+import { execFileSync } from "node:child_process";
+import { logWarning } from "../workflow-logger.js";
+
+// ─── Types ──────────────────────────────────────────────────────────────────
+
+export interface FileViolation {
+  severity: "info" | "warning";
+  file: string;
+  reason: string;
+}
+
+export interface FileChangeAudit {
+  expectedFiles: string[];
+  actualFiles: string[];
+  unexpectedFiles: string[];
+  missingFiles: string[];
+  violations: FileViolation[];
+}
+
+// ─── Public API ─────────────────────────────────────────────────────────────
+
+/**
+ * Validate file changes after auto-commit for an execute-task unit.
+ * Returns null if task data is unavailable or DB is not loaded.
+ *
+ * @param basePath - Working directory (worktree or project root)
+ * @param expectedOutput - JSON array from tasks.expected_output DB column
+ * @param plannedFiles - JSON array from tasks.files DB column
+ */
+export function validateFileChanges(
+  basePath: string,
+  expectedOutput: string[],
+  plannedFiles: string[],
+): FileChangeAudit | null {
+  const allExpected = new Set([...expectedOutput, ...plannedFiles]);
+
+  // If no expected files were planned, skip validation
+  if (allExpected.size === 0) return null;
+
+  // Get actual changed files from last commit
+  const actualFiles = getChangedFilesFromLastCommit(basePath);
+  if (!actualFiles) return null;
+
+  // Filter out .gsd/ internal files — only validate project source files
+  const projectFiles = actualFiles.filter(f => !f.startsWith(".gsd/") && !f.startsWith(".gsd\\"));
+
+  // Normalize expected paths (strip leading ./ or /)
+  const normalizedExpected = new Set(
+    [...allExpected].map(f => f.replace(/^\.\//, "").replace(/^\//, "")),
+  );
+
+  // Compute symmetric difference
+  const unexpectedFiles = projectFiles.filter(f => !normalizedExpected.has(f));
+  const missingFiles = [...normalizedExpected].filter(f => !projectFiles.includes(f));
+
+  const violations: FileViolation[] = [];
+
+  for (const f of unexpectedFiles) {
+    violations.push({
+      severity: "warning",
+      file: f,
+      reason: "Modified but not in task plan's expected output",
+    });
+  }
+
+  for (const f of missingFiles) {
+    violations.push({
+      severity: "info",
+      file: f,
+      reason: "Listed in task plan but not modified",
+    });
+  }
+
+  return {
+    expectedFiles: [...normalizedExpected],
+    actualFiles: projectFiles,
+    unexpectedFiles,
+    missingFiles,
+    violations,
+  };
+}
+
+// ─── Internals ──────────────────────────────────────────────────────────────
+
+function getChangedFilesFromLastCommit(basePath: string): string[] | null {
+  try {
+    const result = execFileSync(
+      "git",
+      ["diff", "--name-only", "HEAD~1", "HEAD"],
+      { cwd: basePath, stdio: ["ignore", "pipe", "pipe"], encoding: "utf-8" },
+    ).trim();
+    return result ? result.split("\n").filter(Boolean) : [];
+  } catch (e) {
+    logWarning("safety", `git diff failed in file-change-validator: ${(e as Error).message}`);
+    return null;
+  }
+}

