crewforth 3.2.0-rc.1 → 3.2.0-rc.3

package/kit/eval/smoke-test.sh

@@ -1674,8 +1674,8 @@ if [ "$IS_KIT" = 1 ]; then
   _ch4_t="$(printf '%s\n' "all fo""ur channels carry it" "Dö""rt kanal da taşıyor" "of fo""ur distribution channels" "three channels" "four channel types of RGB" | grep -ciE "$_ch4_re")"
   if [ "$_ch4_t" != 3 ]; then fail "channel-count check cannot tell its twins apart (caught $_ch4_t of the 3 planted, want 3 and none of the 2 clean)"
   elif git -C "$KR" rev-parse --git-dir >/dev/null 2>&1; then
-    _ch4_hits="$(git -C "$KR" grep -niIE "$_ch4_re" -- . ':!CHANGELOG.md' ':!site/src/content/docs/changelog.md' ':!site/src/content/docs/tr/changelog.md' 2>/dev/null)"
-    [ -z "$_ch4_hits" ] && pass "no tracked file gives the old channel count — three ship Crewforth (CHANGELOG exempt; twins 3/3 caught, 0/2 flagged)" \
+    _ch4_hits="$(git -C "$KR" grep -niIE "$_ch4_re" -- . ':!CHANGELOG.md' ':!site/src/content/docs/changelog.md' ':!site/src/content/docs/tr/changelog.md' ':!packaging/review-bench/cases' 2>/dev/null)"
+    [ -z "$_ch4_hits" ] && pass "no tracked file gives the old channel count — three ship Crewforth (CHANGELOG and the review bench's past-version cases exempt; twins 3/3 caught, 0/2 flagged)" \
       || fail "a tracked file still gives the old channel count (there are three): $(printf '%s\n' "$_ch4_hits" | head -3 | tr '\n' ' ')"
   else skip scope "channel-count check not run — not a git checkout of Crewforth's source"; fi
   # THE CLIENT SIDE IS STACK-NEUTRAL (3.0.1): the install summary, the help and the install page named the scope as
@@ -3407,6 +3407,46 @@ pass|rm -f file.txt
 pass|rm file.txt
 pass|rmdir empty/
 RMCASES
+# V4: a stopped or asked command says what it would have touched (eval/lib/preview.sh): the changes a reset discards,
+# the paths a clean deletes, the files under an rm target inside the project, what a commit takes and a push sends.
+# Read only, and after the decision: a preview library that fails or ends its shell leaves the decision as it was.
+PVR="$(mktemp -d)"; PVR="$(cd -P "$PVR" && pwd)"; PVO="$(mktemp)"; PVE="$(mktemp)"; PVX="$(mktemp -d)"
+( cd "$PVR" && git init -q . && git -c user.email=a@b -c user.name=a commit -q --allow-empty -m base \
+  && printf '1\n' > a.txt && printf '2\n' > b.txt && git add . && git -c user.email=a@b -c user.name=a commit -q -m two \
+  && printf 'x\n' >> a.txt && printf 'y\n' >> b.txt && mkdir -p out/sub && : > out/sub/1 && : > out/2 && : > junk.log ) >/dev/null 2>&1
+_pvrun(){ printf '{"session_id":"s","permission_mode":"default","hook_event_name":"PreToolUse","tool_name":"Bash","tool_input":{"command":"%s"},"cwd":"%s"}' "$1" "$PVR" \
+  | ( cd "$PVR" && CLAUDE_PROJECT_DIR="$PVR" CREW_GATE_LOG=/dev/null bash "${2:-$HOOKS}/guard-bash.sh" ) > "$PVO" 2> "$PVE"; _pvrc=$?; }
+_pvb=""
+_pvrun "git reset --hard"; { [ "$_pvrc" = 2 ] && grep -q 'Preview: it would discard 2 uncommitted change' "$PVE"; } || _pvb="$_pvb [reset: rc $_pvrc · $(grep Preview "$PVE")]"
+_pvrun "git reset --hard HEAD~1"; grep -q 'and drop 1 commit(s) from the branch' "$PVE" || _pvb="$_pvb [reset HEAD~1: $(grep Preview "$PVE")]"
+_pvrun "git clean -fd"; { [ "$_pvrc" = 2 ] && grep -q 'lists 2 path(s) it would delete' "$PVE"; } || _pvb="$_pvb [clean: $(grep Preview "$PVE")]"
+_pvrun "rm -rf out/"; { [ "$_pvrc" = 2 ] && grep -q 'Preview: out/: 2 file(s)' "$PVE"; } || _pvb="$_pvb [rm inside: $(grep Preview "$PVE")]"
+_pvrun "rm -rf $PVX/"; { [ "$_pvrc" = 2 ] && grep -q 'outside the project, not counted' "$PVE"; } || _pvb="$_pvb [rm outside: $(grep Preview "$PVE")]"
+# The asked ones: a commit (past §4.6, with the review record) and a push to a remote of its own.
+( cd "$PVR" && git add a.txt && mkdir -p .claude && D=$(git diff --cached --no-ext-diff --no-textconv | git hash-object --stdin) && H=$(git rev-parse HEAD) \
+  && printf '{"diff_oid":"%s","head":"%s","ts":"%s"}\n' "$D" "$H" "2026-01-01T00:00:00Z" > .claude/review-pass.json ) >/dev/null 2>&1
+_pvrun "git commit -m wip"; { [ "$_pvrc" = 0 ] && grep -q 'Preview: 1 file changed, 1 insertion' "$PVO"; } || _pvb="$_pvb [commit ask: rc $_pvrc · $(grep -o 'Preview[^\\]*' "$PVO")]"
+( cd "$PVR" && git init -q --bare "$PVX/r.git" && git remote add origin "$PVX/r.git" && git push -q -u origin HEAD:main && git branch -q -u origin/main \
+  && git -c user.email=a@b -c user.name=a commit -q --allow-empty -m "one ahead" ) >/dev/null 2>&1
+_pvrun "git push"; { [ "$_pvrc" = 0 ] && grep -q 'Preview: it sends 1 commit(s) to origin/main' "$PVO"; } || _pvb="$_pvb [push ask: rc $_pvrc · $(grep -o 'Preview[^\\]*' "$PVO")]"
+# A push that names its branch sends to that one: an existing one is counted, an unknown one would be created.
+_pvrun "git push origin main"; grep -q 'Preview: it sends 1 commit(s) to origin/main' "$PVO" || _pvb="$_pvb [named push: $(grep -o 'Preview[^\\]*' "$PVO")]"
+_pvrun "git push origin feat/new"; grep -q 'Preview: origin/feat/new is not known here; the push would create it' "$PVO" || _pvb="$_pvb [push to a new branch: $(grep -o 'Preview[^\\]*' "$PVO")]"
+# A push of another local branch counts that branch's commits, not the current one's (current: 1 ahead; side: 3).
+( cd "$PVR" && _cur="$(git rev-parse --abbrev-ref HEAD)" && git checkout -q -b side && git -c user.email=a@b -c user.name=a commit -q --allow-empty -m s1 \
+  && git -c user.email=a@b -c user.name=a commit -q --allow-empty -m s2 && git checkout -q "$_cur" ) >/dev/null 2>&1
+_pvrun "git push origin side:main"; grep -q 'Preview: it sends 3 commit(s) to origin/main' "$PVO" || _pvb="$_pvb [push of another branch: $(grep -o 'Preview[^\\]*' "$PVO")]"
+# A pattern is said, not expanded by the hook's own shell into the files it matches.
+_pvrun "rm -rf out/*"; { [ "$_pvrc" = 2 ] && grep -q 'Preview: out/\*: a pattern, not counted' "$PVE"; } || _pvb="$_pvb [rm pattern: rc $_pvrc · $(grep Preview "$PVE")]"
+# A broken preview library: one that exits, one that does not parse. The decision stays rc 2 and nothing is said.
