Skip to content

P21_AGENT_CONSOLE_AUDIT.md

Latest commit

 

History

History
77 lines (56 loc) · 6.53 KB

P21_AGENT_CONSOLE_AUDIT.md

File metadata and controls

77 lines (56 loc) · 6.53 KB
Raw

Pasted21 Agent Console Audit

Scope

Pasted21 replaces placeholder Agent Console data with safe, backend-connected operational metadata while preserving Freya’s existing canonical architecture. The implementation does not create a second task registry, memory system, scheduler, router, browser stack, or autonomy manager. It reads the existing BackgroundJobService, WorkflowOrchestrator, AutonomyManager, MemoryCoordinator, ObservabilityHub, local provider, and canonical SafetyGate.

Implemented backend integration

The existing UI server now exposes read-only aggregation routes:

Route Purpose
/api/agent-console Bounded combined snapshot for Tasks, Memory, System Status, and Autonomy
/api/tasks Current background jobs, bounded job history, active autonomous work, and active workflows
/api/memory/status Safe counts and readiness metadata only; no conversation text, prompts, embeddings, or private records
/api/system/status Readiness/component state plus real CPU, RAM, disk, process-memory, GPU-if-available, and backend uptime values
/api/autonomy/status Authoritative lifecycle state and worker counts
POST /api/autonomy/start Explicitly starts the canonical AutonomyManager after UI confirmation
POST /api/autonomy/stop Stops the canonical autonomy workers and learning pipeline

Task records are normalized from the existing job and workflow owners. They preserve real identifiers, origin, status, timestamps, duration when known, goal/workflow/trace identifiers when present, waiting reasons, approval flags, and failure summaries. Unknown progress is returned as null; no progress percentage is fabricated. Job summaries expose only sanitized provenance fields and never return raw metadata payloads.

Memory status is intentionally metadata-only. The surface reports working-memory activity, bounded conversation-context count, store/retrieval readiness, and learning-pipeline availability/counts where the canonical service exposes them. It does not expose message content, hidden reasoning, embeddings, raw records, or private summaries.

System Status now consumes the existing ObservabilityHub flat system.* metrics. The live service reports measured CPU, RAM, disk, current process memory, and backend-process uptime. GPU and VRAM remain explicitly Unavailable on this host because no GPU measurement was available. The canonical orchestrator SafetyGate is surfaced as Ready, including its balanced policy mode and assessment/pending-approval counts.

Autonomy lifecycle and safety

Normal UI startup constructs the real autonomy manager in OFF and does not start autonomous workers. The initializer now supports an explicit start_autonomy_on_boot flag, defaulting to False. The Agent Console’s Enable control requires a confirmation dialog that explains autonomous initiation and states that SafetyGate still controls authorization. The backend then performs the real start/stop operation through the existing AutonomyManager.

The AutonomyManager now exposes authoritative OFF, STARTING, ON, STOPPING, and ERROR state metadata, along with started_at, last_error, learning-pipeline state, watchdog state, self-initiated work count, and maintenance state. Start failures roll back already-started components and return an error state. Stop remains idempotent and clears worker state. Existing SafetyGate enforcement and approval boundaries are unchanged.

Frontend integration

The existing Agent Console now uses /api/agent-console polling every eight seconds and continues to consume the existing /api/avatar-events stream for operational activity events. The Activity tab preserves the existing timeline and trace display. The Tasks tab renders active and recent backend task records, origin labels, status, durations, waiting-for-approval markers, and bounded failure summaries. The Memory tab renders safe metadata and unavailable states with retry behavior. The System Status section renders real component readiness, hardware metrics, autonomy state, and lifecycle controls. The existing chat, attachment, microphone, scrolling, image-result, and backend event behavior remains intact.

Verification

Check Result
Modified Python syntax checks Passed
Frontend production build (pnpm --dir client run build) Passed
Permanent tests/test_pasted21_agent_console.py 6 passed
Focused combined suite: Pasted21, pasted18 live paths, research, routing, browser, and SafetyGate tests 100% passed; no failures
Normal startup autonomy state OFF, running: false
Explicit autonomy start API 200, authoritative state ON, learning/watchdog/self-initiated/maintenance workers running
Explicit autonomy stop API 200, authoritative state OFF, workers stopped
Live Agent Console Tasks tab Real recurring background jobs displayed; placeholder-unavailable copy absent
Live Agent Console Memory tab Real metadata displayed; placeholder-unavailable copy absent
Live Agent Console confirmation Confirmation dialog and SafetyGate explanation displayed
Live autonomy transition OFF → ON → OFF passed
Live chat after console interaction Assistant response displayed and composer remained present
Browser console errors during live UI probe None observed
Final live SafetyGate status Ready, balanced mode
Final live hardware status CPU, RAM, disk, process memory, and uptime populated; GPU/VRAM explicitly unavailable

Honest limitations

The host did not expose a GPU device to the existing monitoring stack, so GPU and VRAM are shown as Unavailable. Browser and Research do not currently register dedicated readiness components in the existing health surface, so the console reports those component-specific states as Unavailable rather than claiming they are ready. This does not replace or disable the existing browser/research capabilities; it is a limitation of the currently exposed readiness metadata. No new provider or fake health result was added.

Repository state

Pasted21 source changes are intended for commit and push only from the following files:

  • app/ui/agent_console.py
  • app/autonomy/manager.py
  • app/core/background_jobs.py
  • app/core/initializer.py
  • app/core/protocols.py
  • app/memory/coordinator.py
  • ui_server.py
  • client/src/pages/Home.tsx
  • client/src/index.css
  • tests/test_pasted21_agent_console.py
  • P21_AGENT_CONSOLE_AUDIT.md

Temporary Pasted21 probes and patch scripts were removed. Unrelated/pre-existing working-tree files remain untouched.