@opengsd/gsd-pi 1.2.0-dev.b1abb545 → 1.2.0-dev.d6c5343c

package/src/resources/extensions/gsd/prompts/run-uat.md

@@ -27,21 +27,23 @@ You are the UAT runner. Execute every check defined in `{{uatPath}}` as deeply a
 ### Automation rules by mode
 
 - `artifact-driven` — verify with shell commands, scripts, file reads, and artifact structure checks.
-- `browser-executable` — use browser tools to navigate to the target URL and verify expected behavior. Prefer direct `browser_*` tools when available. Capture screenshots as evidence. Record pass/fail with specific assertions.
-- `runtime-executable` — execute the specified command or script. Capture stdout/stderr as evidence. Record pass/fail based on exit code and output.
+- `browser-executable` — use browser tools to navigate to the target URL and verify expected behavior. Prefer direct `browser_*` tools when available. Capture screenshots as evidence. Record pass/fail with specific assertions. When the UAT **Evidence** section names a self-contained runtime harness (for example `npm run test:uat` or `node tests/browser/search-uat.mjs`), treat the effective mode as **runtime-executable** instead: run that command only and do not start a separate server or `browser_navigate` to a hardcoded localhost port.
+- `runtime-executable` — execute the specified command or script. Capture stdout/stderr as evidence. Record pass/fail based on exit code and output. When the verification script starts its own dev/static server (for example `node tests/browser/search-uat.mjs`), do **not** start a separate server with `uat-service-start` — the script owns server lifecycle and binds an ephemeral port. If a server is already running, pass its URL via `UAT_BASE_URL` or `PORT` in the `gsd_uat_exec` environment instead of hardcoding a fixed port like 4173.
 - `live-runtime` — exercise the real runtime path. Start or connect to the app/service if needed, use browser/runtime/network checks, and verify observable behavior.
 - `mixed` — run all automatable artifact-driven and live-runtime checks. Separate any remaining human-only checks explicitly.
 - `human-experience` — automate setup, preconditions, screenshots, logs, and objective checks, but do **not** invent subjective PASS results. Mark taste-based, experiential, or purely human-judgment checks as `NEEDS-HUMAN`. Use an overall verdict of `PASS` when all automatable checks succeed (even if human-only checks remain as `NEEDS-HUMAN`). Use `PARTIAL` only when automatable checks themselves were inconclusive.
 
 ### Evidence tools
 
-Choose the lightest tool that proves the check honestly:
+The **Tool Surface** block prepended above lists unavailable tools for this unit. In short:
 
 - Run automated checks with `gsd_uat_exec`
   - Use `uat-artifact-check` as `intent` for static file, grep, structure, or artifact checks.
   - Use `uat-runtime-check` as `intent` for executing tests, scripts, or runtime assertions.
   - Use `uat-browser-check` as `intent` for browser interaction or screenshot-backed UI checks.
-  - Use `uat-service-start` as `intent` only when starting or connecting to an app/service.
+  - Use `uat-service-start` as `intent` only when starting or connecting to an app/service that the UAT checks will not start themselves.
+  - Do not start a dev/static server separately when a runtime test script owns server lifecycle — run the script directly instead.
+  - When you do start a service, capture the URL from its stdout (for example `Ready at http://127.0.0.1:PORT`) and pass it to downstream checks via `UAT_BASE_URL` rather than assuming a fixed port.
   - Use `uat-log-inspection` as `intent` for checking logs or captured output files.
   - The result-table evidence mode is separate; do not use `artifact`, `runtime`, or `human-follow-up` as `intent`.
 - Run `grep` / `rg` checks against files

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

@@ -110,9 +110,12 @@ Templates are in `{{templatesDir}}`.
 
 **External facts:** Use `search-the-web` + `fetch_page`, or `search_and_read`; use `freshness` for recency. Never state current facts from training data without verification.
 
