gsd-pi 2.76.0 → 2.77.0-dev.1594b263e

package/dist/resources/extensions/gsd/prompts/discuss-headless.md

@@ -162,7 +162,20 @@ Preserve the specification's exact terminology, emphasis, and specific framing.
 6. For each architectural or pattern decision, call `gsd_decision_save` — the tool auto-assigns IDs and regenerates `.gsd/DECISIONS.md` automatically.
 7. {{commitInstruction}}
 
-After writing the files, say exactly: "Milestone {{milestoneId}} ready." — nothing else. Auto-mode will start automatically.
+### Ready-phrase pre-condition (NON-BYPASSABLE)
+
+Before emitting the ready phrase, verify in the CURRENT turn that you have:
+
+- [ ] Written `.gsd/PROJECT.md` (step 2)
+- [ ] Written `.gsd/REQUIREMENTS.md` (step 3)
+- [ ] Written `{{contextPath}}` (step 4)
+- [ ] Called `gsd_plan_milestone` (step 5)
+
+If ANY box is unchecked, **STOP**. Do NOT emit the ready phrase. Emit the missing tool calls in this same turn. The system detects missing artifacts and will reject premature ready signals — you will be asked again and retries are capped.
+
+Do not announce the ready phrase as something you are "about to" do. Do not narrate "now writing the files" as a substitute for actually writing them. The ready phrase is a post-write signal, not an intent signal.
+
+After completing steps 1–7 above, say exactly: "Milestone {{milestoneId}} ready." — nothing else. Auto-mode will start automatically.
 
 ### Multi-Milestone
 
@@ -234,7 +247,21 @@ For single-milestone projects, do NOT write this file.
 
 7. {{multiMilestoneCommitInstruction}}
 
-After writing the files, say exactly: "Milestone {{milestoneId}} ready." — nothing else. Auto-mode will start automatically.
+### Ready-phrase pre-condition (NON-BYPASSABLE)
+
+Before emitting the ready phrase, verify in the CURRENT turn that you have:
+
+- [ ] Written `.gsd/PROJECT.md`
+- [ ] Written `.gsd/REQUIREMENTS.md`
+- [ ] Written the primary milestone `CONTEXT.md`
+- [ ] Called `gsd_plan_milestone` for the primary milestone
+- [ ] Written `.gsd/DISCUSSION-MANIFEST.json` with `gates_completed === total`
+
+If ANY box is unchecked, **STOP**. Do NOT emit the ready phrase. Emit the missing tool calls in this same turn. The system detects missing artifacts and will reject premature ready signals — you will be asked again and retries are capped.
+
+Do not announce the ready phrase as something you are "about to" do. Do not narrate "now writing the files" as a substitute for actually writing them. The ready phrase is a post-write signal, not an intent signal.
+
+After completing every step above, say exactly: "Milestone {{milestoneId}} ready." — nothing else. Auto-mode will start automatically.
 
 ## Critical Rules
 

package/dist/resources/extensions/gsd/prompts/discuss.md

@@ -339,7 +339,20 @@ These sections are in addition to whatever other context the discussion surfaced
 6. For each architectural or pattern decision made during discussion, call `gsd_decision_save` — the tool auto-assigns IDs and regenerates `.gsd/DECISIONS.md` automatically.
 7. {{commitInstruction}}
 
-After writing the files, say exactly: "Milestone {{milestoneId}} ready." — nothing else. Auto-mode will start automatically.
+### Ready-phrase pre-condition (NON-BYPASSABLE)
+
+Before emitting the ready phrase, verify in the CURRENT turn that you have:
+
+- [ ] Written `.gsd/PROJECT.md` (step 2)
+- [ ] Written `.gsd/REQUIREMENTS.md` (step 3)
+- [ ] Written `{{contextPath}}` (step 4)
+- [ ] Called `gsd_plan_milestone` (step 5)
+
+If ANY box is unchecked, **STOP**. Do NOT emit the ready phrase. Emit the missing tool calls in this same turn. The system detects missing artifacts and will reject premature ready signals — you will be asked again and retries are capped.
+
+Do not announce the ready phrase as something you are "about to" do. Do not narrate "now writing the files" as a substitute for actually writing them. The ready phrase is a post-write signal, not an intent signal.
+
+After completing steps 1–7 above, say exactly: "Milestone {{milestoneId}} ready." — nothing else. Auto-mode will start automatically.
 
