@windyroad/risk-scorer 0.12.7-preview.583 → 0.12.7-preview.591

package/README.md

@@ -79,6 +79,16 @@ The plugin includes six specialised agents:
 | `/wr-risk-scorer:bootstrap-catalog` | Bootstrap `docs/risks/` register from existing `.risk-reports/` corpus per ADR-059 — walks reports, dedupes by ADR-056 slug, emits one `R<NNN>-<slug>.active.md` per unique slug. Idempotent. Auto-triggers from `/install-updates` Step 6.5.1 when register is empty + `RISK-POLICY.md` present + `.risk-reports/` non-empty |
 | `/wr-risk-scorer:update-policy` | Generate or update `RISK-POLICY.md` |
 
+### Internal-use wrapper skills
+
+These wrappers exist so consumer SKILLs can invoke the scoring agents via the Skill tool with `skill: wr-risk-scorer:<name>` per ADR-015's Confirmation literal phrasing. End users should invoke the `/wr-risk-scorer:assess-*` skills above; the wrappers are internal plumbing.
+
+| Wrapper skill | Purpose |
+|---------------|---------|
+| `wr-risk-scorer:pipeline` | Skill-tool wrapper around the `wr-risk-scorer:pipeline` agent (consumer: `/wr-risk-scorer:assess-release`) |
+| `wr-risk-scorer:wip` | Skill-tool wrapper around the `wr-risk-scorer:wip` agent (consumer: `/wr-risk-scorer:assess-wip`) |
+| `wr-risk-scorer:external-comms` | Skill-tool wrapper around the `wr-risk-scorer:external-comms` agent (consumer: `/wr-risk-scorer:assess-external-comms`) |
+
 ## External-comms gate
 
 The `external-comms-gate.sh` hook intercepts outbound prose tool calls and the

package/hooks/git-push-gate.sh

@@ -45,6 +45,16 @@ if echo "$COMMAND" | grep -qE '(^|;|&&|\|\|)\s*npm run push:watch(\s|$)'; then
         if [ -f "${RDIR}/clean" ]; then
             exit 0
         fi
+        # CI-status precondition (P208): a within-appetite predicted-risk
+        # score is necessary but not sufficient — the lagging CI signal
+        # must also be green (or no-history-yet for the documented
+        # first-push case). Fail-closed on gh errors. Ordered AFTER the
+        # one-shot bypass markers and BEFORE the predicted-risk gate so
+        # incident workflows and clean-tree pushes are unaffected.
+        if ! check_ci_status "$SESSION_ID" "push"; then
+            risk_gate_deny "Push blocked: ${CI_GATE_REASON}"
+            exit 0
+        fi
         if ! check_risk_gate "$SESSION_ID" "push"; then
             if [ "$RISK_GATE_CATEGORY" = "threshold" ]; then
                 risk_gate_deny "Push blocked: Push risk score ${RISK_GATE_SCORE}/25 (Medium or above). To proceed: (1) release first via \`npm run release:watch\`, (2) split the push, or (3) add risk-reducing measures. If risk-neutral or risk-reducing, delegate to wr-risk-scorer:pipeline (subagent_type: 'wr-risk-scorer:pipeline') — it will create a bypass marker."
@@ -83,6 +93,8 @@ if echo "$COMMAND" | grep -qE '(^|;|&&|\|\|)\s*npm run release:watch(\s|$)'; the
         # Live-incident bypass: if an incident marker exists, allow release
         # regardless of risk score. Used when addressing outages, security
         # incidents, or information disclosure that requires immediate deployment.
+        # Per JTBD-201, this MUST short-circuit BEFORE the CI-status check
+        # so the hotfix path is unaffected by red CI on master.
         if [ -f "${RDIR}/incident-release" ]; then
             rm -f "${RDIR}/incident-release"
             exit 0
@@ -92,6 +104,12 @@ if echo "$COMMAND" | grep -qE '(^|;|&&|\|\|)\s*npm run release:watch(\s|$)'; the
             rm -f "${RDIR}/reducing-release"
             exit 0
         fi
+        # CI-status precondition (P208): a green CI run on the target
+        # branch is required before shipping. Fail-closed on gh errors.
+        if ! check_ci_status "$SESSION_ID" "release"; then
+            risk_gate_deny "Release blocked: ${CI_GATE_REASON}"
+            exit 0
+        fi
         if ! check_risk_gate "$SESSION_ID" "release"; then
             risk_gate_deny "Release blocked: ${RISK_GATE_REASON}"
             exit 0

package/hooks/lib/risk-gate.sh

@@ -153,6 +153,148 @@ print(('yes' if score > N else 'no') + ' ' + str(N))
   return 0
 }
 
