gsd-pi 2.75.0 → 2.76.0-dev.4c866b677

package/dist/resources/extensions/gsd/auto-prompts.js

@@ -21,12 +21,53 @@ import { formatDecisionsCompact, formatRequirementsCompact } from "./structured-
 import { readPhaseAnchor, formatAnchorForPrompt } from "./phase-anchor.js";
 import { logWarning } from "./workflow-logger.js";
 import { inlineGraphSubgraph } from "./graph-context.js";
+import { buildExtractionStepsBlock } from "./commands-extract-learnings.js";
 // ─── Preamble Cap ─────────────────────────────────────────────────────────────
+/**
+ * Historical static ceiling for the preamble cap. Kept as an upper bound even
+ * after context-window-aware sizing so large-window users don't suddenly see
+ * 10× looser caps than before. Small-window users get a tighter cap derived
+ * from their configured executor window.
+ */
 const MAX_PREAMBLE_CHARS = 30_000;
+/**
+ * Resolve prompt budgets from the configured executor context window.
+ *
+ * The prompt builders here don't have access to the runtime model registry
+ * (they're called from many non-ctx sites), so `resolveExecutorContextWindow`
+ * is fed the user-configurable `context_window_override` preference as the
+ * `sessionContextWindow` fallback. That preference exists specifically to
+ * cover small-window local models (e.g. 32K lemonade/llama.cpp servers) whose
+ * n_ctx is not discoverable through the model registry. Issue #4435.
+ */
+function resolvePromptBudgets() {
+    try {
+        const prefs = loadEffectiveGSDPreferences();
+        const sessionWindow = prefs?.preferences.context_window_override;
+        const windowTokens = resolveExecutorContextWindow(undefined, prefs?.preferences, sessionWindow);
+        return computeBudgets(windowTokens);
+    }
+    catch (e) {
+        logWarning("prompt", `resolvePromptBudgets failed: ${e.message}`);
+        return computeBudgets(200_000);
+    }
+}
+/**
+ * Character budget for dependency/prior slice summaries injected into dispatch
+ * prompts. Scales with the executor's configured context window (issue #4435).
+ */
+function resolveSummaryBudgetChars() {
+    return resolvePromptBudgets().summaryBudgetChars;
+}
 function capPreamble(preamble) {
-    if (preamble.length <= MAX_PREAMBLE_CHARS)
+    // Cap inlined context at min(historical 30K ceiling, scaled inline budget).
+    // The ceiling preserves pre-fix behavior for large-window users; the scaled
+    // budget tightens the cap for small-window users whose true safe limit is
+    // below 30K. `computeBudgets` allocates 40% of total chars to inline context.
+    const budget = Math.min(MAX_PREAMBLE_CHARS, resolvePromptBudgets().inlineContextBudgetChars);
+    if (preamble.length <= budget)
         return preamble;
-    return truncateAtSectionBoundary(preamble, MAX_PREAMBLE_CHARS).content;
+    return truncateAtSectionBoundary(preamble, budget).content;
 }
 // ─── Executor Constraints ─────────────────────────────────────────────────────
 /**
@@ -34,15 +75,17 @@ function capPreamble(preamble) {
  * Uses the budget engine to compute task count ranges and inline context budgets
  * based on the configured executor model's context window.
  */
-function formatExecutorConstraints() {
+function formatExecutorConstraints(sessionContextWindow, modelRegistry) {
     let windowTokens;
     try {
         const prefs = loadEffectiveGSDPreferences();
-        windowTokens = resolveExecutorContextWindow(undefined, prefs?.preferences);
+        windowTokens = resolveExecutorContextWindow(modelRegistry, prefs?.preferences, sessionContextWindow);
     }
     catch (e) {
         logWarning("prompt", `resolveExecutorContextWindow failed: ${e.message}`);
-        windowTokens = 200_000; // safe default
+        // Delegate to the budget engine without prefs (the path that just threw)
+        // so DEFAULT_CONTEXT_WINDOW stays the single source of truth.
+        windowTokens = resolveExecutorContextWindow(undefined, undefined, sessionContextWindow);
     }
     const budgets = computeBudgets(windowTokens);
     const { min, max } = budgets.taskCountRange;
@@ -57,7 +100,18 @@ function formatExecutorConstraints() {
         `- Keep individual tasks completable within a single context window — if a task needs more context than fits, split it`,
     ].join("\n");
 }
-function buildSourceFilePaths(base, mid, sid) {
+/**
+ * Returns a markdown bullet list of known context file paths for the given
+ * milestone (and optionally slice). Falls back to a generic tool-agnostic
+ * instruction when no GSD artifacts are found.
