@kontourai/flow-agents 3.3.0 → 3.4.0

package/evals/static/test_ci_integration_coverage.sh

@@ -0,0 +1,115 @@
+#!/usr/bin/env bash
+set -uo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+
+node - "$ROOT" <<'NODE'
+const fs = require('fs');
+const path = require('path');
+
+const root = process.argv[2];
+const rel = (...parts) => path.join(...parts).replace(/\\/g, '/');
+const read = (file) => fs.readFileSync(path.join(root, file), 'utf8');
+
+// Baseline-red in #297: stop-goal-fit fixture still emits no disputed-claim
+// gate hint. Do not wire until fixed by the owning gate/claim-lookup work.
+// Reason is duplicated in the value so this machine check can fail closed.
+const EXEMPTIONS = {
+  'evals/integration/test_claim_lookup.sh':
+    'Baseline-red in #297: stop-goal-fit fixture emits no disputed-claim gate hint.',
+
+  // Baseline-red in #297: goal-fit block streak fixture now exits 0 instead of
+  // the expected block/release sequence. This touches hook/gate behavior, which
+  // is explicitly out of scope for #297.
+  'evals/integration/test_goal_fit_escape_hatch.sh':
+    'Baseline-red in #297: goal-fit block streak fixture exits 0 instead of block/release sequence.',
+};
+
+function slugify(label) {
+  return label.toLowerCase().replace(/[^a-z0-9]+/g, '-').replace(/^-|-$/g, '');
+}
+
+function integrationTestsIn(text) {
+  return [...text.matchAll(/evals\/integration\/test_[A-Za-z0-9_-]+\.sh/g)].map((m) => m[0]);
+}
+
+const integrationDir = path.join(root, 'evals/integration');
+const allIntegration = fs
+  .readdirSync(integrationDir)
+  .filter((file) => /^test_.*\.sh$/.test(file))
+  .map((file) => rel('evals/integration', file))
+  .sort();
+
+const runBaseline = read('evals/ci/run-baseline.sh');
+const checks = new Map();
+for (const match of runBaseline.matchAll(/"([^"\n]+)\|([^"\n]+)"/g)) {
+  const [, label, command] = match;
+  checks.set(slugify(label), { label, command });
+}
+
+const ci = read('.github/workflows/ci.yml');
+const ciCheckIds = [
+  ...ci.matchAll(/run:\s*bash evals\/ci\/run-baseline\.sh --check ([a-z0-9-]+)/g),
+].map((match) => match[1]);
+
+const covered = new Map();
+const unknownChecks = [];
+
+for (const checkId of ciCheckIds) {
+  const check = checks.get(checkId);
+  if (!check) {
+    unknownChecks.push(checkId);
+    continue;
+  }
+
+  const directTests = integrationTestsIn(check.command);
+  for (const test of directTests) covered.set(test, check.label);
+
+  const suiteScripts = [...check.command.matchAll(/evals\/ci\/[A-Za-z0-9_-]+\.sh/g)].map((m) => m[0]);
+  for (const suiteScript of suiteScripts) {
+    const suitePath = path.join(root, suiteScript);
+    if (!fs.existsSync(suitePath)) {
+      throw new Error(`CI check '${check.label}' references missing suite script: ${suiteScript}`);
+    }
+    for (const test of integrationTestsIn(fs.readFileSync(suitePath, 'utf8'))) {
+      covered.set(test, `${check.label} (${suiteScript})`);
+    }
+  }
+}
+
+const problems = [];
+for (const checkId of unknownChecks) {
+  problems.push(`CI invokes unknown run-baseline check id: ${checkId}`);
+}
+
+for (const [test, reason] of Object.entries(EXEMPTIONS)) {
+  if (!allIntegration.includes(test)) {
+    problems.push(`Exemption references missing integration eval: ${test}`);
+  }
+  if (typeof reason !== 'string' || reason.trim().length < 12) {
+    problems.push(`Exemption for ${test} must include a concrete reason.`);
+  }
+  if (covered.has(test)) {
+    problems.push(`Exempted integration eval is also CI-covered; remove exemption: ${test}`);
+  }
+}
+
+for (const test of allIntegration) {
+  if (covered.has(test) || Object.hasOwn(EXEMPTIONS, test)) continue;
+  problems.push(`Integration eval is not covered by CI and has no reasoned exemption: ${test}`);
+}
+
+for (const test of covered.keys()) {
+  if (!allIntegration.includes(test)) {
+    problems.push(`CI references missing integration eval: ${test}`);
+  }
+}
+
+if (problems.length) {
+  console.error('CI integration coverage audit failed:');
+  for (const problem of problems.sort()) console.error(`  - ${problem}`);
+  process.exit(1);
+}
+
+console.log(`CI integration coverage audit passed: ${covered.size} covered, ${Object.keys(EXEMPTIONS).length} exempted.`);
+NODE

