@lifeaitools/rdc-skills 0.35.3 → 0.35.5

package/guides/agent-bootstrap.md

@@ -1,11 +1,11 @@
-# Agent Bootstrap — Read This First
-> Every dispatched agent reads this before their role-specific guide.
-> Base guide for rdc-skills — provides credential, git, and reporting patterns across projects.
-
----
-
-## Who You Are
-
+# Agent Bootstrap — Read This First
+> Every dispatched agent reads this before their role-specific guide.
+> Base guide for rdc-skills — provides credential, git, and reporting patterns across projects.
+
+---
+
+## Who You Are
+
 You are a subagent dispatched by the rdc:build supervisor. You have a specific
 scope (files, package, feature) that will be in your prompt. Stay in that scope.
 NEVER modify files outside it.
@@ -15,198 +15,242 @@ RDC implementation posture for assumptions, minimal changes, surgical scope,
 verification evidence, and escalation.
 
 ---
-
-## Credentials — Daemon Access Pattern
-
-You do NOT have access to cloud MCP connectors. Instead, all credentials
-come from a daemon running locally (typically on localhost:52437).
-
-**Ping first to confirm availability:**
-```bash
-curl -s http://127.0.0.1:52437/ping
-```
-If it doesn't respond — report BLOCKED, do not proceed.
-
-**Get a credential:**
-```bash
-curl -s http://127.0.0.1:52437/get/<service>
-```
-
-**Pattern for extracting key/value without printing:**
-```bash
-# Correct pattern — never echo the key
-KEY=$(curl -s http://127.0.0.1:52437/get/<service> | python3 -c "import sys,json; print(json.load(sys.stdin)['key'])")
-curl -s -H "Authorization: Bearer $KEY" https://api.example.com/...
-```
-
-**Never print credentials to stdout.** Capture to a variable, use inline, discard.
-
----
-
-## Project Directory Convention
-
-This plugin uses the `.rdc/` directory convention. Check for it first:
-
-```bash
-# Check if .rdc/ exists at project root
-ls {PROJECT_ROOT}/.rdc/config.json 2>/dev/null && echo "using .rdc/" || echo "using docs/ fallback"
-```
-
-**Path resolution rule:**
-- Guides: `{PROJECT_ROOT}/.rdc/guides/` → fallback: `{PROJECT_ROOT}/docs/guides/`
-- Plans: `{PROJECT_ROOT}/.rdc/plans/` → fallback: `{PROJECT_ROOT}/docs/plans/`
-- Reports: `{PROJECT_ROOT}/.rdc/reports/` → fallback: `{PROJECT_ROOT}/docs/reports/`
-- Research: `{PROJECT_ROOT}/.rdc/research/` → fallback: `{PROJECT_ROOT}/docs/research/`
-
-If `.rdc/config.json` exists, read it for project metadata (name, description, conventions).
-
----
-
-## Database Access — Check Project Overlay
-
-The project overlay guide will specify:
-- Database project reference / instance name
-- Whether to use MCP connectors or daemon
-- Available RPC functions
-- Work item management patterns
-
-Read the project-specific agent-bootstrap.md overlay for exact connection details.
-
----
-
-## Git Rules
-
-- Branch: Always use the project's primary development branch (typically `develop` or `main`)
-- Auto-commit after completing your scope — no confirmation needed
-- Commit message must use conventional format: `feat/fix/chore/refactor(<scope>): description`
-- Push to origin after committing
-- NEVER force-push
-
----
-
+
+## Credentials — Daemon Access Pattern
+
+You do NOT have access to cloud MCP connectors. Instead, all credentials
+come from a daemon running locally (typically on localhost:52437).
+
+**Ping first to confirm availability:**
+```bash
+curl -s http://127.0.0.1:52437/ping
+```
+If it doesn't respond — report BLOCKED, do not proceed.
+
+**Get a credential:**
+```bash
+curl -s http://127.0.0.1:52437/get/<service>
+```
+
+**Pattern for extracting key/value without printing:**
+```bash
+# Correct pattern — never echo the key
+KEY=$(curl -s http://127.0.0.1:52437/get/<service> | python3 -c "import sys,json; print(json.load(sys.stdin)['key'])")
+curl -s -H "Authorization: Bearer $KEY" https://api.example.com/...
+```
+
+**Never print credentials to stdout.** Capture to a variable, use inline, discard.