-**Background processes:** Use `bg_shell` `start` + `wait_for_ready` for servers/watchers/daemons. Never use `bash` with `&` or `nohup`; inherited stdout can hang. Never poll with `sleep`; use `wait_for_ready`. For status use `digest`, `highlights` for significant lines, and `output` only when debugging.
+**Choosing a shell tool — decide by how the command runs, not how long it takes:**
 
-**One-shot commands:** Use `async_bash` for builds, tests, and installs. Results are pushed on exit; use `await_job` only to block on a specific job.
+- **Runs and exits (a batch command):** anything that does work and finishes — `terraform apply`, DB migrations, builds, tests, installs, long scripts. Use synchronous `bash` when you want to block and read the result now (uncapped; in auto-mode genuine hangs are caught by the stalled-tool watchdog, interactively they end on human ESC). Use `async_bash` when you want it non-blocking — it returns a job ID immediately and you `await_job` later. **Never** put a run-to-completion command under `bg_shell` `wait_for_ready`: it exits instead of staying alive, so readiness never trips and a clean exit looks like a failure.
+- **Stays alive and you interact with it over time:** servers, watchers, daemons, REPLs. Use `bg_shell` `start`, then `wait_for_ready` (block until it listens/prints its ready pattern), `output`/`digest`/`highlights` to inspect, `send`/`send_and_wait` to drive it, and `kill`/`restart` to manage it. Never use `bash` with `&` or `nohup` (inherited stdout can hang); never poll with `sleep` (use `wait_for_ready`).
+
+Quick rule of thumb: if the command would exit on its own, it's `bash`/`async_bash`; if it would keep running until you stop it, it's `bg_shell`.
 
 **Stale job hygiene:** After editing source to fix a failure, `cancel_job` every in-flight `async_bash` job before rerunning. Changed inputs make in-flight outputs untrusted.
 

package/src/resources/extensions/gsd/provider-error-guidance.ts

@@ -1,7 +1,10 @@
 /**
  * Actionable remediation hints when auto-mode pauses on provider/model errors.
+ * Render with formatGuidance from the Guidance module.
  */
 
+import type { Guidance } from "./guidance.js";
+
 export interface ProviderErrorGuidanceInput {
   errorMsg: string;
   provider?: string;
@@ -12,10 +15,7 @@ export interface ProviderErrorGuidanceInput {
   hasConfiguredFallbacks?: boolean;
 }
 
-export interface ProviderErrorGuidance {
-  summary: string;
-  steps: string[];
-}
+export type ProviderErrorGuidance = Guidance;
 
 /** Map auto unit types to the `models:` key in PREFERENCES.md. */
 export function unitTypeToPrefsPhaseKey(unitType: string | undefined): string | undefined {
@@ -129,8 +129,3 @@ export function resolveProviderErrorGuidance(input: ProviderErrorGuidanceInput):
   return { summary, steps };
 }
 
-/** Flatten guidance into a pause banner / notification string. */
-export function formatProviderErrorGuidance(guidance: ProviderErrorGuidance): string {
-  const numbered = guidance.steps.map((step, index) => `${index + 1}. ${step}`).join("\n");
-  return `${guidance.summary}\n\n${numbered}`;
-}

package/src/resources/extensions/gsd/provider-switch-observer.ts

@@ -128,7 +128,7 @@ function emitAudit(report: ProviderSwitchReport): void {
 
 function emitNotification(report: ProviderSwitchReport): void {
   try {
-    appendNotification(summarize(report), "warning", "workflow-logger");
+    appendNotification(summarize(report), "warning", "workflow-logger", { kind: "provider-switch" });
   } catch {
     // Notification persistence is best-effort.
   }

