@cxi-lmai/ci-agent-platform 3.1.4 → 3.1.5

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cxi-lmai/ci-agent-platform",
-  "version": "3.1.4",
+  "version": "3.1.5",
   "description": "Autonomous dev pipeline on plain GitLab CI or GitHub Actions, driven by Claude Code. A labeled issue goes in, an open merge request comes out.",
   "keywords": [
     "claude",

package/payload/ci-templates/scripts/lib/pipeline-common.sh

@@ -277,10 +277,16 @@ pipe_run_claude() {
   # $3 = allowed tools (optional), $4 = model for the main loop (optional,
   # e.g. $PIPE_MODEL_REVIEW; subagents keep their frontmatter models),
   # stdin = extra prompt (optional, usually empty)
+  #
+  # Always returns 0, so a caller running under `set -e` keeps control over what
+  # a failed agent means for its job. The agent's real exit code lands in
+  # PIPE_AGENT_RC. A caller that must not read a crashed agent as a clean run
+  # checks that variable, or calls pipe_require_agent_ran below.
   local label="$1" skill="$2" tools="${3:-Agent,Read,Write,Edit,Glob,Grep,Bash}" model="${4:-}"
   pipe_require_harness claude
   local stdin_file; stdin_file=$(mktemp)
   cat > "$stdin_file" || true
+  PIPE_AGENT_RC=0
 
   if [ "$(id -u)" -eq 0 ]; then
     local runuser="pipelinebot"
@@ -315,7 +321,7 @@ pipe_run_claude() {
     } > "$runner"
     chmod +x "$runner"
     chown "$runuser:$runuser" "$runner"
-    su -m "$runuser" -s /bin/bash "$runner" || true
+    su -m "$runuser" -s /bin/bash "$runner" || PIPE_AGENT_RC=$?
     chown -R root:root "$PWD" 2>/dev/null || true
     rm -f "$runner"
   else
@@ -327,9 +333,32 @@ pipe_run_claude() {
       else
         capture_claude "$label" -- --dangerously-skip-permissions --allowedTools "$tools" -p "$skill" < "$stdin_file"
       fi
-    ) || true
+    ) || PIPE_AGENT_RC=$?
   fi
   rm -f "$stdin_file"
+  [ "$PIPE_AGENT_RC" -eq 0 ] || pipe_log "agent exited with code $PIPE_AGENT_RC"
+  return 0
+}
+
+pipe_require_agent_ran() {
+  # $1 = file the skill was supposed to write (empty to check the exit code
+  # only, for skills whose marker is legitimately optional), $2 = human label.
+  # Fails the job when the agent crashed or wrote nothing. Without this an
+  # agent that never started is indistinguishable from one that ran and found
+  # nothing: every caller reads its marker file, misses it, and falls back to
+  # the "all clear" default, so the job goes green without a review, fix, or
+  # triage having happened.
+  local marker="$1" what="${2:-agent}"
+  if [ "${PIPE_AGENT_RC:-0}" -ne 0 ]; then
+    pipe_log "ERROR: $what did not complete (exit code $PIPE_AGENT_RC). See the job log above."
+    return 1
+  fi
+  [ -n "$marker" ] || return 0
+  if [ ! -s "$marker" ]; then
+    pipe_log "ERROR: $what produced no output ($marker is missing or empty)."
+    return 1
+  fi
+  return 0
 }
 
 # --- omp invocation -----------------------------------------------------
@@ -418,6 +447,9 @@ pipe_run_omp() {
   # internally), $4 = model for the main loop (optional, e.g.
   # $PIPE_MODEL_REVIEW; subagents keep the models from their own
   # frontmatter), stdin = extra prompt (optional, usually empty).
+  #
+  # Same contract as pipe_run_claude: always returns 0 and reports the agent's
+  # real exit code in PIPE_AGENT_RC, which pipe_require_agent_ran checks.
   local label="$1" skill="$2" tools="${3:-Agent,Read,Write,Edit,Glob,Grep,Bash}" model="${4:-}"
   local omp_tools; omp_tools=$(pipe_translate_tools_to_omp "$tools")
   # Bound the run below the job's own hard kill (see pipe_agent_max_time).
@@ -431,6 +463,7 @@ pipe_run_omp() {
   pipe_require_harness omp
   local stdin_file; stdin_file=$(mktemp)
   cat > "$stdin_file" || true
+  PIPE_AGENT_RC=0
 
