@kontourai/flow-agents 0.1.1

package/evals/README.md

@@ -0,0 +1,128 @@
+# Flow Agents Eval Suite
+
+Evaluation coverage for the canonical Flow Agents source tree and generated universal bundles.
+
+## Quick Start
+
+```bash
+npm install
+
+# Run the fast local gate: source validation, static package checks, integration checks
+bash evals/run.sh
+
+# Run only source/static checks
+bash evals/run.sh static
+
+# Run only integration checks
+bash evals/run.sh integration
+
+# Run harness-native acceptance checks
+bash evals/run.sh acceptance
+
+# Claude acceptance is cheap by default. Opt in to prompt-mode Claude usage only when needed.
+FLOW_AGENTS_ACCEPTANCE_CLAUDE_LLM=1 bash evals/run.sh acceptance claude
+
+# Run behavioral evals through the default Kiro runtime
+bash evals/run.sh llm
+
+# Run one behavioral suite through Codex as subject runtime and judge
+bash evals/run.sh llm dev --runtime codex
+
+# Run Claude Code as the subject runtime while Codex judges rubrics
+bash evals/run.sh llm dev --runtime claude --judge-runtime codex
+
+# Run cheaper behavioral subsets
+bash evals/run.sh llm dev --suite smoke
+bash evals/run.sh llm dev --suite regression
+
+# View promptfoo results
+npm run promptfoo:view
+```
+
+## Layers
+
+### Layer 1: Static (`bash evals/run.sh static`)
+
+Validates the source tree and generated bundle exports:
+- canonical source validation via `npm run validate:source --`
+- package shape, schemas, resources, hooks, routing, MCP server references, write-tool invariants, and agent cards
+- universal bundle build/export checks for Kiro, Claude Code, and Codex
+
+Runs in seconds and has no LLM cost.
+
+### Layer 2: Integration (`bash evals/run.sh integration`)
+
+Validates runtime-adjacent contracts:
+- telemetry event schemas, type mapping, field presence, prompt capture, tool capture, redaction, and agent discovery
+- workflow artifact quality and deterministic end-to-end delivery chain fixtures
+- bundle install smoke tests for Kiro, Claude Code, and Codex temp installs
+
+Runs in seconds and has no LLM cost.
+
+### Layer 3: Behavioral (`bash evals/run.sh llm`)
+
+Runs selected agents through an eval runtime and scores responses with deterministic telemetry assertions plus LLM rubrics. Kiro is the default subject runtime. Pass `--runtime codex` or `--runtime claude` to run Codex or Claude Code where supported.
+
+Subject runtime and judge runtime are separate:
+
+```bash
+bash evals/run.sh llm dev --runtime claude --judge-runtime codex
+bash evals/run.sh llm dev --runtime claude --judge-runtime claude
+```
+
+Use `--suite smoke`, `--suite regression`, or `--suite capability` to avoid running the full behavioral suite when a targeted gate is enough. `smoke` runs the first few cases, `regression` filters `metadata.type=regression`, and `capability` filters `metadata.type=capability`.
+
+Current behavioral suites:
+- `dev`
+
+The root `evals/promptfooconfig.yaml` is a legacy combined promptfoo config for targeted manual runs. Prefer `bash evals/run.sh llm <agent>` or the per-agent configs in `evals/cases/<agent>/promptfooconfig.yaml`.
+
+### Layer 4: Acceptance (`bash evals/run.sh acceptance`)
+
+Runs harness-native smoke tests against generated bundles:
+- `Claude Code` discovers workspace agents and can answer through `dev`
+- `claude` lists project agents and verifies exported telemetry hook configuration without model usage by default
+- `codex exec` loads the exported `.codex` bundle and returns a final response
+
+This layer is environment-dependent and requires installed, authenticated CLIs.
+
+Claude prompt-mode acceptance is opt-in with `FLOW_AGENTS_ACCEPTANCE_CLAUDE_LLM=1`. Real Claude CLI hook telemetry assertions are opt-in with `FLOW_AGENTS_ACCEPTANCE_REQUIRE_CLAUDE_TELEMETRY=1`; deterministic integration tests cover the telemetry wrapper without spending Claude usage.