+ *
+ * @param base - Absolute path to the project root.
+ * @param mid  - Milestone ID (e.g. `"M001"`).
+ * @param sid  - Optional slice ID (e.g. `"S01"`). When provided, the slice
+ *   RESEARCH file is preferred over the milestone-level one.
+ * @returns Markdown string of file path bullets, or a fallback instruction.
+ */
+export function buildSourceFilePaths(base, mid, sid) {
     const paths = [];
     const projectPath = resolveGsdRootFile(base, "PROJECT");
     if (existsSync(projectPath)) {
@@ -97,7 +151,7 @@ function buildSourceFilePaths(base, mid, sid) {
     }
     return paths.length > 0
         ? paths.join("\n")
-        : "- Use `rg --files` and targeted reads to identify the relevant source files before planning.";
+        : "- Use the Grep/Glob/Read tools to identify the relevant source files before planning.";
 }
 // ─── Inline Helpers ───────────────────────────────────────────────────────
 /**
@@ -872,13 +926,13 @@ export async function checkNeedsRunUat(base, mid, state, prefs) {
  * as a seed when present. The discussion agent interviews the user, writes
  * a full CONTEXT.md, and the phase transitions to pre-planning automatically.
  */
-export async function buildDiscussMilestonePrompt(mid, midTitle, base) {
+export async function buildDiscussMilestonePrompt(mid, midTitle, base, structuredQuestionsAvailable = "false") {
     const discussTemplates = inlineTemplate("context", "Context");
     const basePrompt = loadPrompt("guided-discuss-milestone", {
         milestoneId: mid,
         milestoneTitle: midTitle,
         inlinedTemplates: discussTemplates,
-        structuredQuestionsAvailable: "false",
+        structuredQuestionsAvailable,
         commitInstruction: "Do not commit planning artifacts — .gsd/ is managed externally.",
         fastPathInstruction: "",
     });
@@ -1046,7 +1100,7 @@ export async function buildResearchSlicePrompt(mid, _midTitle, sid, sTitle, base
     if (graphBlockRS)
         inlined.push(graphBlockRS);
     inlined.push(inlineTemplate("research", "Research"));
-    const depContent = await inlineDependencySummaries(mid, sid, base);
+    const depContent = await inlineDependencySummaries(mid, sid, base, resolveSummaryBudgetChars());
     const activeOverrides = await loadActiveOverrides(base);
     const overridesInline = formatOverridesSection(activeOverrides);
     if (overridesInline)
@@ -1073,26 +1127,35 @@ export async function buildResearchSlicePrompt(mid, _midTitle, sid, sTitle, base
         ...buildSkillDiscoveryVars(),
     });
 }
-export async function buildPlanSlicePrompt(mid, _midTitle, sid, sTitle, base, level) {
-    const inlineLevel = level ?? resolveInlineLevel();
+/**
+ * Shared assembly for plan-slice and refine-slice prompts. Both builders need
+ * the same inlined context (roadmap excerpt, slice context, research, decisions,
+ * requirements, knowledge, graph subgraph, templates, dependency summaries,
+ * overrides). Extracted to prevent drift between the two sites.
+ *
+ * `prependBlocks` are pushed onto the start of the inlined array BEFORE any
+ * shared content, so callers can add unit-specific headers (e.g., the refine
+ * sketch-scope constraint).
