@opengsd/gsd-pi 1.2.0-dev.844675c9 → 1.2.0-dev.955e4da0

package/src/resources/extensions/gsd/prompts/guided-discuss-project.md

@@ -36,6 +36,8 @@ After the opening answer, classify project shape as **`simple`** or **`complex`*
 
 Persist the verdict through `gsd_summary_save` with `artifact_type: "PROJECT"` into `## Project Shape`; downstream `discuss-requirements`, `discuss-milestone`, and `discuss-slice` read the rendered projection.
 
+When the project is browser-facing (web app, SPA, static site with UI, Next/React/Vue, etc.), also record **`Web stack:`** under `## Project Shape` (framework + whether Playwright or browser UAT is expected). Downstream planning uses this to choose `browser-executable` / `runtime-executable` UAT instead of deferring UI proof to humans.
+
 ### Before deeper rounds
 
 Ground your questions in the **Preparation Context snapshot above** (stack, structure, greenfield/brownfield and framework signals) plus any prior `.planning/` or `.gsd/` artifacts — those are authoritative. **Do not survey the codebase** with `rg`/`find`/`scout` before asking; read a specific file only when a question's answer genuinely hinges on it.
@@ -103,7 +105,7 @@ Once the user confirms depth:
 
 1. Use the **Project** output template (inlined above).
 2. Call `gsd_summary_save` with `artifact_type: "PROJECT"` and full project markdown as `content`; omit `milestone_id`. The tool persists the DB-backed PROJECT artifact and renders `.gsd/PROJECT.md`. Preserve the user's terms and framing.
-3. The `## Project Shape` section MUST contain `**Complexity:** simple` or `**Complexity:** complex` (matching the verdict you announced) plus a one-line `**Why:**` rationale. Downstream stages read this line.
+3. The `## Project Shape` section MUST contain `**Complexity:** simple` or `**Complexity:** complex` (matching the verdict you announced) plus a one-line `**Why:**` rationale. For browser-facing projects, also include `**Web stack:**` (framework and expected browser/Playwright UAT approach). Downstream stages read these lines.
 4. The `## Capability Contract` section MUST reference `.gsd/REQUIREMENTS.md` — that file does not yet exist; the next stage (`discuss-requirements`) will produce it.
 5. The `## Milestone Sequence` MUST list at least M001 with title and one-liner. Subsequent milestones may be listed as known intents; they will be elaborated in their own discuss-milestone stages.
 6. Do NOT use `artifact_type: "CONTEXT"` and do NOT pass `milestone_id: "PROJECT"`; that creates a fake milestone named PROJECT.

package/src/resources/extensions/gsd/prompts/plan-milestone.md

@@ -47,6 +47,8 @@ If milestone research is inlined, trust it and skip redundant exploration. If fi
 
 Narrate decomposition reasoning in complete sentences: grouping, risk order, verification strategy.
 
+**Web apps:** when inlined Web App UAT guidance is present, set milestone `Verification Classes` → UAT to browser-observable acceptance (Playwright spec or `browser_*` checks). Order an early slice to add Playwright smoke scaffolding when the dependency is missing.
+
 Then:
 1. Use the **Roadmap** output template from the inlined context above
 2. {{skillActivation}}

