@kontourai/flow-agents 3.12.1 → 4.0.0

package/dist/claude-code/evals/integration/test_telemetry_tool_outcome.sh

@@ -0,0 +1,279 @@
+#!/usr/bin/env bash
+# test_telemetry_tool_outcome.sh — Layer 2: per-tool-result enrichment
+# (kontourai/flow-agents#580, split from #568)
+#
+# Proves that tool.result (postToolUse/PostToolUse/PostToolUseFailure) telemetry
+# events carry three new fields on their .tool object:
+#   - .tool.duration_ms — non-negative wall-clock ms between this tool's invoke
+#     and its result (correlated per tool call), or null when the matching start
+#     record is absent (never a fabricated 0 / stale value).
+#   - .tool.outcome — a deterministic tri-state pass|fail|ambiguous, a jq port
+#     of scripts/hooks/evidence-capture.js observeResult (never from stdout text).
+#   - .tool.status — the host exit code int when present, else null.
+#
+# The highest-value assertion is the DRIFT-GUARD (case j): a shared fixture
+# battery is fed through BOTH the emitter jq path AND node observeResult, and
+# their {exitCode, observedResult} verdicts must be identical — so the hermetic
+# jq re-implementation can never silently diverge from the canonical source.
+#
+# Uses the same TELEMETRY_DIR resolution and FLOW_AGENTS_TELEMETRY_FOREGROUND
+# convention as test_telemetry_tool_usage.sh so assertions never race a
+# backgrounded subshell.
+set -uo pipefail
+
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+if [[ -d "$ROOT_DIR/context/scripts/telemetry" ]]; then
+  TELEMETRY_DIR="$ROOT_DIR/context/scripts/telemetry"
+else
+  TELEMETRY_DIR="<bundle-root>/context/scripts/telemetry"
+fi
+TELEMETRY_SH="${TELEMETRY_DIR}/telemetry.sh"
+SOURCE_TELEMETRY_SH="$ROOT_DIR/scripts/telemetry/telemetry.sh"
+FIXTURE_TRANSCRIPT="$ROOT_DIR/evals/fixtures/telemetry/usage-transcript-sample.jsonl"
+PRICING_FILE="$ROOT_DIR/scripts/telemetry/pricing.json"
+EVIDENCE_CAPTURE="$ROOT_DIR/scripts/hooks/evidence-capture.js"
+
+TMPDIR_EVAL=$(mktemp -d /tmp/eval-telemetry-tool-outcome.XXXXXX)
+TMPLOG="${TMPDIR_EVAL}/test-output.jsonl"
+
+pass=0; fail=0
+_pass() { echo "  ✓ $1"; pass=$((pass + 1)); }
+_fail() { echo "  ✗ $1"; fail=$((fail + 1)); }
+
+echo "=== Layer 2: Telemetry Tool-Result Enrichment (#580: duration/outcome/status) ==="
+echo ""
+
+if [[ ! -f "$TELEMETRY_SH" ]]; then
+  _fail "telemetry.sh not found at $TELEMETRY_SH"
+  rm -rf "$TMPDIR_EVAL"; exit 1
+fi
+if [[ ! -f "$EVIDENCE_CAPTURE" ]]; then
+  _fail "evidence-capture.js not found at $EVIDENCE_CAPTURE (needed for drift guard)"
+  rm -rf "$TMPDIR_EVAL"; exit 1
+fi
+
+# Run one hook event against telemetry.sh (foreground) and return the resulting
+# event object (jq-compact, one line) from the full-channel log.