+
+## Coverage
+
+Covered now:
+- source/package drift and bundle export drift
+- telemetry schema and redaction contracts
+- install smoke tests for generated bundles
+- normalized telemetry for Kiro, Codex, and Claude Code hook events
+- behavioral routing and workflow checks for the supported per-agent suites
+
+Deferred:
+- multi-turn conversation evals
+- adversarial/red-team evals
+- behavioral coverage for every exported tool agent
+- full LLM-driven end-to-end delivery runs on every edit; deterministic artifact-chain E2E coverage runs in integration
+- direct token usage assertions, because CLI-backed exec providers do not expose reliable token counts today
+
+## Adding Eval Cases
+
+Add behavioral cases to `evals/cases/<agent>/promptfooconfig.yaml`. Each test should include:
+- `vars.prompt` with the user prompt
+- `options.provider.id` or suite default provider
+- deterministic assertions when telemetry can prove the behavior
+- an `llm-rubric` for workflow quality when needed
+- `metadata.type` set to `capability` or `regression`
+
+Run the affected suite with:
+
+```bash
+bash evals/run.sh llm <agent>
+```
+
+## Prerequisites
+
+- `jq` for static and integration checks
+- `Claude Code` for default behavioral and Kiro acceptance checks
+- `codex` for `--runtime codex`, `--judge-runtime codex`, and Codex acceptance checks
+- `claude` for `--runtime claude`, `--judge-runtime claude`, and Claude Code acceptance checks
+- `promptfoo` for behavioral evals and result viewing, installed with `npm install` from the repo root

package/evals/acceptance/run.sh

@@ -0,0 +1,29 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+ACCEPT_DIR="$ROOT_DIR/evals/acceptance"
+TARGET="${1:-all}"
+
+run_one() {
+  local name="$1"
+  echo ""
+  bash "$ACCEPT_DIR/test_${name}_harness.sh"
+}
+
+case "$TARGET" in
+  kiro|claude|codex)
+    run_one "$TARGET"
+    ;;
+  all)
+    status=0
+    run_one kiro || status=1
+    run_one claude || status=1
+    run_one codex || status=1
+    exit "$status"
+    ;;
+  *)
+    echo "Usage: bash evals/acceptance/run.sh [all|kiro|claude|codex]"
+    exit 1
+    ;;
+esac

package/evals/acceptance/test_claude_harness.sh

@@ -0,0 +1,242 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+source "$ROOT_DIR/evals/lib/node.sh"
+TMP_WORK=""
+pass=0
+fail=0
+skip=0
+
+cleanup() {
+  [[ -n "$TMP_WORK" ]] && rm -rf "$TMP_WORK"
+}
+trap cleanup EXIT
+
+_pass() { echo "  ✓ $1"; pass=$((pass + 1)); }
+_fail() { echo "  ✗ $1"; fail=$((fail + 1)); }
+_skip() { echo "  ○ $1"; skip=$((skip + 1)); }
+
+wait_for_telemetry() {
+  local file="$1"
+  local i=0
+  while [[ $i -lt 50 ]]; do
+    [[ -s "$file" ]] && return 0
+    sleep 0.1
+    i=$((i + 1))
+  done
+  return 1
+}
+
+echo "=== Harness Acceptance: Claude Code ==="
+echo ""
+echo "Set FLOW_AGENTS_ACCEPTANCE_CLAUDE_LLM=1 to run prompt-mode Claude checks."
+echo "Set FLOW_AGENTS_ACCEPTANCE_REQUIRE_CLAUDE_TELEMETRY=1 to require real Claude CLI hook telemetry."