+# Check CI health for the current branch (P208).
+#
+# Returns 0 if push/release may proceed, 1 if denied. Sets CI_GATE_REASON
+# on deny with a human-readable message that names the CI conclusion and
+# the run URL. Sets CI_GATE_CATEGORY ∈ {bypass, no-history, allow, red,
+# pending, gh-error}.
+#
+# Consults `gh run list --branch <current-branch> --limit 1 --json
+# status,conclusion,databaseId,url` for the working branch's most recent
+# CI run.
+#
+# Decision table:
+#   - bypass marker present (${RDIR}/ci-bypass-${ACTION}) → allow, consume
+#   - gh failure (auth / timeout / API error) → DENY (fail-CLOSED, per
+#     P208 safe-high-fix-risk classifier — a buggy harden must NOT
+#     degrade to allow)
+#   - empty result `[]` → allow (no CI history yet; first push triggers
+#     CI naturally)
+#   - status ∈ {queued, in_progress, pending, requested, waiting} → deny
+#   - conclusion ∈ {failure, cancelled, timed_out, action_required,
+#     startup_failure} → deny
+#   - conclusion ∈ {success, skipped, neutral} or unknown → allow
+#
+# Usage: check_ci_status "$SESSION_ID" "push"   # or "release"
+check_ci_status() {
+  local SESSION_ID="$1"
+  local ACTION="$2"
+  local RDIR
+  RDIR=$(_risk_dir "$SESSION_ID")
+  local BYPASS_MARKER="${RDIR}/ci-bypass-${ACTION}"
+
+  CI_GATE_REASON=""
+  CI_GATE_CATEGORY=""
+
+  # One-shot bypass marker — consumed on use, same family as
+  # reducing-push / incident-release. Documented override for the
+  # legitimate "first push triggers CI" edge case and infra incidents.
+  if [ -f "$BYPASS_MARKER" ]; then
+    rm -f "$BYPASS_MARKER"
+    CI_GATE_CATEGORY="bypass"
+    return 0
+  fi
+
+  # Resolve current branch. If we're not in a git repo or HEAD is
+  # detached, skip the CI check (the surrounding push/release gate
+  # would itself fail at the git layer with a clearer error).
+  local BRANCH
+  BRANCH=$(git rev-parse --abbrev-ref HEAD 2>/dev/null || echo "")
+  if [ -z "$BRANCH" ] || [ "$BRANCH" = "HEAD" ]; then
+    CI_GATE_CATEGORY="allow"
+    return 0
+  fi
+
+  # Query GitHub. Bounded at 10s wall-clock so a network stall cannot
+  # hang push:watch indefinitely. `command -v timeout` because macOS
+  # default install does not ship GNU `timeout`.
+  local JSON GH_EXIT
+  if command -v timeout >/dev/null 2>&1; then
+    JSON=$(timeout 10s gh run list --branch "$BRANCH" --limit 1 \
+        --json status,conclusion,databaseId,url 2>/dev/null) || GH_EXIT=$?
+  else
+    JSON=$(gh run list --branch "$BRANCH" --limit 1 \
+        --json status,conclusion,databaseId,url 2>/dev/null) || GH_EXIT=$?
+  fi
+
+  if [ -n "${GH_EXIT:-}" ] && [ "$GH_EXIT" != "0" ]; then
+    CI_GATE_CATEGORY="gh-error"
+    CI_GATE_REASON="CI status check failed (gh exit ${GH_EXIT}: auth / timeout / API error). Fail-closed per P208 safe-high-fix-risk. Fix the underlying gh failure, or to override for a legitimate first-push-triggers-CI run, create the bypass marker: touch ${BYPASS_MARKER}"
+    return 1
+  fi
+
+  # Empty array = no CI history for this branch yet. Natural allow for
+  # the documented "first push triggers CI" case — no marker needed.
+  local TRIMMED
+  TRIMMED=$(printf '%s' "$JSON" | tr -d '[:space:]')
+  if [ -z "$TRIMMED" ] || [ "$TRIMMED" = "[]" ]; then
+    CI_GATE_CATEGORY="no-history"
+    return 0
+  fi
+
+  # Parse status, conclusion, url. Fail-closed on parse error.
+  local PARSED
+  PARSED=$(echo "$JSON" | python3 -c "
+import sys, json
+try:
+    runs = json.load(sys.stdin)
+    if not isinstance(runs, list) or not runs:
+        print('||')
+        sys.exit(0)
+    r = runs[0]
+    print('{}|{}|{}'.format(r.get('status') or '', r.get('conclusion') or '', r.get('url') or ''))
+except Exception:
+    print('PARSE_ERROR||')
+" 2>/dev/null || echo "PARSE_ERROR||")
+
+  local STATUS CONCLUSION URL
+  STATUS="${PARSED%%|*}"
+  local REST="${PARSED#*|}"
+  CONCLUSION="${REST%%|*}"
+  URL="${REST#*|}"
+
+  if [ "$STATUS" = "PARSE_ERROR" ]; then
+    CI_GATE_CATEGORY="gh-error"
+    CI_GATE_REASON="CI status check returned unparseable response. Fail-closed per P208 safe-high-fix-risk. To override for a legitimate first-push case, create the bypass marker: touch ${BYPASS_MARKER}"
+    return 1
+  fi
+
+  case "$STATUS" in
+    queued|in_progress|pending|requested|waiting)
+      CI_GATE_CATEGORY="pending"
+      CI_GATE_REASON="Latest CI run on branch '${BRANCH}' is still in flight (status: ${STATUS}). Wait for it to settle: ${URL}. To override, create the bypass marker: touch ${BYPASS_MARKER}"
+      return 1
+      ;;
+    completed)
+      case "$CONCLUSION" in
+        success|skipped|neutral|"")
+          CI_GATE_CATEGORY="allow"
+          return 0
+          ;;
+        failure|cancelled|timed_out|action_required|startup_failure)
+          CI_GATE_CATEGORY="red"
+          CI_GATE_REASON="Latest CI run on branch '${BRANCH}' concluded ${CONCLUSION}: ${URL}. Fix CI before pushing/releasing. To override for a legitimate first-push or infra-incident case, create the bypass marker: touch ${BYPASS_MARKER}"
+          return 1
+          ;;
+        *)
+          # Unknown conclusion — allow rather than block on a value we
+          # don't recognise. New GitHub conclusion values are infrequent.
+          CI_GATE_CATEGORY="allow"
+          return 0
+          ;;
+      esac
+      ;;
+    *)
+      # Unknown status — allow rather than block on a value we don't
+      # recognise. Conservative tilts toward the threshold check below
+      # catching the actual risk.
+      CI_GATE_CATEGORY="allow"
+      return 0
+      ;;
+  esac
+}
+
 # Emit fail-closed deny JSON for PreToolUse hooks.
 risk_gate_deny() {
   local REASON="$1"

