Skip to content

ROADMAP.md

Latest commit

 

History

History
180 lines (152 loc) · 9.02 KB

ROADMAP.md

File metadata and controls

180 lines (152 loc) · 9.02 KB

Roadmap

Naming. This project is Agents Shipgate (display name) / agents-shipgate (package, CLI, repo). See AGENTS.md § Naming (canonical) for the full convention.

Agents Shipgate is preparing the v0.8.0 release. Releases v0.2 through v0.7 are complete and retained here as release history. Active public planning starts with the v0.8 release-decision close-out.

Completed

v0.2

  • Improved onboarding with agents-shipgate init, doctor, self-check, fixtures, and richer examples.
  • Stabilized JSON report compatibility, finding fingerprints, and agent-friendly JSON command output.
  • Added baseline save/apply workflow for strict CI adoption.
  • Hardened OpenAPI, MCP, plugin-loader, YAML resource-limit, and coverage test paths.
  • Added SBOM generation, release signing workflow, and dependency audit checks.
  • Expanded manifest-aware checks and severity overrides.

v0.3

  • Added Google ADK static adapter MVP for Tool-Use Readiness:
    • Supported google_adk as a manifest tool source.
    • Parsed ADK Agent Config YAML and statically extracted Python Agent / LlmAgent definitions, function tools, OpenAPIToolset, McpToolset, callbacks, plugins, sub-agents, eval references, and explicit local inventories.
    • Normalized discovered ADK tools into the existing Tool schema and reused MCP/OpenAPI loaders where possible.
    • Added ADK checks for dynamic or unresolved toolsets, unfiltered MCP toolsets, missing function-tool metadata, long-running tool contracts, guardrail evidence, and eval coverage.
  • Added SARIF output.
  • Added baseline diff mode for PRs.
  • Added optional trace normalization.
  • Added GitLab CI, CircleCI, and Jenkins examples.

v0.4

  • Added external policy/check packs:
    • Declarative YAML policy packs under checks.policy_packs.
    • CLI and GitHub Action policy-pack overrides for CI.
    • Policy-pack findings flow through suppressions, severity overrides, baselines, Markdown, JSON, and SARIF.
  • Hardened multi-framework adapter support:
    • Introduced a shared framework adapter interface where it reduces duplication.
    • Stabilized ADK report schema fields.
    • Documented explicit runtime inventory as a future trust-gated command; it is not part of default CI.
    • Kept TypeScript, Go, and Java ADK support as post-Python-MVP investigation items.
  • Split bundled OpenAI API operational readiness findings into atomic check IDs.
  • Removed the legacy top-level check_severity_overrides alias.
  • Revisited container image distribution and kept it deferred until there is an exercised build-and-test path.

v0.5.0 LangChain/CrewAI And Focused CI

  • Added static Python coverage for LangChain/LangGraph and CrewAI while preserving the default static trust model.
  • Promoted GitLab CI and CircleCI from examples to first-class integration recipes:
    • documented strict/advisory gating;
    • baseline artifact handling;
    • SARIF or native security report guidance where supported;
    • copy-pasteable workflows for monorepos and multi-manifest scans.
  • Added a framework adapter checklist so new platform support is consistent:
    • static extraction only by default;
    • no agent execution, model call, tool call, network call, or MCP connection;
    • deterministic tool inventory normalization;
    • source warnings for dynamic or unresolved tools;
    • framework surface summary in JSON, Markdown, and SARIF-compatible metadata.
  • Bumped the additive report schema to report_schema_version: "0.5" while keeping manifest version: "0.1".

v0.6.0 Agent-Friendly Adoption

Goal: compress the 5-step setup (install → init → edit YAML → scan → read findings → wire CI) into a single tool-using turn for AI coding agents.

  • Added agents-shipgate detect for read-only workspace classification (which framework, which agent-name candidates, which suggested sources).
  • Made agents-shipgate init auto-detect by default; framework-specific manifests are produced for LangChain, CrewAI, Google ADK, OpenAI Agents SDK, Anthropic, and OpenAI API. --minimal preserves the pre-v0.6 template byte-exact.
  • Added agents-shipgate init --ci for opt-in workflow generation, with cross-workflow shipgate detection.
  • Added agents-shipgate scan --suggest-patches and agents-shipgate apply-patches for machine-applicable manifest fixes (stale-manifest removals at high confidence; scope-coverage appends at medium confidence opt-in).
  • Bumped report schema to v0.6 (additive: per-finding patches, top-level manifest_dir).

v0.7.0 Adoption Activation

Goal: make the v0.6 features visible to humans and AI coding agents on real repos, plus expose per-check remediation metadata so agents can decide what's safe to auto-fix.

