@opengsd/gsd-pi 1.2.0-dev.4c756166 → 1.2.0-dev.822c9439

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

@@ -37,4 +37,4 @@ You are executing a GSD quick task — a lightweight, focused unit of work outsi
 - <what was tested/verified>
 ```
 
-When done, say: "Quick task {{taskNum}} complete."
+When done, say: "Quick task {{taskNum}} complete." Say this exactly once — if you already said it in a prior message, do not repeat it.

package/dist/resources/extensions/gsd/prompts/reassess-roadmap.md

@@ -67,4 +67,4 @@ If `.gsd/REQUIREMENTS.md` exists and requirement ownership or status changed, up
 
 **DB access safety:** Do NOT query `.gsd/gsd.db` directly via `sqlite3` or `node -e require('better-sqlite3')`. Use `gsd_milestone_status` to read current milestone and slice state. All roadmap mutations go through `gsd_reassess_roadmap` — the tool writes to the DB and re-renders ROADMAP.md atomically.
 
-When done, say: "Roadmap reassessed."
+When done, say: "Roadmap reassessed." Say this exactly once — if you already said it in a prior message, do not repeat it.

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

@@ -65,7 +65,7 @@ Then:
 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.
+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. Never list GSD planning artifacts — anything under `.gsd/`, `.planning/`, or `.audits/`, or artifact names like `M001-CONTEXT.md` / `S01-PLAN.md` — in `inputs`, `files`, or `expectedOutput`: their content is preloaded as context, and they are written by workflow tools, not by 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.
 7. **Self-audit the plan.** If every task were completed exactly as written, the slice goal/demo should be true. Every must-have maps to a task. Inputs and Expected Output are backtick-wrapped file paths.
 8. If refinement produced structural decisions that diverge from the sketch, call `gsd_decision_save` for each; the tool persists the decision and regenerates `.gsd/DECISIONS.md`.
@@ -77,4 +77,4 @@ The slice directory and tasks/ subdirectory already exist. Do NOT mkdir.
 
 **You MUST call `gsd_plan_slice` to persist planning state before finishing.** After success, the pipeline clears the sketch flag on next state derivation; the on-disk PLAN file is the signal.
 
-When done, say: "Slice {{sliceId}} refined."
+When done, say: "Slice {{sliceId}} refined." Say this exactly once — if you already said it in a prior message, do not repeat it.

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

@@ -34,8 +34,8 @@ Consider these captures when rewriting the remaining tasks — they represent th
 
 1. Use the inlined blocker summary excerpt first. Read the full blocker task summary only if the excerpt is absent, marked truncated, or lacks the specific blocker evidence needed to replan.
 2. Analyze the remaining `[ ]` tasks in the slice plan. Determine which are still valid, which need modification, and which should be replaced.
-3. **Persist replan state through `gsd_replan_slice`.** Call it with: `milestoneId`, `sliceId`, `blockerTaskId`, `blockerDescription`, `whatChanged`, `updatedTasks` (array of task objects with taskId, title, description, estimate, files, verify, inputs, expectedOutput), `removedTaskIds` (array of task ID strings). The tool structurally enforces preservation of completed tasks, writes replan history to the DB, re-renders `{{planPath}}`, and renders `{{replanPath}}`. Preserve or update the Threat Surface and Requirement Impact sections if the replan changes the slice's security posture or requirement coverage.
+3. **Persist replan state through `gsd_replan_slice`.** Call it with: `milestoneId`, `sliceId`, `blockerTaskId`, `blockerDescription`, `whatChanged`, `updatedTasks` (array of task objects with taskId, title, description, estimate, files, verify, inputs, expectedOutput), `removedTaskIds` (array of task ID strings). Never list GSD planning artifacts — anything under `.gsd/`, `.planning/`, or `.audits/`, or artifact names like `M001-CONTEXT.md` / `S01-PLAN.md` — in `inputs`, `files`, or `expectedOutput`: their content is preloaded as context, and they are written by workflow tools, not by tasks. The tool structurally enforces preservation of completed tasks, writes replan history to the DB, re-renders `{{planPath}}`, and renders `{{replanPath}}`. Preserve or update the Threat Surface and Requirement Impact sections if the replan changes the slice's security posture or requirement coverage.
