diff --git a/README.md b/README.md
index 3172bf2..43b888d 100644
--- a/README.md
+++ b/README.md
@@ -178,7 +178,7 @@ bd create "Fix the login timeout" --description "..."     # Shep (the Grok bot)
 
 - **Prompt** — one line pointing at the issue (`Your job is beads issue <id>. Run bd show <id> …`), the same for every harness.
 - **Context** — before it starts the agent, bercail adds a `bd comment` to the issue with the worktree and, headless, the result file and session id.
-- **Result** — a headless run leaves `{status: ready|blocked|failed, summary, session_id}`. Read that file, not terminal output. `--resume SESSION` continues the same session (`claude --resume`, `codex exec resume`, `cursor-agent --resume`, `opencode run --session`).
+- **Result** — a headless run leaves `{status: ready|blocked|failed, summary, session_id}`. A plan usage limit also sets `reason` to `usage-limit`, so Shep can propose another plan. Bercail does not switch or rerun. Read that file, not terminal output. `--resume SESSION` continues the same session (`claude --resume`, `codex exec resume`, `cursor-agent --resume`, `opencode run --session`).
 - **Blocked, not switched** — if the stage's binary, `bd`, or the issue is missing, the result is `blocked` and nothing starts. Bercail never falls through to another harness. `--binary claude|codex|cursor-agent|opencode` runs a headless stage on another CLI only when you ask (with `opencode`, `start` stays on Claude and `codex` on GPT, through your opencode plans).
 - **Dry run** — `--dry-run` prints the exact argv for any stage, even when that binary is not installed.
 - `bercail harness` reports which of claude, codex, cursor-agent, and opencode are on PATH, with a local model hint. It does not choose.
diff --git a/docs/bstack/07-handoff.md b/docs/bstack/07-handoff.md
index 686f50d..37e6a94 100644
--- a/docs/bstack/07-handoff.md
+++ b/docs/bstack/07-handoff.md
@@ -19,7 +19,7 @@ spec. The agent's prompt is one line that points at it.
   review it in that space's tuicr.
 - **No branch:** a headless `opencode run --agent bstack`. It writes
   `.bercail/handoff-result.json` (`status` ready, blocked, or failed, a summary,
-  the session id) and nothing opens.
+  the session id; `reason` is `usage-limit` when a plan is exhausted) and nothing opens.
   It edits and stages but can't commit, push, or merge: a headless run rejects
   anything that would ask. Stage `autopilot` is the exception (see
   [Ship with autopilot](./06-autopilot.md)).
diff --git a/skills/handoff/SKILL.md b/skills/handoff/SKILL.md
index eb99d47..bbe0b0a 100644
--- a/skills/handoff/SKILL.md
+++ b/skills/handoff/SKILL.md
@@ -80,7 +80,9 @@ If the requested harness, bd, or the issue is missing, the stage writes
 `<worktree>/.bercail/handoff-result.json` with status `blocked`, starts
 nothing, and exits nonzero. It does not fall through to another harness. The
 agent writes the same result file (`status`, `summary`, and bercail fills
-`session_id`). Read that file. Do not scrape terminal output. For Codex,
+`session_id`). When the run stops on a plan usage limit, bercail adds `reason`:
+`usage-limit`. It does not rerun or switch plans. Read that file. Do not scrape
+terminal output. For Codex,
 `session_id` is the JSONL `thread.started` thread id. For cursor-agent, it is
 the print JSON `session_id`. When there is no resumable session (blocked before
 codex or cursor-agent ran, or no id in its log), `session_id` is `none`;
diff --git a/skills/handoff/scripts/handoff-stages.sh b/skills/handoff/scripts/handoff-stages.sh
index 1c39105..73545b9 100644
--- a/skills/handoff/scripts/handoff-stages.sh
+++ b/skills/handoff/scripts/handoff-stages.sh
@@ -106,14 +106,18 @@ handoff_issue_prompt() {
   printf 'Your job is beads issue %s. Run `bd show %s` (read its comments too) and do the work it describes.' "$1" "$1"
 }
 
+# reason is optional. Empty omits the field; Shep reads it, and unknown fields
+# do not fail handoff_result_ok. The only value the runner writes is usage-limit.
