## TL;DR

Human-readable logs are not a stable API; their format changes. The agent should get structured data from the run artifacts API (`run_results.json` equivalents) and treat log text as best-effort: tolerant patterns, skip-what-you-cannot-parse, never crash.

## Error

```text
agent's diff parser broke with error on the dbt cloud job log format
```

## Steps

1. Switch the primary data source from raw logs to the structured artifacts endpoint for the run. Expected: stable, versioned fields instead of prose.
2. Where logs are still needed, parse with tolerant patterns (search for known markers, ignore unknown lines). Expected: format drift degrades detail, not the whole parse.
3. Pin to documented fields only; never depend on the exact wording of a log line. Expected: the parser survives wording changes.
4. Wrap parsing in error handling that logs the unparsable section and continues. Expected: one weird line never kills the run.
5. Re-run the agent's analysis against the artifacts. Expected: it completes on structured data.

## When to use

- An agent parser breaks when dbt Cloud changes log formatting.
- Any automation that scrapes human-readable logs.

## When not to use

- The API itself returns malformed JSON (a real API bug; report it).
- The agent needs live streaming output (then tolerate the format explicitly).

## Tool compatibility

- dbt Cloud API artifacts endpoints; any parsing language.

## Variant phrasings

### Log format changed between dbt versions

Version-pin the parser's expectations, or better, stop parsing logs.

### Regex parser fails on a new log line

Tolerant parsing (find markers, skip the rest) is the durable fix.

## Why it happens

Log lines are written for humans and change without notice. Strict parsers encode today's format as law, so every upstream wording tweak becomes an agent crash.

## Edge cases

- Artifacts may lag the live run by a bit; poll for completion before fetching them.
- Some detail only exists in logs (adapter debug lines); extract it with loose searches, not strict grammars.
- Keep a sample of the new format when the parser degrades, so a human can update the markers.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_9F8mZlsaekIGOKtKUjGNMw