gsd-pi 2.76.0-dev.82e249f7b → 2.76.0-dev.97807402

package/dist/resources/extensions/gsd/gsd-db.js

@@ -135,7 +135,7 @@ function openRawDb(path) {
     const Database = providerModule;
     return new Database(path);
 }
-const SCHEMA_VERSION = 21;
+const SCHEMA_VERSION = 22;
 function indexExists(db, name) {
     return !!db.prepare("SELECT 1 as present FROM sqlite_master WHERE type = 'index' AND name = ?").get(name);
 }
@@ -495,7 +495,9 @@ function initSchema(db, fileBacked) {
         // that fresh installs already have. Add only columns needed by bootstrap
         // indexes so old DBs can open far enough for the normal migration chain.
         ensureBootstrapIndexColumns(db);
-        db.exec("CREATE INDEX IF NOT EXISTS idx_memories_scope ON memories(scope)");
+        if (columnExists(db, "memories", "scope")) {
+            db.exec("CREATE INDEX IF NOT EXISTS idx_memories_scope ON memories(scope)");
+        }
         db.exec("CREATE INDEX IF NOT EXISTS idx_memory_sources_kind ON memory_sources(kind)");
         db.exec("CREATE INDEX IF NOT EXISTS idx_memory_sources_scope ON memory_sources(scope)");
         db.exec("CREATE INDEX IF NOT EXISTS idx_memory_relations_from ON memory_relations(from_id)");
@@ -854,19 +856,24 @@ function migrateSchema(db) {
             });
         }
         if (currentVersion < 12) {
+            // NOTE: The original DDL used COALESCE(task_id, '') in the PRIMARY KEY
+            // expression, which is invalid SQLite syntax and causes startup errors on
+            // DBs that migrate through v12. The corrected DDL uses
+            // task_id TEXT NOT NULL DEFAULT '' with a plain column list PK. DBs that
+            // were created with the broken DDL are repaired by the v22 migration below.
             db.exec(`
         CREATE TABLE IF NOT EXISTS quality_gates (
           milestone_id TEXT NOT NULL,
           slice_id TEXT NOT NULL,
           gate_id TEXT NOT NULL,
           scope TEXT NOT NULL DEFAULT 'slice',
-          task_id TEXT DEFAULT NULL,
+          task_id TEXT NOT NULL DEFAULT '',
           status TEXT NOT NULL DEFAULT 'pending',
           verdict TEXT NOT NULL DEFAULT '',
           rationale TEXT NOT NULL DEFAULT '',
           findings TEXT NOT NULL DEFAULT '',
           evaluated_at TEXT DEFAULT NULL,
-          PRIMARY KEY (milestone_id, slice_id, gate_id, COALESCE(task_id, '')),
+          PRIMARY KEY (milestone_id, slice_id, gate_id, task_id),
           FOREIGN KEY (milestone_id, slice_id) REFERENCES slices(milestone_id, id)
         )
       `);
@@ -1085,6 +1092,51 @@ function migrateSchema(db) {
                 ":applied_at": new Date().toISOString(),
             });
         }
+        if (currentVersion < 22) {
+            // v22: Repair quality_gates tables that were created by the broken v12
+            // migration (which used COALESCE(task_id, '') as a PK expression — invalid
+            // SQLite DDL). Those DBs have task_id nullable (dflt_value NULL, notnull 0).
+            // Rebuild the table with the correct schema, migrating existing rows via
+            // COALESCE so no data is lost.
+            const qgInfo = db.prepare("PRAGMA table_info(quality_gates)").all();
+            const taskIdCol = qgInfo.find((r) => r["name"] === "task_id");
+            const needsRepair = taskIdCol && (taskIdCol["notnull"] === 0 || taskIdCol["notnull"] === "0");
+            if (needsRepair) {
+                db.exec(`
+          CREATE TABLE quality_gates_new (
+            milestone_id TEXT NOT NULL,
+            slice_id TEXT NOT NULL,
+            gate_id TEXT NOT NULL,
+            scope TEXT NOT NULL DEFAULT 'slice',
+            task_id TEXT NOT NULL DEFAULT '',
+            status TEXT NOT NULL DEFAULT 'pending',
+            verdict TEXT NOT NULL DEFAULT '',
+            rationale TEXT NOT NULL DEFAULT '',
+            findings TEXT NOT NULL DEFAULT '',
+            evaluated_at TEXT DEFAULT NULL,
+            PRIMARY KEY (milestone_id, slice_id, gate_id, task_id),
+            FOREIGN KEY (milestone_id, slice_id) REFERENCES slices(milestone_id, id)
+          )
+        `);
+                db.exec(`
+          INSERT OR IGNORE INTO quality_gates_new
+            (milestone_id, slice_id, gate_id, scope, task_id, status, verdict, rationale, findings, evaluated_at)
+          SELECT milestone_id, slice_id, gate_id, scope, COALESCE(task_id, ''), status, verdict, rationale, findings, evaluated_at
+          FROM quality_gates
+        `);
+                db.exec("DROP TABLE quality_gates");
+                db.exec("ALTER TABLE quality_gates_new RENAME TO quality_gates");
+                db.exec("CREATE INDEX IF NOT EXISTS idx_quality_gates_pending ON quality_gates(milestone_id, slice_id, status)");
