sparkecoder 0.1.138 → 0.1.139

package/dist/agent/index.d.ts

@@ -1,6 +1,6 @@
 import 'ai';
 import '../schema-Dz-wABVY.js';
-export { A as Agent, a as AgentOptions, b as AgentRunOptions, c as AgentStreamResult, C as ContextManager, M as MessageAttachment, d as buildSystemPrompt, e as buildTaskPromptAddendum } from '../index-BM99kjgq.js';
+export { A as Agent, a as AgentOptions, b as AgentRunOptions, c as AgentStreamResult, C as ContextManager, M as MessageAttachment, d as buildSystemPrompt, e as buildTaskPromptAddendum } from '../index-BAsQWqZj.js';
 import '../search-CVVfuBPZ.js';
 import 'drizzle-orm/sqlite-core';
 import 'zod';

package/dist/agent/index.js

@@ -2220,7 +2220,29 @@ function markRespondedForThread(slackChannel2, threadTs) {
   }
 }
 function resolveBatchOnTurnEnd(events, ok) {
-  if (!ok) return;
+  if (!ok) {
+    for (const ev of events) {
+      const key2 = eventKey(ev);
+      const entry2 = ledger.get(key2);
+      if (!entry2 || TERMINAL.has(entry2.state)) continue;
+      entry2.state = "failed";
+      entry2.updatedAt = Date.now();
+      if (entry2.channel === "slack" && entry2.slackChannel) {
+        if (entry2.messageTs) fireResultReaction(entry2.slackChannel, entry2.messageTs, "failed");
+        if (entry2.threadTs) maybePostFallback(entry2.slackChannel, entry2.threadTs);
+      }
+      recordEvent({
+        source: "daemon",
+        status: "failed",
+        channel: entry2.channel,
+        sessionId: entry2.sessionId,
+        error: "turn failed; not replayed (avoids context-overflow spiral)",
+        textSnippet: entry2.event.content.slice(0, 200),
+        meta: { ackKey: entry2.key, ackState: "failed" }
+      });
+    }
+    return;
+  }
   for (const ev of events) {
     const key2 = eventKey(ev);
     const entry2 = ledger.get(key2);
@@ -2272,11 +2294,7 @@ function failEntry(entry2) {
   if (entry2.channel === "slack" && entry2.slackChannel && entry2.messageTs) {
     fireResultReaction(entry2.slackChannel, entry2.messageTs, "failed");
     if (entry2.threadTs) {
-      fireFallback(
-        entry2.slackChannel,
-        entry2.threadTs,
-        `:warning: I wasn't able to handle this after ${entry2.attempts} attempt(s). It may need a human \u2014 flagging it here so it isn't lost.`
-      );
+      maybePostFallback(entry2.slackChannel, entry2.threadTs);
     }
   }
   recordEvent({
@@ -2319,6 +2337,17 @@ function fireFallback(channel, threadTs, text) {
   } catch {
   }
 }
+function maybePostFallback(channel, threadTs) {
+  const now = Date.now();
+  const last = lastFallbackAt.get(channel) ?? 0;
+  if (now - last < FALLBACK_COOLDOWN_MS) return;
+  lastFallbackAt.set(channel, now);
+  fireFallback(
+    channel,
+    threadTs,
+    "\u26A0\uFE0F I'm having trouble processing messages right now (likely an overloaded context or a backend error). A human may need to check on me \u2014 I'll pick back up once it's resolved."
+  );
+}
 function startReconciler() {
   if (reconcileTimer) return;
   reconcileTimer = setInterval(() => {
@@ -2340,8 +2369,9 @@ function __listAcks() {
 }
 function __resetAcks() {
   ledger.clear();
+  lastFallbackAt.clear();
 }
-var REPLAY_AFTER_MS, RECONCILE_EVERY_MS, MAX_ATTEMPTS, PRUNE_AFTER_MS, MAX_ENTRIES, TERMINAL, SEP, ledger, reconcileTimer;
+var REPLAY_AFTER_MS, RECONCILE_EVERY_MS, MAX_ATTEMPTS, PRUNE_AFTER_MS, FALLBACK_COOLDOWN_MS, lastFallbackAt, MAX_ENTRIES, TERMINAL, SEP, ledger, reconcileTimer;
 var init_inbox_acks = __esm({
   "src/orchestrator/inbox-acks.ts"() {
     "use strict";
@@ -2352,6 +2382,8 @@ var init_inbox_acks = __esm({
     RECONCILE_EVERY_MS = 6e4;
     MAX_ATTEMPTS = 2;
     PRUNE_AFTER_MS = 60 * 6e4;
+    FALLBACK_COOLDOWN_MS = 15 * 6e4;
+    lastFallbackAt = /* @__PURE__ */ new Map();
     MAX_ENTRIES = 5e3;
     TERMINAL = /* @__PURE__ */ new Set(["responded", "skipped", "handed_off", "failed"]);
     SEP = "\u241F";
@@ -7706,7 +7738,11 @@ function capImageCount(messages, max = MAX_IMAGES_IN_CONTEXT) {
 var MODEL_LIMITS = {
   "claude-opus-4-8": { contextWindow: 2e5, rollingTarget: 15e4 },