+_pvh="$(mktemp -d)"; cp -R "$HOOKS" "$_pvh/hooks"; mkdir -p "$_pvh/eval"; cp -R "$ROOT/eval/lib" "$_pvh/eval/lib"
+printf 'crew_preview(){ exit 0; }\n' > "$_pvh/eval/lib/preview.sh"; _pvrun "git reset --hard" "$_pvh/hooks"
+{ [ "$_pvrc" = 2 ] && ! grep -q Preview "$PVE"; } || _pvb="$_pvb [a preview that exits changed the decision: rc $_pvrc]"
+printf 'crew_preview(){ ( \n' > "$_pvh/eval/lib/preview.sh"; _pvrun "rm -rf out/" "$_pvh/hooks"
+[ "$_pvrc" = 2 ] || _pvb="$_pvb [a preview that does not parse changed the decision: rc $_pvrc]"
+[ -z "$_pvb" ] && pass "a stopped command says what it would have touched: reset (2 changes, and 1 commit for HEAD~1), clean (2 paths, dry run), rm (2 files inside the project; a target outside it not counted); an asked commit and push say what they take and send (a push that names a new branch says it would create it, one of another local branch counts that branch's commits); a pattern is said, not expanded; a preview library that exits or does not parse leaves the decision at rc 2" \
+  || fail "risk preview:$_pvb"
+rm -rf "$PVR" "$PVX" "$_pvh"; rm -f "$PVO" "$PVE"
 # §4.5 whole-tree revert. `reset --hard` was gated from the start; `git checkout -- .` does the same damage and
 # was not — a verification subagent used it over uncommitted work in this repo and the tree went with it. The
 # single-file form stays allowed (everyday, recoverable), and the two ways around a naive matcher — `git -C`
@@ -4492,6 +4532,41 @@ else fail "session-rehydrate git-hooks states read '$_got' (want '--UO---': no h
 rm -rf "$RHG"
 _rj=0; for _o in "$_o1" "$_o2" "$_o4" "$_o6"; do printf '%s' "$_o" > "$RHD/o.json"; awk -v op=validate -f "$ROOT/eval/lib/settings-json.awk" "$RHD/o.json" 2>/dev/null || _rj=$((_rj+1)); done
 [ "$_rj" = 0 ] && pass "each language/handover output is valid JSON (Crewforth's reader, 4 of 4)" || fail "$_rj session-rehydrate output(s) are not valid JSON"
+# V3: just before a compaction, precompact-snapshot.sh writes the state of the work from the files and the gates;
+# after it, session-rehydrate gives that state back, to the same session only. The snapshot never blocks the
+# compaction (PreCompact exit 2 would): rc 0 and silent in every case, a broken one included.
+PCS="$HOOKS/precompact-snapshot.sh"
+if [ ! -x "$PCS" ]; then fail "precompact-snapshot.sh missing/not executable"
+else
+  PCD="$(mktemp -d)"; PCD="$(cd -P "$PCD" && pwd)"; _pco="$(mktemp)"; _pcr="$(mktemp)"   # outside the repository: a file there is uncommitted
+  ( cd "$PCD" && git init -q . && git -c user.email=a@b -c user.name=a commit -q --allow-empty -m base ) >/dev/null 2>&1
+  mkdir -p "$PCD/.claude" "$PCD/.crewforth" "$PCD/docs"
+  printf 'x\n' > "$PCD/plain.txt"; printf 'x\n' > "$PCD/odd\"na	me.txt"; printf 'x\n' > "$PCD/ıİşğü.txt"
+  printf '{\n  "findings": [\n    {\n      "rule": "a"\n    },\n    {\n      "rule": "b"\n    }\n  ]\n}\n' > "$PCD/.crewforth/review.json"
+  printf 'BLOCK\t§4.2\tvendor name\t\n' > "$PCD/.git/crewforth-gate-log.tsv"
+  printf '{"session_id":"s-1","cwd":"%s","hook_event_name":"PreCompact","trigger":"auto","custom_instructions":null}' "$PCD" | CLAUDE_PROJECT_DIR="$PCD" bash "$PCS" > "$_pco" 2>&1; _pcrc=$?
+  _snap="$PCD/.git/crewforth-session-snapshot.json"; _sb=""
+  [ -f "$_snap" ] && IFS= read -r _sb < "$_snap"
+  _rs(){ printf '{"session_id":"%s","source":"%s","cwd":"%s"}' "$1" "$2" "$PCD" | CLAUDE_PROJECT_DIR="$PCD" bash "$HOOKS/session-rehydrate.sh" 2>/dev/null; }
+  _same="$(_rs s-1 compact)"; _other="$(_rs s-2 compact)"; _start="$(_rs s-1 startup)"
+  printf '%s' "$_same" > "$_pcr"
+  _b=""
+  [ "$_pcrc" = 0 ] && [ ! -s "$_pco" ] || _b="$_b [rc $_pcrc, printed: $(head -c 120 "$_pco")]"
+  awk -v op=validate -f "$ROOT/eval/lib/settings-json.awk" "$_snap" 2>/dev/null || _b="$_b [the snapshot is not valid JSON: ${_sb:0:160}]"
+  case "$_sb" in *'"session_id":"s-1"'*'"trigger":"auto"'*'branch '*'4 uncommitted file(s)'*'ıİşğü.txt'*'§4.2 vendor name'*'review 2'*) ;; *) _b="$_b [snapshot: ${_sb:0:200}]" ;; esac
+  case "$_same" in *'Before this context was compacted'*'review 2'*) ;; *) _b="$_b [not given back after compaction: ${_same:0:120}]" ;; esac
+  awk -v op=validate -f "$ROOT/eval/lib/settings-json.awk" "$_pcr" 2>/dev/null || _b="$_b [the SessionStart output with an odd file name is not valid JSON]"
+  case "$_other$_start" in *'Before this context was compacted'*) _b="$_b [given to another session or on a plain start]" ;; esac
+  # Fail-open: no repository, a cwd that is not there, no JSON at all, and the hook alone without eval/lib.
+  _nr="$(mktemp -d)"; _alone="$(mktemp -d)"; cp "$PCS" "$_alone/precompact-snapshot.sh"
+  printf '{"session_id":"s-9","cwd":"%s","trigger":"manual"}' "$_nr" | CLAUDE_PROJECT_DIR="$_nr" bash "$PCS" >/dev/null 2>&1 || _b="$_b [non-repo rc $?]"
+  printf '{"session_id":"s-9","cwd":"/nonexistent/x"}' | CLAUDE_PROJECT_DIR= bash "$PCS" >/dev/null 2>&1 || _b="$_b [missing cwd rc $?]"
+  printf 'not json' | CLAUDE_PROJECT_DIR="$_nr" bash "$PCS" >/dev/null 2>&1 || _b="$_b [garbage rc $?]"
+  printf '{"session_id":"s-9","cwd":"%s"}' "$PCD" | CLAUDE_PROJECT_DIR="$PCD" bash "$_alone/precompact-snapshot.sh" >/dev/null 2>&1 || _b="$_b [no eval/lib rc $?]"
+  [ -z "$_b" ] && pass "PreCompact writes the state of the work (branch, uncommitted files, the last gate refusal, the findings on record) as valid JSON, even for a file name with a quote and a tab, and names a Turkish file name as written; SessionStart gives it back after a compaction to the same session only, never to another or on a plain start; the hook is rc 0 and silent outside a repository, with no cwd, with no JSON and without eval/lib" \
+    || fail "precompact snapshot:$_b"
+  rm -rf "$PCD" "$_nr" "$_alone" "$_pco" "$_pcr"
+fi
 # Asked of the entry that carries session-rehydrate, not of the file: skill-trust's entry has the same matcher.
 _rw(){ _n="$(awk -v op=len -v path=hooks.SessionStart -f "$ROOT/eval/lib/settings-json.awk" "$1" 2>/dev/null)"; _i=0
        while [ "$_i" -lt "${_n:-0}" ]; do _e="$(awk -v op=get -v path="hooks.SessionStart.$_i" -f "$ROOT/eval/lib/settings-json.awk" "$1" 2>/dev/null)"
@@ -4971,6 +5046,8 @@ sec "== 7u) update prompt: ASKS about a release, never installs one, never spend
 UPD="$(mktemp -d)"; mkdir -p "$UPD/.claude/.state" "$UPD/.claude/hooks"
 printf '3.0.0\n' > "$UPD/.claude/VERSION"
 UH="$HOOKS/session-update-check.sh"; cp "$UH" "$UPD/.claude/hooks/"
+# The hook reads and orders the install's version through eval/lib/version.sh, beside it in an install.
+mkdir -p "$UPD/.claude/eval/lib"; cp "$ROOT/eval/lib/version.sh" "$UPD/.claude/eval/lib/"
 uc(){ ( printf '{"hook_event_name":"SessionStart","source":"startup","cwd":"%s"}' "$UPD" \
         | env -u CI -u CREW_NO_UPDATE_CHECK -u CREW_LANG CLAUDE_CODE_SESSION_ATTENDED=1 LANG=C LC_ALL= LC_MESSAGES= \
           CLAUDE_PROJECT_DIR="$UPD" CREW_UPDATE_URL="${1:-http://10.255.255.1/blackhole}" bash "$UH" 2>/dev/null ); }
@@ -5099,10 +5176,25 @@ cp "$UH" "$UPD/uh-twin.sh"; printf 'npx --yes crewforth@latest update --here --y
 if [ "$_ni" = 0 ] && [ "$(_noinst "$UPD/uh-twin.sh")" -ge 1 ]; then pass "the update hook runs no installer (npx, npm, adopt.sh, start.sh): 0 lines; the twin with one is caught"
 else fail "the update hook contains an installer call ($_ni line(s)) — it must only ask, or the twin was not caught"; fi
 
+# 12b) AN INSTALLED RC IS OFFERED ITS FINAL. The installer copies VERSION into the project, and an rc's says
+#      X.Y.Z-rc.N: 3.2.0 published over an installed 3.2.0-rc.1 is newer (it used to read as the same 3.2.0, and the
+#      rc's user was never told the final shipped). An older final is not offered, and what the network answers is
+#      still held to X.Y.Z: an rc off the network is no version.
