diff --git a/agentkit/.claude-plugin/plugin.json b/agentkit/.claude-plugin/plugin.json
index 9ea636ca..8ec78423 100644
--- a/agentkit/.claude-plugin/plugin.json
+++ b/agentkit/.claude-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "agentkit",
-  "version": "0.9.11",
+  "version": "0.9.12",
   "description": "Board-aware parallel issue and PR review skills, with lifecycle hooks and a per-repository contract.",
   "author": {
     "name": "wrzonance",
diff --git a/agentkit/.codex-plugin/plugin.json b/agentkit/.codex-plugin/plugin.json
index 57947170..fe92f129 100644
--- a/agentkit/.codex-plugin/plugin.json
+++ b/agentkit/.codex-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "agentkit",
-  "version": "0.9.11",
+  "version": "0.9.12",
   "description": "Board-aware parallel issue and PR review skills, with lifecycle hooks and a per-repository contract.",
   "author": {
     "name": "wrzonance",
diff --git a/agentkit/skills/.shared/scripts/agent-run.sh b/agentkit/skills/.shared/scripts/agent-run.sh
index 9d27fda0..d5324d94 100755
--- a/agentkit/skills/.shared/scripts/agent-run.sh
+++ b/agentkit/skills/.shared/scripts/agent-run.sh
@@ -14,19 +14,75 @@ if [[ -z ${BASH_VERSION:-} || ${BASH_VERSINFO[0]:-0} -lt 4 ]]; then
     exit 2
 fi
 
+current_process_start() {
+    local pid=$1 identity
+    if [[ -r /proc/$pid/stat ]] && identity=$(awk '{print $22}' "/proc/$pid/stat" 2> /dev/null) &&
+        [[ $identity =~ ^[0-9]+$ ]]; then
+        printf '%s' "$identity"
+        return 0
+    fi
+    identity=$(LC_ALL=C ps -o lstart= -p "$pid" 2>/dev/null | tr -d '[:space:]') || return 1
+    [[ $identity =~ ^[[:alnum:]:]+$ ]] || return 1
+    printf '%s' "$identity"
+}
+
+epoch_seconds() { date -u +%s 2>/dev/null || printf '0'; }
+
+status_agent_log() {
+    local requested=$1 log fallback_dir last header pid start epoch current elapsed now
+    [[ -f $requested && ! -L $requested && -O $requested ]] || {
+        printf 'agent-run: error: status requires an owned regular log: %s\n' "$requested" >&2; exit 2;
+    }
+    log=$(realpath -e -- "$requested") || exit 2
+    case $log in
+        */.agent/logs/*.log) ;;
+        *)
+            fallback_dir=$(realpath -e -- "${TMPDIR:-/tmp}/agent-logs-$(id -u)" 2>/dev/null || true)
+            [[ -n $fallback_dir && $log == "$fallback_dir/"*.log ]] || {
+                printf 'agent-run: error: status path is not an agent log: %s\n' "$requested" >&2; exit 2;
+            }
+            ;;
+    esac
+    last=$(tail -n 1 -- "$log")
+    if [[ $last =~ ^===\ agent-run\ exited\ rc=([0-9]+)\ after\ [0-9]+s$ ]]; then
+        ((BASH_REMATCH[1] == 0)) && printf 'pass\n' || printf 'fail rc=%s\n' "${BASH_REMATCH[1]}"
+        exit 0
+    fi
+    [[ $last != '=== agent-run interrupted by '* ]] || { printf 'interrupted\n'; exit 0; }
+    header=$(sed -n '2p' -- "$log")
+    pid=$(sed -n 's/.* pid=\([0-9][0-9]*\) .*/\1/p' <<< "$header")
+    start=$(sed -n 's/.* process-start=\([^ ]*\) .*/\1/p' <<< "$header")
+    epoch=$(sed -n 's/.* epoch=\([0-9][0-9]*\) .*/\1/p' <<< "$header")
+    current=$(current_process_start "$pid" 2>/dev/null || true)
+    if [[ -n $pid && -n $start && $current == "$start" && $epoch =~ ^[0-9]+$ ]]; then
+        now=$(epoch_seconds); elapsed=$((now - epoch)); ((elapsed >= 0)) || elapsed=0
+        printf 'running pid=%s elapsed=%ss\n' "$pid" "$elapsed"
+    else
+        printf 'interrupted\n'
+    fi
+    exit 0
+}
+
+if [[ ${1:-} == status ]]; then
+    (($# == 2)) || { printf 'agent-run: error: usage: agent-run.sh status LOG\n' >&2; exit 2; }
+    status_agent_log "$2"
+fi
+
 usage() {
     cat <<'EOF'
-Usage: agent-run.sh [--dir PATH] [--label NAME] [--resolve NAME] [--force] [--summary] [--only NAME[,NAME...]]
+Usage: agent-run.sh status LOG
+       agent-run.sh [--dir PATH] [--label NAME] [--resolve NAME] [--force] [--summary] [--only NAME[,NAME...]]
                     [--baseline-ref REF --baseline-path PATH --baseline-id ID]
                     (--cmd NAME | [--] <command> ...)
 
 Runs one command with a sandbox-safe environment and a compact result summary.
+  status LOG     Print running, pass, fail, or interrupted for an agent log.
   --dir PATH     Working directory for the command (default: current directory).
                  --repo-root PATH is a silent alias, accepted for
                  compatibility with the kit's other checkout-path helpers.
   --label NAME   Label used in the log file name (default: the command's basename).
   --force        Require fresh execution, including recovery of unknown evidence.
-                 An identical in-flight local command still returns its handle.
+                 An identical in-flight command is still refused with its log.
   --summary      End with status, exit code, duration, log path, digest, and receipt.
   --verification-key  Read-only query for one local, generic, full-checkout
                  command. Prints only its current fingerprint; creates no execution
@@ -41,8 +97,8 @@ Runs one command with a sandbox-safe environment and a compact result summary.
   --if-declared  With --cmd, exit 0 quietly when the repository declares no such
                  command. For a command a skill treats as optional.
   --resolve NAME Query a named command without executing it. Prints declared,
-                 runner, or unresolved and exits 0, 4, or 3 respectively; exit 2
-                 is reserved for a fatal unsupported-interpreter guard.
+                 runner, or unresolved and exits 0, 4, or 3 respectively. The
+                 fatal unsupported-interpreter guard exits 2.
   --cmd NAME     Run the command this repository declares under that name, instead
                  of spelling one out. Repeatable: each --cmd runs only after the
                  previous one exits 0 (re-execs itself for the rest); --if-declared
@@ -93,7 +149,8 @@ Output:
   verification miss: no reusable evidence; a following command executes freshly,
                      or a completed result was not stored for later reuse
   verification current/reused: prior evidence, never a fresh PASS
-  verification running/unknown: durable handle and exit 75; inspect before retry
+  active duplicate exits 2 with its original log; wait instead of relaunching
+  verification unknown: unknown abandoned handle exits 75; inspect before retry
 
 Examples:
   agent-run.sh --cmd test
@@ -117,6 +174,7 @@ failure_result() {
             "${failure_evidence:-${log_file:-stderr}}" "$failure_state" "$failure_action" >&2
     fi
     if declare -F cleanup_suite_run >/dev/null; then cleanup_suite_run; fi
+    if declare -F cleanup_active_run >/dev/null; then cleanup_active_run; fi
     if ((summary_ready)); then
         printf 'agent-run-summary status=%s rc=%s duration_seconds=%s log=%q log-sha256=%s receipt=%q\n' \
             "$summary_status" "$status" "$elapsed" "$log_file" "$log_sha256" "$log_sha256_receipt"
@@ -162,6 +220,7 @@ finish() {
     [[ -z ${verification_fd:-} ]] || exec {verification_fd}>&-
     if ((rc == 0)) && ((${#remaining_queue[@]})); then
         if declare -F cleanup_suite_run >/dev/null; then cleanup_suite_run; fi
+        if declare -F cleanup_active_run >/dev/null; then cleanup_active_run; fi
         build_chain_argv
         exec "$0" "${chain_argv[@]}"
     fi
@@ -1051,7 +1110,8 @@ apply_test_focus() {
 # --------------------------------------------------------------------- logs ---
 choose_log() {
     local log_dir stamp log
-    if [[ -n $git_top ]] && dir_writable "$git_top/.agent/logs"; then
+    if [[ -n $git_top && ! -L $git_top/.agent && ! -L $git_top/.agent/logs ]] &&
+        dir_writable "$git_top/.agent/logs"; then
         log_dir=$git_top/.agent/logs
     else
         # Failing commands routinely echo tokens and credentialed URLs into these
@@ -1083,22 +1143,10 @@ concurrent_suites=1
 timeout_scale=1
 suite_marker_dir=${TMPDIR:-/tmp}/agent-run-suites-$(id -u)
 
-current_process_start() {
-    local pid=$1
-    if [[ -r /proc/$pid/stat ]]; then
-        awk '{print $22}' "/proc/$pid/stat" 2> /dev/null
-    else
-        # macOS has no /proc; kill -0 is the portable liveness fallback. The
-        # marker is still short-lived and is removed by the EXIT trap.
-        kill -0 "$pid" 2> /dev/null || return 1
-        printf 'alive'
-    fi
-}
-
 suite_marker_live() {
     local marker=$1 pid start current
     read -r pid start < "$marker" 2> /dev/null || return 1
-    [[ $pid =~ ^[0-9]+$ && ($start =~ ^[0-9]+$ || $start == alive) ]] || return 1
+    [[ $pid =~ ^[0-9]+$ && $start =~ ^[[:alnum:]:]+$ ]] || return 1
     current=$(current_process_start "$pid")
     [[ -n $current && $current == "$start" ]]
 }
@@ -1116,7 +1164,7 @@ register_suite_run() {
         die "cannot create active-suite marker in $suite_marker_dir"
     pid=$$
     start=$(current_process_start "$pid")
-    [[ $start =~ ^[0-9]+$ || $start == alive ]] || {
+    [[ $start =~ ^[[:alnum:]:]+$ ]] || {
         rm -f -- "$marker"
         die "cannot identify active-suite process $pid"
     }
@@ -1146,6 +1194,40 @@ cleanup_suite_run() {
     suite_marker=''
 }
 
+active_run_handle='' active_run_fd='' active_run_owned=0
+claim_active_run() {
+    local root key prior
+    [[ -n ${git_top:-} && -z $verification_handle ]] || return 0
+    command -v flock >/dev/null || return 0
+    [[ ! -L $git_top/.agent ]] || return 0
+    root=$git_top/.agent/run-records
+    assert_private_dir "$root"
+    key=$(printf '%s\0' "$work_dir" "${cmd[@]}" | sha256sum | awk '{print $1}')
+    active_run_handle=$root/$key
+    assert_private_dir "$active_run_handle"
+    [[ ! -L $active_run_handle/lock && ! -L $active_run_handle/running ]] ||
+        refuse_boundary "active run record is a symlink: $active_run_handle"
+    exec {active_run_fd}>"$active_run_handle/lock" || refuse_boundary "cannot open active run lock: $active_run_handle/lock"
+    if ! flock -n "$active_run_fd"; then
+        IFS= read -r prior < "$active_run_handle/running" 2>/dev/null || true
+        [[ -n $prior ]] || prior=$active_run_handle
+        failure_class=usage failure_state=command-already-running failure_action=wait-for-existing-log
+        failure_evidence=$prior
+        printf 'agent-run: already running: %s; wait for its exited line or run status on it\n' "$prior" >&2
+        finish 2
+    fi
+    printf '%s\n' "$log_file" > "$active_run_handle/running"
+    active_run_owned=1
+}
+
+# shellcheck disable=SC2329  # Invoked indirectly by the EXIT trap.
+cleanup_active_run() {
+    ((active_run_owned)) || return 0
+    rm -f -- "$active_run_handle/running" 2>/dev/null || true
+    [[ -z $active_run_fd ]] || exec {active_run_fd}>&-
+    active_run_owned=0
+}
+
 # A worker may ask for one failed verification to be checked against the chain
 # base. The source path must be the same blob at both commits, and the base
 # checkout must produce matching failure evidence. Deliberately opt-in: ordinary
@@ -1647,9 +1729,16 @@ claim_verification() {
     [[ ! -L $verification_handle/running ]] || refuse_boundary "verification running record is a symlink: $verification_handle/running"
     exec {verification_fd}>"$verification_handle/lock" || refuse_boundary "cannot open verification lock: $verification_handle/lock"
     if ! flock -n "$verification_fd"; then
+        local running_log=''
+        IFS= read -r running_log < "$verification_handle/running" 2>/dev/null || true
         failure_evidence=$verification_handle
-        failure_state=verification-running
-        failure_action=inspect-running-handle-before-retry
+        failure_state=command-already-running
+        failure_action=wait-for-existing-log
+        if [[ $running_log == "$git_top/.agent/logs/"*.log ]]; then
+            failure_evidence=$running_log
+            printf 'agent-run: already running: %s; wait for its exited line or run status on it\n' "$running_log" >&2
+            finish 2
+        fi
         printf 'agent-run: verification running: handle=%s\n' "$verification_handle"
         finish 75
     fi
@@ -1822,25 +1911,27 @@ fi
 
 log_file=$(choose_log)
 [[ -z $verification_handle ]] || printf '%s\n' "$log_file" > "$verification_handle/running"
+claim_active_run
 register_suite_run
 trap failure_result EXIT
+process_start=$(current_process_start "$$") || die "cannot identify active process $$"
 
 # Announce the log before captured output makes a long run look hung.
-if ((summary_cmd)); then
-    printf 'running: %s (wait for the terminal agent-run-summary marker)\n' "$cmd_str" >&2
-else
-    printf 'running: %s\n  log: %s (grows while this runs; tail it instead of waiting blind)\n' \
-        "$cmd_str" "$log_file" >&2
-fi
-printf '  a log with no "=== agent-run exited" line has NOT finished\n' >&2
+printf 'running: %s\n  log: %s\n' "$cmd_str" "$log_file" >&2
+printf '  if this call returns before "=== agent-run exited", the run is still going: resume this same call; never relaunch; status: %q status %q\n' \
+    "$0" "$log_file" >&2
 