 handoff_write_result() {
-  local dest="$1" status="$2" summary="$3" session="$4"
+  local dest="$1" status="$2" summary="$3" session="$4" reason="${5:-}"
   mkdir -p "$(dirname "$dest")"
   jq -nc \
     --arg status "$status" \
     --arg summary "$summary" \
     --arg session_id "$session" \
-    '{status:$status, summary:$summary, session_id:$session_id}' >"$dest"
+    --arg reason "$reason" \
+    '{status:$status, summary:$summary, session_id:$session_id}
+     + if $reason == "" then {} else {reason:$reason} end' >"$dest"
 }
 
 handoff_result_ok() {
@@ -270,6 +274,42 @@ handoff_opencode_session_id() {
   printf '%s' "$id"
 }
 
+# usage-limit when LOG is a plan limit, else nothing. Prints the reason; does not
+# rerun. ChatGPT is the session error type provider.chatgpt-limit, or the code
+# subscription_sharing_usage_limit_exceeded in that error's body. Claude is a
+# sentence Claude Code 2.1.316 composes ("You've hit your 5-hour limit", the
+# weekly, Opus, Sonnet, and Fable lines, "You're out of usage credits") or the
+# recorded extra-usage billing sentence. That sentence is read from the error
+# event, or from stderr when the process exits before one. Other JSON is not a
+# limit. A provider.rate-limit, including a bare HTTP 429, is not one: xAI's
+# weekly-limit body is unverified.
+handoff_usage_limit_reason() {
+  local log="$1" events
+  local phrase="You've hit your (5-hour|weekly|Opus|Sonnet|Fable) limit|You've hit your usage limit|You've hit your limit|You've reached your Fable limit|You're out of usage credits|Third-party apps now draw from your extra usage, not your plan limits"
+  [[ -f "$log" ]] || return 0
+  events="$(jq -c -R 'fromjson? | select(type == "object" and .type == "error" and (.error | type == "object")) | .error' "$log" 2>/dev/null || true)"
+  # A JSON error event decides. Stderr counts only when the process exits first.
+  if [[ -n "$events" ]]; then
+    if jq -e -s --arg phrase "$phrase" '
+      def hit: test($phrase);
+      any(.[];
+        .type == "provider.chatgpt-limit"
+        or ((.response.body // "") | if type == "string" then contains("subscription_sharing_usage_limit_exceeded") else false end)
+        or ((.message // "") | hit)
+        or ((.response.body // "") | if type == "string" then hit else false end)
+      )
+    ' <<<"$events" >/dev/null 2>&1; then
+      printf 'usage-limit'
+    fi
+    return 0
+  fi
+  # grep reads to the end (no -q): with pipefail, an early exit would SIGPIPE jq
+  # on a long log and hide the match.
+  if jq -R -r 'select(try (fromjson | type == "object") catch false | not)' "$log" | grep -E "$phrase" >/dev/null; then
+    printf 'usage-limit'
+  fi
+}
+
 handoff_is_opencode_session() {
   [[ "$1" =~ ^ses_[A-Za-z0-9]+$ ]]
 }
@@ -459,7 +499,7 @@ handoff_headless_comment_text() {
 handoff_run_headless() {
   local stage="$1" worktree="$2" issue="$3" resume="$4" dry="$5" binary_override="${6:-}"
   local mode binary native model_spec model session="" resumed=0 prompt bd beads_dir=""
-  local result log bin cmd_json out rc=0 tid present=false bd_present=false
+  local result log bin cmd_json out rc=0 tid present=false bd_present=false reason=""
 
   command -v jq >/dev/null 2>&1 || handoff_die "jq is required"
   mode="$(handoff_stage_field "$stage" mode)" || handoff_die "unknown stage: $stage"
@@ -593,12 +633,17 @@ handoff_run_headless() {
     session="$tid"
     out="$(jq --arg session_id "$session" '.session_id = $session_id' <<<"$out")"
   fi
+  # The exhausted plan cannot be trusted to name its own limit. Classify the CLI
+  # log and stamp reason; do not start another command.
