Updated Guide and Build Fix

Author: Jitraj277

Asset/SCR-20260423-oqsn.png

@@ -44,12 +44,41 @@ A real-time visualization dashboard at `/dashboard` shows:
 - Action timeline with reward indicators
 - **Accumulated run results** that persist across tiers with expandable per-tier reports
 
+![MissionCtrl Dashboard](Asset/l%20Missionctrl%20TASK.png)
+
+#### Run Results Breakdown
+The Run Results panel now supports expandable per-tier drilldowns for:
+- Score breakdown contributions by signal
+- Hallucination stats (injected/caught/TP/FP)
+- Action-by-action reward history
+
+![Run Results Panel](Asset/SCR-20260423-ordo.png)
+
 ### 🔄 Deterministic Replay
 Every episode can be deterministically replayed via seeded randomness, enabling reproducible debugging and benchmarking.
 
 ### 🐳 Single-Container Deployment
 Server + inference in one Docker image. No orchestration, no external databases — just `docker run` and go.
 
+### 🧾 Verbose LLM Trace View
+When `VERBOSE_TRACE=1`, inference prints compact boxed traces for each step:
+- Prompt metadata (including char count)
+- Prompt preview for fast debugging
+- Action normalization and guardrail rewrites
+- Step transition outcomes and rewards
+
+![LLM Prompt Trace](Asset/SCR-20260423-oqsn.png)
+
+![LLM Response Trace](Asset/SCR-20260423-orzh.png)
+
+### 🚦 Token-Budget Guardrails
+Inference now includes hardening for provider token limits:
+- **Stateless per-step LLM requests** (fresh system + current observation only)
+- **No retry loop for permanent oversized-request errors**
+- Retry/backoff remains enabled for transient provider throttling
+
+This prevents late-step context blowups (for example, step 5 payload growth) from repeatedly failing with the same "request too large" response.
+
 ---
 
 ## 📈 Baseline Results
@@ -204,13 +233,29 @@ pytest tests/ -v
 
 | Variable | Default | Description |
 |---|---|---|
-| `API_BASE_URL` | `https://api.groq.com/openai/v1` | LLM API endpoint |
-| `MODEL_NAME` | `llama-3.3-70b-versatile` | Model to use |
+| `API_BASE_URL` | `https://router.huggingface.co/v1` | OpenAI-compatible LLM API endpoint |
+| `MODEL_NAME` | `openai/gpt-oss-120b` | Model to use |
 | `HF_TOKEN` | — | API key |
-| `STEP_DELAY_S` | `0.5` | Delay between steps (reduce for speed) |
+| `ENV_BASE_URL` | `http://localhost:8000` | MissionCtrl server base URL |
+| `STEP_DELAY_S` | `4.0` | Delay between steps (reduce for speed) |
 | `VERBOSE_TRACE` | `1` | Show detailed step traces |
+| `PROMPT_PREVIEW_CHARS` | `200` | Prompt preview truncation length in trace logs |
+| `TRACE_WRAP_WIDTH` | `76` | Text wrap width for trace block content |
+| `TRACE_BOX_WIDTH` | `76` | Width of the boxed trace output |
+| `SPINNER_ENABLED` | `0` | Enable CLI spinner while waiting for LLM response |
 | `MAX_STEPS` | `5` | Steps per episode |
 
+### Troubleshooting: Request Too Large / TPM Errors
+
+If your provider returns errors like:
+`Request too large ... tokens per minute ... Requested > Limit`
+
+Use this checklist:
+1. Ensure you are running the latest image/code with stateless per-step requests.
+2. Reduce verbosity/observation size if needed (fewer long output snippets).
+3. Switch to a model/tier with higher TPM limits.
+4. Keep retries for transient rate limits; oversized requests are now treated as non-retryable.
+
 ---
 
 ## 📋 Task Tiers

Asset/SCR-20260423-ordo.png

@@ -31,7 +31,7 @@
 import httpx
 from dotenv import load_dotenv
 from openai import OpenAI
-from tenacity import retry, stop_after_attempt, wait_exponential, before_sleep_log
+from tenacity import retry, stop_after_attempt, wait_exponential, before_sleep_log, retry_if_not_exception_type
 import logging as _logging
 
 # ---------------------------------------------------------------------------
@@ -96,6 +96,11 @@ def _validate_env() -> None:
 _retry_logger = _logging.getLogger("missionctrl.retry")
 
 
+class PromptTooLargeError(RuntimeError):
+    """Raised when provider rejects a request as permanently oversized."""
+
+
+
 def _append_bounded_unique(bucket: List[str], value: str, limit: int) -> None:
     value = value.strip()
     if not value:
@@ -685,6 +690,7 @@ def _spin():
     stop=stop_after_attempt(LLM_MAX_RETRIES),
     wait=wait_exponential(multiplier=2, min=2, max=30),
     before_sleep=before_sleep_log(_retry_logger, _logging.WARNING),
+    retry=retry_if_not_exception_type(PromptTooLargeError),
     reraise=True,
 )
 def _call_llm(messages: List[Dict[str, str]]) -> str:
@@ -698,6 +704,9 @@ def _call_llm(messages: List[Dict[str, str]]) -> str:
         )
     except Exception as exc:
         msg = str(exc)
+        lower_msg = msg.lower()
+        if "request too large" in lower_msg or ("tokens per minute" in lower_msg and "requested" in lower_msg):
+            raise PromptTooLargeError(f"Prompt too large: {msg.splitlines()[0]}") from exc
         if "429" in msg or "rate_limit" in msg.lower():
             raise RuntimeError(f"Rate-limited: {msg.splitlines()[0]}") from exc
         raise
@@ -834,7 +843,7 @@ def run_task(task_id: str, policy_memory: PolicyMemory) -> float:
     done = False
 
     try:
-        messages: List[Dict[str, str]] = [{"role": "system", "content": SYSTEM_PROMPT}]
+        system_message = {"role": "system", "content": SYSTEM_PROMPT}
         action_history: List[str] = []
         episode_memory = EpisodeMemory()
 
