@kontourai/flow-agents 2.4.0 → 3.1.0

package/evals/integration/test_pull_work_assignment_join.sh

@@ -0,0 +1,132 @@
+#!/usr/bin/env bash
+# test_pull_work_assignment_join.sh — Two-session disjoint-selection simulation (#290).
+#
+# HONESTY NOTE (plan Design Decision / Stop-Short Risks): this eval simulates the assignment ⋈
+# liveness join's exclusion property with two actor structs evaluating the same candidate
+# subject id fixture in sequence, inside ONE process. It does NOT stand up two real concurrent
+# agent runtimes, and does not prove true concurrency (a real race between two simultaneous
+# writers). It is the closest *local* proof available for the issue's literal acceptance line
+# ("two sessions running pull-work against the same backlog select disjoint issues") without a
+# multi-runtime test harness this repo does not have — see the plan's Stop-Short Risks section.
+# What IS proven here, deterministically: once session A's assignment claim is recorded (local-
+# file or a rendered-then-status-confirmed GitHub fixture), session B's subsequent
+# `assignment-provider status` / join check for the SAME subject reports it `held` (excluded),
+# never `free` — so `pull-work`'s selection loop (kits/builder/skills/pull-work/SKILL.md,
+# "### 1. Read Board State") would skip it for session B.
+#
+# Supports AC12 (disjoint-selection-simulated), AC10 (docs pointer — see
+# docs/workflow-usage-guide.md's new Assignment Ownership subsection).
+set -uo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+CLI="$ROOT/build/src/cli.js"
+FIXTURES="$ROOT/evals/fixtures/assignment-provider"
+ACTOR_A="$FIXTURES/actor-a.json"
+ACTOR_B="$FIXTURES/actor-b.json"
+ISSUE_CLAIMED="$FIXTURES/github-issue-claimed.json"
+LIVENESS_FRESH="$FIXTURES/liveness-fresh.json"
+
+TMPDIR_EVAL="$(mktemp -d)"
+trap 'rm -rf "$TMPDIR_EVAL"' EXIT
+ARTIFACT_ROOT="$TMPDIR_EVAL/artifact-root"
+
+errors=0
+pass() { echo "  ✓ $1"; }
+fail() { echo "  ✗ $1"; errors=$((errors + 1)); }
+
+json_query() {
+  node -e 'const fs=require("fs"); let cur=JSON.parse(fs.readFileSync(process.argv[1],"utf8")); for (const part of process.argv[2].split(".")) cur=part==="length" ? cur.length : (Array.isArray(cur) ? cur[Number(part)] : cur[part]); console.log(cur);' "$1" "$2"
+}
+
+if [[ ! -f "$CLI" ]]; then
+  echo "build/src/cli.js not found — run 'npm run build' first" >&2
+  exit 1
+fi
+
+echo "=== Two-session disjoint-selection simulation (assignment ⋈ liveness join) ==="
+
+EMPTY_LIVENESS="$TMPDIR_EVAL/liveness-empty.json"
+echo '[]' > "$EMPTY_LIVENESS"
+
+# --- local-file leg: session A claims, session B's join check on the same subject is excluded ---
+
+SUBJECT_ID="kontourai/flow-agents#9301"
+
+# "Session A" (actor A) evaluates the candidate subject before anyone has claimed it: free.
+node "$CLI" assignment-provider status \
+  --provider local-file --artifact-root "$ARTIFACT_ROOT" --subject-id "$SUBJECT_ID" \
+  --self-actor "claude-code:eval-actor-a-session:eval-host" --liveness-events-json "$EMPTY_LIVENESS" \
+  > "$TMPDIR_EVAL/session-a-preclaim.json"
+[[ "$(json_query "$TMPDIR_EVAL/session-a-preclaim.json" "effective.effective_state")" == "free" ]] && pass "local-file: candidate subject is free before either session claims" || fail "local-file: candidate subject is free before either session claims"
+
+# Session A selects and claims it (pull-work's "Assignment Claim On Selection" step).
+node "$CLI" assignment-provider claim \
+  --provider local-file --artifact-root "$ARTIFACT_ROOT" --subject-id "$SUBJECT_ID" \
+  --actor-json "$ACTOR_A" --branch "agent/claude-code-eval-actor-a-session-eval-host/flow-agents-9301" \
+  --artifact-dir ".kontourai/flow-agents/flow-agents-9301" > /dev/null
+status=$?
+[[ "$status" -eq 0 ]] && pass "local-file: session A's claim on the candidate subject succeeds" || fail "local-file: session A's claim on the candidate subject succeeds"
+
+# "Session B" (actor B, a distinct actor struct — the second concurrent pull-work session)
+# evaluates the SAME candidate subject next. Its own liveness heartbeat is fresh (it is alive),
+# but that must not matter: the join is computed against the HOLDER's freshness, not the
+# reader's — session B must see the subject excluded.
+NOW_ISO="$(date -u +%Y-%m-%dT%H:%M:%SZ)"
+cat > "$TMPDIR_EVAL/liveness-actor-a-fresh.json" <<JSON
+[
+  {"type":"claim","subjectId":"$SUBJECT_ID","actor":"claude-code:eval-actor-a-session:eval-host","at":"$NOW_ISO","ttlSeconds":1800}
+]
+JSON
+node "$CLI" assignment-provider status \
+  --provider local-file --artifact-root "$ARTIFACT_ROOT" --subject-id "$SUBJECT_ID" \
+  --self-actor "claude-code:eval-actor-b-session:eval-host" \