 4. If any incomplete task had a `T0x-PLAN.md`, remove or rewrite it to match the new task description.
 5. Do not commit manually — the system auto-commits your changes after this unit completes.
 
-When done, say: "Slice {{sliceId}} replanned."
+When done, say: "Slice {{sliceId}} replanned." Say this exactly once — if you already said it in a prior message, do not repeat it.

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

@@ -46,4 +46,4 @@ Then research the codebase and relevant technologies. Narrate key findings and s
 
 **You MUST call `gsd_summary_save` with the research content before finishing.**
 
-When done, say: "Milestone {{milestoneId}} researched."
+When done, say: "Milestone {{milestoneId}} researched." Say this exactly once — if you already said it in a prior message, do not repeat it.

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

@@ -55,4 +55,4 @@ The slice directory already exists at `{{slicePath}}/`. Do NOT mkdir.
 
 **You MUST call `gsd_summary_save` with the research content before finishing.**
 
-When done, say: "Slice {{sliceId}} researched."
+When done, say: "Slice {{sliceId}} researched." Say this exactly once — if you already said it in a prior message, do not repeat it.

package/dist/resources/extensions/gsd/prompts/rewrite-docs.md

@@ -30,4 +30,4 @@ An override was issued by the user that changes a fundamental decision or approa
 
 **You MUST update the relevant documents AND mark overrides as resolved in `{{overridesPath}}` before finishing.**
 
-When done, say: "Override applied across all documents."
+When done, say: "Override applied across all documents." Say this exactly once — if you already said it in a prior message, do not repeat it.

package/dist/resources/extensions/gsd/prompts/run-uat.md

@@ -27,21 +27,23 @@ You are the UAT runner. Execute every check defined in `{{uatPath}}` as deeply a
 ### Automation rules by mode
 
 - `artifact-driven` — verify with shell commands, scripts, file reads, and artifact structure checks.
-- `browser-executable` — use browser tools to navigate to the target URL and verify expected behavior. Prefer direct `browser_*` tools when available. Capture screenshots as evidence. Record pass/fail with specific assertions.
-- `runtime-executable` — execute the specified command or script. Capture stdout/stderr as evidence. Record pass/fail based on exit code and output.
+- `browser-executable` — use browser tools to navigate to the target URL and verify expected behavior. Prefer direct `browser_*` tools when available. Capture screenshots as evidence. Record pass/fail with specific assertions. When the UAT **Evidence** section names a self-contained runtime harness (for example `npm run test:uat` or `node tests/browser/search-uat.mjs`), treat the effective mode as **runtime-executable** instead: run that command only and do not start a separate server or `browser_navigate` to a hardcoded localhost port.
+- `runtime-executable` — execute the specified command or script. Capture stdout/stderr as evidence. Record pass/fail based on exit code and output. When the verification script starts its own dev/static server (for example `node tests/browser/search-uat.mjs`), do **not** start a separate server with `uat-service-start` — the script owns server lifecycle and binds an ephemeral port. If a server is already running, pass its URL via `UAT_BASE_URL` or `PORT` in the `gsd_uat_exec` environment instead of hardcoding a fixed port like 4173.
 - `live-runtime` — exercise the real runtime path. Start or connect to the app/service if needed, use browser/runtime/network checks, and verify observable behavior.
 - `mixed` — run all automatable artifact-driven and live-runtime checks. Separate any remaining human-only checks explicitly.
 - `human-experience` — automate setup, preconditions, screenshots, logs, and objective checks, but do **not** invent subjective PASS results. Mark taste-based, experiential, or purely human-judgment checks as `NEEDS-HUMAN`. Use an overall verdict of `PASS` when all automatable checks succeed (even if human-only checks remain as `NEEDS-HUMAN`). Use `PARTIAL` only when automatable checks themselves were inconclusive.