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

@@ -41,6 +41,7 @@ If slice research is inlined, trust its architectural findings, but verify every
 3. Use the inlined Output Template sections already present in this prompt. Do not read template files from disk.
 4. {{skillActivation}} Record expected executor skills in each task plan's `skills_used` frontmatter.
 5. Define slice verification before tasks. Non-trivial slices need real tests or executable assertions; boundary contracts need contract-exercising checks. Tests must not read .gitignore/gitignored paths such as `.gsd/`, `.planning/`, or `.audits/`.
+   **Web apps:** when inlined Web App UAT guidance is present, follow it — add Playwright smoke scaffolding on the first UI slice if missing, name localhost preconditions, and plan verification commands that match browser-capable UAT modes at slice closeout.
 6. Include Threat Surface (Q3), Requirement Impact (Q4), proof level, observability, integration closure, Failure Modes (Q5), Load Profile (Q6), and Negative Tests (Q7) only where applicable.
 7. Right-size tasks. Simple slices can be one task; split only when context, ownership, or verification boundaries justify it.
 8. Task `verify` commands must be safe, simple commands. Do not use shell pipes, redirects, semicolons, backticks, command substitution, output trimming, or grep regex alternation with `|`. If multiple checks are needed, create a small test file and run it with `node --test` or a package test script, or use separate simple commands joined only with `&&`. For absence checks, verify a pattern does not exist with `! grep -q 'pattern' file` or `! rg -q 'pattern' file`; do not use `grep -c` or `rg -c` to assert zero matches because count commands exit 1 when they find zero matches, and the verification gate treats that as failure.

package/src/resources/extensions/gsd/prompts/refine-slice.md

@@ -63,6 +63,7 @@ Then:
    - `{{taskPlanTemplatePath}}`
 2. {{skillActivation}} Record the installed skills you expect executors to use in each task plan's `skills_used` frontmatter.
 3. Define slice-level verification: the objective stopping condition. Plan real test files with real assertions; for simple slices, executable commands are fine.
+   **Web apps:** when inlined Web App UAT guidance is present, follow it for Playwright scaffolding and browser-capable verification commands.
 4. For non-trivial slices, plan observability / proof level / integration closure, threat surface, and requirement impact. Omit entirely for simple slices.
 5. Decompose the slice into tasks that fit one context window each. Every task passed to `gsd_plan_slice` must use the exact keys `taskId`, `title`, `description`, `estimate`, `files`, `verify`, `inputs`, `expectedOutput`, and optional `observabilityImpact`. Put Why / Do / Done-when detail in `description`. `files`, `inputs`, and `expectedOutput` must be JSON arrays of strings, even for one path (for example, `"expectedOutput": ["src/index.ts"]`, never `"expectedOutput": "src/index.ts"`). `expectedOutput` is path-only: list only files the task creates or overwrites, and use `[]` for pure verification tasks.
 6. **Persist planning state through `gsd_plan_slice`.** Call it with the full payload. The tool writes to the DB and renders `{{outputPath}}` and `{{slicePath}}/tasks/T##-PLAN.md` automatically. Do NOT rely on direct `PLAN.md` writes.

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

@@ -158,7 +158,7 @@ Fix root causes, not symptoms. If applying temporary mitigation, label it and pr
 - 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". If it fits a commit comment or standup note, it is acceptable.
 - 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 summary and 2-4 numbered next-step options; last option is always "Other". Omit the list for strict output formats.
+- After completing a task, give a brief summary and 2-4 numbered next-step options; last option is always "Other". Omit the list for strict output formats, or when the active workflow prompt already ends with its own explicit "Next steps:" handoff block — in that case follow the workflow's handoff and do not add a second list.
 
   If any next step is destructive/outward-facing, present it via `ask_user_questions` and wait for the user's answer before execution. Do not execute a next-step item from a prior plain-text numbered list without fresh confirmation.
 