+printf '3.2.0-rc.1\n' > "$UPD/.claude/VERSION"
+ucache 3.2.0
+case "$(uc)" in *'"Crewforth v3.2.0 is out (installed: v3.2.0-rc.1). Update now?"'*) pass "an installed 3.2.0-rc.1 is offered the final 3.2.0" ;;
+                *) fail "an installed 3.2.0-rc.1 was not offered the final 3.2.0 (got: $(uc | head -c 120))" ;; esac
+ucache 3.1.1
+[ -z "$(uc)" ] && pass "an installed rc is not offered an older final (3.1.1 under 3.2.0-rc.1)" || fail "3.1.1 was offered over an installed 3.2.0-rc.1"
+ucache 3.2.0-rc.2
+[ -z "$(uc)" ] && pass "an rc answered off the network is no version: nothing asked" || fail "a pre-release from the network was put in a question"
+printf '3.0.0\n' > "$UPD/.claude/VERSION"; ustate
+
 # 13) THE PLUGIN EDITION. Its version is its own manifest, its release is the marketplace repo's copy, its state is
 #     user-level (a plugin serves every project), and its Update path is the plugin command — measured to work from
 #     inside a session, applying on restart.
-PLG="$UPD/plugin"; mkdir -p "$PLG/.claude-plugin" "$PLG/hooks"; cp "$UH" "$PLG/hooks/"
+PLG="$UPD/plugin"; mkdir -p "$PLG/.claude-plugin" "$PLG/hooks" "$PLG/eval/lib"; cp "$UH" "$PLG/hooks/"
+cp "$ROOT/eval/lib/version.sh" "$PLG/eval/lib/"
 printf '{"name":"crewforth","version":"3.0.0"}\n' > "$PLG/.claude-plugin/plugin.json"
 XDG="$UPD/xdg"; mkdir -p "$XDG/crewforth"
 printf '3.1.0 %s\n' "$(date +%s)" > "$XDG/crewforth/update-check"
@@ -6363,6 +6455,14 @@ GNI="$(mktemp -d)"; ( cd "$GNI" && git init -q . && mkdir -p .claude )
 ( cd "$GNI" && gjson 'git reset --hard' | bash "$ROOT/hooks/guard-bash.sh" >/dev/null 2>&1 )
 [ -e "$GNI/.claude/gate-log.tsv" ] && fail "gate log wrote into a repo where the path is not gitignored" \
                                    || pass "gate log declines a path git would track"
+# ...and since 3.2.0 writes inside the repository's .git instead, which is never committed: a project that tracks
+# .claude/ had no record of a refusal at all (seen on a real project: no 🛡, no note of a lowered model).
+# The readers find it there (eval/lib/gate-log.sh, the one rule), and `git status` stays clean.
+_gni_read="$( cd "$GNI" && . "$ROOT/eval/lib/gate-log.sh" && crew_gate_log_read && printf '%s' "$_CREW_GL" )"
+_gni_st="$( cd "$GNI" && git status --porcelain --untracked-files=all 2>/dev/null | grep -c gate-log )"
+{ [ -s "$GNI/.git/crewforth-gate-log.tsv" ] && grep -q '^BLOCK' "$GNI/.git/crewforth-gate-log.tsv" && [ "$_gni_st" = 0 ] && [ "$_gni_read" = ./.git/crewforth-gate-log.tsv ]; } \
+  && pass "where git would track it, the gate log goes inside .git: written, found by the readers, and git status stays clean" \
+  || fail "gate log inside .git: file $([ -s "$GNI/.git/crewforth-gate-log.tsv" ] && echo written || echo MISSING), git status shows it ${_gni_st}x, reader found '${_gni_read}'"
 ( cd "$GNI" && echo '.claude/' > .gitignore && gjson 'git reset --hard' | bash "$ROOT/hooks/guard-bash.sh" >/dev/null 2>&1 )
 [ -s "$GNI/.claude/gate-log.tsv" ] && pass "gate log writes once the path is gitignored" \
                                    || fail "gate log stayed silent even though the path is ignored"
@@ -8218,7 +8318,7 @@ _v2new(){ rm -rf "$_v2p" "$_v2/home"; mkdir -p "$_v2/home" "$_v2p/.claude" "$_v2
 _v2e(){ env -u CREW_MODEL_ROUTING -u CREW_ALLOW_FABLE -u CREW_MODEL_CAL_N -u CREW_MODEL_CAL_PCT -u CREW_VERIFY_NO_SETSID -u CREW_VERIFY_TIMEOUT -u CLAUDE_CONFIG_DIR -u CREW_MANAGED_SETTINGS HOME="$_v2/home" CLAUDE_PROJECT_DIR="$_v2p" CREW_GATE_LOG=/dev/null ${_V2ENV:-} "$@"; }
 # $1 agent, $2 model, $3 files, $4 change, $5 verify, $6 extra fields -> _v2rc, stderr in $_v2/err
 _v2call(){ printf '{"session_id":"s","permission_mode":"auto","hook_event_name":"PreToolUse","tool_name":"Agent","tool_input":{"description":"d","prompt":"files: %s\\nchange: %s\\nverify: %s\\n\\nDo the work.","subagent_type":"%s","model":"%s"%s},"tool_use_id":"tu%s"}' "$3" "$4" "$5" "$1" "$2" "${6:-}" "$RANDOM$RANDOM" > "$_v2/c.json"
-  ( cd "$_v2p" && _v2e bash "$_V2G" < "$_v2/c.json" >/dev/null 2>"$_v2/err" ); _v2rc=$?; }
+  ( cd "$_v2p" && _v2e bash "$_V2G" < "$_v2/c.json" >"$_v2/out" 2>"$_v2/err" ); _v2rc=$?; }
 _v2start(){ printf '{"session_id":"s","hook_event_name":"SubagentStart","agent_id":"%s","agent_type":"%s"}' "$1" "${2:-crew-backend-expert}" | ( cd "$_v2p" && _v2e bash "$_V2O" >/dev/null 2>&1 ); }
 _v2stop(){ printf '{"session_id":"s","permission_mode":"%s","hook_event_name":"SubagentStop","agent_id":"%s","agent_type":"x","stop_hook_active":false,"agent_transcript_path":"%s/none.jsonl","last_assistant_message":"did it\\n%s"}' "${_V2PM:-auto}" "$1" "$_v2" "${2:-confidence: high}" | ( cd "$_v2p" && _v2e bash "$_V2O" 2>/dev/null ); }
 _v2post(){ printf '{"session_id":"s","hook_event_name":"PostToolUse","tool_name":"Agent","tool_input":{"prompt":"p"},"tool_response":{"status":"completed","agentId":"%s","content":"x"}}' "$1" | ( cd "$_v2p" && _v2e bash "$_V2O" 2>/dev/null ); }
@@ -8251,21 +8351,26 @@ done <<'V2'
 2|crew-backend-expert|sonnet|prisma/schema.prisma|feature|none|
 2|crew-backend-expert|sonnet|config/SECRETS_local.env|text|none|
 0|crew-backend-expert|sonnet|docs/author.md, src/cryptography/hash.ts|feature|none|
-2|crew-backend-expert|sonnet|legal/terms-of-service.md|text|none|
-2|crew-backend-expert|sonnet|docs/privacy-policy.md|text|none|
+0|crew-backend-expert|sonnet|legal/terms-of-service.md|text|none|
+0|crew-backend-expert|sonnet|docs/privacy-policy.md|text|none|
 2|crew-backend-expert|sonnet|src/PrivacyController.cs|feature|none|
 2|crew-backend-expert|sonnet|src/consent/banner.ts|feature|none|
-2|crew-backend-expert|sonnet|content/terms/index.md|text|none|
-2|crew-backend-expert|sonnet|fastlane/metadata/en-US/description.txt|text|none|
-2|crew-backend-expert|sonnet|store/listing.md|text|none|
-2|crew-backend-expert|sonnet|metadata/android/en-US/full_description.txt|text|none|
+0|crew-backend-expert|sonnet|content/terms/index.md|text|none|
+0|crew-backend-expert|sonnet|fastlane/metadata/en-US/description.txt|text|none|
+0|crew-backend-expert|sonnet|store/listing.md|text|none|
+0|crew-backend-expert|sonnet|metadata/android/en-US/full_description.txt|text|none|
 0|crew-backend-expert|opus|legal/terms-of-service.md|text|none|
+2|crew-backend-expert|haiku|legal/terms-of-service.md|text|none|
+2|crew-backend-expert|sonnet|legal/terms-of-service.md, src/legal/consent.ts|text|none|
+2|crew-backend-expert|sonnet|legal/terms-of-service.md|feature|none|
 0|crew-backend-expert|sonnet|src/searchTerms.ts|feature|none|
 0|crew-backend-expert|sonnet|src/store/index.ts|feature|none|
 0|crew-backend-expert|sonnet|fastlane/metadata/en-US/images/icon.png|text|none|
 0|crew-backend-expert|sonnet|docs/legalese.md|text|none|
 0|crew-backend-expert|sonnet|src/ui/a.ts|feature|none|
-2|crew-backend-expert|haiku|src/ui/a.ts|text|none|
+0|crew-backend-expert|haiku|src/ui/a.ts|text|none|
+2|crew-backend-expert|haiku|src/ui/a.ts|fix-known|none|
+0|crew-commit-agent|haiku|src/ui/a.ts|fix-known|none|
 0|crew-backend-expert|sonnet|src/ui/a.ts|text|none|
 0|crew-backend-expert|haiku|src/ui/a.ts|text|true|
 0|crew-test-expert|haiku|src/ui/**|test-run|true|
@@ -8297,9 +8402,48 @@ done <<'V2'
 0|crew-backend-expert|sonnet|src/ui/a.ts|feature|git commit -m x|
 0|crew-backend-expert|sonnet|src/ui/a.ts|feature|rm -rf /|