package/evals/static/test_declared_scope_forms_documented.sh

@@ -0,0 +1,114 @@
+#!/usr/bin/env bash
+# test_declared_scope_forms_documented.sh — AC8 (issue #303/#304 residue): delivery/README.md
+# documents all four DECLARED scope forms, the compound (space-separated AND) form, and the
+# author:/branch-prefix: combining requirement for identity exemptions.
+#
+# Sourced from scripts/ci/trust-reconcile.js's matchesScopeCondition()/matchesScope() (the
+# actual implemented semantics — string equality/prefix only, never RegExp) and ADR 0022 §2 +
+# the 2026-07-03 addendum. This is a documentation-coverage lock: if delivery/README.md's
+# scope-forms section is ever removed or one of the four forms silently drops out, this eval
+# fails loudly, naming what's missing.
+#
+# Deterministic, no model spend, no fixtures.
+# Usage: bash evals/static/test_declared_scope_forms_documented.sh
+
+set -uo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+DOC="$ROOT/delivery/README.md"
+
+errors=0
+pass() { echo "  ✓ $1"; }
+fail() { echo "  ✗ $1"; errors=$((errors + 1)); }
+
+echo "=== delivery/README.md DECLARED scope-forms documentation coverage (AC8) ==="
+
+if [[ -f "$DOC" ]]; then
+  pass "delivery/README.md exists"
+else
+  fail "delivery/README.md is missing — cannot check scope-forms documentation"
+  echo ""
+  echo "FAIL: 1 check(s) failed"
+  exit 1
+fi
+
+# --- The four single-condition scope forms --------------------------------------
+if grep -qF 'ref:' "$DOC"; then
+  pass "documents the ref: scope form"
+else
+  fail "delivery/README.md does not mention the ref: scope form"
+fi
+
+if grep -qF 'commit:' "$DOC"; then
+  pass "documents the commit: scope form"
+else
+  fail "delivery/README.md does not mention the commit: scope form"
+fi
+
+if grep -qF 'author:' "$DOC"; then
+  pass "documents the author: scope form"
+else
+  fail "delivery/README.md does not mention the author: scope form"
+fi
+
+if grep -qF 'branch-prefix:' "$DOC"; then
+  pass "documents the branch-prefix: scope form"
+else
+  fail "delivery/README.md does not mention the branch-prefix: scope form"
+fi
+
+# --- commit: range form (commit:a..b), not just the bare single-sha form --------
+if grep -qE 'commit:[^[:space:]]*\.\.' "$DOC"; then
+  pass "documents the commit:<from>..<to> RANGE form, not just a bare single-sha example"
+else
+  fail "delivery/README.md does not show a commit:<from>..<to> range example"
+fi
+
+# --- Compound (space-separated AND) scope form -----------------------------------
+if grep -qiE 'compound|space-separated' "$DOC"; then
+  pass "documents the compound (space-separated) scope form"
+else
+  fail "delivery/README.md does not mention a compound/space-separated scope form"
+fi
+
+if grep -qiE '\bAND\b|ANDed' "$DOC"; then
+  pass "documents that compound conditions are ANDed (all must match)"
+else
+  fail "delivery/README.md does not state that compound scope conditions are ANDed"
+fi
+
+# --- author:/branch-prefix: combining requirement for identity exemptions -------
+if grep -qiE 'pusher-controlled|pusher-chosen' "$DOC"; then
+  pass "documents the fork-PR pusher-controlled ref/branch-prefix weakness motivating the combining requirement"
+else
+  fail "delivery/README.md does not explain why ref:/branch-prefix: alone are insufficient for identity exemptions (pusher-controlled branch names)"
+fi
+
+if grep -qiE 'MUST combine|combine.*author:|author:.*combine' "$DOC"; then
+  pass "documents that an identity-binding scope MUST combine author: with a second condition"
+else
+  fail "delivery/README.md does not state the author:+branch-prefix:/ref: combining requirement"
+fi
+
+# --- The real worked example (release-please compound scope) --------------------
+if grep -qF 'author:github-actions[bot] branch-prefix:release-please--' "$DOC"; then
+  pass "shows the real worked compound-scope example (release-please: author: + branch-prefix:)"
+else
+  fail "delivery/README.md does not show the real release-please compound-scope worked example"
+fi
+
+# --- Matching semantics: string equality/prefix only, never RegExp --------------
+if grep -qiE 'no.*RegExp|never.*RegExp' "$DOC"; then
+  pass "documents that matching is string equality/prefix only — no RegExp is ever constructed from marker content"
+else
+  fail "delivery/README.md does not state the no-RegExp matching-semantics note"
+fi
+
+echo ""
+if [[ "$errors" -eq 0 ]]; then
+  echo "PASS: delivery/README.md documents all four DECLARED scope forms, compound-AND, and the identity-combining requirement"
+  exit 0
+else
+  echo "FAIL: $errors check(s) failed"
+  exit 1
+fi