package/hooks/test/ci-status-gate.bats

@@ -0,0 +1,240 @@
+#!/usr/bin/env bats
+# Tests for the CI-status precondition in the push/release gate.
+#
+# Closes P208 (known-error): git-push-gate.sh did not consult CI health
+# before scoring push/release risk, so a push could land on a CI-red
+# master and a release could ship broken code.
+#
+# Contract:
+# - `check_ci_status` queries `gh run list --branch <current-branch>
+#   --limit 1 --json status,conclusion,databaseId,url` for the current
+#   branch and returns 0 (allow) / 1 (deny).
+# - Deny on conclusion ∈ {failure, cancelled, timed_out, action_required,
+#   startup_failure}.
+# - Deny on status ∈ {queued, in_progress, pending, requested, waiting}.
+# - Allow on conclusion ∈ {success, skipped, neutral} or unknown.
+# - Empty array (no CI history yet) → allow. Handles the documented
+#   "first push triggers CI" case naturally — no bypass marker required.
+# - `gh` failure (auth/timeout/API error) → DENY (fail-closed per the
+#   safe-high-fix-risk classifier on P208).
+# - `${RDIR}/ci-bypass-${ACTION}` one-shot bypass marker — consumed on
+#   use, same family as reducing-push / incident-release.
+# - Integration: in git-push-gate.sh, the ordering is bypass-markers →
+#   CI status → risk gate. The `incident-release` bypass MUST short-
+#   circuit BEFORE the CI check fires (per JTBD-201 + ADR-018).
+
+setup() {
+  HOOKS_DIR="$(cd "$(dirname "$BATS_TEST_FILENAME")/.." && pwd)"
+  source "$HOOKS_DIR/lib/gate-helpers.sh"
+  source "$HOOKS_DIR/lib/risk-gate.sh"
+
+  TEST_SESSION="bats-ci-gate-$$-${BATS_TEST_NUMBER}"
+  RDIR=$(_risk_dir "$TEST_SESSION")
+  rm -rf "$RDIR"
+  mkdir -p "$RDIR"
+
+  # Stand up a fake git repo so `git rev-parse --abbrev-ref HEAD` resolves.
+  TEST_REPO="$(mktemp -d)"
+  ( cd "$TEST_REPO" && git init -q -b main && \
+      git -c user.email=t@e -c user.name=t commit --allow-empty -q -m "init" )
+
+  # Stub `gh` on PATH. The stub reads $FAKE_GH_OUTPUT and $FAKE_GH_EXIT
+  # for behaviour. PATH ordering: stub dir first.
+  STUB_DIR="$(mktemp -d)"
+  cat > "$STUB_DIR/gh" <<'STUB'
+#!/bin/bash
+if [ -n "${FAKE_GH_DELAY:-}" ]; then sleep "$FAKE_GH_DELAY"; fi
+if [ -n "${FAKE_GH_OUTPUT:-}" ]; then
+  printf '%s' "$FAKE_GH_OUTPUT"
+fi
+exit "${FAKE_GH_EXIT:-0}"
+STUB
+  chmod +x "$STUB_DIR/gh"
+  # `timeout` may not exist on the path on some macOS setups — stub a
+  # passthrough for portability. Tests inject FAKE_GH_DELAY only when
+  # they specifically test timeout behaviour.
+  ORIG_PATH="$PATH"
+  export PATH="$STUB_DIR:$PATH"
+  export TEST_REPO STUB_DIR
+}
+
+teardown() {
+  rm -rf "$RDIR" "$TEST_REPO" "$STUB_DIR"
+  export PATH="$ORIG_PATH"
+  unset FAKE_GH_OUTPUT FAKE_GH_EXIT FAKE_GH_DELAY CI_GATE_REASON CI_GATE_CATEGORY 2>/dev/null || true
+}
+
+# Run check_ci_status inside the fake repo so branch resolution works.
+_run_check() {
+  local action="$1"
+  CI_GATE_REASON=""
+  CI_GATE_CATEGORY=""
+  ( cd "$TEST_REPO" && \
+      FAKE_GH_OUTPUT="${FAKE_GH_OUTPUT:-}" FAKE_GH_EXIT="${FAKE_GH_EXIT:-0}" \
+      bash -c "source '$HOOKS_DIR/lib/gate-helpers.sh'; source '$HOOKS_DIR/lib/risk-gate.sh'; \
+               if check_ci_status '$TEST_SESSION' '$action'; then echo ALLOW; \
+               else echo \"DENY: \$CI_GATE_REASON\"; fi" )
+}
+
+@test "check_ci_status allows when latest CI run concluded success" {
+  export FAKE_GH_OUTPUT='[{"status":"completed","conclusion":"success","databaseId":1,"url":"https://github.com/x/y/actions/runs/1"}]'
+  result=$(_run_check "push")
+  [[ "$result" == "ALLOW" ]]
+}
+
+@test "check_ci_status denies when latest CI run concluded failure (names conclusion + URL)" {
+  export FAKE_GH_OUTPUT='[{"status":"completed","conclusion":"failure","databaseId":2,"url":"https://github.com/x/y/actions/runs/2"}]'
+  result=$(_run_check "push")
+  [[ "$result" == DENY:* ]]
+  [[ "$result" == *"failure"* ]]
+  [[ "$result" == *"https://github.com/x/y/actions/runs/2"* ]]
+}
+
+@test "check_ci_status denies when latest CI run concluded cancelled" {
+  export FAKE_GH_OUTPUT='[{"status":"completed","conclusion":"cancelled","databaseId":3,"url":"https://github.com/x/y/actions/runs/3"}]'
+  result=$(_run_check "release")
+  [[ "$result" == DENY:* ]]
+  [[ "$result" == *"cancelled"* ]]
+}
+
+@test "check_ci_status denies when latest CI run concluded timed_out" {
+  export FAKE_GH_OUTPUT='[{"status":"completed","conclusion":"timed_out","databaseId":4,"url":"https://github.com/x/y/actions/runs/4"}]'
+  result=$(_run_check "push")
+  [[ "$result" == DENY:* ]]
+  [[ "$result" == *"timed_out"* ]]
+}
+
+@test "check_ci_status denies when latest CI run status is in_progress" {
+  export FAKE_GH_OUTPUT='[{"status":"in_progress","conclusion":null,"databaseId":5,"url":"https://github.com/x/y/actions/runs/5"}]'
+  result=$(_run_check "push")
+  [[ "$result" == DENY:* ]]
+  [[ "$result" == *"in_progress"* ]]
+}
+
+@test "check_ci_status denies when latest CI run status is queued" {
+  export FAKE_GH_OUTPUT='[{"status":"queued","conclusion":null,"databaseId":6,"url":"https://github.com/x/y/actions/runs/6"}]'
+  result=$(_run_check "release")
+  [[ "$result" == DENY:* ]]
+  [[ "$result" == *"queued"* ]]
+}
+