 # The closing marker distinguishes completed logs; exclude bookkeeping lines.
 readonly LOG_HEADER_LINES=2
+log_head=none log_clean=no
+[[ -z $git_top ]] || log_head=$(git -C "$git_top" rev-parse --verify -q HEAD 2> /dev/null) || log_head=none
+[[ -z $git_top || -n $(git -C "$git_top" status --porcelain --untracked-files=no 2> /dev/null || printf x) ]] || log_clean=yes
 {
     printf '=== agent-run %s\n' "$cmd_str"
-    printf '=== started %s  pid=%s  cwd=%s  concurrent-suites=%s\n' \
-        "$(date -u +%Y-%m-%dT%H:%M:%SZ 2> /dev/null || printf 'unknown')" "$$" "$work_dir" \
-        "$concurrent_suites"
+    printf '=== started %s  pid=%s  process-start=%s  epoch=%s  cwd=%s  concurrent-suites=%s  head=%s  tracked-clean=%s\n' \
+        "$(date -u +%Y-%m-%dT%H:%M:%SZ 2> /dev/null || printf 'unknown')" "$$" \
+        "$process_start" "$(epoch_seconds)" "$work_dir" "$concurrent_suites" \
+        "$log_head" "$log_clean"
 } > "$log_file"
 
 started_at=$SECONDS
@@ -1859,7 +1950,10 @@ trap 'log_interrupted SIGINT' INT
 trap 'log_interrupted SIGTERM' TERM
 rc=0
 attempt_start_line=3
-(cd -- "$work_dir" && exec "${cmd[@]}") >> "$log_file" 2>&1 || rc=$?
+(
+    [[ -z $active_run_fd ]] || exec {active_run_fd}>&-
+    cd -- "$work_dir" && exec "${cmd[@]}"
+) >> "$log_file" 2>&1 || rc=$?
 load_flake_retry=0
 if ((rc != 0)) && probe_timeout_load_flake "$log_file"; then
     load_flake_retry=1
@@ -1867,7 +1961,10 @@ if ((rc != 0)) && probe_timeout_load_flake "$log_file"; then
         "$concurrent_suites" >> "$log_file"
     attempt_start_line=$(($(wc -l < "$log_file" | tr -d '[:space:]') + 1))
     rc=0
-    (cd -- "$work_dir" && exec "${cmd[@]}") >> "$log_file" 2>&1 || rc=$?
+    (
+        [[ -z $active_run_fd ]] || exec {active_run_fd}>&-
+        cd -- "$work_dir" && exec "${cmd[@]}"
+    ) >> "$log_file" 2>&1 || rc=$?
 fi
 trap - INT TERM
 elapsed=$((SECONDS - started_at))
diff --git a/agentkit/skills/.shared/scripts/bootstrap-repo.sh b/agentkit/skills/.shared/scripts/bootstrap-repo.sh
index 0e2cced5..06501654 100755
--- a/agentkit/skills/.shared/scripts/bootstrap-repo.sh
+++ b/agentkit/skills/.shared/scripts/bootstrap-repo.sh
@@ -389,9 +389,8 @@ done
 # infer from the forge. Keep the proposal visible until a valid declaration is
 # selected; a carried declaration then appears once as active config below.
 review_providers_declared=0
-if [[ -f $repo_root/.agent/config.env && $reset -eq 0 ]]; then
-    if grep -qE '^[[:space:]]*AGENT_REVIEW_PROVIDERS=(coderabbit|github-code-quality|none|coderabbit,github-code-quality|github-code-quality,coderabbit)[[:space:]]*$' \
-        "$repo_root/.agent/config.env" 2> /dev/null; then
+if [[ -f $repo_root/.agent/config.env && $reset -eq 0 && -x $resolver ]]; then
+    if "$resolver" --repo-root "$repo_root" --get AGENT_REVIEW_PROVIDERS > /dev/null 2>&1; then
         review_providers_declared=1
     fi
 fi
@@ -488,7 +487,8 @@ fi
     fi
     if ((review_providers_declared == 0)); then
         printf '\n# Automated review providers expected on pull requests. Choose one or more.\n'
-        printf '# Supported choices: coderabbit, github-code-quality, or none (none is exclusive).\n'
+        printf '# Known providers get provider-specific settlement. Any other valid name is observed\n'
+        printf '# and triaged on the merits, and is never triggered. Use none alone to disable reviews.\n'
         printf '# AGENT_REVIEW_PROVIDERS=\n'
     fi
     if [[ -n $proposal_inventory ]]; then
diff --git a/agentkit/skills/.shared/scripts/lib/review-provider-catalog.sh b/agentkit/skills/.shared/scripts/lib/review-provider-catalog.sh
index e45d2130..88d632fd 100644
--- a/agentkit/skills/.shared/scripts/lib/review-provider-catalog.sh
+++ b/agentkit/skills/.shared/scripts/lib/review-provider-catalog.sh
@@ -1,9 +1,8 @@
 #!/usr/bin/env bash
 # Single source of review-provider capabilities and canonical identities.
 
-# Canonical accepted provider identities, in display order. Every function in
-# this file recognizes exactly these values; review_provider_names() is what
-# a rejection message names instead of leaving a caller to read this file.
+# Provider-specific identities, in display order. Names outside this list use
+# the generic observe-only defaults when they match the declaration grammar.
 #
 # Not `readonly`: this file is sourced more than once within a single process
 # in practice (the pr-to-green transition engine sources it directly, then again
@@ -19,12 +18,41 @@ review_provider_names() {
     printf '%s\n' "$out"
 }
 
+review_provider_name_valid() {
+    [[ ${1:-} =~ ^[a-z][a-z0-9-]*$ && $1 != coderabbitai ]]
+}
+
+review_provider_login_alias_reserved() {
+    case ${1,,} in
+        coderabbitai|github-code-quality) return 0 ;;
+        *) return 1 ;;
+    esac
+}
+
+review_provider_known() {
+    case ${1:-} in
+        coderabbit|github-code-quality|none) return 0 ;;
+        *) return 1 ;;
+    esac
+}
+
+review_provider_override_login() {
+    local provider=$1 suffix key value
+    suffix=${provider^^}
+    suffix=${suffix//-/_}
+    key=AGENT_REVIEW_PROVIDER_${suffix}_LOGIN
+    value=${!key-}
+    [[ -n $value && $value =~ ^[A-Za-z0-9]([A-Za-z0-9_.-]{0,37}[A-Za-z0-9])?$ ]] || return 1
+    review_provider_login_alias_reserved "$value" && return 1
+    printf '%s\n' "${value,,}"
+}
+
 review_provider_mode() {
     case ${1:-} in
         coderabbit) printf '%s\n' triggerable ;;
         github-code-quality) printf '%s\n' observe-only ;;
         none) printf '%s\n' disabled ;;
-        *) return 1 ;;
+        *) review_provider_name_valid "${1:-}" && printf '%s\n' observe-only || return 1 ;;
     esac
 }
 
@@ -33,7 +61,7 @@ review_provider_lifecycle() {
         coderabbit) printf '%s\n' reply-settlement ;;
         github-code-quality) printf '%s\n' provider-rescan ;;
         none) printf '%s\n' disabled ;;
-        *) return 1 ;;
+        *) review_provider_name_valid "${1:-}" && printf '%s\n' generic-settlement || return 1 ;;
     esac
 }
 
@@ -42,18 +70,44 @@ review_provider_login() {
         coderabbit) printf '%s\n' coderabbitai ;;
         github-code-quality) printf '%s\n' github-code-quality ;;
         none) printf '%s\n' none ;;
-        *) return 1 ;;
+        *)
+            review_provider_name_valid "${1:-}" || return 1
+            review_provider_override_login "$1" 2>/dev/null || printf '%s\n' "$1"
+            ;;
     esac
 }
 
 review_provider_from_login() {
-    case ${1,,} in
+    local login=${1,,} provider expected bot_suffix=0
+    local -a declared_providers=()
+    case $login in
         coderabbitai|coderabbitai\[bot\]) printf '%s\n' coderabbit ;;
         github-code-quality|github-code-quality\[bot\]) printf '%s\n' github-code-quality ;;
-        *) return 1 ;;
+        *)
+            [[ $login == *'[bot]' ]] && bot_suffix=1
+            login=${login%\[bot\]}
+            IFS=, read -r -a declared_providers <<< "${AGENT_REVIEW_PROVIDERS:-}"
+            for provider in "${declared_providers[@]}"; do
+                review_provider_name_valid "$provider" || continue
+                expected=$(review_provider_login "$provider" 2>/dev/null) || continue
+                [[ $login == "$expected" ]] || continue
+                printf '%s\n' "$provider"
+                return 0
+            done
+            ((bot_suffix)) || return 1
+            review_provider_name_valid "$login" || return 1
+            [[ $login != none ]] || return 1
+            printf '%s\n' "$login"
+            ;;
     esac
 }
 
+review_provider_lane() {
+    review_provider_name_valid "${1:-}" || return 1
+    review_provider_known "$1" && return 1
+    printf '%s\n' generic-automated
+}
+
 review_provider_request_marker() {
     case ${1:-} in
         coderabbit) printf '%s\n' '<!-- pr-to-green:provider-request provider=coderabbit -->' ;;
diff --git a/agentkit/skills/.shared/scripts/onboard-state.sh b/agentkit/skills/.shared/scripts/onboard-state.sh
index fccaebbb..288f09a9 100755
--- a/agentkit/skills/.shared/scripts/onboard-state.sh
+++ b/agentkit/skills/.shared/scripts/onboard-state.sh
@@ -22,16 +22,24 @@ done
 [[ -n $repo_root && -d $repo_root ]] || usage
 (( want_report || want_next || want_preflight || want_next_steps )) || usage
 repo_root=$(cd -- "$repo_root" && pwd -P) || die "cannot resolve repository root: $repo_root"
+self_dir=$(dirname -- "$(readlink -f -- "${BASH_SOURCE[0]}")")
 
 state='not onboarded'; next='discover'
 config=$repo_root/.agent/config.env; board=$repo_root/.agent/board.json
 marker=$repo_root/.agent/cache/onboarding-stage
+config_valid=yes
+if [[ -r $config && -x $self_dir/repo-config.sh ]] &&
+    ! "$self_dir/repo-config.sh" --repo-root "$repo_root" --validate >/dev/null 2>&1; then
+    config_valid=no
+fi
 if [[ ! -r $config ]]; then
     if [[ $(head -n 1 "$marker" 2> /dev/null || true) == discovered ]]; then
         state=discovered; next=declare
     fi
 elif [[ ! -r $board ]]; then
     state=discovered; next=declare
+elif [[ $config_valid == no ]]; then
+    state=declared; next=verify
 else
     command_declared=no
     grep -qE '^AGENT_CMD_[A-Z][A-Z0-9_]*=' "$config" 2> /dev/null && command_declared=yes
@@ -75,8 +83,6 @@ else
     fi
 fi
 
-self_dir=$(dirname -- "$(readlink -f -- "${BASH_SOURCE[0]}")")
-
 # --next is the bare stage/next line, exclusive of every other section.
 if ((want_next)); then
     printf 'stage=%s next=%s repo-root=%s\n' "$state" "$next" "$repo_root"
diff --git a/agentkit/skills/.shared/scripts/repo-config.sh b/agentkit/skills/.shared/scripts/repo-config.sh
index 23b893b2..9affec56 100755
--- a/agentkit/skills/.shared/scripts/repo-config.sh
+++ b/agentkit/skills/.shared/scripts/repo-config.sh
@@ -91,6 +91,7 @@ readonly BASE_TRUSTED_KEYS=(
 # The NAME is what `agent-run.sh --cmd <name>` takes, so it is constrained to a
 # shape that survives being lowercased into a filename and an argument.
 readonly CMD_KEY_PATTERN='^AGENT_CMD_[A-Z][A-Z0-9_]*$'
+readonly PROVIDER_LOGIN_KEY_PATTERN='^AGENT_REVIEW_PROVIDER_[A-Z][A-Z0-9_]*_LOGIN$'
 
 # Mirrors codex-adversarial-review.sh / claude-adversarial-review.sh's own
 # `--effort` enum (no `ultra`): a value AGENT_WORKER_EFFORT would accept but
@@ -212,6 +213,7 @@ done
 if [[ $mode == keys ]]; then
     printf '%s\n' "${ACCEPTED_KEYS[@]}"
     printf 'AGENT_CMD_<NAME>\n'
+    printf 'AGENT_REVIEW_PROVIDER_<NAME>_LOGIN\n'
     printf 'AGENT_RUNDIR_<NAME>\n'
     printf 'AGENT_VERIFY_<NAME>_MODE\nAGENT_VERIFY_<NAME>_INPUTS\nAGENT_VERIFY_<NAME>_TOOLCHAIN\n'
     exit 0
@@ -274,6 +276,7 @@ is_accepted() {
     # here, and only a test that asked for the hole by name caught it.
     [[ ! $candidate =~ $SECRET_PATTERN ]] || return 1
     [[ ! $candidate =~ $CMD_KEY_PATTERN ]] || return 0
+    [[ ! $candidate =~ $PROVIDER_LOGIN_KEY_PATTERN ]] || return 0
     [[ ! $candidate =~ $RUNDIR_KEY_PATTERN ]] || return 0
     [[ ! $candidate =~ $VERIFY_KEY_PATTERN ]] || return 0
     for key in "${ACCEPTED_KEYS[@]}"; do
@@ -405,42 +408,40 @@ generated_paths_valid() {
     done
 }
 
-# Mirrors lib/review-provider-catalog.sh's REVIEW_PROVIDER_NAMES, for warning
-# text only. Kept as a local literal rather than sourcing that file: this
+# Mirrors lib/review-provider-catalog.sh's provider-specific names. Kept as a
+# local literal rather than sourcing that file: this
 # parser is deliberately self-contained (see the file header) so a missing or
 # broken lib file can never turn one bad declaration into a hard failure for
 # every accepted key. test-repo-config.sh pins this against the catalog.
+# shellcheck disable=SC2034  # test-repo-config.sh reads this schema mirror.
 readonly REVIEW_PROVIDER_ACCEPTED_NAMES=(coderabbit github-code-quality none)
 
-providers_display() {
-    local out='' name
-    for name in "${REVIEW_PROVIDER_ACCEPTED_NAMES[@]}"; do
-        out+="${out:+, }$name"
-    done
-    printf '%s' "$out"
+provider_login_alias_reserved() {
+    case ${1,,} in
+        coderabbitai|github-code-quality) return 0 ;;
+        *) return 1 ;;
+    esac
 }
 
 providers_valid() {
-    local item saw_none=0 seen_coderabbit=0 seen_code_quality=0
+    local item saw_none=0
+    local -A seen=()
     [[ -n $1 && $1 != ,* && $1 != *, && $1 != *,,* ]] || return 1
     local -a items=()
     IFS=, read -ra items <<< "$1"
     ((${#items[@]})) || return 1
     for item in "${items[@]}"; do
         case $item in
-            coderabbit)
-                ((saw_none == 0 && seen_coderabbit == 0)) || return 1
-                seen_coderabbit=1
-                ;;
-            github-code-quality)
-                ((saw_none == 0 && seen_code_quality == 0)) || return 1
-                seen_code_quality=1
-                ;;
             none)
                 ((saw_none == 0 && ${#items[@]} == 1)) || return 1
                 saw_none=1
                 ;;
-            *) return 1 ;;
+            *)
+                [[ $item =~ ^[a-z][a-z0-9-]*$ && $item != coderabbitai ]] || return 1
+                ((saw_none == 0)) || return 1
+                [[ -z ${seen[$item]+yes} ]] || return 1
+                seen[$item]=1
+                ;;
         esac
     done
     return 0
@@ -839,6 +840,12 @@ validate() {
             ;;
         AGENT_GENERATED_PATHS) generated_paths_valid "$value" ;;
         AGENT_REVIEW_PROVIDERS) providers_valid "$value" ;;
+        AGENT_REVIEW_PROVIDER_*_LOGIN)
+            [[ $key =~ $PROVIDER_LOGIN_KEY_PATTERN &&
+                ! $key =~ ^AGENT_REVIEW_PROVIDER_(CODERABBIT|GITHUB_CODE_QUALITY|NONE)_LOGIN$ &&
+                $value =~ ^[A-Za-z0-9]([A-Za-z0-9_.-]{0,37}[A-Za-z0-9])?$ ]] &&
+                ! provider_login_alias_reserved "$value"
+            ;;
         AGENT_WORKER_MODEL | AGENT_WORKER_MODEL_FALLBACK) worker_model_valid "$value" ;;
         AGENT_WORKER_MODELS | AGENT_WORKER_MODELS_FALLBACK) worker_models_roster_valid "$value" ;;
         AGENT_WORKER_EFFORT)
@@ -975,7 +982,11 @@ while IFS= read -r line || [[ -n $line ]]; do
         if [[ -z $value ]]; then
             warn "empty value for $key on line $lineno, ignoring -- to record that this repository has none, comment the line out instead"
         elif [[ $key == AGENT_REVIEW_PROVIDERS ]]; then
-            warn "invalid value for $key on line $lineno, ignoring -- accepted: $(providers_display)"
+            if [[ ,$value, == *,coderabbitai,* ]]; then
+                warn "invalid value for $key on line $lineno, ignoring -- coderabbitai is a reserved built-in login alias; declare coderabbit instead"
+            else
+                warn "invalid value for $key on line $lineno, ignoring -- each name must match [a-z][a-z0-9-]*; none is exclusive; empty and duplicate items are forbidden"
+            fi
         elif [[ $key == AGENT_ADVERSARIAL_REVIEWER ]]; then
             warn "invalid value for $key on line $lineno, ignoring -- accepted: $(names_display "${ADVERSARIAL_REVIEWER_ACCEPTED_NAMES[@]}")"
         elif [[ $key == AGENT_ADVERSARIAL_REVIEW_EFFORT ]]; then
diff --git a/agentkit/skills/.shared/scripts/review-provider-config.sh b/agentkit/skills/.shared/scripts/review-provider-config.sh
index f20ff116..58b48bbe 100755
--- a/agentkit/skills/.shared/scripts/review-provider-config.sh
+++ b/agentkit/skills/.shared/scripts/review-provider-config.sh
@@ -158,5 +158,10 @@ for provider in "${providers[@]}"; do
                 ;;
         esac
     fi
-    printf 'provider=%s mode=%s source=declared\n' "$provider" "$mode"
+    lane=$(review_provider_lane "$provider" 2>/dev/null || true)
+    if [[ -n $lane ]]; then
+        printf 'provider=%s mode=%s source=declared lane=%s\n' "$provider" "$mode" "$lane"
+    else
+        printf 'provider=%s mode=%s source=declared\n' "$provider" "$mode"
+    fi
 done
diff --git a/agentkit/skills/.shared/scripts/run-state.sh b/agentkit/skills/.shared/scripts/run-state.sh
index 7a9d5ebd..d93f71f9 100755
--- a/agentkit/skills/.shared/scripts/run-state.sh
+++ b/agentkit/skills/.shared/scripts/run-state.sh
@@ -234,6 +234,8 @@ print_summary() {
         positive_ids("queued"; true) as $queued |
         positive_ids("receipt_prs"; true) as $receipts |
         positive_ids("skipped_prs"; true) as $skipped |
+        has("auto_review") as $has_auto_review |
+        .auto_review as $auto_review |
         has("root_turns") as $has_root_turns |
         has("first_completion") as $has_first_completion |
         .root_turns as $root_turns |
@@ -242,6 +244,8 @@ print_summary() {
         elif (($skipped - $prs) | length) > 0 then error("skipped_prs must be a subset of opened_prs")
         elif (($receipts + $skipped | length) != ($receipts + $skipped | unique | length))
             then error("receipt_prs and skipped_prs must be disjoint")
+        elif $has_auto_review and ($auto_review | type) != "boolean"
+            then error("auto_review must be a boolean")
         elif $has_root_turns and (($has_first_completion | not) or ($root_turns | type) != "array"
             or any($root_turns[]; . != true) or ($first_completion | type) != "boolean")
             then error("invalid root-turn summary evidence")
@@ -249,7 +253,10 @@ print_summary() {
             then error("invalid first-completion evidence")
         else (if $has_root_turns | not then "unavailable"
               elif $first_completion then ($root_turns | length | tostring) else "unlatched" end) as $telemetry |
-            [($prs | length), ($receipts | length), ($skipped | length), ($queued | length), $telemetry] | @tsv end
+            ($prs - ($receipts + $skipped)) as $missing |
+            [($prs | length), ($receipts | length), ($skipped | length), ($queued | length),
+             (if $has_auto_review then ($auto_review | tostring) else "false" end),
+             ($missing | if length == 0 then "-" else map(tostring) | join(",") end), $telemetry] | @tsv end
     ' <<<"$STATE" 2>/dev/null) ||
         die 'summary state requires valid opened_prs, queued, receipt_prs, and skipped_prs collections'
     [[ ! -L $LEDGER && -f $LEDGER && -r $LEDGER && -O $LEDGER ]] ||
@@ -271,22 +278,35 @@ print_summary() {
             then . else error("invalid handback evidence") end] | sort_by(.issue)
     ' "$LEDGER" 2>/dev/null) || die "unparseable active-workers evidence: $LEDGER"
     parked_count=$(jq 'length' <<<"$parked_rows")
-    local prs receipts skipped queued root_turns
-    IFS=$'\t' read -r prs receipts skipped queued root_turns <<<"$counts"
+    local prs receipts skipped queued auto_review missing_review_prs root_turns review_resume coverage_failure=''
+    IFS=$'\t' read -r prs receipts skipped queued auto_review missing_review_prs root_turns <<<"$counts"
+    if [[ $auto_review == true && $missing_review_prs != - ]]; then
+        review_resume="/review-remote-pr --auto-review ${missing_review_prs//,/; /review-remote-pr --auto-review }"
+        coverage_failure="auto-review coverage missing for PRs: $missing_review_prs; resume: $review_resume"
+    fi
     printf 'coverage= prs=%s receipts=%s skipped=%s parked=%s queued=%s root-turns-before-first-completion=%s\n' \
         "$prs" "$receipts" "$skipped" "$parked_count" "$queued" "$root_turns"
     jq -r '.[] | "blocked=\(.issue):\(.evidence)"' <<<"$parked_rows"
 