 
 ### Evidence tools
 
-Choose the lightest tool that proves the check honestly:
+The **Tool Surface** block prepended above lists unavailable tools for this unit. In short:
 
 - Run automated checks with `gsd_uat_exec`
   - Use `uat-artifact-check` as `intent` for static file, grep, structure, or artifact checks.
   - Use `uat-runtime-check` as `intent` for executing tests, scripts, or runtime assertions.
   - Use `uat-browser-check` as `intent` for browser interaction or screenshot-backed UI checks.
-  - Use `uat-service-start` as `intent` only when starting or connecting to an app/service.
+  - Use `uat-service-start` as `intent` only when starting or connecting to an app/service that the UAT checks will not start themselves.
+  - Do not start a dev/static server separately when a runtime test script owns server lifecycle — run the script directly instead.
+  - When you do start a service, capture the URL from its stdout (for example `Ready at http://127.0.0.1:PORT`) and pass it to downstream checks via `UAT_BASE_URL` rather than assuming a fixed port.
   - Use `uat-log-inspection` as `intent` for checking logs or captured output files.
   - The result-table evidence mode is separate; do not use `artifact`, `runtime`, or `human-follow-up` as `intent`.
 - Run `grep` / `rg` checks against files
@@ -98,4 +100,4 @@ checks: [{
 
 **You MUST call `gsd_uat_result_save` before finishing. Do not write the assessment file directly, and do not call `gsd_summary_save` as a substitute.**
 
-When done, say: "UAT {{sliceId}} complete."
+When done, say: "UAT {{sliceId}} complete." Say this exactly once — if you already said it in a prior message, do not repeat it.

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

@@ -110,9 +110,12 @@ Templates are in `{{templatesDir}}`.
 
 **External facts:** Use `search-the-web` + `fetch_page`, or `search_and_read`; use `freshness` for recency. Never state current facts from training data without verification.
 
-**Background processes:** Use `bg_shell` `start` + `wait_for_ready` for servers/watchers/daemons. Never use `bash` with `&` or `nohup`; inherited stdout can hang. Never poll with `sleep`; use `wait_for_ready`. For status use `digest`, `highlights` for significant lines, and `output` only when debugging.
+**Choosing a shell tool — decide by how the command runs, not how long it takes:**
 
-**One-shot commands:** Use `async_bash` for builds, tests, and installs. Results are pushed on exit; use `await_job` only to block on a specific job.
+- **Runs and exits (a batch command):** anything that does work and finishes — `terraform apply`, DB migrations, builds, tests, installs, long scripts. Use synchronous `bash` when you want to block and read the result now (uncapped; in auto-mode genuine hangs are caught by the stalled-tool watchdog, interactively they end on human ESC). Use `async_bash` when you want it non-blocking — it returns a job ID immediately and you `await_job` later. **Never** put a run-to-completion command under `bg_shell` `wait_for_ready`: it exits instead of staying alive, so readiness never trips and a clean exit looks like a failure.
+- **Stays alive and you interact with it over time:** servers, watchers, daemons, REPLs. Use `bg_shell` `start`, then `wait_for_ready` (block until it listens/prints its ready pattern), `output`/`digest`/`highlights` to inspect, `send`/`send_and_wait` to drive it, and `kill`/`restart` to manage it. Never use `bash` with `&` or `nohup` (inherited stdout can hang); never poll with `sleep` (use `wait_for_ready`).
+
+Quick rule of thumb: if the command would exit on its own, it's `bash`/`async_bash`; if it would keep running until you stop it, it's `bg_shell`.
 
 **Stale job hygiene:** After editing source to fix a failure, `cancel_job` every in-flight `async_bash` job before rerunning. Changed inputs make in-flight outputs untrusted.
 

package/dist/resources/extensions/gsd/prompts/triage-captures.md

@@ -65,4 +65,4 @@ Classify each capture as one of:
 
 **Important:** Do NOT execute any resolutions. Only classify and update CAPTURES.md. Resolution execution happens separately (in auto-mode dispatch or manually by the user).