+@test "check_ci_status allows when CI history is empty (first push triggers CI)" {
+  export FAKE_GH_OUTPUT='[]'
+  result=$(_run_check "push")
+  [[ "$result" == "ALLOW" ]]
+}
+
+@test "check_ci_status denies when gh exits non-zero (fail-closed, safe-high-fix-risk)" {
+  export FAKE_GH_OUTPUT=''
+  export FAKE_GH_EXIT=1
+  result=$(_run_check "push")
+  [[ "$result" == DENY:* ]]
+  # Must point at the ci-bypass marker for the documented override path.
+  [[ "$result" == *"ci-bypass-push"* ]]
+}
+
+@test "check_ci_status allows when ci-bypass marker is present and consumes it" {
+  : > "$RDIR/ci-bypass-push"
+  export FAKE_GH_OUTPUT='[{"status":"completed","conclusion":"failure","databaseId":7,"url":"https://github.com/x/y/actions/runs/7"}]'
+  result=$(_run_check "push")
+  [[ "$result" == "ALLOW" ]]
+  # Bypass markers are one-shot — same family as reducing-push / incident-release.
+  [ ! -f "$RDIR/ci-bypass-push" ]
+}
+
+@test "check_ci_status bypass marker is action-scoped (push marker does not bypass release)" {
+  : > "$RDIR/ci-bypass-push"
+  export FAKE_GH_OUTPUT='[{"status":"completed","conclusion":"failure","databaseId":8,"url":"https://github.com/x/y/actions/runs/8"}]'
+  result=$(_run_check "release")
+  [[ "$result" == DENY:* ]]
+  # push bypass must not have been consumed by a release check
+  [ -f "$RDIR/ci-bypass-push" ]
+}
+
+@test "check_ci_status allows when conclusion is skipped" {
+  export FAKE_GH_OUTPUT='[{"status":"completed","conclusion":"skipped","databaseId":9,"url":"https://github.com/x/y/actions/runs/9"}]'
+  result=$(_run_check "push")
+  [[ "$result" == "ALLOW" ]]
+}
+
+@test "check_ci_status allows when conclusion is neutral" {
+  export FAKE_GH_OUTPUT='[{"status":"completed","conclusion":"neutral","databaseId":10,"url":"https://github.com/x/y/actions/runs/10"}]'
+  result=$(_run_check "release")
+  [[ "$result" == "ALLOW" ]]
+}
+
+# ---------------------------------------------------------------------------
+# Integration: git-push-gate.sh ordering — bypass-markers → CI status → risk
+# gate. JTBD-201 demands the incident-release bypass MUST short-circuit
+# BEFORE the new CI-status check fires.
+# ---------------------------------------------------------------------------
+
+# Helper: build a PreToolUse Bash input with a given command
+_build_input() {
+  local cmd="$1"
+  cat <<JSON
+{
+  "session_id": "$TEST_SESSION",
+  "tool_name": "Bash",
+  "tool_input": {
+    "command": "$cmd"
+  }
+}
+JSON
+}
+
+@test "git-push-gate.sh denies push:watch when CI is red even if risk score is within appetite" {
+  # Within-appetite risk score
+  echo "1" > "$RDIR/push"
+  # Disable drift check (no stored hash file)
+  export FAKE_GH_OUTPUT='[{"status":"completed","conclusion":"failure","databaseId":11,"url":"https://github.com/x/y/actions/runs/11"}]'
+
+  INPUT=$(_build_input "npm run push:watch")
+  output=$( cd "$TEST_REPO" && echo "$INPUT" | \
+    FAKE_GH_OUTPUT="$FAKE_GH_OUTPUT" PATH="$STUB_DIR:$PATH" \
+    "$HOOKS_DIR/git-push-gate.sh" )
+  [[ "$output" == *"permissionDecision"* ]]
+  [[ "$output" == *"deny"* ]]
+  [[ "$output" == *"failure"* ]]
+}
+
+@test "git-push-gate.sh denies release:watch when CI is red even if risk score is within appetite" {
+  echo "1" > "$RDIR/release"
+  export FAKE_GH_OUTPUT='[{"status":"completed","conclusion":"failure","databaseId":12,"url":"https://github.com/x/y/actions/runs/12"}]'
+
+  INPUT=$(_build_input "npm run release:watch")
+  output=$( cd "$TEST_REPO" && echo "$INPUT" | \
+    FAKE_GH_OUTPUT="$FAKE_GH_OUTPUT" PATH="$STUB_DIR:$PATH" \
+    "$HOOKS_DIR/git-push-gate.sh" )
+  [[ "$output" == *"permissionDecision"* ]]
+  [[ "$output" == *"deny"* ]]
+  [[ "$output" == *"failure"* ]]
+}
+
+@test "git-push-gate.sh allows release:watch when incident-release bypass is set, even if CI is red (JTBD-201)" {
+  echo "9" > "$RDIR/release"
+  : > "$RDIR/incident-release"
+  # Even with a red CI conclusion, the incident bypass must short-circuit
+  # both the CI check and the risk threshold.
+  export FAKE_GH_OUTPUT='[{"status":"completed","conclusion":"failure","databaseId":13,"url":"https://github.com/x/y/actions/runs/13"}]'
+
+  INPUT=$(_build_input "npm run release:watch")
+  output=$( cd "$TEST_REPO" && echo "$INPUT" | \
+    FAKE_GH_OUTPUT="$FAKE_GH_OUTPUT" PATH="$STUB_DIR:$PATH" \
+    "$HOOKS_DIR/git-push-gate.sh" )
+  # No permissionDecision means allow (exit 0 with no JSON).
+  [[ "$output" != *"permissionDecision"* ]]
+  # incident-release marker is one-shot — must be consumed
+  [ ! -f "$RDIR/incident-release" ]
+}
+
+@test "git-push-gate.sh allows push:watch when CI history is empty (first push)" {
+  echo "1" > "$RDIR/push"
+  export FAKE_GH_OUTPUT='[]'
+
+  INPUT=$(_build_input "npm run push:watch")
+  output=$( cd "$TEST_REPO" && echo "$INPUT" | \
+    FAKE_GH_OUTPUT="$FAKE_GH_OUTPUT" PATH="$STUB_DIR:$PATH" \
+    "$HOOKS_DIR/git-push-gate.sh" )
+  [[ "$output" != *"permissionDecision"* ]]
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@windyroad/risk-scorer",
-  "version": "0.12.7-preview.583",
+  "version": "0.12.7-preview.591",
   "description": "Pipeline risk scoring, commit/push gates, and secret leak detection",
   "bin": {
     "windyroad-risk-scorer": "./bin/install.mjs"