 ### Multi-Milestone
 
@@ -418,6 +431,20 @@ For single-milestone projects, do NOT write this file — it is only for multi-m
 
 7. {{multiMilestoneCommitInstruction}}
 
-After writing the files, say exactly: "Milestone M001 ready." — nothing else. Auto-mode will start automatically.
+### Ready-phrase pre-condition (NON-BYPASSABLE)
+
+Before emitting the ready phrase, verify in the CURRENT turn that you have:
+
+- [ ] Written `.gsd/PROJECT.md` (Phase 1)
+- [ ] Written `.gsd/REQUIREMENTS.md` (Phase 1)
+- [ ] Written primary-milestone `CONTEXT.md` (Phase 2)
+- [ ] Called `gsd_plan_milestone` for the primary milestone (Phase 2)
+- [ ] Written `.gsd/DISCUSSION-MANIFEST.json` with `gates_completed === total` (Phase 3)
+
+If ANY box is unchecked, **STOP**. Do NOT emit the ready phrase. Emit the missing tool calls in this same turn. The system detects missing artifacts and will reject premature ready signals — you will be asked again and retries are capped.
+
+Do not announce the ready phrase as something you are "about to" do. Do not narrate "now writing the files" as a substitute for actually writing them. The ready phrase is a post-write signal, not an intent signal.
+
+After completing all phases above, say exactly: "Milestone M001 ready." — nothing else. Auto-mode will start automatically.
 
 {{inlinedTemplates}}

package/dist/resources/extensions/gsd/prompts/doctor-heal.md

@@ -7,9 +7,10 @@ Rules:
 2. Read before edit.
 3. Prefer fixing authoritative artifacts over masking warnings.
 4. For missing summaries or UAT files, generate the real artifact from existing slice/task context when possible — do not leave placeholders if you can reconstruct the real content.
-5. After each repair cluster, verify the relevant invariant directly from disk.
-6. When done, rerun `/gsd doctor {{doctorCommandSuffix}}` mentally by ensuring the remaining issue set for this scope is reduced or cleared.
-7. Do NOT query `.gsd/gsd.db` directly via `sqlite3` or `node -e require('better-sqlite3')` — use `gsd_milestone_status` to inspect DB state. Direct access bypasses the WAL connection owned by the engine and can corrupt in-flight writes.
+5. For a missing milestone `CONTEXT.md` when the milestone is already past `pre-planning` (phase is `executing`, `summarizing`, `validating-milestone`, or `completing-milestone`): the artifact was skipped during bootstrap and must be reconstructed before execution can resume. Read `PROJECT.md`, `REQUIREMENTS.md`, the milestone's `ROADMAP.md`, and any slice-level context on disk, then write `.gsd/milestones/<milestone-id>/<milestone-id>-CONTEXT.md` with the real context for the scoped milestone. Do not leave a stub — the plan gate will reject it on the next cycle.
+6. After each repair cluster, verify the relevant invariant directly from disk.
+7. When done, rerun `/gsd doctor {{doctorCommandSuffix}}` mentally by ensuring the remaining issue set for this scope is reduced or cleared.
+8. Do NOT query `.gsd/gsd.db` directly via `sqlite3` or `node -e require('better-sqlite3')` — use `gsd_milestone_status` to inspect DB state. Direct access bypasses the WAL connection owned by the engine and can corrupt in-flight writes.
 
 ## Doctor Summary
 
@@ -27,4 +28,4 @@ Then:
 - Repair the unresolved issues in scope
 - Keep changes minimal and targeted
 - If unresolved issues remain outside scope, leave them untouched and mention them briefly
-- End with: "GSD doctor heal complete."
+- End with: "GSD doctor heal complete."

package/dist/resources/extensions/gsd/prompts/execute-task.md

@@ -32,6 +32,7 @@ A researcher explored the codebase and a planner decomposed the work — you are
 
 Then:
 0. Narrate step transitions, key implementation decisions, and verification outcomes as you work. Keep it terse — one line between tool-call clusters, not between every call — but write complete sentences in user-facing prose, not shorthand notes or scratchpad fragments.
+0a. Call `memory_query` with 2-4 keywords drawn from the task title and the files it touches. Surface any prior gotchas, conventions, or architectural decisions before you start changing code. Skip only if the task is trivially mechanical.
 1. {{skillActivation}} Follow any activated skills before writing code. If no skills match this task, skip this step.
 2. Execute the steps in the inlined task plan, adapting minor local mismatches when the surrounding code differs from the planner's snapshot
 3. Before any `Write` that creates an artifact or output file, check whether that path already exists. If it does, read it first and decide whether the work is already done, should be extended, or truly needs replacement. "Create" in the plan does **not** mean the file is missing — a prior session may already have started it.
@@ -81,8 +82,8 @@ Then:
     **Scope:** Escalation is instrumented only in `execute-task`. Refine-slice escalation is deferred. Reactive-execute batches run to completion before escalations are surfaced — the dispatch pause happens on the next loop iteration, not mid-batch.
 