+echo ""
+
+if ! command -v claude >/dev/null 2>&1; then
+  _skip "claude CLI not installed"
+  echo ""
+  echo "Results: ${pass}/$((pass + fail)) passed, ${fail} failed, ${skip} skipped"
+  exit 0
+fi
+
+cd "$ROOT_DIR"
+flow_agents_node scripts/build-universal-bundles.js >/dev/null
+
+TMP_WORK="$(mktemp -d /tmp/claude-acceptance-work.XXXXXX)"
+bash dist/claude-code/install.sh "$TMP_WORK" >/dev/null
+
+echo "--- Agent List ---"
+list_output="$(cd "$TMP_WORK" && claude agents --setting-sources local,project,user 2>&1 || true)"
+if echo "$list_output" | grep -q "Project agents:"; then
+  _pass "claude lists project agents"
+else
+  _fail "claude did not list project agents"
+fi
+
+if echo "$list_output" | grep -q "dev ·"; then
+  _pass "claude project agent list includes dev"
+else
+  _fail "claude project agent list did not include dev"
+fi
+
+if [[ -f "$TMP_WORK/.claude/settings.json" ]] && grep -q "claude-telemetry-hook.js" "$TMP_WORK/.claude/settings.json"; then
+  _pass "claude project settings include telemetry hooks"
+else
+  _fail "claude project settings missing telemetry hooks"
+fi
+
+if [[ "${FLOW_AGENTS_ACCEPTANCE_CLAUDE_LLM:-0}" != "1" ]]; then
+  _skip "Claude prompt-mode checks skipped to avoid model usage"
+  echo ""
+  echo "==========================="
+  total=$((pass + fail))
+  echo "Results: ${pass}/${total} passed, ${fail} failed, ${skip} skipped"
+  [[ "$fail" -gt 0 ]] && exit 1
+  exit 0
+fi
+
+echo ""
+echo "--- Print Smoke ---"
+print_output="$(cd "$TMP_WORK" && claude -p --agent dev --permission-mode bypassPermissions --add-dir "$TMP_WORK" --output-format text "Reply with READY only." 2>&1 || true)"
+if echo "$print_output" | grep -qx "READY"; then
+  _pass "dev agent replied READY in print mode"
+else
+  _fail "dev agent did not return plain READY in print mode"
+fi
+
+echo ""
+echo "--- Behavioral Route ---"
+route_output="$(cd "$TMP_WORK" && node - <<'NODE'
+const { spawnSync } = require("node:child_process");
+const result = spawnSync("claude", [
+  "-p",
+  "--agent",
+  "dev",
+  "--permission-mode",
+  "bypassPermissions",
+  "--add-dir",
+  ".",
+  "--output-format",
+  "text",
+  "A user asks: 'Explore the codebase and explain what it does.' Which skill should you activate first? Reply with only the skill name or NONE.",
+], { encoding: "utf8", timeout: 30000 });
+process.stdout.write(result.stdout || "");
+process.stdout.write(result.stderr || "");
+NODE
+)"
+route_output_trimmed="$(printf '%s' "$route_output" | tr -d '\r' | tail -n 1 | sed 's/^[[:space:]]*//;s/[[:space:]]*$//')"
+if [[ "$route_output_trimmed" == "explore" ]]; then
+  _pass "claude dev selects explore for repository exploration"
+else
+  _fail "claude dev did not select explore (got: $route_output_trimmed)"
+fi
+
+echo ""
+echo "--- deliver Route ---"
+sa_build_output="$(cd "$TMP_WORK" && node - <<'NODE'
+const { spawnSync } = require("node:child_process");
+const result = spawnSync("claude", [
+  "-p",
+  "--agent",
+  "dev",
+  "--permission-mode",
+  "bypassPermissions",
+  "--add-dir",
+  ".",
+  "--output-format",
+  "text",
+  "A user asks: 'Build a CLI tool that converts markdown files to HTML'. Which skill should you activate first? Reply with only the skill name or NONE.",