+_run_tool_event() {
+  local hook_type="$1" input="$2"; shift 2
+  local extra_env=("$@")
+  local common_env=(
+    HOME="${TMPDIR_EVAL}/home"
+    TELEMETRY_ENABLED=true
+    TELEMETRY_CHANNELS=full
+    TELEMETRY_CHANNEL_FULL_LOG_FILE="$TMPLOG"
+    FLOW_AGENTS_TELEMETRY_FOREGROUND=true
+    TELEMETRY_CONFIG_FILE="$TMPDIR_EVAL/telemetry.conf"
+    TELEMETRY_DATA_DIR="$TMPDIR_EVAL"
+    TELEMETRY_SESSION_DIR="$TMPDIR_EVAL/sessions"
+    TELEMETRY_USAGE_TRACKING=true
+  )
+  mkdir -p "${TMPDIR_EVAL}/home" "${TMPDIR_EVAL}/sessions"
+
+  local before_lines
+  touch "$TMPLOG"
+  before_lines=$(wc -l < "$TMPLOG" | tr -d ' ')
+
+  echo "$input" | env "${common_env[@]}" ${extra_env[@]+"${extra_env[@]}"} bash "${RUN_TELEMETRY_SH:-$TELEMETRY_SH}" "$hook_type" dev 2>/dev/null
+
+  tail -n +"$((before_lines + 1))" "$TMPLOG" 2>/dev/null | tail -1
+}
+
+# --- (a) invoke -> result: duration_ms is a non-negative number ---------------
+echo "--- (a) invoke->result duration_ms is a non-negative number ---"
+inv_a=$(jq -nc '{session_id:"dur-a",hook_event_name:"PreToolUse",tool_use_id:"tu-a",tool_name:"Bash",tool_input:{command:"echo hi"}}')
+_run_tool_event "preToolUse" "$inv_a" >/dev/null
+res_a=$(jq -nc '{session_id:"dur-a",hook_event_name:"PostToolUse",tool_use_id:"tu-a",tool_name:"Bash",tool_input:{command:"echo hi"},tool_response:{exitCode:0}}')
+out_a=$(_run_tool_event "postToolUse" "$res_a")
+dur_a=$(echo "$out_a" | jq -r '.tool.duration_ms')
+if [[ "$dur_a" =~ ^[0-9]+$ ]]; then
+  _pass "duration_ms is a non-negative integer ($dur_a) when the matching invoke was recorded"
+else
+  _fail "expected numeric duration_ms, got '$dur_a'"
+fi
+
+# --- (b) result with NO prior invoke -> duration_ms is exactly null -----------
+echo ""
+echo "--- (b) result with no prior invoke -> duration_ms:null (never 0/stale) ---"
+res_b=$(jq -nc '{session_id:"dur-b",hook_event_name:"PostToolUse",tool_use_id:"tu-b-orphan",tool_name:"Bash",tool_input:{command:"echo hi"},tool_response:{exitCode:0}}')
+out_b=$(_run_tool_event "postToolUse" "$res_b")
+dur_b=$(echo "$out_b" | jq -r '.tool.duration_ms')
+[[ "$dur_b" == "null" ]] && _pass "duration_ms is exactly null (not 0) when no invoke start record exists" || _fail "expected duration_ms=null, got '$dur_b'"
+
+# --- (c) exit 0 -> outcome:pass status:0 -------------------------------------
+echo ""
+echo "--- (c) host exit code 0 -> outcome:pass status:0 ---"
+res_c=$(jq -nc '{session_id:"oc-c",hook_event_name:"PostToolUse",tool_use_id:"tu-c",tool_name:"Bash",tool_input:{command:"echo hi"},tool_response:{exitCode:0}}')
+out_c=$(_run_tool_event "postToolUse" "$res_c")
+oc_c=$(echo "$out_c" | jq -r '.tool.outcome'); st_c=$(echo "$out_c" | jq -r '.tool.status')
+[[ "$oc_c" == "pass" ]] && _pass "outcome=pass on exit 0" || _fail "expected outcome=pass, got '$oc_c'"
+[[ "$st_c" == "0" ]] && _pass "status=0 on exit 0" || _fail "expected status=0, got '$st_c'"
+
+# --- (d) non-zero exit -> outcome:fail status:<code> -------------------------
+echo ""
+echo "--- (d) non-zero host exit code -> outcome:fail status:<code> ---"
+res_d=$(jq -nc '{session_id:"oc-d",hook_event_name:"PostToolUse",tool_use_id:"tu-d",tool_name:"Bash",tool_input:{command:"false"},tool_response:{exitCode:2}}')
+out_d=$(_run_tool_event "postToolUse" "$res_d")
+oc_d=$(echo "$out_d" | jq -r '.tool.outcome'); st_d=$(echo "$out_d" | jq -r '.tool.status')
+[[ "$oc_d" == "fail" ]] && _pass "outcome=fail on exit 2" || _fail "expected outcome=fail, got '$oc_d'"
+[[ "$st_d" == "2" ]] && _pass "status=2 carries the real exit code" || _fail "expected status=2, got '$st_d'"
+
+# --- (e) failure signal, NO clean code -> outcome:fail status:null -----------
+echo ""
+echo "--- (e) is_error:true / success:false with no exit code -> fail, status:null ---"
+res_e1=$(jq -nc '{session_id:"oc-e1",hook_event_name:"PostToolUse",tool_use_id:"tu-e1",tool_name:"Bash",tool_input:{command:"x"},tool_response:{is_error:true,output:"oops"}}')
+out_e1=$(_run_tool_event "postToolUse" "$res_e1")
+oc_e1=$(echo "$out_e1" | jq -r '.tool.outcome'); st_e1=$(echo "$out_e1" | jq -r '.tool.status')
+[[ "$oc_e1" == "fail" ]] && _pass "outcome=fail on is_error:true (no code)" || _fail "expected fail, got '$oc_e1'"
+[[ "$st_e1" == "null" ]] && _pass "status=null when no clean exit code is present (never fabricated)" || _fail "expected status=null, got '$st_e1'"
