[JaweriaAsif]

Summary

Part of REL-1133 (plugin marketplace discovery). This rewrites the root README as a landing page for people who find the plugin through a marketplace, a directory, or an AI assistant's web search:

• The problem RELAI solves in the first two lines, using the same one-liner as the plugin manifests in relai-cli.
• Install commands for each host. The Claude Code and Codex commands were run against this repo.
• A first prompt to try, a "What you get" section, benchmark results, and an FAQ phrased as real questions.
• UTM parameters (utm_source=github&utm_medium=plugin_readme&utm_campaign=plugin_discovery) on every relai.ai, docs, and platform link. The signup link also carries the site's own source=plugin_readme parameter.
• The maintainer notes on generated packages and provenance move to CONTRIBUTING.md. A short "Repository layout" table stays in the README.

No generated paths are touched.

Behavior changes

None. This is a documentation-only change; the generated packages, catalogs, and workflows are unchanged.

Verification

• python3 scripts/validate_release.py --root . passes, and python3 -m unittest discover -s tests passes (4 tests).
• Claude Code 2.1.170, with an empty CLAUDE_CONFIG_DIR: claude plugin marketplace add relai-ai/coding-agent-plugins registered relai-claude, and claude plugin install relai@relai-claude installed relai 0.2.10. The README shows the equivalent /plugin commands.
• Codex CLI 0.155.1, with an empty CODEX_HOME and HOME: codex plugin marketplace add relai-ai/coding-agent-plugins registered relai-codex, and codex plugin add relai@relai-codex installed relai 0.2.10.
• Benchmark numbers and FAQ answers match relai.ai as of 2026-10-23: Terminal-Bench 2.0 62.5% → 79.2%; AutomationBench +3–10 pass-rate points and 22–44% lower cost per successful task.
• All 8 links in the README return 200.

Still to verify or decide

• Public install path. These tests cloned the repo with my GitHub credentials, because the repo is still internal. Re-run both installs anonymously once REL-1000 makes it public.
• Cursor and Copilot. The README sends these users to relai setup --agent cursor|copilot and mentions Cursor's team marketplace import, which I took from Cursor's docs but did not test. I did not document gh skill install, because the repo has several copies of each skill (one per host package).
• Marketplace name collision. relai setup --agent claude|codex registers marketplaces with the same names (relai-claude, relai-codex), so the README tells marketplace users not to also run relai setup --agent. In my Codex test, an existing personal relai-codex marketplace shadowed the GitHub one until I isolated HOME.
• Missing CLI. The published v0.2.10 skills do not yet install a missing CLI themselves (relai-ai/relai-cli#722 is not released), so the README has users install the CLI explicitly.
• Latency. The README says the optimizer works toward quality and cost, not latency, because the CLI docs describe only a quality/cost objective, although relai.ai mentions latency.

🤖 Generated with Claude Code

[linear-code]

REL-1133 Plugin marketplace discovery: relevance suggestions, AEO-ready listings, and attribution

Context

REL-1000, REL-1001 and REL-999 cover getting listed (Claude, OpenAI, Cursor). This issue covers making those listings actually drive discovery and adoption.

Key finding: AI agents don't proactively recommend uninstalled marketplace plugins to arbitrary users.

• Claude Code suggests plugins based on session signals (cwd, files read, commands run), but only from marketplaces an org admin allowlists in managed settings (pluginSuggestionMarketplaces). It never auto-installs and is rate-limited. Docs: https://code.claude.com/docs/en/plugin-relevance
• Codex/ChatGPT Plugin Directory is a browse/search surface; no documented proactive suggestion of uninstalled plugins.
• The real AEO upside: listings, repo READMEs and directory pages are public and crawlable, so assistants using web search can cite them.

Scope

0. CLI install hints (needs decision)

• Claude Code's hint protocol lets relai-cli print a one-line stderr tag when CLAUDECODE is set, triggering a one-time install prompt for any interactive Claude Code user, no admin allowlist needed. Docs: https://code.claude.com/docs/en/plugin-hints
• Only works for plugins in the official claude-plugins-official marketplace (Anthropic's discretion, via partner contact), which REL-1000 lists as out of scope. Community marketplace doesn't qualify.
• Decide whether to pursue an official listing; hint code can ship early (harmless until listed)

1. Relevance block for in-account expansion (Claude Code)

• Add a relevance block to the RELAI entry in our marketplace.json (signals: e.g. Harbor task dirs, EvalsWiki files, relai CLI commands, agent framework files)
• Validate with claude plugin validate
• Write a short admin guide for customers on allowlisting the RELAI marketplace so their engineers get suggestions

2. Listing copy aligned with the current landing page

Positioning: continual learning / eval engineering + harness optimization for agents.

• One consistent one-line description across all listings and the plugin repo
• Lead with concrete, verifiable claims: Harbor-format evals, Terminal-Bench 2.0 and AutomationBench results, works with existing observability (Arize, Braintrust, LangSmith, Langfuse, Helicone)
• Plugin repo README as a landing page: problem in first 2 lines, one-command install, 30-second example, FAQ phrased as real questions
• GitHub topics (e.g. agent-evaluation, harbor, claude-code-plugin, codex-plugin)
• Keep skill descriptions narrow and accurate (no over-broad triggers)

3. Target query list

• Define 10–15 questions target users ask assistants (e.g. "how to build evals for my AI agent", "Terminal-Bench / Harbor compatible eval tools", "optimize agent prompts without regressions")
• Map each to a listing, README section or docs page that answers it directly

4. Aggregator presence

• Submit to Anthropic community marketplace, awesome-codex-plugins / HOL registry, and plugin indexes

5. Attribution (PostHog)

• UTM params on all links from plugin, listings and README to relai.ai / platform signup
• Add "AI assistant" and "plugin marketplace" options to "how did you hear about us"
• Track activation of plugin-originated signups separately
• Monthly manual check: ask target queries to major assistants with web search, record whether RELAI is cited

Done when

• Relevance block shipped and validated, admin guide published
• Listing copy and README updated and consistent across all marketplaces
• Target query list documented
• Attribution live in PostHog

Review in Linear