fix(llm): preflight compact oversized requests

src/main/java/me/golemcore/bot/adapter/outbound/llm/Langchain4jAdapter.java

@@ -31,6 +31,7 @@
 import me.golemcore.bot.domain.model.ToolDefinition;
 import me.golemcore.bot.domain.service.ToolArtifactService;
 import me.golemcore.bot.domain.service.RuntimeConfigService;
+import me.golemcore.bot.domain.system.LlmErrorPatterns;
 import me.golemcore.bot.port.outbound.ModelConfigPort;
 import me.golemcore.bot.port.outbound.LlmPort;
 import com.fasterxml.jackson.core.type.TypeReference;
@@ -71,13 +72,14 @@
 import org.springframework.stereotype.Component;
 import reactor.core.publisher.Flux;
 
+import java.time.Duration;
 import java.util.ArrayDeque;
 import java.util.ArrayList;
 import java.util.Base64;
-import java.time.Duration;
 import java.util.Collections;
 import java.util.Deque;
 import java.util.HashSet;
+import java.util.IdentityHashMap;
 import java.util.LinkedHashMap;
 import java.util.List;
 import java.util.Locale;
@@ -138,6 +140,13 @@
     private static final java.util.regex.Pattern RESET_SECONDS_PATTERN = java.util.regex.Pattern
             .compile("\"reset_seconds\"\\s*:\\s*(\\d+)");
     private static final String TOOL_ATTACHMENT_REOPEN_HINT = "Re-open the file with a tool if deeper inspection is needed.";
+    private static final Set<String> RATE_LIMIT_MARKERS = Set.of(
+            "rate_limit",
+            "rate limit",
+            "token_quota_exceeded",
+            "too many requests",
+            "model_cooldown",
+            "cooling down");
 
     private record MessageConversionResult(List<ChatMessage> messages, boolean hydratedToolImages) {
     }
@@ -500,21 +509,49 @@
     }
 
     private boolean isRateLimitError(Throwable e) {
-        // Walk the cause chain looking for rate limit indicators
-        Throwable current = e;
-        while (current != null) {
-            // langchain4j maps HTTP 429 to RateLimitException regardless of body content
+        // Use identity semantics for the visited set: the intent is "have I seen this
+        // exact cause object before" (cycle guard), not value equality. A future
+        // provider exception that overrides equals/hashCode - or a wrapped cause that
+        // happens to compare equal to an earlier frame - must not collapse into a
+        // single visited entry and prematurely terminate the cause walk.
+        Set<Throwable> visited = Collections.newSetFromMap(new IdentityHashMap<>());
+        for (Throwable current = e; current != null && visited.add(current); current = current.getCause()) {
+            // langchain4j maps HTTP 429 to RateLimitException regardless of body content.
             if (current instanceof dev.langchain4j.exception.RateLimitException) {
                 return true;
             }
-            String msg = current.getMessage();
-            if (msg != null && (msg.contains("rate_limit") || msg.contains("token_quota_exceeded")
-                    || msg.contains("too_many_tokens") || msg.contains("Too Many Requests")
-                    || msg.contains("429") || msg.contains("model_cooldown")
-                    || msg.contains("cooling down"))) {
+            String raw = current.getMessage();
+            if (raw == null) {
+                continue;
+            }
+            String normalized = raw.toLowerCase(Locale.ROOT);
+            // Context-overflow errors sometimes carry rate-limit-adjacent wording
+            // ("too many tokens") - they must NOT be retried as rate limits.
+            if (LlmErrorPatterns.matchesContextOverflow(normalized)) {
+                continue;
+            }
+            if (containsAny(normalized, RATE_LIMIT_MARKERS) || looksLikeHttp429(normalized)) {
+                return true;
+            }
+        }
+        return false;
+    }
+
+    private static boolean looksLikeHttp429(String normalized) {
+        // Avoid matching an arbitrary "429" in stack traces, byte counts, or IDs.
+        return normalized.contains("http 429")
+                || normalized.contains("status 429")
+                || normalized.contains("status: 429")
+                || normalized.contains("status_code=429")
+                || normalized.contains("code 429")
+                || normalized.contains("429 too many");
+    }
+
+    private static boolean containsAny(String haystack, Set<String> needles) {
+        for (String needle : needles) {
+            if (haystack.contains(needle)) {
                 return true;
             }
-            current = current.getCause();
         }
         return false;
     }
@@ -705,7 +742,7 @@
 