+
+---
+
+## Project Directory Convention
+
+This plugin uses the `.rdc/` directory convention. Check for it first:
+
+```bash
+# Check if .rdc/ exists at project root
+ls {PROJECT_ROOT}/.rdc/config.json 2>/dev/null && echo "using .rdc/" || echo "using docs/ fallback"
+```
+
+**Path resolution rule:**
+- Guides: `{PROJECT_ROOT}/.rdc/guides/` → fallback: `{PROJECT_ROOT}/docs/guides/`
+- Plans: `{PROJECT_ROOT}/.rdc/plans/` → fallback: `{PROJECT_ROOT}/docs/plans/`
+- Reports: `{PROJECT_ROOT}/.rdc/reports/` → fallback: `{PROJECT_ROOT}/docs/reports/`
+- Research: `{PROJECT_ROOT}/.rdc/research/` → fallback: `{PROJECT_ROOT}/docs/research/`
+
+If `.rdc/config.json` exists, read it for project metadata (name, description, conventions).
+
+---
+
+## Database Access — Check Project Overlay
+
+The project overlay guide will specify:
+- Database project reference / instance name
+- Whether to use MCP connectors or daemon
+- Available RPC functions
+- Work item management patterns
+
+Read the project-specific agent-bootstrap.md overlay for exact connection details.
+
+---
+
+## Git Rules
+
+- Branch: Always use the project's primary development branch (typically `develop` or `main`)
+- Auto-commit after completing your scope — no confirmation needed
+- Commit message must use conventional format: `feat/fix/chore/refactor(<scope>): description`
+- Push to origin after committing
+- NEVER force-push
+
+---
+
 ## Build Rules
 
 Never run `pnpm build` or equivalent full builds locally — they consume excessive memory.
 Type-check only: `npx tsc --noEmit --project <path>/tsconfig.json`
 Run tests only for modified packages: modify tests in isolation, not whole suite.
 
-### No Foreground Windows
+### Terminal/process launches — caller-logged, not hard-blocked
 
-Agent-launched processes must not steal focus. This is a hard local-operator
-rule, not a preference.
+**Narrowed 2026-08-26** (epic 688ad6da, lifeai-env; direct operator
+instruction: "remove the PreToolUse foreground-window guard entirely...
+replace it with caller-logging for traceability"). The old hard block on
+every raw `Start-Process`/`cmd /c start`/window-focus API/bare `.ps1` launch
+is retired — `foreground-process-gate.js` (rdc-skills) no longer enforces any
+of it. The replacement isn't a weaker rule; it's a different mechanism:
+lifeai-env's `lib/TermLaunch.psm1` (`Invoke-TermLaunchHidden` /
+`Invoke-TermLaunchInteractive`) logs every caller (script + line + argv) to
+`C:/Dev/.logs` **before** spawning, for every launch — which answers "who
+launched this and with what" for every case, not just the ones a regex
+pattern happened to catch.
 
-- Playwright must run headless. Do not use `--headed`, `--ui`, `codegen`, `open`, `show-report`, or `PWDEBUG=1` in agent sessions.
-- Use list/dot/json reporters and saved trace/report artifacts instead of opening the Playwright UI.
-- PowerShell helpers must use `-WindowStyle Hidden -NonInteractive`, or a hidden wrapper.
-- `Start-Process` must include `-WindowStyle Hidden` or `-WindowStyle Minimized`.
-- `cmd /c start` must use `/min` for intentionally visible tools or `/b` for background work.
-- Node/cmd/ps1 helpers launched by hooks must go through the RDC hidden hook runner.
+- Use `Invoke-TermLaunchHidden -Command <cmd> [-Arguments][-WorkingDirectory]`
+  for a background/no-window process, and `Invoke-TermLaunchInteractive
+  [-Title]` for a visible new terminal tab — never a raw `Start-Process`/
+  `wt.exe` call. Neither primitive accepts an arbitrary inline multi-line
+  command string, only a flat command + argument array or a `-File <script>`
+  path — this structurally prevents the `wt.exe -Command` argv-mangling bug
+  class, which no amount of pattern-blocking ever fully closed.
+- Node/cmd/ps1 helpers launched by hooks must still go through the RDC hidden
+  hook runner (`hooks/run-bash-hidden.ps1` and its RdcRun contract) — that
+  convention is unchanged.
