# Truncated responses are detected and handled gracefully

- **Pattern:** `ab-000201` (`ai-response-quality.response-formatting.truncation-handling`)
- **Severity:** low
- **Lifecycle:** active
- **Last modified:** 2026-04-18
- **Canonical URL:** https://auditbuffet.com/patterns/ab-000201
- **License:** CC-BY-4.0 — attribute to AuditBuffet Pattern Catalog (https://auditbuffet.com/patterns/ab-000201)

## Why it matters

When `finish_reason: "length"` fires and the application ignores it, users receive half-written code, truncated JSON that fails to parse, or advice that ends mid-clause — and they have no signal that anything was cut. They act on incomplete information, file bug reports against the wrong component, or lose trust in the assistant entirely. The error-resilience taxon requires surfacing partial-output states; silent truncation is the worst possible UX for a recoverable failure because it looks like a complete answer.

## Severity rationale

Low because truncation is infrequent in practice and users can retry, but the failure mode is silent and misleading.

## Remediation

After every non-streaming API call, inspect `finish_reason` (OpenAI) or `stop_reason` (Anthropic) and propagate a truncation flag to the UI so the user sees a "Response may be incomplete" banner with a continue button. Wire this through `api/chat/route.ts`.

```ts
const choice = response.choices[0]
return { content: choice.message.content, truncated: choice.finish_reason === 'length' }
```

## Detection

- **ID:** `truncation-handling`
- **Severity:** `low`
- **What to look for:** Enumerate all relevant files and Check API call sites for handling of the `finish_reason` field (OpenAI/Anthropic APIs return `finish_reason: "length"` when a response is cut short by `max_tokens`). Look for code that reads `response.choices[0].finish_reason` or `response.stop_reason` and branches on it. Check whether the UI surfaces any indication to the user when a response was truncated.
- **Pass criteria:** At least 1 implementation must be present. The application checks `finish_reason`/`stop_reason` after receiving a response. When the value is `"length"` (or equivalent), the application either prompts for continuation or displays a notice to the user indicating the response was cut short.
- **Fail criteria:** The application never checks `finish_reason` — truncated responses are silently delivered to users as if complete.
- **Skip (N/A) when:** Application uses streaming with no max_tokens limit, making truncation impossible.
- **Detail on fail:** `"finish_reason not checked after API call in api/chat/route.ts — truncated responses delivered silently"` (max 500 chars)
- **Remediation:** Check the finish reason after every non-streaming API call:

  ```ts
  const response = await openai.chat.completions.create({ ... })
  const choice = response.choices[0]

  if (choice.finish_reason === 'length') {
    return { content: choice.message.content, truncated: true }
  }

  return { content: choice.message.content, truncated: false }
  ```

  Surface `truncated: true` in the UI with a "Response may be incomplete" notice.

---

Taxons: error-resilience, inference-contract

HTML version: https://auditbuffet.com/patterns/ab-000201