package/skills/assess-external-comms/SKILL.md

@@ -72,14 +72,14 @@ The orchestrator does NOT pre-compute the key — the hook derives it from the p
 
 ### 4. Delegate to wr-risk-scorer:external-comms
 
-Invoke the subagent via the `Skill` tool:
+Invoke the external-comms reviewer via the `Skill` tool. The `wr-risk-scorer:external-comms` SKILL is a thin wrapper around the external-comms agent (per ADR-015 — see `packages/risk-scorer/skills/external-comms/SKILL.md`):
 
 ```
-subagent_type: wr-risk-scorer:external-comms
+skill: wr-risk-scorer:external-comms
 prompt: <constructed review prompt from step 3>
 ```
 
-Wait for the subagent to complete. The subagent outputs a structured verdict block (`EXTERNAL_COMMS_RISK_VERDICT: PASS|FAIL` + optional `EXTERNAL_COMMS_RISK_REASON: ...` on FAIL). The `PostToolUse:Agent` hook (`risk-score-mark.sh`) parses the verdict, derives the marker key from the prompt's `SURFACE:` + `<draft>` structure, and writes the marker automatically on PASS.
+Wait for the wrapper to return. The wrapper invokes the external-comms agent internally; the agent's structured verdict block (`EXTERNAL_COMMS_RISK_VERDICT: PASS|FAIL` + optional `EXTERNAL_COMMS_RISK_REASON: ...` on FAIL) flows back verbatim. The `PostToolUse:Agent` hook (`risk-score-mark.sh`) fires on the wrapper's inner Agent invocation, derives the marker key from the prompt's `SURFACE:` + `<draft>` structure, and writes the marker automatically on PASS.
 
 **Do not write to `${TMPDIR:-/tmp}/claude-risk-*` yourself.** The hook is the only correct mechanism.
 

package/skills/assess-release/SKILL.md

@@ -64,14 +64,14 @@ Build a self-contained prompt for the pipeline subagent that includes:
 
 ### 5. Delegate to wr-risk-scorer:pipeline
 
-Invoke the pipeline subagent via the `Skill` tool:
+Invoke the pipeline scorer via the `Skill` tool. The `wr-risk-scorer:pipeline` SKILL is a thin wrapper around the pipeline agent (per ADR-015 — see `packages/risk-scorer/skills/pipeline/SKILL.md`):
 
 ```
-subagent_type: wr-risk-scorer:pipeline
+skill: wr-risk-scorer:pipeline
 prompt: <constructed assessment prompt from step 4>
 ```
 
-Wait for the subagent to complete. The subagent will output a structured `RISK_SCORES:` block. The `PostToolUse:Agent` hook (`risk-score-mark.sh`) reads that output and writes the bypass marker files automatically.
+Wait for the wrapper to return. The wrapper invokes the pipeline agent internally; the agent's structured `RISK_SCORES:` block flows back through the wrapper verbatim. The `PostToolUse:Agent` hook (`risk-score-mark.sh`) fires on the wrapper's inner Agent invocation and writes the bypass marker files automatically.
 