+  --liveness-events-json "$TMPDIR_EVAL/liveness-actor-a-fresh.json" > "$TMPDIR_EVAL/session-b-postclaim.json"
+[[ "$(json_query "$TMPDIR_EVAL/session-b-postclaim.json" "assignment.assignee")" == "claude-code:eval-actor-a-session:eval-host" ]] && pass "local-file: session B's status read shows actor A as the holder" || fail "local-file: session B's status read shows actor A as the holder"
+[[ "$(json_query "$TMPDIR_EVAL/session-b-postclaim.json" "effective.effective_state")" == "held" ]] && pass "local-file: session B's join check reports held (excluded) for the same subject session A just claimed" || fail "local-file: session B's join check reports held (excluded) for the same subject session A just claimed"
+[[ "$(json_query "$TMPDIR_EVAL/session-b-postclaim.json" "effective.effective_state")" != "free" ]] && pass "local-file: subject is definitively NOT free for session B (disjoint selection holds)" || fail "local-file: subject is definitively NOT free for session B (disjoint selection holds)"
+
+# Session B, correctly excluding the held subject, selects a DIFFERENT subject instead — the
+# literal "select disjoint issues" property, one selection loop iteration at a time.
+SUBJECT_ID_2="kontourai/flow-agents#9302"
+node "$CLI" assignment-provider status \
+  --provider local-file --artifact-root "$ARTIFACT_ROOT" --subject-id "$SUBJECT_ID_2" \
+  --self-actor "claude-code:eval-actor-b-session:eval-host" --liveness-events-json "$EMPTY_LIVENESS" \
+  > "$TMPDIR_EVAL/session-b-alternate.json"
+[[ "$(json_query "$TMPDIR_EVAL/session-b-alternate.json" "effective.effective_state")" == "free" ]] && pass "local-file: session B's alternate candidate is free — it can select a disjoint subject" || fail "local-file: session B's alternate candidate is free — it can select a disjoint subject"
+node "$CLI" assignment-provider claim \
+  --provider local-file --artifact-root "$ARTIFACT_ROOT" --subject-id "$SUBJECT_ID_2" \
+  --actor-json "$ACTOR_B" --branch "agent/claude-code-eval-actor-b-session-eval-host/flow-agents-9302" \
+  --artifact-dir ".kontourai/flow-agents/flow-agents-9302" > /dev/null
+[[ $? -eq 0 ]] && pass "local-file: session B claims the disjoint subject successfully" || fail "local-file: session B claims the disjoint subject successfully"
+
+# Final proof: the two sessions' claimed subject sets are disjoint.
+node "$CLI" assignment-provider list --provider local-file --artifact-root "$ARTIFACT_ROOT" --actor-json "$ACTOR_A" > "$TMPDIR_EVAL/list-a.json"
+node "$CLI" assignment-provider list --provider local-file --artifact-root "$ARTIFACT_ROOT" --actor-json "$ACTOR_B" > "$TMPDIR_EVAL/list-b.json"
+DISJOINT="$(node -e '
+const fs = require("fs");
+const a = JSON.parse(fs.readFileSync(process.argv[1], "utf8")).subject_ids;
+const b = JSON.parse(fs.readFileSync(process.argv[2], "utf8")).subject_ids;
+const overlap = a.filter((id) => b.includes(id));
+console.log(overlap.length === 0 && a.length === 1 && b.length === 1 ? "yes" : "no");
+' "$TMPDIR_EVAL/list-a.json" "$TMPDIR_EVAL/list-b.json")"
+[[ "$DISJOINT" == "yes" ]] && pass "session A and session B's claimed subject sets are disjoint (one each, no overlap)" || fail "session A and session B's claimed subject sets are disjoint (one each, no overlap)"
+
+# --- GitHub leg: a rendered-and-status-confirmed claim on a fixture also excludes a second reader ---
+
+# The already-claimed GitHub fixture stands in for "session A already claimed and the render
+# was executed + status-confirmed" (SKILL.md's "Assignment Claim On Selection" round trip).
+# "Session B" reads the SAME fixture next with a fresh liveness fixture for the holder: held.
+node "$CLI" assignment-provider status --provider github --issue-json "$ISSUE_CLAIMED" \
+  --self-actor "claude-code:eval-actor-b-session:eval-host" \
+  --liveness-events-json "$LIVENESS_FRESH" --now "2026-06-01T12:20:00Z" \
+  > "$TMPDIR_EVAL/github-session-b.json"
+[[ "$(json_query "$TMPDIR_EVAL/github-session-b.json" "effective.effective_state")" == "held" ]] && pass "github: session B's join check on an already-claimed fixture issue reports held (excluded)" || fail "github: session B's join check on an already-claimed fixture issue reports held (excluded)"
+[[ "$(json_query "$TMPDIR_EVAL/github-session-b.json" "effective.effective_state")" != "free" ]] && pass "github: subject is definitively NOT free for session B" || fail "github: subject is definitively NOT free for session B"
+
+echo ""
+if [[ "$errors" -eq 0 ]]; then
+  echo "test_pull_work_assignment_join: all checks passed."
+else
+  echo "test_pull_work_assignment_join: $errors check(s) failed."
+fi
+exit "$errors"