+            }
+            // Ensure scope column exists on quality_gates and assessments (guard
+            // against DBs that somehow lack it after a partial migration).
+            ensureColumn(db, "quality_gates", "scope", "ALTER TABLE quality_gates ADD COLUMN scope TEXT NOT NULL DEFAULT 'slice'");
+            ensureColumn(db, "assessments", "scope", "ALTER TABLE assessments ADD COLUMN scope TEXT NOT NULL DEFAULT ''");
+            db.prepare("INSERT INTO schema_version (version, applied_at) VALUES (:version, :applied_at)").run({
+                ":version": 22,
+                ":applied_at": new Date().toISOString(),
+            });
+        }
         db.exec("COMMIT");
     }
     catch (err) {
@@ -1097,6 +1149,8 @@ let currentPath = null;
 let currentPid = 0;
 let _exitHandlerRegistered = false;
 let _dbOpenAttempted = false;
+let _lastDbError = null;
+let _lastDbPhase = null;
 export function getDbProvider() {
     loadProvider();
     return providerName;
@@ -1113,13 +1167,54 @@ export function isDbAvailable() {
 export function wasDbOpenAttempted() {
     return _dbOpenAttempted;
 }
+export function getDbStatus() {
+    loadProvider();
+    return {
+        available: currentDb !== null,
+        provider: providerName,
+        attempted: _dbOpenAttempted,
+        lastError: _lastDbError,
+        lastPhase: _lastDbPhase,
+    };
+}
 export function openDatabase(path) {
     _dbOpenAttempted = true;
     if (currentDb && currentPath !== path)
         closeDatabase();
     if (currentDb && currentPath === path)
         return true;
-    const rawDb = openRawDb(path);
+    // Reset error state only when a new open attempt is actually going to run.
+    _lastDbError = null;
+    _lastDbPhase = null;
+    let rawDb;
+    let fallbackProvider = null;
+    let fallbackModule = null;
+    try {
+        rawDb = openRawDb(path);
+    }
+    catch (primaryErr) {
+        _lastDbPhase = "open";
+        _lastDbError = primaryErr instanceof Error ? primaryErr : new Error(String(primaryErr));
+        // node:sqlite loaded but failed to open this file — try better-sqlite3 as fallback.
+        if (providerName === "node:sqlite") {
+            try {
+                const mod = _require("better-sqlite3");
+                const Db = (mod && mod.default) ? mod.default : mod;
+                if (typeof Db === "function") {
+                    rawDb = new Db(path);
+                    fallbackProvider = "better-sqlite3";
+                    fallbackModule = Db;
+                    _lastDbError = null;
+                    _lastDbPhase = null;
+                }
+            }
+            catch {
+                // fallback unavailable; surface original error
+            }
+        }
+        if (!rawDb)
+            throw primaryErr;
+    }
     if (!rawDb)
         return false;
     const adapter = createAdapter(rawDb);
@@ -1137,6 +1232,8 @@ export function openDatabase(path) {
                 process.stderr.write("gsd-db: recovered corrupt database via VACUUM\n");
             }
             catch (retryErr) {
+                _lastDbPhase = "vacuum-recovery";
+                _lastDbError = retryErr instanceof Error ? retryErr : new Error(String(retryErr));
                 try {
                     adapter.close();
                 }
@@ -1147,15 +1244,22 @@ export function openDatabase(path) {
             }
         }
         else {
+            _lastDbPhase = "initSchema";
+            _lastDbError = err instanceof Error ? err : new Error(String(err));
             try {
                 adapter.close();
             }
             catch (e) {
-                logWarning("db", `close after VACUUM failed: ${e.message}`);
+                logWarning("db", `close after initSchema failed: ${e.message}`);
             }
             throw err;
         }
     }
+    // Commit fallback provider switch only after open + schema both succeeded.
+    if (fallbackProvider) {
+        providerName = fallbackProvider;
+        providerModule = fallbackModule;
+    }
     currentDb = adapter;
     currentPath = path;
     currentPid = process.pid;
@@ -1194,8 +1298,12 @@ export function closeDatabase() {
         currentDb = null;
         currentPath = null;
         currentPid = 0;
-        _dbOpenAttempted = false;
     }
+    // Reset session-scoped state unconditionally so stale error info from a
+    // failed open doesn't persist into the next open attempt or status check.
+    _dbOpenAttempted = false;
+    _lastDbError = null;
+    _lastDbPhase = null;
 }
 /** Run a full VACUUM — call sparingly (e.g. after milestone completion). */
 export function vacuumDatabase() {