     The `escalation` payload is ignored unless `phases.mid_execution_escalation` is enabled; populate it anyway for audit logs.
-17. If you made an architectural, pattern, library, or observability decision during this task that downstream work should know about, append it to `.gsd/DECISIONS.md` (read the template at `~/.gsd/agent/extensions/gsd/templates/decisions.md` if the file doesn't exist yet). Not every task produces decisions — only append when a meaningful choice was made.
-18. If you discover a non-obvious rule, recurring gotcha, or useful pattern during execution, append it to `.gsd/KNOWLEDGE.md`. Only add entries that would save future agents from repeating your investigation. Don't add obvious things.
+17. If you made an architectural, pattern, library, or observability decision during this task that downstream work should know about, call `capture_thought` with `category: "architecture"` (or `"pattern"`). For decisions, populate `structuredFields` with `{ scope, decision, choice, rationale, made_by: "agent", revisable }` so future projection back to a human-visible decisions register stays lossless. Not every task produces decisions — only capture when a meaningful choice was made.
+18. If you discover a non-obvious rule, recurring gotcha, or useful pattern during execution, call `capture_thought` with `category: "gotcha"`, `"convention"`, `"pattern"`, or `"environment"` as appropriate. Only capture entries that would save future agents from repeating your investigation — don't capture obvious things. The memory store is the single source of truth for cross-session knowledge (ADR-013); do not append to `.gsd/DECISIONS.md` or `.gsd/KNOWLEDGE.md` directly.
 19. Read the template at `~/.gsd/agent/extensions/gsd/templates/task-summary.md`
 20. Use that template to prepare the completion content you will pass to `gsd_complete_task` using the camelCase fields `milestoneId`, `sliceId`, `taskId`, `oneLiner`, `narrative`, `verification`, and `verificationEvidence`. Do **not** manually write `{{taskSummaryPath}}` — the DB-backed tool is the canonical write path and renders the summary file for you.
 21. Call `gsd_complete_task` with milestoneId, sliceId, taskId, and the completion fields derived from the template. This is your final required step — do NOT manually edit PLAN.md checkboxes. The tool marks the task complete, updates the DB, renders `{{taskSummaryPath}}`, and updates PLAN.md automatically.

package/dist/resources/extensions/gsd/prompts/parallel-research-slices.md

@@ -15,8 +15,11 @@ Dispatch ALL slices simultaneously using the `subagent` tool in **parallel mode*
 1. Call `subagent` with `tasks: [...]` containing one entry per slice below
 2. Wait for ALL subagents to complete
 3. Verify each slice's RESEARCH file was written (check the `.gsd/{{mid}}/` directory)
-4. If any subagent failed to write its RESEARCH file, re-run it individually
-5. Report which slices completed research and which (if any) failed
+4. If a subagent failed to write its RESEARCH file, retry it **once** individually
+5. If it fails a second time, write a partial RESEARCH file for that slice with a `## BLOCKER` section explaining the failure — do NOT retry again
+6. Report which slices completed research and which (if any) needed a blocker note
+
+**Important**: Each failed slice gets exactly one retry. After that, write the blocker and move on. Never retry the same slice more than once.
 
 ## Subagent Prompts
 

package/dist/resources/extensions/gsd/prompts/plan-slice.md

@@ -20,9 +20,22 @@ Pay particular attention to **Forward Intelligence** sections — they contain h
 
 You have full tool access. Before decomposing, explore the relevant code to ground your plan in reality.
 