package/evals/integration/test_pull_work_liveness_preflight.sh

@@ -151,7 +151,12 @@ if flow_agents_node "$WRITER" ensure-session \
   --title "Parity Check" \
   --summary "resolve-slug parity fixture." \
   --timestamp "2026-07-01T00:00:00Z" >"$TMPDIR_EVAL/b-ensure.out" 2>"$TMPDIR_EVAL/b-ensure.err"; then
-  B_SESSION_DIR="$(find "$B_ROOT" -mindepth 1 -maxdepth 1 -type d ! -name liveness | head -n1)"
+  # #291: ensure-session (when an ambient actor resolves, even without an explicit --actor)
+  # now also creates artifact-root-level "assignment/" (the ownership-guard's durable local-file
+  # claim record dir) and "current/" (the per-actor current.json projection dir) sibling
+  # directories, alongside the pre-existing "liveness/" dir -- all three must be excluded here so
+  # this find still lands on the real session slug directory, not a sibling artifact dir.
+  B_SESSION_DIR="$(find "$B_ROOT" -mindepth 1 -maxdepth 1 -type d ! -name liveness ! -name assignment ! -name current | head -n1)"
   B_SLUG_FROM_DIR="$(basename "$B_SESSION_DIR")"
   B_SLUG_FROM_RESOLVE="$(flow_agents_node "$WRITER" resolve-slug 'owner/repo#42' 2>"$TMPDIR_EVAL/b-resolve.err")"
   if [[ -n "$B_SESSION_DIR" ]] && [[ "$B_SLUG_FROM_DIR" == "$B_SLUG_FROM_RESOLVE" ]]; then
@@ -329,7 +334,7 @@ require_text "$PULL" 'liveness claim <subjectId>' "pull-work references liveness
 require_text "$PULL" 'reading each row.s raw .status. field \(never .label.\)' "pull-work instructs consuming raw status, never label (AC1, AC8)"
 require_text "$PULL" 'a .verified. row for an actor other than self.*.held.' "pull-work documents held classification (AC1, AC2, AC8)"
 require_text "$PULL" 'excluded from the ready set by default' "pull-work excludes held candidates by default (AC2, AC8)"
-require_text "$PULL" 'a .stale. row \(any actor\).*.reclaimable.' "pull-work documents reclaimable classification (AC1, AC3, AC8)"
+require_text "$PULL" 'effective_state: .reclaimable.' "pull-work documents reclaimable classification via the assignment-provider join (AC1, AC3, AC8, #290)"
 require_text "$PULL" 'requires an explicit recorded opt-in' "pull-work requires an explicit recorded opt-in for reclaimable selection (AC3, AC8)"
 require_text "$PULL" 'never a silent normal pick' "pull-work states reclaimable is never a silent normal pick (AC3, AC8)"
 require_text "$PULL" 'a .verified. row for the resolved self actor.*.mine.' "pull-work documents mine classification (AC1, AC5, AC8)"
@@ -348,10 +353,15 @@ require_text "$PICKUP_PROBE" 'flags the selected item .reclaimable., treat the r
 # ─── F2-F6 (fix-plan iteration 1) static assertions ─────────────────────────
 require_text "$PULL" 'group all rows for that .subjectId. by subject' "pull-work classifies per-subject (grouped rows), not per-row (F2, AC1, AC8)"
 require_text "$PULL" 'classify in this precedence order' "pull-work states the classification is an ordered precedence rule (F2, AC8)"
-require_text "$PULL" '^1\. a .verified. row for an actor other than self' "pull-work's precedence rule step 1 (held) is numbered verbatim (F2, AC8)"
-require_text "$PULL" '^2\. else, a .verified. row for the resolved self actor' "pull-work's precedence rule step 2 (mine) is numbered verbatim (F2, AC8)"
-require_text "$PULL" '^3\. else, a .stale. row' "pull-work's precedence rule step 3 (reclaimable) is numbered verbatim (F2, AC8)"
-require_text "$PULL" '^4\. else \(a .revoked.-only row' "pull-work's precedence rule step 4 (free) is numbered verbatim (F2, AC8)"
+# F2's original numbered precedence list was liveness-only (#166); #290 (Wave 3) replaced it with