+res_e2=$(jq -nc '{session_id:"oc-e2",hook_event_name:"PostToolUse",tool_use_id:"tu-e2",tool_name:"Bash",tool_input:{command:"x"},tool_response:{success:false}}')
+out_e2=$(_run_tool_event "postToolUse" "$res_e2")
+oc_e2=$(echo "$out_e2" | jq -r '.tool.outcome')
+[[ "$oc_e2" == "fail" ]] && _pass "outcome=fail on success:false (no code)" || _fail "expected fail, got '$oc_e2'"
+
+# --- (f) no signal at all -> outcome:ambiguous status:null -------------------
+echo ""
+echo "--- (f) no exit code and no failure signal -> ambiguous, status:null ---"
+res_f=$(jq -nc '{session_id:"oc-f",hook_event_name:"PostToolUse",tool_use_id:"tu-f",tool_name:"Read",tool_input:{file:"a"},tool_response:{output:"contents"}}')
+out_f=$(_run_tool_event "postToolUse" "$res_f")
+oc_f=$(echo "$out_f" | jq -r '.tool.outcome'); st_f=$(echo "$out_f" | jq -r '.tool.status')
+[[ "$oc_f" == "ambiguous" ]] && _pass "outcome=ambiguous when there is no positive success evidence (never pass)" || _fail "expected ambiguous, got '$oc_f'"
+[[ "$st_f" == "null" ]] && _pass "status=null in the no-signal case" || _fail "expected status=null, got '$st_f'"
+
+# --- (g) PostToolUseFailure event folds to fail -----------------------------
+echo ""
+echo "--- (g) PostToolUseFailure event folds to outcome:fail ---"
+res_g=$(jq -nc '{session_id:"oc-g",hook_event_name:"PostToolUseFailure",tool_use_id:"tu-g",tool_name:"Bash",tool_input:{command:"x"},tool_response:{output:"partial"}}')
+out_g=$(_run_tool_event "PostToolUseFailure" "$res_g")
+et_g=$(echo "$out_g" | jq -r '.event_type'); oc_g=$(echo "$out_g" | jq -r '.tool.outcome')
+[[ "$et_g" == "tool.result" ]] && _pass "PostToolUseFailure maps to tool.result" || _fail "expected tool.result, got '$et_g'"
+[[ "$oc_g" == "fail" ]] && _pass "PostToolUseFailure folds to outcome=fail regardless of payload signal" || _fail "expected fail, got '$oc_g'"
+
+# --- (h) invoke and permission_request carry NO duration/outcome/status ------
+echo ""
+echo "--- (h) tool.invoke and tool.permission_request carry no result meta ---"
+inv_h=$(jq -nc '{session_id:"h-inv",hook_event_name:"PreToolUse",tool_use_id:"tu-h",tool_name:"Read",tool_input:{file:"a"}}')
+out_h_inv=$(_run_tool_event "preToolUse" "$inv_h")
+has_dur_inv=$(echo "$out_h_inv" | jq -r '.tool | has("duration_ms")')
+has_oc_inv=$(echo "$out_h_inv" | jq -r '.tool | has("outcome")')
+[[ "$has_dur_inv" == "false" && "$has_oc_inv" == "false" ]] && _pass "tool.invoke has no duration_ms/outcome/status (invoke has no result yet)" || _fail "tool.invoke unexpectedly carries result meta (duration=$has_dur_inv outcome=$has_oc_inv)"
+perm_h=$(jq -nc '{session_id:"h-perm",hook_event_name:"PermissionRequest",tool_use_id:"tu-hp",tool_name:"Bash",tool_input:{command:"x",description:"run"}}')
+out_h_perm=$(_run_tool_event "permissionRequest" "$perm_h")
+has_oc_perm=$(echo "$out_h_perm" | jq -r '.tool | has("outcome")')
+[[ "$has_oc_perm" == "false" ]] && _pass "tool.permission_request carries no outcome (not a tool result)" || _fail "tool.permission_request unexpectedly carries an outcome"
+
+# --- (i) slice-1 .usage still present and correct (no regression) ------------
+echo ""
+echo "--- (i) slice-1 .usage enrichment still present on tool.result (no regression) ---"
+res_i=$(jq -nc --arg tp "$FIXTURE_TRANSCRIPT" '{session_id:"i-usage",transcript_path:$tp,hook_event_name:"PostToolUse",tool_use_id:"tu-i",tool_name:"Bash",tool_input:{command:"echo hi"},tool_response:{exitCode:0}}')
+out_i=$(_run_tool_event "postToolUse" "$res_i" TELEMETRY_PRICING_FILE="$PRICING_FILE")
+model_i=$(echo "$out_i" | jq -r '.usage.model // empty'); oc_i=$(echo "$out_i" | jq -r '.tool.outcome')
+[[ "$model_i" == "claude-fable-5" ]] && _pass "usage.model is still the last-turn model (claude-fable-5) alongside the new fields" || _fail "slice-1 regression: expected usage.model=claude-fable-5, got '$model_i'"
+[[ "$oc_i" == "pass" ]] && _pass "new outcome field coexists with slice-1 .usage on the same record" || _fail "expected outcome=pass, got '$oc_i'"
+
+# --- (j) DRIFT-GUARD: emitter jq outcome == node observeResult across battery -
+echo ""
+echo "--- (j) DRIFT-GUARD: emitter jq outcome/status == node observeResult (shared battery) ---"
+# Shared fixture battery: [name, tool_input_json_or_null, tool_response_json_or_null, top_level_extra_json]
+# Each row is fed BOTH through the full emitter (postToolUse) and node observeResult.