package/dist/resources/extensions/gsd/guided-flow.js

@@ -79,6 +79,13 @@ function runPlanV2Gate(ctx, basePath, state) {
 function buildDocsCommitInstruction(_message) {
     return "Do not commit planning artifacts — .gsd/ is managed externally.";
 }
+// #4573: cap for how many times we nudge the LLM after a premature ready
+// phrase before giving up and asking the user to re-run /gsd.
+const MAX_READY_REJECTS = 2;
+// #4573: matches the canonical ready phrase the discuss prompt asks the LLM
+// to emit. Accepts any M-prefixed milestone ID (three digits + optional
+// suffix) with optional trailing punctuation.
+const READY_PHRASE_RE = /\bMilestone\s+M\d{3}[A-Z0-9-]*\s+ready\.?/i;
 const pendingAutoStartMap = new Map();
 /**
  * Backward-compat bridge: returns a mutable reference to the entry matching
@@ -225,6 +232,188 @@ export function checkAutoStartAfterDiscuss() {
     startAutoDetached(ctx, pi, basePath, false, { step });
     return true;
 }
+/**
+ * Extract the concatenated text content from an assistant message, whether it
+ * stores content as a string or as an array of text blocks.
+ */
+function extractAssistantText(msg) {
+    if (!msg)
+        return "";
+    const content = msg.content;
+    if (typeof content === "string")
+        return content;
+    if (!Array.isArray(content))
+        return "";
+    const parts = [];
+    for (const block of content) {
+        if (!block || typeof block !== "object")
+            continue;
+        if (block.type === "text" && typeof block.text === "string")
+            parts.push(block.text);
+    }
+    return parts.join("\n");
+}
+/**
+ * Return true if the assistant message contains any tool-use block.
+ */
+function hasToolUse(msg) {
+    if (!msg)
+        return false;
+    const content = msg.content;
+    if (!Array.isArray(content))
+        return false;
+    return content.some((b) => b && typeof b === "object" && (b.type === "tool_use" || b.type === "tool-use"));
+}
+/**
+ * #4573 — Detect and recover from the "ready phrase without files" failure mode.
+ *
+ * When the LLM emits "Milestone {{id}} ready." but has not written CONTEXT.md
+ * or ROADMAP.md, `checkAutoStartAfterDiscuss()` silently returns false and the
+ * next /gsd invocation loops into the "All milestones complete" warning.
+ *
+ * This function, called from `handleAgentEnd` after `checkAutoStartAfterDiscuss`
+ * returns false, pattern-matches the ready phrase on the last assistant message.
+ * If it fired AND neither CONTEXT.md nor ROADMAP.md exists, it:
+ *   1. Notifies the user that the signal was rejected.
+ *   2. Injects a system message via `pi.sendMessage(..., {triggerTurn:true})`
+ *      telling the LLM the signal was premature and to emit the writes now.
+ *   3. Caps at `MAX_READY_REJECTS` per-entry; beyond that, gives up and asks
+ *      the user to re-run /gsd.
+ *
+ * Returns true when a nudge (or give-up) was emitted, signaling the caller to
+ * skip `resolveAgentEnd`.
+ */
+export function maybeHandleReadyPhraseWithoutFiles(event) {
+    const entry = _getPendingAutoStart();
+    if (!entry)
+        return false;
+    const { ctx, pi, basePath, milestoneId } = entry;
+    // Gate: last assistant message must contain the ready phrase
+    const lastMsg = event.messages[event.messages.length - 1];
+    const text = extractAssistantText(lastMsg);