+# the full ADR 0021 §1 assignment ⋈ liveness join, computed via `assignment-provider status`'s
+# `effective_state`/`reason` fields. These assertions are updated to match the new five-row list
+# (mine/held/reclaimable/human-held/free) rather than the old four-row liveness-only list.
+require_text "$PULL" '^1\. .effective_state: .held.. with .reason: .self_is_holder..' "pull-work's precedence rule step 1 (mine, via self_is_holder) is numbered verbatim (F2, #290, AC8)"
+require_text "$PULL" '^2\. else .effective_state: .held..' "pull-work's precedence rule step 2 (held, excluded) is numbered verbatim (F2, #290, AC8)"
+require_text "$PULL" '^3\. else .effective_state: .reclaimable..' "pull-work's precedence rule step 3 (reclaimable) is numbered verbatim (F2, #290, AC8)"
+require_text "$PULL" '^4\. else .effective_state: .human-held..' "pull-work's precedence rule step 4 (human-held) is numbered verbatim (F2, #290, AC11, AC8)"
+require_text "$PULL" '^5\. else .effective_state: .free..' "pull-work's precedence rule step 5 (free) is numbered verbatim (F2, #290, AC8)"
 require_text "$PULL" 'double-hold' "pull-work names the own-actor + other-actor co-existing row combination a double-hold (F2, AC8)"
 require_text "$PULL" 'explicit conflict warning \(per ADR 0012 §4 detection' "pull-work cites ADR 0012 §4 detection for the double-hold conflict warning (F2, AC8)"
 require_text "$PULL" 'never silently resolve it to .mine.' "pull-work states a double-hold is never silently resolved to mine (F2, AC8)"
@@ -373,7 +383,10 @@ require_text "$PULL" '### Post-Claim Conflict Re-check' "pull-work documents the
 require_text "$PULL" 'coexists with this session.s own just-emitted claim on the same subject' "pull-work's post-claim re-check detects a double-hold against the just-emitted claim (F6, AC8)"
 require_text "$PULL" 'surface the conflict prominently in the user-facing output and instruct the user to coordinate before proceeding' "pull-work instructs surfacing the post-claim conflict prominently and coordinating before proceeding, not silently continuing (F6, AC8)"
 require_text "$PULL" 'does not provide true mutual exclusion across the read-then-write race window itself' "pull-work's honesty note states the exclusion guarantee still does not provide true mutual exclusion across the read-then-write race window (F6, AC8; updated #320 wording — see also test_liveness_verdict.sh)"
-require_text "$PULL" 'True mutual exclusion arrives with the provider assignment lease \(#290\) and the .verify-hold. publish gate \(#293\)' "pull-work names #290/#293 as where true mutual exclusion actually arrives (F6, AC8)"
+require_text "$PULL" '.#290.\x27s provider assignment lease closes this gap for the .{0,2}local-file.{0,2} provider only' "pull-work scopes true mutual exclusion to the local-file provider only, via #290's lock (fix-plan iteration 1, F5, AC8)"
+require_text "$PULL" 'two concurrent local-file .claim. calls on the same subject genuinely cannot both win' "pull-work states local-file claim/claim races genuinely cannot both win under #290's lock (fix-plan iteration 1, F5, AC8)"
+require_text "$PULL" 'The .{0,2}GitHub.{0,2} provider \(assignee/label/claim-comment\) remains advisory/last-writer' "pull-work does NOT overclaim true mutual exclusion for the GitHub provider — advisory/last-writer, detect not prevent (fix-plan iteration 1, F5, AC8)"
+require_text "$PULL" 'Do not read .provider assignment lease. as closing the GitHub race; only the .verify-hold. publish gate \(#293\)' "pull-work explicitly warns not to read #290 as closing the GitHub race — only #293's verify-hold gate would (fix-plan iteration 1, F5, AC8)"
 