package/src/resources/extensions/gsd/provider-payload-policy.ts

@@ -0,0 +1,140 @@
+/**
+ * Provider Payload Policy - ordered shaping of provider request payloads.
+ *
+ * The order is intentional:
+ * 1. observation budgeting masks old tool results in auto-mode,
+ * 2. display truncation caps tool-result text for every mode,
+ * 3. the protected Source Context Block is appended after truncation,
+ * 4. supported models receive the configured service tier.
+ */
+
+import type { ContextManagementConfig } from "./preferences-types.js";
+import type { ServiceTierSetting } from "./service-tier.js";
+
+import {
+  createObservationMask,
+  createResponsesInputObservationMask,
+  truncateContextResultMessages,
+  truncateResponsesInputResultItems,
+} from "./context-masker.js";
+import { getSourceObservationStore, isAutoActive } from "./auto-runtime-state.js";
+import { loadEffectiveGSDPreferences } from "./preferences.js";
+import { getEffectiveServiceTier, supportsServiceTier } from "./service-tier.js";
+import { injectSourceContextBlockIntoPayload } from "./source-observations.js";
+
+const DEFAULT_OBSERVATION_MASK_TURNS = 8;
+const DEFAULT_TOOL_RESULT_MAX_CHARS = 800;
+
+type MessagePayload = Parameters<ReturnType<typeof createObservationMask>>[0];
+type ResponsesInputPayload = Parameters<ReturnType<typeof createResponsesInputObservationMask>>[0];
+
+export interface ProviderPayloadPolicyDeps {
+  isAutoActive(): boolean;
+  loadContextManagementConfig(): ContextManagementConfig | undefined;
+  renderSourceContextBlock(): string | null;
+  getEffectiveServiceTier(): ServiceTierSetting;
+  supportsServiceTier(modelId: string): boolean;
+}
+
+export interface ProviderPayloadPolicyInput {
+  payload: Record<string, unknown>;
+  modelId?: string;
+  deps?: Partial<ProviderPayloadPolicyDeps>;
+}
+
+export const DEFAULT_PROVIDER_PAYLOAD_POLICY_DEPS: ProviderPayloadPolicyDeps = {
+  isAutoActive,
+  loadContextManagementConfig: () => loadEffectiveGSDPreferences()?.preferences.context_management,
+  renderSourceContextBlock: () => getSourceObservationStore().renderActiveBlock(),
+  getEffectiveServiceTier,
+  supportsServiceTier,
+};
+
+export function applyProviderPayloadPolicy({
+  payload,
+  modelId,
+  deps: overrides,
+}: ProviderPayloadPolicyInput): Record<string, unknown> {
+  const deps = { ...DEFAULT_PROVIDER_PAYLOAD_POLICY_DEPS, ...overrides };
+
+  try {
+    applyContextManagement(payload, deps);
+  } catch {
+    // Provider payload shaping should not block a request when optional
+    // context management preferences or adapters fail.
+  }
+
+  try {
+    applySourceContextBlock(payload, deps);
+  } catch {
+    // Source observations are opportunistic; execution can continue without
+    // an injected block.
+  }
+
+  applyServiceTier(payload, modelId, deps);
+  return payload;
+}
+
+function applyContextManagement(
+  payload: Record<string, unknown>,
+  deps: ProviderPayloadPolicyDeps,
+): void {
+  const config = deps.loadContextManagementConfig();
+  applyObservationBudget(payload, config, deps.isAutoActive());
+  applyDisplayTruncation(payload, config);
+}
+
+function applyObservationBudget(
+  payload: Record<string, unknown>,
+  config: ContextManagementConfig | undefined,
+  autoActive: boolean,
+): void {
+  if (!autoActive || config?.observation_masking === false) return;
+
+  const keepTurns = config?.observation_mask_turns ?? DEFAULT_OBSERVATION_MASK_TURNS;
+  if (Array.isArray(payload.messages)) {
+    payload.messages = createObservationMask(keepTurns)(payload.messages as MessagePayload);
+  }
+  if (Array.isArray(payload.input)) {
+    payload.input = createResponsesInputObservationMask(keepTurns)(payload.input as ResponsesInputPayload);
+  }
+}
+
+function applyDisplayTruncation(
+  payload: Record<string, unknown>,
+  config: ContextManagementConfig | undefined,
+): void {
+  const maxChars = config?.tool_result_max_chars ?? DEFAULT_TOOL_RESULT_MAX_CHARS;
+
+  if (Array.isArray(payload.messages)) {
+    payload.messages = truncateContextResultMessages(payload.messages as MessagePayload, maxChars);
+  }
+  if (Array.isArray(payload.input)) {
+    payload.input = truncateResponsesInputResultItems(payload.input as ResponsesInputPayload, maxChars);
+  }
+}
+
+function applySourceContextBlock(
+  payload: Record<string, unknown>,
+  deps: ProviderPayloadPolicyDeps,
+): void {
+  if (!deps.isAutoActive()) return;
+
+  const sourceContextBlock = deps.renderSourceContextBlock();
+  if (!sourceContextBlock) return;
+
+  Object.assign(payload, injectSourceContextBlockIntoPayload(payload, sourceContextBlock));
+}
+
+function applyServiceTier(
+  payload: Record<string, unknown>,
+  modelId: string | undefined,
+  deps: ProviderPayloadPolicyDeps,
+): void {
+  if (!modelId) return;
+
+  const tier = deps.getEffectiveServiceTier();
+  if (!tier || !deps.supportsServiceTier(modelId)) return;
+
+  payload.service_tier = tier;
+}