-### Verify Roadmap Assumptions
+### Verify Roadmap Assumptions (JIT Reassessment — ADR-003 §4)
 
-Check prior slice summaries (inlined above as dependency summaries, if present). If prior slices discovered constraints, changed approaches, or flagged fragility, adjust your plan accordingly. The roadmap description may be stale — verify it against the current codebase state.
+Before planning this slice, verify that the roadmap's assumptions still hold given prior slice summaries. Check inlined dependency summaries (below) for discovered constraints, changed approaches, or flagged fragility.
+
+**If the remaining roadmap needs adjustment, modify it before proceeding:**
+
+- If a downstream slice's title/demo/dependencies are now wrong, call `gsd_reassess_roadmap` with the corrected `sliceChanges.modified` entry.
+- If new work surfaced that deserves its own slice, add it via `sliceChanges.added`.
+- If a downstream slice is now redundant or out of scope, remove it via `sliceChanges.removed`.
+- **Bias strongly toward "roadmap is fine."** Most slice completions produce no structural change. Only adjust when there is concrete evidence a downstream slice is wrong — not speculative concern. Over-reassessment is costlier than a later mid-slice replan.
+
+Completed slices are immutable: never modify or remove a slice whose status is complete.
+
+Then proceed with planning this slice against the (possibly updated) roadmap.
+
+The roadmap description may be stale — verify it against the current codebase state.
 
 ### Explore Slice Scope
 
@@ -44,6 +57,7 @@ Narrate your decomposition reasoning — why you're grouping work this way, what
 
 Then:
 0. If `REQUIREMENTS.md` was preloaded above, identify which Active requirements the roadmap says this slice owns or supports. These are the requirements this plan must deliver — every owned requirement needs at least one task that directly advances it, and verification must prove the requirement is met.
+0a. Call `memory_query` with keywords from the slice title and the source files listed below. Prior architectural decisions, conventions, and gotchas in this area should inform task decomposition — not be re-derived during execution.
 1. Read the templates:
    - `~/.gsd/agent/extensions/gsd/templates/plan.md`
    - `~/.gsd/agent/extensions/gsd/templates/task-plan.md`

package/dist/resources/extensions/gsd/prompts/research-slice.md

@@ -42,6 +42,7 @@ An honest "this is straightforward, here's the pattern to follow" is more valuab
 
 Research what this slice needs. Narrate key findings and surprises as you go — what exists, what's missing, what constrains the approach.
 0. If `REQUIREMENTS.md` was preloaded above, identify which Active requirements this slice owns or supports. Research should target these requirements — surfacing risks, unknowns, and implementation constraints that could affect whether the slice actually delivers them.
+0a. Call `memory_query` with keywords from the slice title and scope. Prior slices may have captured architecture notes, conventions, or gotchas in subsystems you're about to explore — pulling them now prevents re-discovering known constraints.
 1. {{skillActivation}} Reference specific rules from loaded skills in your findings where they inform the implementation approach.
 2. **Skill Discovery ({{skillDiscoveryMode}}):**{{skillDiscoveryInstructions}}
 3. Explore relevant code for this slice's scope. For targeted exploration, use `rg`, `find`, and reads. For broad or unfamiliar subsystems, use `scout` to map the relevant area first.

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

@@ -213,6 +213,7 @@ Fix the root cause, not symptoms. When applying a temporary mitigation, label it
 - State uncertainty plainly: "Not sure this handles X - testing it." No performed confidence, no hedging paragraphs.
 - All user-visible narration must be grammatical English. Do not emit compressed planner notes like "Need inspect X" or "Maybe read Y first". If it would look acceptable in a commit comment or standup note, it's acceptable here.
 - When debugging, stay curious. Problems are puzzles. Say what's interesting about the failure before reaching for fixes.
+- After completing a task, give a brief completion summary and present 2-4 contextual next-step options as a numbered list (the last option is always "Other"). This reduces cognitive load by letting the user pick rather than formulate what's next. Omit the numbered list when the response must follow a strict output format (JSON, patches, commit messages, structured data).
 
 Good narration: "Three existing handlers follow a middleware pattern - using that instead of a custom wrapper."
 Good narration: "Tests pass. Running slice-level verification."

