Skip to content

V1_READINESS.md

Latest commit

 

History

History
78 lines (60 loc) · 4.24 KB

V1_READINESS.md

File metadata and controls

78 lines (60 loc) · 4.24 KB
Raw

AgentGuard Graph 1.0 Readiness

This document records the stable v1 baseline. Release operations such as tag signing, package publication, and partner fixture expansion are handled separately.

Decision

AgentGuard Graph is ready as the 1.0 stable local-first baseline.

The product is an offline evidence analyzer. It is not a live scanner, policy enforcement layer, MCP proxy, telemetry service, or exploitability engine. A user supplies local evidence packs and receives JSON, Markdown, and self-contained HTML reports that separate possible, supported, observed, blocked, and unknown paths.

Stable Surface

  • CLI commands: scan, inventory, explain, compare, portfolio, validate, init, collect, doctor, and demo.
  • doctor prioritizes missing exports, writes collection plans, supports persona repair views, and blocks packaging when likely secrets are present.
  • collect writes an offline evidence-manifest.json, and doctor reports manifest drift when files are changed, missing, or unmanifested.
  • Scanner runtime uses the Python standard library only.
  • Evidence inputs are JSON, JSONL, and local policy source files for Rego or Cedar.
  • collect imports data classification JSON exports from data catalogs, DLP scanners, sensitivity-label systems, and table/object classification reports into data-catalog.json.
  • collect imports OPA/Rego and Cedar policy source or evaluation JSON into approval-policy.json.
  • collect imports runtime JSON exports from tracing stores, approval brokers, MCP hosts, CI systems, and cloud audit logs into events.jsonl.
  • Malformed or unsupported inputs produce explicit errors or warnings.
  • Reports include review decisions, evidence quality, path state, runtime observation, runtime event-quality scoring, runtime correlation diagnostics, offline control analysis, policy rule-risk analysis, owner-routed remediation plans, privacy analysis, accepted-risk status, IAM binding coverage, unused grants, least-privilege suggestions, visibility gaps, remediation, policy snippets, runtime reconstruction, evidence collection guidance, evidence-manifest and remediation compare drift, and saved-report portfolio rollups.
  • HTML reports are self-contained and escape untrusted evidence strings.
  • Public scores are capped at 100; raw_points remains available for diagnostics.
  • Missing identity, permission, approval, data classification, owner, environment, or runtime evidence is reported as a visibility gap.

Frozen Contracts

The v1 freeze is recorded in ../schemas/v1/manifest.json. Product version is 1.0.0; evidence and report payloads keep schema_version: "0.1" until a future breaking evidence-format change.

The freeze covers:

  • evidence schemas in schemas/*.schema.json
  • report schema in schemas/findings.schema.json
  • evidence manifest fields
  • rule ids for current attack-path families
  • evidence quality and path state vocabulary
  • offline control analysis fields
  • policy analysis fields, policy rule-risk fields, and policy evaluation import flags
  • remediation plan fields
  • compare drift fields for evidence manifests and remediation plans
  • portfolio rollup fields for evidence manifests and remediation plans
  • privacy analysis fields and data-category filter labels
  • runtime event-quality fields and correlation diagnostic types
  • visibility gap priority vocabulary
  • review decision vocabulary
  • score and tier semantics
  • no-network, no-telemetry, no-enforcement boundaries

Verification Baseline

python -m compileall -q src tests
python -m unittest discover -s tests -v
coverage run --source=src/agentguard_graph -m unittest discover -s tests
coverage report -m --fail-under=85
PYTHONPATH=src python -m agentguard_graph.cli demo

Package build gate:

python -m pip wheel . --no-deps --no-build-isolation --wheel-dir dist

Non-Goals

  • live cloud, SaaS, repository, or identity crawling
  • MCP tool execution
  • OpenAPI operation execution
  • runtime policy enforcement
  • telemetry upload
  • exploitability or breach confirmation claims
  • treating absent runtime events as proof of safety

Release Operations

Use RELEASE_CHECKLIST.md before publishing a package or tag. Those checks are release hygiene, not open v1 product gaps.