@opengsd/gsd-pi 1.2.0-dev.844675c9 → 1.2.0-dev.8e6112e9

package/dist/resources/extensions/gsd/doctor.js

@@ -2,11 +2,13 @@ import { existsSync, mkdirSync, lstatSync, readdirSync, readFileSync } from "nod
 import { join } from "node:path";
 import { loadFile, parseSummary, saveFile, parseTaskPlanMustHaves, countMustHavesMentionedInSummary } from "./files.js";
 import { parseRoadmap as parseLegacyRoadmap, parsePlan as parseLegacyPlan } from "./parsers-legacy.js";
-import { isDbAvailable, openDatabase, getMilestoneSlices, getSliceTasks } from "./gsd-db.js";
-import { resolveMilestoneFile, resolveMilestonePath, resolveSliceFile, resolveSlicePath, resolveTaskFile, resolveTasksDir, milestonesDir, gsdRoot, relMilestoneFile, relSliceFile, relTaskFile, relSlicePath, relGsdRootFile, resolveGsdRootFile, relMilestonePath, resolveGsdPathContract } from "./paths.js";
+import { isDbAvailable, getMilestoneSlices, getSliceTasks } from "./gsd-db.js";
+import { openExistingWorkflowDatabase } from "./db-workspace.js";
+import { resolveMilestoneFile, resolveMilestonePath, resolveSliceFile, resolveSlicePath, resolveTaskFile, resolveTasksDir, milestonesDir, gsdRoot, relMilestoneFile, relSliceFile, relTaskFile, relSlicePath, relGsdRootFile, resolveGsdRootFile, relMilestonePath } from "./paths.js";
 import { deriveState, isMilestoneComplete } from "./state.js";
 import { invalidateAllCaches } from "./cache.js";
 import { loadEffectiveGSDPreferences } from "./preferences.js";
+import { collectPreferenceDiagnostics, formatPreferenceDiagnosticDetail } from "./preferences-diagnostics.js";
 import { isClosedStatus } from "./status-guards.js";
 import { GLOBAL_STATE_CODES } from "./doctor-types.js";
 import { checkGitHealth, checkRuntimeHealth, checkGlobalHealth, checkEngineHealth } from "./doctor-checks.js";
@@ -286,13 +288,7 @@ export async function runGSDDoctor(basePath, options) {
     // CLI doctor can run before any tool handler has opened the DB. Runtime
     // health checks need the existing project DB to surface DB-backed crash
     // locks, paused sessions, and coordination rows.
-    const dbPath = resolveGsdPathContract(basePath).projectDb;
-    if (existsSync(dbPath)) {
-        try {
-            openDatabase(dbPath);
-        }
-        catch { /* surfaced later as db_unavailable */ }
-    }
+    openExistingWorkflowDatabase(basePath);
     // Issue codes that represent completion state transitions — creating summary
     // stubs, marking slices/milestones done in the roadmap. These belong to the
     // dispatch lifecycle (complete-slice, complete-milestone units), not to
@@ -307,6 +303,17 @@ export async function runGSDDoctor(basePath, options) {
         return true;
     };
     const prefs = loadEffectiveGSDPreferences(basePath);
+    for (const diagnostic of collectPreferenceDiagnostics(basePath)) {
+        issues.push({
+            severity: diagnostic.severity,
+            code: "invalid_preferences",
+            scope: "project",
+            unitId: "project",
+            message: `GSD preferences ${diagnostic.kind}: ${formatPreferenceDiagnosticDetail(diagnostic)}`,
+            file: diagnostic.path,
+            fixable: false,
+        });
+    }
     if (prefs) {
         const prefIssues = validatePreferenceShape(prefs.preferences);
         for (const issue of prefIssues) {

