[zzkamzn]

Problem

Two record types could grow past DynamoDB's 400 KB item limit, and in both cases the failure surfaced as something else.

Pipeline runs.

• Every agent() call list_appended an entry, about 300–400 bytes, onto the run item.
• A run can last 45 minutes with a fan-out of 8, so hundreds or thousands of calls are reachable.
• The same item also holds up to 200 log lines of 500 characters and a 24,000-character result. With CJK text those alone come to roughly 370 KB.
• Once the item was full, the append raised and the run failed. A busy workflow was the one that broke.

Service-entry results.

• result was capped at 200,000 characters, which is up to 600 KB of CJK.
• The final update_item then failed and the worker recorded the invocation as failed. The caller polling for the answer of a successful call got a failure.

Change

Pipeline runs

• Each agent call is its own item in the run's partition: PK=PIPELINERUN#<run_id>, SK=AGENT#<ts>#<rand>.
• The run item gets atomic agents_total and cost_usd_total counters, and keeps the first 100 calls as a preview, enforced by a size(agents) < :cap condition. Run lists therefore cost the same as before and still show calls inline.
• GET /api/v1/pipeline-runs/{id} returns every call.
• When a run ends, completed or failed, its per-phase stats are computed over all calls and stored as phases_summary. The Insights view=summary reads that instead of the preview. Running runs and runs recorded before this change fall back to the inline list, exactly as before.
• The result and logs are capped in UTF-8 bytes: 72 KB and 96 KB.
• The Pipeline page loads the full call list when a large run's card is opened, and shows the exact cost total.

Service-entry results

• Capped at 300,000 bytes of UTF-8, with result_truncated returned to the caller. This is documented in docs/user-guide.md.

There are no Terraform or IAM changes; the backend already has PutItem/Query on the table.

Test plan

• New backend/tests/test_item_size_limits.py, 5 tests:
  • 1,500 calls keep the run item under 100 KB, with a 100-call preview, exact totals and an exact cost.
  • The list shows the preview plus totals while the detail shows every call.
  • phases_summary covers every call.
  • The result and logs are capped in bytes.
  • A 900 KB CJK service-entry result is cut to 300 KB, flagged, and still recorded as succeeded.
• Ran the new writes against moto's DynamoDB: the ADD counters with a Decimal cost, the conditional list_append with size(), and the nested phases_summary update. The preview stopped at 100, 130 rows were stored, and the cost totals and phase counts were exact.
• tsc -b and vite build pass.
• ruff reports no new findings on the changed files.

pipeline_service.py is also touched by #52 and #53, in different functions. A trivial rebase may be needed depending on merge order.

This came out of an inventory of DynamoDB access patterns done for the #44 review. It is independent of #44.

🤖 Generated with Claude Code

[odinwang]

Thanks! This change is included in #66 (published together with the related platform fixes). This repository is now updated by snapshot PRs only, so closing this one.