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

@@ -13,8 +13,9 @@
  * Auto-triggered: after each milestone completion (when auto_report: true).
  * Manual: /gsd export --html
  */
-import { writeFileSync, readFileSync, mkdirSync, existsSync } from 'node:fs';
+import { readFileSync, mkdirSync, existsSync } from 'node:fs';
 import { join } from 'node:path';
+import { atomicWriteSync } from './atomic-write.js';
 import { gsdRoot } from './paths.js';
 import { formatCost, formatTokenCount } from './metrics.js';
 import { formatDateShort, formatDuration } from '../shared/format-utils.js';
@@ -43,7 +44,7 @@ export function loadReportsIndex(basePath) {
 function saveReportsIndex(basePath, index) {
     const dir = reportsDir(basePath);
     mkdirSync(dir, { recursive: true });
-    writeFileSync(reportsIndexPath(basePath), JSON.stringify(index, null, 2) + '\n', 'utf-8');
+    atomicWriteSync(reportsIndexPath(basePath), JSON.stringify(index, null, 2) + '\n', 'utf-8');
 }
 /**
  * Write a report snapshot to .gsd/reports/, update reports.json, regenerate index.html.
@@ -56,7 +57,7 @@ export function writeReportSnapshot(args) {
     const prefix = args.milestoneId === 'final' ? 'final' : args.milestoneId;
     const filename = `${prefix}-${timestamp}.html`;
     const filePath = join(dir, filename);
-    writeFileSync(filePath, args.html, 'utf-8');
+    atomicWriteSync(filePath, args.html, 'utf-8');
     // Load or init registry
     const existing = loadReportsIndex(args.basePath);
     const index = existing ?? {
@@ -97,7 +98,7 @@ export function writeReportSnapshot(args) {
 // ─── HTML Index Generator ─────────────────────────────────────────────────────
 export function regenerateHtmlIndex(basePath, index) {
     const html = buildIndexHtml(index);
-    writeFileSync(reportsHtmlIndexPath(basePath), html, 'utf-8');
+    atomicWriteSync(reportsHtmlIndexPath(basePath), html, 'utf-8');
 }
 function buildIndexHtml(index) {
     const { projectName, projectPath, gsdVersion, entries } = index;

package/dist/resources/extensions/gsd/safety/evidence-collector.js

@@ -3,9 +3,17 @@
  * Tracks every bash command, file write, and file edit during a unit execution.
  * Evidence is compared against LLM completion claims in evidence-cross-ref.ts.
  *
+ * Evidence is persisted to .gsd/safety/evidence-<mid>-<sid>-<tid>.json so it
+ * survives session restarts (pause/resume, crash recovery). On unit start,
+ * call resetEvidence() then loadEvidenceFromDisk(). On every new tool call,
+ * saveEvidenceToDisk() is called automatically by recordToolCall/recordToolResult.
+ *
  * Follows the same module-level Map pattern as auto-tool-tracking.ts.
  * Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
  */
+import { existsSync, mkdirSync, readFileSync, writeFileSync, renameSync, unlinkSync, } from "node:fs";
+import { join, dirname } from "node:path";
+import { randomBytes } from "node:crypto";
 // ─── Module State ───────────────────────────────────────────────────────────
 let unitEvidence = [];
 // ─── Public API ─────────────────────────────────────────────────────────────
@@ -27,6 +35,94 @@ export function getFilePaths() {
         .filter((e) => e.kind === "write" || e.kind === "edit")
         .map(e => e.path);
 }
+// ─── Persistence (Bug #4385 — evidence must survive session restarts) ────────
+/**
+ * Build the path for the evidence JSON file for a given unit.
+ * Lives under .gsd/safety/ which is gitignored and session-scoped.
+ */
+function evidencePath(basePath, milestoneId, sliceId, taskId) {
+    return join(basePath, ".gsd", "safety", `evidence-${milestoneId}-${sliceId}-${taskId}.json`);
+}
+/**
+ * Validate that a parsed value is an array of EvidenceEntry objects.
+ * Rejects corrupt / schema-mismatch data rather than letting it poison state.
+ */
+function isEvidenceArray(data) {
+    if (!Array.isArray(data))
+        return false;
+    return data.every((e) => {
+        if (e === null || typeof e !== "object")
+            return false;
+        const rec = e;
+        if (typeof rec.toolCallId !== "string")
+            return false;
+        if (typeof rec.timestamp !== "number")
+            return false;
+        if (rec.kind === "bash") {
+            return (typeof rec.command === "string" &&
+                typeof rec.exitCode === "number" &&
+                typeof rec.outputSnippet === "string");
+        }
+        if (rec.kind === "write" || rec.kind === "edit") {
+            return typeof rec.path === "string";
+        }
+        return false;
+    });
+}
+/**
+ * Persist the current in-memory evidence to disk so it survives a session
+ * restart. Called from saveEvidenceToDisk after recordToolCall/recordToolResult.
+ * Non-fatal — persistence failures must never break unit execution.
+ */
+export function saveEvidenceToDisk(basePath, milestoneId, sliceId, taskId) {
+    try {
+        const path = evidencePath(basePath, milestoneId, sliceId, taskId);
+        mkdirSync(dirname(path), { recursive: true });
+        const tmp = `${path}.tmp.${randomBytes(4).toString("hex")}`;
+        writeFileSync(tmp, JSON.stringify(unitEvidence, null, 2) + "\n", "utf-8");
+        renameSync(tmp, path);
+    }
+    catch {
+        // Non-fatal — don't let persistence failures break unit execution
+    }
+}
+/**
+ * Load persisted evidence from disk into the in-memory array.
+ * Call after resetEvidence() on session resume to restore context for a
+ * partially-executed unit. If the file does not exist (fresh unit), this
+ * is a no-op — getEvidence() will return [] which is correct.
+ */
+export function loadEvidenceFromDisk(basePath, milestoneId, sliceId, taskId) {
+    try {
+        const path = evidencePath(basePath, milestoneId, sliceId, taskId);
+        if (!existsSync(path))
+            return;
+        const raw = readFileSync(path, "utf-8");
+        const parsed = JSON.parse(raw);
+        if (isEvidenceArray(parsed)) {
+            unitEvidence = parsed;
+        }
+    }
+    catch {
+        // Non-fatal — corrupt / missing file is treated as empty evidence
+    }
+}
+/**
+ * Delete the persisted evidence file for a unit after it has been fully
+ * processed. Prevents stale evidence from affecting future retries of
+ * the same unit ID.
+ */
+export function clearEvidenceFromDisk(basePath, milestoneId, sliceId, taskId) {
+    try {
+        const path = evidencePath(basePath, milestoneId, sliceId, taskId);
+        if (existsSync(path)) {
+            unlinkSync(path);
+        }
+    }
+    catch {
+        // Non-fatal
+    }
+}
 // ─── Recording (called from register-hooks.ts) ─────────────────────────────
 /**
  * Record a tool call at dispatch time (before execution).

