> ## Documentation Index
> Fetch the complete documentation index at: https://docs.kaireonai.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Flow Error Inspector

> Failure summary + DLQ row sample + Fix-with-AI loop for the latest failed pipeline run.

The Error Inspector gives you a single place to see why your pipeline
failed and a one-click on-ramp to ask the AI to propose an IR change
that handles the failure pattern.

<Note>
  **Surface availability.** All three surfaces are live: the
  failure-inspection **API** (`GET /api/v1/pipeline-runs/:id/dlq`), the
  **`inspectFlowError`** MCP tool, and the **Errors tab** in the Flow
  Editor. The Errors tab — including the **Fix with AI** button described
  below — is mounted in the editor's tab bar (Visual, JSON IR, SQL Preview,
  Lineage, Schedule, **Runs**, **Errors**), alongside a **Runs** tab that
  lists recent runs with per-node metrics. Open a pipeline at
  `/data/flow-pipelines/[id]/edit` and switch to the **Errors** tab.
</Note>

## What it surfaces

For the latest failed (or `completed_with_errors`) run:

1. **Failure summary** — failed node id + the error message + run id +
   timestamp.
2. **DLQ row sample** — when the failed node is a `validate` node with
   `quarantine.table` configured, the top 10 rows from that quarantine
   table for this run, with the rule id that caught each row.
3. **Fix-with-AI button** — composes a structured prompt:

   ```
   The latest run of pipeline <id> failed.
   Failed node: <nodeId>
   Error: <message>
   Sample failed rows (top 5):
     - {...}  // failed rule: rule-a
   …
   Propose an IR change that handles this error pattern …
   ```

   Then opens the AI Panel with the prompt prefilled in the input.

## API

```http theme={null}
GET /api/v1/pipeline-runs/:id/dlq?node=<validateNodeId>&limit=50
```

**Tenant scoping:** the run must belong to the current tenant.
**IR resolution:** the route looks up the IR version via
`pipeline_runs.metadata.irVersion`, finds the named validate node, and
queries its configured `quarantine.table` filtered by `runId + nodeId`.

**Response shape:**

```json theme={null}
{
  "runId": "...",
  "nodeId": "v",
  "quarantineTable": "dlq.orders",
  "count": 5,
  "rows": [
    { "id": 1, "ruleId": "rule-a", "row": {...}, "createdAt": "..." }
  ]
}
```

**Honest residuals (200 with `residual` message, never silent failure):**

* Validate node has no `quarantine.table` configured → "failed rows are
  not persisted for this node"
* Quarantine table doesn't exist yet → "no rows ever failed for this
  node"

## What's not in 6.3

* **Selective replay-from-DLQ** runtime (re-running only the quarantined
  rows via `POST /pipelines/:id/run`): Phase 6.6 hardening. This flag is
  not implemented yet, and the Errors tab does NOT ship a fake button
  that does nothing.
* **DLQ for non-validate failures** (source format errors land in
  `.failed/` folder, not the DLQ table): Wave 4 polish.