ui: turn the local session page into an evidence review surface #242

Goal

Make the existing local web dashboard useful for reviewing one agent session: what ran, what changed, what failed, what recovered, and what the instrumentation could not see.

This implements the review-experience sequence in the Verio Telemetry.dev learning. It should replace the current flat event list with a review-oriented session surface, not grow into a generic hosted analytics dashboard.

Dependencies

• Use the verified provider/capture metadata from product: add verified setup diagnostics and a provider capture matrix #239.
• Show the evidence-health result and reason codes from product: calculate and surface per-session evidence health #240.
• Reuse existing replay, timeline, tree, diff/compare, annotations, share, and OTLP export behavior instead of inventing parallel semantics.

MVP scope

Review header

Show:

• session/provider/capture identity;
• start/end/duration;
• evidence-health status and known gaps;
• redaction/content mode; and
• quick actions for compare, minimized/shareable review packet, and OTLP export.

Ordered evidence view

Render a deterministic execution sequence with expandable raw details and source provenance. Preserve parent/child relationships where they are recorded.

Provide focused filters for:

• files read or changed;
• commands and tool calls;
• tests;
• failures and errors;
• retries and recovery;
• human annotations or corrections;
• privacy transformations; and
• capture gaps or ambiguous boundaries.

A reviewer must be able to distinguish an observed failure from a missing result and a provider blind spot.

Inspector behavior

• Link calls to their results when identifiers allow it.
• Show orphaned or ambiguous relationships without guessing.
• Make timestamps, event/source identifiers, redaction state, and raw fields inspectable.
• Preserve provider-specific fields.
• Support keyboard and URL-addressable navigation to an event.
• Keep the page local and usable without a hosted service.

Out of scope

• Organization-wide productivity ranking.
• A new generic trends or token/cost dashboard.
• Hosted collaboration, accounts, billing, or retention.
• AI-generated causal claims presented as facts.
• Flattening provider-specific events into a falsely uniform story.

Acceptance criteria

• A reviewer can follow a representative session from start through failure, retry, recovery, and final effect.
• The header displays evidence health, capture method, redaction mode, and gaps consistently with product: calculate and surface per-session evidence health #240.
• Files, commands/tools, tests, failures, retries, annotations/corrections, privacy transformations, and gaps can be isolated.
• Tool/result and parent/child relationships are shown when supported.
• Missing or ambiguous relationships remain visibly unresolved.
• Every displayed fact can reveal its source event and raw/provider-specific fields.
• Compare, minimized/share output, and OTLP export are reachable as primary actions.
• The representative fixture used by activation and compatibility tests is covered by UI tests.
• The page remains fully local and introduces no hosted-service dependency.