+    if (!READY_PHRASE_RE.test(text))
+        return false;
+    // Gate: artifacts must still be missing — if they exist, the happy path
+    // already fired and we have nothing to do.
+    const contextFile = resolveMilestoneFile(basePath, milestoneId, "CONTEXT");
+    const roadmapFile = resolveMilestoneFile(basePath, milestoneId, "ROADMAP");
+    if (contextFile || roadmapFile)
+        return false;
+    entry.readyRejectCount = (entry.readyRejectCount ?? 0) + 1;
+    if (entry.readyRejectCount > MAX_READY_REJECTS) {
+        // Give up: clear state and tell the user to re-run /gsd. Avoids an
+        // infinite nudge loop when the LLM never produces the writes.
+        pendingAutoStartMap.delete(basePath);
+        ctx.ui.notify(`Milestone ${milestoneId}: LLM signaled "ready" ${entry.readyRejectCount} times without writing files. ` +
+            `Stopping auto-nudge. Run /gsd to try again.`, "error");
+        return true;
+    }
+    ctx.ui.notify(`Milestone ${milestoneId}: "ready" signal rejected — CONTEXT.md and ROADMAP.md are missing. Asking the LLM to complete the writes.`, "warning");
+    const nudge = `You emitted "Milestone ${milestoneId} ready." but neither ` +
+        `.gsd/milestones/${milestoneId}/${milestoneId}-CONTEXT.md nor ` +
+        `.gsd/milestones/${milestoneId}/${milestoneId}-ROADMAP.md exists on disk. ` +
+        `The ready phrase is a POST-WRITE signal and has been rejected. ` +
+        `In this turn: (1) write PROJECT.md, REQUIREMENTS.md, and the milestone ` +
+        `CONTEXT.md, (2) call gsd_plan_milestone, then (3) emit the ready phrase. ` +
+        `Do not describe these steps — execute them as tool calls. ` +
+        `This is retry ${entry.readyRejectCount}/${MAX_READY_REJECTS}; further ` +
+        `premature signals will clear the session.`;
+    try {
+        pi.sendMessage({ customType: "gsd-ready-no-files", content: nudge, display: false }, { triggerTurn: true });
+    }
+    catch (e) {
+        logWarning("guided", `ready-phrase nudge sendMessage failed: ${e.message}`);
+        return false;
+    }
+    return true;
+}
+/**
+ * #4573 — Detect and recover from the "announces tool, never calls it" stall.
+ *
+ * The LLM emits text like "I'll now write the CONTEXT.md file" but the turn
+ * ends with zero tool-use blocks. The harness has no post-turn tool-call
+ * validation, so the unit promise resolves and the user sees a stalled state.
+ *
+ * This function, called from `handleAgentEnd`, inspects the last assistant
+ * message. If ALL of the following are true, it injects a recovery message:
+ *   - Text-only (no tool-use blocks)
+ *   - Contains a commit-intent phrase ("I'll write", "I'll call", etc.)
+ *   - Auto-mode is active OR a discussion autostart is pending
+ *   - `emptyTurnRetryCount` is under the cap
+ *
+ * Per-handler state is held on the `PendingAutoStartEntry` when present, and
+ * on a module-level map otherwise. The counter resets on any successful
+ * tool-use turn via `resetEmptyTurnCounter`.
+ */
+const emptyTurnCounterByBase = new Map();
+const MAX_EMPTY_TURN_RETRIES = 2;
+// Phrases that indicate the LLM is about to do something but has not yet.
+// Kept tight to avoid flagging legitimate narration like "I'll wait for your answer."