+  if [[ "$rc" -ne 0 ]]; then
+    reason="$(handoff_usage_limit_reason "$log")"
+  fi
   # bercail knows the session id; the agent only has to write status and summary.
   # Keep the agent's own uuid if the CLI log had none. Never invent one: a
   # made-up id would look resumable. Say so in the summary instead.
   if [[ -f "$result" ]] && jq -e 'type == "object"' "$result" >/dev/null 2>&1; then
     local filled
-    filled="$(jq --arg session_id "$session" --arg none "$HANDOFF_NO_SESSION" '
+    filled="$(jq --arg session_id "$session" --arg none "$HANDOFF_NO_SESSION" --arg reason "$reason" '
       (.session_id | type == "string" and test("^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}$")) as $own
       | .status = (if (.status == "ready" or .status == "blocked" or .status == "failed") then .status else "failed" end)
       | .summary = (if (.summary | type == "string" and length > 0) then .summary else "headless stage finished without a usable summary" end)
@@ -606,13 +651,14 @@ handoff_run_headless() {
         elif $own then .
         else .session_id = $none | .summary += " (bercail found no session id in the CLI log, so this run cannot be resumed.)"
         end
+      | if $reason != "" then .reason = $reason else del(.reason) end
     ' "$result")"
     printf '%s\n' "$filled" >"$result"
   fi
   if [[ ! -f "$result" ]] || ! handoff_result_ok "$result"; then
     [[ "$rc" -eq 0 ]] && rc=1
     [[ -n "$session" ]] || session="$HANDOFF_NO_SESSION"
-    handoff_write_result "$result" failed "headless stage exited ${rc} without a result file" "$session"
+    handoff_write_result "$result" failed "headless stage exited ${rc} without a result file" "$session" "$reason"
   fi
   if [[ "$rc" -ne 0 ]]; then
     printf 'handoff: %s exited %s; result is %s\n' "$bin" "$rc" "$result" >&2
diff --git a/tests/handoff_spawn_test.sh b/tests/handoff_spawn_test.sh
index e5cf5b8..2a5ceac 100755
--- a/tests/handoff_spawn_test.sh
+++ b/tests/handoff_spawn_test.sh
@@ -133,13 +133,11 @@ test_info_json_reports_graphite_and_dirty() {
   git_init "$repo"
   printf 'dirty\n' >>"$repo/file.txt"
   printf 'loose\n' >"$repo/untracked.txt"
-  mkdir -p "$(dirname "$(git -C "$repo" rev-parse --git-path .graphite_repo_config)")"
+  # git -C prints the path relative to $repo, not to the cwd.
   printed="$(git -C "$repo" rev-parse --git-path .graphite_repo_config)"
-  if [[ "$printed" == /* ]]; then
-    printf '{}\n' >"$printed"
-  else
-    printf '{}\n' >"$repo/$printed"
-  fi
+  [[ "$printed" == /* ]] || printed="$repo/$printed"
+  mkdir -p "$(dirname "$printed")"
+  printf '{}\n' >"$printed"
   printf '#!/bin/sh\nexit 1\n' >"$TMP_DIR/no-herdr"
   chmod +x "$TMP_DIR/no-herdr"
   unset HC_TERMINAL_ID HANDOFF_SPACE
diff --git a/tests/usage_limit_reason_test.sh b/tests/usage_limit_reason_test.sh
new file mode 100644
index 0000000..5a8607c
--- /dev/null
+++ b/tests/usage_limit_reason_test.sh
@@ -0,0 +1,169 @@
+#!/usr/bin/env bash
+# shellcheck shell=bash
+# A stub opencode replays recorded plan-limit errors. No live provider calls.