 V2
-if [ "$_v2n" != 60 ]; then fail "FIXTURE: the card table has $_v2n rows, not 60"
-elif [ -z "$_v2b" ]; then pass "the card decides the model: critical work (payments/, migrations, a .sql file, AuthController, userAuth, schema.prisma, and the four critical kinds of change) is refused below opus and passes on opus; verify none is refused on haiku; test-run passes on haiku, also on a critical path; test-write, audit and review are held to sonnet in normal work and to opus in critical work; critical work with a verify command is refused when the call says run_in_background true, and passes with false and with the field left out (a session whose Agent tool has no such field cannot send it); a card that is not one is refused; a verify command never refuses the call, whatever it is (the hook decides whether to run it) (47 rows)"
+if [ "$_v2n" != 65 ]; then fail "FIXTURE: the card table has $_v2n rows, not 65"
+elif [ -z "$_v2b" ]; then pass "the card decides the model: critical work (payments/, migrations, a .sql file, AuthController, userAuth, schema.prisma, and the four critical kinds of change) is refused below opus and passes on opus; a text change whose every file is prose (legal, privacy, terms, a store listing) passes on sonnet and is refused on haiku, and one card file that is not prose holds it to opus; verify none is refused on haiku, except for a text change and the commit agent; test-run passes on haiku, also on a critical path; test-write, audit and review are held to sonnet in normal work and to opus in critical work; critical work with a verify command is refused when the call says run_in_background true, and passes with false and with the field left out (a session whose Agent tool has no such field cannot send it); a card that is not one is refused; a verify command never refuses the call, whatever it is (the hook decides whether to run it) ($_v2n rows)"
 else fail "card table:$_v2b"; fi
+# The card's floor is also the default: a call above it runs ON it (the gate hands back the same input with `model`
+# replaced, PreToolUse updatedInput), unless the card's next line says `above: <why>`. Measured on a real project:
+# 73 of 186 crew tasks ran above their floor and none on haiku; measured on Claude Code 2.1.299: the rewrite holds.
+_v2new; _v2b=""
+_v2last(){ ls -t "$_v2p/.claude/state/crew-model/pending/"* 2>/dev/null | head -1; }
+_v2call crew-review-agent opus 'src/ui/a.ts' audit none ',"run_in_background":true,"isolation":"worktree"'
+_o="$(cat "$_v2/out" 2>/dev/null)"; _r="$(_v2last)"
+case "$_o" in *'"permissionDecision":"allow"'*'"updatedInput":'*) ;; *) _v2b="$_v2b [above the floor, no above: -> no rewrite: rc=$_v2rc ${_o:0:80}]" ;; esac
+_u="$(printf '%s' "$_o" | awk -v op=get -v path=hookSpecificOutput.updatedInput -f "$ROOT/eval/lib/settings-json.awk" - 2>/dev/null | tr -d ' \n')"
+case "$_u" in *'"model":"sonnet"'*'"run_in_background":true'*'"isolation":"worktree"'*|*'"run_in_background":true'*'"isolation":"worktree"'*) ;; *) _v2b="$_v2b [the rewrite lost a field: ${_u:0:120}]" ;; esac
+case "$_u" in *'"model":"sonnet"'*) ;; *) _v2b="$_v2b [the rewrite is not to the floor: ${_u:0:120}]" ;; esac
+{ grep -q '^tier=sonnet$' "$_r" && grep -q '^lowered=opus$' "$_r"; } 2>/dev/null || _v2b="$_v2b [record: $(tr '\n' ' ' < "$_r" 2>/dev/null | cut -c1-80)]"
+rm -f "$_v2p/.claude/state/crew-model/pending/"*
+_v2call crew-review-agent opus 'src/ui/a.ts' audit 'none\nabove: the audit covers the whole session layer'
+{ [ "$_v2rc" = 0 ] && [ ! -s "$_v2/out" ] && grep -q '^tier=opus$' "$(_v2last)"; } 2>/dev/null || _v2b="$_v2b [above: did not keep opus: rc=$_v2rc $(head -c 60 "$_v2/out")]"
+rm -f "$_v2p/.claude/state/crew-model/pending/"*
+_v2call crew-review-agent sonnet 'src/ui/a.ts' audit none
+{ [ "$_v2rc" = 0 ] && [ ! -s "$_v2/out" ]; } || _v2b="$_v2b [a call on its floor was rewritten: $(head -c 60 "$_v2/out")]"
+_v2call crew-frontend-expert sonnet 'src/ui/a.ts' text 'test -d src'
+case "$(tr -d ' ' < "$_v2/out")" in *'"model":"haiku"'*) ;; *) _v2b="$_v2b [work a command checks did not go to haiku: $(head -c 80 "$_v2/out")]" ;; esac
+rm -f "$_v2p/.claude/state/crew-model/pending/"*
+# Not a crew agent: no card and no floor, only the expensive end is held. opus or fable with no `above:` line in the
+# task runs on sonnet; with one it stays; a call that names no model, or sonnet, is not touched.
+_v2o(){ printf '{"session_id":"s","permission_mode":"auto","hook_event_name":"PreToolUse","tool_name":"Agent","tool_input":{"description":"d","prompt":"Find where X is defined.%s","subagent_type":"general-purpose"%s},"tool_use_id":"tu%s"}' "$1" "$2" "$RANDOM" | ( cd "$_v2p" && _v2e bash "$_V2G" 2>/dev/null ) | tr -d ' '; }
+case "$(_v2o '' ',"model":"opus"')" in *'"updatedInput":'*'"model":"sonnet"'*) ;; *) _v2b="$_v2b [general-purpose on opus was not run on sonnet]" ;; esac
+[ -z "$(_v2o '\nabove: a design question across the codebase' ',"model":"opus"')" ] || _v2b="$_v2b [above: did not keep opus for general-purpose]"
+[ -z "$(_v2o '' ',"model":"sonnet"')" ] || _v2b="$_v2b [a call naming sonnet was rewritten]"
+# Naming no model: general-purpose takes the session's model, so it is given sonnet (CLAUDE_CODE_SUBAGENT_MODEL in
+# the project's settings env did not do it, measured on 2.1.299), unless the session is on haiku; an agent whose
+# definition names its own model is left to it; a project agent with no model line is treated like general-purpose.
+_v2ot(){ printf '{"session_id":"s","permission_mode":"auto","hook_event_name":"PreToolUse","tool_name":"Agent","transcript_path":"%s","tool_input":{"description":"d","prompt":"Find X.","subagent_type":"%s"},"tool_use_id":"tu%s"}' "$2" "$1" "$RANDOM" | ( cd "$_v2p" && _v2e bash "$_V2G" 2>/dev/null ) | tr -d ' '; }
+printf '{"type":"assistant","message":{"model":"claude-opus-5-5"}}\n' > "$_v2/tr-opus.jsonl"; printf '{"type":"assistant","message":{"model":"claude-haiku-5-5"}}\n' > "$_v2/tr-haiku.jsonl"
+mkdir -p "$_v2p/.claude/agents"; printf -- '---\nname: mine\ndescription: x\n---\nDo it.\n' > "$_v2p/.claude/agents/mine.md"; printf -- '---\nname: cheap\nmodel: haiku\n---\nDo it.\n' > "$_v2p/.claude/agents/cheap.md"
+case "$(_v2ot general-purpose "$_v2/tr-opus.jsonl")" in *'"updatedInput":'*'"model":"sonnet"'*) ;; *) _v2b="$_v2b [general-purpose naming no model under opus was not given sonnet]" ;; esac
+[ -z "$(_v2ot general-purpose "$_v2/tr-haiku.jsonl")" ] || _v2b="$_v2b [a haiku session's general-purpose was raised to sonnet]"
+[ -z "$(_v2ot Explore "$_v2/tr-opus.jsonl")$(_v2ot cheap "$_v2/tr-opus.jsonl")" ] || _v2b="$_v2b [an agent with a model of its own was given one]"
+case "$(_v2ot mine "$_v2/tr-opus.jsonl")" in *'"model":"sonnet"'*) ;; *) _v2b="$_v2b [a project agent with no model line was not given sonnet]" ;; esac
+[ -z "$_v2b" ] && pass "above the floor runs on the floor: opus for a sonnet card is rewritten to sonnet with every other field kept and recorded as lowered; 'above: <why>' keeps opus; a call on its floor is untouched; checkable text work goes to haiku; an agent outside the crew asking for opus, or inheriting an opus session's model, runs on sonnet unless its task says why; a haiku session and an agent with its own model are left alone" \
+               || fail "model on the floor:$_v2b"
