Skip to content

operator-guide.md

Latest commit

 

History

History
81 lines (55 loc) · 3.49 KB

operator-guide.md

File metadata and controls

81 lines (55 loc) · 3.49 KB

Operator Guide

Operator-facing APIs explain runtime state, safe recovery actions, lifecycle reports, and capability visibility. They do not provide a UI or a separate management plane.

Inspect

use bm_sdk::{MemoryInspectionRequest, PressureLevel, RuntimeLifecycleModeInput};

let report = runtime.inspect(MemoryInspectionRequest {
    query: "release status".to_string(),
    system_max_len: 4096,
    pressure: PressureLevel::Normal,
    mode_input: RuntimeLifecycleModeInput::default(),
})?;

assert!(report.capabilities.inspection.visible);

MemoryInspectionReport includes working recall inspection, capability catalog data, deferred governance queue state for the current runtime scope, operator action report, and lifecycle report.

Use inspect after migration dry-run/apply to confirm selected ids, recall planes, deferred job status, lifecycle diagnosis, and safe recovery actions. Operator surfaces may show this report, but they must not read store internals or invent their own plane-count logic.

Projection diagnostics come from MemoryProjectionReport.audit, including source planes, selected ids, section chars, budget, and private gate decisions. Conservative compaction comes from MemoryRuntime::run_retention_compaction(), whose report explicitly forbids host-side deletion of accepted memory.

Runtime Skill And Standard Agent Skill

Standalone consoles and CLI operators manage runtime Skill Memory only: procedural memory records learned during runtime. They are not executors, marketplaces, workflow runners, or managers for standard Agent Skill directories.

SDK entry points:

  • MemoryRuntime::list_runtime_skills
  • MemoryRuntime::get_runtime_skill
  • MemoryRuntime::edit_runtime_skill
  • MemoryRuntime::set_runtime_skill_enabled
  • MemoryRuntime::delete_runtime_skill

Every mutation enters MemoryRuntime, then core skill governance and the configured store backend. The HTTP console only routes runtime Skill view/edit/enable/delete operations under /console/skills*; the CLI only calls the entry facade. Neither path reads or writes skill files directly.

Standard Agent Skills stay host-managed. The SDK provides MemoryRuntimeBuilder::agent_skill_dirs / add_agent_skill_dir for read-only mounts, and standalone HTTP/CLI deployments can use BM_AGENT_SKILL_DIRS. Recall and projection use only SKILL.md summaries, resource counts, and fingerprints; scripts are not executed, assets are not read, and the directory never becomes a memory-owned management object. ESP profiles reject standard Agent Skill directory mounts.

Recover

use bm_sdk::{MemoryRecoverRequest, RuntimeLifecycleModeInput, RuntimeLifecycleTrigger};

let recovered = runtime.recover(MemoryRecoverRequest {
    trigger: RuntimeLifecycleTrigger::OperatorRequested,
    mode_input: RuntimeLifecycleModeInput::default(),
})?;

Recover acts on recoverable runtime/lifecycle state. It does not skip store repair reports or rewrite the persistence schema.

Close

use bm_sdk::MemoryCloseRequest;

let closed = runtime.close(MemoryCloseRequest {
    reason: "release smoke complete".to_string(),
})?;

Close emits a lifecycle event. A process supervisor can decide whether to exit based on the returned report.

CLI Inspection

cargo run -p bm-cli --bin bm -- \
  platform capability-snapshot \
  --profile profile-esp-standalone-memory

Memory commands also go through bm-entry:

cargo run -p bm-cli --bin bm -- \
  memory capabilities \
  --profile profile-server-linux-dev-full