 
 echo ""

package/evals/integration/test_reconcile_soundness.sh

@@ -9,8 +9,13 @@
 #  B. REAL-VERIFY-CLOSED:     build passes but a fake "real verify" (eval:static substitute)
 #     fails → trust-reconcile exits 1 (PRE-FIX this would exit 0 because only build ran).
 #
-#  C. REAL-VERIFY-PASSES:     the comprehensive verify resolves green → exits 0 (legit
-#     work is not false-blocked when package.json trust-reconcile-verify is present).
+#  C. REAL-VERIFY-PASSES:     the comprehensive verify resolves green, no bundle, but a
+#     well-formed in-scope delivery/DECLARED marker exempts Step 2 (ADR 0022 §1) → exits 0
+#     with the DECLARED (no-agent-delivery) line (legit work is not false-blocked; the
+#     bundle-absence branch is bundle-required by default post-ADR-0022, so this case is
+#     now anchored by a DECLARED marker rather than by bundle absence alone — see
+#     evals/integration/test_trust_reconcile_negatives.sh section 7 for the no-marker /
+#     out-of-scope-marker fail-closed cases this test does NOT re-cover).
 #
 #  D. CHECKPOINT-BYPASS-CLOSED: a checkpoint-only bundle (no evidence/claims, statusByClaimId
 #     all-passed) → exits 1 with checkpoint-bypass divergence (not a silent skip).
@@ -56,8 +61,14 @@ fs.writeFileSync('$PKG_NO_VERIFY/package.json', JSON.stringify(pkg, null, 2));
 "
 
 # ─── Minimal package.json WITH trust-reconcile-verify ─────────────────────────
+# ADR 0022 §1 (bundle-required by default): no bundle is written for this fixture, so a
+# well-formed, in-scope delivery/DECLARED marker (ADR 0022 §2) is seeded alongside it —
+# this is the new legitimate no-bundle path (a real human/bot PR author with no agent
+# delivery), preserving TEST C's original "no bundle" shape more faithfully than
+# switching it to a --bundle fixture would (a bundle would test Step 2 reconcile, which
+# TEST C was never about — it proves Step 1 fresh-verify success is honored).
 PKG_WITH_VERIFY="$TMP/pkg_with_verify"
-mkdir -p "$PKG_WITH_VERIFY"
+mkdir -p "$PKG_WITH_VERIFY/delivery"
 node -e "
 const fs = require('fs');
 const pkg = {
@@ -68,6 +79,13 @@ const pkg = {
   }
 };
 fs.writeFileSync('$PKG_WITH_VERIFY/package.json', JSON.stringify(pkg, null, 2));
+const declared = {
+  scope: 'ref:test-verify-passes',
+  reason: 'test fixture: real verify passes, no agent delivery bundle',
+  approved_by: 'test-harness',
+  declared_at: '2026-06-27T00:00:00Z'
+};
+fs.writeFileSync('$PKG_WITH_VERIFY/delivery/DECLARED', JSON.stringify(declared, null, 2));
 "
 
 # ─── Minimal package.json with failing verify ─────────────────────────────────
@@ -255,22 +273,28 @@ fi
 # package.json has passing trust-reconcile-verify → exits 0 (not false-blocked)
 # ═══════════════════════════════════════════════════════════════════════════════
 echo ""
-echo "=== TEST C: REAL-VERIFY-PASSES — passing verify + no bundle → exits 0 ==="
+echo "=== TEST C: REAL-VERIFY-PASSES — passing verify + in-scope delivery/DECLARED marker (no bundle) → exits 0 with DECLARED exemption ==="
 