     @Override
     public List<String> getSupportedModels() {
-        // Build from models.json — provider/modelName for each configured provider
+        // Build from models.json - provider/modelName for each configured provider.
         List<String> models = new ArrayList<>();
         Map<String, ModelCatalogEntry> modelsConfig = modelConfig
                 .getAllModels();

src/main/java/me/golemcore/bot/domain/model/CompactionReason.java

@@ -4,5 +4,5 @@
  * Trigger source for conversation compaction.
  */
 public enum CompactionReason {
-    AUTO_THRESHOLD, MANUAL_COMMAND, CONTEXT_OVERFLOW_RECOVERY
+    AUTO_THRESHOLD, MANUAL_COMMAND, CONTEXT_OVERFLOW_RECOVERY, REQUEST_PREFLIGHT
 }

src/main/java/me/golemcore/bot/domain/model/ContextAttributes.java

@@ -27,14 +27,14 @@ public final class ContextAttributes {
     private ContextAttributes() {
     }
 
-    /** Boolean — signal to stop the agent loop after current iteration. */
-    /** LlmResponse — response from LLM execution. */
+    /** Boolean - signal to stop the agent loop after current iteration. */
+    /** LlmResponse - response from LLM execution. */
     public static final String LLM_RESPONSE = "llm.response";
 
-    /** String — LLM error message. */
+    /** String - LLM error message. */
     public static final String LLM_ERROR = "llm.error";
 
-    /** String — machine-readable LLM error classification code. */
+    /** String - machine-readable LLM error classification code. */
     public static final String LLM_ERROR_CODE = "llm.error.code";
 
     /** String ? active request trace identifier. */
@@ -78,16 +78,16 @@ private ContextAttributes() {
      */
     public static final String CONTEXT_SCOPED_TOOLS = "context.scoped.tools";
 
-    /** Boolean — tools were executed in this iteration. */
+    /** Boolean - tools were executed in this iteration. */
 
     /**
-     * String — LLM model name that last generated tool calls in the session
+     * String - LLM model name that last generated tool calls in the session
      * (persisted in session metadata).
      */
     public static final String LLM_MODEL = "llm.model";
 
     /**
-     * String — reasoning effort used for the current LLM call (e.g. "low",
+     * String - reasoning effort used for the current LLM call (e.g. "low",
      * "medium", "high").
      */
     public static final String LLM_REASONING = "llm.reasoning";
@@ -109,10 +109,10 @@ private ContextAttributes() {
     /** Boolean - session-scoped lock for the selected model tier. */
     public static final String SESSION_MODEL_TIER_FORCE = "session.modelTier.force";
 
-    /** Boolean — plan mode is active for the current session. */
+    /** Boolean - plan mode is active for the current session. */
     public static final String PLAN_MODE_ACTIVE = "plan.mode.active";
 
-    /** String — plan ID that needs user approval before execution. */
+    /** String - plan ID that needs user approval before execution. */
     public static final String PLAN_APPROVAL_NEEDED = "plan.approval.needed";
 
     /**
@@ -141,7 +141,7 @@ private ContextAttributes() {
     public static final String ITERATION_LIMIT_REACHED = "iteration.limit.reached";
 
     /**
-     * Boolean — tool loop stopped due to internal limit (LLM calls / tool
+     * Boolean - tool loop stopped due to internal limit (LLM calls / tool
      * executions / deadline).
      */
     public static final String TOOL_LOOP_LIMIT_REACHED = "toolloop.limit.reached";
@@ -152,16 +152,16 @@ private ContextAttributes() {
     public static final String TOOL_LOOP_LIMIT_REASON = "toolloop.limit.reason";
 
     /**
-     * List<RuntimeEvent> — runtime execution events for the current turn.
+     * List<RuntimeEvent> - runtime execution events for the current turn.
      */
     public static final String RUNTIME_EVENTS = "runtime.events";
 
     /**
-     * Boolean — stop tool execution between tool calls for current turn.
+     * Boolean - stop tool execution between tool calls for current turn.
      */
     public static final String TURN_INTERRUPT_REQUESTED = "turn.interrupt.requested";
 
