feat(jafar-perf): add the jafar-perf plugin and attach the jfr-mcp plugins to a shared daemon

[jbachorik]

What & why

Moves the jafar-perf plugin here from btraceio/jafar-perf-box, so the btraceio agent plugins live in one marketplace instead of two with overlapping names, and points both jfr-mcp plugins at the shared daemon from btraceio/jafar#119.

A separate plugin, not folded into jfr-analyzer. The overlap turned out to be small. jafar-perf is task playbooks (triage, cpu, latency, gc, memory-leak, heap-diff, compare) plus specialist subagents and a perf-lead fan-out. jfr-analyzer is the USE/TSA method and the async-profiler/BTrace interop. They clash only on the agent name perf-engineer, which plugin namespacing already keeps apart, and on the MCP wiring below.

Commit	Change
fce91c1	add jafar-perf (both manifests, both catalogs, package.json), --attach for both jfr-mcp plugins, fix jfr-analyzer's .mcp.json, the Node drift check, a relaxed skill front-matter check
c160726	make the drift check resolve the published server with jbang --fresh
85fc3ad	scheduled tool-drift.yml workflow

• --attach. Both plugins now start the server as jbang jfr-mcp@btraceio --stdio --attach, so every session shares one daemon instead of starting a JVM each. An older jfr-mcp ignores the flag and runs a private server, so this degrades instead of breaking.
• jfr-analyzer's .mcp.json was broken. It pointed at an SSE daemon on localhost:3000 with no token, which the daemon rejects. It now uses the bridge, keeping the jfr-mcp server key so the agent's mcp__jfr-mcp__* tool names still match. Its README also named a jafar-mcp alias that does not exist.
• Drift check (scripts/check-tool-references.js, ported from jafar-perf-box in Node like the other validation scripts): starts the server, asks for tools/list, and fails naming every tool a skill or agent mentions that the server lacks, with file and line. It only reads plugins whose .mcp.json uses jfr-mcp, and ignores jfr_session/jfr_window/jfr_evidence, which are field names in the interop skill's evidence record, not tools.
• Relaxed validate-repository.js. Its front-matter regex only allowed name and description, so it rejected 7 jafar-perf skills that carry allowed-tools (which pre-approves their MCP tools). Stripping the field would have changed how the skills behave, so the check now allows further single-line fields. A skill with no description is still rejected.
• --fresh in the drift check. jbang caches the catalog; a stale alias started a server reporting 0.10.0 with 36 tools and the check failed on jfr_compare, drift that did not exist. The published 0.28.0 exposes 37 tools and all 35 referenced ones.

How verified

Command	Result
scripts/validate-marketplace.sh	passes: 5 plugins, 31 skills (only the existing interface warnings for jfr-analyzer and perf-engineer)
claude plugin validate plugins/jafar-perf	passed
node scripts/check-tool-references.js against the published server	OK: all 35 referenced tools exist (37 exposed)
same, with a planted jfr_bogus_tool in an agent	FAIL naming the tool and plugins/jafar-perf/agents/cpu-analyst.md:21, exit 1
a skill with no description under the relaxed check	still rejected
actionlint .github/workflows/tool-drift.yml	clean
GitHub Actions on this PR	Skills reference tools that exist (the new job, first run) pass in 21 s; both validate jobs pass
real Claude Code 2.1.288 (claude -p, strict MCP config) driving the built bridge	server connected, jfr_open/jfr_summary/jfr_close succeed, two concurrent sessions with the same alias both succeed (btraceio/jafar#119 has the details)

Coverage notes

The validators cover manifests, catalogs, skill front matter and the Pi skill paths. The drift check covers every tool name in the two jfr-mcp plugins' skills and agents against the live server. No eval case was added: the corpus covers btrace-observability only.

Not verified / known limitations

• The plugin installed through the marketplace with the real jbang jfr-mcp@btraceio --stdio --attach: the published jar has no --attach yet, so the Claude run above used the built jar with an equivalent scratch config.
• The workflow's issue-opening step, which fires only on a failed scheduled run. The rest of the job ran and passed on this PR.
• The branch is one commit behind main (docs: invoke JBang scripts directly, note jbang fallback #7); it merges cleanly and was not rebased.

Why this is a draft: it needs a jfr-mcp release that contains --attach to do what it says (until then it works, without sharing), and the issue-opening step is unexercised. It flips to ready once btraceio/jafar#119 is released.

Breaking changes / migration

• jfr-analyzer users no longer need a daemon already running on localhost:3000; the plugin starts the server itself through JBang, which must be on PATH. A hand-started daemon is attached to automatically if one is already running.
• jafar-perf-box users move with: /plugin marketplace remove btraceio, /plugin marketplace add btraceio/agent-plugins, /plugin install jafar-perf@btraceio-agent-plugins. A README pointer in that repository is prepared locally and will be pushed, and the repository archived, only after this merges.

Links

Depends on btraceio/jafar#119 (the --attach flag). Changelogs and plugin versions untouched.

🤖 Generated with Claude Code