-    [[ -n $REPORTS_DIR ]] || return 0
+    if [[ -z $REPORTS_DIR ]]; then
+        [[ -z $coverage_failure ]] || die "$coverage_failure"
+        return 0
+    fi
     [[ ! -L $REPORTS_DIR ]] || die "verification reports directory must not be a symlink: $REPORTS_DIR"
     [[ ! -e $REPORTS_DIR || (-d $REPORTS_DIR && -O $REPORTS_DIR) ]] ||
         die "verification reports must be an owned directory: $REPORTS_DIR"
-    [[ -e $REPORTS_DIR ]] || return 0
+    if [[ ! -e $REPORTS_DIR ]]; then
+        [[ -z $coverage_failure ]] || die "$coverage_failure"
+        return 0
+    fi
     local reports_mode report report_mode report_text report_issue content_issue
     reports_mode=$(stat -c %a -- "$REPORTS_DIR") || die "could not inspect verification reports: $REPORTS_DIR"
     (( (8#$reports_mode & 8#077) == 0 )) || die "verification reports directory must be owner-private: $REPORTS_DIR"
     local -a reports=("$REPORTS_DIR"/issue-*.report)
-    [[ -e ${reports[0]} ]] || return 0
+    if [[ ! -e ${reports[0]} ]]; then
+        [[ -z $coverage_failure ]] || die "$coverage_failure"
+        return 0
+    fi
     for report in "${reports[@]}"; do
         [[ ${report##*/} =~ ^issue-([1-9][0-9]*)\.report$ ]] ||
             die "verification report filename must be issue-POSITIVE_INTEGER.report: $report"
@@ -304,6 +324,7 @@ print_summary() {
             die "verification report filename issue does not match content issue: $report"
         cat -- "$report"
     done
+    [[ -z $coverage_failure ]] || die "$coverage_failure"
 }
 
 main() {
diff --git a/agentkit/skills/parallel-issues/SKILL.md b/agentkit/skills/parallel-issues/SKILL.md
index 66251cd1..0d29bbf7 100755
--- a/agentkit/skills/parallel-issues/SKILL.md
+++ b/agentkit/skills/parallel-issues/SKILL.md
@@ -117,7 +117,7 @@ The scope, flags, repository, and base are fixed before the first receipt and su
 changes after compaction/resume: `scope=57,54` and `scope=57,62` cannot share an ID, nor can
 `auto-review=false` and `auto-review=true`; the same exact tuple may intentionally resume. Reuse this
 `RUN_ID` for all issues; never use a worker-local value. Immediately append each grant, steer, or board adjudication with `printf '%s' "$QUOTE" | "$agentkit/.shared/scripts/session-ledger.sh" append --ledger "$LEDGER" --run-id "$RUN_ID" --skills-path "$agentkit" --procedure-set parallel-issues --decision "$DECISION" --scope "$SCOPE" --quote-stdin || exit 1`.
-After establishing `RUN_ID`, run `"$agentkit/.shared/scripts/run-state.sh" init-summary --run-id "$RUN_ID" --repo-root "$repository_root"`; it preserves existing records.
+After establishing `RUN_ID`, run `"$agentkit/.shared/scripts/run-state.sh" init-summary --run-id "$RUN_ID" --repo-root "$repository_root"`; it preserves existing records. Persist the fixed invocation fact, never an unset shell default: when the invocation carried `--auto-review`, run `"$agentkit/.shared/scripts/run-state.sh" set --run-id "$RUN_ID" --repo-root "$repository_root" --path auto_review --json true`; otherwise run `"$agentkit/.shared/scripts/run-state.sh" set --run-id "$RUN_ID" --repo-root "$repository_root" --path auto_review --json false`. The handoff summary can then enforce review coverage after compaction.
 `QUOTE` is the human's verbatim quote; never put secrets or credentials in any field.
 After any compaction/resume, before taking another action, run `"$agentkit/.shared/scripts/session-ledger.sh" read --ledger "$LEDGER" --run-id "$RUN_ID"` and treat its output as the durable decision state.
 
@@ -510,7 +510,7 @@ Composer publishes once; root installs and verifies its hashed `uncoveredVerific
 ### Collect (per-completion — never wait for the slowest issue)
 
 `worker-result=PATH` uses the [result contract](references/worker-prompts.md#structured-result-contract): validate dispatch, ownership, Git and logs before accepting. Keep root CI/review obligations; unknown or blocked evidence is never green; unchanged accepted receipts resume without repeated work. Text fallbacks stay unknown.
-`agentkit activation-blocked: {...}` keeps ownership. Validate worker, worktree and workflow, then follow `.shared/spawn-contract.md` once to redeliver current bytes to the same context. The leaf acknowledges and resumes; unavailable or repeated delivery parks with work preserved.
+`agentkit activation-blocked: {...}` keeps ownership. Validate worker, worktree and workflow, then follow `.shared/spawn-contract.md` once to redeliver current bytes to the same context. The leaf acknowledges and resumes; unavailable or repeated delivery parks with work preserved. Before either PR-open path, including after a resumed Collect, restore the fixed invocation fact with `auto_review_state=$("$agentkit/.shared/scripts/run-state.sh" get --run-id "$RUN_ID" --repo-root "$repository_root" --path auto_review) || exit 1`; validate it with `case $auto_review_state in true|false) ;; *) printf 'invalid durable auto_review: %s\n' "$auto_review_state" >&2; exit 1 ;; esac`.
 
 - **Cross-write check first** → run the root-checkout Collect check against the immutable
   dispatch snapshot before trusting the worker's handback. Keep the helper's incident line,
@@ -518,8 +518,8 @@ Composer publishes once; root installs and verifies its hashed `uncoveredVerific
   worker's evidence. A dirty path is never an "unrelated local change" until the check proves
   otherwise.
 
-- **Completion report (branch + pushed SHA)** → review pushed diff; run `$agentkit/parallel-issues/scripts/compose-pr-body.sh`, then `"$agentkit/.shared/scripts/gh-body.sh" pr create --draft`; record with `"$agentkit/.shared/scripts/run-state.sh" record-summary --run-id "$RUN_ID" --repo-root "$repository_root" --path opened_prs --json "$pr"`; move issue to `In review`; start Phase 3. Diff size is never a reason to withhold this; see Diff-size facts.
-- **BLOCKED** → preserve the text handback, set `blocker_file="$worktree/.agent/logs/partial-blockers.list"`, and run `"$agentkit/.shared/scripts/validate-handback.sh" --classify-completion --worktree "$worktree" --handback-file "$completion_file" --blocker-file "$blocker_file"`. `disposition=partial-pushed pr=open blocker-file=written verification=unbound` proves the queried remote HEAD, not log attribution: review the diff, open the draft, record it with `"$agentkit/.shared/scripts/run-state.sh" record-summary --run-id "$RUN_ID" --repo-root "$repository_root" --path opened_prs --json "$pr"`, and pass `--blocker-file "$blocker_file"` to `$agentkit/parallel-issues/scripts/compose-pr-body.sh`; its `## Operator action required` section preserves blocker paths and discloses limited verification. Dispatch chained successors from the pushed SHA on both completion paths. Otherwise gate redrive on `"$agentkit/.shared/scripts/run-state.sh" get --run-id "$RUN_ID" --path redrive.<N>` and proceed only on exit 11 (absent); clear the blocker (`write-set`: widen the fence, recheck every active worker); only after the blocker clears, run one `tools.send`, then record `"$agentkit/.shared/scripts/run-state.sh" set --run-id "$RUN_ID" --path redrive.<N>`. If the same lead is unavailable, give a fresh lead the exact resume command; other blockers park. `baseline-red` gets one automatic re-drive. A sole `needs-paths: <glob>[,<glob>...]` drives that recheck; otherwise preserve the worktree and blocker evidence.
+- **Completion report (branch + pushed SHA)** → review pushed diff; run `$agentkit/parallel-issues/scripts/compose-pr-body.sh`, then `"$agentkit/.shared/scripts/gh-body.sh" pr create --draft`; record with `"$agentkit/.shared/scripts/run-state.sh" record-summary --run-id "$RUN_ID" --repo-root "$repository_root" --path opened_prs --json "$pr"`; print `printf 'next: dispatch draft-phase loop for #%s (Step 3a); auto-review=%s\n' "$pr" "$auto_review_state"`; move issue to `In review`; start Phase 3. Diff size is never a reason to withhold this; see Diff-size facts.
+- **BLOCKED** → preserve the text handback, set `blocker_file="$worktree/.agent/logs/partial-blockers.list"`, and run `"$agentkit/.shared/scripts/validate-handback.sh" --classify-completion --worktree "$worktree" --handback-file "$completion_file" --blocker-file "$blocker_file"`. `disposition=partial-pushed pr=open blocker-file=written verification=unbound` proves the queried remote HEAD, not log attribution: review the diff, open the draft, record it with `"$agentkit/.shared/scripts/run-state.sh" record-summary --run-id "$RUN_ID" --repo-root "$repository_root" --path opened_prs --json "$pr"`, print `printf 'next: dispatch draft-phase loop for #%s (Step 3a); auto-review=%s\n' "$pr" "$auto_review_state"`, and pass `--blocker-file "$blocker_file"` to `$agentkit/parallel-issues/scripts/compose-pr-body.sh`; its `## Operator action required` section preserves blocker paths and discloses limited verification. Dispatch chained successors from the pushed SHA on both completion paths. Otherwise gate redrive on `"$agentkit/.shared/scripts/run-state.sh" get --run-id "$RUN_ID" --path redrive.<N>` and proceed only on exit 11 (absent); clear the blocker (`write-set`: widen the fence, recheck every active worker); only after the blocker clears, run one `tools.send`, then record `"$agentkit/.shared/scripts/run-state.sh" set --run-id "$RUN_ID" --path redrive.<N>`. If the same lead is unavailable, give a fresh lead the exact resume command; other blockers park. `baseline-red` gets one automatic re-drive. A sole `needs-paths: <glob>[,<glob>...]` drives that recheck; otherwise preserve the worktree and blocker evidence.
 - **Queued issue** → spawn it immediately into the freed slot.
 
 **Stall detection:** record the next check at last progress + `STALL_THRESHOLD_MINUTES` (default 12 minutes). Before the threshold elapses, do not call
diff --git a/agentkit/skills/parallel-issues/references/worker-prompts.md b/agentkit/skills/parallel-issues/references/worker-prompts.md
index 7a834889..7d550db8 100644
--- a/agentkit/skills/parallel-issues/references/worker-prompts.md
+++ b/agentkit/skills/parallel-issues/references/worker-prompts.md
@@ -544,12 +544,13 @@ metadata, comments, replies, board moves, ready-flips — stays with the root.
    silent.
 3. Follow this composed verification runbook:
    __VERIFY_RUNBOOK__
-   Run every focused and full verification command through `agent-run.sh`; retain the fresh
-   green marker-bearing log path and do not rerun a failed command outside the wrapper.
-4. When verification is green, commit with `"$shared/worktree-commit.sh"` (explicit file
+   Run every verification command through `agent-run.sh`; use focused checks during TDD, but
+   commit the repair before the final unfocused run. Do not rerun a failed command outside the wrapper.
+4. When focused verification is green, commit with `"$shared/worktree-commit.sh"` (explicit file
    operands, Conventional Commit subject, the expanded `--trailer "$worker_attribution"`
-   -- or omitted, letting the helper derive it from the contract), then
-   push the branch. If unrelated dirt appears, stop and surface its files, diffstat, and
+   -- or omitted, letting the helper derive it from the contract). Run the unfocused full command
+   through `agent-run.sh` on that clean commit, retain its green marker-bearing log, and
+   push the branch only after that clean committed-HEAD run passes. If unrelated dirt appears, stop and surface its files, diffstat, and
    whether the checkpoint manifest explains it — never commit it.
 5. Return a completion report: branch, full commit SHA from the helper's success line,
    diffstat, and the green verification log path. If the helper exits 2 (nothing
diff --git a/agentkit/skills/pr-to-green/scripts/review-transition.sh b/agentkit/skills/pr-to-green/scripts/review-transition.sh
index 3a2f69c5..7affb98e 100755
--- a/agentkit/skills/pr-to-green/scripts/review-transition.sh
+++ b/agentkit/skills/pr-to-green/scripts/review-transition.sh
@@ -305,9 +305,11 @@ trap cleanup EXIT HUP INT TERM
 [[ -s $work_dir/providers.txt ]] || die 'provider capability resolver returned an empty plan'
 
 while IFS= read -r line; do
-    if [[ $line =~ ^provider=([a-z0-9-]+)[[:space:]]mode=([a-z-]+)[[:space:]]source=([a-z]+)$ ]]; then
+    lane=''
+    if [[ $line =~ ^provider=([a-z0-9-]+)[[:space:]]mode=([a-z-]+)[[:space:]]source=([a-z]+)([[:space:]]lane=([a-z-]+))?$ ]]; then
         provider=${BASH_REMATCH[1]}
         mode=${BASH_REMATCH[2]}
+        lane=${BASH_REMATCH[5]-}
     else
         die 'provider capability resolver returned a malformed record'
     fi
@@ -316,6 +318,9 @@ while IFS= read -r line; do
         die "provider capability plan contains unsupported provider: $provider"
     [[ $mode == "$expected_mode" ]] ||
         die "provider capability plan contains unsupported capability: $provider:$mode"
+    expected_lane=$(review_provider_lane "$provider" 2>/dev/null || true)
+    [[ $lane == "$expected_lane" ]] ||
+        die "provider capability plan contains unsupported lane: $provider:${lane:-none}"
     providers+=("$provider")
     modes[$provider]=$mode
 done <"$work_dir/providers.txt"
diff --git a/agentkit/skills/review-remote-pr/SKILL.md b/agentkit/skills/review-remote-pr/SKILL.md
index 306f1afc..442ab9e5 100755
--- a/agentkit/skills/review-remote-pr/SKILL.md
+++ b/agentkit/skills/review-remote-pr/SKILL.md
@@ -335,24 +335,25 @@ failure; stop with evidence unavailable. A receipt marker is authoritative from
 
 ## Step 2: Fix CI Failures
 
-**Step 1c — batch pushes:** review behavior after a push is provider configuration, not a
-workflow guarantee — still batch each cycle's fixes into **one** push; never post
-`@coderabbitai pause`/`resume`.
+**Step 1c — batch pushes:** batch each cycle's fixes into **one** push; never post `@coderabbitai pause`/`resume`. Provider behavior after a push is configuration, not a guarantee.
 
