@rpamis/comet 0.4.0-beta.7 → 0.4.0-beta.9

package/eval/pyproject.toml

@@ -29,10 +29,10 @@ dev = [
     "matplotlib>=3.8",
 ]
 
-[tool.pytest.ini_options]
-testpaths = ["local/tests"]
-pythonpath = ["."]
-markers = [
+[tool.pytest.ini_options]
+testpaths = ["local/tests"]
+pythonpath = ["."]
+markers = [
     "langsmith: marks test as using LangSmith tracing",
 ]
 

package/eval/scaffold/python/aligned_comparison.py

@@ -269,6 +269,7 @@ def build_execution_identity(
         "simulator_prompt": getattr(interaction, "simulator_prompt", None),
         "decision_patterns": list(getattr(interaction, "decision_patterns", ()) or ()),
         "decision_reply": getattr(interaction, "decision_reply", None),
+        "decision_replies": list(getattr(interaction, "decision_replies", []) or []),
         "continue_prompt": getattr(interaction, "continue_prompt", None),
         "fresh_resume_marker": getattr(interaction, "fresh_resume_marker", None),
     }

package/eval/scaffold/python/tasks.py

@@ -100,6 +100,7 @@ class InteractionConfig:
     simulator_prompt: str | None = None
     decision_patterns: list[str] = field(default_factory=list)
     decision_reply: str | None = None
+    decision_replies: list[str] = field(default_factory=list)
     continue_prompt: str = "Please continue with the next phase of the workflow."
     fresh_resume_marker: str | None = None
 
@@ -287,6 +288,7 @@ def load_task(name: str, tasks_dir: Path | None = None) -> Task:
         simulator_prompt=interaction.get("simulator_prompt"),
         decision_patterns=interaction.get("decision_patterns", []),
         decision_reply=interaction.get("decision_reply"),
+        decision_replies=interaction.get("decision_replies", []),
         continue_prompt=interaction.get(
             "continue_prompt",
             "Please continue with the next phase of the workflow.",

package/eval/scaffold/shell/docker.sh

@@ -456,7 +456,7 @@ docker_run_claude_loop() {
                 expected_image_id="$2"
                 shift 2
                 ;;
-            --max-turns|--model|--simulator-prompt-file|--decision-reply|--continue-prompt|--decision-pattern|--fresh-resume-marker)
+            --max-turns|--model|--simulator-prompt-file|--decision-reply|--decision-reply-step|--continue-prompt|--decision-pattern|--fresh-resume-marker)
                 loop_args+=("$1" "$2")
                 shift 2
                 ;;

package/eval/scaffold/shell/run-claude-loop.sh

@@ -11,6 +11,7 @@
 #   --model MODEL             Model for both subject and simulator
 #   --simulator-prompt-file   File containing the simulator system prompt
 #   --decision-reply TEXT     Deterministic reply for each detected decision point
+#   --decision-reply-step TEXT  Queue one deterministic reply for the next decision point
 #   --continue-prompt TEXT    Nudge used when the workflow should continue
 #   --decision-pattern TEXT   Extra case-insensitive substring to treat as a decision point
 #   --fresh-resume-marker TEXT  Start the following turn in a new subject session
@@ -37,6 +38,7 @@ MAX_TURNS=12
 MODEL="${ANTHROPIC_MODEL:-}"
 SIMULATOR_PROMPT=""
 DECISION_REPLY=""
+DECISION_REPLY_STEPS=()
 CONTINUE_PROMPT="Please continue with the next phase of the comet workflow."
 FRESH_RESUME_MARKER=""
 DECISION_PATTERNS=()
@@ -50,10 +52,22 @@ while [[ $# -gt 0 ]]; do
             shift 2
             ;;
         --simulator-prompt-file)
-            SIMULATOR_PROMPT="$(cat "$2")"
+            SIMULATOR_PROMPT_FILE="$2"
+            if ! SIMULATOR_PROMPT="$(cat -- "$SIMULATOR_PROMPT_FILE")"; then
+                echo "[loop] unable to read private simulator prompt" >&2
+                exit 2
+            fi
+            if ! rm -f -- "$SIMULATOR_PROMPT_FILE"; then
+                echo "[loop] unable to remove private simulator prompt" >&2
+                exit 2
+            fi
             shift 2
             ;;
         --decision-reply) DECISION_REPLY="$2"; shift 2 ;;
+        --decision-reply-step)
+            DECISION_REPLY_STEPS+=("$2")
+            shift 2
+            ;;
         --continue-prompt) CONTINUE_PROMPT="$2"; shift 2 ;;
         --fresh-resume-marker) FRESH_RESUME_MARKER="$2"; shift 2 ;;
         --decision-pattern)
@@ -107,6 +121,7 @@ SESSION_ID=""
 FRESH_PROMPT=""
 COMBINED_OUT=""
 TURN=0
+DECISION_REPLY_STEP_INDEX=0
 
 while [[ $TURN -lt $MAX_TURNS ]]; do
     TURN=$((TURN + 1))
@@ -187,7 +202,15 @@ except: print('')
 
     if bash "$SCRIPT_DIR/decision-point.sh" "$RESULT_TEXT" "${DECISION_PATTERNS[@]}"; then
         echo "[loop] decision point detected; simulating user reply" >&2
-        if [[ -n "$DECISION_REPLY" ]]; then
+        if [[ ${#DECISION_REPLY_STEPS[@]} -gt 0 ]]; then
+            if [[ $DECISION_REPLY_STEP_INDEX -ge ${#DECISION_REPLY_STEPS[@]} ]]; then
+                echo "[loop] deterministic decision reply queue exhausted" >&2
+                exit 3
+            fi
+            USER_REPLY="${DECISION_REPLY_STEPS[$DECISION_REPLY_STEP_INDEX]}"
+            DECISION_REPLY_STEP_INDEX=$((DECISION_REPLY_STEP_INDEX + 1))
+            echo "[loop] deterministic decision reply applied" >&2
+        elif [[ -n "$DECISION_REPLY" ]]; then
             USER_REPLY="$DECISION_REPLY"
             echo "[loop] deterministic decision reply applied" >&2
         else

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@rpamis/comet",
-  "version": "0.4.0-beta.7",
+  "version": "0.4.0-beta.9",
   "description": "Agent Skill Harness For Turning Ideas Into Evaluated Workflows",
   "keywords": [
     "comet",
@@ -91,8 +91,15 @@
   },
   "packageManager": "pnpm@10.18.3",
   "overrides": {
+    "brace-expansion": "5.0.7",
     "esbuild": "0.28.1"
   },
+  "pnpm": {
+    "overrides": {
+      "brace-expansion": "5.0.7",
+      "esbuild": "0.28.1"
+    }
+  },
   "engines": {
     "node": ">=22"
   },