Skip to content

Changelog

All twelve plugins and the marketplace share a version. Contracts are versioned separately, by filename.

First public release.

Twelve plugins under one marketplace (fedcal/foundry), each declaring dependencies: [foundry-core] except the kernel itself. foundry-core is the only plugin with defaultEnabled: true.

Plugin Agents Skills
foundry-core 4 7
foundry-research 5 5
foundry-ai 4 4
foundry-data 4 4
foundry-dev 19 17
foundry-quality 6 5
foundry-ops 6 5
foundry-pmo 6 6
foundry-economics 5 5
foundry-legal 5 4
foundry-growth 6 6
foundry-oss 4 5
  • Governed memory in four tiers: session scratch, atomic facts, runbooks and ADRs. Only .foundry/memory/INDEX.md is loaded by default, hard-capped at 4000 tokens; everything else is retrieved on demand.
  • Eleven contracts as JSON Schema 2020-12: fact.v1, finding.v1, review.v1, adr.v1, plan.v1, requirement.v1, risk.v1, estimate.v1, compliance-check.v1, handoff.v1, tracker-item.v1.
  • MCP server (foundry): nine tools — memory_search, memory_write, memory_index, runbook_list, runbook_get, contract_validate, blackboard_write, blackboard_read, token_report — plus a memory-index resource, a contracts resource and one resource per runbook.
  • Nine hook entries across eight events: session state injection, targeted prompt-time recall, a Bash guard with eight named rules, a write guard for credentials and protected paths, blackboard contract validation, the subagent context firewall, the verify-before-claiming gate, pre-compaction persistence, worktree preparation and session-end metrics.
  • CLI: foundry init, doctor, memory index|search|prune, tokens, runbooks, validate, profile.
  • Three dynamic workflows: foundry-feature-delivery, foundry-audit-sweep, foundry-compliance-sweep.
  • Two playbooks: feature-delivery.yaml, audit.yaml.
  • Three output styles: Foundry Senior Engineer, Foundry Analyst, Foundry PMO.

Five profiles that set plugins, permissions and enforcement level together: angular-spring-enterprise, oss-library, pa-italia, startup-mvp, full.

Five packs, 147 controls in total: global-baseline (40), eu (39), it (16), north-america (26), uk-apac-latam (26). Every pack ships with lastReviewed: null and an empty sources array; citations are unverified and the engine says so in every report.

Thirteen OSS templates in foundry-oss, an authoring contract enforced in CI by scripts/validate-assets.mjs, kernel unit tests, and a bilingual EN/IT documentation site.

  • Jurisdiction packs are unverified against official texts.
  • memory_search is keyword scoring, not semantic search.
  • Token figures are estimated at roughly four characters per token, not tokenizer output.
  • Pipeline scaffolding targets GitHub Actions only; infrastructure code targets Terraform and OpenTofu only; repository operations require an authenticated gh.
  • The database agents assume PostgreSQL.

Checked against the Claude Code 2.1.250 binary on 2026-08-28, rather than against bundled documentation. Three consequences are baked into this release:

  • PreToolUse accepts only allow, deny, ask and defer. escalate is rejected by the schema and fails open, so the protected-path gate asks instead.
  • Stop and SubagentStop block with a top-level {"decision":"block","reason":…}, not with permissionDecision.
  • Marketplace source paths resolve against the marketplace root, so every entry carries its full path and metadata.pluginRoot is not used.

The blocking gates stay silent until a project runs foundry init, so installing the kernel does not arm them on every project on the machine.

Claude Code 2.1.x or later; Node.js 20 or later. superpowers optional, with graceful degradation. Licensed Apache-2.0.