package/dist/resources/extensions/gsd/engine-hook-contract.js

@@ -0,0 +1,70 @@
+// Project/App: gsd-pi
+// File Purpose: Typed contract for which tool lifecycle hooks fire under which
+// engine, plus the single seam for tool-name normalization.
+//
+// ── Why this contract exists ────────────────────────────────────────────────
+// External engines (claude-code-cli) pre-execute tools inside the vendor CLI
+// and hand the agent loop a `toolCall.externalResult`. The loop short-circuits
+// before `beforeToolCall`/`afterToolCall` ever run for those calls
+// (packages/pi-agent-core/src/agent-loop.ts, prepareToolCall: the
+// externalResult branch returns an "immediate" outcome before the
+// config.beforeToolCall invocation, and "immediate" outcomes skip
+// finalizeExecutedToolCall where config.afterToolCall lives).
+// `beforeToolCall`/`afterToolCall` are exactly what the extension runner maps
+// to the `tool_call`/`tool_result` extension events
+// (packages/gsd-agent-core/src/session/agent-session-extensions.ts,
+// installAgentToolHooks).
+//
+// `tool_execution_start` is emitted unconditionally before prepareToolCall
+// and `tool_execution_end` after every finalized outcome — immediate or
+// executed — so those two are the ONLY tool lifecycle hooks that fire for
+// every tool call on every engine.
+//
+// Consequence: any enforcement attached only to `tool_call` (blocking) or
+// `tool_result` (rewriting) is silently dead under external engines. Safety
+// enforcement must ride `tool_execution_start`/`tool_execution_end`, or
+// mirror across both with toolCallId dedup. See the per-registration contract
+// comments in bootstrap/register-hooks.ts for how each registered hook
+// honors (or knowingly violates) this contract.
+import { stripMcpToolPrefix } from "./mcp-tool-name.js";
+import { canonicalWorkflowSurfaceToolName } from "./workflow-tool-surface.js";
+/**
+ * Tool lifecycle hooks that fire for EVERY tool call on EVERY engine,
+ * including external engines (claude-code-cli) that pre-execute tools.
+ * Attach safety-critical enforcement and evidence collection here.
+ */
+export const UNIVERSAL_TOOL_HOOKS = ["tool_execution_start", "tool_execution_end"];
+/**
+ * Tool lifecycle hooks that fire ONLY for natively executed tools. External
+ * engines pre-execute tools (externalResult), short-circuiting the agent
+ * loop's beforeToolCall/afterToolCall — so handlers on these events never run
+ * for those calls. Blocking guards attached only here are dead under external
+ * engines; they need a tool_execution_start mirror to be universal.
+ */
+export const NATIVE_ONLY_TOOL_HOOKS = ["tool_call", "tool_result"];
+// Non-tool lifecycle events (session_start, agent_end, message_update, ...)
+// are intentionally NOT classified here: they are emitted by the session
+// host / extension runner independent of tool execution, so the external
+// engine short-circuit above does not apply to them. Only classify events
+// whose engine behavior has been verified against the agent loop.
+/**
+ * Canonical tool name: strips the `mcp__<server>__` prefix when present,
+ * nothing else. Use for identity checks against host/native tool names and
+ * for any guard that must NOT conflate workflow aliases with their canonical
+ * tools (e.g. write gates, evidence keys).
+ *
+ * Malformed MCP names (empty server or empty tool segment, e.g.
+ * `mcp____tool` or `mcp__server__`) are returned unchanged.
+ */
+export function canonicalToolName(toolName) {
+    return stripMcpToolPrefix(toolName);
+}
+/**
+ * Workflow-aware canonical tool name: strips the MCP prefix AND resolves
+ * workflow tool aliases to their canonical contract names. Use whenever the
+ * name is compared against the workflow tool surface (scoping, presentation,
+ * dispatch) — plain {@link canonicalToolName} would miss alias spellings.
+ */
+export function canonicalWorkflowToolName(toolName) {
+    return canonicalWorkflowSurfaceToolName(toolName);
+}