-    /** String — queue kind for inbound message while a turn is running. */
+    /** String - queue kind for inbound message while a turn is running. */
     public static final String TURN_QUEUE_KIND = "turn.queue.kind";
 
     /**
@@ -180,7 +180,7 @@ private ContextAttributes() {
     public static final String TURN_QUEUE_KIND_DELAYED_ACTION = "delayed_action";
 
     /**
-     * Boolean — current turn scheduled an internal retry instead of user feedback.
+     * Boolean - current turn scheduled an internal retry instead of user feedback.
      */
     public static final String TURN_INTERNAL_RETRY_SCHEDULED = "turn.internal.retry.scheduled";
 
@@ -217,7 +217,7 @@ private ContextAttributes() {
     /** Boolean ? current inbound message is an internal runtime-only message. */
     public static final String TURN_INPUT_INTERNAL = "turn.input.internal";
 
-    /** WebSocketSession — reference to WebSocket session for streaming. */
+    /** WebSocketSession - reference to WebSocket session for streaming. */
     public static final String WEB_STREAM_SINK = "web.stream.sink";
 
     /** Boolean ? webhook delivery should be routed to an external channel. */
@@ -373,6 +373,12 @@ private ContextAttributes() {
     /** Map<String,Object> ? latest structured compaction details. */
     public static final String COMPACTION_LAST_DETAILS = "compaction.last.details";
 
+    /** Map<String,Object> - latest LLM request token preflight diagnostics. */
+    public static final String LLM_REQUEST_PREFLIGHT = "llm.request.preflight";
+
+    /** Map<String,Object> - latest LLM context-overflow recovery diagnostics. */
+    public static final String LLM_CONTEXT_OVERFLOW_RECOVERY = "llm.context.overflow.recovery";
+
     /** List<Map<String,Object>> ? per-turn edited file stats for UI hints. */
     public static final String TURN_FILE_CHANGES = "turn.file.changes";
 

src/main/java/me/golemcore/bot/domain/service/CompactionOrchestrationService.java

@@ -18,7 +18,19 @@
 import java.util.Optional;
 
 /**
- * Central compaction pipeline: prepare boundaries, summarize, persist details.
+ * Central compaction pipeline shared by every trigger source (auto-compaction,
+ * manual command, preflight, context-overflow recovery).
+ *
+ * <p>
+ * The service coordinates four collaborators: a
+ * {@link CompactionPreparationService} that decides which messages to drop vs.
+ * keep, a {@link CompactionService} that produces the summary text, a
+ * {@link CompactionDetailsExtractor} that builds the structured audit trail,
+ * and the {@link SessionPort} that persists the compacted session. The flow
+ * produces a single {@link CompactionResult} whether the compaction removed
+ * messages, became a no-op, or failed to locate the session, so callers can
+ * record diagnostics from one result shape.
+ * </p>
  */
 @Service
 public class CompactionOrchestrationService {
@@ -46,6 +58,35 @@ public CompactionOrchestrationService(SessionPort sessionPort,
         this.clock = clock;
     }
 
+    /**
+     * Compact the named session, keeping at most {@code keepLast} trailing messages
+     * and summarizing the rest if the preparation service can identify messages
+     * that may be removed.
+     *
+     * <p>
+     * Returns a {@link CompactionResult} with:
+     * </p>
+     * <ul>
+     * <li>{@code removed = -1} when the session cannot be located - the caller
+     * still receives a well-formed result and records diagnostics rather than
+     * throwing.</li>
+     * <li>{@code removed = 0} with no-op details when there was nothing to remove
+     * (keepLast larger than the session, or the preparation service refused to
+     * split a turn).</li>
+     * <li>{@code removed > 0} when messages were dropped. A summary message is
+     * present only when the summarizer returns non-blank text; otherwise
+     * {@code usedSummary=false} and the kept messages are persisted without a
+     * synthetic summary.</li>
+     * </ul>
+     *
+     * @param sessionId
+     *            the session to compact
+     * @param reason
+     *            classification used downstream for telemetry and summary prompt
+     *            selection
+     * @param keepLast
+     *            trailing messages to retain intact
+     */
     public CompactionResult compact(String sessionId, CompactionReason reason, int keepLast) {
         Optional<AgentSession> sessionOptional = sessionPort.get(sessionId);
         if (sessionOptional.isEmpty()) {
@@ -79,13 +120,14 @@ public CompactionResult compact(String sessionId, CompactionReason reason, int k
                     false,
                     0,
                     runtimeConfigService.getCompactionDetailsMaxItemsPerCategory());
-            persistDetails(session, emptyDetails);
-            return CompactionResult.builder()
+            CompactionResult emptyResult = CompactionResult.builder()
                     .removed(0)
                     .usedSummary(false)
                     .summaryMessage(null)
                     .details(emptyDetails)
                     .build();
+            persistDetails(session, emptyResult);
+            return emptyResult;
         }
 