package/src/resources/extensions/gsd/pull-request-process.ts

@@ -0,0 +1,41 @@
+// Project/App: gsd-pi
+// File Purpose: Process-level pull request policy for GSD-generated PRs.
+
+import { createDraftPR } from "./git-service.js";
+import {
+  buildPrEvidence,
+  type PrEvidence,
+  type PrEvidenceInput,
+} from "./pr-evidence.js";
+
+export interface DraftPullRequestOptions {
+  head: string;
+  base: string;
+}
+
+export interface DraftPullRequestDeps {
+  createDraftPR: (
+    basePath: string,
+    milestoneId: string,
+    title: string,
+    body: string,
+    opts: DraftPullRequestOptions,
+  ) => string | null;
+}
+
+export function buildPullRequestEvidence(input: PrEvidenceInput): PrEvidence {
+  return buildPrEvidence({
+    ...input,
+    aiAssisted: false,
+  });
+}
+
+export function createDraftPullRequestFromEvidence(
+  basePath: string,
+  milestoneId: string,
+  evidence: PrEvidence,
+  options: DraftPullRequestOptions,
+  deps: DraftPullRequestDeps = { createDraftPR },
+): string | null {
+  return deps.createDraftPR(basePath, milestoneId, evidence.title, evidence.body, options);
+}