package/dist/resources/extensions/gsd/error-classifier.js

@@ -9,6 +9,7 @@
  *
  * @see https://github.com/open-gsd/gsd-pi/issues/2577
  */
+import { TOOL_SURFACE_NOT_READY } from "./tool-surface-readiness.js";
 export function createRetryState() {
     return { networkRetryCount: 0, consecutiveTransientCount: 0, currentRetryModelId: undefined };
 }
@@ -32,7 +33,7 @@ const AFFORDABILITY_RE = /requires more credits|can only afford|insufficient cre
 // are emitted by SDK/harness transports for mid-stream disconnects.
 // These indicate transient network-level interruptions.
 // See: https://github.com/open-gsd/gsd-pi/issues/4558
-const NETWORK_RE = /network|ECONNRESET|ETIMEDOUT|ECONNREFUSED|socket hang up|web ?socket|fetch failed|connection.*reset|dns|unexpected eof|stream idle timeout|partial response received/i;
+const NETWORK_RE = /network|ECONNRESET|ETIMEDOUT|ECONNREFUSED|socket hang up|web ?socket|fetch failed|connection.*reset|dns|unexpected eof|stream idle timeout|partial response received|stream ended without finish_reason/i;
 // Context overflow errors (context window/length exceeded) should be treated as server-class
 // transient errors so auto-mode can retry with reduced budget or fall back to a larger-context model.
 // See: https://github.com/open-gsd/gsd-pi/issues/4528
@@ -44,6 +45,10 @@ const CONNECTION_RE = /terminated|connection.?(?:refused|error)|other side close
 const STREAM_RE = /in JSON at position \d+|Unexpected end of JSON|SyntaxError.*JSON/i;
 const RESET_DELAY_RE = /reset in (\d+)s/i;
 const TOOL_SCHEMA_RE = /schema overload|consecutive tool validation failures/i;
+// GSD tool-surface readiness abort (claude-code stream adapter): the workflow
+// MCP server had not registered the Unit's required tools at SDK init. The
+// server typically finishes connecting within seconds — same-model retry.
+const TOOL_SURFACE_NOT_READY_RE = new RegExp(TOOL_SURFACE_NOT_READY, "i");
 // Provider rejected the request shape for the selected model (400 bad request,
 // grammar limits, etc.). Not transient — try a different model/fallback.
 // Context-window 400s stay in SERVER_RE (checked earlier).
@@ -77,6 +82,10 @@ export function classifyError(errorMsg, retryAfterMs) {
     if (TOOL_SCHEMA_RE.test(errorMsg)) {
         return { kind: "tool-schema", retryAfterMs: 0 };
     }
+    // Tool-surface readiness abort — transient; retry the same model shortly.
+    if (TOOL_SURFACE_NOT_READY_RE.test(errorMsg)) {
+        return { kind: "network", retryAfterMs: retryAfterMs ?? 3_000 };
+    }
     const isPermanent = PERMANENT_RE.test(errorMsg);
     const isRateLimit = RATE_LIMIT_RE.test(errorMsg) || AFFORDABILITY_RE.test(errorMsg);
     const isUnsupportedModel = UNSUPPORTED_MODEL_MODEL_RE.test(errorMsg) &&