+const COMMIT_INTENT_RE = /\b(?:I['’]ll|I will|Next,? I['’]ll|Now I['’]ll|Let me|I['’]m going to|I am going to)\s+(?:now\s+)?(?:write|create|call|invoke|update|add|make|run|execute|generate|produce|emit|compose|implement|save|apply|commit)\b/i;
+/**
+ * Reset the empty-turn counter for a basePath after a successful tool-use turn.
+ * Called from handleAgentEnd when the last message contains tool_use blocks.
+ */
+export function resetEmptyTurnCounter(basePath) {
+    if (basePath)
+        emptyTurnCounterByBase.delete(basePath);
+    else
+        emptyTurnCounterByBase.clear();
+}
+export function maybeHandleEmptyIntentTurn(event, isAuto) {
+    // Gate: only fire when there is system-driven work in flight. Interactive
+    // /gsd discuss (user-driven) produces legitimate text-only turns.
+    if (!isAuto && pendingAutoStartMap.size === 0)
+        return false;
+    const lastMsg = event.messages[event.messages.length - 1];
+    if (!lastMsg)
+        return false;
+    if (hasToolUse(lastMsg))
+        return false;
+    const text = extractAssistantText(lastMsg).trim();
+    if (!text)
+        return false;
+    // Skip if the LLM is emitting the ready phrase — that is the ready-no-files
+    // path, handled by maybeHandleReadyPhraseWithoutFiles.
+    if (READY_PHRASE_RE.test(text))
+        return false;
+    // Skip if the LLM is clearly handing back to the user. Keep the heuristic
+    // tight: a trailing question mark on the last non-empty line is the common
+    // signal for "I asked the user a question and stopped."
+    const lines = text.split(/\r?\n/).map((l) => l.trim()).filter(Boolean);
+    const lastLine = lines[lines.length - 1] ?? "";
+    if (lastLine.endsWith("?"))
+        return false;
+    // Must contain a commit-intent phrase — this is the stall we care about.
+    if (!COMMIT_INTENT_RE.test(text))
+        return false;
+    // Resolve the target basePath + pi for injection. Prefer the pending
+    // autostart entry (discuss flow); otherwise we cannot inject.
+    const entry = _getPendingAutoStart();
+    if (!entry)
+        return false;
+    const { ctx, pi, basePath } = entry;
+    const count = (emptyTurnCounterByBase.get(basePath) ?? 0) + 1;
+    emptyTurnCounterByBase.set(basePath, count);
+    if (count > MAX_EMPTY_TURN_RETRIES) {
+        ctx.ui.notify(`Empty-turn recovery: LLM announced intent ${count} times without calling any tool. ` +
+            `Stopping auto-nudge.`, "error");
+        return false; // let the normal flow resolve/pause the unit
+    }
+    ctx.ui.notify(`Empty-turn detected: LLM announced intent but called no tool. Prompting it to execute.`, "info");
+    const nudge = `Your last turn announced an action (e.g. "I'll write…" or "Let me call…") ` +
+        `but contained no tool call. The system records zero tool-use blocks for ` +
+        `that turn. Execute the announced action NOW as a tool call in this turn. ` +
+        `Do not describe it again. Retry ${count}/${MAX_EMPTY_TURN_RETRIES}.`;
+    try {
+        pi.sendMessage({ customType: "gsd-empty-turn-recovery", content: nudge, display: false }, { triggerTurn: true });
+    }
+    catch (e) {
+        logWarning("guided", `empty-turn nudge sendMessage failed: ${e.message}`);
+        return false;
+    }
+    return true;
+}
 /**
  * Extract milestone IDs from PROJECT.md milestone sequence table.
  * Looks for rows like "| M001 | Name | Status |" and extracts the ID column.