package/evals/static/test_universal_bundles.sh

@@ -206,6 +206,40 @@ else
   _fail "Codex agent TOML required-field check failed"
 fi
 
+if node - "$DIST_DIR/codex/.codex/agents" <<'NODE'
+const fs = require("node:fs");
+const path = require("node:path");
+const root = process.argv[2];
+function stringAssignment(text, key, name) {
+  const matches = [...text.matchAll(new RegExp(`^${key}\\s*=\\s*"([^"]*)"\\s*$`, "gm"))];
+  if (matches.length !== 1) throw new Error(`${name}: expected exactly one ${key} assignment, found ${matches.length}`);
+  return matches[0][1];
+}
+const expected = {
+  "tool-planner.toml": ["gpt-5.6-sol", "high"],
+  "tool-worker.toml": ["gpt-5.6-terra", "high"],
+  "tool-code-reviewer.toml": ["gpt-5.6-sol", "high"],
+  "tool-security-reviewer.toml": ["gpt-5.6-sol", "high"],
+  "tool-verifier.toml": ["gpt-5.6-sol", "high"],
+};
+for (const [name, [model, effort]] of Object.entries(expected)) {
+  const text = fs.readFileSync(path.join(root, name), "utf8");
+  if (stringAssignment(text, "model", name) !== model) throw new Error(`${name}: expected model ${model}`);
+  if (stringAssignment(text, "model_reasoning_effort", name) !== effort) throw new Error(`${name}: expected reasoning ${effort}`);
+}
+const fallback = fs.readFileSync(path.join(root, "tool-playwright.toml"), "utf8");
+if (stringAssignment(fallback, "model", "tool-playwright.toml") !== "gpt-5.5" ||
+    stringAssignment(fallback, "model_reasoning_effort", "tool-playwright.toml") !== "low") {
+  throw new Error("unmapped Codex agent no longer uses family fallback");
+}
+console.log("ok");
+NODE
+then
+  _pass "Codex Builder specialist roles use deterministic 5.6 models and unmapped agents retain family fallback"
+else
+  _fail "Codex Builder specialist role routing is missing, stale, or bypasses family fallback"
+fi
+
 if node - "$DIST_DIR/codex/.codex/hooks.json" <<'NODE'
 const fs = require("node:fs");
 const hooks = JSON.parse(fs.readFileSync(process.argv[2], "utf8")).hooks || {};