+battery_json=$(cat <<'JSON'
+[
+  {"name":"exit0","ti":{"command":"echo hi"},"tr":{"exitCode":0},"extra":{}},
+  {"name":"exit2","ti":{"command":"false"},"tr":{"exitCode":2},"extra":{}},
+  {"name":"str_code_0","ti":{"command":"x"},"tr":{"exit_code":"0"},"extra":{}},
+  {"name":"is_error","ti":{"command":"x"},"tr":{"is_error":true},"extra":{}},
+  {"name":"success_false","ti":{"command":"x"},"tr":{"success":false},"extra":{}},
+  {"name":"error_top","ti":{"command":"x"},"tr":{"output":"ok"},"extra":{"error":"boom"}},
+  {"name":"error_top_array","ti":{"command":"x"},"tr":{"output":"ok"},"extra":{"error":["boom"]}},
+  {"name":"no_signal","ti":{"file":"a"},"tr":{"output":"contents"},"extra":{}},
+  {"name":"grep_absence","ti":{"command":"grep -r foo src"},"tr":{"exitCode":1},"extra":{}},
+  {"name":"grep_or_true","ti":{"command":"grep -r foo src || true"},"tr":{"exitCode":1},"extra":{}},
+  {"name":"grep_negated","ti":{"command":"! grep foo x"},"tr":{"exitCode":1},"extra":{}},
+  {"name":"diff_absence","ti":{"command":"diff a b"},"tr":{"exitCode":1},"extra":{}},
+  {"name":"diff_err2","ti":{"command":"diff a b"},"tr":{"exitCode":2},"extra":{}},
+  {"name":"bashlc_grep","ti":{"command":"bash -lc 'grep foo x'"},"tr":{"exitCode":1},"extra":{}},
+  {"name":"keyval_grep","ti":{"command":"FOO=bar grep foo x"},"tr":{"exitCode":1},"extra":{}},
+  {"name":"piped_grep","ti":{"command":"grep foo x | wc -l"},"tr":{"exitCode":1},"extra":{}},
+  {"name":"returnCode0","ti":{"command":"x"},"tr":{"returnCode":0},"extra":{}},
+  {"name":"top_status1","ti":{"command":"x"},"tr":{"output":"noise"},"extra":{"status":1}},
+  {"name":"stderr_nostdout","ti":{"command":"x"},"tr":{"stderr":"boom"},"extra":{}},
+  {"name":"stderr_withstdout","ti":{"command":"x"},"tr":{"stderr":"boom","stdout":"ok"},"extra":{}},
+  {"name":"plain_string_out","ti":{"command":"x"},"tr":"just a string","extra":{}}
+]
+JSON
+)
+battery_count=$(echo "$battery_json" | jq 'length')
+drift_ok=0; drift_bad=0
+while IFS= read -r fixture; do
+  bname=$(echo "$fixture" | jq -r '.name')
+  # Build the hook payload (unique tool_use_id so no cross-case duration leak).
+  payload=$(echo "$fixture" | jq -c '{session_id:("drift-"+.name),hook_event_name:"PostToolUse",tool_use_id:("tu-drift-"+.name),tool_name:"Bash",tool_input:.ti,tool_response:.tr} + .extra')
+  # Emitter path.
+  emit_out=$(_run_tool_event "postToolUse" "$payload")
+  emit_oc=$(echo "$emit_out" | jq -r '.tool.outcome')
+  emit_st=$(echo "$emit_out" | jq -c '.tool.status')
+  # Node canonical path — mirror observeResult's exact call shape at the run() site.