package/src/resources/extensions/gsd/quality-gate-closure.ts

@@ -0,0 +1,140 @@
+// Project/App: gsd-pi
+// File Purpose: Canonical quality-gate closure from durable DB and artifact evidence.
+
+import { existsSync, readFileSync } from "node:fs";
+
+import { extractSection } from "./files.js";
+import { getGateDefinition } from "./gate-registry.js";
+import { getMilestoneSlices, getPendingGates, saveGateResult } from "./gsd-db.js";
+import { resolveSliceFile, resolveTaskFile } from "./paths.js";
+import type { GateId, GateRow, GateVerdict } from "./types.js";
+
+export interface QualityGateClosureOptions {
+  artifactBasePath?: string;
+  milestoneValidationPassed?: boolean;
+}
+
+export interface QualityGateClosureResult {
+  repaired: Array<{ gateId: GateId; sliceId: string; taskId?: string; verdict: GateVerdict }>;
+  unresolved: GateRow[];
+}
+
+interface GateEvidence {
+  verdict: GateVerdict;
+  rationale: string;
+  findings: string;
+}
+
+const GATE_SECTION_HEADINGS: Partial<Record<GateId, string[]>> = {
+  Q3: ["Threat Surface", "Abuse Surface"],
+  Q4: ["Requirement Impact", "Broken Promises"],
+  Q5: ["Failure Modes"],
+  Q6: ["Load Profile"],
+  Q7: ["Negative Tests"],
+  Q8: ["Operational Readiness"],
+};
+
+function readFile(path: string | null): string | null {
+  if (!path || !existsSync(path)) return null;
+  return readFileSync(path, "utf-8");
+}
+
+function firstSection(content: string | null, gateId: GateId): string | null {
+  if (!content) return null;
+  for (const heading of GATE_SECTION_HEADINGS[gateId] ?? []) {
+    const section = extractSection(content, heading);
+    if (section) return section;
+  }
+  return null;
+}
+
+function evidenceArtifactContent(row: GateRow, basePath: string): string | null {
+  const def = getGateDefinition(row.gate_id);
+  switch (def?.ownerTurn) {
+    case "gate-evaluate":
+      return readFile(resolveSliceFile(basePath, row.milestone_id, row.slice_id, "PLAN"));
+    case "execute-task":
+      return readFile(resolveTaskFile(basePath, row.milestone_id, row.slice_id, row.task_id, "SUMMARY"));
+    case "complete-slice":
+      return readFile(resolveSliceFile(basePath, row.milestone_id, row.slice_id, "SUMMARY"));
+    default:
+      return null;
+  }
+}
+
+function closureEvidence(row: GateRow, options: QualityGateClosureOptions): GateEvidence | null {
+  const def = getGateDefinition(row.gate_id);
+  if (!def) return null;
+
+  if (def.ownerTurn === "validate-milestone" && options.milestoneValidationPassed) {
+    return {
+      verdict: "pass",
+      rationale: `${def.promptSection} covered by passing milestone validation`,
+      findings: "",
+    };
+  }
+
+  if (!options.artifactBasePath) return null;
+
+  const section = firstSection(evidenceArtifactContent(row, options.artifactBasePath), row.gate_id);
+  if (section) {
+    return {
+      verdict: "pass",
+      rationale: `${def.promptSection} evidence found in durable artifact`,
+      findings: section,
+    };
+  }
+
+  if (!options.milestoneValidationPassed) return null;
+  return {
+    verdict: "omitted",
+    rationale: `${def.promptSection} has no durable artifact section; milestone validation passed`,
+    findings: "",
+  };
+}
+
+function closeGate(row: GateRow, evidence: GateEvidence): void {
+  saveGateResult({
+    milestoneId: row.milestone_id,
+    sliceId: row.slice_id,
+    gateId: row.gate_id,
+    taskId: row.task_id,
+    verdict: evidence.verdict,
+    rationale: evidence.rationale,
+    findings: evidence.findings,
+  });
+}
+
+export function closeQualityGatesFromEvidence(
+  milestoneId: string,
+  options: QualityGateClosureOptions = {},
+): QualityGateClosureResult {
+  const repaired: QualityGateClosureResult["repaired"] = [];
+  const unresolved: GateRow[] = [];
+
+  for (const slice of getMilestoneSlices(milestoneId)) {
+    const sliceId = slice.id;
+    for (const row of getPendingGates(milestoneId, sliceId)) {
+      if (!getGateDefinition(row.gate_id)) {
+        unresolved.push(row);
+        continue;
+      }
+
+      const evidence = closureEvidence(row, options);
+      if (!evidence) {
+        unresolved.push(row);
+        continue;
+      }
+
+      closeGate(row, evidence);
+      repaired.push({
+        gateId: row.gate_id,
+        sliceId: row.slice_id,
+        ...(row.task_id ? { taskId: row.task_id } : {}),
+        verdict: evidence.verdict,
+      });
+    }
+  }
+
+  return { repaired, unresolved };
+}