package/dist/resources/extensions/gsd/safety/file-change-validator.js

@@ -4,13 +4,18 @@
  *
  * Uses tasks.expected_output (DB column, populated from per-task ## Expected Output)
  * and tasks.files (from slice PLAN.md - Files: subline) as the expected set.
- * Compares against git diff HEAD~1 --name-only after auto-commit.
+ * Compares against `git diff-tree --root --no-commit-id -r --name-only HEAD` after auto-commit.
+ * Using diff-tree --root handles initial commits, shallow clones, and merge commits correctly
+ * (Bug #4385 — git diff HEAD~1 failed on initial commits).
  *
  * Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
  */
+import { createRequire } from "node:module";
 import { execFileSync } from "node:child_process";
 import { normalizePlannedFileReference } from "../files.js";
 import { logWarning } from "../workflow-logger.js";
+const _require = createRequire(import.meta.url);
+const picomatch = _require("picomatch");
 // ─── Public API ─────────────────────────────────────────────────────────────
 /**
  * Validate file changes after auto-commit for an execute-task unit.
@@ -20,7 +25,7 @@ import { logWarning } from "../workflow-logger.js";
  * @param expectedOutput - JSON array from tasks.expected_output DB column
  * @param plannedFiles - JSON array from tasks.files DB column
  */
-export function validateFileChanges(basePath, expectedOutput, plannedFiles) {
+export function validateFileChanges(basePath, expectedOutput, plannedFiles, fileChangeAllowlist = []) {
     const allExpected = new Set([...expectedOutput, ...plannedFiles]);
     // If no expected files were planned, skip validation
     if (allExpected.size === 0)
@@ -33,8 +38,11 @@ export function validateFileChanges(basePath, expectedOutput, plannedFiles) {
     const projectFiles = actualFiles.filter(f => !f.startsWith(".gsd/") && !f.startsWith(".gsd\\"));
     // Normalize expected paths (strip leading ./ or /)
     const normalizedExpected = new Set([...allExpected].map((f) => normalizePlannedFileReference(f).replace(/^\.\//, "").replace(/^\//, "")));
-    // Compute symmetric difference
-    const unexpectedFiles = projectFiles.filter(f => !normalizedExpected.has(f));
+    // Build allowlist matchers once (dot: true so patterns like `**/.hidden` work).
+    const allowlistMatchers = fileChangeAllowlist.map(p => picomatch(p, { dot: true }));
+    const isAllowlisted = (f) => allowlistMatchers.some(m => m(f));
+    // Compute symmetric difference, excluding allowlisted files
+    const unexpectedFiles = projectFiles.filter(f => !normalizedExpected.has(f) && !isAllowlisted(f));
     const missingFiles = [...normalizedExpected].filter(f => !projectFiles.includes(f));
     const violations = [];
     for (const f of unexpectedFiles) {
@@ -62,7 +70,7 @@ export function validateFileChanges(basePath, expectedOutput, plannedFiles) {
 // ─── Internals ──────────────────────────────────────────────────────────────
 function getChangedFilesFromLastCommit(basePath) {
     try {
-        const result = execFileSync("git", ["diff", "--name-only", "HEAD~1", "HEAD"], { cwd: basePath, stdio: ["ignore", "pipe", "pipe"], encoding: "utf-8" }).trim();
+        const result = execFileSync("git", ["diff-tree", "--root", "--no-commit-id", "-r", "--name-only", "HEAD"], { cwd: basePath, stdio: ["ignore", "pipe", "pipe"], encoding: "utf-8" }).trim();
         return result ? result.split("\n").filter(Boolean) : [];
     }
     catch (e) {