package/dist/resources/extensions/gsd/health-widget.js

@@ -91,6 +91,7 @@ export function initHealthWidget(ctx) {
         let data = initialData;
         let cachedLines;
         let refreshInFlight = false;
+        let isDisposed = false;
         const refresh = async () => {
             if (refreshInFlight)
                 return;
@@ -98,7 +99,8 @@ export function initHealthWidget(ctx) {
             try {
                 data = loadHealthWidgetData(basePath);
                 cachedLines = undefined;
-                _tui.requestRender();
+                if (!isDisposed)
+                    _tui.requestRender();
             }
             catch { /* non-fatal */ }
             finally {
@@ -122,6 +124,7 @@ export function initHealthWidget(ctx) {
             },
             invalidate() { cachedLines = undefined; cachedWidth = undefined; },
             dispose() {
+                isDisposed = true;
                 clearInterval(refreshTimer);
             },
         };

package/dist/resources/extensions/gsd/init-wizard.js

@@ -8,7 +8,7 @@
 import { existsSync, mkdirSync, writeFileSync, readFileSync } from "node:fs";
 import { join } from "node:path";
 import { showNextAction } from "../shared/tui.js";
-import { nativeInit } from "./native-git-bridge.js";
+import { nativeInit, nativeAddAll, nativeCommit } from "./native-git-bridge.js";
 import { ensureGitignore, untrackRuntimeFiles } from "./gitignore.js";
 import { gsdRoot } from "./paths.js";
 import { assertSafeDirectory } from "./validate-directory.js";
@@ -40,6 +40,7 @@ export async function showProjectInit(ctx, pi, basePath, detection) {
         ctx.ui.notify(`Project detected:\n${detectionSummary.join("\n")}`, "info");
     }
     // ── Step 2: Git setup ──────────────────────────────────────────────────────
+    let didInitGit = false;
     if (!signals.isGitRepo) {
         const gitChoice = await showNextAction(ctx, {
             title: "GSD — Project Setup",
@@ -54,6 +55,7 @@ export async function showProjectInit(ctx, pi, basePath, detection) {
             return { completed: false, bootstrapped: false };
         if (gitChoice === "init_git") {
             nativeInit(basePath, prefs.mainBranch);
+            didInitGit = true;
         }
     }
     else {
@@ -244,6 +246,18 @@ export async function showProjectInit(ctx, pi, basePath, detection) {
     // Ensure .gitignore
     ensureGitignore(basePath);
     untrackRuntimeFiles(basePath);
+    // Create initial commit so git log and git worktree work immediately (#4530).
+    // Without this, the branch is "unborn" (zero commits) and downstream operations
+    // like `git log` and `git worktree add` fail.
+    if (didInitGit) {
+        try {
+            nativeAddAll(basePath);
+            nativeCommit(basePath, "chore: init project");
+        }
+        catch {
+            // Non-fatal — user can commit manually; don't block project init
+        }
+    }
     // Auto-generate codebase map for instant agent orientation
     try {
         const result = generateCodebaseMap(basePath);

package/dist/resources/extensions/gsd/key-manager.js

@@ -12,6 +12,12 @@ import { getErrorMessage } from "./error-utils.js";
 export const PROVIDER_REGISTRY = [
     // LLM Providers
     { id: "anthropic", label: "Anthropic (Claude)", category: "llm", envVar: "ANTHROPIC_API_KEY", prefixes: ["sk-ant-"], hasOAuth: true, dashboardUrl: "console.anthropic.com" },
+    // Claude Code CLI: routes through the local `claude` binary — no API key,
+    // authentication is handled by the CLI's own OAuth flow.
+    // Referenced by doctor-providers.ts, auto-model-selection.ts, and others;
+    // must be in the canonical registry so all consumers see the same catalog.
+    // See: https://github.com/gsd-build/gsd-2/issues/4541
+    { id: "claude-code", label: "Claude Code CLI", category: "llm", hasOAuth: true },
     { id: "openai", label: "OpenAI", category: "llm", envVar: "OPENAI_API_KEY", prefixes: ["sk-"], dashboardUrl: "platform.openai.com/api-keys" },
     { id: "github-copilot", label: "GitHub Copilot", category: "llm", envVar: "GITHUB_TOKEN", hasOAuth: true },
     { id: "openai-codex", label: "ChatGPT Plus/Pro (Codex)", category: "llm", hasOAuth: true },