+
+**One safety property survives, unrelated to the above and never retired:**
+Playwright must still run headless in agent sessions. Do not use `--headed`,
+`--ui`, `codegen`, `open`, `show-report`, or `PWDEBUG=1`. Use list/dot/json
+reporters and saved trace/report artifacts instead of opening the Playwright
+UI. `foreground-process-gate.js` still hard-blocks this — it was never a
+terminal-launch-primitive question, so narrowing the launch rules above never
+touched it.
 
 Check the project overlay for specific language, package manager, and build constraints.
-
----
-
-## RDC_TEST Sandbox Contract
-
-When `RDC_TEST=1` is set, the skill is running inside the Tier 2 sandbox harness. Short-circuit all destructive external side effects.
-
-**How to detect:** `process.env.RDC_TEST === '1'` (Node/JS) or `[ "$RDC_TEST" = "1" ]` (bash). The Tier 2 runner sets this before invoking `claude --print`.
-
-**Allowed under RDC_TEST** — run normally, no guard:
-
-| Operation | Why |
-|-----------|-----|
-| Local git commits | Commit to the worktree branch; never pushed |
-| Local filesystem writes | Inside the sandbox worktree |
-| Supabase reads | Assertions read prod DB; no writes |
-| Service reads | Coolify status, CF DNS lookup, GitHub read — all safe |
-| `clauth` credential reads | Read-only against the daemon |
-
-**Short-circuit under RDC_TEST** — must be wrapped in a guard:
-
-- Supabase INSERT/UPDATE/DELETE (work items, prototype_registry, design_context, etc.) — runner uses main-db mode, no test branch
-- Coolify app deploys / restarts / config writes
-- Cloudflare DNS writes / cache purges
-- R2 object writes / deletes
-- `git push` to any remote
-- `gh pr create`, `gh release create`, tag pushes
-- `npm publish`
-- Any webhook fire, email send, Slack post
-- MCP connector tools that mutate external state (e.g. `mcp__coolify__deploy`)
-
-**Bash pattern:**
-```bash
-if [ "$RDC_TEST" != "1" ]; then
+
+### Harness Use — Global Policy (applies to every rdc:* skill)
+
+If ANY step of the skill you are executing — regardless of which skill —
+needs to materialize a real product shape, open a signed edit session, run a
+target's own declared build gates, or deploy to dev-PM2/npm-registry, you
+MUST use the real, tested `rdc-harness` CLI instead of hand-rolled bash/curl:
+
+```bash
+node C:/Dev/rdc-harness/bin/rdc-harness.mjs <create|open|edit|build|deploy> <slug> --monorepo-root <your own worktree, never the shared checkout>
+```
+
+One JSON receipt per call, exit 0/1 — use it as evidence for whatever
+checklist/commit step it satisfies. This is a genuinely FORCED requirement,
+not a suggestion: a skill that hand-rolls its own create/open/build/deploy
+logic when this applies is doing exactly the duplicated, unproven work this
+policy exists to stop.
+
+This does NOT apply to a skill with no create/open/build/deploy step at all
+(status/report/help/analysis/media/conversation skills) — do not invent one.
+`open`/`edit` require `RDC_HARNESS_ISSUER_SECRET` set explicitly, no default.
+`deploy` has zero Coolify awareness (Coolify stays `rdc:deploy`'s own path)
+and no live co-editing surface outside `site-html`/`site-ts` targets.
+
+---
+
+## RDC_TEST Sandbox Contract
+
+When `RDC_TEST=1` is set, the skill is running inside the Tier 2 sandbox harness. Short-circuit all destructive external side effects.
+
+**How to detect:** `process.env.RDC_TEST === '1'` (Node/JS) or `[ "$RDC_TEST" = "1" ]` (bash). The Tier 2 runner sets this before invoking `claude --print`.
