Plugin marketplace: build out the DSPy.ts Claude Code / Codex plugins

[ruvnet]

Tracking the DSPy.ts plugin marketplace (see docs/adr/ADR-0001-claude-code-plugin-marketplace.md, branch feat/plugins-marketplace).

The repo publishes a Claude Code / Codex plugin marketplace (.claude-plugin/marketplace.json) with one plugin per directory under plugins/. Each plugin = .claude-plugin/plugin.json + commands/*.md + agents/*.md + skills/<name>/SKILL.md, and where useful a bundled MCP server (mcp/server.js, referenced from plugin.json's mcpServers) exposing DSPy.ts as MCP tools/resources (dspy_scaffold, dspy_compile, dspy_eval, dspy_retrieve; dspy://docs/api, dspy://examples, design guides) — Codex picks up the same MCP servers.

Build cadence: one plugin per iteration — scaffold → validate (plugin.json parses, frontmatter present, referenced files exist) → commit on feat/plugins-marketplace. Closes with a single PR.

Plugins — practical → exotic

• dspy-core — scaffold / compile / evaluate DSPy.ts programs; dspy-architect agent; signature-design + metric-design skills; MCP tools for the library (shipped)
• dspy-optimize — deep optimizer workflows: MIPROv2 + experience replay (replayStore), GEPA Pareto evolution, BootstrapFewShot dynamic demos; optimizer-engineer agent; optimizer-selection skill; MCP dspy_optimize/dspy_replay tools
• dspy-rag — RetrieveModule over AgentDB (HNSW + RaBitQ + MMR) → ChainOfThought in a Pipeline; corpus-indexing command; rag-architect agent; chunking/grounding skills; MCP dspy_index/dspy_retrieve tools
• dspy-react — ReAct agents with tool registries + ReActReflexion (recall lessons, record episodes, promote skills); react-engineer agent; tool-design + reflexion skills; MCP dspy_react_run tool
• dspy-observability — CompilationTracer causal-chain traces + AgentDB persistence + optional MLflow; CachingLM setup; observability-engineer agent; tracing skill; MCP dspy_trace/dspy_runs tools + dspy://traces resource
• dspy-evolution (exotic) — GEPA-driven self-evolution loops: evolve a program's prompts (and the program) against a benchmark across runs, persisting frontiers; evolution-coordinator agent; benchmark-design skill; MCP dspy_evolve tool

Vertical appliances — pre-wired end-to-end programs

• dspy-appliance-support-bot — RAG over a knowledge base → ChainOfThought answer → quality metric → MIPROv2 tuning; bring-your-own corpus
• dspy-appliance-code-review — Retrieve(repo context) → ChainOfThought(review) → actionability metric → GEPA-tuned
• dspy-appliance-research-assistant — ReAct(search tools) + reflexion → synthesize → cite; optimized for grounded answers
• dspy-appliance-data-pipeline — typed Predict modules in a Pipeline, BootstrapFewShot from a labeled CSV, evaluated batch

Done criteria

• Every plugin above checked, validated, committed
• .claude-plugin/marketplace.json status flipped planned → available for each
• PR feat/plugins-marketplace → main opened and merged
• README.md gains a short "Plugins" section pointing at the marketplace

[ruvnet]

✅ dspy-optimize landed on feat/plugins-marketplace (commit 3c4eaf2) — MIPROv2 replay / GEPA evolution / BootstrapFewShot dynamic-demo commands, optimizer-engineer agent, optimizer-selection + experience-replay skills, MCP server scaffold (dspy_mipro/dspy_gepa/dspy_bootstrap/dspy_replay_status), AGENTS.md. Marketplace: 2/10 available. Next up: dspy-rag.

[ruvnet]

✅ dspy-rag landed on feat/plugins-marketplace — /dspy-index (chunk+embed+store an AgentDB corpus: HNSW, RaBitQ, tiers), /dspy-rag-new (scaffold a Retrieve→ChainOfThought Pipeline + grounding metric), /dspy-retrieve (debug top-k + MMR), rag-architect agent, chunking-strategy + grounding-and-citations skills, MCP server (dspy_index/dspy_retrieve/dspy_rag_run/dspy_corpus_status), AGENTS.md. Marketplace: 3/10 available. Next: dspy-react.

[ruvnet]

✅ dspy-react landed on feat/plugins-marketplace — /dspy-react-new (scaffold a ReAct agent + tool registry [+ reflexion on AgentDB]), /dspy-react-run (run on a task, full thought/action/observation trace), /dspy-reflexion (lessons / promoted skills / episodes), react-engineer agent, tool-design + reflexion-loop skills, MCP server (dspy_react_new/dspy_react_run/dspy_reflexion_recall/dspy_reflexion_status), AGENTS.md. Marketplace: 4/10 available. Next: dspy-observability.

[ruvnet]

✅ dspy-observability landed on feat/plugins-marketplace — /dspy-trace (compile with a CompilationTracer → run summary + causal chain), /dspy-runs (list/inspect compile runs in an AgentDB trace store), /dspy-cache (wrap an LM with CachingLM + hit stats), observability-engineer agent, compile-tracing + llm-caching skills, MCP server (dspy_trace/dspy_runs/dspy_run_chain/dspy_cache_stats), AGENTS.md. Marketplace: 5/10 available. Next (exotic): dspy-evolution.

[ruvnet]

✅ dspy-evolution (exotic) landed on feat/plugins-marketplace — /dspy-evolve (N-generation GEPA evolution vs a held-out benchmark: reflect → mutate instructions → persistent AgentDB Pareto frontier with warm-start → optional structural exploration), /dspy-benchmark (score raw/optimized/evolved variants, record results), /dspy-evolution-status (learning curve / reflections / lineage), evolution-coordinator agent, benchmark-design + self-evolution-loop skills, MCP server (dspy_evolve/dspy_benchmark_run/dspy_evolution_status/dspy_frontier_inspect), AGENTS.md. Marketplace: 6/10 available. Next: the vertical appliances — dspy-appliance-support-bot.

[ruvnet]

✅ dspy-appliance-support-bot landed on feat/plugins-marketplace — vertical appliance: ships a ready-to-run program template (RetrieveModule MMR over an AgentDB KB → ChainOfThought answer with citations + a quality metric) plus /support-bot-init (scaffold + create KB), /support-bot-ask (retrieve + answer + cite), /support-bot-tune (MIPROv2 vs a Q/A set + the metric, replay + tracer), support-bot-builder agent, support-kb-curation + support-answer-quality skills, MCP server (support_bot_init/support_bot_ask/support_bot_tune/support_bot_status), AGENTS.md. Marketplace: 7/10 available. Next: dspy-appliance-code-review.

[ruvnet]

✅ dspy-appliance-code-review landed on feat/plugins-marketplace — vertical appliance: ships a ready-to-run program template (RetrieveModule pulls repo context: conventions/ADRs/related code/prior reviews → ChainOfThought structured review with severity-tagged located findings → an actionability metric) plus /code-review-init (scaffold + build context index), /code-review-run (review a diff/file), /code-review-tune (GEPA/MIPROv2 vs labelled reviews + the metric, frontier + cache), code-review-builder agent, review-context-indexing + review-actionability-metric skills, MCP server (code_review_init/code_review_run/code_review_tune/code_review_status), AGENTS.md. Marketplace: 8/10 available. Next: dspy-appliance-research-assistant.