package/dist/resources/extensions/gsd/model-router.js

@@ -440,6 +440,20 @@ function getModelCost(modelId) {
 function bareModelId(modelId) {
     return modelId.includes("/") ? modelId.split("/").pop() : modelId;
 }
+// ─── Provider-specific Tool Limits ─────────────────────────────────────────
+/**
+ * Groq enforces a hard limit of 128 tools per request.
+ * Requests exceeding this limit receive a 400 error:
+ * "maximum number of items is 128"
+ * @see https://console.groq.com/docs/tool-use
+ */
+export const GROQ_MAX_TOOLS = 128;
+/**
+ * Provider IDs that map to the Groq API backend.
+ * Used to detect Groq at the GSD routing layer where only the provider string
+ * is available (the pi-ai openai-completions adapter is shared across providers).
+ */
+const GROQ_PROVIDER_IDS = new Set(["groq"]);
 // ─── Tool Compatibility Filter (ADR-005 Phase 3) ───────────────────────────
 /**
  * Check if a tool is compatible with a provider's capabilities.
@@ -461,8 +475,14 @@ export function isToolCompatibleWithProvider(toolName, providerCaps) {
 /**
  * Filter a list of tool names to only those compatible with a provider.
  * Used by the routing pipeline to adjust tool sets when switching providers.
+ *
+ * @param toolNames - The full list of active tool names to filter.
+ * @param providerApi - The pi-ai API string (e.g. "openai-completions").
+ * @param provider - Optional provider ID (e.g. "groq"). Used to apply
+ *   provider-specific limits that can't be expressed as API-level capabilities
+ *   (e.g. Groq's 128-tool hard limit on the shared openai-completions adapter).
  */
-export function filterToolsForProvider(toolNames, providerApi) {
+export function filterToolsForProvider(toolNames, providerApi, provider) {
     const providerCaps = getProviderCapabilities(providerApi);
     // Provider doesn't support tool calling at all
     if (!providerCaps.toolCalling) {
@@ -478,6 +498,14 @@ export function filterToolsForProvider(toolNames, providerApi) {
             filtered.push(name);
         }
     }
+    // Groq enforces a hard limit of 128 tools per request (#4376).
+    // Trim the compatible list to GROQ_MAX_TOOLS and move the excess to filtered.
+    if (provider && GROQ_PROVIDER_IDS.has(provider) && compatible.length > GROQ_MAX_TOOLS) {
+        const trimmed = compatible.splice(GROQ_MAX_TOOLS);
+        filtered.push(...trimmed);
+        console.warn(`[gsd] Groq tool limit: ${compatible.length + trimmed.length} tools active but Groq allows at most ${GROQ_MAX_TOOLS}. ` +
+            `Trimming to the first ${GROQ_MAX_TOOLS} tools. Removed: ${trimmed.join(", ")}`);
+    }
     return { compatible, filtered };
 }
 /**
@@ -486,8 +514,13 @@ export function filterToolsForProvider(toolNames, providerApi) {
  *
  * This is a hard filter only — it removes tools that would fail at the
  * provider level. It does NOT remove tools based on soft heuristics.
+ *
+ * @param activeToolNames - The full list of currently active tool names.
+ * @param selectedModelApi - The pi-ai API string for the selected model.
+ * @param provider - Optional provider ID (e.g. "groq") for provider-specific
+ *   limits beyond what the API-level capability profile expresses.
  */
-export function adjustToolSet(activeToolNames, selectedModelApi) {
-    const { compatible, filtered } = filterToolsForProvider(activeToolNames, selectedModelApi);
+export function adjustToolSet(activeToolNames, selectedModelApi, provider) {
+    const { compatible, filtered } = filterToolsForProvider(activeToolNames, selectedModelApi, provider);
     return { toolNames: compatible, removedTools: filtered };
 }