 **Do not write to `$TMPDIR/claude-risk-*` yourself.** The hook is the only correct mechanism.
 

package/skills/assess-release/test/assess-skills-delegate-via-skill-tool.bats

@@ -0,0 +1,162 @@
+#!/usr/bin/env bats
+# Contract guard: the on-demand assessment SKILLs (assess-release, assess-wip,
+# assess-external-comms) MUST delegate to their scoring agent via the Skill
+# tool — not via the Agent tool — matching ADR-015's Confirmation literal
+# phrasing ("the skill delegates to wr-risk-scorer:<agent> via the Skill
+# tool"). Closes the P205 contradiction surfaced by ADR-015 Confirmation
+# vs. SKILL.md prose mismatch.
+#
+# Structural assertions — Permitted Exception to the source-grep ban
+# (ADR-005 / P011), same framing as risk-scorer-register-hint.bats. SKILL.md
+# prose IS the contract document the orchestrator (Claude) consumes when
+# executing the SKILL; an LLM-output behavioural check is out of scope for
+# bats and is the responsibility of the promptfoo harness (ADR-075).
+#
+# What is asserted (contract, not implementation):
+#   1. Each assess-* SKILL's step 5 (release) / step 3 (wip) / step 4
+#      (external-comms) names `skill:` as the delegation tool parameter
+#      with the correct wrapper SKILL name as the target.
+#   2. None of the assess-* SKILLs name `subagent_type:` as the delegation
+#      tool parameter (the P205 contradiction class).
+#   3. Each wrapper SKILL (`pipeline`, `wip`, `external-comms`) exists at
+#      its expected path, is namespaced `wr-risk-scorer:<name>`, and
+#      delegates to its sibling agent via `subagent_type:`.
+#
+# Cross-reference:
+#   P205:    docs/problems/known-error/205-wr-risk-scorer-assess-release-skill-md-step-5-prose-says-skill-tool-but-provides-subagent-type.md
+#   ADR-015: docs/decisions/015-on-demand-assessment-skills.proposed.md (Confirmation criteria 189-193)
+#   ADR-052: docs/decisions/052-behavioural-tests-default.proposed.md (Permitted Exception)
+#   @jtbd JTBD-005 (invoke governance assessments on demand)
+#   @jtbd JTBD-101 (extend the suite — plugins expose corresponding skills)
+
+setup() {
+  SKILLS_DIR="$(cd "$(dirname "$BATS_TEST_FILENAME")/../.." && pwd)"
+
+  ASSESS_RELEASE="${SKILLS_DIR}/assess-release/SKILL.md"
+  ASSESS_WIP="${SKILLS_DIR}/assess-wip/SKILL.md"
+  ASSESS_EXTERNAL_COMMS="${SKILLS_DIR}/assess-external-comms/SKILL.md"
+
+  WRAPPER_PIPELINE="${SKILLS_DIR}/pipeline/SKILL.md"
+  WRAPPER_WIP="${SKILLS_DIR}/wip/SKILL.md"
+  WRAPPER_EXTERNAL_COMMS="${SKILLS_DIR}/external-comms/SKILL.md"
+}
+
+# ──────────────────────────────────────────────────────────────────────────────
+# Consumer SKILLs delegate via Skill tool (skill: parameter)
+# ──────────────────────────────────────────────────────────────────────────────
+
+@test "assess-release delegates via skill: wr-risk-scorer:pipeline" {
+  [ -f "$ASSESS_RELEASE" ]
+  run grep -E "^skill: wr-risk-scorer:pipeline$" "$ASSESS_RELEASE"
+  [ "$status" -eq 0 ]
+}
+
+@test "assess-release does NOT use subagent_type: in its delegation block" {
+  [ -f "$ASSESS_RELEASE" ]
+  # The P205 contradiction was: prose says "Skill tool" but provides
+  # subagent_type: wr-risk-scorer:pipeline. After the fix, no
+  # `subagent_type:` line may appear in the delegation block.
+  run grep -E "^subagent_type: wr-risk-scorer:pipeline$" "$ASSESS_RELEASE"
+  [ "$status" -ne 0 ]
+}
+
+@test "assess-wip delegates via skill: wr-risk-scorer:wip" {
+  [ -f "$ASSESS_WIP" ]
+  run grep -E "^skill: wr-risk-scorer:wip$" "$ASSESS_WIP"
+  [ "$status" -eq 0 ]
+}
+
+@test "assess-wip does NOT use subagent_type: in its delegation block" {
+  [ -f "$ASSESS_WIP" ]
+  run grep -E "^subagent_type: wr-risk-scorer:wip$" "$ASSESS_WIP"
+  [ "$status" -ne 0 ]
+}
+
+@test "assess-external-comms delegates via skill: wr-risk-scorer:external-comms" {
+  [ -f "$ASSESS_EXTERNAL_COMMS" ]
+  run grep -E "^skill: wr-risk-scorer:external-comms$" "$ASSESS_EXTERNAL_COMMS"
+  [ "$status" -eq 0 ]
+}
+
+@test "assess-external-comms does NOT use subagent_type: in its delegation block" {
+  [ -f "$ASSESS_EXTERNAL_COMMS" ]
+  run grep -E "^subagent_type: wr-risk-scorer:external-comms$" "$ASSESS_EXTERNAL_COMMS"
+  [ "$status" -ne 0 ]
+}
+
+# ──────────────────────────────────────────────────────────────────────────────
+# Wrapper SKILLs exist with correct names and delegate to the agent
+# ──────────────────────────────────────────────────────────────────────────────
+
+@test "wrapper SKILL packages/risk-scorer/skills/pipeline/SKILL.md exists" {
+  [ -f "$WRAPPER_PIPELINE" ]
+}
+
+@test "wrapper SKILL pipeline declares name: wr-risk-scorer:pipeline" {
+  [ -f "$WRAPPER_PIPELINE" ]
+  run grep -E "^name: wr-risk-scorer:pipeline$" "$WRAPPER_PIPELINE"
+  [ "$status" -eq 0 ]
+}
+
+@test "wrapper SKILL pipeline delegates to the pipeline agent via subagent_type:" {
+  [ -f "$WRAPPER_PIPELINE" ]
+  run grep -E "^subagent_type: wr-risk-scorer:pipeline$" "$WRAPPER_PIPELINE"
+  [ "$status" -eq 0 ]
+}
+
+@test "wrapper SKILL packages/risk-scorer/skills/wip/SKILL.md exists" {
+  [ -f "$WRAPPER_WIP" ]
+}
+
+@test "wrapper SKILL wip declares name: wr-risk-scorer:wip" {
+  [ -f "$WRAPPER_WIP" ]
+  run grep -E "^name: wr-risk-scorer:wip$" "$WRAPPER_WIP"
+  [ "$status" -eq 0 ]
+}
+
+@test "wrapper SKILL wip delegates to the wip agent via subagent_type:" {
+  [ -f "$WRAPPER_WIP" ]
+  run grep -E "^subagent_type: wr-risk-scorer:wip$" "$WRAPPER_WIP"
+  [ "$status" -eq 0 ]
+}
+
+@test "wrapper SKILL packages/risk-scorer/skills/external-comms/SKILL.md exists" {
+  [ -f "$WRAPPER_EXTERNAL_COMMS" ]
+}
+
+@test "wrapper SKILL external-comms declares name: wr-risk-scorer:external-comms" {
+  [ -f "$WRAPPER_EXTERNAL_COMMS" ]
+  run grep -E "^name: wr-risk-scorer:external-comms$" "$WRAPPER_EXTERNAL_COMMS"
+  [ "$status" -eq 0 ]
+}
+
+@test "wrapper SKILL external-comms delegates to the external-comms agent via subagent_type:" {
+  [ -f "$WRAPPER_EXTERNAL_COMMS" ]
+  run grep -E "^subagent_type: wr-risk-scorer:external-comms$" "$WRAPPER_EXTERNAL_COMMS"
+  [ "$status" -eq 0 ]
+}
+
+# ──────────────────────────────────────────────────────────────────────────────
+# Wrapper SKILLs disambiguate from end-user assess-* surfaces
+# ──────────────────────────────────────────────────────────────────────────────
+
+@test "wrapper pipeline description names assess-release as the end-user surface" {
+  [ -f "$WRAPPER_PIPELINE" ]
+  # JTBD-005 persona-fit: solo developer must not land on the raw wrapper
+  # and miss the assess-* gate-satisfaction wrap-up. Description must
+  # disambiguate.
+  run grep -E "assess-release" "$WRAPPER_PIPELINE"
+  [ "$status" -eq 0 ]
+}
+
+@test "wrapper wip description names assess-wip as the end-user surface" {
+  [ -f "$WRAPPER_WIP" ]
+  run grep -E "assess-wip" "$WRAPPER_WIP"
+  [ "$status" -eq 0 ]
+}
+
+@test "wrapper external-comms description names assess-external-comms as the end-user surface" {
+  [ -f "$WRAPPER_EXTERNAL_COMMS" ]
+  run grep -E "assess-external-comms" "$WRAPPER_EXTERNAL_COMMS"
+  [ "$status" -eq 0 ]
+}

