feat(dashboards): SQL-defined datasets, publish gate, list/load tools - #346
Merged
8 commits merged intoSep 9, 2026
Merged
Conversation
…ed result
A refresh failed in staging because the agent recorded a base query and then
derived the chart columns in pandas, so the recorded SQL could not reproduce
the dataset. A dataset is now {connection, sql} (or static rows); the
snapshot at artifacts/datasets/<name>.csv is written only by the new
sp.dashboard_dataset helper, which runs the query through the governed path
and records a sidecar with the SQL hash and columns.
- notebook-server: sp.dashboard_dataset; the check tools report
snapshot_missing and snapshot_stale (SQL or columns differ from the
snapshot); CSV-only snapshots; skill rewritten around SQL-only datasets
with a TTM window and CASE example.
- gateway: publish runs each dataset's SQL once (row_limit 1) and rejects
with 422 not_repeatable / sql_failed when the result cannot feed the
charts; refresh drops file carry-forward; one query seam for gate and
refresh; edit chats no longer inline dataset bytes.
- web: renderer and publish dialog use the new dataset shape, show
"SQL · <connection>" or "Static", resolve snapshots by convention, and
render the per-dataset gate errors; fixtures converted.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XfZZecsk7TeJvfaLxD2d3g
A fresh chat can now find and edit a published dashboard. Two sandbox tools, dashboard_list_published and dashboard_load_published, read the gallery with the run's identity and materialize artifacts/<slug>.dashboard.json plus its snapshots and sidecars, so the file passes the freshness checks and renders at once. Publishing stays with the user: the publish dialog preselects the matching dashboard by slug and the file view shows "Loaded from". Also: the sandbox run token could reach the dashboard write routes; every write route now refuses execution identities. Edit chats no longer inline the spec; they ask the agent to load it. The screenshot projection parses past the runtime's "[image]" marker so the card shows rendered charts. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XfZZecsk7TeJvfaLxD2d3g
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
…book skill The SDK query route parsed SQL with the default Postgres dialect before it knew the connection, so valid T-SQL such as `select top 1` was rejected with "Query blocked" while the MCP query_database path accepted it. The governed executor already validates with the connection's dialect; the pre-check is removed and covered by regression tests. The notebook-specific instructions move out of the chat system prompt into a new `signalpilot-dbt:notebook` skill (tools, sp SDK, cell rules, dialect rule, files, charts, troubleshooting). The prompt keeps the basics and points at the skill. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XfZZecsk7TeJvfaLxD2d3g
…ilters Three defects an agent hit while building a dashboard on SQL Server: - Limit injection on a T-SQL UNION/EXCEPT/INTERSECT wrapped the query as a derived table with its trailing ORDER BY inside, which SQL Server rejects (error 1033). The ORDER BY, OFFSET, and FETCH now move to the outer select; the cap is TOP, or FETCH when an OFFSET exists; a branch's own TOP stays on its branch; CTEs are hoisted. Covered by an 18-case matrix. - The sandbox SDK client made one attempt, so a transient 502/503/504 from the tunnel aborted a dataset build. Requests and downloads now retry three times with backoff; other HTTP errors stay final. - A dashboard filter could bind to only one dataset. `dataset` is now optional: an unbound filter applies to every dataset that has the column, so one control drives all charts sharing it. Bound filters are unchanged. Mirrored in the renderer, the sandbox checks, and the publish gate. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XfZZecsk7TeJvfaLxD2d3g
…in batches Dashboards: - KPI tiles accept text cells: a value or comparison without a numeric format renders the cell as text, so a top-region name comes from SQL instead of a hard-coded description. Same rule in the renderer, the sandbox checks, and the publish gate. - New chart types: combo (grouped or stacked bars with lines on a secondary axis) and heatmap (category axes with a continuous colour scale), in the schema, renderer, print path, checks, and the plugin example. - Skill: dataset reuse guidance, TOP/LIMIT in SQL for small snapshots, and the text-KPI rule. Chat streaming: the worker persisted every streamed chunk as its own run event with six database round trips, which on a remote database delivered about one chunk per second. Deltas now coalesce per run and flush every 150 ms, at 2,000 characters, or before any other event (including the terminal status), and append_event allocates the sequence with one atomic UPDATE ... RETURNING. Measured on the local staging stack: 5 chars/s and ~900 events per answer became 214 chars/s and 9 events. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XfZZecsk7TeJvfaLxD2d3g
…itle
The analysis notebook no longer starts with placeholder scaffold cells
("Analysis output / Pending governed notebook analysis"). Every notebook
seeds with the hidden setup cells and one empty visible cell; the notebook
skill tells the agent to put a short title there as the first visible cell.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XfZZecsk7TeJvfaLxD2d3g
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Follow-up to #344/#345 on the same branch.
{connection, sql}or staticrows); the snapshot is the cached result written bysp.dashboard_dataset. The sandbox check tools flagsnapshot_missing/snapshot_stale.not_repeatable/sql_failedwhen it cannot feed the charts. This is what the failed staging refresh needed.dashboard_list_publishedanddashboard_load_publishedso a fresh chat can edit a published dashboard; the user still publishes (dialog preselects the matching dashboard).[image]marker.Verified live on the local staging stack (Dumpsters project): build with the helper, publish through the gate, SQL refresh to v2 with rows; then a fresh chat loaded and edited that dashboard without publishing.
Deployment: no new env vars; the staging notebook image rebuilds automatically on merge (notebook-server changed).
🤖 Generated with Claude Code
https://claude.ai/code/session_01XfZZecsk7TeJvfaLxD2d3g