+  node_json=$(EVCAP="$EVIDENCE_CAPTURE" PAYLOAD="$payload" node -e '
+    const {observeResult}=require(process.env.EVCAP);
+    const i=JSON.parse(process.env.PAYLOAD);
+    const r=observeResult({tool_response:i.tool_response,tool_output:i.tool_output,error:i.error,exitCode:i.exitCode,exit_code:i.exit_code,status:i.status,code:i.code,command:(i.tool_input&&i.tool_input.command)});
+    process.stdout.write(JSON.stringify({outcome:r.observedResult,status:(r.exitCode===null?null:r.exitCode)}));
+  ' 2>/dev/null)
+  node_oc=$(echo "$node_json" | jq -r '.outcome')
+  node_st=$(echo "$node_json" | jq -c '.status')
+  if [[ "$emit_oc" == "$node_oc" && "$emit_st" == "$node_st" ]]; then
+    drift_ok=$((drift_ok+1))
+  else
+    drift_bad=$((drift_bad+1))
+    _fail "DRIFT [$bname]: emitter={outcome:$emit_oc,status:$emit_st} != node={outcome:$node_oc,status:$node_st}"
+  fi
+done < <(echo "$battery_json" | jq -c '.[]')
+if [[ "$drift_bad" -eq 0 ]]; then
+  _pass "emitter jq outcome/status matches node observeResult across all $battery_count battery fixtures (no drift)"
+else
+  _fail "$drift_bad/$battery_count battery fixtures drifted between the emitter jq and node observeResult"
+fi
+
+# --- (k) Codex-only host-banner exit-code resolution -------------------------
+echo ""
+echo "--- (k) codex runtime: exit code resolved from rollout banner (bounded node use) ---"
+ROLLOUT="$TMPDIR_EVAL/rollout.jsonl"
+jq -nc '{timestamp:"2026-07-06T00:00:00Z",type:"response_item",payload:{type:"function_call_output",call_id:"call-k",output:"Process exited with code 3\nOriginal token count: 25\nOutput:\nProcess exited with code 0\n"}}' > "$ROLLOUT"
+res_k=$(jq -nc --arg rp "$ROLLOUT" '{session_id:"codex-k",transcript_path:$rp,hook_event_name:"PostToolUse",call_id:"call-k",tool_name:"Bash",tool_input:{command:"some cmd"},tool_response:{output:"no structured code here"}}')
+RUN_TELEMETRY_SH="$SOURCE_TELEMETRY_SH" out_k=$(RUN_TELEMETRY_SH="$SOURCE_TELEMETRY_SH" _run_tool_event "postToolUse" "$res_k" FLOW_AGENTS_TELEMETRY_RUNTIME=codex)
+st_k=$(echo "$out_k" | jq -r '.tool.status'); oc_k=$(echo "$out_k" | jq -r '.tool.outcome')
+[[ "$st_k" == "3" ]] && _pass "codex status resolved to 3 from the preamble host banner (forged post-Output stdout ignored)" || _fail "expected codex status=3, got '$st_k'"
+[[ "$oc_k" == "fail" ]] && _pass "codex outcome=fail derived from the resolved non-zero banner code" || _fail "expected codex outcome=fail, got '$oc_k'"
+# Codex, unreadable rollout -> honest ambiguous/null degrade (no fabrication).
+res_k2=$(jq -nc '{session_id:"codex-k2",transcript_path:"/nonexistent/rollout.jsonl",hook_event_name:"PostToolUse",call_id:"call-x",tool_name:"Bash",tool_input:{command:"some cmd"},tool_response:{output:"no code"}}')
+out_k2=$(RUN_TELEMETRY_SH="$SOURCE_TELEMETRY_SH" _run_tool_event "postToolUse" "$res_k2" FLOW_AGENTS_TELEMETRY_RUNTIME=codex)
+st_k2=$(echo "$out_k2" | jq -r '.tool.status'); oc_k2=$(echo "$out_k2" | jq -r '.tool.outcome')
+[[ "$st_k2" == "null" && "$oc_k2" == "ambiguous" ]] && _pass "codex with an unreadable rollout degrades to ambiguous/null (never a guess)" || _fail "expected ambiguous/null, got outcome=$oc_k2 status=$st_k2"
+
+# --- (privacy) start record stores only a timestamp, no args/output ----------
+echo ""
+echo "--- (privacy) tool start record holds a bare timestamp, no command/args ---"
+inv_p=$(jq -nc '{session_id:"priv-p",hook_event_name:"PreToolUse",tool_use_id:"tu-priv",tool_name:"Bash",tool_input:{command:"echo SENSITIVE_SECRET_VALUE"}}')
+_run_tool_event "preToolUse" "$inv_p" >/dev/null
+start_file=$(ls "${TMPDIR_EVAL}/sessions/"toolstart-* 2>/dev/null | head -1)
+if [[ -n "$start_file" ]]; then
+  contents=$(cat "$start_file")
+  if [[ "$contents" =~ ^[0-9]+$ ]]; then
+    _pass "start record contains only a numeric timestamp ($contents)"
+  else
+    _fail "start record is not a bare timestamp: '$contents'"
+  fi
+  if grep -q "SENSITIVE_SECRET_VALUE" "$start_file" 2>/dev/null; then
+    _fail "PRIVACY LEAK: start record contains the command args"
+  else
+    _pass "start record does not leak command args"
+  fi
+else
+  _fail "no tool start record written on preToolUse"
+fi
+
+rm -rf "$TMPDIR_EVAL"
+
+echo ""
+echo "Telemetry tool-result enrichment: $pass passed, $fail failed"
+[ "$fail" -eq 0 ]

package/dist/claude-code/evals/integration/test_workflow_sidecar_writer.sh

@@ -3573,6 +3573,61 @@ else
   _fail "compose: record-gate-claim (selected-work) failed: $(cat "$TMPDIR_EVAL/compose-gate1.out" "$TMPDIR_EVAL/compose-gate1.err")"
 fi
 
+# ─── #634: writer-observed execution — die-before-append + append-on-success ────────────────
+# A rejected expected-status mismatch must leave command-log.jsonl untouched (no writer
+# observation may exist for a claim that was refused); a successful gate claim with
+# --command must append a chain-linked canonical-writer-execution observation.