+], { encoding: "utf8", timeout: 30000 });
+process.stdout.write(result.stdout || "");
+process.stdout.write(result.stderr || "");
+NODE
+)"
+sa_build_trimmed="$(printf '%s' "$sa_build_output" | tr -d '\r' | tail -n 1 | sed 's/^[[:space:]]*//;s/[[:space:]]*$//')"
+if [[ "$sa_build_trimmed" == "deliver" ]]; then
+  _pass "claude dev selects deliver for broad build requests"
+else
+  _fail "claude dev did not select deliver (got: $sa_build_trimmed)"
+fi
+
+echo ""
+echo "--- Live Hook Influence ---"
+mkdir -p "$TMP_WORK/.flow-agents/live-hook" "$TMP_WORK/docs"
+printf '# Context Map\n' > "$TMP_WORK/docs/context-map.md"
+cat > "$TMP_WORK/.flow-agents/live-hook/state.json" <<'JSON'
+{
+  "schema_version": "1.0",
+  "task_slug": "live-hook",
+  "status": "not_verified",
+  "phase": "verification",
+  "updated_at": "2026-05-10T00:00:00Z",
+  "next_action": {
+    "status": "needs_user",
+    "summary": "Acknowledge live hook guidance.",
+    "target_phase": "verification"
+  }
+}
+JSON
+cat > "$TMP_WORK/.flow-agents/live-hook/critique.json" <<'JSON'
+{
+  "schema_version": "1.0",
+  "task_slug": "live-hook",
+  "status": "fail",
+  "required": true,
+  "updated_at": "2026-05-10T00:01:00Z",
+  "critiques": [
+    {
+      "id": "live-hook-review",
+      "reviewer": "tool-code-reviewer",
+      "reviewed_at": "2026-05-10T00:01:00Z",
+      "verdict": "fail",
+      "summary": "Live hook guidance must be acknowledged.",
+      "findings": [
+        {
+          "id": "open-live-hook",
+          "severity": "high",
+          "status": "open",
+          "description": "Report the unfinished workflow state."
+        }
+      ]
+    }
+  ]
+}
+JSON
+hook_output="$(cd "$TMP_WORK" && node - <<'NODE'
+const { spawnSync } = require("node:child_process");
+const result = spawnSync("claude", [
+  "-p",
+  "--agent",
+  "dev",
+  "--permission-mode",
+  "bypassPermissions",
+  "--add-dir",
+  ".",
+  "--output-format",
+  "text",
+  "Use a harmless tool first, such as listing the current directory. After that, if Flow Agents hook guidance mentions WORKFLOW STATE ATTENTION or task live-hook, reply exactly HOOK_GUIDANCE_SEEN live-hook. If no such guidance is visible, reply exactly HOOK_GUIDANCE_MISSING.",
+], { encoding: "utf8", timeout: 45000 });
+process.stdout.write(result.stdout || "");
+process.stdout.write(result.stderr || "");
+NODE
+)"
+hook_output_trimmed="$(printf '%s' "$hook_output" | tr -d '\r' | tail -n 1 | sed 's/^[[:space:]]*//;s/[[:space:]]*$//')"
+if [[ "$hook_output_trimmed" == "HOOK_GUIDANCE_SEEN live-hook" ]]; then
+  _pass "claude live session responds to workflow hook guidance"
+else
+  _fail "claude live session did not respond to workflow hook guidance (got: $hook_output_trimmed)"
+fi
+
+echo ""
+echo "--- Telemetry ---"
+telemetry_file="$TMP_WORK/.telemetry/full.jsonl"
+if [[ "${FLOW_AGENTS_ACCEPTANCE_REQUIRE_CLAUDE_TELEMETRY:-0}" != "1" ]]; then
+  _skip "real Claude CLI telemetry assertion skipped"
+else
+  if wait_for_telemetry "$telemetry_file"; then
+    _pass "claude telemetry log was written"
+  else
+    _fail "claude telemetry log was not written"
+  fi
+