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

@@ -0,0 +1,122 @@
+// Project/App: gsd-pi
+// File Purpose: Publication module — push and draft-PR creation for a merged milestone (ADR-034).
+//
+// Publication is distinct from the merge verb: merging a milestone branch is a
+// Worktree Lifecycle concern (worktree, branch, lease, stash discipline);
+// publication needs only the resulting commit, a remote, and preferences.
+// Extracted from the tail of `mergeMilestoneToMain` (auto-worktree.ts) so it is
+// testable without a merge fixture and reachable from any closeout path.
+//
+// Publication failure is non-fatal to a completed local merge: every failure is
+// logged and reported in the result, never thrown.
+
+import { execFileSync } from "node:child_process";
+
+import {
+  buildPullRequestEvidence,
+  createDraftPullRequestFromEvidence,
+} from "./pull-request-process.js";
+import { logWarning } from "./workflow-logger.js";
+
+export interface PublicationPrefs {
+  autoPush: boolean;
+  autoPr: boolean;
+  remote?: string;
+  prTargetBranch?: string;
+}
+
+export interface PublicationRequest {
+  basePath: string;
+  milestoneId: string;
+  milestoneTitle: string;
+  /** The branch the milestone was merged to (push target for auto-push). */
+  integrationBranch: string;
+  /** The milestone branch (push + PR head for auto-PR). */
+  milestoneBranch: string;
+  /** Pre-rendered per-slice summary blocks for the PR body. */
+  sliceSummaries: readonly string[];
+  /** True when the merge found nothing to commit — publication is skipped. */
+  nothingToCommit: boolean;
+  prefs: PublicationPrefs;
+}
+
+export interface PublicationResult {
+  pushed: boolean;
+  prCreated: boolean;
+  prUrl?: string;
+}
+
+export function gitRemoteExists(basePath: string, remote: string): boolean {
+  try {
+    execFileSync("git", ["remote", "get-url", remote], {
+      cwd: basePath,
+      stdio: ["ignore", "pipe", "pipe"],
+      encoding: "utf-8",
+    });
+    return true;
+  } catch {
+    return false;
+  }
+}
+
+/**
+ * Publish a merged milestone per preferences: push the integration branch
+ * (auto-push, suppressed when auto-PR owns the remote interaction) and/or
+ * push the milestone branch and open a draft PR (auto-PR).
+ */
+export function publishMilestone(request: PublicationRequest): PublicationResult {
+  const result: PublicationResult = { pushed: false, prCreated: false };
+  if (request.nothingToCommit) return result;
+
+  const { basePath, prefs } = request;
+  const remote = prefs.remote ?? "origin";
+
+  if (prefs.autoPush && !prefs.autoPr && gitRemoteExists(basePath, remote)) {
+    try {
+      execFileSync("git", ["push", remote, request.integrationBranch], {
+        cwd: basePath,
+        stdio: ["ignore", "pipe", "pipe"],
+        encoding: "utf-8",
+      });
+      result.pushed = true;
+    } catch (err) {
+      // Push failure is non-fatal
+      logWarning("worktree", `git push failed: ${err instanceof Error ? err.message : String(err)}`);
+    }
+  }
+
+  // #2302: PR creation is not gated on pushed/auto-push.
+  if (prefs.autoPr && gitRemoteExists(basePath, remote)) {
+    try {
+      // Push the milestone branch to remote first
+      execFileSync("git", ["push", remote, request.milestoneBranch], {
+        cwd: basePath,
+        stdio: ["ignore", "pipe", "pipe"],
+        encoding: "utf-8",
+      });
+      const prEvidence = buildPullRequestEvidence({
+        milestoneId: request.milestoneId,
+        milestoneTitle: request.milestoneTitle,
+        changeType: "feat",
+        summaries: [...request.sliceSummaries],
+        testsRun: ["Auto-created after milestone merge. Run `npm run verify:merge` before marking this draft ready."],
+        rollbackNotes: ["Close the draft PR or revert the merge commit if review finds a behavior regression."],
+        how: "Generated by git.auto_pr after the milestone branch was pushed and merged locally.",
+      });
+      const prUrl = createDraftPullRequestFromEvidence(basePath, request.milestoneId, prEvidence, {
+        head: request.milestoneBranch,
+        base: prefs.prTargetBranch ?? request.integrationBranch,
+      });
+      if (!prUrl) {
+        throw new Error("gh pr create returned no URL");
+      }
+      result.prCreated = true;
+      result.prUrl = prUrl;
+    } catch (err) {
+      // PR creation failure is non-fatal — gh may not be installed or authenticated
+      logWarning("worktree", `PR creation failed: ${err instanceof Error ? err.message : String(err)}`);
+    }
+  }
+
+  return result;
+}

package/src/resources/extensions/gsd/reactive-graph.ts

@@ -12,7 +12,11 @@
 import type { TaskIO, DerivedTaskNode, ReactiveExecutionState } from "./types.js";
 import { loadFile, parseTaskPlanIO } from "./files.js";
 import { isDbAvailable, getSliceTasks } from "./gsd-db.js";
+// Explicit degraded-mode fallback ONLY (ADR-017): parsePlan is consulted when
+// the DB is unavailable or has no task rows for the slice; the DB path in
+// loadSliceTaskIO always wins when rows exist, and the fallback logs a warning.
 import { parsePlan } from "./parsers-legacy.js";
+import { logWarning } from "./workflow-logger.js";
 import { resolveTasksDir, resolveTaskFiles } from "./paths.js";
 import { join } from "node:path";
 import { loadJsonFileOrNull, saveJsonFile } from "./json-persistence.js";
@@ -224,7 +228,13 @@ export async function loadSliceTaskIO(
   } catch { /* fall through */ }
 