-Diagnose the causal failure (`gh run view --log-failed "$run_id" | grep -E "FAIL|error|Error"`,
-run ID from the `gh pr checks` URL column), then run the **Implementation-worker gate** above.
-The worker verifies independently before its cycle push, through `agent-run.sh`:
+Diagnose the causal failure from the run ID in `gh pr checks`, then run the **Implementation-worker gate** above; the worker verifies before its cycle push:
 
 ```bash
 [ -d "${agentkit:-}/.shared/scripts" ] && [ "${agentkit_provenance:-}" = ok ] || { printf "%s\n" "agentkit unresolved: prepend THE CACHE REHYDRATION block" >&2; exit 1; }
 agent_run="$agentkit/.shared/scripts/agent-run.sh"
-"$agent_run" --cmd lint --if-declared --cmd test
+"$agent_run" --cmd lint --if-declared
 ```
 
-For red/green iterations the worker uses `"$agent_run" --cmd test --only NAME[,NAME...]` (forwards through the
-repo's `AGENT_CMD_TEST_FOCUS` declaration); after the final tree change, the worker must run the unfocused `"$agent_run" --cmd test` once for the full-suite verdict
-before worker publication. A successful run prints one `PASS:` line; a failure prints `FAIL(rc=N):`,
-context, `note:` lines, matched errors, and the log path. **Never push without local verification passing** — on `FAIL`, having set `check`, `log`, and `failing_paths` from its output:
+Commit the repair through the worker gate. After the worker-gate commit and before push, run the full test on clean committed HEAD:
+
+```bash
+[ -d "${agentkit:-}/.shared/scripts" ] && [ "${agentkit_provenance:-}" = ok ] || { printf "%s\n" "agentkit unresolved: prepend THE CACHE REHYDRATION block" >&2; exit 1; }
+agent_run="$agentkit/.shared/scripts/agent-run.sh"
+"$agent_run" --cmd test
+```
+
+For red/green iterations use `"$agent_run" --cmd test --only NAME[,NAME...]` through `AGENT_CMD_TEST_FOCUS`. After the final edit, commit, then run the unfocused `"$agent_run" --cmd test` once on clean committed HEAD for the full-suite verdict; push only after `PASS:`. On `FAIL`, having set `check`, `log`, and `failing_paths` from its output:
 
 ```bash
 [ -d "${agentkit:-}/.shared/scripts" ] && [ "${agentkit_provenance:-}" = ok ] || { printf "%s\n" "agentkit unresolved: prepend THE CACHE REHYDRATION block" >&2; exit 1; }
diff --git a/agentkit/skills/review-remote-pr/references/adversarial-review.md b/agentkit/skills/review-remote-pr/references/adversarial-review.md
index f26b01fc..8a2ffec6 100644
--- a/agentkit/skills/review-remote-pr/references/adversarial-review.md
+++ b/agentkit/skills/review-remote-pr/references/adversarial-review.md
@@ -445,7 +445,8 @@ covered rather than `stale` with zero additional review spends — see
 Update the same title after repair. Produce its evidence with
 `finding-ledger.sh evidence --title TITLE --path AFFECTED_PATH --log GREEN_LOG --repo-root WORKTREE --repair-sha REPAIR_SHA > FILE`:
 the log must be the green, unfocused `agent-run.sh --cmd test` run (a focused `--only` or red log is
-refused), `--head` defaults to the checkout's HEAD, and `REPAIR_SHA` is the commit that changed that
+refused) made after the repair commit and before push. Its header binds the tested head and tracked-tree cleanliness;
+that head must be the checkout's clean current HEAD. `REPAIR_SHA` is the commit that changed the
 path (not a later formatting-only commit). Then record it with
 `add --verdict fixed --sha "$(jq -r .repairSha FILE)" --evidence FILE --repo-root WORKTREE --head CURRENT_SHA`
 (also supply title and severity). One evidence file per finding. The helper checks commit ancestry,
diff --git a/agentkit/skills/review-remote-pr/references/worker-gate.md b/agentkit/skills/review-remote-pr/references/worker-gate.md
index 7e6f8682..84bf109b 100644
--- a/agentkit/skills/review-remote-pr/references/worker-gate.md
+++ b/agentkit/skills/review-remote-pr/references/worker-gate.md
@@ -52,11 +52,11 @@ grep -Fq -- '--cmd test' "$repair_prompt" || exit 1
 
 ## Worker-owned publication
 
-Workers commit and push their own branch after focused/full verification and a completion report
-(branch, full SHA, diffstat, and the path of the green unfocused `agent-run.sh --cmd test` log);
-`worktree-commit.sh` uses explicit files and trailer. Root turns that log into repair evidence with
-`finding-ledger.sh evidence`, which refuses a focused or red log -- send the worker back to verify
-rather than accepting the handback.
+Workers commit and push their own branch. Between those actions, run unfocused `agent-run.sh --cmd test`
+after the repair commit and before push. Use focused checks while editing; commit via `worktree-commit.sh`
+with explicit files and a trailer; push only after the clean committed HEAD passes. Return a completion report
+with branch, full SHA, diffstat, and green log. Root passes it to `finding-ledger.sh evidence`, which refuses
+focused, red, dirty, unbound, or different-HEAD logs; send the worker back to verify.
 The root owns the pushed `base...HEAD` review, PR metadata, board, replies, and next cycle.
 
 ## Environment-refusal fallback
diff --git a/agentkit/skills/review-remote-pr/scripts/finding-ledger.sh b/agentkit/skills/review-remote-pr/scripts/finding-ledger.sh
index d0612e47..1856ef16 100755
--- a/agentkit/skills/review-remote-pr/scripts/finding-ledger.sh
+++ b/agentkit/skills/review-remote-pr/scripts/finding-ledger.sh
@@ -7,6 +7,7 @@ readonly PROGNAME=${0##*/}
 readonly RECEIPT_MARKER='<!-- adversarial-review:spent -->'
 readonly DOC_MARKER='<!-- review-remote-pr:agent-doc -->'
 readonly SHA_RE='^[[:xdigit:]]{7,64}(,[[:xdigit:]]{7,64})*$'
+readonly FULL_SHA_RE='^([[:xdigit:]]{40}|[[:xdigit:]]{64})$'
 readonly ORDER_RC=13
 readonly FINDING_SLUG_JQ='ascii_downcase | gsub("[^a-z0-9]+"; "-") | ltrimstr("-") | rtrimstr("-") | if . == "" then "finding" else . end'
 SCRIPT_DIR=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)
@@ -32,7 +33,8 @@ Usage: $PROGNAME add --title TITLE --severity P1|P2 --verdict fixed --sha SHA
        $PROGNAME ids --file FILE    (prints ID<TAB>TITLE; review-ledger.sh cover --reason fix:ID names one)
        $PROGNAME evidence --title T --path P --log LOG --repo-root DIR --repair-sha SHA [--head SHA]
                  (prints fixed-verdict evidence JSON; LOG must be a green unfocused agent-run.sh --cmd test
-                 log; head defaults to DIR's HEAD; SHA is the commit that changed P)
+                 log run on DIR's clean committed HEAD; head defaults to that HEAD; SHA is the
+                 commit that changed P)
 
 Terminal evidence: --evidence FILE --repo-root DIR --head SHA. Evidence JSON
 binds finding (title) to decision rejected|accepted-risk and rationale, or to
@@ -361,7 +363,7 @@ validate_repairs() {
         log=$(jq -r .evidence.log <<<"$row")
         digest=$(jq -r .evidence.logSha256 <<<"$row")
         command=$(jq -r .evidence.command <<<"$row")
-        [[ $head =~ ^[0-9a-f]{40}$ && $sha =~ ^[0-9a-f]{40}$ && $tested =~ ^[0-9a-f]{40}$ && $digest =~ ^[0-9a-f]{64}$ ]] ||
+        [[ $head =~ $FULL_SHA_RE && $sha =~ $FULL_SHA_RE && $tested =~ $FULL_SHA_RE && $digest =~ ^[0-9a-f]{64}$ ]] ||
             die_evidence 'repair evidence requires full commit and log hashes'
         if ! git -C "$root" merge-base --is-ancestor "$sha" "$tested" 2>/dev/null ||
             ! git -C "$root" merge-base --is-ancestor "$tested" "$head" 2>/dev/null; then
@@ -441,11 +443,26 @@ resolve_commit() {
     git -C "$1" rev-parse --verify -q "$2^{commit}" 2>/dev/null || die_evidence "not a commit in $1: $2"
 }
 
+require_tested_head() {
+    local log=$1 current=$2 header tested clean
+    header=$(sed -n '2p' "$log")
+    if [[ $header =~ '  head='([[:xdigit:]]{40}|[[:xdigit:]]{64})'  tracked-clean='(yes|no)$ ]]; then
+        tested=${BASH_REMATCH[1]}
+        clean=${BASH_REMATCH[2]}
+    else
+        die_evidence 'verification log has no tested-head metadata; commit the repair, then run agent-run.sh --cmd test'
+    fi
+    [[ $tested == "$current" ]] ||
+        die_evidence "verification log tested $tested, not the current head $current"
+    [[ $clean == yes ]] ||
+        die_evidence 'verification log ran with uncommitted tracked changes; commit the repair, then run agent-run.sh --cmd test'
+}
+
 # Emit fixed-verdict evidence for one finding, refusing anything add would
 # later reject: the log must be the green, unfocused declared test run, and the
 # named repair commit must change the finding's path.
 cmd_evidence() {
-    local title='' path='' log='' root='' repair_sha='' head='' declared command digest row
+    local title='' path='' log='' root='' repair_sha='' head='' actual_head='' declared command digest row
     shift
     while (($#)); do
         case $1 in
@@ -464,7 +481,11 @@ cmd_evidence() {
     [[ $path != /* && $path != -* && $path != *'..'* ]] || die_evidence 'repair path must be repository relative'
     [[ -f $log && ! -L $log ]] || die_evidence "verification log is unavailable: $log"
     log=$(cd -- "$(dirname -- "$log")" && pwd -P)/${log##*/}
-    head=$(resolve_commit "$root" "${head:-HEAD}") && repair_sha=$(resolve_commit "$root" "$repair_sha") || exit 1
+    head=$(resolve_commit "$root" "${head:-HEAD}") &&
+        actual_head=$(resolve_commit "$root" HEAD) &&
+        repair_sha=$(resolve_commit "$root" "$repair_sha") || exit 1
+    [[ $head == "$actual_head" ]] ||
+        die_evidence "evidence head $head is not the current head $actual_head"
     command=$(sed -n '1s/^=== agent-run //p' "$log")
     declared=$("$SCRIPT_DIR/../../.shared/scripts/repo-config.sh" --repo-root "$root" \
         --get-argv AGENT_CMD_TEST | tr '\0' ' ') || die_evidence 'the repository declares no AGENT_CMD_TEST'
@@ -472,6 +493,9 @@ cmd_evidence() {
     [[ -n $declared ]] || die_evidence 'the repository declares no AGENT_CMD_TEST'
     [[ -n $command && $command == "$declared" ]] ||
         die_evidence "log is not the unfocused declared test run (log: ${command:-<none>}; declared: $declared); run agent-run.sh --cmd test without --only"
+    [[ $(tail -n 1 -- "$log") == '=== agent-run exited rc=0 '* ]] ||
+        die_evidence 'verification log has no final successful agent-run result'
+    require_tested_head "$log" "$head"
     digest=$(verification_digest "$log") || die_evidence 'verification log digest unavailable (requires sha256sum or shasum)'
     row=$(jq -cn --arg finding "$title" --arg sha "$repair_sha" --arg head "$head" \
         --arg path "$path" --arg command "$command" --arg log "$log" --arg digest "${digest%% *}" \
diff --git a/opencode/package-lock.json b/opencode/package-lock.json
index 7b62caf6..cb349ad8 100644
--- a/opencode/package-lock.json
+++ b/opencode/package-lock.json
@@ -1,12 +1,12 @@
 {
   "name": "@wrzonance/agentkit-opencode",
-  "version": "0.9.11",
+  "version": "0.9.12",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "@wrzonance/agentkit-opencode",
-      "version": "0.9.11",
+      "version": "0.9.12",
       "license": "MIT",
       "devDependencies": {
         "@opencode-ai/plugin": "1.18.18"
diff --git a/opencode/package.json b/opencode/package.json
index 3f9f077c..c47d56bf 100644
--- a/opencode/package.json
+++ b/opencode/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@wrzonance/agentkit-opencode",
-  "version": "0.9.11",
+  "version": "0.9.12",
   "description": "Agent Kit plugin for OpenCode CLI: injects the environment contract into the model's system prompt at session start.",
   "type": "module",
   "main": "./index.js",
diff --git a/plugin/agentkit/.claude-plugin/plugin.json b/plugin/agentkit/.claude-plugin/plugin.json
index 9ea636ca..8ec78423 100644
--- a/plugin/agentkit/.claude-plugin/plugin.json
+++ b/plugin/agentkit/.claude-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "agentkit",
-  "version": "0.9.11",
+  "version": "0.9.12",
   "description": "Board-aware parallel issue and PR review skills, with lifecycle hooks and a per-repository contract.",
   "author": {
     "name": "wrzonance",
diff --git a/plugin/agentkit/.codex-plugin/plugin.json b/plugin/agentkit/.codex-plugin/plugin.json
index 57947170..fe92f129 100644
--- a/plugin/agentkit/.codex-plugin/plugin.json
+++ b/plugin/agentkit/.codex-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "agentkit",
-  "version": "0.9.11",
+  "version": "0.9.12",
   "description": "Board-aware parallel issue and PR review skills, with lifecycle hooks and a per-repository contract.",
   "author": {
     "name": "wrzonance",
diff --git a/plugin/opencode/package.json b/plugin/opencode/package.json
index 3f9f077c..c47d56bf 100644
--- a/plugin/opencode/package.json
+++ b/plugin/opencode/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@wrzonance/agentkit-opencode",
-  "version": "0.9.11",
+  "version": "0.9.12",
   "description": "Agent Kit plugin for OpenCode CLI: injects the environment contract into the model's system prompt at session start.",
   "type": "module",
   "main": "./index.js",
diff --git a/tests/lint-helper-size.sh b/tests/lint-helper-size.sh
index 9ebc983a..8f07f61e 100755
--- a/tests/lint-helper-size.sh
+++ b/tests/lint-helper-size.sh
@@ -26,12 +26,15 @@ declare -A KNOWN_OVERSIZE=(
     # #777: complete Step 0 recipe moved from injected prose into --help.
     # #777 absolute-path guard + #778 harness-bound runtime-tool record.
     [skills/.shared/scripts/agent-preflight.sh]="1401:17157:800"
-    # #731/#732/#776/#809: records, summaries, and truthful reuse diagnostics.
-    [skills/.shared/scripts/agent-run.sh]="1920:20761:800"
+    # #731/#732/#776/#809/#874/PR #877: yielded-run status and lease lifecycle.
+    # #873: bind repair evidence to the clean committed head tested before push.
+    # #874 review repair: stable process identity and canonical fallback boundaries.
+    [skills/.shared/scripts/agent-run.sh]="2017:21859:800"
     # #865: scope the generated regeneration hint to plugin-backed onboarding.
