From 0dad681a5d03228d98828832355238cf3d683a79 Mon Sep 17 00:00:00 2001
From: Joe Saunderson <joe.saunderson@mention-me.com>
Date: Fri, 10 Jul 2026 13:31:18 +0100
Subject: [PATCH] Fix worker: skill not invoked headlessly; false success

Two bugs stopped the sweep actually running:
- `--setting-sources ""` disabled skill discovery, so `/babysit-prs`
  errored "Unknown command". Drop it; the tool allowlist still constrains,
  and the sweep is a skill so it must be discoverable.
- exit 0 was treated as success even when nothing ran. Switch to
  `--output-format json` and gate success on is_error/subtype/result.

Verified: headless dry-run now returns is_error=false, subtype=success,
and a real classified sweep summary.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HjC3rKLYTSqVqf6YG2taPc
---
 bin/babysit-prs-worker.sh | 34 +++++++++++++++++++++++++++-------
 1 file changed, 27 insertions(+), 7 deletions(-)

diff --git a/bin/babysit-prs-worker.sh b/bin/babysit-prs-worker.sh
index a1bb7ba..9f9586a 100755
--- a/bin/babysit-prs-worker.sh
+++ b/bin/babysit-prs-worker.sh
@@ -52,8 +52,11 @@ if [[ "$PERMISSION_MODE" == "bypassPermissions" ]]; then
   log_warn "Running with bypassPermissions (operator opt-in): the approval gate is OFF."
   permission_args=(--permission-mode bypassPermissions)
 else
-  # Explicit allowlist, no settings files loaded. Anything unlisted is denied.
-  permission_args=(--permission-mode default --setting-sources "" \
+  # Explicit allowlist. Do NOT pass --setting-sources "": that disables skill
+  # discovery, and the sweep itself is a skill (/babysit-prs), so the run would
+  # no-op with "Unknown command". The allowlist still constrains tools; any
+  # rules in ~/.claude/settings.json also apply.
+  permission_args=(--permission-mode default \
     --allowedTools "${BABYSIT_ALLOWED_TOOLS[@]}")
 fi
 
@@ -73,20 +76,37 @@ else
 fi
 
 set +e
-"${run_prefix[@]}" \
+output="$("${run_prefix[@]}" \
   claude --print \
     --session-id "$SESSION_ID" \
     "${permission_args[@]}" \
-    --output-format text \
-    "$PROMPT"
+    --output-format json \
+    "$PROMPT")"
 exit_code=$?
 set -e
 
 if [[ $exit_code -eq 124 ]]; then
   log_error "Sweep timed out after ${BABYSIT_RUN_TIMEOUT_SECONDS}s (session ${SESSION_ID})"
-elif [[ $exit_code -ne 0 ]]; then
+  exit 124
+fi
+
+# claude --print exits 0 even when it did nothing (e.g. an unknown slash
+# command), so success is judged from the JSON envelope, not the exit code.
+is_error="$(jq -r '.is_error // true' <<<"$output" 2>/dev/null || echo true)"
+subtype="$(jq -r '.subtype // "unknown"' <<<"$output" 2>/dev/null || echo unknown)"
+result="$(jq -r '.result // ""' <<<"$output" 2>/dev/null || echo "")"
+
+log_info "Sweep result: ${result}"
+
+if [[ $exit_code -ne 0 ]]; then
   log_error "Sweep failed with exit code ${exit_code} (resume: cd ${REPO_ROOT} && claude --resume ${SESSION_ID})"
+  exit "$exit_code"
+elif [[ "$is_error" == "true" || "$subtype" != "success" ]]; then
+  log_error "Sweep did not complete cleanly (is_error=${is_error}, subtype=${subtype})."
+  exit 3
+elif [[ "$result" == *"Unknown command"* || -z "$result" ]]; then
+  log_error "Sweep produced no result; the skill was likely not invoked."
+  exit 3
 else
   log_success "Sweep finished (session ${SESSION_ID})"
 fi
-exit "$exit_code"