Changelog
All notable changes to this project will be documented in this file.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
[Unreleased]
Added
- Context Engineering foundation (
foundations/context-engineering.md). Promotes "context-as-finite-resource" from a memory-doc subtopic to a first-class architectural framing: the four levers (select/compress/prune/persist), memory hierarchy, per-pattern context shape, context-window awareness, compaction, and budgets. Cross-linked fromfoundations/README.mdandprimitives/memory/overview.md. - Guardrails modifier (
modifiers/guardrails/). The second modifier (alongsidehuman_in_the_loop). Layered input / tool / output policy checks plus an explicit dual-LLM split (privileged actor + quarantined reader) that breaks the indirect-prompt-injection path. Ships the six tier docs (overview / design / implementation / evolution / observability / cost-and-latency),schemas/state.py, and two typed prompts (quarantined-summarizer.md,policy-rewriter.md).appliesTo: [any]. - Sub-agents primitive (
primitives/sub_agents/). The fourth primitive (alongsidetool_use,memory,skills). Models named, role-scoped agent instances with their own context windows, tool grants, and (optionally) models. The 2026 production default formulti_agentworkers and the unit-of-work insideplan_and_executeandreact. Ships six tier docs,schemas/state.py(SubAgentSpec,ContextEnvelope,SubAgentInvocation,SubAgentResult,SubAgentsState), and two typed prompts (delegator.md,sub-agent-base.md). - Long-Horizon agent pattern (
patterns/long_horizon/). Multi-session tasks that span hours-to-weeks; checkpoint-and-resume across crashes, deploys, and external waits. Distinct fromsaga(compensation) andmemory(storage). The deep-agents shape (planner + virtual filesystem + sub-agents) is the canonical composition. Ships six tier docs,schemas/state.py(LongHorizonState,Plan,StepRecord,EventLogEntry,Checkpoint), and theplanner.mdprompt. Evolves fromsaga+event_driven. - Agentic RAG agent pattern (
patterns/agentic_rag/). RAG where the agent plans retrievals, decomposes queries, routes across sources, reflects on sufficiency, enforces citation-bound answers, and cross-checks across sources to defend against single-source RAG poisoning. Ships six tier docs,schemas/state.py(SubQuestion,RetrievalAttempt,EvidenceChunk,CrossSourceConflict,Citation,VerificationResult,AgenticRagState), and two typed prompts (decomposer.md,sufficiency-reflector.md). Evolves fromrag+plan_and_execute.
Changed
patterns-catalog.yamlregenerated (now 14 Patterns + 4 Primitives + 2 Modifiers; derived workflows view unchanged at 4).website/src/data/patterns.tsregenerated to register the new entries.- README.md,
patterns/README.md,primitives/README.md,modifiers/README.md, andfoundations/choosing-a-pattern.mdauto-blocks regenerated to include the new entries.
[0.2.330] - 2026-06-09
First tagged release. Bundles two work streams: the three-tier taxonomy refactor (12 patterns + 3 primitives + 1 modifier under a single taxonomy.yaml source of truth, with byte-identical drift-checked generators) and an AI-tool-discoverability hygiene round (llms.txt, agents.md, root-level CHANGELOG/ROADMAP, action version unification, Conventional Commits convention).
Added
- Three-tier taxonomy (
patterns/,primitives/,modifiers/) backed by a newtaxonomy.yamlat the repo root. Adding a new cohort (e.g.guardrails/) is now one entry intaxonomy.yaml+ a directory — no code change. The validator, catalog emitter, schemas test, docs generator, and website data generator all read taxonomy.yaml. Catalogschema_versionbumped to 2;workflows[]becomes a derived view ofpatterns[]for backward compat. - Generic taxonomy infrastructure: new
meta/generate-docs.js(replacesAUTO:marker blocks across markdown), newmeta/generate-website-data.js(regenerateswebsite/src/data/patterns.tsfrom the catalog), refactoredmeta/validate-metadata.jsto be taxonomy-driven. Drift gates incatalog-drift.ymlnow cover catalog + docs + website TS together. - Contributor manual
meta/HOW_TO_ADD_AN_ENTRY.md— quick-start commands, per-cohort recipes, prompts for Claude Code / Cursor / Copilot, post-AI verification checklist.CONTRIBUTING.md+meta/contributing.mdrefreshed to lead with the three-step flow. - Per-cohort READMEs at
primitives/README.mdandmodifiers/README.md. - AI-tool discovery files at repo root:
llms.txt(llmstxt.org spec) +agents.md(consumer-focused guide for AI tools reading the catalog). - CHANGELOG.md and ROADMAP.md moved to repo root (from
meta/). Stubs remain at the old paths. patterns-catalog.yaml— top-level machine-readable index aggregating every pattern + primitive + modifier + composition edge into one file. Generated bynode meta/validate-metadata.js --emit patterns-catalog.yaml. Primary downstream consumer is theagent-deploymentsCI generator. See `PATTERNS_CATALOG_SCHEMA.md`.meta/validate-metadata.js --emit <path>flag — aggregates validated metadata into the catalog. Without the flag, validator behavior is unchanged..github/workflows/catalog-drift.yml— hard CI gate that regenerates the catalog on every PR and fails if the committed file diverges.composition/anti-compositions.md— standalone reference for pattern pairs that fight, overlap, or leak state, with concrete what-to-use-instead guidance.- Two new reference architectures in
composition/reference-architectures.md: high-stakes content moderation pipeline (Routing + Tool Use + Reflection + HITL) and event-driven ingestion + RAG enrichment (Event-Driven + Tool Use + RAG + Memory). meta/contributing.md— "Good first contributions" section listing newcomer-friendly tasks.- Sitemap (
/sitemap-index.xml) and RSS feed (/changelog.rss) on the documentation website. - Refreshed
.github/ISSUE_TEMPLATE/blueprint-proposal.yml— current "pattern" terminology and added when-to-use / when-NOT-to-use fields.