 
-When done, say: "Triage complete."
+When done, say: "Triage complete." Say this exactly once — if you already said it in a prior message, do not repeat it.

package/dist/resources/extensions/gsd/prompts/validate-milestone.md

@@ -87,4 +87,4 @@ If verdict is `needs-remediation`:
 
 **File system safety:** When scanning milestone directories, use `ls` or `find` first. Never pass a directory path such as `tasks/` or `slices/` directly to `read`; it only accepts files.
 
-When done, say: "Milestone {{milestoneId}} validation complete — verdict: <verdict>."
+When done, say: "Milestone {{milestoneId}} validation complete — verdict: <verdict>." Say this exactly once — if you already said it in a prior message, do not repeat it.

package/dist/resources/extensions/gsd/provider-error-guidance.js

@@ -1,5 +1,6 @@
 /**
  * Actionable remediation hints when auto-mode pauses on provider/model errors.
+ * Render with formatGuidance from the Guidance module.
  */
 /** Map auto unit types to the `models:` key in PREFERENCES.md. */
 export function unitTypeToPrefsPhaseKey(unitType) {
@@ -91,8 +92,3 @@ export function resolveProviderErrorGuidance(input) {
     steps.push("Run /gsd next to resume the paused unit.");
     return { summary, steps };
 }
-/** Flatten guidance into a pause banner / notification string. */
-export function formatProviderErrorGuidance(guidance) {
-    const numbered = guidance.steps.map((step, index) => `${index + 1}. ${step}`).join("\n");
-    return `${guidance.summary}\n\n${numbered}`;
-}

package/dist/resources/extensions/gsd/provider-switch-observer.js

@@ -92,7 +92,7 @@ function emitAudit(report) {
 }
 function emitNotification(report) {
     try {
-        appendNotification(summarize(report), "warning", "workflow-logger");
+        appendNotification(summarize(report), "warning", "workflow-logger", { kind: "provider-switch" });
     }
     catch {
         // Notification persistence is best-effort.

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

@@ -0,0 +1,87 @@
+// Project/App: gsd-pi
+// File Purpose: Publication module — push and draft-PR creation for a merged milestone (ADR-034).
+//
+// Publication is distinct from the merge verb: merging a milestone branch is a
+// Worktree Lifecycle concern (worktree, branch, lease, stash discipline);
+// publication needs only the resulting commit, a remote, and preferences.
+// Extracted from the tail of `mergeMilestoneToMain` (auto-worktree.ts) so it is
+// testable without a merge fixture and reachable from any closeout path.
+//
+// Publication failure is non-fatal to a completed local merge: every failure is
+// logged and reported in the result, never thrown.
+import { execFileSync } from "node:child_process";
+import { buildPullRequestEvidence, createDraftPullRequestFromEvidence, } from "./pull-request-process.js";
+import { logWarning } from "./workflow-logger.js";
+export function gitRemoteExists(basePath, remote) {
+    try {
+        execFileSync("git", ["remote", "get-url", remote], {
+            cwd: basePath,
+            stdio: ["ignore", "pipe", "pipe"],
+            encoding: "utf-8",
+        });
+        return true;
+    }
+    catch {
+        return false;
+    }
+}
+/**
+ * Publish a merged milestone per preferences: push the integration branch
+ * (auto-push, suppressed when auto-PR owns the remote interaction) and/or
+ * push the milestone branch and open a draft PR (auto-PR).
+ */
+export function publishMilestone(request) {
+    const result = { pushed: false, prCreated: false };
+    if (request.nothingToCommit)
+        return result;
+    const { basePath, prefs } = request;
+    const remote = prefs.remote ?? "origin";
+    if (prefs.autoPush && !prefs.autoPr && gitRemoteExists(basePath, remote)) {
+        try {
+            execFileSync("git", ["push", remote, request.integrationBranch], {
+                cwd: basePath,
+                stdio: ["ignore", "pipe", "pipe"],
+                encoding: "utf-8",
+            });
+            result.pushed = true;
+        }
+        catch (err) {
+            // Push failure is non-fatal
+            logWarning("worktree", `git push failed: ${err instanceof Error ? err.message : String(err)}`);
+        }
+    }
+    // #2302: PR creation is not gated on pushed/auto-push.
