helm-agent-ops 0.9.0

Stability-first operations CLI for long-lived agent workspaces.

Helm

Stop long-running coding agents from losing context, making unsafe edits, and becoming impossible to audit.

Helm is a local operations layer for AI agent workspaces: profiles before commands, checkpoints before risky work, durable task history after the chat is gone.

Current release: v0.9.0

Landing page · 한국어 README

Quickstart · Why Helm · What Helm Adds · Workflows · Docs · Landing Page

Quickstart

Install from PyPI:

python -m pip install helm-agent-ops
helm --help

Or use the workspace bootstrap installer:

curl -fsSL https://raw.githubusercontent.com/JDeun/Helm/main/install.sh | bash
helm doctor --path ~/.helm/workspace
helm profile --path ~/.helm/workspace run inspect_local --task-name "first Helm inspection" -- git status --short
helm status --path ~/.helm/workspace --brief
helm dashboard --path ~/.helm/workspace

The installer installs Helm and creates ~/.helm/workspace. If helm is not found afterward, use the PATH line printed by the installer.

Need a different workspace?

curl -fsSL https://raw.githubusercontent.com/JDeun/Helm/main/install.sh | bash -s -- \
  --workspace ~/work/helm

Why Helm

Helm is for developers who already use coding agents for real work and need the session to leave behind something more durable than chat history.

Use Helm when you want to:

• run agent-adjacent commands under explicit risk profiles
• block destructive or out-of-profile commands before they execute
• create visible recovery points before broad edits
• keep task and command history in local files
• rehydrate future runs from workspace state instead of memory alone
• review what happened after a long session ends

Helm is not another agent runtime. It is the operating layer around the one you already run.

Use it when an OpenClaw/Hermes-style workspace, or a similar self-hosted agent service, has moved past demos and needs repeated work to stay:

• bounded by explicit execution profiles
• recoverable through checkpoints
• inspectable through task and command logs
• resumable from files instead of chat history
• governed by skill contracts and local policy

If the agent only runs one-off demos, Helm is probably unnecessary.

Three-Minute Demo

helm profile --path ~/.helm/workspace run inspect_local \
  --task-name "inspect current repository" \
  -- git status --short

helm checkpoint create --path ~/.helm/workspace \
  --label before-risky-work \
  --include ~/.helm/workspace

helm report --path ~/.helm/workspace --format markdown
helm dashboard --path ~/.helm/workspace

This leaves a task ledger, command log, checkpoint record, and dashboard summary on disk.

How Helm Fits

Category	Better for	Helm adds
Agent frameworks	prompts, planners, tool loops, agent graphs	profiles, guard decisions, checkpoints, task ledgers
Observability tools	hosted traces, service metrics, telemetry correlation	pre-execution policy and local recovery state
Eval tools	scoring model output or task success	operational history around repeated human-agent work
Shell wrappers	command convenience	workspace state, memory capture, reports, and recovery discipline

What Helm Adds

Core ideas:

• Profile: declares the allowed blast radius before a command runs, such as inspect-only, workspace edit, or risky edit.
• Guardrail: checks command shape against local policy before execution, blocking dangerous or out-of-profile actions.
• Checkpoint: preserves a visible recovery point before work that may need rollback.
• Audit trail: records what ran, under which profile, with what guard decision, and what task it belonged to.
• File-backed memory: keeps reusable context in files so later runs resume from durable state instead of chat history.
• Context retrieval: ranks notes, memory, ontology, tasks, commands, and checkpoints through one inspectable query surface.
• Privacy boundary: scans and tokenizes private text before it crosses tool, API, report, or remote handoff boundaries.

Repeated-agent problem	Helm adds
The agent forgets prior work	Context hydration from notes, memory, tasks, commands, and checkpoints
Risky edits happen too fast	Profiles, command guard, and checkpoint discipline
Runs are hard to explain later	Task ledger, command log, status, dashboard, and reports
Private context may leak into tools	helm privacy scan/tokenize/restore with local vault and audit events
Retrieval feels like a black box	helm context --explain-ranking with field, recency, graph, adapter, and source scores
Skill rules live in prompts	SKILL.md guidance plus contract.json execution policy
Model fallback is ad hoc	File-backed health checks and fallback selection
Operational state is scattered	Workspace layout, adopted sources, and SQLite query index

Helm is runtime-agnostic, but it is built first for persistent workspaces with state, memory, profiles, checkpoints, and task history.