     [skills/.shared/scripts/bootstrap-repo.sh]="818:10363:800"
     # #777: repository-facts recipe moved from injected prose into --help.
-    [skills/.shared/scripts/repo-config.sh]="1154:12008:800"
+    # #876 review repair: reject normalized aliases of built-in provider logins.
+    [skills/.shared/scripts/repo-config.sh]="1165:12203:800"
     [skills/.shared/scripts/worktree-commit.sh]="847:8772:800"
     # #852: distinguish configured missing scans at the retarget boundary.
     [skills/parallel-issues/scripts/chain-advance.sh]="1086:13108:800"
@@ -183,7 +186,17 @@ readonly MAX_HELPER_TOKENS=10000
 # #834 review repair: exact combined helper tree measurement.
 # #865 reference-use clauses and bounded no-run diagnostics: 1,891,483 bytes / 4.
 # #873 (trimmed): IDs, evidence producer, cover preconditions: 1,898,236 bytes / 4.
-readonly MAX_TREE_TOKENS=474559
+# #875 review: coverage failures preserve handoff evidence across every report path.
+# #874 PR #877: portable status and bounded active-run lease inheritance.
+# #873 chain: tested-head metadata and repair handback validation.
+# #873 review repair: real SHA-256 evidence fixture and split verification recipe.
+# #873/#874 + #875 final merge-down: exact combined helper tree measurement.
+# #874 review repair: stable identity and symlink-safe log selection.
+# #874 review + #875 merge: exact combined helper tree measurement.
+# #873 PR repair: bind evidence overrides to the checkout's actual HEAD.
+# #873 evidence repair + #874/#875 merge: exact combined helper tree measurement.
+# #873/#874 + #875 + #876 provider-alias repair: exact combined helper tree measurement.
+readonly MAX_TREE_TOKENS=477181
 
 violations=0
 checked=0
diff --git a/tests/lint-skill-size.sh b/tests/lint-skill-size.sh
index acc9a74c..e6beaef7 100755
--- a/tests/lint-skill-size.sh
+++ b/tests/lint-skill-size.sh
@@ -46,7 +46,10 @@ declare -A KNOWN_OVERSIZE=(
     # #845 + #865: state the cold-start/reference-use boundary at workflow activation.
     # #873: Step 0 names each helper's flags separately, with a canonical preflight path.
     # #873 review: preflight's flag set includes --activation-origin.
-    [parallel-issues]="761:16265:500"
+    # #875 review: preserve summary evidence and persist an invocation-derived review mode.
+    # #875 follow-up: Collect restores that durable review mode in resumed shells.
+    # #875 integration: retain the inline restore across the final predecessor merge.
+    [parallel-issues]="772:16666:500"
 )
 
 readonly MAX_BODY_LINES=500
diff --git a/tests/test-agent-run-cmd.sh b/tests/test-agent-run-cmd.sh
index 6ec82364..3e979c0b 100755
--- a/tests/test-agent-run-cmd.sh
+++ b/tests/test-agent-run-cmd.sh
@@ -394,7 +394,29 @@ assert_contains "$log" '=== agent-run echo hello' 'the log names the command it
 assert_contains "$log" '=== started' 'and when it started'
 assert_contains "$log" 'concurrent-suites=1' 'the log records the active full-suite count'
 assert_contains "$log" '=== agent-run exited rc=0' 'and terminates with the verdict'
-assert_contains "$out" 'has NOT finished' 'and the caller is told what an unterminated log means'
+assert_contains "$out" 'resume this same call; never relaunch' \
+    'and the caller is told how to continue an unterminated run'
+
+# Issue #873: a repair handback must prove which committed head the full run
+# tested. The header records HEAD and tracked-tree cleanliness so a stale or
+# dirty log cannot certify the branch tip that was pushed.
+bound_repo=$(make_repo)
+printf 'AGENT_CMD_OK=echo hello\n' >"$bound_repo/.agent/config.env"
+printf '.agent/\n' >"$bound_repo/.gitignore"
+printf 'a\n' >"$bound_repo/tracked.txt"
+git -C "$bound_repo" add .gitignore tracked.txt
+git -C "$bound_repo" -c user.name=Test -c user.email=test@example.invalid commit -qm init
+bound_sha=$(git -C "$bound_repo" rev-parse HEAD)
+(cd "$bound_repo" && "$real_run_sh" --cmd ok >/dev/null 2>&1)
+log=$(cat "$bound_repo"/.agent/logs/*-ok.log)
+assert_contains "$log" "head=$bound_sha  tracked-clean=yes" \
+    'the log header binds a clean run to its committed head'
+printf 'b\n' >"$bound_repo/tracked.txt"
+rm -f -- "$bound_repo"/.agent/logs/*-ok.log*
+(cd "$bound_repo" && "$real_run_sh" --cmd ok >/dev/null 2>&1)
+log=$(cat "$bound_repo"/.agent/logs/*-ok.log)
+assert_contains "$log" "head=$bound_sha  tracked-clean=no" \
+    'the log header records uncommitted tracked changes'
 
 # The suppressed-line count must report the command output, not the markers.
 assert_contains "$out" '(1 lines suppressed' 'the line count excludes the log bookkeeping'
@@ -414,12 +436,19 @@ cat >"$marker_repo/bin/awk" <<'EOF'
 #!/usr/bin/env bash
 set -uo pipefail
 if [[ ${1:-} == '{print $22}' ]]; then
-    printf 'alive\n'
+    exit 1
 else
     exec /usr/bin/awk "$@"
 fi
 EOF
 chmod +x -- "$marker_repo/bin/awk"
+cat >"$marker_repo/bin/ps" <<'EOF'
+#!/usr/bin/env bash
+set -uo pipefail
+[[ ${*: -1} != 999999 ]] || exit 1
+printf 'Tue Sep 22 12:34:56 2026\n'
+EOF
+chmod +x -- "$marker_repo/bin/ps"
 cat >"$marker_repo/bin/mktemp" <<'EOF'
 #!/usr/bin/env bash
 set -uo pipefail
@@ -438,9 +467,12 @@ printf '999999 1\n' >"$marker_suite_dir/agent-run.stale"
 marker_out=$(cd "$marker_repo" && PATH="$marker_repo/bin:$PATH" \
     TMPDIR="$tmp" "$real_run_sh" --force --cmd test 2>&1)
 marker_log=$(find "$marker_repo/.agent/logs" -type f -name '*-test.log' -print -quit)
-assert_contains "$marker_out" 'marker-ok' 'a no-proc marker accepts the alive fallback'
+assert_contains "$marker_out" 'PASS: printf marker-ok' \
+    'a failed proc read uses a stable process-start fallback'
 assert_not_contains "$marker_out" 'marker template is not BSD-compatible' \
     'the active marker uses a BSD-compatible trailing-X template'
+assert_not_contains "$(cat "$marker_log")" 'process-start=alive' \
+    'the process identity never degrades to a liveness sentinel'
 assert_contains "$(cat "$marker_log")" 'concurrent-suites=1' \
     'a stale new-format marker is removed before counting active suites'
 
@@ -740,7 +772,7 @@ assert_contains "$out" 'declared-test-ran' \
 # runner-resolved link; finding 2 carries --force into build_chain_argv. Both
 # were offset by further comment trims elsewhere, holding the line count at 1627.
 # #612 adds paired formatter resolution and bounded cargo failure summaries.
-assert_eq yes "$([[ $(wc -l < "$root/agentkit/skills/.shared/scripts/agent-run.sh") -le 1920 ]] && printf yes || printf no)" \
-    'agent-run.sh stays at or under 1920 lines (#809 reuse diagnostics)'
+assert_eq yes "$([[ $(wc -l < "$root/agentkit/skills/.shared/scripts/agent-run.sh") -le 2017 ]] && printf yes || printf no)" \
+    'agent-run.sh stays at or under 2017 lines (#873 metadata + #874 review repair)'
 
 finish
diff --git a/tests/test-agent-run-verification-cache.sh b/tests/test-agent-run-verification-cache.sh
index 66309e9f..6f58e72c 100755
--- a/tests/test-agent-run-verification-cache.sh
+++ b/tests/test-agent-run-verification-cache.sh
@@ -564,15 +564,17 @@ printf 'AGENT_CMD_TEST=tools/run\nAGENT_VERIFY_TEST_MODE=local\nAGENT_VERIFY_TES
 printf '%s\n' '#!/bin/sh' 'echo run >> "$COUNT_FILE"' 'sleep 2' 'exit 0' > "$local_repo/tools/run"
 local_run > "$tmp/owner-output" & owner=$!
 for ((attempt=0; attempt<100; attempt++)); do
-    [[ -d $local_repo/.agent/verification-records ]] &&
-        find "$local_repo/.agent/verification-records" -name running -print | grep -q . && break
+    running_record=$(find "$local_repo/.agent/verification-records" -name running -print -quit 2>/dev/null || true)
+    [[ -n $running_record ]] && grep -q '/.agent/logs/.*\.log$' "$running_record" && break
     sleep 0.05
 done
 out=$(local_run); local_rc=$?
-assert_eq '75' "$local_rc" 'running identical command returns non-success status'
-assert_contains "$out" 'verification running: handle=' 'running identical command returns existing handle'
-running_handle=$(printf '%s\n' "$out" | sed -n 's/^agent-run: verification running: handle=//p')
+assert_eq '2' "$local_rc" 'running identical command is refused before it can duplicate work'
+assert_contains "$out" 'already running:' 'running identical command returns the existing log'
 wait "$owner"
+# An unrelated record proves later checks never rediscover the handle by directory order.
+mkdir -p "$local_repo/.agent/verification-records/000-decoy"
+running_handle=${running_record%/running}
 out=$(local_run)
 assert_contains "$out" 'verification current:' 'a completed concurrent owner is reusable'
 record=$running_handle/result
diff --git a/tests/test-agent-run-yield.sh b/tests/test-agent-run-yield.sh
new file mode 100755
index 00000000..f19c8737
--- /dev/null
+++ b/tests/test-agent-run-yield.sh
@@ -0,0 +1,174 @@
+#!/usr/bin/env bash
+# Suite: yielded agent-run calls remain observable and cannot be duplicated.
+set -uo pipefail
+
+TEST_NAME='agent-run-yield'
+here=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)
+root=$(dirname -- "$here")
+# shellcheck source=lib/assert.sh
+source "$here/lib/assert.sh"
+
+run_sh="$root/agentkit/skills/.shared/scripts/agent-run.sh"
+tmp=$(mktemp -d)
+trap 'rm -rf -- "$tmp"' EXIT
+
+help=$($run_sh --help)
+assert_contains "$help" 'active duplicate exits 2' \
+    'help documents the accepted duplicate-run exit status'
+assert_contains "$help" 'unknown abandoned handle exits 75' \
+    'help distinguishes abandoned verification evidence from a live duplicate'
+assert_not_contains "$help" 'exit 2 is reserved' \
+    'help no longer reserves duplicate-run status for interpreter errors'
+assert_contains "$help" 'in-flight command is still refused with its log' \
+    '--force documentation does not promise a duplicate-run bypass'
+
+epoch_repo=$tmp/epoch-repo
+git -C "$tmp" init -q epoch-repo
+mkdir -p "$epoch_repo/.agent" "$epoch_repo/tools"
+cat > "$epoch_repo/tools/check" <<'EOF'
+#!/usr/bin/env bash
+set -uo pipefail
+printf 'started\n' > "${STARTED_FILE:?}"
+for ((attempt=0; attempt<500; attempt++)); do
+    [[ -e ${RELEASE_FILE:?} ]] && exit 0
+    sleep 0.02
+done
+exit 1
+EOF
+chmod +x -- "$epoch_repo/tools/check"
+printf 'AGENT_CMD_TEST=tools/check\n' > "$epoch_repo/.agent/config.env"
+printf 'unset EPOCHSECONDS\n' > "$tmp/unset-epoch"
+BASH_ENV=$tmp/unset-epoch STARTED_FILE=$tmp/epoch-started RELEASE_FILE=$tmp/epoch-release \
+    "$run_sh" --dir "$epoch_repo" --cmd test > "$tmp/epoch.out" 2>&1 &
+epoch_owner=$!
+for ((attempt=0; attempt<100; attempt++)); do
+    epoch_log=$(find "$epoch_repo/.agent/logs" -type f -name '*-test.log' -print -quit 2>/dev/null || true)
+    [[ -n ${epoch_log:-} && -f $tmp/epoch-started ]] && break
+    sleep 0.02
+done
+epoch_status=$(BASH_ENV=$tmp/unset-epoch "$run_sh" status "$epoch_log" 2>&1)
+assert_contains "$epoch_status" 'running pid=' \
+    'a live run remains observable when EPOCHSECONDS is unavailable'
+printf 'release\n' > "$tmp/epoch-release"
+wait "$epoch_owner"
+assert_eq 0 "$?" 'a declared command runs when EPOCHSECONDS is unavailable'
+
+repo=$tmp/repo
+git -C "$tmp" init -q repo
+mkdir -p "$repo/.agent" "$repo/tools"
+cat > "$repo/tools/slow-check" <<'EOF'
+#!/usr/bin/env bash
+set -uo pipefail
+printf 'started\n' > "${STARTED_FILE:?}"
+for ((attempt=0; attempt<500; attempt++)); do
+    [[ -e ${RELEASE_FILE:?} ]] && { printf 'finished\n'; exit 0; }
+    sleep 0.02
+done
+exit 1
+EOF
+chmod +x -- "$repo/tools/slow-check"
+printf 'AGENT_CMD_TEST=tools/slow-check\n' > "$repo/.agent/config.env"
+
+owner_out=$tmp/owner.out
+STARTED_FILE=$tmp/started RELEASE_FILE=$tmp/release \
+    "$run_sh" --dir "$repo" --cmd test > "$owner_out" 2>&1 &
+owner=$!
+for ((attempt=0; attempt<100; attempt++)); do
+    log=$(find "$repo/.agent/logs" -type f -name '*-test.log' -print -quit 2>/dev/null || true)
+    [[ -n ${log:-} && -f $tmp/started ]] && break
+    sleep 0.02
+done
+
+assert_contains "$(tail -n 1 -- "$owner_out")" 'resume this same call; never relaunch' \
+    'the final pre-block line tells a yielded caller to resume the same call'
+status=$($run_sh status "$log" 2>&1)
+assert_contains "$status" 'running pid=' 'status identifies an unfinished live run'
+assert_contains "$status" 'elapsed=' 'running status includes elapsed seconds'
+
+duplicate=''
+duplicate_rc=0
+duplicate=$(STARTED_FILE=$tmp/duplicate RELEASE_FILE=$tmp/release \
+    "$run_sh" --dir "$repo" --cmd test 2>&1) || duplicate_rc=$?
+assert_eq 2 "$duplicate_rc" 'an identical active launch is refused as usage'
+assert_contains "$duplicate" "already running: $log" 'duplicate refusal names the original log'
+assert_eq no "$([[ -e $tmp/duplicate ]] && printf yes || printf no)" \
+    'the refused duplicate never starts the declared command'
+
+printf 'release\n' > "$tmp/release"
+wait "$owner"
+assert_eq pass "$($run_sh status "$log")" 'status reports pass after the owner completes'
+assert_eq pass "$(TMPDIR=$tmp/nonexistent-fallback "$run_sh" status "$log")" \
+    'repository-log status does not require a fallback directory'
+
+printf 'AGENT_CMD_TEST=false\n' > "$repo/.agent/config.env"
+"$run_sh" --dir "$repo" --label failing --cmd test > /dev/null 2>&1 || true
+fail_log=$(find "$repo/.agent/logs" -type f -name '*-failing.log' -print -quit)
+assert_eq 'fail rc=1' "$($run_sh status "$fail_log")" 'status preserves a terminal failure code'
+
+background_repo=$tmp/background-repo
+git -C "$tmp" init -q background-repo
+mkdir -p "$background_repo/.agent" "$background_repo/tools"
+cat > "$background_repo/tools/check" <<'EOF'
+#!/usr/bin/env bash
+set -uo pipefail
+printf 'run\n' >> "${COUNT_FILE:?}"
+(
+    for ((attempt=0; attempt<500; attempt++)); do
+        [[ -e ${DESCENDANT_RELEASE:?} ]] && exit 0
+        sleep 0.02
+    done
+    exit 1
+) &
+printf '%s\n' "$!" > "${DESCENDANT_FILE:?}"
+EOF
+chmod +x -- "$background_repo/tools/check"
+printf 'AGENT_CMD_TEST=tools/check\n' > "$background_repo/.agent/config.env"