package/src/resources/extensions/gsd/question-transport.ts

@@ -0,0 +1,138 @@
+// Project/App: gsd-pi
+// File Purpose: Resolve how workflow units should ask the user for input.
+
+import { parseMcpToolName, toMcpToolName } from "./mcp-tool-name.js";
+
+export type StructuredQuestionsFlag = "true" | "false";
+
+export interface QuestionTransportOptions {
+  activeTools: string[];
+  authMode?: "apiKey" | "oauth" | "externalCli" | "none";
+  baseUrl?: string;
+  env?: NodeJS.ProcessEnv;
+}
+
+export interface QuestionTransportResolution {
+  structuredQuestionsAvailable: StructuredQuestionsFlag;
+  questionToolAvailable: boolean;
+  usesWorkflowMcp: boolean;
+  reason: "question-tool-available" | "question-tool-missing" | "workflow-mcp-disabled";
+}
+
+export interface WorkflowQuestionToolSurfaceOptions {
+  workflowServerName?: string;
+  workflowExplicitlyBlocked?: boolean;
+  workflowMcpTools: string[];
+  exactWorkflowMcpTools: string[];
+  env?: NodeJS.ProcessEnv;
+}
+
+export interface WorkflowQuestionToolSurfaceResolution extends QuestionTransportResolution {
+  questionToolName?: string;
+  workflowQuestionsEnabled: boolean;
+  disallowedTools: string[];
+}
+
+function isWorkflowMcpServerName(serverName: string): boolean {
+  const normalized = serverName.toLowerCase();
+  return normalized === "gsd" || normalized.includes("workflow");
+}
+
+export function usesWorkflowMcpTransport(
+  authMode: QuestionTransportOptions["authMode"],
+  baseUrl: string | undefined,
+): boolean {
+  return authMode === "externalCli" && typeof baseUrl === "string" && baseUrl.startsWith("local://");
+}
+
+export function hasAskUserQuestionsTool(activeTools: string[]): boolean {
+  return activeTools.some((toolName) => {
+    if (toolName === "ask_user_questions") return true;
+    const mcp = parseMcpToolName(toolName);
+    if (!mcp) return false;
+    if (mcp.toolName === "ask_user_questions") return true;
+    return mcp.toolName === "*" && isWorkflowMcpServerName(mcp.serverName);
+  });
+}
+
+function workflowMcpStructuredQuestionsEnabled(env: NodeJS.ProcessEnv = process.env): boolean {
+  const value = env.GSD_WORKFLOW_MCP_STRUCTURED_QUESTIONS?.trim().toLowerCase();
+  return value !== "0" && value !== "false" && value !== "off";
+}
+
+export function resolveQuestionTransport(
+  options: QuestionTransportOptions,
+): QuestionTransportResolution {
+  const questionToolAvailable = hasAskUserQuestionsTool(options.activeTools);
+  const usesWorkflowMcp = usesWorkflowMcpTransport(options.authMode, options.baseUrl);
+
+  if (!questionToolAvailable) {
+    return {
+      structuredQuestionsAvailable: "false",
+      questionToolAvailable,
+      usesWorkflowMcp,
+      reason: "question-tool-missing",
+    };
+  }
+
+  if (options.authMode === "externalCli" && !workflowMcpStructuredQuestionsEnabled(options.env)) {
+    return {
+      structuredQuestionsAvailable: "false",
+      questionToolAvailable,
+      usesWorkflowMcp,
+      reason: "workflow-mcp-disabled",
+    };
+  }
+
+  return {
+    structuredQuestionsAvailable: "true",
+    questionToolAvailable,
+    usesWorkflowMcp,
+    reason: "question-tool-available",
+  };
+}
+
+export function supportsStructuredQuestions(
+  activeTools: string[],
+  options: Omit<QuestionTransportOptions, "activeTools"> = {},
+): boolean {
+  return resolveQuestionTransport({
+    ...options,
+    activeTools,
+  }).structuredQuestionsAvailable === "true";
+}
+
+export function resolveWorkflowQuestionToolSurface(
+  options: WorkflowQuestionToolSurfaceOptions,
+): WorkflowQuestionToolSurfaceResolution {
+  const questionToolName = options.workflowServerName && !options.workflowExplicitlyBlocked
+    ? toMcpToolName(options.workflowServerName, "ask_user_questions")
+    : undefined;
+  const activeTools = [
+    ...options.exactWorkflowMcpTools,
+    ...options.workflowMcpTools,
+    ...(questionToolName ? [questionToolName] : []),
+  ];
+  const transport = resolveQuestionTransport({
+    activeTools,
+    authMode: "externalCli",
+    baseUrl: "local://claude-code",
+    env: options.env,
+  });
+  const exactQuestionToolAllowed =
+    !!questionToolName && options.exactWorkflowMcpTools.includes(questionToolName);
+  const workflowQuestionsEnabled =
+    transport.structuredQuestionsAvailable === "true" &&
+    (options.workflowMcpTools.length > 0 || exactQuestionToolAllowed);
+  const disallowedTools =
+    options.workflowServerName && transport.questionToolAvailable && !workflowQuestionsEnabled
+      ? [toMcpToolName(options.workflowServerName, "ask_user_questions")]
+      : [];
+
+  return {
+    ...transport,
+    questionToolName,
+    workflowQuestionsEnabled,
+    disallowedTools,
+  };
+}