+CMDLOG_634="$COMPOSE_DIR/command-log.jsonl"
+CMDLOG_634_BEFORE=""
+[[ -f "$CMDLOG_634" ]] && CMDLOG_634_BEFORE="$(shasum -a 256 "$CMDLOG_634" | cut -d' ' -f1)"
+if flow_agents_node "$WRITER" record-gate-claim "$COMPOSE_DIR" \
+  --actor compose-actor \
+  --status pass \
+  --summary "634 mismatch: claiming pass while the command fails must die before any append" \
+  --command "test -f .kontourai/flow-agents/compose-270/definitely-not-a-real-file-634" \
+  --evidence-ref-json '{"kind":"artifact","file":".kontourai/flow-agents/compose-270/compose-270--pull-work.md","summary":"634 mismatch fixture."}' \
+  --timestamp "2026-07-14T15:00:00Z" >"$TMPDIR_EVAL/wo-mismatch.out" 2>"$TMPDIR_EVAL/wo-mismatch.err"; then
+  _fail "#634: record-gate-claim accepted a passing claim for a failing command"
+else
+  if grep -q "failed (exit" "$TMPDIR_EVAL/wo-mismatch.err"; then
+    _pass "#634: expected-status mismatch dies with the exit-code diagnostic"
+  else
+    _fail "#634: mismatch died with an unexpected diagnostic: $(cat "$TMPDIR_EVAL/wo-mismatch.err")"
+  fi
+fi
+CMDLOG_634_AFTER=""
+[[ -f "$CMDLOG_634" ]] && CMDLOG_634_AFTER="$(shasum -a 256 "$CMDLOG_634" | cut -d' ' -f1)"
+if [[ "$CMDLOG_634_BEFORE" == "$CMDLOG_634_AFTER" ]]; then
+  _pass "#634: rejected mismatch left command-log.jsonl byte-identical (die-before-append)"
+else
+  _fail "#634: rejected mismatch modified command-log.jsonl"
+fi
+
+if flow_agents_node "$WRITER" record-gate-claim "$COMPOSE_DIR" \
+  --actor compose-actor \
+  --status pass \
+  --summary "634 success: writer observation must append on a genuinely passing command" \
+  --command "test -f .kontourai/flow-agents/compose-270/compose-270--pull-work.md" \
+  --evidence-ref-json '{"kind":"artifact","file":".kontourai/flow-agents/compose-270/compose-270--pull-work.md","summary":"634 success fixture."}' \
+  --timestamp "2026-07-14T15:01:00Z" >"$TMPDIR_EVAL/wo-success.out" 2>"$TMPDIR_EVAL/wo-success.err"; then
+  if [[ -f "$CMDLOG_634" ]] && node -e "
+    const fs = require('fs');
+    const entries = fs.readFileSync(process.argv[1], 'utf8').split('\n').filter(Boolean).map(JSON.parse);
+    const writer = entries.filter((e) => e.source === 'canonical-writer-execution');
+    if (writer.length === 0) process.exit(1);
+    const last = writer[writer.length - 1];
+    if (last.observedResult !== 'pass' || last.exitCode !== 0) process.exit(1);
+    if (!last._chain || typeof last._chain.hash !== 'string') process.exit(1);
+  " "$CMDLOG_634"; then
+    _pass "#634: successful gate claim appended a chain-linked canonical-writer-execution pass"
+  else
+    _fail "#634: no chain-linked writer observation found after a successful gate claim"
+  fi
+else
+  _fail "#634: success-path record-gate-claim failed: $(cat "$TMPDIR_EVAL/wo-success.err")"
+fi
+
+
 if node --input-type=module <<NODEOF 2>"$TMPDIR_EVAL/compose-gate1-assert.err"
 import { readFileSync } from 'node:fs';
 const bundle = JSON.parse(readFileSync('${COMPOSE_DIR}/trust.bundle', 'utf8'));
@@ -3863,6 +3918,14 @@ const bundle = JSON.parse(fs.readFileSync(bundleFile, 'utf8'));
 const gate = bundle.claims.find((claim) => claim.metadata?.gate_claim?.expectation_id === 'tests-evidence');
 const review = bundle.claims.find((claim) => claim.metadata?.origin === 'critique');
 if (!gate || gate.metadata?.observed_commands?.length !== 2) process.exit(2);
+const criteria = bundle.claims.filter((claim) => claim.metadata?.origin === 'acceptance');
+if (criteria.length !== 2 || criteria.some((claim) => claim.value !== 'pass' || claim.status !== 'verified')) process.exit(4);
+for (const claim of criteria) {
+  const evidence = bundle.evidence.filter((item) => item.claimId === claim.id);
+  const event = bundle.events.find((item) => item.claimId === claim.id && item.status === 'verified');
+  if (evidence.length !== 1 || evidence[0].evidenceType !== 'test_output' || evidence[0].passing !== true) process.exit(5);