+set -euo pipefail
+
+unset BASH_ENV
+export __MISE_BASH_ENV_LOADED=1
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+TMP_DIR="$(mktemp -d)"
+trap 'rm -rf "$TMP_DIR"' EXIT
+
+fail() { printf 'FAIL: %s\n' "$*" >&2; exit 1; }
+
+# shellcheck disable=SC1090
+source "$ROOT/skills/handoff/scripts/handoff-spawn"
+
+ISSUE=bercail-0s1
+
+mkdir -p "$TMP_DIR/bin"
+cat >"$TMP_DIR/bin/bd" <<'STUB'
+#!/usr/bin/env bash
+set -euo pipefail
+case "$1" in
+  show)
+    [[ "$2" == "${BD_KNOWN:?}" ]] || { echo "no issue $2" >&2; exit 1; }
+    printf '[{"id":"%s"}]\n' "$2"
+    ;;
+  comment)
+    [[ "${3:-}" == --stdin ]] || exit 3
+    cat >/dev/null
+    ;;
+  context)
+    printf '{"beads_dir":"%s"}\n' "$PWD/.beads"
+    ;;
+  *) exit 2 ;;
+esac
+STUB
+chmod +x "$TMP_DIR/bin/bd"
+export HANDOFF_BD_BIN="$TMP_DIR/bin/bd"
+export BD_KNOWN="$ISSUE"
+
+# Prints $HANDOFF_STUB_PAYLOAD and exits 1. Counts invocations so a limit cannot
+# start a second command.
+cat >"$TMP_DIR/bin/opencode" <<'STUB'
+#!/usr/bin/env bash
+set -euo pipefail
+printf '%s\n' "$*" >>"${HANDOFF_ARGV_LOG:?}"
+printf '1\n' >>"${HANDOFF_RUNS:?}"
+cat "${HANDOFF_STUB_PAYLOAD:?}"
+exit 1
+STUB
+chmod +x "$TMP_DIR/bin/opencode"
+export HANDOFF_OPENCODE_BIN="$TMP_DIR/bin/opencode"
+export HANDOFF_ARGV_LOG="$TMP_DIR/argv.log"
+export HANDOFF_RUNS="$TMP_DIR/runs"
+export HANDOFF_STUB_PAYLOAD="$TMP_DIR/payload"
+
+run_stage() {
+  local stage="$1" name="$2" rc=0
+  local wt="$TMP_DIR/$name"
+  mkdir -p "$wt"
+  : >"$HANDOFF_ARGV_LOG"
+  : >"$HANDOFF_RUNS"
+  (cd "$wt" && "$ROOT/skills/handoff/scripts/handoff-spawn" \
+    --stage "$stage" --issue "$ISSUE" --worktree "$wt" >/dev/null 2>&1) || rc=$?
+  [[ "$rc" -ne 0 ]] || fail "$name should exit nonzero"
+  [[ "$(wc -l <"$HANDOFF_RUNS")" -eq 1 ]] || fail "$name started more than once: $(cat "$HANDOFF_RUNS")"
+  printf '%s' "$wt"
+}
+
+assert_limit() {
+  local wt="$1" label="$2"
+  jq -e '.status == "failed"
+    and .reason == "usage-limit"
+    and (.summary | type == "string" and length > 0)
+    and (.session_id | type == "string" and length > 0)' \
+    "$wt/.bercail/handoff-result.json" >/dev/null \
+    || fail "$label: $(cat "$wt/.bercail/handoff-result.json")"
+  handoff_result_ok "$wt/.bercail/handoff-result.json" \
+    || fail "$label result should still pass handoff_result_ok"
+}
+
+test_chatgpt_limit_sets_reason() {
+  # Recorded ChatGPT session error: type provider.chatgpt-limit, code in the body.
+  # Live capture was HTTP 200, not a 429.
+  jq -nc \
+    --arg body '{"error":{"code":"subscription_sharing_usage_limit_exceeded","message":"Limit"}}' \
+    '{type:"error", sessionID:"ses_gptlimit", error:{
+      type:"provider.chatgpt-limit",
+      message:"ChatGPT usage limit reached. Check ChatGPT Settings → Usage for details.",
+      status:200,
+      response:{body:$body}
+    }}' >"$HANDOFF_STUB_PAYLOAD"
+  local wt
+  wt="$(run_stage bstack chatgpt)"
+  assert_limit "$wt" "chatgpt"
+  jq -e '.session_id == "ses_gptlimit"' "$wt/.bercail/handoff-result.json" >/dev/null \
+    || fail "chatgpt session: $(cat "$wt/.bercail/handoff-result.json")"
+  wt="$(run_stage autopilot chatgpt-autopilot)"
+  assert_limit "$wt" "autopilot"
+  printf 'PASS: ChatGPT usage-limit error yields failed with reason usage-limit\n'
+}
+
+test_claude_text_sets_reason() {
+  # The sentence Claude Code 2.1.316 composes, as opencode's plain error, and as
+  # stderr when the process exits before a JSON error event.
