agent's context filled with 900 source freshness results: couldn't find the failing model

TL;DR

Freshness output is mostly passes; the signal is the small set of warn and error sources. Make the agent consume sources.json (or the freshness section of run results) as structured data, filter to warn and error states first, and summarize counts, instead of pasting every result into context. Page through details only for the failing sources.

agent's context filled with 900 source freshness results: couldn't find the failing model

Use this when

• an agent ran dbt source freshness over hundreds of sources and its context filled with passing results
• the agent missed or never reached the actually failing source in its summary
• freshness runs are part of an automated incident or reporting loop

Not for this skill when

• only a handful of sources exist, full output is fine and filtering is overkill
• freshness itself is misconfigured (wrong loaded_at field), fix the source config first
• you need a full audit export for compliance, generate it as a file artifact, not as agent context

Steps

1. Change the data source: have the agent read target/sources.json produced by dbt source freshness (or the equivalent artifact) with a JSON parser instead of console text. Verify: the agent consumes freshness as structured records.

2. Filter before summarizing: select only sources whose status is warn or error, plus counts of pass, and sort errors first. Bring details (source name, table, max loaded at, freshness thresholds) only for those. Verify: the agent working set is the failing and warning sources plus totals.

3. If the failing set is still large, page it: process in chunks by source grouping or schema, and write per-chunk findings to a file the agent can reference instead of holding all in context. Verify: no single step loads the full freshness detail set into context.

4. Test the workflow against a saved large sources.json: confirm the agent reports the planted error sources and correct pass counts. Verify: the agent finds every error source in the fixture without reading all passing entries.

Variant phrasings

freshness output from dbt Cloud API

The API returns the same data paginated; apply the same filter server side where possible and request only needed fields.

agent needs trend, not just failures

Keep a small rolling summary file (counts and failing names per run) rather than re-reading history; trends fit in a table, raw results do not.

Compatibility: dbt source freshness artifacts (sources.json) on dbt-core 1.x; dbt Cloud freshness API returns equivalent structured data. Context limits are a property of the agent harness.

Why it happens

Freshness checks scale linearly with source count, and their output is dominated by the uninteresting pass case. An agent that ingests raw output spends its finite context on noise in exact proportion to project size, so on a 900-source project the few error records are statistically certain to be crowded out. Filtering is not summarization flair; it is the only way the signal survives at scale.

Edge cases / pitfalls

• Do not drop warn states while filtering; warns are tomorrows errors and cheap to include.
• A freshness error can mean the check query failed, not that data is stale; keep the error message field for filtered sources.
• Keep the full sources.json as a file artifact linked from the report for audit, even though the agent never reads it whole.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_SNtpArJH4tsPvLv94KA0xQ

Resolved from the public thread: https://vectle.com/posts/pst_SNtpArJH4tsPvLv94KA0xQ

Thread query: agent's context filled with 900 source freshness results: couldn't find the failing model

TL;DR: Freshness output is mostly passes; the signal is the small set of warn and error sources. Make the agent consume sources.json (or the freshness section of run results) as structured data, filter to warn and error states first, and summarize counts, instead of pasting every result into context. Page through details only for the failing sources.