+ */
+async function renderSlicePrompt(options) {
+    const { mid, sid, sTitle, base, level, promptTemplate, prependBlocks = [], extraVars = {}, sessionContextWindow, modelRegistry, } = options;
     const roadmapPath = resolveMilestoneFile(base, mid, "ROADMAP");
     const roadmapRel = relMilestoneFile(base, mid, "ROADMAP");
     const researchPath = resolveSliceFile(base, mid, sid, "RESEARCH");
     const researchRel = relSliceFile(base, mid, sid, "RESEARCH");
     const sliceContextPath = resolveSliceFile(base, mid, sid, "CONTEXT");
     const sliceContextRel = relSliceFile(base, mid, sid, "CONTEXT");
-    const inlined = [];
-    // Inject phase handoff anchor from research phase (if available)
+    const inlined = [...prependBlocks];
+    // Phase handoff anchor from research phase (if available)
     const researchSliceAnchor = readPhaseAnchor(base, mid, "research-slice");
     if (researchSliceAnchor)
         inlined.push(formatAnchorForPrompt(researchSliceAnchor));
-    // Use roadmap excerpt instead of full roadmap for context reduction
-    const roadmapExcerptPS = await inlineRoadmapExcerpt(base, mid, sid);
-    if (roadmapExcerptPS) {
-        inlined.push(roadmapExcerptPS);
+    // Roadmap excerpt with full-roadmap fallback
+    const roadmapExcerpt = await inlineRoadmapExcerpt(base, mid, sid);
+    if (roadmapExcerpt) {
+        inlined.push(roadmapExcerpt);
     }
     else {
-        // Fall back to full roadmap if excerpt fails
         inlined.push(await inlineFile(roadmapPath, roadmapRel, "Milestone Roadmap"));
     }
     const sliceCtxInline = await inlineFileOptional(sliceContextPath, sliceContextRel, "Slice Context (from discussion)");
@@ -1101,40 +1164,34 @@ export async function buildPlanSlicePrompt(mid, _midTitle, sid, sTitle, base, le
     const researchInline = await inlineFileOptional(researchPath, researchRel, "Slice Research");
     if (researchInline)
         inlined.push(researchInline);
-    if (inlineLevel !== "minimal") {
-        // Derive scope from slice title for decision filtering (R005)
-        const derivedScopePS = deriveSliceScope(sTitle);
-        const decisionsInline = await inlineDecisionsFromDb(base, mid, derivedScopePS, inlineLevel);
+    if (level !== "minimal") {
+        const derivedScope = deriveSliceScope(sTitle);
+        const decisionsInline = await inlineDecisionsFromDb(base, mid, derivedScope, level);
         if (decisionsInline)
             inlined.push(decisionsInline);
-        const requirementsInline = await inlineRequirementsFromDb(base, mid, sid, inlineLevel);
+        const requirementsInline = await inlineRequirementsFromDb(base, mid, sid, level);
         if (requirementsInline)
             inlined.push(requirementsInline);
     }
-    // Use scoped knowledge based on slice title keywords
-    const keywordsPS = extractKeywords(sTitle);
-    const knowledgeInlinePS = await inlineKnowledgeScoped(base, keywordsPS);
-    if (knowledgeInlinePS)
-        inlined.push(knowledgeInlinePS);
-    // Knowledge graph: subgraph for this slice (graceful — skipped if no graph.json)
-    const graphBlockPS = await inlineGraphSubgraph(base, `${sid} ${sTitle}`, { budget: 3000 });
-    if (graphBlockPS)
-        inlined.push(graphBlockPS);
+    const knowledgeInline = await inlineKnowledgeScoped(base, extractKeywords(sTitle));
+    if (knowledgeInline)
+        inlined.push(knowledgeInline);
+    const graphBlock = await inlineGraphSubgraph(base, `${sid} ${sTitle}`, { budget: 3000 });
+    if (graphBlock)
+        inlined.push(graphBlock);
     inlined.push(inlineTemplate("plan", "Slice Plan"));
-    if (inlineLevel === "full") {
+    if (level === "full") {
         inlined.push(inlineTemplate("task-plan", "Task Plan"));
     }
-    const depContent = await inlineDependencySummaries(mid, sid, base);
-    const planActiveOverrides = await loadActiveOverrides(base);
-    const planOverridesInline = formatOverridesSection(planActiveOverrides);
-    if (planOverridesInline)
-        inlined.unshift(planOverridesInline);
+    const depContent = await inlineDependencySummaries(mid, sid, base, resolveSummaryBudgetChars());
+    const overridesInline = formatOverridesSection(await loadActiveOverrides(base));
+    if (overridesInline)
+        inlined.unshift(overridesInline);
     const inlinedContext = capPreamble(`## Inlined Context (preloaded — do not re-read these files)\n\n${inlined.join("\n\n---\n\n")}`);
-    // Build executor context constraints from the budget engine
-    const executorContextConstraints = formatExecutorConstraints();
+    const executorContextConstraints = formatExecutorConstraints(sessionContextWindow, modelRegistry);
     const outputRelPath = relSliceFile(base, mid, sid, "PLAN");
     const commitInstruction = "Do not commit — .gsd/ planning docs are managed externally and not tracked in git.";
-    return loadPrompt("plan-slice", {
+    return loadPrompt(promptTemplate, {
         workingDirectory: base,
         milestoneId: mid, sliceId: sid, sliceTitle: sTitle,
         slicePath: relSlicePath(base, mid, sid),
@@ -1153,6 +1210,60 @@ export async function buildPlanSlicePrompt(mid, _midTitle, sid, sTitle, base, le
             sliceTitle: sTitle,
             extraContext: [inlinedContext, depContent],
         }),
+        ...extraVars,
+    });
+}
+export async function buildPlanSlicePrompt(mid, _midTitle, sid, sTitle, base, level, options) {
+    const prependBlocks = [];
+    // ADR-011: when the refining-phase dispatch rule gracefully downgrades to
+    // plan-slice (progressive_planning was toggled off mid-milestone), it
+    // forwards the stored sketch_scope as a SOFT hint — context, not a hard
+    // constraint. The planner is free to expand beyond it.