         long startedAt = clock.millis();
@@ -123,62 +165,29 @@ public CompactionResult compact(String sessionId, CompactionReason reason, int k
             Map<String, Object> metadata = summaryMessage.getMetadata() != null
                     ? new LinkedHashMap<>(summaryMessage.getMetadata())
                     : new LinkedHashMap<>();
-            metadata.put(METADATA_KEY_COMPACTION_DETAILS, toDetailsMap(details));
+            metadata.put(METADATA_KEY_COMPACTION_DETAILS, CompactionPayloadMapper.toDetailsMap(details));
             summaryMessage.setMetadata(metadata);
         }
 
-        persistDetails(session, details);
-        sessionPort.save(session);
-
-        return CompactionResult.builder()
+        CompactionResult result = CompactionResult.builder()
                 .removed(preparation.messagesToCompact().size())
                 .usedSummary(usedSummary)
                 .summaryMessage(summaryMessage)
                 .details(details)
                 .build();
+        persistDetails(session, result);
+        sessionPort.save(session);
+        return result;
     }
 
-    private void persistDetails(AgentSession session, CompactionDetails details) {
-        if (session == null || details == null) {
+    private void persistDetails(AgentSession session, CompactionResult result) {
+        if (session == null || result == null || result.details() == null) {
             return;
         }
         if (session.getMetadata() == null) {
             session.setMetadata(new LinkedHashMap<>());
         }
-        session.getMetadata().put(ContextAttributes.COMPACTION_LAST_DETAILS, toDetailsMap(details));
-    }
-
-    private Map<String, Object> toDetailsMap(CompactionDetails details) {
-        Map<String, Object> result = new LinkedHashMap<>();
-        result.put("schemaVersion", details.schemaVersion());
-        result.put("reason", details.reason() != null ? details.reason().name() : null);
-        result.put("summarizedCount", details.summarizedCount());
-        result.put("keptCount", details.keptCount());
-        result.put("usedLlmSummary", details.usedLlmSummary());
-        result.put("summaryLength", details.summaryLength());
-        result.put("toolCount", details.toolCount());
-        result.put("readFilesCount", details.readFilesCount());
-        result.put("modifiedFilesCount", details.modifiedFilesCount());
-        result.put("durationMs", details.durationMs());
-        result.put("toolNames", details.toolNames());
-        result.put("readFiles", details.readFiles());
-        result.put("modifiedFiles", details.modifiedFiles());
-        result.put("splitTurnDetected", details.splitTurnDetected());
-        result.put("fallbackUsed", details.fallbackUsed());
-
-        List<Map<String, Object>> fileChanges = new ArrayList<>();
-        if (details.fileChanges() != null) {
-            for (CompactionDetails.FileChangeStat fileChange : details.fileChanges()) {
-                Map<String, Object> fileChangeMap = new LinkedHashMap<>();
-                fileChangeMap.put("path", fileChange.path());
-                fileChangeMap.put("addedLines", fileChange.addedLines());
-                fileChangeMap.put("removedLines", fileChange.removedLines());
-                fileChangeMap.put("deleted", fileChange.deleted());
-                fileChanges.add(fileChangeMap);
-            }
-        }
-        result.put("fileChanges", fileChanges);
-
-        return result;
+        session.getMetadata().put(ContextAttributes.COMPACTION_LAST_DETAILS,
+                CompactionPayloadMapper.toPayload(result));
     }
 }