+  jq -nc '{type:"error", sessionID:"ses_claude", error:{
+    type:"provider.unknown",
+    message:"You'\''ve hit your 5-hour limit · resets 6:10pm"
+  }}' >"$HANDOFF_STUB_PAYLOAD"
+  local wt
+  wt="$(run_stage bstack claude-json)"
+  assert_limit "$wt" "claude json"
+  printf '%s\n' "You've hit your 5-hour limit" >"$HANDOFF_STUB_PAYLOAD"
+  wt="$(run_stage bstack claude-text)"
+  assert_limit "$wt" "claude text"
+  printf 'PASS: Claude 5-hour limit text yields failed with reason usage-limit\n'
+}
+
+test_rate_limit_has_no_reason() {
+  # A transient 429. Status alone is not a plan limit (xAI's weekly body is unverified).
+  jq -nc \
+    --arg body '{"error":{"type":"rate_limit_error","message":"Too many requests"}}' \
+    '{type:"error", sessionID:"ses_rate", error:{
+      type:"provider.rate-limit",
+      message:"Too many requests",
+      status:429,
+      response:{body:$body}
+    }}' >"$HANDOFF_STUB_PAYLOAD"
+  local wt
+  wt="$(run_stage bstack rate)"
+  jq -e '.status == "failed"
+    and (.summary | type == "string" and length > 0)
+    and .session_id == "ses_rate"
+    and (has("reason") | not)' \
+    "$wt/.bercail/handoff-result.json" >/dev/null \
+    || fail "rate limit must not set reason: $(cat "$wt/.bercail/handoff-result.json")"
+  handoff_result_ok "$wt/.bercail/handoff-result.json" \
+    || fail "rate-limit result should still pass handoff_result_ok"
+
+  # A mention in other JSON, or a bare 429 with no code, is not a limit.
+  printf '%s\n' '{"type":"text","text":"subscription_sharing_usage_limit_exceeded"}' >"$TMP_DIR/mention.log"
+  [[ -z "$(handoff_usage_limit_reason "$TMP_DIR/mention.log")" ]] \
+    || fail "a mention of the code is not a usage limit"
+  printf '%s\n' '{"type":"text","text":"You'"'"'ve hit your 5-hour limit"}' >"$TMP_DIR/quote.log"
+  [[ -z "$(handoff_usage_limit_reason "$TMP_DIR/quote.log")" ]] \
+    || fail "a quoted phrase in other JSON is not a usage limit"
+  printf '%s\n' ' {"type":"text","text":"You'"'"'ve hit your 5-hour limit"}' >"$TMP_DIR/indented.log"
+  [[ -z "$(handoff_usage_limit_reason "$TMP_DIR/indented.log")" ]] \
+    || fail "indented JSON is not stderr"
+  jq -nc '{type:"error", error:{type:"provider.unknown", message:"You'\''re out of usage credits"}}' >"$TMP_DIR/credits.log"
+  [[ "$(handoff_usage_limit_reason "$TMP_DIR/credits.log")" == usage-limit ]] \
+    || fail "out of usage credits should be a plan limit"
+  jq -nc '{type:"error", error:{type:"provider.unknown", message:"HTTP 429", status:429}}' >"$TMP_DIR/bare429.log"
+  [[ -z "$(handoff_usage_limit_reason "$TMP_DIR/bare429.log")" ]] \
+    || fail "status 429 alone is not a usage limit"
+  # The error event decides. A phrase on another line does not override a 429.
+  printf '%s\n' \
+    '{"type":"error","error":{"type":"provider.rate-limit","message":"Too many requests","status":429}}' \
+    "You've hit your 5-hour limit" >"$TMP_DIR/mixed.log"
+  [[ -z "$(handoff_usage_limit_reason "$TMP_DIR/mixed.log")" ]] \
+    || fail "a rate-limit event must not pick up a phrase elsewhere"
+  printf 'PASS: a plain 429 yields no usage-limit reason\n'
+}
+
+test_chatgpt_limit_sets_reason
+test_claude_text_sets_reason
+test_rate_limit_has_no_reason