+    if (options?.softScopeHint && options.softScopeHint.trim().length > 0) {
+        prependBlocks.push(`## Prior Sketch Scope (soft hint — non-binding)\n\n${options.softScopeHint.trim()}\n\n` +
+            `This scope was captured during an earlier progressive-planning pass that was later disabled. Treat it as context only — you may plan beyond it if the work genuinely requires more scope. Do NOT treat this as a hard boundary.`);
+    }
+    return renderSlicePrompt({
+        mid, sid, sTitle, base,
+        level: level ?? resolveInlineLevel(),
+        promptTemplate: "plan-slice",
+        prependBlocks,
+        sessionContextWindow: options?.sessionContextWindow,
+        modelRegistry: options?.modelRegistry,
+    });
+}
+/**
+ * ADR-011 refine-slice: expand a sketch into a full plan using the current
+ * codebase state and prior slice summary. Mechanically similar to plan-slice
+ * but framed as a *transformation* (sketch → full plan) rather than a
+ * blank-sheet planning pass. Reuses inlineDependencySummaries for prior
+ * slice SUMMARY and inlines the stored sketch_scope as a hard constraint.
+ */
+export async function buildRefineSlicePrompt(mid, _midTitle, sid, sTitle, base, level, options) {
+    // Pull the stored sketch scope from the DB — the hard constraint we plan within.
+    let sketchScope = "";
+    try {
+        const { isDbAvailable, getSlice } = await import("./gsd-db.js");
+        if (isDbAvailable()) {
+            sketchScope = getSlice(mid, sid)?.sketch_scope ?? "";
+        }
+    }
+    catch {
+        sketchScope = "";
+    }
+    const prependBlocks = [];
+    if (sketchScope.trim().length > 0) {
+        prependBlocks.push(`## Sketch Scope (hard constraint)\n\n${sketchScope.trim()}\n\n` +
+            `Treat this as the authoritative boundary for the slice. Do not plan work outside this scope; if the scope is too narrow, surface it as a deviation rather than expanding silently.`);
+    }
+    return renderSlicePrompt({
+        mid, sid, sTitle, base,
+        level: level ?? resolveInlineLevel(),
+        promptTemplate: "refine-slice",
+        prependBlocks,
+        extraVars: { sketchScope },
+        sessionContextWindow: options?.sessionContextWindow,
+        modelRegistry: options?.modelRegistry,
     });
 }
 export async function buildExecuteTaskPrompt(mid, sid, sTitle, tid, tTitle, base, level) {
@@ -1218,7 +1329,7 @@ export async function buildExecuteTaskPrompt(mid, sid, sTitle, tid, tTitle, base
     const overridesSection = formatOverridesSection(activeOverrides);
     // Compute verification budget for the executor's context window (issue #707)
     const prefs = loadEffectiveGSDPreferences();
-    const contextWindow = resolveExecutorContextWindow(undefined, prefs?.preferences);
+    const contextWindow = resolveExecutorContextWindow(opts.modelRegistry, prefs?.preferences, opts.sessionContextWindow);
     const budgets = computeBudgets(contextWindow);
     const verificationBudget = `~${Math.round(budgets.verificationBudgetChars / 1000)}K chars`;
     // Truncate carry-forward section when it exceeds 40% of inline context budget.
@@ -1233,7 +1344,28 @@ export async function buildExecuteTaskPrompt(mid, sid, sTitle, tid, tTitle, base
     const runtimeContext = runtimeContent
         ? `### Runtime Context\nSource: \`.gsd/RUNTIME.md\`\n\n${runtimeContent.trim()}`
         : "";
-    const phaseAnchorSection = planAnchor ? formatAnchorForPrompt(planAnchor) : "";
+    let phaseAnchorSection = planAnchor ? formatAnchorForPrompt(planAnchor) : "";
+    // ADR-011 Phase 2: inject any resolved-but-unapplied escalation override
+    // into this task's prompt. Claim is atomic via DB UPDATE WHERE IS NULL, so
+    // if a parallel build already injected it, we skip. Feature-gated by
+    // phases.mid_execution_escalation. Prepended to phaseAnchorSection so it
+    // appears near the top of the prompt above planning anchors.