package/evals/static/test_validate_source_kit_asset_scope.sh

@@ -0,0 +1,259 @@
+#!/usr/bin/env bash
+# test_validate_source_kit_asset_scope.sh — src/tools/validate-source-tree.ts's
+# validateLegacyRefs() kit-owned-asset exemption stays EXACT-PATH scoped.
+#
+# Context (fix pass, issue #303 follow-up): a prior draft of this exemption widened it to
+# "any file in the same directory as a registered kit.json asset" (a directory-level
+# exemption) to make an eval's reference to a kit skill's own helper script pass
+# validate:source. That widening was reverted — the helper is now registered as its own
+# kit.json "assets" entry (data-driven, no validator special-casing) — but the REAL risk
+# the reviewer flagged deserves a permanent regression lock: a directory-level exemption
+# would silence real missing-path detection for every OTHER (bogus, typo'd, deleted) file
+# that happens to sit in the same directory as a legitimately registered asset, not just
+# the intended one.
+#
+# This eval proves the negative: a reference to a FILE THAT DOES NOT EXIST, sitting in the
+# SAME directory as a real, registered kit.json asset (kits/veritas-governance/skills/
+# exemption-usage-review/, which legitimately owns a registered SKILL.md AND a registered
+# assets[] helper), IS STILL FLAGGED by validate:source as a missing source path — never
+# silently exempted just because it shares a directory with real registered assets.
+#
+# Mechanism: writes a throwaway eval-shaped fixture file (self-cleaning, removed in the
+# trap) under evals/integration/ whose CONTENT (built at runtime, not hardcoded as a
+# literal path string in THIS file's own source — see note below) references a bogus
+# filename in that real kit skill directory, runs the REAL `npm run validate:source`
+# against the REAL repo tree, asserts it fails and names the exact bogus path, then
+# removes the throwaway fixture and re-confirms the repo is clean again. Deterministic, no
+# model spend.
+#
+# Self-scan hazard: validate-source-tree's legacy-ref scanner walks EVERY .sh file under
+# evals/, including this one. If the bogus path string were written as a literal in this
+# file's own source text, this eval would trip its own probe merely by existing. The bogus
+# ref/probe-file path strings below are therefore assembled at runtime from parts (never
+# written as one contiguous literal anywhere in this file) so this file's own presence
+# never becomes a false positive.
+#
+# Crash-resilience hazard (fix pass, second review round): the probe fixture below MUST
+# live in a scanned directory (evals/integration/) for the negative probe to mean anything
+# — but a SIGKILL/OOM/CI-timeout landing between the probe's write and this script's EXIT
+# trap would leave it behind, permanently failing validate:source repo-wide for every
+# future run until someone manually deletes it. Two defenses: (1) a SELF-HEALING
+# PRE-FLIGHT, at the very top of this script before anything else runs (including the
+# positive control), that deletes any file matching this eval's own narrowly-scoped
+# probe-name glob and logs a loud line when it heals one — this is the defense that
+# actually matters if the crash window is ever hit; (2) the EXIT trap is kept as the
+# normal-path cleanup, and the probe write is placed as late as possible (immediately
+# before the validate:source call that depends on it) to narrow the window. The glob is
+# deliberately a literal, distinctive prefix this eval alone uses (never a generic
+# wildcard) so the self-heal can never match — and therefore never delete — any real,
+# non-probe file.
+
+set -uo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+cd "$ROOT"
+
+errors=0
+pass() { echo "  ✓ $1"; }
+fail() { echo "  ✗ $1"; errors=$((errors + 1)); }
+
+# Narrowly-scoped probe-name glob prefix — every probe fixture this eval ever writes
+# (negative-probe or self-heal-regression-test fixtures alike) MUST start with this exact
+# literal prefix, and nothing else in the repo may. This is what makes the self-heal safe:
+# it can only ever match files THIS eval itself creates.
+PROBE_GLOB_PREFIX="test__scope_probe_"
+
+# Delete any stale probe fixture(s) matching PROBE_GLOB_PREFIX under evals/integration/,
+# logging a loud line per file healed. Called at pre-flight (below) AND directly by the
+# self-heal regression assertion later in this file (as a lightweight in-process function
+# call, not a recursive re-invocation of this whole script — re-running the full script,
+# with its several npm run validate:source calls, is too heavy to use as the mechanism for
+# a single targeted assertion).
+heal_stale_probes() {
+  local stale_probes=("$ROOT"/evals/integration/${PROBE_GLOB_PREFIX}*)
+  if [[ -e "${stale_probes[0]}" ]]; then
+    for stale in "${stale_probes[@]}"; do
+      echo "  ⚠ SELF-HEAL: removing stale probe fixture left behind by an interrupted prior run: $stale"
