Four coordinated capabilities extending the bees domain — giving Worker Bees a safe execution environment, human-readable memory, resilient failure recovery, and autonomous skill authoring.
Each capability ships behind an independent per-hive feature flag. All four are designed to coexist without disrupting existing memory, circuit-breaker, or tool-retry systems.
A persistent execution environment mirroring the main agent's Daytona pattern. Each hive gets its own isolated container with a persistent volume at /workspace/hive/.
A human-readable procedural knowledge layer using MEMORY.md, daily trajectory logs, and a skills library. Complements — never replaces — the existing MemoryService.
A supervisor wrapping Worker execution: mechanical rules → single Reflexion pass → failure note. Hard cost cap of 1 Haiku call per task. Wraps, never replaces, existing resilience systems.
Workers autonomously author markdown playbooks (plus optional sandboxed Python) on successful tasks and self-improve them on re-use. No RL or fine-tuning required.
Each hive gets exactly one Daytona sandbox with a persistent volume. The sandbox mirrors the main agent's pattern with bees-specific configuration for longer idle times and conversational workloads.
A markdown-based procedural knowledge layer that complements the existing pgvector MemoryService. Files are the source of truth for procedural knowledge; the database remains authoritative for declarative facts.
YYYY-MM-DD.md — Today's log (append-only)YYYY-MM-DD.N.md — Roll suffix on 5MB overflowarchive/YYYY-MM.tar.gz — Nightly compression (>30 days)index.md — Auto-maintained 30-day index<slug>.md — Playbook with YAML frontmatter<slug>.py — Optional executable helper.archive/ — Skills over the 100-skill soft capAn append-only bees.memory_events table ships with v1 as migration-readiness infrastructure. Zero behavior change for readers — but enables future replay and file-based migration.
MEMORY.md (above scratchpad)
Worker (RO in prompt)
MarkdownMemoryProjector only
MEMORY.md scratchpad
Worker (RO in prompt)
Worker via SandboxFilesTool
memory/<today>.md
Worker (RO excerpt)
HiveWorkspace.append_trajectory only
memory/<past>.md
Worker (RO via files tool)
nobody (immutable)
skills/<slug>.md
Worker (RO via prompt)
SkillLibraryService only
The HealingSupervisor wraps Worker execution with a three-tier escalation strategy. It sits above — and never replaces — the existing CircuitBreaker, tool-retry, and confidence-escalation systems.
healing.tier1.rule_fired · healing.tier2.reflection · healing.tier3.failure_noteWorkers autonomously author markdown playbooks on successful tasks and self-improve them on re-use. No RL, no fine-tuning — just structured knowledge accumulation through experience.
skill_linter_rules.py. Changes require (a) Change Management RFC, (b) security review by CODEOWNERS, (c) version bump of BEES_SKILL_LINTER_VERSION.
skills/draft_linkedin_post.{md,py}The complete flow for a single task with all four feature flags enabled. Post-task jobs run asynchronously via Dramatiq — zero impact on user-visible latency.
All 16 flag combinations produce a valid, non-crashing system. Each feature degrades gracefully when its flag is off.
Four sequenced PRs, each behind its own feature flag. Dev → staging → prod with 48h dashboard watch between each environment.