src/resources/extensions/gsd/safety/git-checkpoint.ts

@@ -0,0 +1,106 @@
+/**
+ * Pre-unit git checkpoint and rollback for auto-mode safety harness.
+ * Uses the existing refs/gsd/ namespace (already pruned by doctor).
+ *
+ * Creates a lightweight ref at HEAD before unit execution. On failure,
+ * the ref can be used to rollback the branch to the pre-unit state.
+ *
+ * Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
+ */
+
+import { execFileSync } from "node:child_process";
+import { logWarning } from "../workflow-logger.js";
+
+// ─── Constants ──────────────────────────────────────────────────────────────
+
+const CHECKPOINT_PREFIX = "refs/gsd/checkpoints/";
+
+// ─── Public API ─────────────────────────────────────────────────────────────
+
+/**
+ * Create a checkpoint ref at the current HEAD for the given unit.
+ * Returns the SHA of HEAD, or null if the operation fails.
+ */
+export function createCheckpoint(basePath: string, unitId: string): string | null {
+  try {
+    const sha = execFileSync("git", ["rev-parse", "HEAD"], {
+      cwd: basePath,
+      stdio: ["ignore", "pipe", "pipe"],
+      encoding: "utf-8",
+    }).trim();
+
+    if (!sha || sha.length < 7) return null;
+
+    // Sanitize unitId for use in ref path (replace / with -)
+    const safeUnitId = unitId.replace(/\//g, "-");
+
+    execFileSync("git", ["update-ref", `${CHECKPOINT_PREFIX}${safeUnitId}`, sha], {
+      cwd: basePath,
+      stdio: ["ignore", "pipe", "pipe"],
+    });
+
+    return sha;
+  } catch (e) {
+    logWarning("safety", `checkpoint creation failed: ${(e as Error).message}`);
+    return null;
+  }
+}
+
+/**
+ * Rollback the current branch to a checkpoint SHA.
+ * Returns true on success, false on failure.
+ *
+ * WARNING: This is a destructive operation — it discards all changes
+ * since the checkpoint. Only call when the user has opted in via
+ * safety_harness.auto_rollback or an explicit manual trigger.
+ */
+export function rollbackToCheckpoint(
+  basePath: string,
+  unitId: string,
+  sha: string,
+): boolean {
+  try {
+    // Get current branch name
+    const branch = execFileSync("git", ["rev-parse", "--abbrev-ref", "HEAD"], {
+      cwd: basePath,
+      stdio: ["ignore", "pipe", "pipe"],
+      encoding: "utf-8",
+    }).trim();
+
+    if (!branch || branch === "HEAD") {
+      logWarning("safety", "rollback: detached HEAD state, cannot rollback");
+      return false;
+    }
+
+    // Reset branch pointer and working tree to checkpoint SHA in one step.
+    // Using `git reset --hard <sha>` works on the currently checked-out branch
+    // (unlike `git branch -f` which is rejected for checked-out branches).
+    execFileSync("git", ["reset", "--hard", sha], {
+      cwd: basePath,
+      stdio: ["ignore", "pipe", "pipe"],
+    });
+
+    // Cleanup checkpoint ref
+    cleanupCheckpoint(basePath, unitId);
+
+    return true;
+  } catch (e) {
+    logWarning("safety", `rollback failed: ${(e as Error).message}`);
+    return false;
+  }
+}
+
+/**
+ * Remove a checkpoint ref after successful unit completion.
+ */
+export function cleanupCheckpoint(basePath: string, unitId: string): void {
+  try {
+    const safeUnitId = unitId.replace(/\//g, "-");
+    execFileSync("git", ["update-ref", "-d", `${CHECKPOINT_PREFIX}${safeUnitId}`], {
+      cwd: basePath,
+      stdio: ["ignore", "pipe", "pipe"],
+    });
+  } catch {
+    // Non-fatal — ref may already have been cleaned up
+  }
+}