   if (!taskEntries) {
-    // File-based fallback: parse slice plan for task entries
+    // Degraded-mode fallback (ADR-017): no DB task rows for this slice —
+    // parse the PLAN.md projection instead. Warn so silent drift from the
+    // DB-authoritative path is visible in logs.
+    logWarning(
+      "projection",
+      `reactive-graph: DB unavailable or no task rows for ${mid}/${sid}; falling back to PLAN.md parsing`,
+    );
     const parsed = parsePlan(planContent);
     if (parsed.tasks.length > 0) {
       taskEntries = parsed.tasks.map(t => ({

package/src/resources/extensions/gsd/recovery-classification.ts

@@ -1,18 +1,23 @@
 // Project/App: gsd-pi
 // File Purpose: ADR-015 Recovery Classification module for runtime failure taxonomy.
 
+import { isToolUnavailableError } from "./auto-tool-tracking.js";
 import { classifyError, isTransient, type ErrorClass } from "./error-classifier.js";
+import { recoveryRemediation } from "./guidance.js";
 import { ReconciliationFailedError } from "./state-reconciliation.js";
+import { IllegalPhaseTransitionError } from "./state-transition-matrix.js";
 
 export type RecoveryFailureKind =
   | "tool-schema"
   | "tool-contract"
+  | "tool-unavailable"
   | "deterministic-policy"
   | "lifecycle-progression"
   | "stale-worker"
   | "worktree-invalid"
   | "verification-drift"
   | "reconciliation-drift"
+  | "illegal-transition"
   | "provider"
   | "runtime-unknown";
 
@@ -43,99 +48,53 @@ export function classifyFailure(input: RecoveryClassificationInput): RecoveryCla
   const failureKind =
     input.error instanceof ReconciliationFailedError
       ? "reconciliation-drift"
-      : input.failureKind ?? inferFailureKind(message);
+      : input.error instanceof IllegalPhaseTransitionError
+        ? "illegal-transition"
+        : input.failureKind ?? inferFailureKind(message);
 
-  switch (failureKind) {
-    case "tool-schema":
-      return {
-        failureKind,
-        action: "stop",
-        reason: `Tool schema failure${unitSuffix(input)}: ${message}`,
-        exitReason: "tool-schema",
-        remediation: "Fix the Unit Tool Contract or tool schema before retrying.",
-      };
-    case "tool-contract":
-      return {
-        failureKind,
-        action: "stop",
-        reason: `Tool Contract failure${unitSuffix(input)}: ${message}`,
-        exitReason: "tool-contract",
-        remediation: "Fix the Unit Tool Contract or prompt so the Unit is only asked to use tools owned by its phase.",
-      };
-    case "deterministic-policy":
-      return {
-        failureKind,
-        action: "stop",
-        reason: `Deterministic policy failure${unitSuffix(input)}: ${message}`,
-        exitReason: "deterministic-policy",
-        remediation: "Resolve the policy blocker; retrying the same Unit will repeat the failure.",
-      };
-    case "lifecycle-progression":
-      return {
-        failureKind,
-        action: "stop",
-        reason: `Lifecycle progression failure${unitSuffix(input)}: ${message}`,
-        exitReason: "lifecycle-progression",
-        remediation: "Route to the required owning Unit or restore the missing artifact before advancing lifecycle state.",
-      };
-    case "stale-worker":
-      return {
-        failureKind,
-        action: "stop",
-        reason: `Stale worker failure${unitSuffix(input)}: ${message}`,
-        exitReason: "stale-worker",
-        remediation: "Clear or reconcile the stale worker before dispatching another Unit.",
-      };
-    case "worktree-invalid":
-      return {
-        failureKind,
-        action: "stop",
-        reason: `Worktree invalid${unitSuffix(input)}: ${message}`,
-        exitReason: "worktree-invalid",
-        remediation: "Repair or recreate the milestone worktree before launching source-writing Units.",
-      };
-    case "verification-drift":
-      return {
-        failureKind,
-        action: "escalate",
-        reason: `Verification drift${unitSuffix(input)}: ${message}`,
-        exitReason: "verification-drift",
-        remediation: "Inspect the verification artifact and reconcile the state snapshot before resuming.",
-      };
-    case "reconciliation-drift":
-      return {
-        failureKind,
-        action: "escalate",
-        reason: `Reconciliation drift${unitSuffix(input)}: ${message}`,
-        exitReason: "reconciliation-drift",
-        remediation:
-          "Inspect the persistent or repair-failed drift kinds reported by the State Reconciliation Module before resuming.",
-      };
-    case "provider": {
-      const providerClass = classifyError(message, input.retryAfterMs);
-      return {
-        failureKind,
-        action: isTransient(providerClass) ? "retry" : "escalate",
-        reason: message,
-        exitReason: `provider-${providerClass.kind}`,
-        remediation: isTransient(providerClass)
-          ? "Retry after the provider/network condition clears."
-          : "Inspect provider credentials, model entitlement, or request shape.",
-        providerClass: providerClass.kind,
-      };
-    }
-    case "runtime-unknown":
-      return {
-        failureKind,
-        action: "escalate",
-        reason: message,
-        exitReason: "runtime-unknown",
-        remediation: "Inspect the runtime error and add a dedicated classification if it is repeatable.",
-      };
+  if (failureKind === "provider") {
+    const providerClass = classifyError(message, input.retryAfterMs);
+    const transient = isTransient(providerClass);
+    return {
+      failureKind,
+      action: transient ? "retry" : "escalate",
+      reason: message,
+      exitReason: `provider-${providerClass.kind}`,
+      remediation: recoveryRemediation(transient ? "provider-transient" : "provider-permanent"),
+      providerClass: providerClass.kind,
+    };
   }
+
+  const { action, label } = FAILURE_TAXONOMY[failureKind];
+  return {
+    failureKind,
+    action,
+    reason: label ? `${label}${unitSuffix(input)}: ${message}` : message,
+    exitReason: failureKind,
+    remediation: recoveryRemediation(failureKind),
+  };
 }
 