+COUNT_FILE=$tmp/background-count DESCENDANT_FILE=$tmp/descendant DESCENDANT_RELEASE=$tmp/descendant-release \
+    "$run_sh" --dir "$background_repo" --cmd test > /dev/null 2>&1
+descendant=$(<"$tmp/descendant")
+assert_eq yes "$([[ $descendant =~ ^[0-9]+$ ]] && kill -0 "$descendant" 2>/dev/null && printf yes || printf no)" \
+    'fixture leaves a live background descendant after agent-run completes'
+second_rc=0
+COUNT_FILE=$tmp/background-count DESCENDANT_FILE=$tmp/descendant-2 DESCENDANT_RELEASE=$tmp/descendant-release \
+    "$run_sh" --dir "$background_repo" --cmd test > /dev/null 2>&1 || second_rc=$?
+assert_eq 0 "$second_rc" 'a completed command descendant does not retain the active-run lease'
+assert_eq 2 "$(wc -l < "$tmp/background-count" | tr -d '[:space:]')" \
+    'the identical command executes again after its prior wrapper completes'
+descendant_two=$(<"$tmp/descendant-2")
+assert_eq yes "$([[ $descendant_two =~ ^[0-9]+$ ]] && kill -0 "$descendant_two" 2>/dev/null && printf yes || printf no)" \
+    'the second completed wrapper also leaves its descendant alive until release'
+printf 'release\n' > "$tmp/descendant-release"
+for ((attempt=0; attempt<100; attempt++)); do
+    kill -0 "$descendant" 2>/dev/null || { kill -0 "$descendant_two" 2>/dev/null || break; }
+    sleep 0.02
+done
+kill "$descendant" "$descendant_two" 2>/dev/null || true
+
+symlink_repo=$tmp/symlink-repo
+symlink_agent=$tmp/symlink-agent
+git -C "$tmp" init -q symlink-repo
+mkdir -p "$symlink_agent"
+ln -s "$symlink_agent" "$symlink_repo/.agent"
+"$run_sh" --dir "$symlink_repo" -- true > /dev/null 2>&1
+assert_eq no "$([[ -e $symlink_agent/run-records ]] && printf yes || printf no)" \
+    'a symlinked .agent parent receives no active-run records'
+assert_eq no "$([[ -e $symlink_agent/logs ]] && printf yes || printf no)" \
+    'a symlinked .agent parent receives no command logs'
+
+logs_repo=$tmp/logs-repo
+logs_target=$tmp/logs-target
+git -C "$tmp" init -q logs-repo
+mkdir -p "$logs_repo/.agent" "$logs_target"
+ln -s "$logs_target" "$logs_repo/.agent/logs"
+"$run_sh" --dir "$logs_repo" -- true > /dev/null 2>&1
+assert_eq '' "$(find "$logs_target" -mindepth 1 -print -quit)" \
+    'a symlinked .agent/logs directory receives no command log'
+
+mkdir -p "$tmp/relative-tmp"
+(cd "$tmp" && TMPDIR=relative-tmp/// "$run_sh" --dir "$symlink_repo" -- true > /dev/null 2>&1)
+fallback_log=$(find "$tmp/relative-tmp/agent-logs-$(id -u)" -type f -name '*-true.log' -print -quit)
+fallback_status=$(cd "$tmp" && TMPDIR=relative-tmp/// "$run_sh" status "$fallback_log" 2>&1)
+assert_eq pass "$fallback_status" \
+    'status accepts a canonical fallback log with relative trailing-slash TMPDIR'
+
+finish
diff --git a/tests/test-bootstrap-repo.sh b/tests/test-bootstrap-repo.sh
index 85ec5e45..263f4f10 100755
--- a/tests/test-bootstrap-repo.sh
+++ b/tests/test-bootstrap-repo.sh
@@ -108,8 +108,11 @@ assert_contains "$listed" 'AGENT_STATUS_VOCAB=Backlog,Ready,In progress,In revie
 assert_contains "$(cat "$repo/.agent/config.env")" '# AGENT_REVIEW_PROVIDERS=' \
     'generated config proposes an explicit automated review provider choice'
 assert_contains "$(cat "$repo/.agent/config.env")" \
-    'coderabbit, github-code-quality, or none' \
-    'provider proposal explains the supported choices'
+    'Known providers get provider-specific settlement' \
+    'provider proposal explains provider-specific handling'
+assert_contains "$(cat "$repo/.agent/config.env")" \
+    'Any other valid name is observed' \
+    'provider proposal explains the generic observe-only default'
 expected_generator=$(jq -r .version "$root/agentkit/.codex-plugin/plugin.json")
 assert_contains "$(cat "$repo/.agent/config.env")" "AGENT_ONBOARDED_BY=agentkit/$expected_generator" \
     'generated config records the installed generator version'
@@ -490,6 +493,17 @@ assert_eq '1' "$(grep -c '^AGENT_REVIEW_PROVIDERS=none$' "$repo/.agent/config.en
 assert_eq '0' "$(grep -c '^# AGENT_REVIEW_PROVIDERS=' "$repo/.agent/config.env" || true)" \
     'refresh does not leave a proposal beside a selected provider'
 
+# Arbitrary valid names are declarations too; refresh must preserve the choice
+# and must not reintroduce a commented proposal beside it.
+sed -i 's/^AGENT_REVIEW_PROVIDERS=none$/AGENT_REVIEW_PROVIDERS=chatgpt-codex-connector/' \
+    "$repo/.agent/config.env"
+assert_rc 0 'refresh accepts an unknown observe-only provider choice' -- run_bs \
+    --repo-root "$repo" --project 7 --force
+assert_eq '1' "$(grep -c '^AGENT_REVIEW_PROVIDERS=chatgpt-codex-connector$' "$repo/.agent/config.env")" \
+    'refresh preserves one unknown provider declaration'
+assert_eq '0' "$(grep -c '^# AGENT_REVIEW_PROVIDERS=' "$repo/.agent/config.env" || true)" \
+    'refresh does not propose a replacement for an unknown provider'
+
 # A declared worker model is likewise a declaration, not a proposal: refresh
 # must not nag once a repository has already made the choice. Uses --refresh
 # (not --force) since that is what this case actually describes and exercises
diff --git a/tests/test-finding-ledger.sh b/tests/test-finding-ledger.sh
index 7f99efd5..ee7a6916 100755
--- a/tests/test-finding-ledger.sh
+++ b/tests/test-finding-ledger.sh
@@ -316,8 +316,9 @@ assert_rc 1 'ids refuses a missing findings file as unavailable evidence' -- \
 
 # --- evidence producer (issue #873) -------------------------------------------
 # Plan-level strictness: the log must be the green, unfocused declared test run
-# and the named repair commit must change the finding's path. The log is not
-# bound to a commit; forge-verifiable evidence is a separate follow-up.
+# on the clean current head, and the named repair commit must change the
+# finding's path. Existing ledgers written before tested-head metadata remain
+# readable, but new evidence cannot be produced from an unbound log.
 ev_repo="$tmp/ev-repo"
 git init -q "$ev_repo"
 git -C "$ev_repo" config user.name Test
@@ -338,9 +339,16 @@ printf 'tidy\n' >"$ev_repo/other.txt"
 git -C "$ev_repo" add other.txt
 git -C "$ev_repo" commit -qm 'format follow-up'
 ev_head=$(git -C "$ev_repo" rev-parse HEAD)
-printf '=== agent-run tests/regression.sh\n=== agent-run exited rc=0 after 1s\n' >"$tmp/ev-full.log"
-printf '=== agent-run tests/regression.sh --only one\n=== agent-run exited rc=0 after 1s\n' >"$tmp/ev-focused.log"
-printf '=== agent-run tests/regression.sh\n=== agent-run exited rc=1 after 1s\n' >"$tmp/ev-red.log"
+agent_log() {
+    printf '=== agent-run %s\n=== started 2026-09-22T00:00:00Z  pid=1  cwd=%s  concurrent-suites=1  head=%s  tracked-clean=%s\n=== agent-run exited rc=%s after 1s\n' \
+        "$2" "$ev_repo" "$3" "$4" "$5" >"$1"
+}
+agent_log "$tmp/ev-full.log" tests/regression.sh "$ev_head" yes 0
+agent_log "$tmp/ev-focused.log" 'tests/regression.sh --only one' "$ev_head" yes 0
+agent_log "$tmp/ev-red.log" tests/regression.sh "$ev_head" yes 1
+agent_log "$tmp/ev-other-head.log" tests/regression.sh "$ev_repair" yes 0
+agent_log "$tmp/ev-dirty.log" tests/regression.sh "$ev_head" no 0
+printf '=== agent-run tests/regression.sh\n=== agent-run exited rc=0 after 1s\n' >"$tmp/ev-unbound.log"
 evidence() {
     "$script" evidence --title 'Guard input' --path affected.sh --repo-root "$ev_repo" "$@"
 }
@@ -361,7 +369,7 @@ assert_rc 0 'producer output is accepted by add --verdict fixed unmodified' -- r
     --repo-root "$ev_repo" --head "$ev_head"
 assert_eq complete "$("$script" status --file "$ev_run/findings.ndjson" --repo-root "$ev_repo" \
     --head "$ev_head" | jq -r .remediation)" \
-    'evidence from a header-less agent-run log validates as complete (existing ledgers keep working)'
+    'fresh tested-head evidence validates as complete'
 
 assert_rc 2 'evidence requires an explicit --repair-sha' -- evidence --log "$tmp/ev-full.log"
 assert_rc 1 'a focused log is refused as repair evidence' -- \
@@ -370,10 +378,33 @@ assert_rc 1 'a red log is refused as repair evidence' -- \
     evidence --log "$tmp/ev-red.log" --repair-sha "$ev_repair"
 assert_rc 1 'a repair SHA that does not change the path is refused' -- \
     evidence --log "$tmp/ev-full.log" --repair-sha "$ev_head"
+stale_override_rc=0
+evidence --head "$ev_repair" --log "$tmp/ev-other-head.log" --repair-sha "$ev_repair" \
+    >/dev/null 2>"$tmp/ev-stale-override.err" || stale_override_rc=$?
+assert_eq 1 "$stale_override_rc" \
+    'an explicit reachable old head cannot replace the checkout HEAD for new evidence'
+assert_contains "$(cat "$tmp/ev-stale-override.err")" \
+    "evidence head $ev_repair is not the current head $ev_head" \
+    'the stale override refusal names both the requested and actual heads'
+assert_rc 1 'a log from another head cannot certify the current pushed head' -- \
+    evidence --log "$tmp/ev-other-head.log" --repair-sha "$ev_repair"
+assert_rc 1 'a log from a dirty tree cannot certify the committed head' -- \
+    evidence --log "$tmp/ev-dirty.log" --repair-sha "$ev_repair"
+assert_rc 1 'new repair evidence requires tested-head metadata' -- \
+    evidence --log "$tmp/ev-unbound.log" --repair-sha "$ev_repair"
 focused_err=$(evidence --log "$tmp/ev-focused.log" --repair-sha "$ev_repair" 2>&1 >/dev/null || true)
 assert_contains "$focused_err" 'without --only' 'the focused-log refusal says how to produce valid evidence'
 red_err=$(evidence --log "$tmp/ev-red.log" --repair-sha "$ev_repair" 2>&1 >/dev/null || true)
 assert_contains "$red_err" 'no final successful agent-run result' 'the red-log refusal names the failed run'
+other_head_err=$(evidence --log "$tmp/ev-other-head.log" --repair-sha "$ev_repair" 2>&1 >/dev/null || true)
+assert_contains "$other_head_err" "tested $ev_repair, not the current head $ev_head" \
+    'the stale-log refusal names both the tested and current heads'
+dirty_err=$(evidence --log "$tmp/ev-dirty.log" --repair-sha "$ev_repair" 2>&1 >/dev/null || true)
+assert_contains "$dirty_err" 'uncommitted tracked changes' \
+    'the dirty-log refusal tells the worker to commit before full verification'
+unbound_err=$(evidence --log "$tmp/ev-unbound.log" --repair-sha "$ev_repair" 2>&1 >/dev/null || true)
+assert_contains "$unbound_err" 'no tested-head metadata' \
+    'the unbound-log refusal asks for a current agent-run log'
 abs_err=$("$script" evidence --title 'Guard input' --path "$ev_repo/affected.sh" --log "$tmp/ev-full.log" \
     --repo-root "$ev_repo" --repair-sha "$ev_repair" 2>&1 >/dev/null; printf 'rc=%s' "$?")
 assert_contains "$abs_err" 'repair path must be repository relative' 'an absolute --path is refused by name'
@@ -390,6 +421,43 @@ real_rc=0
 evidence --log "$real_log" --repair-sha "$ev_repair" >/dev/null 2>"$tmp/real.err" || real_rc=$?
 assert_eq 0 "$real_rc" "a real agent-run.sh log certifies the repair ($(cat "$tmp/real.err"))"
 
+# Git repositories can use 64-character SHA-256 object IDs. When this Git
+# supports that object format, exercise the real runner and evidence producer
+# together so the log-header contract stays aligned with Git's full IDs.
+sha256_repo="$tmp/ev-sha256-repo"
+if git init -q --object-format=sha256 "$sha256_repo" 2>/dev/null; then
+    git -C "$sha256_repo" config user.name Test
+    git -C "$sha256_repo" config user.email test@example.invalid
+    mkdir -p "$sha256_repo/.agent" "$sha256_repo/tests"
+    printf 'AGENT_CMD_TEST=tests/regression.sh\n' >"$sha256_repo/.agent/config.env"
+    printf '.agent/\n' >"$sha256_repo/.gitignore"
+    printf '#!/bin/sh\necho regression ok\n' >"$sha256_repo/tests/regression.sh"
+    chmod +x "$sha256_repo/tests/regression.sh"
+    printf 'broken\n' >"$sha256_repo/affected.sh"
+    git -C "$sha256_repo" add .gitignore affected.sh tests/regression.sh
+    git -C "$sha256_repo" commit -qm baseline
+    printf 'repaired\n' >"$sha256_repo/affected.sh"
+    git -C "$sha256_repo" commit -qam repair
+    sha256_repair=$(git -C "$sha256_repo" rev-parse HEAD)
+    assert_eq 64 "${#sha256_repair}" 'the regression fixture uses a full SHA-256 object ID'
+    (cd -- "$sha256_repo" && "$agent_run" --cmd test >/dev/null 2>&1)
+    sha256_log=$(find "$sha256_repo/.agent/logs" -name '*-test.log' -type f -print -quit)
+    sha256_rc=0
+    "$script" evidence --title 'Guard input' --path affected.sh --repo-root "$sha256_repo" \
+        --log "$sha256_log" --repair-sha "$sha256_repair" >/dev/null 2>"$tmp/sha256.err" || sha256_rc=$?
+    assert_eq 0 "$sha256_rc" \
+        "a real SHA-256 repository log certifies the repair ($(cat "$tmp/sha256.err"))"
+fi
+
+# Validation remains backward-compatible with records produced before the
+# tested-head header existed. Only evidence creation requires the new binding.
+legacy_digest=$(sha256sum "$tmp/ev-unbound.log"); legacy_digest=${legacy_digest%% *}
+jq -c --arg log "$tmp/ev-unbound.log" --arg digest "$legacy_digest" \
+    '.evidence.log=$log | .evidence.logSha256=$digest' "$ev_run/findings.ndjson" >"$tmp/legacy-repair.ndjson"
+assert_eq complete "$("$script" status --file "$tmp/legacy-repair.ndjson" --repo-root "$ev_repo" \
+    --head "$ev_head" | jq -r .remediation)" \
+    'existing header-less repair evidence remains readable'
+
 undeclared_repo="$tmp/ev-undeclared"
 git clone -q "$ev_repo" "$undeclared_repo"
 undeclared_err=$("$script" evidence --title 'Guard input' --path affected.sh --log "$tmp/ev-full.log" \
diff --git a/tests/test-onboard-refresh.sh b/tests/test-onboard-refresh.sh
index 2c3fa9d6..02b65d6d 100755
--- a/tests/test-onboard-refresh.sh
+++ b/tests/test-onboard-refresh.sh
@@ -66,7 +66,7 @@ assert_contains "$out" 'toolchains=-1' \
 assert_contains "$out" 'paths=drift' \
     'summary preserves path drift alongside generator drift'
 
-printf '%s\n' 'AGENT_REVIEW_PROVIDERS=not-a-provider' >> "$repo/.agent/config.env"
+printf '%s\n' 'AGENT_REVIEW_PROVIDERS=X-1' >> "$repo/.agent/config.env"