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

@@ -3,6 +3,7 @@
 
 import { classifyError, isTransient, type ErrorClass } from "./error-classifier.js";
 import { ReconciliationFailedError } from "./state-reconciliation.js";
+import { IllegalPhaseTransitionError } from "./state-transition-matrix.js";
 
 export type RecoveryFailureKind =
   | "tool-schema"
@@ -13,6 +14,7 @@ export type RecoveryFailureKind =
   | "worktree-invalid"
   | "verification-drift"
   | "reconciliation-drift"
+  | "illegal-transition"
   | "provider"
   | "runtime-unknown";
 
@@ -43,7 +45,9 @@ export function classifyFailure(input: RecoveryClassificationInput): RecoveryCla
   const failureKind =
     input.error instanceof ReconciliationFailedError
       ? "reconciliation-drift"
-      : input.failureKind ?? inferFailureKind(message);
+      : input.error instanceof IllegalPhaseTransitionError
+        ? "illegal-transition"
+        : input.failureKind ?? inferFailureKind(message);
 
   switch (failureKind) {
     case "tool-schema":
@@ -111,6 +115,15 @@ export function classifyFailure(input: RecoveryClassificationInput): RecoveryCla
         remediation:
           "Inspect the persistent or repair-failed drift kinds reported by the State Reconciliation Module before resuming.",
       };
+    case "illegal-transition":
+      return {
+        failureKind,
+        action: "escalate",
+        reason: `Illegal phase transition${unitSuffix(input)}: ${message}`,
+        exitReason: "illegal-transition",
+        remediation:
+          "A derived Phase edge rejected by the Phase Transition Invariant survived reconciliation; inspect deriveState and the State Reconciliation Module before resuming.",
+      };
     case "provider": {
       const providerClass = classifyError(message, input.retryAfterMs);
       return {

package/src/resources/extensions/gsd/roadmap-slices.ts

@@ -189,9 +189,15 @@ export function parseRoadmapSlices(content: string): RoadmapSliceEntry[] {
       const risk = (riskMatch ? riskMatch[1] : "low") as RiskLevel;
 
       const depsMatch = rest.match(/`depends:\[([^\]]*)\]`/);
-      const depends = depsMatch && depsMatch[1]!.trim()
-        ? expandDependencies(depsMatch[1]!.split(",").map(s => s.trim()))
+      // Recovery fallback: double-bracket form `[[id]]` from serialized bracket-wrapped IDs
+      const fallbackDepsMatch = depsMatch ? null : rest.match(/`depends:\[(\[(?:[^\]]*)\](?:,\[(?:[^\]]*)\])*)\]`/);
+      const rawDepContent = (depsMatch ?? fallbackDepsMatch)?.[1] ?? "";
+      const SLICE_ID_RE = /^[A-Za-z0-9][A-Za-z0-9-]*$/;
+      const RANGE_RE = /^[A-Za-z]+\d+(?:-|\.\.)[A-Za-z]+\d+$/;
+      const rawDepParts = rawDepContent.trim()
+        ? rawDepContent.replace(/\[|\]/g, "").split(",").map(s => s.trim()).filter(s => SLICE_ID_RE.test(s) || RANGE_RE.test(s))
         : [];
+      const depends = expandDependencies(rawDepParts);
 