Workflows

Inspect the workspace.

helm doctor --path ~/.helm/workspace
helm status --path ~/.helm/workspace --brief
helm dashboard --path ~/.helm/workspace

Run under a declared profile.

helm profile --path ~/.helm/workspace run inspect_local \
  --task-name "inspect repository state" \
  -- git status --short

Adopt existing systems as context sources.

helm survey --path ~/.helm/workspace
helm onboard --path ~/.helm/workspace --use-detected --dry-run
helm onboard --path ~/.helm/workspace --use-detected

Check rollback and recent state.

helm checkpoint-recommend --path ~/.helm/workspace
helm checkpoint list --path ~/.helm/workspace
helm task list --path ~/.helm/workspace --status running
helm task doctor --path ~/.helm/workspace
helm report --path ~/.helm/workspace --format markdown

Query durable context with inspectable ranking.

helm context --path ~/.helm/workspace --mode decisions --explain-ranking --json
helm context --path ~/.helm/workspace --mode timeline --since 2026-05-01
helm context --path ~/.helm/workspace --mode entity --entity project_helm
helm context --path ~/.helm/workspace --mode reflect-candidates

Run a privacy boundary preflight.

helm privacy --path ~/.helm/workspace scan --text "Contact [e-mail]" --json
helm privacy --path ~/.helm/workspace tokenize --scope task-123 --text "Contact [e-mail]"

Review stale negative claims in skill instructions.

helm skill-lifecycle negative-claims --path ~/.helm/workspace --persist
helm skill-lifecycle revalidation-due --path ~/.helm/workspace
helm skill-lifecycle revalidate-claim --path ~/.helm/workspace \
  --skill old-skill \
  --claim-id sha256:abc123 \
  --status resolved \
  --note "command now exists"

Probe model health.

helm health --path ~/.helm/workspace state --json
helm health --path ~/.helm/workspace select --json

Try the demo workspace.

helm doctor --path examples/demo-workspace
helm dashboard --path examples/demo-workspace

Workspace Model

Keep Helm in a dedicated workspace. Treat existing systems as read-only context sources first.

• Helm state lives under .helm/
• profiles, notes, policies, and skill rules stay as explicit files
• OpenClaw, Hermes, and notes vaults can be adopted instead of overwritten
• JSONL remains the append-only source of truth; SQLite is a query index

Docs

Start here:

• docs/three-minute-demo.md
• docs/first-run.md
• docs/onboarding.md
• docs/demos.md
• docs/integrations/openclaw.md
• docs/integrations/existing-agent-workspace.md

Core concepts:

• docs/execution-profiles.md
• docs/memory-operations-policy.md
• docs/ops-memory-query.md
• docs/privacy-boundary.md
• docs/task-finalization.md
• docs/task-state.md
• docs/adaptive-harness.md
• docs/evidence-label-convention.md
• docs/hitl-decision-patterns.md
• docs/skill-quality-and-policy.md

Positioning:

• docs/opensource-product-definition.md
• docs/opensource-module-split.md
• docs/helm-dogfooding-reference.md
• docs/public-launch-checklist.md
• docs/comparisons/agent-frameworks.md
• docs/comparisons/observability-tools.md
• docs/comparisons/eval-tools.md

Release details:

• CHANGELOG.md
• CONTRIBUTING.md
• SECURITY.md
• docs/releases/0.9.0.md
• docs/releases/0.8.0.md
• docs/releases/0.7.3.md
• docs/releases/0.7.2.md
• docs/releases/0.7.1.md
• docs/releases/0.7.0.md
• docs/releases/0.6.7.md
• docs/releases/0.6.6.md

Status

Helm v0.9.0 adds append-only task state operations, completion evidence gates, checkpoint retention planning, outcome-aware skill lifecycle reporting, DCI inspection hints, and human-approved HITL decision pattern tracking. See docs/releases/0.9.0.md.

Helm does not include private memory, personal agent overlays, credentials, or private task history.

License

MIT