 # The refusal names the way forward that exists (the field was sent, so it can be sent as false), and a call that
 # left the field out and was launched in the background is told so when the launch comes back.
 _v2new; _v2b=""
@@ -8345,15 +8489,20 @@ done
 _v2new; _v2b=""
 _v2call crew-backend-expert sonnet 'src/ui/a.ts' refactor true; _v2start ag-son
 _v2call crew-backend-expert opus 'src/payments/p.ts' feature none; _v2start ag-op
+_v2call crew-docs-agent sonnet 'legal/terms.md' text none; _v2start ag-txt crew-docs-agent
 _v2write "$_v2p/src/payments/p.ts" ',"agent_id":"ag-son","agent_type":"crew-backend-expert"'; { [ "$_v2wrc" = 2 ] && grep -q 'escalate: src/payments/p.ts' "$_v2/werr" && grep -q '^GUARD (agent model): ' "$_v2/werr"; } || _v2b="$_v2b sonnet-agent-wrote-payments:$_v2wrc"
 _v2write "$_v2p/src/ui/a.ts"        ',"agent_id":"ag-son","agent_type":"crew-backend-expert"'; [ "$_v2wrc" = 0 ] || _v2b="$_v2b sonnet-agent-refused-on-ui:$_v2wrc"
 _v2write "$_v2p/src/payments/p.ts" ',"agent_id":"ag-op","agent_type":"crew-backend-expert"';  [ "$_v2wrc" = 0 ] || _v2b="$_v2b opus-agent-refused:$_v2wrc"
 _v2write "$_v2p/src/payments/p.ts" ',"agent_id":"ag-unknown","agent_type":"general-purpose"'; [ "$_v2wrc" = 2 ] || _v2b="$_v2b unknown-model-wrote-payments:$_v2wrc"
 _v2write "$_v2p/src/payments/p.ts"; [ "$_v2wrc" = 0 ] || _v2b="$_v2b the-session-refused:$_v2wrc"
+# A text change on sonnet writes prose on a critical path; not code there, and not prose for any other change.
+_v2write "$_v2p/legal/terms.md"     ',"agent_id":"ag-txt","agent_type":"crew-docs-agent"';     [ "$_v2wrc" = 0 ] || _v2b="$_v2b sonnet-text-refused-legal-prose:$_v2wrc"
+_v2write "$_v2p/src/payments/p.ts" ',"agent_id":"ag-txt","agent_type":"crew-docs-agent"';     [ "$_v2wrc" = 2 ] || _v2b="$_v2b sonnet-text-wrote-payments-code:$_v2wrc"
+_v2write "$_v2p/legal/terms.md"     ',"agent_id":"ag-son","agent_type":"crew-backend-expert"'; [ "$_v2wrc" = 2 ] || _v2b="$_v2b sonnet-refactor-wrote-legal-prose:$_v2wrc"
 mkdir -p "$_v2/t/subagents"; printf '{"type":"assistant","message":{"model":"claude-opus-5-5"}}\n' > "$_v2/t/subagents/agent-ag-gp.jsonl"
 _v2write "$_v2p/src/payments/p.ts" ',"agent_id":"ag-gp","agent_type":"general-purpose"'; [ "$_v2wrc" = 0 ] || _v2b="$_v2b opus-by-its-transcript-refused:$_v2wrc"
 _V2ENV="CREW_MODEL_ROUTING=off" _v2write "$_v2p/src/payments/p.ts" ',"agent_id":"ag-son","agent_type":"crew-backend-expert"'; [ "$_v2wrc" = 0 ] || _v2b="$_v2b routing-off-refused:$_v2wrc"
-[ -z "$_v2b" ] && pass "an agent on sonnet is refused a write to a critical path its card did not name, and told to report escalate; the same agent writes elsewhere; an agent on opus (by its call, or by its own transcript) writes there; an agent whose model cannot be read does not; the session's own write is not asked; routing off asks nothing" \
+[ -z "$_v2b" ] && pass "an agent on sonnet is refused a write to a critical path its card did not name, and told to report escalate; the same agent writes elsewhere; an agent on opus (by its call, or by its own transcript) writes there; an agent whose model cannot be read does not; the session's own write is not asked; routing off asks nothing; an agent on sonnet with a text card writes prose there (legal/terms.md) and not code, and one with any other card writes neither" \
                || fail "write-time check:$_v2b"
 # B2 (K7): the session's own write to a critical path is not refused, and it is recorded: one NOTE line in the gate
 # log naming the path and the rule. A write elsewhere, one outside the project, and one with routing off add none.
@@ -8387,6 +8536,32 @@ _v2call crew-backend-expert sonnet 'docs/security.md' text none; [ "$_v2rc" = 0
 _v2call crew-backend-expert sonnet 'Api/Services/AuthorService.cs' feature none; [ "$_v2rc" = 0 ] || _v2b="$_v2b AuthorService-read-as-auth"
 [ -z "$_v2b" ] && pass "with no settings file the scan finds the critical paths of a Flutter app and a .NET service (a folder once, not per file; AuthorService and author.md are not auth); crew-model-rules adds and removes paths, at SessionStart too, and the gate follows it" \
                || fail "critical paths:$_v2b"
+# V5: an organisation preset (eval/lib/preset.sh) is a block at the top of crew-model-rules. The project's own lines
+# stay below it and win where the file's rule lets them; applying again changes nothing; an invalid one writes nothing.
+_v2new; _v2b=""; PSH="$ROOT/eval/lib/preset.sh"; _psr="$_v2p/.claude/crew-model-rules"; _psx="$(mktemp -d)"
+printf '# ours\n- infra\n' > "$_psr"
+( cd "$_v2p" && bash "$PSH" --apply strict ) >/dev/null 2>&1 || _v2b="$_v2b [strict did not apply]"
+{ IFS= read -r _l1 < "$_psr"; case "$_l1" in '# ---- preset: strict '*) ;; *) false ;; esac; } || _v2b="$_v2b [first line: ${_l1:-none}]"
+grep -qx '+ deploy' "$_psr" && grep -qx 'floor crew-review-agent \* \* opus' "$_psr" && grep -qx -- '- infra' "$_psr" && grep -qx '# ours' "$_psr" || _v2b="$_v2b [block or own lines missing]"
+[ "$(grep -n -- '^- infra$' "$_psr" | cut -d: -f1)" -gt "$(grep -n '^+ infra$' "$_psr" | cut -d: -f1)" ] || _v2b="$_v2b [the project's own line is not below the block]"
+cp "$_psr" "$_psx/a"; ( cd "$_v2p" && bash "$PSH" --apply strict ) >/dev/null 2>&1; cmp -s "$_psr" "$_psx/a" || _v2b="$_v2b [a second apply changed the file]"
+[ "$( cd "$_v2p" && bash "$PSH" --show )" = strict ] || _v2b="$_v2b [--show]"
+# The gate follows it: review on sonnet is refused under strict, and the project's `- infra` undoes the preset's `+ infra`.
+_v2call crew-review-agent sonnet 'src/ui/a.ts' audit none; { [ "$_v2rc" = 2 ] && grep -q 'floor crew-review-agent \* \* opus' "$_v2/err"; } || _v2b="$_v2b [review on sonnet under strict: rc $_v2rc]"
+_v2call crew-backend-expert sonnet 'infra/main.tf' feature 'test -d src'; [ "$_v2rc" = 0 ] || _v2b="$_v2b [infra stayed critical though the project removed it: rc $_v2rc]"
+_v2call crew-backend-expert sonnet 'deploy/run.sh' feature 'test -d src'; [ "$_v2rc" = 2 ] || _v2b="$_v2b [deploy is not critical under strict: rc $_v2rc]"
+( cd "$_v2p" && bash "$PSH" --apply solo ) >/dev/null 2>&1; ! grep -q '^+ deploy' "$_psr" && grep -qx 'verify npm test' "$_psr" && grep -qx -- '- infra' "$_psr" || _v2b="$_v2b [switching to solo left the old block or lost the own line]"
+# Invalid presets: each refused with rc 1, the file untouched.