+  if (!event || event.evidenceIds.length !== 1 || event.evidenceIds[0] !== evidence[0].id) process.exit(6);
+}
 const snapshot = review?.metadata?.review_target?.workspace_snapshot;
 if (!snapshot || snapshot.version !== 1 || snapshot.algorithm !== 'sha256' || typeof snapshot.digest !== 'string') process.exit(3);
 NODE
@@ -3875,6 +3938,54 @@ else
   _fail "tests-evidence multi-command write failed: $(cat "$TMPDIR_EVAL/multi-pass.out" "$TMPDIR_EVAL/multi-pass.err")"
 fi
 
+if flow_agents_node "$WRITER" record-evidence "$MULTI_DIR" \
+  --verdict not_verified \
+  --check-json '{"id":"acceptance-rebuild-probe","kind":"external","status":"not_verified","summary":"Trigger a valid unrelated bundle rebuild."}' \
+  --timestamp "2026-07-11T11:00:35Z" >"$TMPDIR_EVAL/multi-rebuild.out" 2>"$TMPDIR_EVAL/multi-rebuild.err" \
+  && node - "$MULTI_DIR/trust.bundle" <<'NODE'
+const fs = require('node:fs');
+const bundle = JSON.parse(fs.readFileSync(process.argv[2], 'utf8'));
+const criteria = bundle.claims.filter((claim) => claim.metadata?.origin === 'acceptance');
+if (criteria.length !== 2 || criteria.some((claim) => claim.value !== 'pass' || claim.status !== 'verified')) process.exit(1);
+for (const claim of criteria) {
+  if (claim.createdAt !== '2026-07-11T11:00:30Z' || claim.updatedAt !== '2026-07-11T11:00:30Z') process.exit(2);
+  const event = bundle.events.find((item) => item.claimId === claim.id);
+  if (event?.createdAt !== '2026-07-11T11:00:30Z' || event?.verifiedAt !== '2026-07-11T11:00:30Z') process.exit(3);
+}
+NODE
+then
+  _pass "verified acceptance provenance survives an unrelated rebuild without timestamp laundering"
+else
+  _fail "unrelated rebuild regressed or refreshed verified acceptance provenance: $(cat "$TMPDIR_EVAL/multi-rebuild.out" "$TMPDIR_EVAL/multi-rebuild.err")"
+fi
+
+# A direct acceptance artifact edit cannot replace or erase the durable criterion contract on a
+# later, unrelated bundle write. This reproduces the issue #601 artifact-only erasure vector.
+node - "$MULTI_DIR/acceptance.json" <<'NODE'
+const fs = require('node:fs');
+const file = process.argv[2];
+const acceptance = JSON.parse(fs.readFileSync(file, 'utf8'));
+acceptance.criteria = [{ id: null, description: 'Attempted replacement', status: 'PASS' }];
+fs.writeFileSync(file, JSON.stringify(acceptance, null, 2) + '\n');
+NODE
+if flow_agents_node "$WRITER" record-evidence "$MULTI_DIR" \
+  --verdict not_verified \
+  --check-json '{"id":"acceptance-erasure-probe","kind":"external","status":"not_verified","summary":"Trigger an unrelated bundle rebuild."}' \
+  --timestamp "2026-07-11T11:00:40Z" >"$TMPDIR_EVAL/multi-erasure.out" 2>"$TMPDIR_EVAL/multi-erasure.err"; then
+  _fail "malformed acceptance artifact was accepted during an unrelated bundle rebuild"
+elif node - "$MULTI_DIR/trust.bundle" <<'NODE'
+const fs = require('node:fs');
+const bundle = JSON.parse(fs.readFileSync(process.argv[2], 'utf8'));
+const criteria = bundle.claims.filter((claim) => claim.metadata?.origin === 'acceptance');
+if (criteria.length !== 2) process.exit(1);
+if (criteria.some((claim) => claim.value !== 'pass' || claim.status !== 'verified')) process.exit(2);
+NODE
+then
+  _pass "malformed acceptance replacement fails closed without erasing durable criteria"
+else
+  _fail "rejected acceptance replacement still erased or regressed durable criteria: $(cat "$TMPDIR_EVAL/multi-erasure.out" "$TMPDIR_EVAL/multi-erasure.err")"
+fi
+
 # ─── AC4: compose-three-writer (five writer calls total: evidence x2, gate-claim x2, critique, ──
 # then record-learning) — all claim families present and correctly typed in the final bundle ──
 if flow_agents_node "$WRITER" record-learning "$COMPOSE_DIR" \
@@ -5036,8 +5147,6 @@ else
   _fail "mutation-test setup: could not locate the compiled build/src/cli/workflow-sidecar.js to mutate (ran 'npm run build' first?)"