   if [ "$(id -u)" -eq 0 ]; then
     local runuser="pipelinebot"
@@ -465,7 +498,7 @@ pipe_run_omp() {
     } > "$runner"
     chmod +x "$runner"
     chown "$runuser:$runuser" "$runner"
-    su -m "$runuser" -s /bin/bash "$runner" || true
+    su -m "$runuser" -s /bin/bash "$runner" || PIPE_AGENT_RC=$?
     chown -R root:root "$PWD" 2>/dev/null || true
     rm -f "$runner"
   else
@@ -477,9 +510,11 @@ pipe_run_omp() {
       else
         capture_omp "$label" -- "${max_time_args[@]}" --yolo --no-session --no-title --tools "$omp_tools" -p "$skill" < "$stdin_file"
       fi
-    ) || true
+    ) || PIPE_AGENT_RC=$?
   fi
   rm -f "$stdin_file"
+  [ "$PIPE_AGENT_RC" -eq 0 ] || pipe_log "agent exited with code $PIPE_AGENT_RC"
+  return 0
 }
 
 # --- Harness dispatch ---------------------------------------------------

package/payload/ci-templates/scripts/review-fix.sh

@@ -41,6 +41,9 @@ echo "$NOTES" | jq -r --arg u "${PIPE_BOT_USER:-}" --arg m "$REVIEW_MARKER" \
 #    verifies, writes the result JSON, and the review-fix.env marker. On cap it
 #    runs the shared postmortem flow (writing postmortem.md + postmortem.env).
 pipe_run_agent coder "/fix-review-findings" "Agent,Read,Write,Edit,Glob,Grep,Bash" "$PIPE_MODEL_CODE" < /dev/null
+# A crashed agent writes no marker, which step 6 would read as "nothing changed"
+# and step 7 would report as a completed fix run that had nothing to do.
+pipe_require_agent_ran "" "The review-fix agent" || exit 1
 
 # 6. Read the marker the skill wrote.
 RF_ENV="$PIPE_CONTEXT_DIR/review-fix.env"

package/payload/ci-templates/scripts/review.sh

@@ -87,7 +87,14 @@ pipe_metric_event code-reviewer review_findings \
       --argjson inc "$([ "${PIPE_IS_INCREMENTAL:-false}" = "true" ] && echo true || echo false)" \
       '{findings_total: $t, status: $s, has_fixable_bugs: $b, is_incremental: $inc}')"
 
-# 7. Gate. review.env is exported so the fix job can read REVIEW_HAS_BUGS at
+# 7. Fail the job when no review was produced. The comment above already told
+#    the humans; this tells the pipeline. PIPE_NO_GATE does not apply: it turns
+#    off the verdict gate, and this is a broken run, not a verdict. Without it a
+#    review that never ran (crashed agent, missing ANTHROPIC_API_KEY) reads as a
+#    review that ran and found nothing, and the MR goes green unreviewed.
+pipe_require_agent_ran "$PIPE_RESULT_REVIEW" "The review agent" || exit 1
+
+# 8. Gate. review.env is exported so the fix job can read REVIEW_HAS_BUGS at
 #    runtime. On GitLab the job exits non-zero to block the pipeline. On GitHub
 #    the workflow step owns the gate (it reads the output first), so it sets
 #    PIPE_NO_GATE=1 and this script leaves the exit to the workflow.

package/payload/ci-templates/scripts/test-fix.sh

@@ -27,6 +27,9 @@ COVERAGE_BEFORE=$(pipe_count_fix_commits "$PIPE_COMMIT_COVERAGE")
 #    test-writer, verifies, and writes fix-tests.env. On cap it runs the shared
 #    postmortem flow (writing postmortem.md + postmortem.env).
 pipe_run_agent test-fix "/fix-tests" "Agent,Read,Write,Edit,Glob,Grep,Bash" "$PIPE_MODEL_CODE" < /dev/null
+# A crashed agent writes no marker, which step 4 would read as FIX_BRANCH=none
+# and report as a false-alarm trigger, hiding the failure behind a green job.
+pipe_require_agent_ran "" "The test-fix agent" || exit 1
 
 # 4. Read the marker the skill wrote.
 FT_ENV="$PIPE_CONTEXT_DIR/fix-tests.env"