+  if [[ -f "$telemetry_file" ]] && jq -e 'select(.agent.runtime == "claude-code")' "$telemetry_file" >/dev/null 2>&1; then
+    _pass "claude telemetry uses normalized claude-code runtime"
+  else
+    _fail "claude telemetry did not include claude-code runtime"
+  fi
+
+  if [[ -f "$telemetry_file" ]] && jq -e 'select(.event_type == "turn.user")' "$telemetry_file" >/dev/null 2>&1; then
+    _pass "claude telemetry captures user prompts"
+  else
+    _fail "claude telemetry did not capture user prompts"
+  fi
+fi
+
+echo ""
+echo "==========================="
+total=$((pass + fail))
+echo "Results: ${pass}/${total} passed, ${fail} failed, ${skip} skipped"
+[[ "$fail" -gt 0 ]] && exit 1
+exit 0

package/evals/acceptance/test_codex_harness.sh

@@ -0,0 +1,108 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+source "$ROOT_DIR/evals/lib/node.sh"
+TMP_WORK=""
+TMP_LOG=""
+TMP_LAST=""
+pass=0
+fail=0
+skip=0
+
+cleanup() {
+  [[ -n "$TMP_WORK" ]] && rm -rf "$TMP_WORK"
+  [[ -n "$TMP_LOG" ]] && rm -f "$TMP_LOG"
+  [[ -n "$TMP_LAST" ]] && rm -f "$TMP_LAST"
+}
+trap cleanup EXIT
+
+_pass() { echo "  ✓ $1"; pass=$((pass + 1)); }
+_fail() { echo "  ✗ $1"; fail=$((fail + 1)); }
+_skip() { echo "  ○ $1"; skip=$((skip + 1)); }
+
+echo "=== Harness Acceptance: Codex ==="
+echo ""
+
+if ! command -v codex >/dev/null 2>&1; then
+  _skip "codex CLI not installed"
+  echo ""
+  echo "Results: ${pass}/$((pass + fail)) passed, ${fail} failed, ${skip} skipped"
+  exit 0
+fi
+
+cd "$ROOT_DIR"
+flow_agents_node scripts/build-universal-bundles.js >/dev/null
+
+TMP_WORK="$(mktemp -d /tmp/codex-acceptance-work.XXXXXX)"
+TMP_LOG="$(mktemp /tmp/codex-acceptance-log.XXXXXX)"
+TMP_LAST="$(mktemp /tmp/codex-acceptance-last.XXXXXX)"
+bash dist/codex/install.sh "$TMP_WORK" >/dev/null
+
+echo "--- Exec Smoke ---"
+if codex exec --skip-git-repo-check -C "$TMP_WORK" --sandbox read-only --json --output-last-message "$TMP_LAST" "After any required startup checks, reply with READY only." >"$TMP_LOG" 2>&1; then
+  _pass "codex exec completed successfully"
+else
+  _fail "codex exec exited non-zero"
+fi
+
+if grep -q "Ignoring malformed agent role definition" "$TMP_LOG"; then
+  _fail "codex reported malformed exported agent roles"
+else
+  _pass "codex accepted exported local agent role files"
+fi
+
+if grep -q "failed to stat skills path" "$TMP_LOG"; then
+  _fail "codex could not stat exported skill paths"
+else
+  _pass "codex resolved exported skill paths"
+fi
+
+if grep -q "READY" "$TMP_LAST"; then
+  _pass "codex returned READY in final message"
+else
+  _fail "codex final message did not contain READY"
+fi
+
+echo ""
+echo "--- Behavioral Route ---"
+TMP_ROUTE_LOG="$(mktemp /tmp/codex-acceptance-route.XXXXXX)"
+if node -e 'const fs=require("fs"); const cp=require("child_process"); const [work,log]=process.argv.slice(1); const r=cp.spawnSync("codex",["exec","--skip-git-repo-check","-C",work,"--sandbox","read-only","--json","Before doing anything else, state the exact skill you are activating if any, then explore the codebase and explain what it does."],{encoding:"utf8",timeout:45000}); fs.writeFileSync(log,(r.stdout||"")+(r.stderr||"")); process.exit(r.error?.code==="ETIMEDOUT" ? 0 : (r.status ?? 1));' "$TMP_WORK" "$TMP_ROUTE_LOG"
