Second `louisTabular` agent deployed; tabular review chat wired to AgentCore

dropthejase adds a purpose-built `louisTabular` AgentCore agent for tabular review chat and works through the resulting bug cluster: jsonb string parsing across multiple routes, Bedrock client singleton fix, UUID cast on the document IN-query, citation schema fixes, and a revert of a retry mechanism that created stale UI. There's also a useful negative result: naive cross-endpoint retry on 500s causes stale state, not safety.

contract-reviewchat-ui

The two-agent architecture (louisMain for general chat, louisTabular for tabular review) keeps tool surfaces tight. The tabular agent has a single tool: read_table_cells, with the same ownership guard as the main agent. System prompt is built per-request with review title, column manifest, and document list.

The bug cluster after wiring is instructive. columns_config comes back from Aurora as a jsonb string and needs JSON.parse before use in the generation endpoint (66a5cec). The Bedrock client was being re-instantiated on every call; 66a5cec adds a module-level singleton (let _client: BedrockRuntimeClient | null = null). UUID casts on the document IN-query (::uuid on numbered params) appear in multiple places across the tabular route - same class of bug as the main Aurora migration, just showing up later.

The shared_with and columns_config parse-on-read pattern lands in GET and PATCH tabular review responses (6e5b266, 8a09da7), in workflow responses (714bcbd), and in mapTRMessages for the frontend (8c7e740). Each requires the same three-way check: typeof x === 'string' ? JSON.parse(x) : (x ?? []).

The retry attempt and revert are worth noting. 54fd7a3 adds a doRequest wrapper that retries once after 1.5 seconds on HTTP 500. 720650f reverts it immediately with the explanation that retrying idempotent and non-idempotent endpoints uniformly produces stale UI in unrelated parts of the app. The right fix was a resolve-batch endpoint (1fede12c) that reads S3 once, applies all bulk accept/reject operations in memory, and writes once - removing the race condition at the source.

Tabular chat storage migrates from Aurora to S3 in 29b80f4, matching the main agent's session persistence approach. Short-lived tabular chat turns stored in a relational DB were a mismatch to start with.

So what Worth a look if you're building a split-agent architecture and want to see how tabular review chat separates from general chat. The `resolve-batch` pattern for bulk edit operations is the most portable single piece. Skip the retry code entirely - it was reverted for good reason. The jsonb parse-on-read fixes are necessary infrastructure for anyone on Aurora Data API, not specific to this feature.

View this fork on GitHub →

Spotted something wrong? Or know the PR text has fresher detail than the writeup above?

Commits in this thread

13 commits from dropthejase/louis, oldest first. Source extracted verbatim from the harvested git log.

SHA	Subject	Author	Date
`1a79b1c0`	Add louisTabular agent and wire tabular chat to AgentCore	Jason Lee	2026-05-09	↗ GitHub
commit body - agents/app/tabular/: new Strands agent mirroring main structure; single read_table_cells tool; system prompt built per-request from review context - backend: replace POST /:reviewId/chat (runLLMStream) with two thin endpoints: POST /:reviewId/chats (create) and POST /:reviewId/chats/:chatId/messages (persist) - frontend: streamTabularChat now calls AgentCore louisTabular URL; TRChatPanel pre-creates chat record before streaming, persists turn after [DONE] - config: NEXT_PUBLIC_AGENTCORE_TABULAR_URL added to config + env example - docs: README and ARCHITECTURE updated
`1fede12c`	feat(api): add resolve-batch endpoint to fix bulk accept/reject race condition	Jason Lee	2026-05-12	↗ GitHub
commit body Single Lambda reads S3 once, applies all resolutions in memory, writes once. Frontend handleAll sends one batch request per document instead of N parallel. Also constrains agent to single-paragraph edits to avoid cross-para deletion failures.
`6e5b266e`	fix(api): parse columns_config and shared_with JSON strings in tabular review responses	Jason Lee	2026-05-13	↗ GitHub
`8a09da7b`	fix(api): parse columns_config and shared_with in PATCH tabular review response	Jason Lee	2026-05-13	↗ GitHub
`a0d653c1`	feat(tabular): optimistic update when adding documents to review	Jason Lee	2026-05-13	↗ GitHub
`54fd7a39`	feat(frontend): retry apiRequest once on 500 after 1.5s delay	Jason Lee	2026-05-13	↗ GitHub
`720650fc`	revert(frontend): remove apiRequest retry - caused stale UI across endpoints	Jason Lee	2026-05-13	↗ GitHub
`66a5cecb`	fix(tabular): fix generation failures - columns_config parse, Bedrock client, UUID cast	Jason Lee	2026-05-13	↗ GitHub
commit body - Parse columns_config from RDS Data API string before use in generate endpoint - Remove stale 'const client = client' self-references in bedrock.ts (ReferenceError) - Rename queryGeminiAllColumns/queryGemini to queryBedrockAllColumns/queryBedrock - Add ::uuid cast to :pid${i} params in shared-by-project query
`e2ed854f`	fix(tabular-agent): fix credential bootstrap, columns_config parse, and UUID cast on document IN query	Jason Lee	2026-05-14	↗ GitHub
`714bcbd1`	fix(api): parse columns_config JSON string in workflow responses	Jason Lee	2026-05-14	↗ GitHub
commit body Aurora Data API returns jsonb columns as raw strings. withWorkflowAccess was spreading the raw DB row without parsing columns_config, causing TypeError: .sort is not a function on the WorkflowPage and corrupt double-serialized data when creating a tabular review from a custom workflow. Parse in withWorkflowAccess so all workflow endpoints (GET /, GET /:id, PUT, PATCH) are covered by a single fix.
`8c7e7405`	fix(tabular): parse jsonb content/annotations in mapTRMessages	Jason Lee	2026-05-14	↗ GitHub
`42c8e187`	fix(tabular): suppress CITATIONS from stream, fix citation schema and extraction	Jason Lee	2026-05-14	↗ GitHub
`29b80f48`	feat(tabular): migrate tabular chat storage from Aurora to S3	Jason Lee	2026-05-14	↗ GitHub
commit body Tabular review chat messages now stored in S3 (conversations/{chatId}/messages.json) using the same Strands snapshot format as the main chat agent. This gives the tabular agent full conversation history on every turn (agent memory). - agent.ts: loadMessages (S3 GET) + AfterInvocationEvent hook (S3 PUT) - index.ts: load previous messages, pass to createAgent, tail-buffer suppresses <CITATIONS> - sessions.ts: new shared lib - S3 read/write + snapshotMessagesToSessionMessages - chat.ts: remove ~250 lines of private session logic, import from sessions.ts - tabular.ts: replace Aurora message reads/writes with S3 via sessions.ts - mikeApi.ts + TRChatPanel.tsx: remove Aurora-specific message mapping

Capture this thread into my fork

Download a single Markdown prompt that tells Claude how to port every commit above into your working tree — adapting paths and structure to match your repo. Run it via claude -p < capture-thread-345.md from inside the repo you want the changes in.

⬇ Download capture-thread-345.md