src/resources/extensions/gsd/safety/safety-harness.ts

@@ -0,0 +1,105 @@
+/**
+ * Safety Harness — central module for LLM damage control during auto-mode.
+ * Provides types, preference resolution, and orchestration for all safety components.
+ *
+ * Components:
+ * - evidence-collector.ts: Real-time tool call tracking
+ * - destructive-guard.ts: Bash command classification
+ * - file-change-validator.ts: Post-unit git diff vs plan
+ * - evidence-cross-ref.ts: Claimed vs actual verification evidence
+ * - git-checkpoint.ts: Pre-unit checkpoints + rollback
+ * - content-validator.ts: Output quality validation
+ *
+ * Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
+ */
+
+// ─── Types ──────────────────────────────────────────────────────────────────
+
+export interface SafetyHarnessConfig {
+  enabled: boolean;
+  evidence_collection: boolean;
+  file_change_validation: boolean;
+  evidence_cross_reference: boolean;
+  destructive_command_warnings: boolean;
+  content_validation: boolean;
+  checkpoints: boolean;
+  auto_rollback: boolean;
+  timeout_scale_cap: number;
+}
+
+// ─── Defaults ───────────────────────────────────────────────────────────────
+
+const DEFAULTS: SafetyHarnessConfig = {
+  enabled: true,
+  evidence_collection: true,
+  file_change_validation: true,
+  evidence_cross_reference: true,
+  destructive_command_warnings: true,
+  content_validation: true,
+  checkpoints: true,
+  auto_rollback: false,
+  timeout_scale_cap: 6,
+};
+
+// ─── Public API ─────────────────────────────────────────────────────────────
+
+/**
+ * Resolve safety harness configuration from raw preferences.
+ * Missing fields fall back to defaults.
+ */
+export function resolveSafetyHarnessConfig(
+  raw: Record<string, unknown> | undefined,
+): SafetyHarnessConfig {
+  if (!raw) return { ...DEFAULTS };
+
+  return {
+    enabled: typeof raw.enabled === "boolean" ? raw.enabled : DEFAULTS.enabled,
+    evidence_collection: typeof raw.evidence_collection === "boolean" ? raw.evidence_collection : DEFAULTS.evidence_collection,
+    file_change_validation: typeof raw.file_change_validation === "boolean" ? raw.file_change_validation : DEFAULTS.file_change_validation,
+    evidence_cross_reference: typeof raw.evidence_cross_reference === "boolean" ? raw.evidence_cross_reference : DEFAULTS.evidence_cross_reference,
+    destructive_command_warnings: typeof raw.destructive_command_warnings === "boolean" ? raw.destructive_command_warnings : DEFAULTS.destructive_command_warnings,
+    content_validation: typeof raw.content_validation === "boolean" ? raw.content_validation : DEFAULTS.content_validation,
+    checkpoints: typeof raw.checkpoints === "boolean" ? raw.checkpoints : DEFAULTS.checkpoints,
+    auto_rollback: typeof raw.auto_rollback === "boolean" ? raw.auto_rollback : DEFAULTS.auto_rollback,
+    timeout_scale_cap: typeof raw.timeout_scale_cap === "number" ? raw.timeout_scale_cap : DEFAULTS.timeout_scale_cap,
+  };
+}
+
+/**
+ * Check if the safety harness is enabled.
+ * Used as a fast gate at hook registration and phase integration points.
+ */
+export function isHarnessEnabled(
+  raw: Record<string, unknown> | undefined,
+): boolean {
+  if (!raw) return DEFAULTS.enabled;
+  if (typeof raw.enabled === "boolean") return raw.enabled;
+  return DEFAULTS.enabled;
+}
+
+// ─── Re-exports ─────────────────────────────────────────────────────────────
+
+export {
+  resetEvidence,
+  getEvidence,
+  getBashEvidence,
+  getFilePaths,
+  recordToolCall,
+  recordToolResult,
+} from "./evidence-collector.js";
+
+export type { EvidenceEntry, BashEvidence, FileWriteEvidence, FileEditEvidence } from "./evidence-collector.js";
+
+export { classifyCommand } from "./destructive-guard.js";
+export type { CommandClassification } from "./destructive-guard.js";
+
+export { validateFileChanges } from "./file-change-validator.js";
+export type { FileChangeAudit, FileViolation } from "./file-change-validator.js";
+
+export { crossReferenceEvidence } from "./evidence-cross-ref.js";
+export type { ClaimedEvidence, EvidenceMismatch } from "./evidence-cross-ref.js";
+
+export { createCheckpoint, rollbackToCheckpoint, cleanupCheckpoint } from "./git-checkpoint.js";
+
+export { validateContent } from "./content-validator.js";
+export type { ContentViolation } from "./content-validator.js";