+cp "$_psr" "$_psx/b"; _bad=0
+for _bj in '{"name":"x","oops":1}' '{"name":"x","floors":[["a","b","c","gpt"]]}' '{"name":"x","verify":["npm test; rm -rf /"]}' '{"name":"X Y"}' 'not json'; do
+  printf '%s' "$_bj" > "$_psx/bad.json"; ( cd "$_v2p" && bash "$PSH" --apply "$_psx/bad.json" ) >/dev/null 2>&1; _rc=$?
+  { [ "$_rc" = 1 ] && cmp -s "$_psr" "$_psx/b"; } || { _bad=$((_bad+1)); _v2b="$_v2b [accepted or wrote: $_bj rc $_rc]"; }
+done
+( cd "$_v2p" && bash "$PSH" --apply no-such-preset ) >/dev/null 2>&1 && _v2b="$_v2b [a preset that is not there applied]"
+_dl="$(cd "$_v2p" && bash "$ROOT/eval/doctor.sh" . 2>/dev/null | grep -c 'organisation preset: solo')"; [ "$_dl" = 1 ] || _v2b="$_v2b [doctor did not name the preset: $_dl]"
+[ -z "$_v2b" ] && pass "an organisation preset is a block at the top of crew-model-rules: the project's own lines stay below it and win (its '- infra' undoes the preset's), the gate follows it (review held to opus under strict), applying again changes no byte, switching replaces the block, five invalid presets and a missing one write nothing, and doctor names the one in force" \
+               || fail "presets:$_v2b"
+rm -rf "$_psx"
 
 # ---- 3. verify, one fix inside the agent, then one model up, then the user -----------------------------------
 _v2new; _v2b=""
@@ -8699,12 +8874,41 @@ if command -v iconv >/dev/null 2>&1; then printf '%s' "$_o" | iconv -f UTF-8 -t
 [ "${#_o}" -le 2600 ] || _v2b="$_v2b [the reason is ${#_o} bytes]"
 [ -z "$_v2b" ] && pass "the one fix is asked once and only where it can be counted: a red verify under stop_hook_active ends as fail with no block, and so does one whose counter cannot be written (both recorded as fail with no fix); the output handed to the agent is cut at 1500 bytes without splitting a character" \
                || fail "the stop hook's one fix:$_v2b"
+# A verify command in a sub-folder, `cd <folder> && <command>` (41 of 71 refusals on one project): judged as the
+# command and run in the folder with no chain. A folder outside the project, one that does not exist, or a second
+# chain after it stays refused. A card that says `none (why)` is none.
+_v2new; _v2b=""; mkdir -p "$_v2p/.claude/state/crew-model/agents" "$_v2p/sub"; : > "$_v2p/sub/ok.flag"
+printf 'verify test -f ok.flag\n' > "$_v2p/.claude/crew-model-rules"
+_v2plant c1 'cd sub && test -f ok.flag'; _v2stop c1 >/dev/null
+[ "$(_v2row c1 | cut -d/ -f2)" = pass ] || _v2b="$_v2b [cd sub && test: $(_v2row c1)]"
+for _cc in 'c2:cd ../sub && test -f ok.flag' 'c3:cd nowhere && test -f ok.flag' 'c4:cd sub && test -f ok.flag && touch x' 'c5:cd /tmp && test -f ok.flag'; do
+  _v2plant "${_cc%%:*}" "${_cc#*:}"; _v2stop "${_cc%%:*}" >/dev/null
+  [ "$(_v2row "${_cc%%:*}" | cut -d/ -f2)" = blocked ] || _v2b="$_v2b [${_cc#*:} was not refused: $(_v2row "${_cc%%:*}")]"
+done
+[ ! -e "$_v2p/sub/x" ] || _v2b="$_v2b [a chain after the folder ran]"
+_v2call crew-backend-expert sonnet 'src/ui/a.ts' text 'none (the author ran it)'
+grep -q '^verify=none$' "$(ls -t "$_v2p/.claude/state/crew-model/pending/"* 2>/dev/null | head -1)" 2>/dev/null || _v2b="$_v2b [none (why) was not none]"
+# doctor lists the commands refused for not being allowed, with the line that allows them, and not a sentence
+mkdir -p "$_v2p/.claude/state/crew-model/results"
+printf 'verify=blocked\nwhynot=it is not a test or build command on this hook'"'"'s list, and no rule names it\ncmd=node tools/build.mjs --check\n' > "$_v2p/.claude/state/crew-model/results/r1"
+printf 'verify=blocked\nwhynot=it is not a test or build command on this hook'"'"'s list, and no rule names it\ncmd=actionlint (if installed)\n' > "$_v2p/.claude/state/crew-model/results/r2"
+# The other way of saying "nothing allows it" (default mode), for a command in a sub-folder: counted, and offered as
+# the rule that matches it, the command without its `cd` (measured on Windows: the line with the cd in it did not).
+printf "verify=blocked\nwhynot=in 'default' mode Claude Code does not run a command the user has not allowed, so neither does a hook\ncmd=cd sub && ./check.sh\n" > "$_v2p/.claude/state/crew-model/results/r3"
+if [ -f "$ROOT/eval/doctor.sh" ]; then
+  mkdir -p "$_v2p/.claude/eval"; cp -R "$ROOT/eval/." "$_v2p/.claude/eval/" 2>/dev/null
+  _dd="$( cd "$_v2p" && env -u CLAUDE_CODE_SUBAGENT_MODEL CREW_NO_STAR=1 CREW_LANG=en CLAUDE_CONFIG_DIR="$_v2/home" bash .claude/eval/doctor.sh 2>/dev/null | grep 'verify commands never run' )"
+  case "$_dd" in *'3 time(s)'*'verify node tools/build.mjs --check'*'verify ./check.sh'*) case "$_dd" in *actionlint*|*'verify cd '*) _v2b="$_v2b [doctor offered a sentence, or a rule with its cd, as a rule: ${_dd}]" ;; esac ;;
+    *) _v2b="$_v2b [doctor: ${_dd:-no line}]" ;; esac
+fi
+[ -z "$_v2b" ] && pass "a verify command in a sub-folder runs there: 'cd sub && <command>' passes, judged as the command alone; a folder outside the project or missing, an absolute one, and a chain after it stay refused; 'none (why)' is none; doctor offers the refused commands, in either mode, as the rule that matches them (no cd prefix), not a sentence" \
+               || fail "verify in a sub-folder:$_v2b"
 # Which record is whose. Two calls of one type made together are matched by order, which can be wrong; the card in
 # the agent's own transcript settles it.
 _v2new; _v2b=""; _v2st="$_v2p/.claude/state/crew-model"
 _v2tr(){ printf '{"type":"user","message":{"role":"user","content":"files: %s\\nchange: %s\\nverify: %s\\n\\nDo the work."}}\n' "$2" "$3" "$4" > "$_v2/tr-$1.jsonl"; }
 _v2stopt(){ printf '{"session_id":"s","permission_mode":"auto","hook_event_name":"SubagentStop","agent_id":"%s","agent_type":"x","stop_hook_active":false,"agent_transcript_path":"%s/tr-%s.jsonl","last_assistant_message":"did it\\nconfidence: high"}' "$1" "$_v2" "$1" | ( cd "$_v2p" && _v2e bash "$_V2O" 2>/dev/null ); }
-_v2call crew-backend-expert haiku 'src/ui/one.ts' text 'test -f one.ok'; _v2call crew-backend-expert opus 'src/ui/two.ts' feature 'test -f two.ok'
+_v2call crew-backend-expert haiku 'src/ui/one.ts' text 'test -f one.ok'; _v2call crew-backend-expert opus 'src/ui/two.ts' feature 'test -f two.ok\nabove: two models pending for one agent type'
 # The order the two records are taken in is made known: the haiku call's first, so each agent gets the OTHER's below.
 for _pf in "$_v2st/pending"/*; do if grep -q 'one.ok' "$_pf"; then mv "$_pf" "$_v2st/pending/1-a"; else mv "$_pf" "$_v2st/pending/2-b"; fi; done
 _pt="$(date +%s)"; mv "$_v2st/pending/1-a" "$_v2st/pending/$_pt-a"; mv "$_v2st/pending/2-b" "$_v2st/pending/$_pt-b"
@@ -8739,6 +8943,16 @@ rm -f "$_v2st"/pending/*
 _v2call crew-backend-expert sonnet 'src/ui/race.ts' text true
 _v2start r1 & _v2start r2 & wait
 _rn=0; [ -f "$_v2st/agents/r1" ] && _rn=$((_rn+1)); [ -f "$_v2st/agents/r2" ] && _rn=$((_rn+1)); [ "$_rn" = 1 ] || _v2b="$_v2b [one record, two agents: $_rn got it]"
+ls "$_v2st/agents" 2>/dev/null | grep -qE '\.(take|claim)$' && _v2b="$_v2b a-lock-or-claim-left-behind"
+# A record another hook holds the lock on is not taken: the agent goes on to the next one (the lock is what keeps
+# two agents off one record where a rename alone did not, on Windows).