 out=$(bash "$refresh_sh" --repo-root "$repo" --summary 2>&1)
 assert_contains "$out" 'review-providers=invalid' \
     'report identifies an invalid provider choice as an advisory gap'
diff --git a/tests/test-onboard-state.sh b/tests/test-onboard-state.sh
index be779498..4eaf79f3 100755
--- a/tests/test-onboard-state.sh
+++ b/tests/test-onboard-state.sh
@@ -49,6 +49,16 @@ out=$($state_sh --repo-root "$repo" --report)
 assert_contains "$out" 'stage=armed' 'ignored local declarations arm after verification'
 assert_contains "$out" 'next=none' 'the local model has no commit step'
 
+# An invalid declaration cannot be called armed merely because it is locally
+# ignored and has a command. It must return to declaration repair first.
+printf 'AGENT_REPO_SLUG=o/r\nAGENT_BASE_BRANCH=main\nAGENT_CMD_VERIFY=true\nAGENT_REVIEW_PROVIDERS=X-1\n' \
+    > "$repo/.agent/config.env"
+out=$("$state_sh" --repo-root "$repo" --report)
+assert_not_contains "$out" 'stage=armed' 'invalid config never reports the repository as armed'
+assert_contains "$out" 'stage=declared' 'invalid config returns to the declaration stage'
+assert_contains "$out" 'next=verify' 'invalid config asks for verification after repair'
+printf 'AGENT_REPO_SLUG=o/r\nAGENT_BASE_BRANCH=main\nAGENT_CMD_VERIFY=true\n' > "$repo/.agent/config.env"
+
 printf '.agent/*\n!.agent/config.env\n!.agent/board.json\n' > "$repo/.gitignore"
 git -C "$repo" add -- .agent/config.env .agent/board.json .gitignore
 out=$("$state_sh" --repo-root "$repo" --report)
diff --git a/tests/test-repo-config.sh b/tests/test-repo-config.sh
index 93cddee8..7f60caef 100755
--- a/tests/test-repo-config.sh
+++ b/tests/test-repo-config.sh
@@ -86,16 +86,14 @@ repo=$(make_repo config-bad.env)
 out=$("$rc_sh" --repo-root "$repo" --export 2> /dev/null)
 err=$("$rc_sh" --repo-root "$repo" --export 2>&1 > /dev/null)
 for key in AGENT_REPO_SLUG AGENT_BASE_BRANCH AGENT_PROJECT_NUMBER \
-    AGENT_ADR_DIR AGENT_WORKTREE_ROOT AGENT_BRANCH_PREFIXES AGENT_REVIEW_PROVIDERS; do
+    AGENT_ADR_DIR AGENT_WORKTREE_ROOT AGENT_BRANCH_PREFIXES; do
     assert_not_contains "$out" "export $key=" "rejects invalid $key"
 done
 assert_contains "$err" 'AGENT_UNKNOWN_KEY' 'warns about an unknown key'
 assert_contains "$err" '--list-keys' 'the unknown-key warning names a discoverable way to list accepted keys'
 assert_contains "$err" 'no equals sign' 'warns about a malformed line'
-assert_contains "$err" 'invalid value for AGENT_REVIEW_PROVIDERS' 'warns about the invalid provider declaration'
-for provider in coderabbit github-code-quality none; do
-    assert_contains "$err" "$provider" "the provider rejection names $provider as accepted"
-done
+assert_contains "$out" "export AGENT_REVIEW_PROVIDERS='coderabbit,not-a-provider'" \
+    'accepts arbitrary syntactically valid provider names in a mixed declaration'
 assert_rc 0 'a fully invalid config still exits 0' -- "$rc_sh" --repo-root "$repo" --export
 
 # Unknown declarations are onboarding drift, not a per-line emergency. One
@@ -139,6 +137,8 @@ for key in AGENT_REPO_SLUG AGENT_BASE_BRANCH AGENT_REVIEW_PROVIDERS AGENT_WORKER
     assert_contains "$list_out" "$key" "--list-keys names the accepted literal key $key"
 done
 assert_contains "$list_out" 'AGENT_CMD_<NAME>' '--list-keys documents the open-ended command pattern'
+assert_contains "$list_out" 'AGENT_REVIEW_PROVIDER_<NAME>_LOGIN' \
+    '--list-keys documents the open-ended provider login pattern'
 assert_contains "$list_out" 'AGENT_RUNDIR_<NAME>' '--list-keys documents the open-ended rundir pattern'
 list_out_repo=$("$rc_sh" --repo-root "$repo" --list-keys 2> /dev/null)
 assert_eq "$list_out" "$list_out_repo" '--list-keys is schema, not affected by which repo it is pointed at'
@@ -152,6 +152,76 @@ assert_eq 'yes' "$([[ -n $repo_config_providers ]] && printf yes || printf no)"
 assert_eq "$catalog_providers" "$repo_config_providers" \
     'repo-config.sh provider display list matches lib/review-provider-catalog.sh exactly'
 
+# --- open provider declarations and optional login overrides ---------------
+repo=$(mktemp -d "$tmp/repo.XXXXXX")
+mkdir -p "$repo/.agent"
+printf '%s\n' \
+    'AGENT_REVIEW_PROVIDERS=chatgpt-codex-connector' \
+    'AGENT_REVIEW_PROVIDER_CHATGPT_CODEX_CONNECTOR_LOGIN=codex-review-bot' \
+    > "$repo/.agent/config.env"
+assert_rc 0 'a valid unknown provider and login override validate' -- \
+    "$rc_sh" --repo-root "$repo" --validate
+out=$("$rc_sh" --repo-root "$repo" --export 2> /dev/null)
+assert_contains "$out" "export AGENT_REVIEW_PROVIDER_CHATGPT_CODEX_CONNECTOR_LOGIN='codex-review-bot'" \
+    'the optional provider login override is exported as data'
+
+for alias in coderabbitai CodeRabbitAI github-code-quality GITHUB-CODE-QUALITY; do
+    printf '%s\n' \
+        'AGENT_REVIEW_PROVIDERS=chatgpt-codex-connector' \
+        "AGENT_REVIEW_PROVIDER_CHATGPT_CODEX_CONNECTOR_LOGIN=$alias" \
+        > "$repo/.agent/config.env"
+    set +e
+    invalid_out=$("$rc_sh" --repo-root "$repo" --validate 2>&1)
+    invalid_rc=$?
+    set -e
+    assert_eq 1 "$invalid_rc" "reserved login alias '$alias' is rejected after normalization"
+    assert_contains "$invalid_out" 'invalid value for AGENT_REVIEW_PROVIDER_CHATGPT_CODEX_CONNECTOR_LOGIN' \
+        "reserved login alias '$alias' identifies the rejected override key"
+done
+
+for key in AGENT_REVIEW_PROVIDER_CODERABBIT_LOGIN \
+    AGENT_REVIEW_PROVIDER_GITHUB_CODE_QUALITY_LOGIN \
+    AGENT_REVIEW_PROVIDER_NONE_LOGIN; do
+    printf '%s=%s\n' "$key" custom-review-bot > "$repo/.agent/config.env"
+    set +e
+    invalid_out=$("$rc_sh" --repo-root "$repo" --validate 2>&1)
+    invalid_rc=$?
+    set -e
+    assert_eq 1 "$invalid_rc" "$key is rejected because built-in identities are fixed"
+    assert_contains "$invalid_out" "invalid value for $key" \
+        "$key reports the dead override instead of silently accepting it"
+    out=$("$rc_sh" --repo-root "$repo" --export 2> /dev/null)
+    assert_not_contains "$out" "export $key=" "$key is never exported as live configuration"
+done
+
+for providers in 'coderabbit,chatgpt-codex-connector' 'x-y,z9'; do
+    printf 'AGENT_REVIEW_PROVIDERS=%s\n' "$providers" > "$repo/.agent/config.env"
+    assert_rc 0 "valid provider list '$providers' validates" -- \
+        "$rc_sh" --repo-root "$repo" --validate
+done
+for providers in 'none,x' 'x,x' 'x,,y' 'X-1' 'coderabbitai'; do
+    printf 'AGENT_REVIEW_PROVIDERS=%s\n' "$providers" > "$repo/.agent/config.env"
+    set +e
+    invalid_out=$("$rc_sh" --repo-root "$repo" --validate 2>&1)
+    invalid_rc=$?
+    set -e
+    assert_eq 1 "$invalid_rc" "invalid provider list '$providers' is rejected"
+    if [[ $providers == coderabbitai ]]; then
+        assert_contains "$invalid_out" 'reserved built-in login alias' \
+            "invalid provider list '$providers' names the alias collision"
+    else
+        assert_contains "$invalid_out" '[a-z][a-z0-9-]*' \
+            "invalid provider list '$providers' names the provider-name rule"
+    fi
+done
+
+printf '%s\n' \
+    'AGENT_REVIEW_PROVIDERS=chatgpt-codex-connector' \
+    'AGENT_REVIEW_PROVIDER_CHATGPT_CODEX_CONNECTOR_LOGIN=bad login' \
+    > "$repo/.agent/config.env"
+assert_rc 1 'an unsafe provider login override is rejected' -- \
+    "$rc_sh" --repo-root "$repo" --validate
+
 # --- secret rejection ------------------------------------------------------
 repo=$(make_repo config-secrets.env)
 out=$("$rc_sh" --repo-root "$repo" --export 2> /dev/null)
@@ -740,8 +810,8 @@ assert_rc 0 '--validate still accepts launchable codex/claude reviewer compounds
 # reviewer_roster_entry_valid now refuses an OpenCode-family compound, which
 # model_family itself recognizes but adversarial-run.sh cannot launch (+4
 # lines). Measured.
-assert_eq yes "$([[ $(wc -l < "$root/agentkit/skills/.shared/scripts/repo-config.sh") -le 1154 ]] && printf yes || printf no)" \
-    'repo-config.sh stays at or under 1154 lines (issue #777 helper-owned recipe)'
+assert_eq yes "$([[ $(wc -l < "$root/agentkit/skills/.shared/scripts/repo-config.sh") -le 1165 ]] && printf yes || printf no)" \
+    'repo-config.sh stays at or under 1165 lines (issue #876 review repair)'
 
 verify_repo=$(mktemp -d "$tmp/verification.XXXXXX")
 mkdir -p "$verify_repo/.agent"
diff --git a/tests/test-review-provider-catalog.sh b/tests/test-review-provider-catalog.sh
index 3b74bb2c..ebd9dc47 100755
--- a/tests/test-review-provider-catalog.sh
+++ b/tests/test-review-provider-catalog.sh
@@ -38,8 +38,45 @@ for provider in coderabbit github-code-quality none; do
     assert_contains "$names_out" "$provider" \
         "review_provider_names names $provider as accepted"
 done
+assert_eq observe-only "$(catalog_call review_provider_mode chatgpt-codex-connector)" \
+    'a syntactically valid unknown provider is observe-only'
+assert_eq generic-settlement "$(catalog_call review_provider_lifecycle chatgpt-codex-connector)" \
+    'an unknown provider uses generic settlement'
+assert_eq chatgpt-codex-connector \
+    "$(catalog_call review_provider_login chatgpt-codex-connector)" \
+    'an unknown provider defaults its login to its declared name'
+assert_eq chatgpt-codex-connector \
+    "$(AGENT_REVIEW_PROVIDERS=chatgpt-codex-connector \
+        catalog_call review_provider_from_login chatgpt-codex-connector)" \
+    'an unknown provider login round-trips without a bot suffix'
+assert_eq chatgpt-codex-connector \
+    "$(catalog_call review_provider_from_login 'chatgpt-codex-connector[bot]')" \
+    'an unknown provider login round-trips with a bot suffix'
 # shellcheck disable=SC2016 # The inner shell expands its own positional parameter.
-assert_rc 1 'unknown providers have no capability entry' -- bash -c \
-    'source "$1"; review_provider_mode unexpected' bash "$catalog"
+assert_rc 1 'an unknown provider has no request marker' -- bash -c \
+    'source "$1"; review_provider_request_marker chatgpt-codex-connector' bash "$catalog"
+assert_eq codex-review-bot "$(AGENT_REVIEW_PROVIDERS=chatgpt-codex-connector \
+    AGENT_REVIEW_PROVIDER_CHATGPT_CODEX_CONNECTOR_LOGIN=codex-review-bot \
+    catalog_call review_provider_login chatgpt-codex-connector)" \
+    'a declared unknown provider accepts an explicit login override'
+assert_eq chatgpt-codex-connector "$(AGENT_REVIEW_PROVIDERS=chatgpt-codex-connector \
+    AGENT_REVIEW_PROVIDER_CHATGPT_CODEX_CONNECTOR_LOGIN=codex-review-bot \
+    catalog_call review_provider_from_login 'codex-review-bot[bot]')" \
+    'an explicit login override round-trips to its declared provider'
+for alias in coderabbitai CodeRabbitAI github-code-quality GITHUB-CODE-QUALITY; do
+    # shellcheck disable=SC2016 # The inner shell expands its own positional parameters.
+    assert_rc 1 "reserved login alias $alias is rejected by the catalog override boundary" -- \
+        env AGENT_REVIEW_PROVIDER_CHATGPT_CODEX_CONNECTOR_LOGIN="$alias" \
+        bash -c 'source "$1"; review_provider_override_login chatgpt-codex-connector' bash "$catalog"
+done
+# shellcheck disable=SC2016 # The inner shell expands its own positional parameter.
+assert_rc 1 'a built-in login alias cannot become an implicit generic provider name' -- bash -c \
+    'source "$1"; review_provider_mode coderabbitai' bash "$catalog"
+# shellcheck disable=SC2016 # The inner shell expands its own positional parameter.
+assert_rc 1 'an undeclared human-shaped login is not promoted to a provider' -- bash -c \
+    'source "$1"; review_provider_from_login ordinary-human' bash "$catalog"
+# shellcheck disable=SC2016 # The inner shell expands its own positional parameter.
+assert_rc 1 'malformed provider names have no capability entry' -- bash -c \
+    'source "$1"; review_provider_mode X-1' bash "$catalog"
 
 finish
diff --git a/tests/test-review-provider-classification.sh b/tests/test-review-provider-classification.sh
new file mode 100755
index 00000000..496ac1a3
--- /dev/null
+++ b/tests/test-review-provider-classification.sh
@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+# Unknown forge bots stay in the generic automated lane regardless of declaration.
+set -uo pipefail
+
+TEST_NAME='review-provider-classification'
+here=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)
+root=$(dirname -- "$here")
+# shellcheck source=lib/assert.sh
+source "$here/lib/assert.sh"
+
+classifier="$root/agentkit/skills/review-remote-pr/scripts/classify-author.sh"
+author='{"login":"chatgpt-codex-connector[bot]","type":"Bot"}'
+
+out=$(printf '%s\n' "$author" | "$classifier")
+assert_contains "$out" '"lane":"generic-automated"' \
+    'an undeclared unknown bot uses the generic automated lane'
+
+out=$(AGENT_REVIEW_PROVIDERS=chatgpt-codex-connector \
+    bash -c 'printf "%s\n" "$1" | "$2"' bash "$author" "$classifier")
+assert_contains "$out" '"lane":"generic-automated"' \
+    'declaring an unknown provider does not grant the known-provider lane'
+
+finish
diff --git a/tests/test-review-provider-config.sh b/tests/test-review-provider-config.sh
index 72487c55..ce8f7090 100755
--- a/tests/test-review-provider-config.sh
+++ b/tests/test-review-provider-config.sh
@@ -67,24 +67,21 @@ assert_eq 'provider=none mode=disabled source=invalid' "$out" \
 
 repo=$(make_repo 'unknown-provider')
 out=$(bash "$resolver" --repo-root "$repo" 2> "$tmp/err")
+assert_eq 'provider=unknown-provider mode=observe-only source=declared lane=generic-automated' "$out" \
+    'unknown providers use the observe-only generic bot lane'
+assert_eq '' "$(<"$tmp/err")" 'a valid unknown provider declaration is silent'
+
+repo=$(make_repo 'coderabbitai')
+out=$(bash "$resolver" --repo-root "$repo" 2> "$tmp/err")
 assert_eq 'provider=none mode=disabled source=invalid' "$out" \
-    'unknown providers cannot select a trigger'
-unrecognized_source=$out
-for provider in coderabbit github-code-quality none; do