+      rm -f "$stale"
+    done
+  fi
+}
+
+# --pre-flight-only: run ONLY the self-healing pre-flight (heal_stale_probes) and exit,
+# skipping every npm run validate:source call in the rest of this script. Exists solely so
+# the end-to-end self-heal assertion below can exercise the REAL pre-flight wiring at the
+# top of this script (not just heal_stale_probes() called in isolation) as a fresh
+# subprocess, WITHOUT paying for a full second run of this eval's several
+# npm run validate:source builds — keeps the added runtime bounded to one lightweight
+# glob-and-delete pass, not a second full eval run.
+if [[ "${1:-}" == "--pre-flight-only" ]]; then
+  heal_stale_probes
+  exit 0
+fi
+
+echo "=== validate-source-tree kit-asset exemption stays exact-path scoped (no directory-level exemption) ==="
+
+# --- Self-healing pre-flight (reviewer HIGH: hard-kill/OOM/CI-timeout resilience) -
+heal_stale_probes
+
+# Real skill directory + its two REAL registered kit.json assets (assembled from parts —
+# see the self-scan hazard note above; these ARE real, existing paths, so splitting them
+# is purely to avoid THIS descriptive assertion accidentally reading as a probe target).
+KIT_DIR="kits/veritas-governance"
+REAL_SKILL_DIR="$KIT_DIR/skills""/exemption-usage-review"
+REAL_SKILL_MD="$REAL_SKILL_DIR/SKILL.md"
+REAL_HELPER="$REAL_SKILL_DIR/review-exemptions.mjs"
+# kit.json's own "path" fields are relative to the KIT directory, not the repo root.
+KIT_RELATIVE_SKILL_MD="${REAL_SKILL_MD#"$KIT_DIR/"}"
+KIT_RELATIVE_HELPER="${REAL_HELPER#"$KIT_DIR/"}"
+
+# --- Sanity: the real skill directory + its two REAL registered assets exist -----
+if [[ -f "$ROOT/$REAL_SKILL_MD" && -f "$ROOT/$REAL_HELPER" ]]; then
+  pass "real skill directory has both registered files present (SKILL.md, review-exemptions.mjs)"
+else
+  fail "expected real skill directory files missing — cannot run the negative probe against real fixtures"
+fi
+
+if grep -qF "\"path\": \"$KIT_RELATIVE_SKILL_MD\"" "$ROOT/$KIT_DIR/kit.json" \
+  && grep -qF "\"path\": \"$KIT_RELATIVE_HELPER\"" "$ROOT/$KIT_DIR/kit.json"; then
+  pass "both files are registered in kit.json (SKILL.md under skills[], review-exemptions.mjs under assets[])"
+else
+  fail "kit.json does not register both real files under skills[]/assets[] as expected"
+fi
+
+# --- Positive control: validate:source passes on the clean, real tree -----------
+if npm run validate:source > /tmp/vskas_positive.out 2>&1; then
+  pass "positive control: npm run validate:source passes on the clean tree (no probe fixture present)"
+else
+  fail "positive control: npm run validate:source unexpectedly failed on the clean tree — cannot trust the negative probe below"; tail -20 /tmp/vskas_positive.out
+fi
+
+# --- Negative probe: a BOGUS sibling filename in the SAME real directory is flagged --
+# Assembled from parts at runtime (see self-scan hazard note above) — this exact
+# contiguous string never appears anywhere else in this file's own source text.
+BOGUS_FILENAME_PART1="this-file-does-not"
+BOGUS_FILENAME_PART2="-exist-probe.mjs"
+# NOTE: legacyRefRe (validate-source-tree.ts) matches starting at recognized top-level
+# prefixes (skills / evals / scripts / etc — see the regex in validate-source-tree.ts),
+# not at "kits" — so the ref as written in the probe fixture (and as reported in
+# validate:source's failure output) is kit-directory-relative, not repo-root-relative. Use
+# the kit-relative form here to match what the scanner actually records. Built from parts
+# (never one contiguous literal in this file's own comments or code) for the same
+# self-scan-hazard reason noted at the top of this file.
+BOGUS_REF="skills""/exemption-usage-review/$BOGUS_FILENAME_PART1$BOGUS_FILENAME_PART2"
+PROBE_FILE_NAME_PART1="${PROBE_GLOB_PREFIX}bogus"
+PROBE_FILE_NAME_PART2="_ref.sh"
+PROBE_FILE="$ROOT/evals/integration/$PROBE_FILE_NAME_PART1$PROBE_FILE_NAME_PART2"
+
+cleanup() {
+  rm -f "$PROBE_FILE"
+}
+trap cleanup EXIT
+
+# Content-writer for a throwaway probe fixture — factored out so the self-heal regression
+# assertion below can plant an equivalent stale fixture without duplicating the heredoc.
+write_probe_fixture() {
+  local target="$1"
+  {
+    echo "#!/usr/bin/env bash"
+    echo "# Throwaway probe fixture (self-removing) — exists only for the duration of"
+    echo "# test_validate_source_kit_asset_scope.sh's negative-probe assertion."
+    echo "HELPER=\"\$KIT/$BOGUS_REF\""
+  } > "$target"
+}
+
+# Write the probe fixture as LATE as possible (immediately before the validate:source call
+# that depends on it) to narrow the window a hard-kill could land in between write and
+# cleanup — this is the second line of defense; the self-healing pre-flight above is the
+# first (and the one that actually matters if this window is ever hit).
+write_probe_fixture "$PROBE_FILE"
+
+if npm run validate:source > /tmp/vskas_negative.out 2>&1; then