package/dist/resources/extensions/gsd/exec-sandbox.js

@@ -11,7 +11,11 @@ import { spawn } from "node:child_process";
 import { existsSync, mkdirSync, writeFileSync } from "node:fs";
 import { randomUUID } from "node:crypto";
 import { resolve } from "node:path";
+import { killProcessTree, SIGKILL_GRACE_MS, HARD_DEADLINE_MS } from "@gsd/pi-coding-agent";
 const ALWAYS_FORWARD_ENV = ["PATH", "HOME"];
+// SIGKILL_GRACE_MS / HARD_DEADLINE_MS are imported from @gsd/pi-coding-agent
+// (shell.ts) — the single source of truth for the graceful-kill timing ladder —
+// so this sandbox can never drift from the canonical kill path it delegates to.
 export const EXEC_DEFAULTS = {
     clampTimeoutMs: 600_000,
     defaultTimeoutMs: 30_000,
@@ -120,6 +124,7 @@ export function runExecSandbox(request, opts) {
                 exit_code: null,
                 signal: null,
                 timed_out: false,
+                force_resolved: false,
                 duration_ms: duration,
                 stdout_bytes: 0,
                 stderr_bytes: Buffer.byteLength(`spawn error: ${message}\n`),
@@ -172,24 +177,38 @@ export function runExecSandbox(request, opts) {
                 stderrTruncated = true;
             }
         });
+        const effectiveGraceMs = opts.kill_grace_ms ?? SIGKILL_GRACE_MS;
+        const effectiveForceResolveDelay = opts.force_resolve_delay_ms ?? (effectiveGraceMs + HARD_DEADLINE_MS);
         let timedOut = false;
+        let settled = false;
+        let forceResolveTimer;
         const timer = setTimeout(() => {
             timedOut = true;
-            if (useProcessGroup && child.pid != null) {
-                try {
-                    process.kill(-child.pid, "SIGKILL");
-                }
-                catch {
-                    child.kill("SIGKILL");
-                }
+            // killProcessTree handles both platforms and kills the whole tree: on Unix
+            // it signals the process group (SIGTERM -> grace -> SIGKILL); on Windows it
+            // force-kills the tree via taskkill /F /T. Using child.kill("SIGTERM") here
+            // would only terminate the direct child on Windows, orphaning grandchildren.
+            if (child.pid != null) {
+                killProcessTree(child.pid, { graceMs: effectiveGraceMs });
             }
             else {
-                child.kill("SIGKILL");
+                child.kill("SIGTERM");
             }
+            // Arm hard-deadline force-resolve in case child never closes (D-state).
+            // The "SIGKILL" here is a synthetic marker (the process may not have actually
+            // received it); force_resolved=true records that this was a deadline, not an exit.
+            forceResolveTimer = setTimeout(() => {
+                finalize(null, "SIGKILL", true);
+            }, effectiveForceResolveDelay);
+            forceResolveTimer.unref?.();
         }, timeoutMs);
         timer.unref?.();
-        const finalize = (exitCode, signal) => {
+        const finalize = (exitCode, signal, forceResolved = false) => {
+            if (settled)
+                return;
+            settled = true;
             clearTimeout(timer);
+            clearTimeout(forceResolveTimer);
             const duration = Date.now() - started;
             const stdoutBuf = Buffer.concat(stdoutChunks);
             const stderrBuf = Buffer.concat(stderrChunks);
@@ -211,6 +230,7 @@ export function runExecSandbox(request, opts) {
                 exit_code: exitCode,
                 signal,
                 timed_out: timedOut,
+                force_resolved: forceResolved,
                 duration_ms: duration,
                 stdout_bytes: stdoutBytes,
                 stderr_bytes: stderrBytes,
@@ -254,6 +274,7 @@ function writeMeta(path, result, request, now) {
         exit_code: result.exit_code,
         signal: result.signal,
         timed_out: result.timed_out,
+        force_resolved: result.force_resolved,
         duration_ms: result.duration_ms,
         stdout_bytes: result.stdout_bytes,
         stderr_bytes: result.stderr_bytes,
@@ -261,7 +282,6 @@ function writeMeta(path, result, request, now) {
         stderr_truncated: result.stderr_truncated,
         stdout_path: result.stdout_path,
         stderr_path: result.stderr_path,
-        ...(request.metadata ? { metadata: request.metadata } : {}),
     };
     writeFileSync(path, `${JSON.stringify(meta, null, 2)}\n`);
 }