       // ADR-011: the renderer writes a `[sketch]` badge for sketch slices.
       // Parse it back so the is_sketch flag survives a markdown → DB re-import

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

@@ -57,9 +57,10 @@ const EXECUTION_TOOL_NAMES = new Set([
   "functions.exec_command",
   "gsd_exec",
   "gsd_exec_search",
+  "gsd_uat_exec",
   "powershell",
 ]);
-const MCP_EXECUTION_TOOL_RE = /^mcp__.+__gsd_exec(?:_search)?$/;
+const MCP_EXECUTION_TOOL_RE = /^mcp__.+__gsd_(?:uat_)?exec(?:_search)?$/;
 
 // ─── Module State ───────────────────────────────────────────────────────────
 
@@ -206,11 +207,17 @@ export function clearEvidenceFromDisk(
  * Exit codes and output are filled in by recordToolResult after execution.
  */
 export function recordToolCall(toolCallId: string, toolName: string, input: Record<string, unknown>): void {
+  // Idempotent by toolCallId: native tools reach this via both
+  // tool_execution_start and tool_call; external (pre-executed) tools only
+  // via tool_execution_start. First recording wins.
+  if (unitEvidence.some(e => e.toolCallId === toolCallId)) return;
   if (isExecutionToolName(toolName)) {
     unitEvidence.push({
       kind: "bash",
       toolCallId,
-      command: String(input.command ?? input.cmd ?? input.query ?? ""),
+      // gsd_exec / gsd_uat_exec carry the script body in `script` (or `code`);
+      // bash-style tools use `command`/`cmd`; gsd_exec_search uses `query`.
+      command: String(input.command ?? input.script ?? input.cmd ?? input.code ?? input.query ?? ""),
       exitCode: -1,
       outputSnippet: "",
       timestamp: Date.now(),
@@ -249,11 +256,36 @@ export function recordToolResult(
   if (entry.kind === "bash") {
     const text = extractResultText(result);
     entry.outputSnippet = text.slice(0, 500);
-    const exitMatch = text.match(/Command exited with code (\d+)/);
-    entry.exitCode = exitMatch ? Number(exitMatch[1]) : (isError ? 1 : 0);
+    entry.exitCode = resolveExitCode(text, isError);
   }
 }
 
+/**
+ * Resolve the exit code from a tool result's text. Handles the bash tool's
+ * prose marker, the gsd_exec / gsd_uat_exec JSON envelope (`"exit_code": N`),
+ * and a last-resort read of the run's persisted `.gsd/exec/<id>.meta.json`
+ * (covers truncated result text).
+ */
+function resolveExitCode(text: string, isError: boolean): number {
+  const proseMatch = text.match(/Command exited with code (\d+)/);
+  if (proseMatch) return Number(proseMatch[1]);
+
+  const jsonMatch = text.match(/"exit_code"\s*:\s*(-?\d+)/);
+  if (jsonMatch) return Number(jsonMatch[1]);
+
+  const metaMatch = text.match(/"meta_path"\s*:\s*"([^"]+)"/);
+  if (metaMatch) {
+    try {
+      const meta = JSON.parse(readFileSync(metaMatch[1], "utf-8")) as Record<string, unknown>;
+      if (typeof meta.exit_code === "number") return meta.exit_code;
+    } catch {
+      // Fall through to the isError heuristic
+    }
+  }
+
+  return isError ? 1 : 0;
+}
+
 // ─── Internals ──────────────────────────────────────────────────────────────
 
 function extractResultText(result: unknown): string {

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

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