+then
+  _pass "codex behavioral route command completed successfully"
+else
+  _fail "codex behavioral route command exited non-zero"
+fi
+
+if grep -Fq 'Activating `$explore`' "$TMP_ROUTE_LOG" || grep -Fq 'Activating skill: `explore`' "$TMP_ROUTE_LOG" || grep -Fq 'Activating skill: explore' "$TMP_ROUTE_LOG"; then
+  _pass "codex dev activates explore on repository exploration"
+else
+  _fail "codex dev did not activate explore on repository exploration"
+fi
+
+rm -f "$TMP_ROUTE_LOG"
+
+echo ""
+echo "--- deliver Route ---"
+TMP_BUILD_LOG="$(mktemp /tmp/codex-acceptance-build.XXXXXX)"
+if node -e 'const fs=require("fs"); const cp=require("child_process"); const [work,log]=process.argv.slice(1); const r=cp.spawnSync("codex",["exec","--skip-git-repo-check","-C",work,"--sandbox","read-only","--json","Before doing anything else, state the exact skill you are activating if any, then begin the deliver workflow for '\''Build a CLI tool that converts markdown files to HTML'\'', but stop after deciding the initial skill and first phase."],{encoding:"utf8",timeout:45000}); fs.writeFileSync(log,(r.stdout||"")+(r.stderr||"")); process.exit(r.error?.code==="ETIMEDOUT" ? 0 : (r.status ?? 1));' "$TMP_WORK" "$TMP_BUILD_LOG"
+then
+  _pass "codex deliver route command completed successfully"
+else
+  _fail "codex deliver route command exited non-zero"
+fi
+
+if grep -Fq 'Activating skill: `$deliver`' "$TMP_BUILD_LOG" || grep -Fq 'Activating skill: `deliver`' "$TMP_BUILD_LOG" || grep -Fq 'Activating skill: deliver' "$TMP_BUILD_LOG"; then
+  _pass "codex dev activates deliver for broad build requests"
+else
+  _fail "codex dev did not activate deliver for broad build requests"
+fi
+
+rm -f "$TMP_BUILD_LOG"
+
+echo ""
+echo "==========================="
+total=$((pass + fail))
+echo "Results: ${pass}/${total} passed, ${fail} failed, ${skip} skipped"
+[[ "$fail" -gt 0 ]] && exit 1
+exit 0

package/evals/acceptance/test_kiro_harness.sh

@@ -0,0 +1,128 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+source "$ROOT_DIR/evals/lib/node.sh"
+TMP_HOME=""
+TMP_WORK=""
+TMP_TELEMETRY=""
+pass=0
+fail=0
+skip=0
+
+cleanup() {
+  [[ -n "$TMP_HOME" ]] && rm -rf "$TMP_HOME"
+  [[ -n "$TMP_WORK" ]] && rm -rf "$TMP_WORK"
+  [[ -n "$TMP_TELEMETRY" ]] && rm -rf "$TMP_TELEMETRY"
+}
+trap cleanup EXIT
+
+_pass() { echo "  ✓ $1"; pass=$((pass + 1)); }
+_fail() { echo "  ✗ $1"; fail=$((fail + 1)); }
+_skip() { echo "  ○ $1"; skip=$((skip + 1)); }
+strip_ansi() {
+  perl -pe 's/\e\[[0-9;?]*[ -\/]*[@-~]//g; s/\e\(B//g'
+}
+
+echo "=== Harness Acceptance: Kiro ==="
+echo ""
+
+if ! command -v kiro-cli >/dev/null 2>&1; then
+  _skip "kiro-cli not installed"
+  echo ""
+  echo "Results: ${pass}/$((pass + fail)) passed, ${fail} failed, ${skip} skipped"
+  exit 0
+fi
+
+cd "$ROOT_DIR"
+flow_agents_node scripts/build-universal-bundles.js >/dev/null
+
+TMP_HOME="$(mktemp -d /tmp/kiro-acceptance-home.XXXXXX)"
+TMP_WORK="$(mktemp -d /tmp/kiro-acceptance-work.XXXXXX)"
+TMP_TELEMETRY="$(mktemp -d /tmp/kiro-acceptance-telemetry.XXXXXX)"
+bash dist/kiro/install.sh "$TMP_HOME" >/dev/null
+mkdir -p "$TMP_WORK/.kiro"