package/skills/assess-wip/SKILL.md

@@ -42,14 +42,14 @@ Build a self-contained prompt for the wip subagent that includes:
 
 ### 3. Delegate to wr-risk-scorer:wip
 
-Invoke the wip subagent via the `Skill` tool:
+Invoke the WIP scorer via the `Skill` tool. The `wr-risk-scorer:wip` SKILL is a thin wrapper around the wip agent (per ADR-015 — see `packages/risk-scorer/skills/wip/SKILL.md`):
 
 ```
-subagent_type: wr-risk-scorer:wip
+skill: wr-risk-scorer:wip
 prompt: <constructed assessment prompt from step 2>
 ```
 
-Wait for the subagent to complete.
+Wait for the wrapper to return. The wrapper invokes the wip agent internally and returns the agent's verdict verbatim.
 
 ### 4. Present results
 

package/skills/external-comms/SKILL.md

@@ -0,0 +1,37 @@
+---
+name: wr-risk-scorer:external-comms
+description: Invokable SKILL wrapper around the wr-risk-scorer:external-comms leak-review agent. Delegates to the agent via the Agent tool and returns the agent's structured EXTERNAL_COMMS_RISK_VERDICT. Internal-use plumbing used by `/wr-risk-scorer:assess-external-comms` per ADR-015's Confirmation literal phrasing. End users should invoke `/wr-risk-scorer:assess-external-comms` instead.
+allowed-tools: Read, Glob, Grep, Bash, Agent
+---
+
+# External-Comms Leak Review Skill (Wrapper)
+
+This SKILL is an **invokable wrapper** around the `wr-risk-scorer:external-comms` agent. It exists so consumer SKILLs can invoke the leak reviewer via the **Skill tool** with `skill: wr-risk-scorer:external-comms` — matching ADR-015's Confirmation literal phrasing.
+
+**End users**: invoke `/wr-risk-scorer:assess-external-comms` instead. This wrapper is internal-use plumbing — calling it directly returns the raw verdict without the structured AskUserQuestion above-appetite handling (Rewrite / Move to private channel / Override / Cancel) that `/wr-risk-scorer:assess-external-comms` provides.
+
+## Contract
+
+- **Input** (`$ARGUMENTS`): a self-contained leak-review prompt structured per `packages/risk-scorer/agents/external-comms.md` § "What you receive":
+  - A leading `SURFACE: <name>` line (one of the canonical surface strings).
+  - The draft body wrapped verbatim inside `<draft>...</draft>` markers (the PostToolUse hook derives the marker key from this).
+  - The destination when known.
+- **Output**: the agent's verbatim verdict — `EXTERNAL_COMMS_RISK_VERDICT: PASS | FAIL` plus, on FAIL, an `EXTERNAL_COMMS_RISK_REASON:` block naming each Confidential Information class and the substrings that triggered it.
+- **Side effects**: the `PostToolUse:Agent` hook (`risk-score-mark.sh`) parses the verdict and writes the `external-comms-gate.sh` marker on PASS. The wrapper itself writes no files.
+
+## Steps
+
+### 1. Pass-through to the external-comms agent
+
+Invoke the external-comms subagent via the Agent tool with the caller's `$ARGUMENTS` verbatim. The `SURFACE:` line and `<draft>...</draft>` markers MUST be preserved exactly — the PostToolUse hook depends on the prompt structure for marker-key derivation:
+
+```
+subagent_type: wr-risk-scorer:external-comms
+prompt: $ARGUMENTS
+```
+
+### 2. Return the agent report verbatim
+
+Return the agent's response to the caller without alteration. Do NOT strip, paraphrase, or post-process the `EXTERNAL_COMMS_RISK_VERDICT:` or `EXTERNAL_COMMS_RISK_REASON:` blocks — the hook parses the verdict and consumer SKILLs surface the reason directly.
+
+$ARGUMENTS

