Skip to content
This repository was archived by the owner on Aug 3, 2026. It is now read-only.

ROADMAP.md

Latest commit

 

History

History
123 lines (91 loc) · 6.11 KB

ROADMAP.md

File metadata and controls

123 lines (91 loc) · 6.11 KB
Raw

Roadmap

The MVP wedge is Claude Code PreToolUse enforcement + deploy/migration verifiers. Everything else builds from that.

Phase 0 — Scaffolding (you are here)

  • Repo created, transferred to eidos-agi org, public.
  • Architecture, runbook, policy, and verifier docs.
  • JSON Schema for runbook templates.
  • Example runbook checked in (examples/).
  • License (Business Source License 1.1, converts to Apache 2.0 on 2030-04-20).

Phase 1 — Single-Node Control Plane

Build the minimum StepProof daemon that can track workflow state.

  • Postgres schema: runbook_templates, workflow_runs, step_runs, policy_decisions, audit_log, liveness_heartbeats.
  • HTTP API:
    • POST /runs — start a workflow from a template.
    • POST /runs/:id/evidence — submit step evidence.
    • POST /runs/:id/heartbeat — register/refresh liveness TTL.
    • POST /policy/evaluate — evaluate a proposed action. Returns {decision, reason, policy_id, suggested_tool, trust_signals}.
    • GET /runs/:id — current state.
    • GET /runs / GET /audit — cursor-paginated listings.
  • YAML rule-set policy engine (OPA/Cedar pluggable later). Supports allow / deny / transform / require_approval / audit decisions, priority-ordered rules, content-pattern matching.
  • Ring classifier — every action maps to Ring 0/1/2/3 per ADR 0002. Unclassified → Ring 3.
  • Three-property trust state per ADR 0003 — identity + authority + liveness as independent gates.
  • Shadow mode — shadow: true on rules and runbooks; log-only evaluation.
  • Append-only audit log with content-addressed payloads and compliance_tags[].
  • Alarm-based step timeouts (no synchronous waits on step completion).

Phase 2 — Claude Code Adapter

The first hook-based enforcement surface. Follows the idioms documented in LESSONS_FROM_HOOKS_MASTERY.md — uv single-file scripts, exit-code contract, matchers in settings.json.

  • PreToolUse adapter that calls StepProof's /policy/evaluate.
  • PermissionRequest adapter — log + updatedInput transform support (not interactive prompts).
  • SubagentStart / SubagentStop adapters — record verifier dispatch lifecycle to audit log.
  • SessionStart adapter — inject active runbook, current step, allowed tools via additionalContext.
  • PreCompact adapter — re-inject runbook state so the worker doesn't forget after compaction.
  • UserPromptSubmit adapter — early soft-nudge when the prompt mentions denied tools.
  • SessionEnd adapter — mark abandoned runs.
  • Setup adapter — first-run installer: creates .stepproof/, drops verifier agents, writes scoped matchers.
  • Deny messages routed back to the agent via exit 2 + stderr, including suggested alternatives.
  • Graceful degradation: control-plane outage must not break the session. Local JSONL audit buffer at .stepproof/audit-buffer.jsonl, flushed on reconnect.
  • Verifier subagent definitions under .claude/agents/stepproof/ with disallowedTools enforced read-only.
  • runbook-author subagent (meta-agent pattern) — generates valid runbook YAML from plain-English descriptions.
  • Custom slash commands: /runbook-start, /runbook-status, /step-complete, /step-evidence, /approve, /runbook-abandon.
  • Per-session state at .stepproof/sessions/<session_id>.json binding session → run → step.
  • stepproof run start <template> CLI to open a workflow.
  • stepproof step complete <step_id> --evidence key=value ... CLI for step completion.

Success criteria: a Claude Code session on a configured runbook cannot run raw psql when the runbook requires cerebro-migrate, and cannot advance to a production deploy without a verifier pass on the staging migration.

Phase 3 — Tier 1 Verifier Library

Ship the deterministic checks that cover most real runbooks.

  • verify_ci_green
  • verify_migration_applied
  • verify_deploy_succeeded
  • verify_git_branch
  • verify_pr_merged
  • verify_env_var_set
  • verify_secret_rotated
  • verify_health_endpoint

Phase 4 — Tier 2 Verifier

Small-model verification for unstructured evidence.

  • Haiku-based verifier worker.
  • Prompt templates per verification_method.
  • Structured output validation; auto-retry on malformed JSON.
  • Cost and latency telemetry per verifier method.

Phase 5 — CI/Deploy Adapters

Push StepProof into the full deployment pipeline.

  • GitHub Actions gate: fail workflow if required StepProof steps are not verified.
  • Railway deploy wrapper.
  • Generic HTTP PreDeploy webhook.
  • PreCommit / PreMerge git hooks.

Phase 6 — Broader Agent Platforms

Generalize beyond Claude Code.

  • OpenAI Agents SDK adapter.
  • Cursor integration.
  • MCP-server mode so any MCP-capable agent can be governed.
  • Agent identity + attestation (who is this worker, really?).

Phase 7 — Human Approval Workflow

Route policy-flagged actions to humans cleanly.

  • Approval UI (web or Slack-first).
  • Time-boxed approvals.
  • Multi-party approval for critical runbooks.
  • Mobile-friendly approval flow for on-call scenarios.

Phase 8 — Tier 3 Guardrail Verifier

Heavy-model verification, opt-in per step.

  • Guardrail library: data-handling, security, compliance, architecture.
  • Per-org policy overrides.
  • Evidence-replay for audit.

Beyond

  • Domain packs: SRE runbooks, security runbooks, data-pipeline runbooks.
  • Learning loop: policies that tighten based on near-miss patterns in the audit log.
  • Multi-agent coordination: workflows with multiple workers, each gated independently.
  • SOC 2 / ISO audit exports straight from the audit log.

Non-Goals (for now)

  • Replacing CI. StepProof augments CI; it does not rebuild it.
  • Being a general-purpose workflow engine. Durable execution is a dependency (Temporal, Cloudflare Workflows, etc.), not a product line.
  • Prompt optimization or agent orchestration. StepProof governs agents; it does not coordinate them.