agent context overflow during pull Metabase query results: context limit error failed the run

TL;DR

Shrink what the agent carries. Dashboard builds accumulate schemas, query results, API responses and retry history until the context overflows, so replace raw payloads with compact summaries, page or filter large reads, and persist build state outside the prompt in a run file the agent re-reads in slices. The Metabase build itself is fine, the agent working memory management is what failed.

agent context overflow during pull Metabase query results: context limit error failed the run

Use this when

• The run fails with a context or length error after many successful steps.
• Failure likelihood grows with dashboard size, schema size, or the number of retries in the run.
• Run logs show very large tool results, such as full schema dumps or complete query outputs.

Not for this skill when

• The run fails on its first step with a single huge input file. Split that input directly rather than redesigning run state.
• The model in use has a known small limit and the task is modest. Switching to a larger context model may be the whole fix.

Steps

1. Measure the context contribution of each tool result type in a failed run log, ranking schemas, API payloads and query results by size. Verify: the largest contributors to the overflow are named
2. Replace full payload reads with projections, requesting only the fields the build decisions need from the API and the schema. Verify: the same build decisions are made from the reduced payloads in a test run
3. Summarise completed step results into a short state record and drop the raw responses from the working context where the framework allows it. Verify: a long test run holds its context near a steady size instead of growing per step
4. Persist the build state, object identifiers, decisions and remaining work, to a run file, and have the agent read back only the slice it needs. Verify: a restarted run resumes from the file without the earlier raw history
5. Page or sample any genuinely large read, such as a wide schema or a long result set, rather than loading it whole. Verify: the build handles the largest target dashboard in the test set
6. Rerun the original Metabase build end to end and confirm it completes within the context budget. Verify: the run finishes and the dashboard renders correctly

Why it happens

Context is the agent working memory, and unmanaged tool output is a memory leak. Frameworks default to keeping everything because dropping information feels risky, but the cost is a hard failure late in the longest, most valuable runs. External state, compact summaries and projected reads are the standard cure, and they have the side benefit of making the run resumable and auditable, since the state file shows what the agent believed at each step.

During pull Metabase query results the agent reads the card and dashboard definitions plus the query result payloads, warehouse schemas, prior dashboards and validation output. Each read lands in the context and stays there, including the verbose JSON the Metabase API returns. Retries make it worse by stacking near duplicate responses. Eventually a routine step pushes the total past the model limit and the run dies with a context error that blames the step, not the accumulation. The pattern scales with dashboard size, which is why small test dashboards build cleanly and production sized ones overflow.

Operational notes for context failures in agent pipelines. Budget the context like money, with a rough allocation per phase of the build, discovery, planning, execution, verification, and a check that warns when one phase consumes the share of another. Most overflows come from one phase, usually discovery reading whole schemas, silently eating the budget execution needed later. A run summary that reports peak context by phase makes the imbalance visible before it becomes a failure, and gives the team the evidence to set the projections and summaries where they pay off most.

Variants and edge cases

context limit error agent pull Metabase query results summarise state

The state management phrasing for long builds.

agent context overflow large schema API payload paging

The payload variant naming projected and paged reads.

long agent run context grows retries external state file

The accumulation variant where retries stack duplicate responses.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_snYVGU5df0itWS4vc7ErwQ