+rm -rf "$_v2st/agents"; rm -f "$_v2st"/pending/*
+_v2call crew-backend-expert sonnet 'src/ui/l1.ts' text 'test -f l1'; _pl1="$(ls "$_v2st/pending")"
+sleep 1; _v2call crew-backend-expert sonnet 'src/ui/l2.ts' text 'test -f l2'
+mkdir -p "$_v2st/agents"; : > "$_v2st/agents/$_pl1.take"
+_v2start k1
+{ grep -q '^verify=test -f l2' "$_v2st/agents/k1" 2>/dev/null && [ -f "$_v2st/pending/$_pl1" ]; } || _v2b="$_v2b a-locked-record-was-taken-or-the-next-one-was-not"
+rm -f "$_v2st/agents/$_pl1.take" "$_v2st"/pending/*
 # A record nothing claimed within five minutes is nobody's.
 mkdir -p "$_v2st/pending"; printf 'agent=crew-backend-expert\ntier=opus\nchange=text\nrisk=normal\ncard=old\nbg=unset\nesc=-\nverify=none\n' > "$_v2st/pending/1000000000-tuold"
 _v2start o1; { [ ! -f "$_v2st/agents/o1" ] && [ ! -f "$_v2st/pending/1000000000-tuold" ]; } || _v2b="$_v2b a-stale-record-was-given-to-an-agent-or-kept"
@@ -8761,7 +8975,7 @@ _kn="$(awk -F'\t' '$2 == "k2" { print $5 "/" $13 }' "$_v2st/../model-outcomes.ts
 # The gate's and the write-time check's refusals are in the gate log; an agent id that cannot be read does not open
 # the write-time check; run_in_background named twice is refused.
 _v2new; _v2b=""; _V2ENV="CREW_GATE_LOG=$_v2/gl.tsv"; : > "$_v2/gl.tsv"
-_v2call crew-database-expert sonnet 'src/payments/p.ts' feature none; _v2call crew-backend-expert haiku 'src/ui/a.ts' text none
+_v2call crew-database-expert sonnet 'src/payments/p.ts' feature none; _v2call crew-backend-expert haiku 'src/ui/a.ts' fix-known none
 _v2write "$_v2p/src/payments/p.ts" ',"agent_id":"nobody-known","agent_type":"x"'
 _v2write "$_v2p/src/payments/p.ts" ',"agent_id":"","agent_type":"x"'; [ "$_v2wrc" = 2 ] || _v2b="$_v2b an-empty-agent-id-opened-the-check:$_v2wrc"
 _v2write "$_v2p/src/payments/p.ts" ',"agent_id":"a b/../c","agent_type":"x"'; [ "$_v2wrc" = 2 ] || _v2b="$_v2b an-unreadable-agent-id-opened-the-check:$_v2wrc"
@@ -8905,11 +9119,11 @@ _v2call crew-backend-expert sonnet 'src/payments/p.ts' feature none; [ "$_v2rc"
 # What a lowering cannot reach: the floors the card and the agent carry. Two raised floors are planted and lowered;
 # the model below still fails where another rule holds it, and passes where only the lowered floor did.
 _v2new; _v2b=""; mkdir -p "$_v2p/.claude/state"
-printf 'crew-backend-expert\ttext\tnormal\tsonnet\t6\t3\t2026-01-01T00:00:00Z\ncrew-database-expert\tfeature\tnormal\topus\t6\t3\t2026-01-01T00:00:00Z\n' > "$_v2fl"
-_o="$(_v2say '/crew-loosen crew-backend-expert text normal')$(_v2say '/crew-loosen crew-database-expert feature normal')"
+printf 'crew-backend-expert\tfix-known\tnormal\tsonnet\t6\t3\t2026-01-01T00:00:00Z\ncrew-database-expert\tfeature\tnormal\topus\t6\t3\t2026-01-01T00:00:00Z\n' > "$_v2fl"
+_o="$(_v2say '/crew-loosen crew-backend-expert fix-known normal')$(_v2say '/crew-loosen crew-database-expert feature normal')"
 case "$_o" in *'from sonnet to haiku'*'from opus to sonnet'*) ;; *) _v2b="$_v2b FIXTURE:the-two-floors-were-not-lowered" ;; esac
-_v2call crew-backend-expert haiku 'src/ui/f1.ts' text true;   [ "$_v2rc" = 0 ] || _v2b="$_v2b the-lowered-class-is-still-refused-on-haiku:$_v2rc"
-_v2call crew-backend-expert haiku 'src/ui/f1.ts' text none;   [ "$_v2rc" = 2 ] || _v2b="$_v2b verify-none-went-to-haiku-after-a-lowering"
+_v2call crew-backend-expert haiku 'src/ui/f1.ts' fix-known true;   [ "$_v2rc" = 0 ] || _v2b="$_v2b the-lowered-class-is-still-refused-on-haiku:$_v2rc"
+_v2call crew-backend-expert haiku 'src/ui/f1.ts' fix-known none;   [ "$_v2rc" = 2 ] || _v2b="$_v2b verify-none-went-to-haiku-after-a-lowering"
 _v2call crew-database-expert sonnet 'src/ui/f2.ts' feature true; [ "$_v2rc" = 0 ] || _v2b="$_v2b the-lowered-class-is-still-refused-on-sonnet:$_v2rc"
 _v2call crew-database-expert haiku 'src/ui/f2.ts' feature true;  { [ "$_v2rc" = 2 ] && grep -q 'runs on sonnet or above' "$_v2/err"; } || _v2b="$_v2b an-agents-own-floor-was-lowered"
 _v2call crew-database-expert sonnet 'db/migrations/1.sql' feature true ',"run_in_background":false'; [ "$_v2rc" = 2 ] || _v2b="$_v2b critical-work-went-below-opus-after-a-lowering"
@@ -11155,7 +11369,7 @@ sec "== 14c) the 3.0 rename left no old name behind — outside history and the
 if [ -n "$SGR" ] && [ -d "$SGR/packaging" ] && [ -f "$SGR/VERSION" ] && [ -d "$SGR/kit" ] && [ -f "$SGR/packaging/build-plugin.sh" ]; then
   RN_ALLOW='CHANGELOG.md	196	history: every entry before 3.0 keeps the name it shipped under
 README*.md site/content/*/install.md	4	migration: the 2.x plugin note names the plugin to uninstall — one line in each README and install page, EN + TR
-evals/results/*	6	history: recorded eval runs stay byte-for-byte
+evals/results/* packaging/review-bench/cases/* packaging/review-bench/results/* packaging/advisor-bench/results/*	729	history: recorded eval runs, and the review bench'"'"'s cases (past versions of Crewforth'"'"'s own files) and results, and the advisor bench'"'"'s results on them, stay byte-for-byte
 adopt.sh kit/legacy-blobs.tsv	347	migration: finds and moves 2.x names (components, CLAUDE.md, board, auto-mode rules, variables), and the generated list of the 1.x/2.x files it may move
 bin/cli.js	4	migration: add accepts a typed <x>-csk and moves an add record written under the old names
 evals/run.sh	4	compat: reads the 2.x trusted eval parent when the 3.0 one is absent — removed in 4.0
@@ -11375,6 +11589,22 @@ $(printf '%s\n%s\n%s\n' "$_ka" "$_kb" "$_kc" | grep . | head -n 6 | sed "s|$KR/|
     [ -n "$(_dn "$_kd/done-twin.sh")" ] && pass "the installers' closing lines name no kit (EN key and TR value); a planted 'kit/ deleted' is caught" \
       || fail "twin: the closing-line check missed a planted 'kit/ deleted' — it reads nothing"
   fi
+  # adopt.sh --help answers before it reads anything. An unknown flag is ignored further down, so --help used to run
+  # the detection (seen on Windows in the 3.2.0-rc.1 trial). Read in a git repo holding only adopt.sh: no kit/, both
+  # languages, and nothing written. Twin: the same script without its --help branch must not print the usage.