 fi
 
-
-
 if [[ "$errors" -eq 0 ]]; then
   echo "Workflow sidecar writer integration passed."
   exit 0

package/dist/claude-code/evals/run.sh

@@ -177,6 +177,14 @@ run_integration() {
   echo ""
   bash "$EVAL_DIR/integration/test_telemetry_tool_usage.sh" || result=1
   echo ""
+  bash "$EVAL_DIR/integration/test_telemetry_tool_outcome.sh" || result=1
+  echo ""
+  bash "$EVAL_DIR/integration/test_telemetry_delegation.sh" || result=1
+  echo ""
+  bash "$EVAL_DIR/integration/test_telemetry_action_class.sh" || result=1
+  echo ""
+  bash "$EVAL_DIR/integration/test_telemetry_sanitize_usage.sh" || result=1
+  echo ""
   bash "$EVAL_DIR/integration/test_telemetry_doctor.sh" || result=1
   echo ""
   bash "$EVAL_DIR/integration/test_usage_feedback_outcomes.sh" || result=1
@@ -308,6 +316,10 @@ run_integration() {
   echo ""
   bash "$EVAL_DIR/integration/test_command_log_integrity.sh" || result=1
   echo ""
+  bash "$EVAL_DIR/integration/test_narrative_source_contract.sh" || result=1
+  echo ""
+  bash "$EVAL_DIR/integration/test_narrative_redaction_failclosed.sh" || result=1
+  echo ""
   bash "$EVAL_DIR/integration/test_gate_lockdown.sh" || result=1
   echo ""
   bash "$EVAL_DIR/integration/test_resolvefirststep_security.sh" || result=1

package/dist/claude-code/evals/static/test_library_exports.sh

@@ -41,6 +41,11 @@ if [[ -f "build/src/index.js" && -f "build/src/index.d.ts" ]]; then
 else
   fail "build is missing index.js or index.d.ts (run npm run build)"
 fi
+if grep -q 'GateActionArtifactBinding' build/src/index.d.ts && grep -q 'GateActionArtifactTarget' build/src/index.d.ts; then
+  pass "package root exports the public gate-action artifact binding and target types"
+else
+  fail "build/src/index.d.ts is missing a public gate-action artifact type"
+fi
 
 # 3. importing the library does not run the CLI, and the public API is present.
 # If importing executed the CLI it would call process.exit before our marker prints.

package/dist/claude-code/install.sh

@@ -72,7 +72,7 @@ if [[ ! -e "$DEST/CLAUDE.md" && ! -L "$DEST/CLAUDE.md" ]]; then
   rm -f "$instruction_tmp"
 fi
 if command -v node >/dev/null 2>&1; then
-  node "$DEST/scripts/install-merge.js" --config "$DEST/.claude/settings.json" --managed-hooks "$SRC/.claude/settings.json" --version "3.12.1" --install-record "$DEST/.flow-agents/install.json" --runtime "claude-code" || true
+  node "$DEST/scripts/install-merge.js" --config "$DEST/.claude/settings.json" --managed-hooks "$SRC/.claude/settings.json" --version "4.0.0" --install-record "$DEST/.flow-agents/install.json" --runtime "claude-code" || true
 fi
 if [[ ${#CONSOLE_CONFIG_ARGS[@]} -gt 0 || -n "${FLOW_AGENTS_TELEMETRY_SINK:-}" || -n "${FLOW_AGENTS_TELEMETRY_SINKS:-}" || -n "${FLOW_AGENTS_CONSOLE_URL:-}" || -n "${CONSOLE_TELEMETRY_URL:-}" || -n "${CONSOLE_URL:-}" || -n "${FLOW_AGENTS_CONSOLE_TOKEN_FILE:-}" || -n "${CONSOLE_TELEMETRY_TOKEN_FILE:-}" ]]; then
   bash "$DEST/scripts/telemetry/install-console-config.sh" "$DEST/scripts/telemetry/telemetry.conf" "${CONSOLE_CONFIG_ARGS[@]}"