fix: restore model activity reasoning stream

2026-07-14 15:03:07 +02:00
parent 1b1554b353
commit c7474f3852
11 changed files with 490 additions and 14 deletions
@@ -0,0 +1,68 @@
+# Workflow UI Regressions Design
+
+## Goal
+
+Restore visible model reasoning and make the F4/F6 review experience deterministic,
+readable, and resilient to malformed model-authored artifacts.
+
+## Diagnosed causes
+
+- Session creation and resume force Pi `thinking` to `off`, regardless of the saved setting.
+- The backend bridge forwards assistant `text_delta` events but drops `thinking_delta` events.
+- The Model Activity panel reads final assistant text rather than a dedicated activity stream.
+- F4 joins use the editable `multiselect` contract; `recommended` is not translated into
+  `selected` outside F2, so every join initially appears unchecked.
+- The CTE result viewer passes one SQL block to a viewer that always displays its
+  multi-block Horizontal/Vertical control.
+- CTE plan cards flatten long filters, table names, keys, and rationale into loosely spaced text.
+- The F6 phase-summary payload accepted an object in `open_questions`; React then attempted to
+  render that object directly and the widget error boundary displayed the generic failure message.
+- The F3 auto-approval fix was present in the image tag but not in the running containers because
+  they were built without being recreated.
+
+## Design
+
+### Model activity
+
+Session creation uses the global `thinking` preference and resume uses the persisted manifest
+preference, falling back to the current global setting. `SessionBridge` maps Pi's nested
+`thinking_delta` into a distinct SSE `activity_delta`. The frontend stores that stream separately
+from final assistant text, and Model Activity renders only activity deltas. This prevents internal
+reasoning from leaking into transcript-oriented state while making the panel accurately reflect
+the configured model activity.
+
+### Join review
+
+F4 calls to `reviewer_decide` whose merit decisions are all `join_modified` become a read-only
+`join-review` widget. Each proposed join is rendered as an informational card with its name,
+join expression, and rationale. `Continue` returns every join id; the gate persists all decisions.
+`Other — specify` returns textual feedback without persisting the current proposal, so the model
+must revise and present the complete join set again. Mixed join/non-join calls remain regular
+multiselects, and the skill instructs the model to keep joins in a separate call.
+
+### CTE presentation
+
+The plan viewer uses a responsive card layout: a compact numbered header, purpose and rationale
+as readable prose, aligned metadata rows for dependencies and keys, wrapped table rows, and a
+structured filter grid with separate column/operator/value fields. Output columns remain compact
+wrapping chips. The SQL layout switch is hidden whenever the viewer receives fewer than two blocks.
+
+### Phase-summary resilience
+
+The gate validates that `open_questions` is an array of strings and reports a corrective error to
+the model before emitting a widget. The frontend also normalizes legacy malformed entries to a
+safe string (preferring `question`, then `label`) so old or externally produced payloads cannot
+crash React. The session skill documents the exact `string[]` contract.
+
+### Deployment
+
+After targeted and full regression tests, build both Docker images and recreate the Compose
+services. Verify that the running containers use the newly built image ids and that both health
+checks pass.
+
+## Non-goals
+
+- Persisting verbatim model reasoning in session artifacts.
+- Allowing individual joins to be removed from the join review.
+- Redesigning multi-block SQL comparison behavior.
+- Changing the eight-phase workflow or F3 auto-approval semantics.