+
+**Allowed under RDC_TEST** — run normally, no guard:
+
+| Operation | Why |
+|-----------|-----|
+| Local git commits | Commit to the worktree branch; never pushed |
+| Local filesystem writes | Inside the sandbox worktree |
+| Supabase reads | Assertions read prod DB; no writes |
+| Service reads | Coolify status, CF DNS lookup, GitHub read — all safe |
+| `clauth` credential reads | Read-only against the daemon |
+
+**Short-circuit under RDC_TEST** — must be wrapped in a guard:
+
+- Supabase INSERT/UPDATE/DELETE (work items, prototype_registry, design_context, etc.) — runner uses main-db mode, no test branch
+- Coolify app deploys / restarts / config writes
+- Cloudflare DNS writes / cache purges
+- R2 object writes / deletes
+- `git push` to any remote
+- `gh pr create`, `gh release create`, tag pushes
+- `npm publish`
+- Any webhook fire, email send, Slack post
+- MCP connector tools that mutate external state (e.g. `mcp__coolify__deploy`)
+
+**Bash pattern:**
+```bash
+if [ "$RDC_TEST" != "1" ]; then
   curl -X POST "$DEPLOY_API_BASE/api/v1/applications/$UUID/deploy" ...
-else
-  echo "[RDC_TEST] skipping Coolify deploy"
-fi
-```
-
-**Node/JS pattern:**
-```js
-if (process.env.RDC_TEST !== '1') {
-  await deployToCoolify(...);
-} else {
-  console.log('[RDC_TEST] skipping Coolify deploy');
-}
-```
-
-**Why this matters:** Tier 2 runs every skill in a throwaway sandbox. If your skill fires a real deploy or DNS change under `RDC_TEST`, the test isn't a test — it's a production incident.
-
-**New-skill contract:** every new `rdc:*` skill MUST honor `RDC_TEST` before shipping. Tier 2 manifests will fail any skill that writes to external state under the flag.
-
+else
+  echo "[RDC_TEST] skipping Coolify deploy"
+fi
+```
+
+**Node/JS pattern:**
+```js
+if (process.env.RDC_TEST !== '1') {
+  await deployToCoolify(...);
+} else {
+  console.log('[RDC_TEST] skipping Coolify deploy');
+}
+```
+
+**Why this matters:** Tier 2 runs every skill in a throwaway sandbox. If your skill fires a real deploy or DNS change under `RDC_TEST`, the test isn't a test — it's a production incident.
+
+**New-skill contract:** every new `rdc:*` skill MUST honor `RDC_TEST` before shipping. Tier 2 manifests will fail any skill that writes to external state under the flag.
+
 **Known blocker:** Project-specific cwd hooks must check `process.env.RDC_TEST === '1'` and call `process.exit(0)` early to allow Tier 2 sandbox runs. Without this bypass, headless self-test invocations can fail before the skill loads. File: `~/.claude/hooks/check-cwd.js`.
-
----
-
-## Completion Report
-
-When your scope is done, return a structured report to the supervisor:
-
-```
-AGENT_COMPLETE: {
-  scope: "<what you were assigned>",
-  files_changed: ["path/to/file", ...],
-  work_item_id: "<id if you had one>",
-  commits: ["<hash> <message>"],
-  blockers: ["<anything that needs supervisor attention>"]
-}
-```
-
-If you hit a blocker mid-task: stop, report it, do not guess or work around it.
-
----
-
-## Self-Check Rules — Prevent Getting Lost
-
-### 10-Minute Rule
-If you have been working on a **single step** for more than 10 minutes without measurable progress (no new files changed, no successful tool calls, no forward movement), **stop immediately**. Do not keep trying variations. Report it as a blocker.
-
-### 2-Retry Rule
-If the **same command or approach fails twice**, stop. Do not attempt a third variation or creative workaround. Report the failure with the exact error output.
-
-### Scope Drift Rule
-If you discover that fixing your assigned task would also require changing files **outside your scope**, stop. Do not fix them. Add them to `blockers` in your AGENT_COMPLETE report. The supervisor assigns them separately.
-
-### What "measurable progress" means
-- A file was created or modified ✅
-- A tool call succeeded and returned useful data ✅
-- A command ran without error ✅
-- Trying the same thing with slightly different parameters ❌
-- Reading the same file again hoping for different insight ❌
-- Rephrasing a failing query ❌
-
----
-
+
+---
+
+## Completion Report
+
+When your scope is done, return a structured report to the supervisor:
+
+```
+AGENT_COMPLETE: {
+  scope: "<what you were assigned>",
+  files_changed: ["path/to/file", ...],
+  work_item_id: "<id if you had one>",
+  commits: ["<hash> <message>"],
+  blockers: ["<anything that needs supervisor attention>"]
+}
+```
+
+If you hit a blocker mid-task: stop, report it, do not guess or work around it.