   "gpt-5.5": { contextWindow: 35e4, rollingTarget: 15e4 },
-  "claude-fable-5": { contextWindow: 2e5, rollingTarget: 15e4 }
+  "claude-fable-5": { contextWindow: 2e5, rollingTarget: 15e4 },
+  // Claude Opus 4.7 advertises a 1M-token context. Keyed WITH the provider
+  // prefix because getModelLimits() matches the full id (e.g.
+  // "anthropic/claude-opus-4.7") before falling back to the prefix table.
+  "anthropic/claude-opus-4.7": { contextWindow: 1e6, rollingTarget: 15e4 }
 };
 var DEFAULT_LIMITS = { contextWindow: 2e5, rollingTarget: 15e4 };
 var PREFIX_DEFAULTS = {
@@ -7778,8 +7814,41 @@ ${summaryContent}`
     messages = ensureToolResultsFollowCalls(messages);
     messages = ensureEndsWithUserOrTool(messages);
     messages = capImageCount(messages);
+    messages = this.enforceHardCap(messages);
     return messages;
   }
+  /**
+   * Drop oldest messages (and, as a last resort, hard-truncate text) until
+   * the prompt fits the model's context window. Preserves any leading
+   * summary system message and always keeps the most recent message, then
+   * repairs tool pairing so dropping can't orphan a tool result.
+   */
+  enforceHardCap(messages) {
+    const { contextWindow } = getModelLimits(this.modelId);
+    const ceiling = Math.max(2e4, Math.floor(contextWindow * 0.9));
+    const tokens = messages.map((m) => this.messageTokens(m));
+    let total = tokens.reduce((a, b) => a + b, 0);
+    if (total <= ceiling) return messages;
+    const hasLeadingSummary = messages.length > 0 && messages[0].role === "system";
+    const firstBody = hasLeadingSummary ? 1 : 0;
+    const lastIndex = messages.length - 1;
+    let start = firstBody;
+    while (start < lastIndex && total > ceiling) {
+      total -= tokens[start];
+      start += 1;
+    }
+    let out = hasLeadingSummary ? [messages[0], ...messages.slice(start)] : messages.slice(start);
+    if (total > ceiling) {
+      out = out.map((m) => hardTruncateMessageText(m));
+    }
+    out = repairToolPairing(out);
+    out = ensureToolResultsFollowCalls(out);
+    out = ensureEndsWithUserOrTool(out);
+    console.warn(
+      `[Context] hard cap engaged for ${this.modelId}: trimmed ${messages.length}\u2192${out.length} msgs (ceiling ${ceiling} tokens).`
+    );
+    return out;
+  }
   // ---------------------------------------------------------------------------
   // Phase 1 – Compact
   // ---------------------------------------------------------------------------
@@ -8006,6 +8075,40 @@ ${summaryContent}`
     this.summaries = [];
   }
 };
+var HARD_TRUNCATE_CHARS = 6e3;
+function truncateMiddle(s, cap) {
+  if (s.length <= cap) return s;
+  const half = Math.floor(cap / 2);
+  return s.slice(0, half) + `
+...[truncated ${s.length - cap} chars to fit context]...
+` + s.slice(-half);
+}
+function truncateToolResultText(part) {
+  const trunc = (r) => {
+    if (typeof r === "string") return truncateMiddle(r, HARD_TRUNCATE_CHARS);
+    if (r && typeof r === "object" && typeof r.text === "string") {
+      return { ...r, text: truncateMiddle(r.text, HARD_TRUNCATE_CHARS) };
+    }
+    return r;
+  };
+  if (Array.isArray(part.result)) return { ...part, result: part.result.map(trunc) };
+  if ("result" in part) return { ...part, result: trunc(part.result) };
+  return part;
+}
+function hardTruncateMessageText(msg) {
+  if (typeof msg.content === "string") {
+    return { ...msg, content: truncateMiddle(msg.content, HARD_TRUNCATE_CHARS) };
+  }
+  if (!Array.isArray(msg.content)) return msg;
+  const parts = msg.content.map((part) => {
+    if (part?.type === "text" && typeof part.text === "string") {
+      return { ...part, text: truncateMiddle(part.text, HARD_TRUNCATE_CHARS) };
+    }
+    if (part?.type === "tool-result") return truncateToolResultText(part);
+    return part;
+  });
+  return { ...msg, content: parts };
+}
 function stripBinaryContentForSummary(value) {
   if (Array.isArray(value)) return value.map(stripBinaryContentForSummary);
   if (!value || typeof value !== "object") return value;