package/skills/pipeline/SKILL.md

@@ -0,0 +1,34 @@
+---
+name: wr-risk-scorer:pipeline
+description: Invokable SKILL wrapper around the wr-risk-scorer:pipeline scoring agent. Delegates to the agent via the Agent tool and returns the agent's structured RISK_SCORES output. Internal-use plumbing used by `/wr-risk-scorer:assess-release` and any other consumer SKILL that needs Skill-tool-shaped invocation of the pipeline scorer per ADR-015's Confirmation literal phrasing. End users should invoke `/wr-risk-scorer:assess-release` instead.
+allowed-tools: Read, Glob, Bash, Agent
+---
+
+# Pipeline Scoring Skill (Wrapper)
+
+This SKILL is an **invokable wrapper** around the `wr-risk-scorer:pipeline` agent. It exists so consumer SKILLs can invoke the pipeline scorer via the **Skill tool** with `skill: wr-risk-scorer:pipeline` — matching ADR-015's Confirmation literal phrasing.
+
+**End users**: invoke `/wr-risk-scorer:assess-release` instead. This wrapper is internal-use plumbing — calling it directly returns raw scoring output without the gate-satisfaction wrap-up, AskUserQuestion above-appetite handling, or release-context resolution that `/wr-risk-scorer:assess-release` provides.
+
+## Contract
+
+- **Input** (`$ARGUMENTS`): a self-contained scoring prompt with pipeline state context. Caller assembles UNCOMMITTED / UNPUSHED / UNRELEASED sections per `packages/risk-scorer/agents/pipeline.md` § Pipeline State.
+- **Output**: the agent's verbatim report, including the structured `RISK_SCORES: commit=N push=N release=N` block, optional `RISK_BYPASS:` line, optional `RISK_REMEDIATIONS:` block, optional `RISK_REGISTER_HINT:` block, and optional `CATALOG_HIT_RATE:` line.
+- **Side effects**: the `PostToolUse:Agent` hook (`risk-score-mark.sh`) reads the agent's output downstream of this wrapper and writes the bypass marker files to `${TMPDIR}/claude-risk-${SESSION_ID}/`. The wrapper itself writes no files.
+
+## Steps
+
+### 1. Pass-through to the pipeline agent
+
+Invoke the pipeline subagent via the Agent tool with the caller's `$ARGUMENTS` verbatim:
+
+```
+subagent_type: wr-risk-scorer:pipeline
+prompt: $ARGUMENTS
+```
+
+### 2. Return the agent report verbatim
+
+Return the agent's response to the caller without alteration. Do NOT strip, paraphrase, or post-process the structured output blocks (`RISK_SCORES:`, `RISK_BYPASS:`, `RISK_REMEDIATIONS:`, `RISK_REGISTER_HINT:`, `CATALOG_HIT_RATE:`). The PostToolUse hook depends on the exact byte sequence to parse.
+
+$ARGUMENTS

package/skills/wip/SKILL.md

@@ -0,0 +1,33 @@
+---
+name: wr-risk-scorer:wip
+description: Invokable SKILL wrapper around the wr-risk-scorer:wip nudge agent. Delegates to the agent via the Agent tool and returns the agent's structured WIP risk verdict. Internal-use plumbing used by `/wr-risk-scorer:assess-wip` per ADR-015's Confirmation literal phrasing. End users should invoke `/wr-risk-scorer:assess-wip` instead.
+allowed-tools: Read, Glob, Bash, Agent
+---
+
+# WIP Scoring Skill (Wrapper)
+
+This SKILL is an **invokable wrapper** around the `wr-risk-scorer:wip` agent. It exists so consumer SKILLs can invoke the WIP nudge scorer via the **Skill tool** with `skill: wr-risk-scorer:wip` — matching ADR-015's Confirmation literal phrasing.
+
+**End users**: invoke `/wr-risk-scorer:assess-wip` instead. This wrapper is internal-use plumbing — calling it directly returns raw nudge output without the present-results layer that `/wr-risk-scorer:assess-wip` provides.
+
+## Contract
+
+- **Input** (`$ARGUMENTS`): a self-contained WIP-scoring prompt — typically the edited file path(s) plus a `git diff HEAD --stat` summary per `packages/risk-scorer/agents/wip.md`.
+- **Output**: the agent's verbatim report, including the WIP Risk Assessment markdown table, the cumulative pipeline risk picture, and the structured `RISK_VERDICT: CONTINUE | PAUSE | COMMIT` line.
+
+## Steps
+
+### 1. Pass-through to the wip agent
+
+Invoke the wip subagent via the Agent tool with the caller's `$ARGUMENTS` verbatim:
+
+```
+subagent_type: wr-risk-scorer:wip
+prompt: $ARGUMENTS
+```
+
+### 2. Return the agent report verbatim
+
+Return the agent's response to the caller without alteration. Do NOT strip, paraphrase, or post-process the `RISK_VERDICT:`, `RISK_REMEDIATIONS:`, or `RISK_COMMIT_REASON:` blocks — consumer SKILLs parse them directly.
+
+$ARGUMENTS