+  fail "negative probe: validate:source PASSED with a bogus sibling-file reference present — the exemption is over-wide (directory-level), not exact-path"
+else
+  pass "negative probe: validate:source FAILS (exit non-zero) with a bogus sibling-file reference present"
+fi
+
+if grep -qF "references missing source path: $BOGUS_REF" /tmp/vskas_negative.out; then
+  pass "negative probe: failure output names the exact bogus path ($BOGUS_REF) — real missing-path detection is not silenced by directory proximity to registered assets"
+else
+  fail "negative probe: expected failure output to name the bogus path — output: $(tail -20 /tmp/vskas_negative.out)"
+fi
+
+# --- Cleanup verification: removing the probe restores a clean, passing tree ----
+rm -f "$PROBE_FILE"
+if npm run validate:source > /tmp/vskas_cleanup.out 2>&1; then
+  pass "cleanup: npm run validate:source passes again once the throwaway probe fixture is removed"
+else
+  fail "cleanup: npm run validate:source did not pass again after removing the probe fixture"; tail -20 /tmp/vskas_cleanup.out
+fi
+
+# --- Self-heal regression lock: a STALE probe left by a simulated hard-kill is healed ---
+# Simulates exactly the failure mode this fix pass addresses: a probe fixture left behind
+# by an interrupted prior run (no EXIT trap ever ran for it, since this write below is
+# deliberately NOT wrapped in this script's own trap/cleanup — it models "the process died
+# before its trap could fire"). Exercises heal_stale_probes() DIRECTLY (a lightweight
+# in-process function call — NOT a recursive re-invocation of this whole script, which
+# would be several redundant npm run validate:source builds just to prove one glob-delete
+# works) and confirms it heals: the file is removed and a loud line names it.
+STALE_PROBE_FILE_PART1="${PROBE_GLOB_PREFIX}stale"
+STALE_PROBE_FILE_PART2="_simulated.sh"
+STALE_PROBE_FILE="$ROOT/evals/integration/$STALE_PROBE_FILE_PART1$STALE_PROBE_FILE_PART2"
+write_probe_fixture "$STALE_PROBE_FILE"
+if [[ -f "$STALE_PROBE_FILE" ]]; then
+  pass "self-heal setup: pre-planted a stale probe fixture (simulating an interrupted prior run with no trap-driven cleanup)"
+else
+  fail "self-heal setup: failed to pre-plant the stale probe fixture — cannot exercise the self-heal assertion below"
+fi
+
+HEAL_OUT="$(heal_stale_probes 2>&1)"
+
+if echo "$HEAL_OUT" | grep -qF "SELF-HEAL: removing stale probe fixture left behind by an interrupted prior run: $STALE_PROBE_FILE"; then
+  pass "self-heal: heal_stale_probes() logged a loud SELF-HEAL line naming the exact stale probe path it removed"
+else
+  fail "self-heal: expected a loud SELF-HEAL line naming the stale probe path — output: $HEAL_OUT"
+fi
+
+if [[ ! -e "$STALE_PROBE_FILE" ]]; then
+  pass "self-heal: the stale probe fixture no longer exists after heal_stale_probes() ran (healed, not just logged)"
+else
+  fail "self-heal: the stale probe fixture STILL EXISTS after heal_stale_probes() ran — self-heal did not actually delete it"
+  rm -f "$STALE_PROBE_FILE"
+fi
+
+# End-to-end confirmation: a FRESH invocation of this eval (its own pre-flight, not a
+# hand-called function) also heals a pre-planted stale probe and still passes overall —
+# proves the pre-flight wiring at the top of this script, not just the helper function in
+# isolation. This is the one genuinely-recursive re-invocation in this file; kept to
+# exactly one (not three, as an earlier draft did) to bound the added runtime.
+write_probe_fixture "$STALE_PROBE_FILE"
+FRESH_RUN_OUT="$(bash "$ROOT/evals/static/test_validate_source_kit_asset_scope.sh" --pre-flight-only 2>&1)"
+FRESH_RUN_CODE=$?
+
+if echo "$FRESH_RUN_OUT" | grep -qF "SELF-HEAL: removing stale probe fixture left behind by an interrupted prior run: $STALE_PROBE_FILE"; then
+  pass "self-heal (end-to-end): a fresh --pre-flight-only invocation of this eval's own pre-flight healed the pre-planted stale probe"
+else
+  fail "self-heal (end-to-end): expected the fresh invocation to log a loud SELF-HEAL line naming the stale probe — output: $(echo "$FRESH_RUN_OUT" | head -10)"
+fi
+
+if [[ ! -e "$STALE_PROBE_FILE" ]]; then
+  pass "self-heal (end-to-end): the stale probe fixture no longer exists after the fresh --pre-flight-only invocation"
+else
+  fail "self-heal (end-to-end): the stale probe fixture STILL EXISTS after the fresh invocation"
+  rm -f "$STALE_PROBE_FILE"
+fi
+
+if [[ "$FRESH_RUN_CODE" -eq 0 ]]; then
+  pass "self-heal (end-to-end): the fresh --pre-flight-only invocation exits 0"
+else
+  fail "self-heal (end-to-end): the fresh --pre-flight-only invocation exited $FRESH_RUN_CODE (expected 0) — output: $(echo "$FRESH_RUN_OUT" | tail -10)"
+fi
+
+echo ""
+if [[ "$errors" -eq 0 ]]; then
+  echo "PASS: validate-source-tree kit-asset exemption is exact-path scoped (no directory-level widening)"
+  exit 0
+else
+  echo "FAIL: $errors check(s) failed"
+  exit 1
+fi