+
+---
+
+## Self-Check Rules — Prevent Getting Lost
+
+### 10-Minute Rule
+If you have been working on a **single step** for more than 10 minutes without measurable progress (no new files changed, no successful tool calls, no forward movement), **stop immediately**. Do not keep trying variations. Report it as a blocker.
+
+### 2-Retry Rule
+If the **same command or approach fails twice**, stop. Do not attempt a third variation or creative workaround. Report the failure with the exact error output.
+
+### Scope Drift Rule
+If you discover that fixing your assigned task would also require changing files **outside your scope**, stop. Do not fix them. Add them to `blockers` in your AGENT_COMPLETE report. The supervisor assigns them separately.
+
+### What "measurable progress" means
+- A file was created or modified ✅
+- A tool call succeeded and returned useful data ✅
+- A command ran without error ✅
+- Trying the same thing with slightly different parameters ❌
+- Reading the same file again hoping for different insight ❌
+- Rephrasing a failing query ❌
+
+---
+
 ## ⛔ Implementation Report + CodeFlow Exit Contract
 
 Every implementation agent MUST follow this protocol before moving a work item
 to `review`. Agents do not close non-epic work as `done`; validators close it
 after fresh verification.
-
-### Step 1 — Tick checklist items as you complete them
-
+
+### Step 1 — Tick checklist items as you complete them
+
 ```sql
 SELECT update_checklist_item(
   '<work-item-id>'::uuid,
@@ -221,10 +265,10 @@ SELECT update_checklist_item(
 Call this for each item AS you complete it — not all at once at the end. The
 database records every tick in `work_item_checklist_events`. Supervisor and
 validator re-ticks are rejected by the exit gate.
-
-### Step 2 — Submit implementation report BEFORE marking done
-
-```sql
+
+### Step 2 — Submit implementation report BEFORE marking done
+
+```sql
 SELECT submit_implementation_report(
   '<work-item-id>'::uuid,
   '{
@@ -276,20 +320,20 @@ SELECT update_work_item_status(
 If any `required: true` checklist item is still unchecked, was re-ticked by a
 supervisor/validator, or was ticked by a different session than the originating
 agent, the DB and PreToolUse hook reject the close.
-
-### Supervisor workflow
-
-- All zeros → clean run, proceed
-- `flags_count > 0` or `deviations_count > 0` → pull full report:
-  ```sql
-  SELECT implementation_report FROM work_items WHERE id = '<id>';
-  ```
-
----
-
-## Now read your role-specific guide
-
-Path: `{PROJECT_ROOT}/.rdc/guides/<type>.md` (e.g., `frontend.md`, `backend.md`, `data.md`)
-Fallback: `{PROJECT_ROOT}/docs/guides/<type>.md` if `.rdc/` does not exist.
-
-The project overlay will specify the exact location if it differs from the convention above.
+
+### Supervisor workflow
+
+- All zeros → clean run, proceed
+- `flags_count > 0` or `deviations_count > 0` → pull full report:
+  ```sql
+  SELECT implementation_report FROM work_items WHERE id = '<id>';
+  ```
+
+---
+
+## Now read your role-specific guide
+
+Path: `{PROJECT_ROOT}/.rdc/guides/<type>.md` (e.g., `frontend.md`, `backend.md`, `data.md`)
+Fallback: `{PROJECT_ROOT}/docs/guides/<type>.md` if `.rdc/` does not exist.
+
+The project overlay will specify the exact location if it differs from the convention above.

package/hooks/foreground-process-gate.js

@@ -1,9 +1,28 @@
 #!/usr/bin/env node
 /**
- * PreToolUse hook — block focus-stealing foreground process launches.
+ * PreToolUse hook — blocks Playwright headed/UI-mode invocations only.
+ *
+ * NARROWED 2026-08-26 (epic 688ad6da WP-4, lifeai-env). This file used to
+ * ALSO block every raw foreground process launch (Start-Process without
+ * -WindowStyle Hidden, cmd /c start without /min, window-focus Win32 APIs,
+ * bare PowerShell .ps1 launches) -- that entire class is retired per Dave's
+ * direct operator instruction: "remove the PreToolUse foreground-window
+ * guard entirely... replace it with caller-logging for traceability."
+ * lifeai-env's lib/TermLaunch.psm1 (Invoke-TermLaunchHidden /
+ * Invoke-TermLaunchInteractive) is the replacement -- it logs every caller
+ * (script + line + argv) to C:/Dev/.logs BEFORE spawning, which is strictly
+ * more informative than a hard block that told a caller only "add
+ * -WindowStyle Hidden" and never recorded who asked.