-outC=$(node "$RECONCILE" \
+outC=$(TRUST_RECONCILE_REF="test-verify-passes" node "$RECONCILE" \
     --repo-root "$PKG_WITH_VERIFY" 2>&1)
 exitC=$?
 
 if [[ $exitC -eq 0 ]]; then
-  _pass "REAL-VERIFY-PASSES: exits 0 when real verify passes and no bundle (got $exitC)"
+  _pass "REAL-VERIFY-PASSES: exits 0 when real verify passes and a well-formed, in-scope delivery/DECLARED marker exempts Step 2 (got $exitC)"
 else
   _fail "REAL-VERIFY-PASSES: expected exit 0, got $exitC — output: $outC"
 fi
 
-if echo "$outC" | grep -q "fresh verify passed"; then
-  _pass "REAL-VERIFY-PASSES: 'fresh verify passed' message present"
+if echo "$outC" | grep -qF "DECLARED (no-agent-delivery): ref:test-verify-passes — test fixture: real verify passes, no agent delivery bundle (approved by test-harness, declared 2026-06-27T00:00:00Z)"; then
+  _pass "REAL-VERIFY-PASSES: exact DECLARED (no-agent-delivery) exemption line present"
+else
+  _fail "REAL-VERIFY-PASSES: expected the exact DECLARED (no-agent-delivery) line, got: $outC"
+fi
+
+if echo "$outC" | grep -q "running: npm run trust-reconcile-verify"; then
+  _pass "REAL-VERIFY-PASSES: Step 1 (fresh verify) still ran the real trust-reconcile-verify command (not skipped by the exemption)"
 else
-  _fail "REAL-VERIFY-PASSES: expected 'fresh verify passed' in output, got: $outC"
+  _fail "REAL-VERIFY-PASSES: expected Step 1 to have run 'npm run trust-reconcile-verify', got: $outC"
 fi
 
 # ═══════════════════════════════════════════════════════════════════════════════

package/evals/integration/test_trust_reconcile.sh

@@ -6,8 +6,9 @@
 #      Exit 1 with "trust divergence" message naming the command.
 #   2. MATCHING-PASSES:    bundle claims a command passed; CI re-runs it and it PASSES.
 #      Exit 0 (no divergence).
-#   3. NO-CHECKPOINT:      no bundle present; canonical verify passes.
-#      Exit 0 (fail-open on bundle absence, enforce fresh verify only).
+#   3. NO-CHECKPOINT:      no bundle present, no delivery/DECLARED marker; canonical
+#      verify passes. Exit 1 — bundle required by default (ADR 0022 §1); the
+#      'bundle-required-no-declared-marker' issue fires (no more fail-open on absence).
 #   4. LAUNDERING-CAUGHT:  bundle claims "something || true" passed.
 #      Exit 1 with laundering message (checked before "CI never ran" check).
 #   YAML-VALID:            .github/workflows/trust-reconcile.yml parses as valid YAML.
@@ -158,16 +159,16 @@ out3=$(TRUST_RECONCILE_COMMANDS="node -e 'process.exit(0)'" \
     --repo-root "$TMP" 2>&1)
 exit3=$?
 
-if [[ $exit3 -eq 0 ]]; then
-  _pass "NO-CHECKPOINT: exits 0 (fresh verify passes, no bundle)"
+if [[ $exit3 -ne 0 ]]; then
+  _pass "NO-CHECKPOINT: exits 1 (bundle required by default, no bundle, no marker) — got $exit3"
 else
-  _fail "NO-CHECKPOINT: expected exit 0, got $exit3 — output: $out3"
+  _fail "NO-CHECKPOINT: expected exit 1 (bundle-required by default, ADR 0022 §1), got 0 — output: $out3"
 fi
 
-if echo "$out3" | grep -q "fail-open"; then
-  _pass "NO-CHECKPOINT: output notes fail-open on bundle absence"
+if echo "$out3" | grep -q "bundle-required-no-declared-marker"; then
+  _pass "NO-CHECKPOINT: output contains 'bundle-required-no-declared-marker'"
 else
-  _fail "NO-CHECKPOINT: expected 'fail-open' in output, got: $out3"
+  _fail "NO-CHECKPOINT: expected 'bundle-required-no-declared-marker' in output, got: $out3"
 fi
 
 # Also verify: no-bundle + failing fresh verify still exits 1