+    if (prefs?.preferences?.phases?.mid_execution_escalation === true) {
+        try {
+            const { claimOverrideForInjection } = await import("./escalation.js");
+            const claimed = claimOverrideForInjection(base, mid, sid);
+            if (claimed) {
+                const block = claimed.injectionBlock + "\n\n---\n\n";
+                phaseAnchorSection = phaseAnchorSection
+                    ? `${block}${phaseAnchorSection}`
+                    : block;
+            }
+        }
+        catch (escalationErr) {
+            // Escalation module unavailable or threw — log and proceed.
+            logWarning("prompt", `escalation override injection failed: ${escalationErr.message}`);
+        }
+    }
     // Task-scoped gates owned by execute-task (Q5/Q6/Q7). Pull only the
     // gates that plan-slice actually seeded for this task — tasks with no
     // external dependencies legitimately skip Q5, tasks with no runtime
@@ -1398,6 +1530,13 @@ export async function buildCompleteMilestonePrompt(mid, midTitle, base, level) {
     inlined.push(inlineTemplate("milestone-summary", "Milestone Summary"));
     const inlinedContext = capPreamble(`## Inlined Context (preloaded — do not re-read these files)\n\n${inlined.join("\n\n---\n\n")}`);
     const milestoneSummaryPath = join(base, `${relMilestonePath(base, mid)}/${mid}-SUMMARY.md`);
+    const learningsRelPath = join(relMilestonePath(base, mid), `${mid}-LEARNINGS.md`);
+    const learningsAbsPath = join(base, learningsRelPath);
+    const extractLearningsSteps = buildExtractionStepsBlock({
+        milestoneId: mid,
+        outputPath: learningsAbsPath,
+        relativeOutputPath: learningsRelPath,
+    });
     return loadPrompt("complete-milestone", {
         workingDirectory: base,
         milestoneId: mid,
@@ -1405,6 +1544,7 @@ export async function buildCompleteMilestonePrompt(mid, midTitle, base, level) {
         roadmapPath: roadmapRel,
         inlinedContext,
         milestoneSummaryPath,
+        extractLearningsSteps,
         skillActivation: buildSkillActivationBlock({
             base,
             milestoneId: mid,
@@ -1724,7 +1864,7 @@ export async function buildReassessRoadmapPrompt(mid, midTitle, completedSliceId
     });
 }
 // ─── Reactive Execute Prompt ──────────────────────────────────────────────
-export async function buildReactiveExecutePrompt(mid, midTitle, sid, sTitle, readyTaskIds, base, subagentModel) {
+export async function buildReactiveExecutePrompt(mid, midTitle, sid, sTitle, readyTaskIds, base, subagentModel, opts) {
     const { loadSliceTaskIO, deriveTaskGraph, graphMetrics } = await import("./reactive-graph.js");
     // Build graph for context
     const taskIO = await loadSliceTaskIO(base, mid, sid);
@@ -1755,7 +1895,11 @@ export async function buildReactiveExecutePrompt(mid, midTitle, sid, sTitle, rea
         // Build dependency-scoped carry-forward paths for this task
         const depPaths = await getDependencyTaskSummaryPaths(mid, sid, tid, node?.dependsOn ?? [], base);
         // Build a full execute-task prompt with dependency-based carry-forward
-        const taskPrompt = await buildExecuteTaskPrompt(mid, sid, sTitle, tid, tTitle, base, { carryForwardPaths: depPaths });
+        const taskPrompt = await buildExecuteTaskPrompt(mid, sid, sTitle, tid, tTitle, base, {
+            carryForwardPaths: depPaths,
+            sessionContextWindow: opts?.sessionContextWindow,
+            modelRegistry: opts?.modelRegistry,
+        });
         const modelSuffix = subagentModel ? ` with model: "${subagentModel}"` : "";
         subagentSections.push([
             `### ${tid}: ${tTitle}`,

package/dist/resources/extensions/gsd/auto-recovery.js

@@ -8,6 +8,7 @@
  */
 import { parseUnitId } from "./unit-id.js";
 import { appendEvent } from "./workflow-events.js";
+import { atomicWriteSync } from "./atomic-write.js";
 import { clearParseCache } from "./files.js";
 import { parseRoadmap as parseLegacyRoadmap, parsePlan as parseLegacyPlan } from "./parsers-legacy.js";
 import { isDbAvailable, getTask, getSlice, getSliceTasks, getPendingGates, updateTaskStatus, updateSliceStatus } from "./gsd-db.js";
@@ -212,6 +213,47 @@ export function verifyExpectedArtifact(unitType, unitId, base) {
         }
         return true;
     }
+    // #4414: research-slice parallel-research sentinel. The unitId
+    // `{mid}/parallel-research` is not a real slice — it triggers a single agent
+    // that fans out research across multiple slices. Verify success by checking
+    // that every slice which was "research-ready" in the roadmap now has a
+    // RESEARCH file. Without this, resolveExpectedArtifactPath returns null and
+    // the retry/escalation machinery silently re-dispatches forever.