package/evals/static/test_workflow_skills.sh

@@ -552,7 +552,7 @@ require_text "$WORKFLOW_STEERING_INTEGRATION" 'workflow steering hook neutralize
 require_text "$WORKFLOW_STEERING_INTEGRATION" 'workflow steering hook appends context-map recovery guidance' "workflow steering integration covers context-map guidance"
 require_text "$WORKFLOW_STEERING_INTEGRATION" 'workflow steering hook emits ambient state guidance at user prompt submit' "workflow steering integration covers ambient state guidance"
 require_text "$WORKFLOW_STEERING_INTEGRATION" 'Claude hook adapter surfaces Builder workflow route for coding prompts' "workflow steering integration covers Claude Builder-route prompt guidance"
-require_text "$WORKFLOW_STEERING_INTEGRATION" 'publish/release-readiness and learning-review' "workflow steering integration covers full Builder lifecycle guidance"
+require_text "$WORKFLOW_STEERING_INTEGRATION" 'release-readiness and learning-review' "workflow steering integration covers full Builder lifecycle guidance"
 require_text "$WORKFLOW_STEERING_INTEGRATION" 'does not route explicit review-only prompts into Builder workflow' "workflow steering integration covers review-only prompt suppression"
 require_text "$WORKFLOW_STEERING_INTEGRATION" 'does not route explicit validation-only prompts into Builder workflow' "workflow steering integration covers validation-only prompt suppression"
 require_text "$WORKFLOW_STEERING_INTEGRATION" 'does not route question-only test prompts into Builder workflow' "workflow steering integration covers question-only test prompt suppression"

