fix: cap search writer context before final answer generation

[stablegenius49]

Summary

• cap the search-result context passed into the final writer call so large result sets do not overflow model context windows
• truncate individual results with a small note while preserving source ordering and citations
• apply the same capped context builder to both the main chat search flow and the API search flow

Testing

• npm install
• npx prettier --write src/lib/utils/splitText.ts src/lib/agents/search/context.ts src/lib/agents/search/index.ts src/lib/agents/search/api.ts
• npx tsc --noEmit
• npx eslint src/lib/utils/splitText.ts src/lib/agents/search/context.ts src/lib/agents/search/index.ts src/lib/agents/search/api.ts

Fixes #1011

---

Summary by cubic

Caps search-result context passed to the final writer to prevent model context window overflow. Truncates long results with a brief note and applies the cap to both chat and API search flows.

• Bug Fixes
  • Add buildSearchResultsContext with a 20k-token total budget and 2.5k per-result cap.
  • Implement token-based truncation via truncateTextByTokens; export getTokenCount.
  • Escape result titles in attributes; preserve source order and citations.
  • Use the shared context builder in both SearchAgent and APISearchAgent.

^{Written for commit 0d09a9a. Summary will update on new commits.}

[cubic-dev-ai]

1 issue found across 4 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/lib/agents/search/context.ts">

<violation number="1" location="src/lib/agents/search/context.ts:48">
P2: Unescaped search result content is inserted into `<result>` blocks, so injected `</result>` or `<result>` text can break prompt boundaries and corrupt citations.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

P2: Unescaped search result content is inserted into <result> blocks, so injected </result> or <result> text can break prompt boundaries and corrupt citations.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At src/lib/agents/search/context.ts, line 48:

<comment>Unescaped search result content is inserted into `<result>` blocks, so injected `</result>` or `<result>` text can break prompt boundaries and corrupt citations.</comment>

<file context>
@@ -0,0 +1,60 @@
+      content = `${truncateTextByTokens(fullContent, noteBudget)}${TRUNCATION_NOTE}`;
+    }
+
+    const entry = `${prefix}${content}${suffix}`;
+    const entryTokens = getTokenCount(entry);
+
</file context>