+ *
+ * checkPlaywright is a SEPARATE, unrelated concern (Design Decision D3,
+ * .rdc/plans/terminal-launch-consolidation.md in lifeai-env): agent
+ * sessions must never pop an interactive Playwright UI. That has nothing to
+ * do with terminal/process launch primitives, so it was deliberately kept
+ * here rather than folded into caller-logging, which would have silently
+ * dropped a real safety property this narrowing was never asked to remove.
+ * The filename is legacy -- kept to avoid an unrelated hookify-manifest
+ * rewire for a rename that changes nothing about what the file does.
  */
-'use strict';
-
 const hookLog = require('./hook-logger');
 
 function readStdin() {
@@ -35,35 +54,6 @@ function toolText(raw) {
   try { return JSON.stringify(raw.tool_input || raw); } catch { return ''; }
 }
 
-function hasHiddenIntent(command) {
-  return /-WindowStyle\s+Hidden/i.test(command) ||
-    /-WindowStyle\s+Minimized/i.test(command) ||
-    /windowsHide\s*:\s*true/i.test(command) ||
-    /CreateNoWindow\s*=\s*\$?true/i.test(command) ||
-    /Start-Job\b/i.test(command) ||
-    /--background\b/i.test(command) ||
-    /\bHEADLESS\s*=\s*(1|true)\b/i.test(command) ||
-    /\bCI\s*=\s*(1|true)\b/i.test(command);
-}
-
-function hasExplicitWindowOverride(command) {
-  return /\bRDC_ALLOW_WINDOW_FOCUS\s*=\s*(1|true)\b/i.test(command) ||
-    /\bRDC_INTERACTIVE_WINDOW\s*=\s*(1|true)\b/i.test(command);
-}
-
-function checkWindowFocusApi(command) {
-  if (hasExplicitWindowOverride(command)) return;
-  const focusApi = /\b(SetForegroundWindow|SwitchToThisWindow|AppActivate|SetWindowPos|ShowWindowAsync?|BringWindowToTop)\b/i;
-  const broadWindowApi = /\b(EnumWindows|Get-Process\s+\|\s*Where-Object|GetWindow|FindWindow)\b/i;
-  const windowMutation = /\b(minimi[sz]e|restore|foreground|focus|activate|collapse)\b/i;
-  if (focusApi.test(command) || (broadWindowApi.test(command) && windowMutation.test(command))) {
-    block(
-      'Window focus/restore/minimize/collapse operations are not allowed in agent-launched commands. Spawn helpers hidden/no-window instead; set RDC_ALLOW_WINDOW_FOCUS=1 only for an explicitly requested interactive recovery action.',
-      { kind: 'window-focus-api' },
-    );
-  }
-}
-
 function checkPlaywright(command) {
   if (!/\b(playwright|@playwright\/test)\b/i.test(command)) return;
 
@@ -82,45 +72,13 @@ function checkPlaywright(command) {
   }
 }
 
-function checkPowerShell(command) {
-  if (!/\bStart-Process\b/i.test(command)) return;
-  if (hasHiddenIntent(command)) return;
-  block(
-    '`Start-Process` must include `-WindowStyle Hidden` or `-WindowStyle Minimized` for agent-launched node/cmd/ps1/test processes. Focus/restore/collapse APIs remain blocked unless explicitly requested.',
-    { kind: 'start-process' },
-  );
-}
-
-function checkCmdStart(command) {
-  if (!/\bcmd(?:\.exe)?\s+\/c\s+start\b/i.test(command)) return;
-  if (/\bcmd(?:\.exe)?\s+\/c\s+start\s+(""|''|`"")?\s*\/b\b/i.test(command)) return;
-  block(
-    '`cmd /c start` must use `/min` or `/b` for background tools. Focus/restore/collapse APIs remain blocked unless explicitly requested.',
-    { kind: 'cmd-start' },
-  );
-}
-
-function checkDirectShellLaunch(command) {
-  if (hasHiddenIntent(command)) return;
-  if (/\bpowershell(?:\.exe)?\b[^|\n]*(?:-File\s+[^|\n]*\.ps1|\.ps1\b)/i.test(command)) {
-    block(
-      'PowerShell script launches from agent tooling must use `-WindowStyle Hidden -NonInteractive` or a hidden wrapper.',
-      { kind: 'powershell-ps1' },
-    );
-  }
-}
-
 async function main() {
   let raw;
   try { raw = JSON.parse(await readStdin()); } catch { process.exit(0); }
   const command = toolText(raw);