+    if (prefs.autoPr && gitRemoteExists(basePath, remote)) {
+        try {
+            // Push the milestone branch to remote first
+            execFileSync("git", ["push", remote, request.milestoneBranch], {
+                cwd: basePath,
+                stdio: ["ignore", "pipe", "pipe"],
+                encoding: "utf-8",
+            });
+            const prEvidence = buildPullRequestEvidence({
+                milestoneId: request.milestoneId,
+                milestoneTitle: request.milestoneTitle,
+                changeType: "feat",
+                summaries: [...request.sliceSummaries],
+                testsRun: ["Auto-created after milestone merge. Run `npm run verify:merge` before marking this draft ready."],
+                rollbackNotes: ["Close the draft PR or revert the merge commit if review finds a behavior regression."],
+                how: "Generated by git.auto_pr after the milestone branch was pushed and merged locally.",
+            });
+            const prUrl = createDraftPullRequestFromEvidence(basePath, request.milestoneId, prEvidence, {
+                head: request.milestoneBranch,
+                base: prefs.prTargetBranch ?? request.integrationBranch,
+            });
+            if (!prUrl) {
+                throw new Error("gh pr create returned no URL");
+            }
+            result.prCreated = true;
+            result.prUrl = prUrl;
+        }
+        catch (err) {
+            // PR creation failure is non-fatal — gh may not be installed or authenticated
+            logWarning("worktree", `PR creation failed: ${err instanceof Error ? err.message : String(err)}`);
+        }
+    }
+    return result;
+}

package/dist/resources/extensions/gsd/reactive-graph.js

@@ -10,7 +10,11 @@
  */
 import { loadFile, parseTaskPlanIO } from "./files.js";
 import { isDbAvailable, getSliceTasks } from "./gsd-db.js";
+// Explicit degraded-mode fallback ONLY (ADR-017): parsePlan is consulted when
+// the DB is unavailable or has no task rows for the slice; the DB path in
+// loadSliceTaskIO always wins when rows exist, and the fallback logs a warning.
 import { parsePlan } from "./parsers-legacy.js";
+import { logWarning } from "./workflow-logger.js";
 import { resolveTasksDir, resolveTaskFiles } from "./paths.js";
 import { join } from "node:path";
 import { loadJsonFileOrNull, saveJsonFile } from "./json-persistence.js";
@@ -178,7 +182,10 @@ export async function loadSliceTaskIO(basePath, mid, sid) {
     }
     catch { /* fall through */ }
     if (!taskEntries) {
-        // File-based fallback: parse slice plan for task entries
+        // Degraded-mode fallback (ADR-017): no DB task rows for this slice —
+        // parse the PLAN.md projection instead. Warn so silent drift from the
+        // DB-authoritative path is visible in logs.
+        logWarning("projection", `reactive-graph: DB unavailable or no task rows for ${mid}/${sid}; falling back to PLAN.md parsing`);
         const parsed = parsePlan(planContent);
         if (parsed.tasks.length > 0) {
             taskEntries = parsed.tasks.map(t => ({

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

@@ -1,7 +1,10 @@
 // Project/App: gsd-pi
 // File Purpose: ADR-015 Recovery Classification module for runtime failure taxonomy.
+import { isToolUnavailableError } from "./auto-tool-tracking.js";
 import { classifyError, isTransient } from "./error-classifier.js";
+import { recoveryRemediation } from "./guidance.js";
 import { ReconciliationFailedError } from "./state-reconciliation.js";
+import { IllegalPhaseTransitionError } from "./state-transition-matrix.js";
 export function classifyFailure(input) {
     const message = errorMessage(input.error);
     // ADR-017: ReconciliationFailedError is a typed throw from the State
@@ -9,96 +12,47 @@ export function classifyFailure(input) {
     // failureKind so the taxonomy stays consistent.