+  _ah_run(){ ( cd "$1" && CREW_LANG=en bash adopt.sh --help; echo "rc=$?"; bash adopt.sh -h --lang tr; echo "rc=$?" ) 2>&1; }
+  mkdir -p "$_kd/ah" "$_kd/ah-twin"; git init -q "$_kd/ah" 2>/dev/null; git init -q "$_kd/ah-twin" 2>/dev/null
+  cp "$KR/adopt.sh" "$_kd/ah/"; sed 's/-h|--help) _hh=1 ;;//' "$KR/adopt.sh" > "$_kd/ah-twin/adopt.sh"
+  _ah="$(_ah_run "$_kd/ah")"; _aht="$(_ah_run "$_kd/ah-twin")"
+  _ahw="$(git -C "$_kd/ah" status --porcelain 2>/dev/null | grep -vx '?? adopt.sh')"
+  case "$_ah" in
+    *'Usage: bash adopt.sh'*'rc=0'*'Kullanım: bash adopt.sh'*'rc=0'*)
+      if [ -n "$_ahw" ]; then fail "adopt.sh --help wrote into the project: $(printf '%s' "$_ahw" | head -n 2 | tr '\n' ' ')"
+      else case "$_aht" in *'Usage: bash adopt.sh'*) fail "twin: adopt.sh without its --help branch still printed the usage — the check reads nothing" ;;
+             *) pass "adopt.sh --help prints its usage in English and Turkish, rc 0, with no kit/ and nothing written (a twin without it does not)" ;; esac
+      fi ;;
+    *) fail "adopt.sh --help did not print its usage in both languages with rc 0: $(printf '%s' "$_ah" | head -n 3 | tr '\n' ' ')" ;;
+  esac
   rm -rf "$_kd"
 else
   skip scope "printed-text wording check skipped (installed project — the installers live in the source repository)"

package/kit/hooks/agent-outcome.sh

@@ -204,27 +204,36 @@ case "$EV" in
     low=9; cands=""; if [ "${BASH_VERSINFO[0]}" -ge 5 ] || { [ "${BASH_VERSINFO[0]}" -eq 4 ] && [ "${BASH_VERSINFO[1]}" -ge 2 ]; }; then printf -v NOWS '%(%s)T' -1; else NOWS="$(date +%s)"; fi
     for f in "$_AMD/pending"/*; do
       [ -f "$f" ] || continue
-      t="${f##*/}"; t="${t%%-*}"; case "$t" in ''|*[!0-9]*) ;; *) if [ $((NOWS - t)) -gt 300 ]; then rm -f "$f" 2>/dev/null; continue; fi ;; esac
+      t="${f##*/}"; t="${t%%-*}"; case "$t" in ''|*[!0-9]*) ;; *) if [ $((NOWS - t)) -gt 300 ]; then rm -f "$f" "$_AMD/agents/${f##*/}.take" 2>/dev/null; continue; fi ;; esac
       _am_rec_get "$f" agent || continue; [ "$_AMV" = "$AT" ] || continue
       cands="$cands$f"$'\n'
       _am_rec_get "$f" tier && { _am_rank "$_AMV"; [ "$_AMR" -lt "$low" ] && low="$_AMR"; }
     done
     [ -n "$cands" ] || exit 0
     mkdir -p "$_AMD/agents" 2>/dev/null || exit 0
-    # TAKEN BY RENAME. Two agents of one type start at the same moment, each in its own hook process, and both see
-    # the same oldest record. A rename either happens or fails, so exactly one of them gets it; the other goes on
-    # to the next record. (Reading it and deleting it afterwards gave both agents the same record.)
-    claim="$_AMD/agents/$AID.claim"; got=0
+    # TAKEN UNDER A LOCK, THEN BY RENAME. Two agents of one type start at the same moment, each in its own hook
+    # process, and both see the same oldest record. (Reading it and deleting it afterwards gave both the same record.)
+    # A rename alone was the lock on POSIX, and was not on Windows: two hook processes both reported the same rename
+    # as done (smoke, windows-latest, "2 got it", twice). Why is not measured; a rename there goes through an open
+    # handle, which two processes can both hold. So each record is first locked by creating <record>.take with
+    # O_EXCL (bash's noclobber, a builtin: no process); only the one that made the lock renames, the others go on to
+    # the next record.
+    claim="$_AMD/agents/$AID.claim"; got=0; take=""
+    set -C
     while IFS= read -r f; do
       [ -n "$f" ] || continue
+      take="$_AMD/agents/${f##*/}.take"
+      { : > "$take"; } 2>/dev/null || { take=""; continue; }
       if mv "$f" "$claim" 2>/dev/null; then got=1; break; fi
+      rm -f "$take" 2>/dev/null; take=""          # locked, but the record had gone (dropped as stale): let it be
     done <<< "$cands"
+    set +C
     [ "$got" = 1 ] || exit 0
     _am_rec_get "$claim" tier; _am_rank "$_AMV"
     if [ "$low" -lt "$_AMR" ]; then _am_name "$low"
       { grep -v '^tier=' "$claim"; printf 'asked=%s\ntier=%s\nambiguous=1\n' "$_AMV" "$_AMN"; } > "$_AMD/agents/$AID" 2>/dev/null
     else { cat "$claim"; printf 'asked=%s\n' "$_AMV"; } > "$_AMD/agents/$AID" 2>/dev/null; fi
-    rm -f "$claim" 2>/dev/null
+    rm -f "$claim" "$take" 2>/dev/null
     exit 0 ;;
 
   SubagentStop)
@@ -290,7 +299,9 @@ case "$EV" in
     RES=none; WHYNOT=""
     if [ -n "$VCMD" ] && [ "$VCMD" != none ]; then
       _json_slice "$INPUT" permission_mode >/dev/null; PM="$_JS"
-      if ! _am_verify_ok "$VCMD" "$PM"; then
+      # `cd <folder> && <command>` is judged as the command and run in the folder (guard-agent-model.sh).
+      _am_verify_split "$VCMD"; VDIR="$_AMVD"; VRUN="$_AMVC"
+      if ! _am_verify_ok "$VRUN" "$PM"; then
         # Judged again here, at the moment it would run: the gate judged it when the agent was called, and this
         # is the place the command actually starts. Not run; nothing is concluded about the work.
         RES=blocked; WHYNOT="$_AMVW"; VRC=0
@@ -305,11 +316,11 @@ case "$EV" in
         # control, which gives a background job its own group in bash (macOS ships no setsid). On Windows the tree
         # is ended by its Windows process id.
         if [ -z "${CREW_VERIFY_NO_SETSID:-}" ] && command -v setsid >/dev/null 2>&1; then
-          ( cd "$_AMP" 2>/dev/null && exec setsid bash -c "$VCMD" ) > "$VF" 2>&1 </dev/null &
+          ( cd "$_AMP/$VDIR" 2>/dev/null && exec setsid bash -c "$VRUN" ) > "$VF" 2>&1 </dev/null &
           VPID=$!
         else
           set -m
-          ( cd "$_AMP" 2>/dev/null && exec bash -c "$VCMD" ) > "$VF" 2>&1 </dev/null &
+          ( cd "$_AMP/$VDIR" 2>/dev/null && exec bash -c "$VRUN" ) > "$VF" 2>&1 </dev/null &
           VPID=$!
           set +m
         fi

package/kit/hooks/board-sync.sh

@@ -79,28 +79,7 @@ MSG="$(cat "$CACHE")
 Board state above is a cached snapshot; $_BS (Bash tool, not PowerShell — or the user can type /crew-board sync) refreshes it."
 
 # ONE PATH, no jq. The JSON is built here on every machine, so a Mac with jq and a Windows box without it emit the
-# same bytes. The escaper is jq-identical, measured on 15 inputs (quote, backslash, tab, CR, C0 controls, DEL,
-# UTF-8, leading/trailing newlines, 5000 chars): 15/15 equal to `jq -cn --arg m`. The one it replaced escaped only
-# the quote, the backslash and the newline, so a tab in an item title produced JSON the CLI cannot parse
-# (measured: jq rc=5) and the board silently vanished from the session — on exactly the no-jq machine this
-# branch existed for. The trailing newline on the printf closes the last record, so a value ending in a newline
-# keeps it. One awk, no subshell per character. A CR that ENDS a line is dropped on purpose: it is a CRLF line
-# ending, not content, and MSYS gawk already drops it on read (text mode) while BSD awk keeps it — measured on
-# Windows. Stripping it here makes every OS emit the same bytes; a CR inside a line is still escaped as \r.
-ESC="$(printf '%s\n' "$MSG" | LC_ALL=C awk 'BEGIN { ORS = ""
-    for (i = 1; i < 32; i++) ctl[i] = sprintf("%c", i)
-    nm[8] = "\\b"; nm[9] = "\\t"; nm[12] = "\\f"; nm[13] = "\\r" }
-  NR > 1 { print "\\n" }
-  { s = $0; sub(/\r$/, "", s)
-    gsub(/\\/, "\\\\", s); gsub(/"/, "\\\"", s)
-    for (i = 1; i < 32; i++) {
-      if (i == 10 || !index(s, ctl[i])) continue
-      r = (i in nm) ? nm[i] : sprintf("\\u%04x", i)
-      out = ""; while ((p = index(s, ctl[i])) > 0) { out = out substr(s, 1, p - 1) r; s = substr(s, p + 1) }
-      s = out s
-    }
-    d = sprintf("%c", 127)
-    if (index(s, d)) { out = ""; while ((p = index(s, d)) > 0) { out = out substr(s, 1, p - 1) "\\u007f"; s = substr(s, p + 1) }; s = out s }
-    print s }')"
+# same bytes, through Crewforth's one escaper (eval/lib/json-string.awk: jq-identical, measured; why, in its header).
+ESC="$(printf '%s\n' "$MSG" | LC_ALL=C awk -f "$HERE/../eval/lib/json-string.awk")"
 printf '{"hookSpecificOutput":{"hookEventName":"SessionStart","additionalContext":"%s"}}\n' "$ESC"
 exit 0