src/resources/extensions/gsd/tests/git-checkpoint.test.ts

@@ -0,0 +1,94 @@
+// GSD2 — Regression tests for git-checkpoint rollback (#3576)
+// Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
+
+import { describe, it } from "node:test";
+import assert from "node:assert/strict";
+import { mkdtempSync, writeFileSync, rmSync } from "node:fs";
+import { join } from "node:path";
+import { tmpdir } from "node:os";
+import { execFileSync } from "node:child_process";
+import { createCheckpoint, rollbackToCheckpoint, cleanupCheckpoint } from "../safety/git-checkpoint.js";
+
+function git(args: string[], cwd: string): string {
+  return execFileSync("git", args, { cwd, stdio: ["ignore", "pipe", "pipe"], encoding: "utf-8" }).trim();
+}
+
+function createTempRepo(): string {
+  const dir = mkdtempSync(join(tmpdir(), "ckpt-test-"));
+  git(["init"], dir);
+  git(["config", "user.email", "test@test.com"], dir);
+  git(["config", "user.name", "Test"], dir);
+  writeFileSync(join(dir, "file.txt"), "initial\n");
+  git(["add", "."], dir);
+  git(["commit", "-m", "init"], dir);
+  git(["branch", "-M", "main"], dir);
+  return dir;
+}
+
+describe("git-checkpoint rollback", () => {
+  it("rolls back to checkpoint on checked-out branch", (t) => {
+    const repo = createTempRepo();
+    t.after(() => rmSync(repo, { recursive: true, force: true }));
+
+    // Create checkpoint at initial commit
+    const sha = createCheckpoint(repo, "unit-1");
+    assert.ok(sha, "checkpoint should return a SHA");
+
+    // Make a second commit
+    writeFileSync(join(repo, "file.txt"), "modified\n");
+    git(["add", "."], repo);
+    git(["commit", "-m", "second"], repo);
+
+    const headBefore = git(["rev-parse", "HEAD"], repo);
+    assert.notEqual(headBefore, sha, "HEAD should have advanced");
+
+    // Rollback — this must work on the checked-out branch
+    const result = rollbackToCheckpoint(repo, "unit-1", sha);
+    assert.equal(result, true, "rollback should succeed");
+
+    const headAfter = git(["rev-parse", "HEAD"], repo);
+    assert.equal(headAfter, sha, "HEAD should match checkpoint SHA after rollback");
+  });
+
+  it("returns false on detached HEAD", (t) => {
+    const repo = createTempRepo();
+    t.after(() => rmSync(repo, { recursive: true, force: true }));
+
+    const sha = git(["rev-parse", "HEAD"], repo);
+    git(["checkout", "--detach", sha], repo);
+
+    const result = rollbackToCheckpoint(repo, "unit-2", sha);
+    assert.equal(result, false, "rollback should fail on detached HEAD");
+  });
+
+  it("cleans up checkpoint ref after rollback", (t) => {
+    const repo = createTempRepo();
+    t.after(() => rmSync(repo, { recursive: true, force: true }));
+
+    const sha = createCheckpoint(repo, "unit-3");
+    assert.ok(sha);
+
+    // Ref should exist
+    const refBefore = git(["for-each-ref", "refs/gsd/checkpoints/unit-3", "--format=%(objectname)"], repo);
+    assert.equal(refBefore, sha);
+
+    rollbackToCheckpoint(repo, "unit-3", sha);
+
+    // Ref should be cleaned up
+    const refAfter = git(["for-each-ref", "refs/gsd/checkpoints/unit-3", "--format=%(objectname)"], repo);
+    assert.equal(refAfter, "", "checkpoint ref should be removed after rollback");
+  });
+
+  it("cleanupCheckpoint removes the ref without error", (t) => {
+    const repo = createTempRepo();
+    t.after(() => rmSync(repo, { recursive: true, force: true }));
+
+    const sha = createCheckpoint(repo, "unit-4");
+    assert.ok(sha);
+
+    cleanupCheckpoint(repo, "unit-4");
+
+    const ref = git(["for-each-ref", "refs/gsd/checkpoints/unit-4", "--format=%(objectname)"], repo);
+    assert.equal(ref, "", "ref should be gone");
+  });
+});