package/dist/resources/extensions/gsd/safety/git-checkpoint.js

@@ -58,6 +58,17 @@ export function rollbackToCheckpoint(basePath, unitId, sha) {
             logWarning("safety", "rollback: detached HEAD state, cannot rollback");
             return false;
         }
+        // Preserve any staged or untracked user work before the hard reset.
+        // The user may have a partial fix staged that they wanted to inspect;
+        // reset --hard wipes both staged and unstaged changes (reflog only
+        // covers committed state). Push a labeled stash first so recovery
+        // is possible. (Issue #4980 HIGH-4)
+        try {
+            execFileSync("git", ["stash", "push", "--include-untracked", "-m", `gsd: pre-rollback-stash ${unitId} ${new Date().toISOString()}`], { cwd: basePath, stdio: ["ignore", "pipe", "pipe"], encoding: "utf-8" });
+        }
+        catch {
+            /* nothing to stash, or stash refused — proceed with reset */
+        }
         // Reset branch pointer and working tree to checkpoint SHA in one step.
         // Using `git reset --hard <sha>` works on the currently checked-out branch
         // (unlike `git branch -f` which is rejected for checked-out branches).

package/dist/resources/extensions/gsd/safety/safety-harness.js

@@ -23,6 +23,7 @@ const DEFAULTS = {
     checkpoints: true,
     auto_rollback: false,
     timeout_scale_cap: 6,
+    file_change_allowlist: [],
 };
 // ─── Public API ─────────────────────────────────────────────────────────────
 /**
@@ -42,6 +43,9 @@ export function resolveSafetyHarnessConfig(raw) {
         checkpoints: typeof raw.checkpoints === "boolean" ? raw.checkpoints : DEFAULTS.checkpoints,
         auto_rollback: typeof raw.auto_rollback === "boolean" ? raw.auto_rollback : DEFAULTS.auto_rollback,
         timeout_scale_cap: typeof raw.timeout_scale_cap === "number" ? raw.timeout_scale_cap : DEFAULTS.timeout_scale_cap,
+        file_change_allowlist: Array.isArray(raw.file_change_allowlist)
+            ? raw.file_change_allowlist.filter((p) => typeof p === "string")
+            : DEFAULTS.file_change_allowlist,
     };
 }
 /**
@@ -56,7 +60,7 @@ export function isHarnessEnabled(raw) {
     return DEFAULTS.enabled;
 }
 // ─── Re-exports ─────────────────────────────────────────────────────────────
-export { resetEvidence, getEvidence, getBashEvidence, getFilePaths, recordToolCall, recordToolResult, } from "./evidence-collector.js";
+export { resetEvidence, getEvidence, getBashEvidence, getFilePaths, recordToolCall, recordToolResult, saveEvidenceToDisk, loadEvidenceFromDisk, clearEvidenceFromDisk, } from "./evidence-collector.js";
 export { classifyCommand } from "./destructive-guard.js";
 export { validateFileChanges } from "./file-change-validator.js";
 export { crossReferenceEvidence } from "./evidence-cross-ref.js";

