## TL;DR
JSONL merge steps lose findings quietly when a parser expects one JSON document, chokes on a long or malformed line, or overwrites keys that collide across hosts. Parse the nuclei file as a stream of independent lines, count lines in, findings out, and lines rejected with reasons, and reconcile those three numbers before the master report is allowed to build.

```text
agent dropped findings when merging nuclei JSONL into the master report
```

## Use this when
- The master report has fewer nuclei findings than a line count of the JSONL file suggests.
- The merge step logs no error, yet whole templates or hosts are missing from the report.

## Not for this skill when
- Findings were intentionally deduplicated or triaged out with an audit trail. Reconcile against that audit, not the raw count.
- The JSONL file itself is incomplete because the scan was interrupted. Fix the scan output first.

## Steps
1. Count the source: lines in the JSONL, and a quick tally by template id and host using a streaming count, not by loading the file as one JSON document. Verify: Verify: you have per-template source counts to compare against.
2. Run the current merge on its own and capture its output counts by the same template and host tally. Verify: Verify: the specific templates or hosts being dropped are named by the diff.
3. Inspect how the merge reads the file: whole-file JSON parsing, a key map that overwrites same-template results, and silent exception handling are the usual culprits; note which one your pipeline does. Verify: Verify: the loss mechanism is identified in the code, not guessed.
4. Rewrite the merge to stream line by line, append results to a list keyed by a unique result id, and collect unparsable lines into a rejected list with the error, never dropping them silently. Verify: Verify: input lines equal merged findings plus true duplicates plus rejected lines, exactly.
5. Add the reconciliation as a standing gate: the master report cannot build while the counts fail to reconcile, and the rejected list ships with the QA pack. Verify: Verify: truncating a test file makes the gate block with a named count gap.

## Variant phrasings
### nuclei jsonl merge dropped findings line parsing
The parsing phrasing.
### report count reconciliation scanner input output gate
The accounting phrasing.
### silent exception merge pipeline losing results agent
The silent-failure phrasing.

Compatibility: Compatibility: any agent pipeline ingesting nuclei JSONL; streaming parsers exist in every common agent language.

## Why it happens
JSONL is a stream format pretending to be a file format. Tools that load it as a single JSON value fail or, worse, succeed partially. Key collisions do the rest: many results share a template id, so any map keyed by template keeps only the last. Both failures are silent because the merge still produces a plausible-looking report, just a smaller one.

## Edge cases / pitfalls
- A truly corrupt line from an interrupted scan belongs in the rejected list with its line number, so the gap is explained, not hidden.
- Deduplication after the merge must use its own audit; never let dedup run inside the merge step where its removals are uncounted.
- Very large JSONL files are another reason to stream: whole-file loads also die on memory, with the same silent-loss symptom.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_EcuIazT0jprJsfnXCGOkJA