   if (!command) pass({ reason: 'no-command' });
 
-  checkWindowFocusApi(command);
   checkPlaywright(command);
-  checkPowerShell(command);
-  checkCmdStart(command);
-  checkDirectShellLaunch(command);
 
   pass({ reason: 'clean' });
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lifeaitools/rdc-skills",
-  "version": "0.35.3",
+  "version": "0.35.5",
   "description": "RDC typed-agent dispatch skill suite for Claude Code - plan, build, review, overnight builds",
   "keywords": [
     "claude-code",
@@ -44,7 +44,7 @@
     "validate": "node tests/validate-skills.js",
     "rdc-design": "node scripts/rdc-design-cli.mjs",
     "test:hooks": "node scripts/test-rdc-hooks.mjs",
-    "test:truth-gate": "node tests/run-evidence-gate.test.mjs && node tests/work-item-exit-gate-l2.test.mjs && node tests/work-item-exit-gate-l3.test.mjs && node tests/require-work-item-on-commit.test.mjs && node tests/harness-gates.test.mjs",
+    "test:truth-gate": "node tests/run-evidence-gate.test.mjs && node tests/work-item-exit-gate-l2.test.mjs && node tests/work-item-exit-gate-l3.test.mjs && node tests/require-work-item-on-commit.test.mjs && node tests/harness-gates.test.mjs && node tests/foreground-process-gate.test.mjs",
     "test:acceptance": "node tests/acceptance.test.mjs && node tests/install-rdc-skills.test.mjs && node tests/help-surface.test.mjs && node tests/manifest-contract-fields.test.mjs && node tests/plugin-namespace-names.test.mjs && node tests/skill-test-matrix.test.mjs && node tests/completion-gate-supervisor-admission.test.mjs && node tests/curl-surface.test.mjs && node tests/clauth-plugin-postinstall.test.mjs",
     "acceptance": "node scripts/acceptance.mjs --changed",
     "test:mcp": "node tests/mcp.test.mjs",

package/scripts/install-rdc-skills.js

@@ -508,8 +508,16 @@ function syncMarketplaceCheckout(mktDir) {
   const git = (...args) =>
     execSync(`git -C "${mktDir}" ${args.join(' ')}`, { encoding: 'utf8', stdio: 'pipe' }).trim();
   try {
-    if (git('status', '--porcelain')) {
-      warn(`marketplace clone has local changes at ${mktDir} — left untouched; it may serve stale commands`);
+    // PRESERVE-DIRTY means "don't discard a real local edit to a TRACKED file" —
+    // `--untracked-files=no` is deliberate. Plain `git status --porcelain` also
+    // reports untracked files (`?? some-cruft.txt`), which `git reset --hard`
+    // never touches and never conflicts with. Treating that as "dirty" made this
+    // sync silently refuse forever the moment ANY stray untracked file appeared
+    // in the clone — which is exactly how this checkout sat 39+ commits behind
+    // for over a week (2026-08-23), still warning "left untouched" on every run,
+    // while a plain `git status` on the same directory read completely clean.
+    if (git('status', '--porcelain', '--untracked-files=no')) {
+      warn(`marketplace clone has local changes to tracked files at ${mktDir} — left untouched; it may serve stale commands`);
       return;
     }
     const branch = git('rev-parse', '--abbrev-ref', 'HEAD') || 'master';
@@ -1347,7 +1355,7 @@ async function main() {
 // scripts/probe-installed-hooks.mjs verify the real install path instead of
 // re-implementing it — a probe that copies its own way proves nothing about what
 // ships. Without this guard, requiring the module would run a full install.
-module.exports = { copyHookFiles, assertHooksLoadable, registerCodexTarget, RDC_ENV_HOOK_TIMEOUT_SEC };
+module.exports = { copyHookFiles, assertHooksLoadable, registerCodexTarget, syncMarketplaceCheckout, RDC_ENV_HOOK_TIMEOUT_SEC };
 
 if (require.main === module) {
   main().catch(e => { fail(e.message); process.exit(1); });