+    //
+    // NOTE: this predicate mirrors the dispatch rule at
+    // auto-dispatch.ts parallel-research-slices — keep the two in sync.
+    if (unitType === "research-slice" && unitId.endsWith("/parallel-research")) {
+        const { milestone: mid } = parseUnitId(unitId);
+        if (!mid)
+            return false;
+        const roadmapFile = resolveMilestoneFile(base, mid, "ROADMAP");
+        if (!roadmapFile || !existsSync(roadmapFile)) {
+            logWarning("recovery", `verify-fail ${unitType} ${unitId}: roadmap missing`);
+            return false;
+        }
+        try {
+            const roadmap = parseLegacyRoadmap(readFileSync(roadmapFile, "utf-8"));
+            const milestoneResearchFile = resolveMilestoneFile(base, mid, "RESEARCH");
+            for (const slice of roadmap.slices) {
+                if (slice.done)
+                    continue;
+                if (milestoneResearchFile && slice.id === "S01")
+                    continue;
+                const depsComplete = (slice.depends ?? []).every((depId) => !!resolveSliceFile(base, mid, depId, "SUMMARY"));
+                if (!depsComplete)
+                    continue;
+                if (!resolveSliceFile(base, mid, slice.id, "RESEARCH")) {
+                    logWarning("recovery", `verify-fail ${unitType} ${unitId}: slice ${slice.id} missing RESEARCH`);
+                    return false;
+                }
+            }
+            return true;
+        }
+        catch (err) {
+            logWarning("recovery", `parallel-research verification failed: ${err instanceof Error ? err.message : String(err)}`);
+            return false;
+        }
+    }
     const absPath = resolveExpectedArtifactPath(unitType, unitId, base);
     // For unit types with no verifiable artifact (null path), the parent directory
     // is missing on disk — treat as stale completion state so the key gets evicted (#313).
@@ -405,6 +447,13 @@ export function writeBlockerPlaceholder(unitType, unitId, base, reason) {
         `Review and replace this file before relying on downstream artifacts.`,
     ].join("\n");
     writeFileSync(absPath, content, "utf-8");
+    // #4414: Clear caches so subsequent dispatch guards (e.g.
+    // resolveMilestoneFile) see the placeholder file. Without this, the
+    // cached directory listing is stale and the dispatch rule re-fires,
+    // producing an infinite loop despite the placeholder being on disk.
+    // Matches the pattern used in verifyExpectedArtifact above.
+    clearPathCache();
+    clearParseCache();
     // Mark the task/slice as complete in the DB so verifyExpectedArtifact passes.
     // Without this, the DB status stays "pending" and the dispatch loop
     // re-derives the same unit indefinitely (#2531, #2653).
@@ -414,6 +463,14 @@ export function writeBlockerPlaceholder(unitType, unitId, base, reason) {
         if (unitType === "execute-task" && mid && sid && tid) {
             try {
                 updateTaskStatus(mid, sid, tid, "complete", ts);
+                const planPath = resolveSliceFile(base, mid, sid, "PLAN");
+                if (planPath && existsSync(planPath)) {
+                    const planContent = readFileSync(planPath, "utf-8");
+                    const updatedPlan = planContent.replace(new RegExp(`^(\\s*-\\s+)\\[ \\]\\s+\\*\\*${tid}:`, "m"), `$1[x] **${tid}:`);
+                    if (updatedPlan !== planContent) {
+                        atomicWriteSync(planPath, updatedPlan);
+                    }
+                }
             }
             catch (e) {
                 logWarning("recovery", `updateTaskStatus failed during context exhaustion: ${e instanceof Error ? e.message : String(e)}`);

package/dist/resources/extensions/gsd/auto-start.js

@@ -30,7 +30,7 @@ import { initRoutingHistory } from "./routing-history.js";
 import { restoreHookState, resetHookState } from "./post-unit-hooks.js";
 import { resetProactiveHealing, setLevelChangeCallback } from "./doctor-proactive.js";
 import { snapshotSkills } from "./skill-discovery.js";
-import { isDbAvailable, getMilestone, openDatabase } from "./gsd-db.js";
+import { isDbAvailable, getMilestone, openDatabase, getDbStatus } from "./gsd-db.js";
 import { hideFooter } from "./auto-dashboard.js";
 import { debugLog, enableDebug, isDebugEnabled, getDebugLogPath, } from "./debug-logger.js";
 import { logWarning, logError } from "./workflow-logger.js";
@@ -190,8 +190,8 @@ export async function bootstrapAutoSession(s, ctx, pi, base, verboseMode, reques
     //
     // Precedence:
     // 1) Explicit session override via /gsd model (this session)
-    // 2) GSD model preferences from PREFERENCES.md (validated against live auth)
-    // 3) Current session model from settings/session restore (if provider ready)
+    // 2) Current session model from settings/session restore (if provider ready)
+    // 3) GSD model preferences from PREFERENCES.md (validated against live auth)
     //
     // This preserves #3517 defaults while honoring explicit runtime model
     // selection for subsequent /gsd runs in the same session.
