* fix(replay): repair saved-workstream tool result rendering + extend audit-trail decoration
Loading a saved workstream silently dropped tool results and missed
verdict / output-guard / truncation signals on replay. Root cause was
in `Pane.prototype.replayHistory`: an assistant message carrying both
content and tool_calls cleared the `lastToolBlock` anchor before the
following tool-result iteration could attach. The fix reorders content
to render before the tool block (matching live SSE order) and
restructures the tool-result branch to anchor by `data-call-id` so
multi-tool batches render `[hdr A][out A][hdr B][out B]` rather than
bunching outputs at the bottom.
Beyond the bug, replay now reaches near-parity with the live UX:
- Persisted intent verdicts and output_assessments flow through both
the SSE replay (`_build_history`) and the `/history` REST endpoint
used by coord. Single shared helper module owns the wire shape.
- Memory/recall calls persist instead of being filtered at storage
time — full audit trail; UI dims them by default with hover-reveal
so heavy memory usage doesn't crowd the narrative.
- Truncation indicator surfaces as a sibling pill (consistent across
interactive + coord) when a tool result hit the 2000-char cap.
- `replayHistory` wraps DOM work in `aria-busy` so screen readers
don't get a chatty announce-flood on long replays.
- `_build_history`'s storage I/O moves off the event loop via a new
`events_replay_prepare` async hook for the SSE path; other async
callers wrap in `asyncio.to_thread`.
Coord parity:
- `/history` REST endpoint decorates tool_calls with verdict +
output_assessment + truncation flag (was previously raw
`load_messages` output).
- Coord JS stamps `judge_verdict` / `heuristic_verdict` from
history-loaded `tc.verdict` so the existing batch render paints
the persisted pill, seeds the verdict cache to dedupe later live
SSE events, and emits an inline `.coord-tool-row-warning` chip
per call instead of a generic chat line.
- Memory/recall dim rule mirrored on `.coord-tool-row[data-tool-name=...]`.
* fix(replay): address PR #461 review feedback + raise tool-result storage cap
Copilot review feedback:
- Sibling-chain dim rule (memory/recall) now adds :focus-within
alongside :hover for .tool-output / .media-embed / .output-warning
/ .tool-output-truncated — keyboard users tabbing into a faded
subtree now get full opacity.
- ``cfg.open_post_load`` is now invoked via ``await asyncio.to_thread``
so its sync ``_build_history`` call (storage I/O for verdict
indexes + message reconstruction) doesn't block the event loop on
every workstream open. Mirrors the SSE replay path that's already
protected via ``events_replay_prepare``.
- Replaced the hardcoded ``2000`` literal in server.py and session.py
with ``TOOL_RESULT_STORAGE_CAP`` from the shared decoration module
so the UI truncation-pill detection can't silently desync from the
storage write side.
While here:
- Raised ``TOOL_RESULT_STORAGE_CAP`` from 2000 → 10000. A 2000-char
clip routinely cut grep / file-read bodies mid-line, leaving the
audit trail useless for retrospective debugging. FTS5 + row size
grow proportionally; the per-tool upper bound is still bounded
upstream by ``_truncate_output``'s context-budget clamp.
- Updated the user-visible truncation-pill tooltip on both
interactive and coord to reflect the new cap.
- ``test_decorates_tool_calls_and_marks_truncated`` now references
the constant instead of a literal so it stays correct on future
cap changes.
Patrick Buckley
eaabc79eb3