package/dist/resources/extensions/gsd/files.js

@@ -576,38 +576,52 @@ export function parseTaskPlanIO(content) {
         outputFiles: extractPaths(outputSection),
     };
 }
+/** Canonical list of recognised UAT types — uat-policy.ts re-exports this as UAT_TYPES. */
+export const UAT_TYPE_KEYWORDS = [
+    'artifact-driven',
+    'browser-executable',
+    'runtime-executable',
+    'live-runtime',
+    'mixed',
+    'human-experience',
+];
+/** Match a value against the recognised UAT type keywords (leading-keyword-only). */
+function matchUatTypeKeyword(value) {
+    const normalized = value.trim().toLowerCase();
+    return UAT_TYPE_KEYWORDS.find(keyword => normalized.startsWith(keyword));
+}
 /**
  * Extract the UAT type from a UAT file's raw content.
  *
  * UAT files have no YAML frontmatter - pass raw file content directly.
  * Classification is leading-keyword-only: e.g. `mixed (artifact-driven + live-runtime)` → `'mixed'`.
  *
+ * The canonical form is a `- UAT mode: <type>` bullet under `## UAT Type`
+ * (case-insensitive prefix, `**bold**` tolerated). When no such line exists,
+ * a line that itself starts with a recognised keyword — e.g. a bare
+ * `browser-executable` under the heading — is accepted, so agent-authored
+ * format drift does not silently fall back to artifact-driven.
+ *
  * Returns `undefined` when:
  * - the `## UAT Type` section is absent
- * - no `UAT mode:` bullet is found in the section
- * - the value does not start with a recognised keyword
+ * - a `UAT mode:` line exists but its value starts with no recognised keyword
+ * - no line in the section starts with `UAT mode:` or a recognised keyword
  */
 export function extractUatType(content) {
     const sectionText = extractSection(content, 'UAT Type');
     if (!sectionText)
         return undefined;
-    const bullets = parseBullets(sectionText);
-    const modeBullet = bullets.find(b => b.startsWith('UAT mode:'));
-    if (!modeBullet)
-        return undefined;
-    const rawValue = modeBullet.slice('UAT mode:'.length).trim().toLowerCase();
-    if (rawValue.startsWith('artifact-driven'))
-        return 'artifact-driven';
-    if (rawValue.startsWith('browser-executable'))
-        return 'browser-executable';
-    if (rawValue.startsWith('runtime-executable'))
-        return 'runtime-executable';
-    if (rawValue.startsWith('live-runtime'))
-        return 'live-runtime';
-    if (rawValue.startsWith('human-experience'))
-        return 'human-experience';
-    if (rawValue.startsWith('mixed'))
-        return 'mixed';
+    const lines = parseBullets(sectionText).map(line => line.replace(/\*\*/g, ''));
+    for (const line of lines) {
+        const declared = /^uat mode:\s*(.*)$/i.exec(line);
+        if (declared)
+            return matchUatTypeKeyword(declared[1]);
+    }
+    for (const line of lines) {
+        const matched = matchUatTypeKeyword(line);
+        if (matched)
+            return matched;
+    }
     return undefined;
 }
 /**

package/dist/resources/extensions/gsd/git-conflict-state.js

@@ -61,7 +61,22 @@ export function gitDiffCheckFailures(basePath) {
             .filter(Boolean)
             .join("\n")
             .trim();
-        failures.push(output || `git diff --check ${args.join(" ")} failed`);
+        // git diff --check exits non-zero for both whitespace errors and leftover
+        // conflict markers. Only conflict markers block the gate — whitespace errors
+        // in staged files (common in auto-generated code) must not be misread as an
+        // unresolved merge conflict. Filter to lines git explicitly labels as
+        // "leftover conflict marker".
+        if (!output) {
+            failures.push(`git diff --check ${args.join(" ")} failed`);
+            continue;
+        }
+        const conflictMarkerLines = output
+            .split("\n")
+            .filter((line) => /leftover conflict marker/i.test(line))
+            .join("\n");
+        if (conflictMarkerLines) {
+            failures.push(conflictMarkerLines);
+        }
     }
     return failures;
 }

package/dist/resources/extensions/gsd/git-service.js

@@ -207,6 +207,7 @@ export class MergeConflictError extends GSDError {
  * This array must stay synchronized with it.
  */
 export const RUNTIME_EXCLUSION_PATHS = [
+    ".gsd-worktrees/",
     ".gsd/activity/",
     ".gsd/audit/",
     ".gsd/forensics/",