+/** Per-kind action and reason label. Remediation lives in the Guidance module. */
+const FAILURE_TAXONOMY: Record<
+  Exclude<RecoveryFailureKind, "provider">,
+  { action: RecoveryAction; label: string | null }
+> = {
+  "tool-schema": { action: "stop", label: "Tool schema failure" },
+  "tool-contract": { action: "stop", label: "Tool Contract failure" },
+  "tool-unavailable": { action: "retry", label: "Tool unavailable" },
+  "deterministic-policy": { action: "stop", label: "Deterministic policy failure" },
+  "lifecycle-progression": { action: "stop", label: "Lifecycle progression failure" },
+  "stale-worker": { action: "stop", label: "Stale worker failure" },
+  "worktree-invalid": { action: "stop", label: "Worktree invalid" },
+  "verification-drift": { action: "escalate", label: "Verification drift" },
+  "reconciliation-drift": { action: "escalate", label: "Reconciliation drift" },
+  "illegal-transition": { action: "escalate", label: "Illegal phase transition" },
+  "runtime-unknown": { action: "escalate", label: null },
+};
+
 function inferFailureKind(message: string): RecoveryFailureKind {
+  if (isToolUnavailableError(message)) return "tool-unavailable";
   if (/tool contract|auto-unit tool scope|phase-boundary gate|not permitted.*own/i.test(message)) return "tool-contract";
   if (/lifecycle progression|required artifact|missing .*assessment|missing .*closeout|cannot legally (?:advance|progress)/i.test(message)) return "lifecycle-progression";
   if (/schema|invalid.*tool|tool.*invalid|enum/i.test(message)) return "tool-schema";

