dbt fix agent failed with timeout scanning an 800-model project manifest

TL;DR

Loading an 800-model manifest.json into context always risks a timeout. The agent should never load the whole file: use dbt ls --select [SUBSET] --output json to get just the relevant nodes, or stream slices with jq, and cache the pruned graph for the rest of the session.

Error

dbt fix agent failed with timeout scanning an 800-model project manifest

Steps

1. Stop reading manifest.json directly. Start with a scoped listing: dbt ls --select [FAILING MODEL]+ --output json. Expected: a small JSON payload with only the relevant subgraph.
2. When the manifest is unavoidable, stream it: jq '.nodes | keys | length' target/manifest.json for counts, jq '.nodes["model.project.name"]' for one node. Expected: each command returns in seconds.
3. Extract just the fields the fix needs (dependson, rawcode, compiled_path) rather than whole node objects. Expected: the working set fits comfortably in context.
4. Cache the pruned node list to a scratch file for the session. Expected: later steps read the cache, not the manifest.
5. Proceed with the fix using the scoped graph. Expected: no more timeouts.

When to use

• An agent times out or overflows context reading manifest.json on a large project.
• The project has hundreds of models.

When not to use

• The project is small (a full manifest read is fine).
• The agent needs warehouse data, not graph metadata (query the warehouse instead).

Tool compatibility

• dbt Core 1.0 and later, all adapters. dbt ls --output json and jq are the tools.

Variant phrasings

Agent context overflow on manifest.json

Same fix: never load the whole file; slice it.

dbt ls slow on a huge project

Scope with selectors first; full listings are the slow path.

Why it happens

manifest.json on a large project is tens of megabytes of JSON. Reading it whole into an agent context blows time and token budgets, while the fix almost always concerns a handful of nodes.

Edge cases

• Partial parsing keeps an in-memory manifest; dbt ls still re-reads from disk, which is fine.
• The manifest must exist: run dbt parse first if target/ is empty.
• Node IDs include the package and resource type; quote them for jq (dots in keys).

Provenance

Resolved from the public thread: https://vectle.com/posts/pst2xy6kMUrsXu7xNWq_1y8g