[0.3.0] - 2026-05-31
Six-PR improvement series. Closes the three-repo connectivity gap, fills the security/production-realities content layer, and evens out pattern depth.
Added
agent-scaffoldconnectivity.composition/blueprints-to-deployments.mdandcomposition/blueprint-to-spec-to-scaffold.md— reverse-lookup table and end-to-end walkthrough makingagent-deploymentsandagent-scaffoldreachable from this repo. README lifecycle diagram replaces the static three-repo block.- Frameworks & MCP map.
foundations/frameworks-and-integrations.md— 15 patterns × 6 frameworks/MCP. New "MCP and tool registries" subsection inpatterns/tool_use/design.md. - Security & production foundations (4 new docs):
foundations/security-and-safety.md,foundations/hallucination-and-grounding.md,foundations/evals-and-quality.md,foundations/cost-and-model-selection.md. - Per-pattern "Production concerns" subsection added to all 15 workflow + pattern design docs. Each names a pattern-specific surface for prompt injection, hallucination, cost, rate limiting/retries, idempotency, observability.
- "When NOT to use" subsection added to all 15 overview.md files with 3 concrete anti-cases each, seeded from
foundations/choosing-a-pattern.md. - "Next steps" footer added to all 15 overview.md files pointing at the deployment mapping and scaffold walkthrough.
- CI rigor.
.github/workflows/docs.ymlnow does dynamic pattern validation and runsnpm run buildagainst the website on every PR.validate-metadata.jscross-checkspatterns.tsregistration..github/CODEOWNERSand.github/dependabot.ymladded.
Changed
patterns/multi_agent/design.mdexpanded from 105 → 210 lines and pinned as the canonical exemplar inmeta/contributing.mdandmeta/style-guide.md. Now the deepest pattern doc, exceedingsaga/design.md.- 8 thin design docs (memory, reflection, routing, plan-and-execute, react, tool-use, rag, parallel-calls) expanded to consistent ~120–165 line floor with explicit Observability Hooks subsections.
website/src/data/patterns.ts— registeredsagaandhuman-in-the-loop(were missing).SECURITY.md— adds adopter-facing security guidance pointing to the new foundations doc.
Fixed
docs.ymlshell loop now covers all 11 agent patterns (previously hardcoded to 8;saga,human-in-the-loop,event-drivenwere uncovered).
[0.2.0] - 2026-04 (approximate)
Pattern additions and website maturity.
Added
patterns/saga/— long-running multi-step processes with compensation.patterns/event_driven/— agents triggered by queue or stream events; idempotency, DLQ, replay.patterns/human_in_the_loop/— agent proposes, human approves, modifies, or denies before commit.- Reference Python implementations under
patterns/*/code/python/andworkflows/*/code/python/. - Astro 4.x documentation website at
website/. Interactive React Flow canvases (HeroViz, PlaygroundCanvas, EvolutionExplorer), Mermaid rendering, Pagefind full-text search. Deploys to GitHub Pages on push tomain. - Per-pattern
observability.mdandcost-and-latency.mdfiles for all patterns.
[0.1.0] - 2026-03-25
Phase 1: Documentation-First Restructure.
Added
/foundations/— Core concepts, terminology, anatomy of an agent, pattern selection guide with decision flowchart./workflows/— 4 workflow patterns (prompt chaining, parallel calls, orchestrator-worker, evaluator-optimizer) with 3-tier documentation./patterns/— 8 agent patterns (ReAct, plan & execute, tool use, memory, RAG, reflection, routing, multi-agent) with 3-tier documentation.- Evolution bridges (
evolution.md) for each agent pattern showing the workflow-to-agent transition. /composition/— Combination matrix and reference architectures./meta/— Contributing guide, style guide, roadmap, changelog.
Changed
- Repository philosophy: shifted from code-first blueprints to architecture-first documentation.
README.md: rewritten to reflect documentation-first approach./patterns/directory: replaced pattern reference docs with full 3-tier pattern documentation.
Removed
- Previous code implementations moved to
legacy/(blueprints, CLI, website, scripts). - Original CI/CD workflows archived; replaced with documentation-focused validation.