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

@@ -0,0 +1,134 @@
+/**
+ * One-shot confirmation token for destructive bash commands.
+ *
+ * The destructive-command guard hard-blocks classified commands (force push,
+ * rm -rf, SQL drop, etc.) in all modes. The block instructs the model to
+ * confirm via ask_user_questions and re-issue the command. This module is the
+ * missing escape hatch: it records the user's confirmation and lets the exact
+ * confirmed command through exactly once.
+ *
+ * Design constraints:
+ *  - In-memory only, never persisted. A confirmation token written to disk
+ *    could silently auto-approve a destructive command in a later session —
+ *    confirmation must be re-obtained every process lifetime.
+ *  - One-shot. Consuming a token clears it, so a second destructive command
+ *    (even an identical one) re-blocks and re-prompts.
+ *  - Command-bound. The token only matches the exact (normalized) command
+ *    string the user confirmed. A reworded command re-blocks, which is safe.
+ *  - Per basePath, so concurrent workspaces in one process never share tokens.
+ *
+ * Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
+ */
+
+import { resolve } from "node:path";
+
+/**
+ * Question-id substring that marks an ask_user_questions call as a
+ * destructive-command confirmation. The tool_result handler promotes the
+ * pending command to a confirmed token when an affirmative answer arrives for
+ * a question whose id contains this marker.
+ */
+export const DESTRUCTIVE_CONFIRM_GATE_MARKER = "destructive_confirm";
+
+interface DestructiveConfirmationState {
+  /** Command awaiting confirmation, captured when the guard blocked it. */
+  pendingCommand: string | null;
+  /** Confirmed command cleared on first matching consume. */
+  confirmedCommand: string | null;
+}
+
+const statesByBasePath = new Map<string, DestructiveConfirmationState>();
+
+function stateKey(basePath: string): string {
+  return resolve(basePath);
+}
+
+function getState(basePath: string): DestructiveConfirmationState {
+  const key = stateKey(basePath);
+  let state = statesByBasePath.get(key);
+  if (!state) {
+    state = { pendingCommand: null, confirmedCommand: null };
+    statesByBasePath.set(key, state);
+  }
+  return state;
+}
+
+/**
+ * Normalize a command for stable matching across block → confirm → retry.
+ * Trims surrounding whitespace and collapses internal runs of whitespace so
+ * cosmetic reformatting of the same command still matches the token.
+ */
+export function normalizeDestructiveCommand(command: string): string {
+  return command.replace(/\s+/g, " ").trim();
+}
+
+/**
+ * Whether an ask_user_questions question id is a destructive-confirm gate.
+ */
+export function isDestructiveConfirmGateId(questionId: unknown): boolean {
+  return typeof questionId === "string" && questionId.includes(DESTRUCTIVE_CONFIRM_GATE_MARKER);
+}
+
+/**
+ * Record that a destructive command was blocked and is awaiting confirmation.
+ * Called by the guard at block time. Overwrites any prior pending command —
+ * only the most recently blocked command can be confirmed.
+ */
+export function requestDestructiveConfirmation(
+  command: string,
+  basePath: string = process.cwd(),
+): void {
+  const state = getState(basePath);
+  state.pendingCommand = normalizeDestructiveCommand(command);
+  // A fresh request invalidates any stale confirmed token for a different
+  // command so confirmation cannot leak across distinct destructive actions.
+  state.confirmedCommand = null;
+}
+
+/**
+ * Promote the pending command to a confirmed, one-shot token. Called by the
+ * tool_result handler when the user gives an affirmative answer to a
+ * destructive-confirm gate. Returns the confirmed command, or null if there
+ * was nothing pending (e.g. confirmation arrived without a preceding block).
+ */
+export function confirmDestructiveCommand(
+  basePath: string = process.cwd(),
+): string | null {
+  const state = getState(basePath);
+  if (!state.pendingCommand) return null;
+  state.confirmedCommand = state.pendingCommand;
+  state.pendingCommand = null;
+  return state.confirmedCommand;
+}
+
+/**
+ * Check whether the given command has been confirmed, consuming the token if
+ * so. Returns true exactly once per confirmation; subsequent calls (or a
+ * non-matching command) return false. Called by the guard before blocking.
+ */
+export function consumeDestructiveConfirmation(
+  command: string,
+  basePath: string = process.cwd(),
+): boolean {
+  const state = getState(basePath);
+  if (!state.confirmedCommand) return false;
+  if (state.confirmedCommand !== normalizeDestructiveCommand(command)) return false;
+  state.confirmedCommand = null;
+  return true;
+}
+
+/**
+ * Inspect the pending command without consuming it (diagnostics/tests).
+ */
+export function peekPendingDestructiveCommand(
+  basePath: string = process.cwd(),
+): string | null {
+  return getState(basePath).pendingCommand;
+}
+
+/**
+ * Clear all destructive-confirmation state for a basePath (tests / flow reset).
+ */
+export function resetDestructiveConfirmation(basePath: string = process.cwd()): void {
+  statesByBasePath.delete(stateKey(basePath));
+}

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

@@ -57,9 +57,10 @@ const EXECUTION_TOOL_NAMES = new Set([
   "functions.exec_command",
   "gsd_exec",
   "gsd_exec_search",
+  "gsd_uat_exec",
   "powershell",
 ]);
-const MCP_EXECUTION_TOOL_RE = /^mcp__.+__gsd_exec(?:_search)?$/;
+const MCP_EXECUTION_TOOL_RE = /^mcp__.+__gsd_(?:uat_)?exec(?:_search)?$/;
 