     const failureKind = input.error instanceof ReconciliationFailedError
         ? "reconciliation-drift"
-        : input.failureKind ?? inferFailureKind(message);
-    switch (failureKind) {
-        case "tool-schema":
-            return {
-                failureKind,
-                action: "stop",
-                reason: `Tool schema failure${unitSuffix(input)}: ${message}`,
-                exitReason: "tool-schema",
-                remediation: "Fix the Unit Tool Contract or tool schema before retrying.",
-            };
-        case "tool-contract":
-            return {
-                failureKind,
-                action: "stop",
-                reason: `Tool Contract failure${unitSuffix(input)}: ${message}`,
-                exitReason: "tool-contract",
-                remediation: "Fix the Unit Tool Contract or prompt so the Unit is only asked to use tools owned by its phase.",
-            };
-        case "deterministic-policy":
-            return {
-                failureKind,
-                action: "stop",
-                reason: `Deterministic policy failure${unitSuffix(input)}: ${message}`,
-                exitReason: "deterministic-policy",
-                remediation: "Resolve the policy blocker; retrying the same Unit will repeat the failure.",
-            };
-        case "lifecycle-progression":
-            return {
-                failureKind,
-                action: "stop",
-                reason: `Lifecycle progression failure${unitSuffix(input)}: ${message}`,
-                exitReason: "lifecycle-progression",
-                remediation: "Route to the required owning Unit or restore the missing artifact before advancing lifecycle state.",
-            };
-        case "stale-worker":
-            return {
-                failureKind,
-                action: "stop",
-                reason: `Stale worker failure${unitSuffix(input)}: ${message}`,
-                exitReason: "stale-worker",
-                remediation: "Clear or reconcile the stale worker before dispatching another Unit.",
-            };
-        case "worktree-invalid":
-            return {
-                failureKind,
-                action: "stop",
-                reason: `Worktree invalid${unitSuffix(input)}: ${message}`,
-                exitReason: "worktree-invalid",
-                remediation: "Repair or recreate the milestone worktree before launching source-writing Units.",
-            };
-        case "verification-drift":
-            return {
-                failureKind,
-                action: "escalate",
-                reason: `Verification drift${unitSuffix(input)}: ${message}`,
-                exitReason: "verification-drift",
-                remediation: "Inspect the verification artifact and reconcile the state snapshot before resuming.",
-            };
-        case "reconciliation-drift":
-            return {
-                failureKind,
-                action: "escalate",
-                reason: `Reconciliation drift${unitSuffix(input)}: ${message}`,
-                exitReason: "reconciliation-drift",
-                remediation: "Inspect the persistent or repair-failed drift kinds reported by the State Reconciliation Module before resuming.",
-            };
-        case "provider": {
-            const providerClass = classifyError(message, input.retryAfterMs);
-            return {
-                failureKind,
-                action: isTransient(providerClass) ? "retry" : "escalate",
-                reason: message,
-                exitReason: `provider-${providerClass.kind}`,
-                remediation: isTransient(providerClass)
-                    ? "Retry after the provider/network condition clears."