-    assert_contains "$(<"$tmp/err")" "$provider" \
-        "an unrecognized declared provider names $provider as accepted"
-done
+    'a built-in login alias cannot declare an implicit generic provider'
+assert_contains "$(<"$tmp/err")" 'reserved built-in login alias' \
+    'the implicit-alias rejection names the collision'
 
 repo=$(make_repo __missing__)
 out=$(bash "$resolver" --repo-root "$repo" 2> "$tmp/err")
 assert_eq 'provider=none mode=disabled source=missing' "$out" \
     'missing configuration uses the safe disabled plan'
-# A declared-but-unrecognized provider (source=invalid, asserted above) and no
-# declaration at all (source=missing, just above) must never collapse to the
-# same reported source -- that would quietly hide an explicit bad choice
-# behind "nothing was asked for".
-assert_not_contains "$unrecognized_source" 'source=missing' \
-    'a declared-but-unrecognized provider never reports as an undeclared repository'
 assert_contains "$(<"$tmp/err")" 'using effective none' \
     'missing configuration warns without blocking'
 
diff --git a/tests/test-review-transition.sh b/tests/test-review-transition.sh
index 00971d3a..e551a9e0 100755
--- a/tests/test-review-transition.sh
+++ b/tests/test-review-transition.sh
@@ -21,6 +21,7 @@ printf '%s\n' provider-resolve >>"$TRANSITION_LOG"
 case ${PROVIDER_MODE:-coderabbit} in
     coderabbit) printf '%s\n' 'provider=coderabbit mode=triggerable source=declared' ;;
     observe) printf '%s\n' 'provider=github-code-quality mode=observe-only source=declared' ;;
+    generic) printf '%s\n' 'provider=chatgpt-codex-connector mode=observe-only source=declared lane=generic-automated' ;;
     none) printf '%s\n' 'provider=none mode=disabled source=declared' ;;
     pair)
         printf '%s\n' 'provider=coderabbit mode=triggerable source=declared'
@@ -343,6 +344,14 @@ assert_contains "$out" 'provider=github-code-quality result=OBSERVE_ONLY' \
 assert_eq '0' "$(grep -c '^comment ' "$tmp/transition.log" || true)" \
     'observe-only provider cannot reach comment posting'
 
+write_auth '[{"name":"chatgpt-codex-connector","action":"observe","source":"capability-default"}]'
+: >"$tmp/transition.log"
+out=$(PROVIDER_MODE=generic run_transition)
+assert_contains "$out" 'provider=chatgpt-codex-connector result=OBSERVE_ONLY' \
+    'an unknown declared provider remains observe-only during ready transition'
+assert_eq '0' "$(grep -c '^comment ' "$tmp/transition.log" || true)" \
+    'an unknown declared provider cannot reach comment posting'
+
 write_auth "$(trigger_action coderabbit)"
 rm -f "$tmp/missing-auth.json"
 set +e
diff --git a/tests/test-rrp-remediation-contract.sh b/tests/test-rrp-remediation-contract.sh
index 2e6cab64..33c09223 100755
--- a/tests/test-rrp-remediation-contract.sh
+++ b/tests/test-rrp-remediation-contract.sh
@@ -67,6 +67,49 @@ assert_contains "$(cat -- "$rrp_skill")" 'finding-ledger.sh" evidence' \
 worker_gate="$skills/review-remote-pr/references/worker-gate.md"
 assert_contains "$(cat -- "$worker_gate")" 'unfocused' \
     'the worker completion report names the unfocused test log'
+assert_contains "$(cat -- "$worker_gate")" 'after the repair commit and before push' \
+    'the repair handback orders full verification on the commit that will be pushed'
+assert_contains "$(cat -- "$worker_gate")" 'clean committed HEAD' \
+    'the repair handback requires the log to bind the committed head'
+fix_prompt=$(sed -n '/## PR-fix-batch worker prompt/,/## Exit Report/p' \
+    "$skills/parallel-issues/references/worker-prompts.md")
+assert_contains "$fix_prompt" 'commit the repair before the final unfocused run' \
+    'the composed fix-worker prompt commits before full verification'
+assert_contains "$fix_prompt" 'push the branch only after that clean committed-HEAD run passes' \
+    'the composed fix-worker prompt cannot push an unverified commit'
+
+# Step 2's runnable phases must not execute the full suite before the worker
+# commits. Execute its first fence with a recording runner, then ensure the
+# committed-head test is a separate fence after an explicit commit boundary.
+ci_fix_section=$(sed -n '/^## Step 2: Fix CI Failures/,/^For red\/green iterations/p' "$rrp_skill")
+fence() {
+    local number=$1
+    awk -v wanted="$number" '
+        /^```bash$/ { count++; capture=(count==wanted); next }
+        /^```$/ { if (capture) exit }
+        capture
+    ' <<<"$ci_fix_section"
+}
+precommit_fence=$(fence 1)
+postcommit_fence=$(fence 2)
+recipe_kit="$tmp/recipe-kit"
+recipe_calls="$tmp/recipe-calls"
+mkdir -p "$recipe_kit/.shared/scripts"
+cat >"$recipe_kit/.shared/scripts/agent-run.sh" <<'EOF'
+#!/usr/bin/env bash
+printf '%s\n' "$*" >>"$AGENTKIT_RECIPE_CALLS"
+EOF
+chmod +x "$recipe_kit/.shared/scripts/agent-run.sh"
+AGENTKIT_RECIPE_CALLS="$recipe_calls" agentkit="$recipe_kit" agentkit_provenance=ok \
+    bash -c "$precommit_fence"
+assert_eq '--cmd lint --if-declared' "$(cat "$recipe_calls")" \
+    'executing the precommit fence cannot run the full test on a dirty repair'
+assert_contains "$postcommit_fence" '"$agent_run" --cmd test' \
+    'committed-head verification has its own runnable fence'
+commit_line=$(grep -nF 'Commit the repair' <<<"$ci_fix_section" | cut -d: -f1 | head -n1)
+test_line=$(grep -nF '"$agent_run" --cmd test' <<<"$ci_fix_section" | cut -d: -f1 | head -n1)
+assert_eq yes "$([[ -n $commit_line && -n $test_line && $commit_line -lt $test_line ]] && printf yes || printf no)" \
+    'the actual commit step precedes the runnable full-test phase'
 
 # --- item 6: the spawn contract names the primary checkout's ledger ----------
 assert_contains "$(cat -- "$skills/.shared/spawn-contract.md")" "primary checkout's \`.agent/runs/active-workers.ndjson\`" \
@@ -123,8 +166,10 @@ assert_contains "$after_repair" 'RUN_DIR="$RUN_DIR" "$agentkit/review-remote-pr/
 assert_contains "$after_repair" '--repair-sha' 'the evidence step names the repair commit'
 assert_not_contains "$after_repair" '--reviewed-head' 'the evidence step needs no reviewed head'
 assert_not_contains "$(cat -- "$adv_ref")" '--reviewed-head' 'the evidence contract needs no reviewed head'
-assert_not_contains "$(cat -- "$worker_gate")" 'after the commit, so its header' \
-    'workers are not told to re-run the suite after committing'
+assert_contains "$(cat -- "$adv_ref")" 'after the repair commit and before push' \
+    'the evidence recipe says when the binding full run occurs'
+assert_contains "$(cat -- "$adv_ref")" 'tested head and tracked-tree cleanliness' \
+    'the evidence recipe explains what the log binding proves'
 assert_not_contains "$(cat -- "$adv_ref")" 'defaults to the last commit' \
     'the evidence contract no longer promises a guessed repair commit'
 
diff --git a/tests/test-run-state-summary.sh b/tests/test-run-state-summary.sh
index 7ff41df3..63898aac 100755
--- a/tests/test-run-state-summary.sh
+++ b/tests/test-run-state-summary.sh
@@ -38,6 +38,60 @@ printf '%s\n' \
     >"$ledger"
 chmod 600 -- "$ledger"
 
+# Auto-review handoff cannot succeed until every opened PR has either an
+# adversarial receipt or a verified skip. Non-auto-review runs keep the same
+# summary behavior.
+printf '%s\n' \
+    '{"opened_prs":[594,595],"queued":[],"receipt_prs":[594],"skipped_prs":[],"auto_review":false}' >"$state"
+assert_rc 0 'non-auto-review summary permits opened PRs without review coverage' -- \
+    "$script" summary --run-id wave --repo-root "$repo"
+
+printf '%s\n' \
+    '{"opened_prs":[594,595],"queued":[],"receipt_prs":[594],"skipped_prs":[],"auto_review":true}' >"$state"
+missing_review_rc=0
+missing_review_out="$tmp/missing-review.out"
+missing_review_err="$tmp/missing-review.err"
+"$script" summary --run-id wave --repo-root "$repo" --reports-dir "$reports" \
+    >"$missing_review_out" 2>"$missing_review_err" || missing_review_rc=$?
+assert_eq 1 "$missing_review_rc" 'auto-review summary refuses uncovered opened PRs'
+assert_contains "$(cat "$missing_review_err")" '595' 'auto-review refusal names every uncovered PR'
+assert_contains "$(cat "$missing_review_err")" '/review-remote-pr --auto-review 595' \
+    'auto-review refusal prints the exact review resume command'
+assert_contains "$(cat "$missing_review_out")" 'coverage= prs=2 receipts=1 skipped=0 parked=2 queued=0' \
+    'auto-review refusal preserves coverage output'
+assert_contains "$(cat "$missing_review_out")" 'blocked=101:src/one.sh' \
+    'auto-review refusal preserves parked-worker evidence'
+assert_contains "$(cat "$missing_review_out")" 'spec-verification= issue=103' \
+    'auto-review refusal preserves durable verification reports'
+
+for report_case in omitted absent; do
+    early_out="$tmp/missing-review-$report_case.out"
+    early_err="$tmp/missing-review-$report_case.err"
+    early_rc=0
+    early_args=()
+    [[ $report_case == omitted ]] || early_args=(--reports-dir "$tmp/absent-reports")
+    "$script" summary --run-id wave --repo-root "$repo" "${early_args[@]}" \
+        >"$early_out" 2>"$early_err" || early_rc=$?
+    assert_eq 1 "$early_rc" "auto-review refusal survives the $report_case reports early-return path"
+    assert_contains "$(cat "$early_out")" 'blocked=101:src/one.sh' \
+        "auto-review refusal preserves parked evidence with $report_case reports"
+done
+
+printf '%s\n' \
+    '{"opened_prs":[594,595],"queued":[],"receipt_prs":[],"skipped_prs":[],"auto_review":true}' >"$state"
+multiple_missing_err=$("$script" summary --run-id wave --repo-root "$repo" 2>&1 >/dev/null) || true
+assert_contains "$multiple_missing_err" \
+    '/review-remote-pr --auto-review 594; /review-remote-pr --auto-review 595' \
+    'auto-review refusal prints one exact review invocation per uncovered PR'
+
+printf '%s\n' \
+    '{"opened_prs":[594,595],"queued":[],"receipt_prs":[594],"skipped_prs":[595],"auto_review":true}' >"$state"
+assert_rc 0 'auto-review summary accepts complete receipt and skip coverage' -- \
+    "$script" summary --run-id wave --repo-root "$repo"
+
+printf '%s\n' \
+    '{"opened_prs":[],"queued":[103],"receipt_prs":[],"skipped_prs":[],"root_turns":[true,true,true,true,true,true,true],"first_completion":true}' >"$state"
+
 expected=$'coverage= prs=0 receipts=0 skipped=0 parked=2 queued=1 root-turns-before-first-completion=7\nblocked=101:src/one.sh\nblocked=102:src/two.sh,tests/two.sh\nspec-verification= issue=103 steps=4 covered=1 uncovered=3 uncovered-steps=2,3,4 coverage=1/4 classification=majority-uncovered'
 assert_eq "$expected" \
     "$(cd -- "$tmp" && "$script" summary --run-id wave --repo-root "$repo" --reports-dir "$reports")" \
@@ -176,4 +230,35 @@ large_output=$(PATH="$fake_bin:$PATH" SUMMARY_REAL_GIT="$(command -v git)" SUMMA
 assert_eq 0 "$large_rc" 'primary worktree selection consumes a large porcelain stream without SIGPIPE'
 assert_contains "$large_output" 'coverage= prs=0' 'large worktree selection still resolves the primary ledger'
 
+skill_text=$(tr '\n' ' ' <"$root/agentkit/skills/parallel-issues/SKILL.md" | tr -s '[:space:]' ' ')
+# shellcheck disable=SC2016
+auto_review_get_recipe='auto_review_state=$("$agentkit/.shared/scripts/run-state.sh" get --run-id "$RUN_ID" --repo-root "$repository_root" --path auto_review) || exit 1'
+# shellcheck disable=SC2016
+auto_review_case_recipe='case $auto_review_state in true|false)'
+# shellcheck disable=SC2016
+auto_review_value='"$auto_review_state"'
+# shellcheck disable=SC2016
+auto_review_default='${auto_review:-false}'
+assert_contains "$skill_text" \
+    "$auto_review_get_recipe" \
+    'Collect restores the durable auto-review mode before either PR-open path'
+assert_contains "$skill_text" "$auto_review_case_recipe" \
+    'Collect refuses a restored auto-review value outside the boolean boundary'
+completion_recipe=$(grep -F -- '- **Completion report (branch + pushed SHA)**' "$root/agentkit/skills/parallel-issues/SKILL.md")
+blocked_recipe=$(grep -F -- '- **BLOCKED**' "$root/agentkit/skills/parallel-issues/SKILL.md")
+assert_contains "$completion_recipe" "$auto_review_value" \
+    'normal PR-open completion prints the restored auto-review mode'
+assert_contains "$blocked_recipe" "$auto_review_value" \
+    'partial-pushed PR-open completion prints the restored auto-review mode'
+assert_not_contains "$completion_recipe$blocked_recipe" "$auto_review_default" \
+    'neither Collect completion path can default a resumed auto-review run to false'
+# The literal expansion is the unsafe recipe under test.
+# shellcheck disable=SC2016
+assert_not_contains "$skill_text" '--path auto_review --json "${auto_review:-false}"' \
+    'auto-review persistence never defaults an unset shell variable to false'
+assert_contains "$skill_text" '--path auto_review --json true' \
+    'auto-review invocation facts persist the literal true value'
+assert_contains "$skill_text" '--path auto_review --json false' \
+    'non-auto-review invocation facts persist the literal false value'
+
 finish
diff --git a/tests/test-skill-size.sh b/tests/test-skill-size.sh
index aa2cebbf..940c7092 100755
--- a/tests/test-skill-size.sh
+++ b/tests/test-skill-size.sh
@@ -220,9 +220,9 @@ mkdir -p "$root/parallel-issues"
 } > "$root/parallel-issues/SKILL.md"
 run_lint "$root"
 assert_eq '1' "$LINT_RC" 'the parallel-issues ratchet fixture exceeds its measured ceiling'
-assert_contains "$LINT_OUT" 'past its ratcheted ceiling of 761 lines' \
+assert_contains "$LINT_OUT" 'past its ratcheted ceiling of 772 lines' \
     'the parallel-issues line ratchet pins the extracted-recipe ceiling'
-assert_contains "$LINT_OUT" 'past its ratcheted ceiling of 16265 tokens' \
+assert_contains "$LINT_OUT" 'past its ratcheted ceiling of 16666 tokens' \
     'the parallel-issues token ratchet pins the extracted-recipe ceiling'
 
 # A bad allowlist field must be named, never evaluated. Under `set -u` these