Agent Observability Replay — Full-Fidelity Session Reload

[Number531]

Summary

Agent observation cards now reconstruct full activity detail on session reload, matching the live streaming view. Previously, reloading a session showed only 2 agents (those with state files) and displayed "No activity recorded" — now all 29+ agents appear with granular tool-by-tool activity logs including durations, success/failure, search queries, and file targets.

Version: v4.10.0
Commit: f58e960

Changes

Backend — Hook Data Enrichment (hookDBBridge.js)

• toolTimers Map: PreToolUse→PostToolUse elapsed time → duration_ms
• agentTimers Map: SubagentStart→SubagentStop elapsed time → duration_ms
• Per-hook-type success derivation (!tool_response.isError, false for failures, true for agents)
• Compact tool_input extraction: query/pattern/command/url into event_data JSONB
• WebSearch/WebFetch native SDK tool name matching (alongside MCP variants)
• AUDIT_SKIP_TOOLS emptied — Read/Grep/Glob now persisted to hook_audit_log
• persistState reads input.duration_ms for agent_states table

Backend — Query & API (dbFrontendRouter.js, postgres.js)

• UNION ALL query: stateless research agents sourced from hook_audit_log SubagentStop
• GET /api/db/sessions/:key/agent-events/:agentType — per-agent tool event feed
• agentType regex validation (/^[a-z0-9_-]+$/i)
• Composite index idx_audit_session_agent_event_time

Frontend — Playing Card Popup (app.js, styles.css)

• Async activity feed: fetches granular events from DB, renders tool/target/context/duration
• renderAgentStats() called on reload (stats grid populated)
• CSS .detail-* rules scoped under .agent-obs-activity (cascade fix)
• Activity feed max-height: 200px → 320px

Remaining Frontend Enhancements

The following items would further close the gap between live and reload views:

• Transcript summary fallback — when DB events are unavailable (pre-v4.10.0 sessions), parse transcript_summary from SubagentStop event_data for a condensed activity view
• Live-to-reload parity for running agents — agents that were actively running at time of page close don't get agent_type written at completion; capture in-progress state
• Rapid popup race condition — if user opens/closes popups quickly, pending fetch could write to wrong #agentObsActivityFeed element; add generation counter or AbortController
• Gate check detail in popup — show individual gate check pass/fail (data exists in event_data.gate_checks_passed/total) as a visual checklist
• Tool usage breakdown chart — render tool category distribution (WebSearch, MCP, Read/Write, Bash) as a mini bar chart in the popup header area
• Duration sparkline — visualize per-tool durations as a horizontal bar or sparkline in the activity feed
• Backfill script — one-time migration to compute duration_ms for existing SubagentStop events from SubagentStart/Stop timestamp deltas in hook_audit_log

Test Plan

• All 4 modules import cleanly (hookDBBridge, hookDBBridgeConfig, dbFrontendRouter, postgres)
• 11/11 hook data flow scenarios pass (PreToolUse deny, PostToolUse success/failure, PostToolUseFailure, SubagentStart, SubagentStop with/without duration, PreCompact, SessionStart)
• Endpoint validation: valid agent types pass, injection attempts blocked
• CSS scoping: 0 unscoped .detail-* rules, 8 scoped rules
• wrapHooksForDB no-pool passthrough: returns original config unchanged
• Composite index created (idempotent CREATE INDEX IF NOT EXISTS)
• Live session verification — run full session post-restart, confirm enriched data in DB and populated cards on reload

🤖 Generated with Claude Code

[Number531]

Update (2026-03-31): DB persistence + session replay shipped

Shipped in v4.12.0–v4.13.0:

• agentProgressSummaries: true on both P0 and main agentQuery() (SDK 0.2.72)
• DB persistence: persistProgressSummary() writes AgentProgress events to hook_audit_log (fire-and-forget, 500-char cap)
• Session replay endpoint: GET /api/db/sessions/:sessionKey/agent-progress — groups by agent_id, 500-row limit
• Frontend hydration: reconstructPhaseProgress() fetches progress history async, populates agentProgress Map, re-renders agent panes
• Analytics exclusion: 3 queries in dbFrontendRouter.js have NOT IN ('AgentProgress')

Still blocked on upstream:

• Live streaming of task_progress messages — SDK 0.2.72 generates them internally but doesn't yield to stdout transport
• All code is inert when messages absent; will activate automatically when upstream ships

New since last update:

• Knowledge Graph extraction runs post-session, visible in session replay via kg_evolution timeline
• Agent pane observation popup shows timestamped activity logs for completed agents

[Number531]

Stale-open: shipped as v4.10.0 (commit f58e960) with full backend + frontend implementation. Closing for hygiene.