@@ -224,11 +224,14 @@ export async function bootstrapAutoSession(s, ctx, pi, base, verboseMode, reques
         }
     }
     const sessionModelReady = ctx.model && ctx.modelRegistry.isProviderRequestReady(ctx.model.provider);
+    const currentSessionModel = (sessionModelReady && ctx.model)
+        ? { provider: ctx.model.provider, id: ctx.model.id }
+        : null;
+    const startThinkingSnapshot = pi.getThinkingLevel();
     const startModelSnapshot = manualSessionOverride
+        ?? currentSessionModel
         ?? validatedPreferredModel
-        ?? (sessionModelReady && ctx.model
-            ? { provider: ctx.model.provider, id: ctx.model.id }
-            : null);
+        ?? null;
     try {
         // Validate GSD_PROJECT_ID early so the user gets immediate feedback
         const customProjectId = process.env.GSD_PROJECT_ID;
@@ -348,10 +351,11 @@ export async function bootstrapAutoSession(s, ctx, pi, base, verboseMode, reques
         // Detect survivor milestone branches in both pre-planning and complete phases.
         // In phase=complete, the milestone artifacts exist but finalization (merge,
         // worktree cleanup) was never run — the survivor branch must be merged.
+        // Applies to both worktree and branch isolation modes.
         let hasSurvivorBranch = false;
         if (state.activeMilestone &&
             (state.phase === "pre-planning" || state.phase === "complete") &&
-            shouldUseWorktreeIsolation() &&
+            getIsolationMode() !== "none" &&
             !detectWorktreeName(base) &&
             !base.includes(`${pathSep}.gsd${pathSep}worktrees${pathSep}`)) {
             const milestoneBranch = `milestone/${state.activeMilestone.id}`;
@@ -505,8 +509,9 @@ export async function bootstrapAutoSession(s, ctx, pi, base, verboseMode, reques
         s.pendingQuickTasks = [];
         s.currentUnit = null;
         s.currentMilestoneId = state.activeMilestone?.id ?? null;
-        s.originalModelId = ctx.model?.id ?? null;
-        s.originalModelProvider = ctx.model?.provider ?? null;
+        s.originalModelId = startModelSnapshot?.id ?? ctx.model?.id ?? null;
+        s.originalModelProvider = startModelSnapshot?.provider ?? ctx.model?.provider ?? null;
+        s.originalThinkingLevel = startThinkingSnapshot ?? null;
         // Register SIGTERM handler
         registerSigtermHandler(base);
         // Capture integration branch
@@ -547,7 +552,7 @@ export async function bootstrapAutoSession(s, ctx, pi, base, verboseMode, reques
             return symlinkRe.test(p);
         };
         if (s.currentMilestoneId &&
-            shouldUseWorktreeIsolation() &&
+            getIsolationMode() !== "none" &&
             !detectWorktreeName(base) &&
             !isUnderGsdWorktrees(base)) {
             buildResolver().enterMilestone(s.currentMilestoneId, {
@@ -592,8 +597,21 @@ export async function bootstrapAutoSession(s, ctx, pi, base, verboseMode, reques
         // call returns "db_unavailable", triggering artifact-retry which
         // re-dispatches the same task — producing an infinite loop (#2419).