package/dist/resources/extensions/gsd/gitignore.js

@@ -23,6 +23,7 @@ import { GIT_NO_PROMPT_ENV } from "./git-constants.js";
  * but retained for backwards compatibility during migration.
  */
 const GSD_RUNTIME_PATTERNS = [
+    ".gsd-worktrees/",
     ".gsd/activity/",
     ".gsd/audit/",
     ".gsd/forensics/",
@@ -45,6 +46,8 @@ const GSD_RUNTIME_PATTERNS = [
 const BASELINE_PATTERNS = [
     // ── GSD state directory (symlink to external storage) ──
     ".gsd",
+    // Worktree container sibling — NOT covered by the ".gsd" pattern above.
+    ".gsd-worktrees/",
     ".gsd-id",
     ".mcp.json",
     ".bg-shell/",

package/dist/resources/extensions/gsd/gsd-command-home.js

@@ -28,13 +28,17 @@ export function buildGsdHomeModel(state, closeout) {
     const workLabel = activeWorkLabel(state);
     const strandedQuick = closeout?.strandedQuick ?? null;
     const unmergedMilestone = closeout?.unmergedMilestones?.[0];
+    const idleResidueHint = closeout?.idleResidueHint ?? null;
+    const hasIdleResidue = Boolean(idleResidueHint);
     const nextReason = complete
         ? "all milestones are complete"
         : blocked
             ? "the active milestone is blocked"
-            : !hasActiveWork
-                ? "there is no active milestone"
-                : "";
+            : hasIdleResidue
+                ? "milestone git residue needs recovery"
+                : !hasActiveWork
+                    ? "there is no active milestone"
+                    : "";
     const canAdvance = hasActiveWork && !blocked && !complete;
     const unmappedActive = complete ? countUnmappedActiveRequirements() : 0;
     const recommended = strandedQuick
@@ -43,11 +47,13 @@ export function buildGsdHomeModel(state, closeout) {
             ? "finish_milestone"
             : blocked
                 ? "fix_recover"
-                : canAdvance
-                    ? "continue_step"
-                    : complete && unmappedActive > 0
-                        ? "review_requirements_backlog"
-                        : "start_configure";
+                : hasIdleResidue
+                    ? "fix_recover"
+                    : canAdvance
+                        ? "continue_step"
+                        : complete && unmappedActive > 0
+                            ? "review_requirements_backlog"
+                            : "start_configure";
     const completionSummary = complete
         ? appendRequirementsBacklogToSummary(state, [
             `All milestones complete${state.lastCompletedMilestone ? ` after ${state.lastCompletedMilestone.id}: ${state.lastCompletedMilestone.title}` : ""}.`,
@@ -57,7 +63,9 @@ export function buildGsdHomeModel(state, closeout) {
         ? [`Quick task Q${strandedQuick.taskNum} finished on ${strandedQuick.quickBranch} but is not merged to ${strandedQuick.originalBranch}.`]
         : unmergedMilestone
             ? [`${unmergedMilestone.milestoneId} is complete but not merged into ${unmergedMilestone.integrationBranch}.`]
-            : completionSummary;
+            : idleResidueHint
+                ? [idleResidueHint.message]
+                : completionSummary;
     return {
         title: "GSD — What now?",
         summary: [
@@ -130,10 +138,12 @@ export function buildGsdHomeModel(state, closeout) {
                 label: "Fix or recover",
                 description: blocked
                     ? "Review the blocker and recovery commands for the active milestone."
-                    : disabled("This becomes active when closeout, validation, or state recovery is needed.", "no blocker is active"),
-                enabled: blocked,
+                    : hasIdleResidue
+                        ? "Review stranded milestone worktrees/branches and run the suggested recovery command."
+                        : disabled("This becomes active when closeout, validation, or state recovery is needed.", "no blocker is active"),
+                enabled: blocked || hasIdleResidue,
                 recommended: recommended === "fix_recover",
-                disabledReason: blocked ? undefined : "no blocker is active",
+                disabledReason: blocked || hasIdleResidue ? undefined : "no blocker is active",
             },
             {
                 id: "start_configure",