 // ─── Module State ───────────────────────────────────────────────────────────
 
@@ -206,11 +207,17 @@ export function clearEvidenceFromDisk(
  * Exit codes and output are filled in by recordToolResult after execution.
  */
 export function recordToolCall(toolCallId: string, toolName: string, input: Record<string, unknown>): void {
+  // Idempotent by toolCallId: native tools reach this via both
+  // tool_execution_start and tool_call; external (pre-executed) tools only
+  // via tool_execution_start. First recording wins.
+  if (unitEvidence.some(e => e.toolCallId === toolCallId)) return;
   if (isExecutionToolName(toolName)) {
     unitEvidence.push({
       kind: "bash",
       toolCallId,
-      command: String(input.command ?? input.cmd ?? input.query ?? ""),
+      // gsd_exec / gsd_uat_exec carry the script body in `script` (or `code`);
+      // bash-style tools use `command`/`cmd`; gsd_exec_search uses `query`.
+      command: String(input.command ?? input.script ?? input.cmd ?? input.code ?? input.query ?? ""),
       exitCode: -1,
       outputSnippet: "",
       timestamp: Date.now(),
@@ -249,11 +256,36 @@ export function recordToolResult(
   if (entry.kind === "bash") {
     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);
+    entry.exitCode = resolveExitCode(text, isError);
   }
 }
 
+/**
+ * Resolve the exit code from a tool result's text. Handles the bash tool's
+ * prose marker, the gsd_exec / gsd_uat_exec JSON envelope (`"exit_code": N`),
+ * and a last-resort read of the run's persisted `.gsd/exec/<id>.meta.json`
+ * (covers truncated result text).
+ */
+function resolveExitCode(text: string, isError: boolean): number {
+  const proseMatch = text.match(/Command exited with code (\d+)/);
+  if (proseMatch) return Number(proseMatch[1]);
+
+  const jsonMatch = text.match(/"exit_code"\s*:\s*(-?\d+)/);
+  if (jsonMatch) return Number(jsonMatch[1]);
+
+  const metaMatch = text.match(/"meta_path"\s*:\s*"([^"]+)"/);
+  if (metaMatch) {
+    try {
+      const meta = JSON.parse(readFileSync(metaMatch[1], "utf-8")) as Record<string, unknown>;
+      if (typeof meta.exit_code === "number") return meta.exit_code;
+    } catch {
+      // Fall through to the isError heuristic
+    }
+  }
+
+  return isError ? 1 : 0;
+}
+
 // ─── Internals ──────────────────────────────────────────────────────────────
 
 function extractResultText(result: unknown): string {

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

@@ -121,9 +121,14 @@ function findMatches(
   const exact = bashCalls.filter(b => b.command.trim() === normalized);
   if (exact.length > 0) return exact;
 
-  // Substring match: claimed is contained in actual or actual in claimed
+  // Substring match: claimed is contained in actual or actual in claimed.
+  // A claimed verification command typically appears verbatim inside a
+  // larger gsd_exec script body (cd prefix, multi-line scripts), so
+  // script-containing-claim is the common direction. Blank-command entries
+  // must be excluded — `"x".includes("")` is true, so they'd match anything.
   const substring = bashCalls.filter(
-    b => b.command.includes(normalized) || normalized.includes(b.command),
+    b => b.command.trim().length > 0 &&
+      (b.command.includes(normalized) || normalized.includes(b.command)),
   );
   if (substring.length > 0) return substring;
 

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

@@ -39,6 +39,20 @@ export interface FileChangeAudit {
 
 // ─── Public API ─────────────────────────────────────────────────────────────
 
+/**
+ * Build the effective allowlist for a unit's file-change audit.
+ *
+ * When GSD manages .gitignore (manage_gitignore unset or true), ensureGitignore()
+ * appends baseline patterns at auto-start and the edit rides into the task's
+ * auto-commit — so .gitignore changes must not be attributed to the task.
+ */
+export function effectiveFileChangeAllowlist(
+  baseAllowlist: string[],
+  manageGitignore: boolean | undefined,
+): string[] {
+  return manageGitignore === false ? baseAllowlist : [...baseAllowlist, ".gitignore"];
+}
+
 /**
  * Validate file changes after auto-commit for an execute-task unit.
  * Returns null if task data is unavailable or DB is not loaded.