-                    : "Inspect provider credentials, model entitlement, or request shape.",
-                providerClass: providerClass.kind,
-            };
-        }
-        case "runtime-unknown":
-            return {
-                failureKind,
-                action: "escalate",
-                reason: message,
-                exitReason: "runtime-unknown",
-                remediation: "Inspect the runtime error and add a dedicated classification if it is repeatable.",
-            };
+        : input.error instanceof IllegalPhaseTransitionError
+            ? "illegal-transition"
+            : input.failureKind ?? inferFailureKind(message);
+    if (failureKind === "provider") {
+        const providerClass = classifyError(message, input.retryAfterMs);
+        const transient = isTransient(providerClass);
+        return {
+            failureKind,
+            action: transient ? "retry" : "escalate",
+            reason: message,
+            exitReason: `provider-${providerClass.kind}`,
+            remediation: recoveryRemediation(transient ? "provider-transient" : "provider-permanent"),
+            providerClass: providerClass.kind,
+        };
     }
+    const { action, label } = FAILURE_TAXONOMY[failureKind];
+    return {
+        failureKind,
+        action,
+        reason: label ? `${label}${unitSuffix(input)}: ${message}` : message,
+        exitReason: failureKind,
+        remediation: recoveryRemediation(failureKind),
+    };
 }
+/** Per-kind action and reason label. Remediation lives in the Guidance module. */
+const FAILURE_TAXONOMY = {
+    "tool-schema": { action: "stop", label: "Tool schema failure" },
+    "tool-contract": { action: "stop", label: "Tool Contract failure" },
+    "tool-unavailable": { action: "retry", label: "Tool unavailable" },
+    "deterministic-policy": { action: "stop", label: "Deterministic policy failure" },
+    "lifecycle-progression": { action: "stop", label: "Lifecycle progression failure" },
+    "stale-worker": { action: "stop", label: "Stale worker failure" },
+    "worktree-invalid": { action: "stop", label: "Worktree invalid" },
+    "verification-drift": { action: "escalate", label: "Verification drift" },
+    "reconciliation-drift": { action: "escalate", label: "Reconciliation drift" },
+    "illegal-transition": { action: "escalate", label: "Illegal phase transition" },
+    "runtime-unknown": { action: "escalate", label: null },
+};
 function inferFailureKind(message) {
+    if (isToolUnavailableError(message))
+        return "tool-unavailable";
     if (/tool contract|auto-unit tool scope|phase-boundary gate|not permitted.*own/i.test(message))
         return "tool-contract";
     if (/lifecycle progression|required artifact|missing .*assessment|missing .*closeout|cannot legally (?:advance|progress)/i.test(message))

package/dist/resources/extensions/gsd/roadmap-slices.js

@@ -161,6 +161,26 @@ function parseTableSlices(section) {
     }
     return slices;
 }
+function looksLikeTable(section) {
+    const lines = section.split("\n");
+    // Checkbox format takes precedence — embedded demo tables must not switch mode (#721).
+    if (lines.some(line => /^\s*-\s+\[[ xX]\]/.test(line))) {
+        return false;
+    }
+    const pipeLines = lines.filter(line => /^\s*\|/.test(line));
+    if (pipeLines.length < 2)
+        return false;
+    const hasSeparatorRow = pipeLines.some((line, index) => {
+        if (index === 0)
+            return false;
+        const cells = line.split("|").map(c => c.trim()).filter(Boolean);
+        return /^\s*\|[\s:-]+\|/.test(line) && cells.length >= 2 && cells.every(c => /^[\s:-]+$/.test(c));
+    });
+    if (hasSeparatorRow)
+        return true;
+    // Tables without a separator row still expose slice IDs in data rows.
+    return pipeLines.some(line => /\bS\d+\b/.test(line));
+}
 export function parseRoadmapSlices(content) {
     const slicesSection = extractSlicesSection(content);
     if (!slicesSection) {
@@ -171,9 +191,11 @@ export function parseRoadmapSlices(content) {
     }
     // Try table format first — if the section contains pipe-delimited rows with
     // slice IDs, parse them as a table (#1736).
-    const tableSlices = parseTableSlices(slicesSection);
-    if (tableSlices.length > 0) {
-        return tableSlices;
+    if (looksLikeTable(slicesSection)) {
+        const tableSlices = parseTableSlices(slicesSection);
+        if (tableSlices.length > 0) {
+            return tableSlices;
+        }
     }
     // Standard checkbox format
     const slices = [];

package/dist/resources/extensions/gsd/safety/destructive-confirmation.js

@@ -0,0 +1,108 @@
+/**
+ * One-shot confirmation token for destructive bash commands.
+ *
+ * The destructive-command guard hard-blocks classified commands (force push,
+ * rm -rf, SQL drop, etc.) in all modes. The block instructs the model to
+ * confirm via ask_user_questions and re-issue the command. This module is the
+ * missing escape hatch: it records the user's confirmation and lets the exact
+ * confirmed command through exactly once.