         if (existsSync(gsdDbPath) && !isDbAvailable()) {
-            ctx.ui.notify("SQLite database exists but failed to open. Auto-mode cannot proceed without a working database provider. " +
-                "Check for corrupt gsd.db or missing native SQLite bindings.", "error");
+            const dbStatus = getDbStatus();
+            const phaseHint = dbStatus.lastPhase === "open"
+                ? "The database file could not be opened"
+                : dbStatus.lastPhase === "initSchema"
+                    ? "The database schema could not be initialized"
+                    : dbStatus.lastPhase === "vacuum-recovery"
+                        ? "Corruption recovery (VACUUM) failed"
+                        : dbStatus.attempted
+                            ? "The database could not be opened (phase unknown)"
+                            : "The database provider could not be loaded";
+            const errorDetail = dbStatus.lastError ? ` (${dbStatus.lastError.message})` : "";
+            const providerHint = dbStatus.provider
+                ? ` Provider: ${dbStatus.provider}.`
+                : " No SQLite provider available — check Node >= 22 or install better-sqlite3.";
+            ctx.ui.notify(`SQLite database exists but failed to open: ${gsdDbPath}. ${phaseHint}${errorDetail}.${providerHint}`, "error");
             return releaseLockAndReturn();
         }
         // Initialize metrics
@@ -607,6 +625,7 @@ export async function bootstrapAutoSession(s, ctx, pi, base, verboseMode, reques
                 id: startModelSnapshot.id,
             };
         }
+        s.autoModeStartThinkingLevel = startThinkingSnapshot ?? null;
         s.manualSessionModelOverride = manualSessionOverride ?? null;
         // Apply worker model override from parallel orchestrator (#worker-model).
         // GSD_WORKER_MODEL is injected by the coordinator when parallel.worker_model
@@ -658,7 +677,8 @@ export async function bootstrapAutoSession(s, ctx, pi, base, verboseMode, reques
         const bannerPrefs = loadEffectiveGSDPreferences()?.preferences;
         const effectiveProvider = s.autoModeStartModel?.provider ?? ctx.model?.provider;
         const effectivelyEnabled = routingConfig.enabled
-            && !(effectiveProvider && isFlatRateProvider(effectiveProvider, buildFlatRateContext(effectiveProvider, ctx, bannerPrefs)));
+            && (routingConfig.allow_flat_rate_providers
+                || !(effectiveProvider && isFlatRateProvider(effectiveProvider, buildFlatRateContext(effectiveProvider, ctx, bannerPrefs))));
         // The actual ceiling may come from tier_models.heavy, not the start model.
         const effectiveCeiling = (routingConfig.enabled && routingConfig.tier_models?.heavy)
             ? routingConfig.tier_models.heavy

package/dist/resources/extensions/gsd/auto-verification.js

@@ -5,9 +5,9 @@
  * dependency audits, handles auto-fix retry logic, and writes
  * verification evidence JSON.
  *
- * Extracted from handleAgentEnd() in auto.ts. Returns a sentinel
- * value instead of calling return/pauseAuto directly — the caller
- * checks the result and handles control flow.
+ * Extracted from the pre-loop agent_end handler in auto.ts. Returns a
+ * sentinel value instead of calling return/pauseAuto directly — the
+ * caller checks the result and handles control flow.
  */
 import { mkdirSync, writeFileSync } from "node:fs";
 import { resolveSlicePath, resolveMilestoneFile } from "./paths.js";
@@ -396,6 +396,39 @@ export async function runPostUnitVerification(vctx, pauseAuto) {
         if (postExecBlockingFailure) {
             result.passed = false;
         }
+        // Emit Layer 2 verify_result event with the final, post-exec verdict so hooks
+        // see the authoritative pass/fail and the complete set of failures.
+        try {
+            const { emitVerifyResult } = await import("./hook-emitter.js");
+            const checkFailures = result.checks
+                .filter((c) => c.exitCode !== 0)
+                .map((c) => ({
+                kind: "gate",
+                message: `${c.command} exited ${c.exitCode}${c.stderr ? `: ${c.stderr.slice(0, 200)}` : ""}`,
+            }));
+            const runtimeFailures = (result.runtimeErrors ?? [])
+                .filter((e) => e.blocking)
+                .map((e) => ({
+                kind: "other",
+                message: `[${e.source}] ${e.message.slice(0, 200)}`,
+            }));
+            const postExecFailures = (postExecChecks ?? [])
+                .filter((c) => !c.passed)
+                .map((c) => ({
+                kind: "other",
+                message: `[${c.category}] ${c.target}: ${c.message}`,
+            }));
+            await emitVerifyResult({
+                passed: result.passed,
+                failures: [...checkFailures, ...runtimeFailures, ...postExecFailures],
+                unitType: s.currentUnit.type,
+                unitId: s.currentUnit.id,
+                cwd: s.basePath,
+            });
+        }
+        catch (hookErr) {
+            logWarning("engine", `verify_result hook emission failed: ${hookErr.message}`);
+        }
         // ── Auto-fix retry logic ──
         if (result.passed) {
             s.verificationRetryCount.delete(s.currentUnit.id);