package/dist/resources/extensions/gsd/pre-execution-checks.js

@@ -68,8 +68,13 @@ export function extractPackageReferences(description) {
             }
         }
     }
-    // require('pkg') or import from 'pkg' in code blocks
-    const importPattern = /(?:require\s*\(\s*['"]|from\s+['"])([a-zA-Z0-9@/_-]+)['"\)]/g;
+    // require('pkg') or `import ... from 'pkg'` in code blocks.
+    // The `from\s+['"]` branch MUST be preceded by an `import` keyword so that
+    // natural-language prose like `from "What's Next"` or `from 'master'` does
+    // not produce false package-existence failures.  Requiring the leading import
+    // keyword anchors the match to JavaScript/TypeScript syntax.
+    // See: https://github.com/gsd-build/gsd-2/issues/4388
+    const importPattern = /(?:require\s*\(\s*['"]|import\b[\s\S]*?\bfrom\s+['"])([a-zA-Z0-9@/_-]+)['"\)]/g;
     let importMatch;
     while ((importMatch = importPattern.exec(description)) !== null) {
         // Skip relative imports and node builtins
@@ -278,7 +283,12 @@ function extractPathFromAnnotation(raw) {
     }
     const annotatedMatch = trimmed.match(/^(.+?)\s+[—–-]\s+.+$/);
     if (annotatedMatch) {
-        return annotatedMatch[1].trim();
+        const prefix = annotatedMatch[1].trim();
+        const prefixBacktickMatch = prefix.match(/`([^`]+)`/);
+        if (prefixBacktickMatch && looksLikePathOrUrl(prefixBacktickMatch[1].trim())) {
+            return prefixBacktickMatch[1].trim();
+        }
+        return prefix.replace(/`/g, "").trim();
     }
     // Fallback: scan all backticked tokens and return the first one that looks
     // like a path or URL. Handles prose-annotated bullets such as:
@@ -333,13 +343,19 @@ function containsGlobPattern(candidate) {
 }
 /**
  * Build a set of files that will be created by tasks up to (but not including) taskIndex.
+ * Also includes outputs of completed tasks at any position — a completed task has already
+ * run and its outputs are available regardless of sequence position or disk state (#4071).
  * All paths are normalized for consistent comparison.
  */
 function getExpectedOutputsUpTo(tasks, taskIndex) {
     const outputs = new Set();
-    for (let i = 0; i < taskIndex; i++) {
-        for (const file of tasks[i].expected_output) {
-            outputs.add(normalizeFilePath(file));
+    for (let i = 0; i < tasks.length; i++) {
+        const task = tasks[i];
+        // Include prior tasks (i < taskIndex) OR completed tasks at any position
+        if (i < taskIndex || task.status === "completed") {
+            for (const file of task.expected_output) {
+                outputs.add(normalizeFilePath(file));
+            }
         }
     }
     return outputs;
@@ -416,8 +432,14 @@ export function checkTaskOrdering(tasks, basePath) {
         const task = tasks[i];
         for (const file of task.expected_output) {
             const normalizedFile = normalizeFilePath(file);
-            if (!fileCreators.has(normalizedFile)) {
-                fileCreators.set(normalizedFile, { taskId: task.id, index: i, originalPath: file });
+            const existing = fileCreators.get(normalizedFile);
+            if (!existing || (!existing.completed && task.status === "completed")) {
+                fileCreators.set(normalizedFile, {
+                    taskId: task.id,
+                    index: i,
+                    originalPath: file,
+                    completed: task.status === "completed",
+                });
             }
         }
     }
@@ -441,7 +463,11 @@ export function checkTaskOrdering(tasks, basePath) {
             const creator = fileCreators.get(normalizedFile);
             const absolutePath = resolve(basePath, normalizedFile);
             const existsOnDisk = existsSync(absolutePath);
-            if (creator && creator.index > i && !existsOnDisk) {
+            // Skip if the creating task has already completed — its output is available
+            // regardless of disk state (e.g. file was a temp artifact cleaned up after
+            // the task ran, or a replan introduced a new earlier-sequence task that
+            // reads this pre-execution output). (#4071)
+            if (creator && creator.index > i && !existsOnDisk && !creator.completed) {
                 // Task reads file that is created later — impossible ordering
                 results.push({
                     category: "file",

package/dist/resources/extensions/gsd/preferences-types.js

@@ -5,6 +5,14 @@
  * both the validation and runtime modules can import them without pulling
  * in filesystem or loading logic.
  */
+/**
+ * Resolve whether context-mode features (gsd_exec sandbox + compaction
+ * snapshot) should be active. Default is ON: missing config or missing
+ * `enabled` is treated as true. Only `enabled: false` disables.
+ */
+export function isContextModeEnabled(prefs) {
+    return prefs?.context_mode?.enabled !== false;
+}
 /** Default preference values for each workflow mode. */
 export const MODE_DEFAULTS = {
     solo: {
@@ -87,6 +95,7 @@ export const KNOWN_PREFERENCE_KEYS = new Set([
     "flat_rate_providers",
     "language",
     "context_window_override",
+    "context_mode",
 ]);
 /** Canonical list of all dispatch unit types. */
 export const KNOWN_UNIT_TYPES = [