+ *
+ * Design constraints:
+ *  - In-memory only, never persisted. A confirmation token written to disk
+ *    could silently auto-approve a destructive command in a later session —
+ *    confirmation must be re-obtained every process lifetime.
+ *  - One-shot. Consuming a token clears it, so a second destructive command
+ *    (even an identical one) re-blocks and re-prompts.
+ *  - Command-bound. The token only matches the exact (normalized) command
+ *    string the user confirmed. A reworded command re-blocks, which is safe.
+ *  - Per basePath, so concurrent workspaces in one process never share tokens.
+ *
+ * Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
+ */
+import { resolve } from "node:path";
+/**
+ * Question-id substring that marks an ask_user_questions call as a
+ * destructive-command confirmation. The tool_result handler promotes the
+ * pending command to a confirmed token when an affirmative answer arrives for
+ * a question whose id contains this marker.
+ */
+export const DESTRUCTIVE_CONFIRM_GATE_MARKER = "destructive_confirm";
+const statesByBasePath = new Map();
+function stateKey(basePath) {
+    return resolve(basePath);
+}
+function getState(basePath) {
+    const key = stateKey(basePath);
+    let state = statesByBasePath.get(key);
+    if (!state) {
+        state = { pendingCommand: null, confirmedCommand: null };
+        statesByBasePath.set(key, state);
+    }
+    return state;
+}
+/**
+ * Normalize a command for stable matching across block → confirm → retry.
+ * Trims surrounding whitespace and collapses internal runs of whitespace so
+ * cosmetic reformatting of the same command still matches the token.
+ */
+export function normalizeDestructiveCommand(command) {
+    return command.replace(/\s+/g, " ").trim();
+}
+/**
+ * Whether an ask_user_questions question id is a destructive-confirm gate.
+ */
+export function isDestructiveConfirmGateId(questionId) {
+    return typeof questionId === "string" && questionId.includes(DESTRUCTIVE_CONFIRM_GATE_MARKER);
+}
+/**
+ * Record that a destructive command was blocked and is awaiting confirmation.
+ * Called by the guard at block time. Overwrites any prior pending command —
+ * only the most recently blocked command can be confirmed.
+ */
+export function requestDestructiveConfirmation(command, basePath = process.cwd()) {
+    const state = getState(basePath);
+    state.pendingCommand = normalizeDestructiveCommand(command);
+    // A fresh request invalidates any stale confirmed token for a different
+    // command so confirmation cannot leak across distinct destructive actions.
+    state.confirmedCommand = null;
+}
+/**
+ * Promote the pending command to a confirmed, one-shot token. Called by the
+ * tool_result handler when the user gives an affirmative answer to a
+ * destructive-confirm gate. Returns the confirmed command, or null if there
+ * was nothing pending (e.g. confirmation arrived without a preceding block).
+ */
+export function confirmDestructiveCommand(basePath = process.cwd()) {
+    const state = getState(basePath);
+    if (!state.pendingCommand)
+        return null;
+    state.confirmedCommand = state.pendingCommand;
+    state.pendingCommand = null;
+    return state.confirmedCommand;
+}
+/**
+ * Check whether the given command has been confirmed, consuming the token if
+ * so. Returns true exactly once per confirmation; subsequent calls (or a
+ * non-matching command) return false. Called by the guard before blocking.
+ */
+export function consumeDestructiveConfirmation(command, basePath = process.cwd()) {
+    const state = getState(basePath);
+    if (!state.confirmedCommand)
+        return false;
+    if (state.confirmedCommand !== normalizeDestructiveCommand(command))
+        return false;
+    state.confirmedCommand = null;
+    return true;
+}
+/**
+ * Inspect the pending command without consuming it (diagnostics/tests).
+ */
+export function peekPendingDestructiveCommand(basePath = process.cwd()) {
+    return getState(basePath).pendingCommand;
+}
+/**
+ * Clear all destructive-confirmation state for a basePath (tests / flow reset).
+ */
+export function resetDestructiveConfirmation(basePath = process.cwd()) {
+    statesByBasePath.delete(stateKey(basePath));
+}