package/kits/builder/flows/build.flow.json

@@ -1,6 +1,6 @@
 {
   "id": "builder.build",
-  "version": "1.0",
+  "version": "1.1",
   "phase_map": {
     "pickup": "pull-work",
     "planning": "plan",
@@ -36,8 +36,7 @@
             "claimType": "builder.pull-work.selected",
             "subjectType": "work-item",
             "accepted_statuses": [
-              "trusted",
-              "accepted"
+              "verified"
             ]
           }
         }
@@ -55,8 +54,7 @@
             "claimType": "builder.design-probe.pickup-readiness",
             "subjectType": "work-item",
             "accepted_statuses": [
-              "trusted",
-              "accepted"
+              "verified"
             ]
           }
         },
@@ -69,8 +67,7 @@
             "claimType": "builder.design-probe.decisions",
             "subjectType": "decision",
             "accepted_statuses": [
-              "trusted",
-              "accepted"
+              "verified"
             ]
           }
         }
@@ -88,8 +85,7 @@
             "claimType": "builder.plan.implementation",
             "subjectType": "artifact",
             "accepted_statuses": [
-              "trusted",
-              "accepted"
+              "verified"
             ]
           }
         }
@@ -107,8 +103,7 @@
             "claimType": "builder.execute.scope",
             "subjectType": "change",
             "accepted_statuses": [
-              "trusted",
-              "accepted"
+              "verified"
             ]
           }
         }
@@ -137,8 +132,7 @@
             "claimType": "builder.verify.tests",
             "subjectType": "flow-step",
             "accepted_statuses": [
-              "trusted",
-              "accepted"
+              "verified"
             ]
           }
         },
@@ -152,9 +146,7 @@
             "claimType": "builder.verify.policy-compliance",
             "subjectType": "artifact",
             "accepted_statuses": [
-              "trusted",
-              "accepted",
-              "advisory"
+              "verified"
             ]
           }
         }
@@ -183,8 +175,7 @@
             "claimType": "builder.merge-ready.readiness",
             "subjectType": "change",
             "accepted_statuses": [
-              "trusted",
-              "accepted"
+              "verified"
             ]
           }
         }

package/kits/builder/flows/publish-learn.flow.json

@@ -1,6 +1,6 @@
 {
   "id": "builder.publish-learn",
-  "version": "1.0",
+  "version": "1.1",
   "description": "Reusable delivery closeout extension: publish the verified change, confirm merge readiness from provider/CI evidence, and record learning feedback.",
   "steps": [
     { "id": "pr-open", "next": "merge-ready-ci" },
@@ -27,6 +27,7 @@
             "claimType": "builder.pr-open.pull-request",
             "subjectType": "pull-request",
             "accepted_statuses": [
+              "verified",
               "trusted",
               "accepted"
             ]
@@ -46,6 +47,7 @@
             "claimType": "builder.merge-ready-ci.readiness",
             "subjectType": "pull-request",
             "accepted_statuses": [
+              "verified",
               "trusted",
               "accepted"
             ]
@@ -65,6 +67,7 @@
             "claimType": "builder.learn.decisions",
             "subjectType": "decision",
             "accepted_statuses": [
+              "verified",
               "trusted",
               "accepted"
             ]
@@ -79,6 +82,7 @@
             "claimType": "builder.learn.evidence",
             "subjectType": "release",
             "accepted_statuses": [
+              "verified",
               "trusted",
               "accepted"
             ]