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

@@ -2,8 +2,8 @@
  * Service Tier — gating, status formatting, icon resolution, and
  * the /gsd fast command handler.
  *
- * Service tiers (priority/flex) are an OpenAI feature that only applies
- * to gpt-5.4 variants. This module centralizes the model-gating logic
+ * Service tiers (priority/flex) are an OpenAI feature that currently only
+ * applies to gpt-5.4 variants in GSD. This module centralizes the model-gating logic
  * so that icons, preferences, and the before_provider_request hook all
  * use a single source of truth.
  */
@@ -21,6 +21,9 @@ const SERVICE_TIER_SCOPE_NOTE = "Only affects gpt-5.4 models, regardless of prov
  * (set via CAPABILITY_PATCHES in packages/pi-ai/src/models.ts). When callers
  * have access to the full Model object, prefer reading capabilities directly.
  *
+ * GPT-5.5 is intentionally excluded until we verify its provider payload
+ * contract instead of assuming `service_tier` support.
+ *
  * See: https://github.com/gsd-build/gsd-2/issues/2546
  */
 const SERVICE_TIER_MODEL_PREFIXES = ["gpt-5.4"];

package/dist/resources/extensions/gsd/session-lock.js

@@ -138,15 +138,16 @@ function ensureExitHandler(_gsdDir) {
         // Lock files accumulate across main project .gsd/, worktree .gsd/,
         // and projects registry paths — cleanup must cover all of them.
         for (const dir of _lockDirRegistry) {
+            const lockFile = join(dir, LOCK_FILE);
+            const ownsRegisteredLock = isLockFileOwnedByCurrentProcess(lockFile);
             try {
-                const lockFile = join(dir, LOCK_FILE);
-                if (existsSync(lockFile))
+                if (ownsRegisteredLock && existsSync(lockFile))
                     unlinkSync(lockFile);
             }
             catch { /* best-effort */ }
             try {
                 const lockDir = join(dir + ".lock");
-                if (existsSync(lockDir))
+                if (ownsRegisteredLock && existsSync(lockDir))
                     rmSync(lockDir, { recursive: true, force: true });
             }
             catch { /* best-effort */ }
@@ -447,10 +448,13 @@ export function releaseSessionLock(basePath) {
         }
         _releaseFunction = null;
     }
-    // Remove the lock file at the current path
+    // Remove the lock file at the current path only if it still belongs to us.
+    // Lost-lock cleanup can run after another process has taken ownership; in
+    // that case deleting auto.lock would erase the newer owner's evidence.
     const lp = lockPath(basePath);
+    const ownsPrimaryLock = isLockFileOwnedByCurrentProcess(lp);
     try {
-        if (existsSync(lp))
+        if (ownsPrimaryLock && existsSync(lp))
             unlinkSync(lp);
     }
     catch {
@@ -462,14 +466,14 @@ export function releaseSessionLock(basePath) {
     const lockTarget = effectiveLockTarget(gsdDir);
     try {
         const lockDir = join(lockTarget + ".lock");
-        if (existsSync(lockDir))
+        if (ownsPrimaryLock && existsSync(lockDir))
             rmSync(lockDir, { recursive: true, force: true });
     }
     catch {
         // Non-fatal
     }
     // Also clean the per-milestone parallel directory itself if it exists
-    if (lockTarget !== gsdDir) {
+    if (ownsPrimaryLock && lockTarget !== gsdDir) {
         try {
             if (existsSync(lockTarget))
                 rmSync(lockTarget, { recursive: true, force: true });
@@ -481,15 +485,16 @@ export function releaseSessionLock(basePath) {
     // Clean ALL registered lock paths (#1578) — lock files accumulate across
     // main project .gsd/, worktree .gsd/, and projects registry paths.
     for (const dir of _lockDirRegistry) {
+        const lockFile = join(dir, LOCK_FILE);
+        const ownsRegisteredLock = isLockFileOwnedByCurrentProcess(lockFile);
         try {
-            const lockFile = join(dir, LOCK_FILE);
-            if (existsSync(lockFile))
+            if (ownsRegisteredLock && existsSync(lockFile))
                 unlinkSync(lockFile);
         }
         catch { /* best-effort */ }
         try {
             const lockDir = join(dir + ".lock");
-            if (existsSync(lockDir))
+            if (ownsRegisteredLock && existsSync(lockDir))
                 rmSync(lockDir, { recursive: true, force: true });
         }
         catch { /* best-effort */ }
@@ -541,6 +546,10 @@ function readExistingLockData(lp) {
         return null;
     }
 }
+function isLockFileOwnedByCurrentProcess(lp) {
+    const existing = readExistingLockData(lp);
+    return existing?.pid === process.pid;
+}
 export function readExistingLockDataWithRetry(lp, options) {
     const maxAttempts = options?.maxAttempts ?? 3;
     const delayMs = options?.delayMs ?? 200;