Epic: Agent platform hardening — learning loop, background automation, runtime safety #11260

Goal

Harden AutoBot's agent platform across three axes where we have the machinery but haven't fully wired it into the primary path: the learning loop, background automation, and runtime safety/governance. Each task below is a self-contained PR.

Motivation

An audit of our agent subsystems found several capabilities that are built and production-grade in the orchestration path but absent from the single-agent chat loop (where most real work happens), plus a few safety gaps with no current coverage. This epic closes those gaps by reusing existing components rather than adding new frameworks.

Tasks

• Task 1 — Chat-loop trajectory learning (priority: high)
  Wire trajectory pre-fetch (search-before) and outcome-capture (store-after) into the single-agent chat loop, reusing TrajectoryStore + TaskOutcomeJudge. Today these only run in the multi-agent orchestration path.

  • Files: autobot-backend/async_chat_workflow.py, autobot-backend/memory/trajectory_store.py, autobot-backend/orchestration/workflow_runner.py (reference)

• Task 2 — Worker priority tiers
  Route audit workers to Celery priority queues so audit preempts low-priority cleanup; reserve a scheduled slot for the consolidation worker (Task 3).

  • Files: autobot-backend/celery_app.py, autobot-backend/workers/audit_tasks.py

• Task 3 — Memory consolidation worker
  New Celery Beat worker that periodically compacts/dedupes the trajectories collection to keep retrieval precision high as N grows.

  • Files: autobot-backend/workers/consolidate_tasks.py (new), autobot-backend/memory/trajectory_store.py, autobot-backend/celery_app.py

• Task 4 — Prompt-injection hard-block mode
  Add a confidence-gated hard-block mode to the injection detector alongside the existing quarantine/flag path, for untrusted multi-modal input.

  • Files: autobot-backend/security/prompt_injection_detector.py, autobot-backend/security/content_firewall.py (integration)

• Task 5 — File integrity fingerprint manifest
  SHA256 manifest of security-critical files, verified at startup, to catch out-of-band tampering the write-time config guard cannot see.

  • Files: autobot_shared/config_guard.py (extend) or new autobot_shared/integrity_manifest.py

• Task 6 — Enable governed delegation behind validation
  Flip AUTOBOT_DELEGATION_ENABLED on behind a documented validation plan; the governed subagent runner (chat_workflow/delegation.py) is already built and depth-bounded but dormant.

  • Files: autobot-backend/chat_workflow/delegation.py, config docs

Invariants (all tasks)

• Reuse existing components; no _v2/_enhanced duplicates.
• All learning/audit paths fire-and-forget — never break the primary workflow.
• Each task gets a discrete sub-issue for a clean Closes #.
• Target branch Dev_new_gui; worktree-isolated; code-reviewer before merge.