+ln -s "$TMP_HOME/agents" "$TMP_WORK/.kiro/agents"
+
+echo "--- Agent List ---"
+list_output="$(cd "$TMP_WORK" && kiro-cli agent list 2>&1 || true)"
+if echo "$list_output" | grep -q "dev[[:space:]]\+Workspace"; then
+  _pass "workspace agent list includes dev"
+else
+  _fail "workspace agent list did not include dev"
+fi
+
+echo ""
+echo "--- Chat Smoke ---"
+chat_output="$(cd "$TMP_WORK" && kiro-cli chat --agent dev --no-interactive "Reply with READY only." 2>&1 || true)"
+if echo "$chat_output" | grep -q "READY"; then
+  _pass "dev agent replied to chat smoke prompt"
+else
+  _fail "dev agent did not reply READY"
+fi
+
+echo ""
+echo "--- Explore Behavior ---"
+explore_output="$(cd "$TMP_WORK" && TELEMETRY_ENABLED=true TELEMETRY_DATA_DIR="$TMP_TELEMETRY" TELEMETRY_SESSION_DIR="$TMP_TELEMETRY/sessions" TELEMETRY_CHANNELS=full TELEMETRY_CHANNEL_FULL_LOG_FILE="$TMP_TELEMETRY/full.jsonl" node - <<'NODE'
+const { spawnSync } = require("node:child_process");
+const result = spawnSync("kiro-cli", [
+  "chat",
+  "--agent",
+  "dev",
+  "--no-interactive",
+  "--trust-all-tools",
+  "Explore the codebase and explain what it does.",
+], { encoding: "utf8", timeout: 30000 });
+process.stdout.write(result.stdout || "");
+process.stdout.write(result.stderr || "");
+NODE
+)"
+explore_clean="$(printf '%s' "$explore_output" | strip_ansi)"
+if echo "$explore_clean" | grep -q "Activating skill: explore"; then
+  _pass "dev activates the explore skill on a plain explore prompt"
+else
+  _fail "dev did not activate the explore skill on a plain explore prompt"
+fi
+
+if echo "$explore_clean" | grep -q "Tool validation failed"; then
+  _fail "explore workflow exceeded harness delegation limits"
+else
+  _pass "explore workflow stayed within harness delegation limits"
+fi
+
+if [[ -f "$TMP_TELEMETRY/full.jsonl" ]] && rg -q '"event_type":"agent.delegate"' "$TMP_TELEMETRY/full.jsonl"; then
+  _pass "telemetry confirms delegated explore execution"
+else
+  _fail "telemetry did not confirm delegated explore execution"
+fi
+
+echo ""
+echo "--- Strict Stop Gate ---"
+mkdir -p "$TMP_WORK/.flow-agents/live-stop"
+cat > "$TMP_WORK/.flow-agents/live-stop/live-stop--deliver.md" <<'MARKDOWN'
+# Live Stop Gate
+
+status: executing
+type: deliver
+
+## Plan
+
+This delivery artifact is intentionally incomplete so the strict stop hook must surface Goal Fit guidance.
+MARKDOWN
+
+stop_output="$(cd "$TMP_WORK" && FLOW_AGENTS_GOAL_FIT_STRICT=true kiro-cli chat --agent dev --no-interactive "Reply with READY only." 2>&1 || true)"
+stop_clean="$(printf '%s' "$stop_output" | strip_ansi)"
+if echo "$stop_clean" | grep -q 'stop "node .*stop:goal-fit stop-goal-fit.js standard,strict" failed with exit code: 2' \
+  && echo "$stop_clean" | grep -q '\[Hook\] Goal Fit warning:' \
+  && echo "$stop_clean" | grep -q 'live-stop--deliver.md is still status:executing'; then
+  _pass "strict Goal Fit stop hook surfaces live Kiro stop gate"
+else
+  _fail "strict Goal Fit stop hook did not surface live Kiro stop gate"
+fi
+
+echo ""
+echo "==========================="
+total=$((pass + fail))
+echo "Results: ${pass}/${total} passed, ${fail} failed, ${skip} skipped"
+[[ "$fail" -gt 0 ]] && exit 1
+exit 0