@@ -850,7 +859,10 @@ def run_task(task_id: str, policy_memory: PolicyMemory) -> float:
                 episode_memory,
                 policy_memory,
             )
-            messages.append({"role": "user", "content": user_msg})
+            messages: List[Dict[str, str]] = [
+                system_message,
+                {"role": "user", "content": user_msg},
+            ]
             before_obs = obs
 
             if VERBOSE_TRACE:
@@ -872,7 +884,6 @@ def run_task(task_id: str, policy_memory: PolicyMemory) -> float:
                 raw_action = "NOOP"
 
             safe_action = _normalize_action(raw_action, obs, episode_memory)
-            messages.append({"role": "assistant", "content": safe_action})
 
             extracted = _extract_action_from_response(raw_action)
             was_cleaned = raw_action.strip() != extracted

Asset/SCR-20260423-orzh.png

@@ -276,6 +276,40 @@ <h1>🛡️ MissionCtrl</h1>
   document.getElementById('kpi-fp').className = 'kpi-value';
 }
 
+function pickMetric(...values) {
+  for (const v of values) {
+    const n = Number(v);
+    if (Number.isFinite(n)) return n;
+  }
+  return 0;
+}
+
+function updateHallucinationKpis(stats, hasTasks) {
+  const drEl = document.getElementById('kpi-detection');
+  const fpEl = document.getElementById('kpi-fp');
+
+  if (!hasTasks) {
+    drEl.textContent = '—';
+    drEl.className = 'kpi-value';
+    fpEl.textContent = '—';
+    fpEl.className = 'kpi-value';
+    return;
+  }
+
+  const totalInjected = pickMetric(stats.total_injected, stats.injected, stats.num_injected);
+  const totalCaught = pickMetric(stats.total_caught, stats.true_positives, stats.caught);
+  const totalFlags = pickMetric(stats.total_flags, stats.flags);
+  const falsePositives = pickMetric(stats.false_positives, stats.false_positive_count, stats.fp);
+
+  const dr = totalInjected > 0 ? (totalCaught / totalInjected) * 100 : 100;
+  drEl.textContent = dr.toFixed(0) + '%';
+  drEl.className = 'kpi-value ' + (dr >= 85 ? 'good' : dr >= 50 ? 'warn' : 'bad');
+
+  const fpRate = totalFlags > 0 ? (falsePositives / totalFlags) * 100 : 0;
+  fpEl.textContent = fpRate.toFixed(0) + '%';
+  fpEl.className = 'kpi-value ' + (fpRate <= 10 ? 'good' : fpRate <= 25 ? 'warn' : 'bad');
+}
+
 // ── Dashboard Update ─────────────────────────────────────────────────────
 
 function updateDashboard(state, history) {
@@ -309,6 +343,10 @@ <h1>🛡️ MissionCtrl</h1>
   const doneCount = tasks.filter(t => t.status === 'DONE').length;
   document.getElementById('kpi-tasks').textContent = `${doneCount} / ${tasks.length}`;
 
+  // Keep detection/FP KPIs live during the episode.
+  const liveStats = state.hallucination_stats || {};
+  updateHallucinationKpis(liveStats, tasks.length > 0);
+
   // Task table
   const tbody = document.getElementById('task-table');
   if (tasks.length > 0) {
@@ -368,17 +406,9 @@ <h1>🛡️ MissionCtrl</h1>
     }
     document.getElementById('score-bars').innerHTML = barsHtml;
 
-    // Detection KPIs
+    // Prefer final stats after episode completion.
     const stats = sb.hallucination_stats || {};
-    const dr = stats.total_injected > 0 ? ((stats.total_caught / stats.total_injected) * 100) : 100;
-    const drEl = document.getElementById('kpi-detection');
-    drEl.textContent = dr.toFixed(0) + '%';
-    drEl.className = 'kpi-value ' + (dr >= 85 ? 'good' : dr >= 50 ? 'warn' : 'bad');
-
-    const fpRate = stats.total_flags > 0 ? ((stats.false_positives / stats.total_flags) * 100) : 0;
-    const fpEl = document.getElementById('kpi-fp');
-    fpEl.textContent = fpRate.toFixed(0) + '%';
-    fpEl.className = 'kpi-value ' + (fpRate <= 10 ? 'good' : fpRate <= 25 ? 'warn' : 'bad');
+    updateHallucinationKpis(stats, tasks.length > 0);
 
     // Incidents
     if (stats.total_injected > stats.total_caught || stats.false_positives > 0) {