  • Added AGENTS.md "Should I run Shipgate?" trigger table with the soft-stop rule (don't skip MCP/OpenAPI-only repos).
  • Added docs/agent-recipes.md, docs/autofix-policy.md, and docs/minimal-real-configs.md as the agent-facing surface.
  • Extended CheckMetadata with autofix_safe, requires_human_review, suggested_patch_kind, populated docs_url for all 45 entries (PR 2 — catalog-conservative defaults; per-check generator targets documented).
  • Added the same four optional fields to Finding, plus the derivation policy: when Finding.patches is non-empty, derive from actual patches (autofix_safe=True only when EVERY patch is non-manual AND high confidence); otherwise seed from CheckMetadata. Unknown check IDs (policy packs, plugins) get the safe-closed fallback. Three patch states (None / [] / non-empty) handled distinctly.
  • Bumped report schema to v0.7 (additive over v0.6 — four new optional Finding fields). v0.6 schema retained as a frozen reference.
  • _run_id excludes the four new derived fields plus patches; scan run_id is identical with or without --suggest-patches.
  • Plugin-loading isolation: every code path that reads the catalog during scan honors the scan's plugins_enabled setting, so --no-plugins is respected even when AGENTS_SHIPGATE_ENABLE_PLUGINS=1 is set in the environment.
  • Onboarding prompt rewritten to lead with the v0.6 4-call flow; byte-parity test enforces dual-copy synchrony between prompts/ and skills/agents-shipgate/prompts/.

Open

v0.8.0 Release Decision Engine

Goal: close v0.8 around a baseline-aware release decision over agent tool surfaces, not broader framework or runtime surface area.

  • Make release_decision.decision the recommended release-gating signal for JSON consumers, CLI output, Markdown, PR comments, and GitHub Action outputs.
  • Preserve summary.status as baseline-blind compatibility for v0.7 consumers.
  • Keep the verdict path deterministic and static: no LLM calls, agent execution, MCP connections, network access, runtime inventory, telemetry, or new framework adapters in the v0.8 close-out.
  • Harden real-repo detection so local/private state, virtualenv fixture installs, generated reports, and gitignored worktrees do not create false framework or source signals.
  • Keep Tool-Use Readiness as the public wedge: OpenAI Agents SDK, MCP exports, and OpenAPI specs are the clearest first path; other supported frameworks remain documented inputs.

v0.8.x Source-Provenance Enrichment (incremental)

Once we have origin (file path, line index for JSONL, list index for arrays) threaded through finding evidence, expand the patch generator catalog beyond manifest-only. Candidate generators:

  • SHIP-API-RETRY-POLICY-MISSING and SHIP-API-TIMEOUT-MISSING targeting policy-rule files (likely with a new create_file patch kind).
  • Trace-event metadata enrichments — but never approval/confirmation flips, which stay manual permanently.

Later Candidates

  • Expand agent-platform coverage beyond the v0.5 framework adapters:
    • AutoGen multi-agent tool surfaces;
    • Semantic Kernel plugins/functions;
    • LlamaIndex tools and workflows;
    • TypeScript/JavaScript agent frameworks where static extraction is practical;
    • additional Google ADK language surfaces after the Python adapter remains stable.
  • Optional trust-gated runtime inventory export as an explicit command, separate from default static CI.
  • Cross-platform CI expansion after v0.8 stabilizes:
    • Jenkins;
    • Buildkite;
    • Azure Pipelines;
    • Bitbucket Pipelines;
    • local pre-commit / pre-push usage;
    • generic POSIX shell integration for unsupported CI systems.
  • Container image distribution if the image has CI coverage, security scanning, and release signing.
  • Homebrew or other package-manager distribution if CLI usage warrants it.
  • Public versions of the private release-readiness research themes once they are shaped into concrete checks and schema changes.

Google ADK Support Principles

ADK support is read-only by default: local file parsing only; no adk run, adk web, adk eval, MCP connection, tool call, model call, or network call. ADK callbacks and plugins are static guardrail evidence only, not proof of runtime enforcement. Dynamic toolsets must produce warnings or findings unless the user provides explicit MCP, OpenAPI, or tool inventory inputs.