src/resources/extensions/gsd/workflow-logger.ts

@@ -48,7 +48,8 @@ export type LogComponent =
   | "bootstrap"     // Extension bootstrap (system-context, agent-end)
   | "guided"        // Guided flow (discuss, plan wizards)
   | "registry"      // Rule registry hook state
-  | "renderer";     // Markdown renderer and projections
+  | "renderer"      // Markdown renderer and projections
+  | "safety";       // LLM safety harness
 
 export interface LogEntry {
   ts: string;

src/resources/extensions/ollama/ollama-chat-provider.ts

@@ -149,7 +149,7 @@ export function streamOllamaChat(
 				// Handle text content — process independently of tool_calls
 				// (a chunk may contain both content and tool_calls)
 				const content = chunk.message?.content ?? "";
-				if (content && !chunk.done) {
+				if (content) {
 					if (thinkParser) {
 						processChunks(thinkParser.push(content));
 					} else {

src/resources/extensions/ollama/tests/ollama-chat-provider-stream.test.ts

@@ -0,0 +1,82 @@
+// GSD2 — Regression test: Ollama streaming must not drop content on done:true chunks (#3576)
+// Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
+
+import { describe, it } from "node:test";
+import assert from "node:assert/strict";
+
+/**
+ * This test validates the streaming logic pattern used in ollama-chat-provider.ts.
+ * The bug: content on the terminal done:true chunk was silently dropped because
+ * the stream loop only emitted content when `!chunk.done`.
+ *
+ * The fix: process chunk.message.content regardless of chunk.done, then handle
+ * done metadata. This test exercises that logic path with a simulated chunk stream.
+ */
+
+interface OllamaChunk {
+  done: boolean;
+  done_reason?: string;
+  message?: { content?: string; tool_calls?: unknown[] };
+  prompt_eval_count?: number;
+  eval_count?: number;
+}
+
+function simulateStreamLoop(chunks: OllamaChunk[]): string {
+  let output = "";
+
+  for (const chunk of chunks) {
+    // This mirrors the fixed logic in ollama-chat-provider.ts
+    const content = chunk.message?.content ?? "";
+    if (content) {
+      output += content;
+    }
+
+    if (chunk.done) {
+      break;
+    }
+  }
+
+  return output;
+}
+
+describe("Ollama stream terminal chunk handling", () => {
+  it("captures content from done:true chunk", () => {
+    const chunks: OllamaChunk[] = [
+      { done: false, message: { content: "Hello " } },
+      { done: false, message: { content: "world" } },
+      { done: true, done_reason: "stop", message: { content: "!" } },
+    ];
+
+    const result = simulateStreamLoop(chunks);
+    assert.equal(result, "Hello world!", "trailing content on done chunk must not be dropped");
+  });
+
+  it("works when done chunk has no content", () => {
+    const chunks: OllamaChunk[] = [
+      { done: false, message: { content: "Hello" } },
+      { done: true, done_reason: "stop", message: {} },
+    ];
+
+    const result = simulateStreamLoop(chunks);
+    assert.equal(result, "Hello");
+  });
+
+  it("works when done chunk has empty string content", () => {
+    const chunks: OllamaChunk[] = [
+      { done: false, message: { content: "data" } },
+      { done: true, done_reason: "stop", message: { content: "" } },
+    ];
+
+    const result = simulateStreamLoop(chunks);
+    assert.equal(result, "data");
+  });
+
+  it("handles single done chunk with content", () => {
+    const chunks: OllamaChunk[] = [
+      { done: true, done_reason: "stop", message: { content: "one-shot" } },
+    ];
+
+    const result = simulateStreamLoop(chunks);
+    assert.equal(result, "one-shot", "single done chunk with content should work");
+  });
+});