[jbachorik]

What & why

Fixes B5–B12 from the verification of https://btrace.io/agent-plugins/, one commit each:

• B5: jfr-analyzer named a JBang alias that doesn't exist; it now says jfr-mcp@btraceio.
• B6: the failure skill printed exception messages, against the data-safety promise. The default probe now prints the type only. It also warns that jstack() and the oneliner stack action omit the probed method, which made one live run name the wrong method.
• B7: list_probes called detached probes "Active" and hid attached ones. main's fix(mcp): list the MCP server's own probe and close the list_probes client #19 has since rewritten it to list this server's connected probe too, so the merge takes that version; this PR keeps only the skill's confirm-the-target rule around it.
• B8: clearer skill descriptions, a confirm-the-target rule, and 2 new eval cases.
• B9: perf-engineer now checks ASYNC_PROFILER_HOME and prefers the supervisor.
• B10: cleanup output now names what it removed and kept. It still deletes the recording, because docs/perf-engineer-design.md specifies that; the alternative was keeping it, which is a design change.
• B11: removed a skill from the README composition table that has moved plugins.
• B12: removed the unknown interface field from 2 manifests.

Also included:

• A fix for a racy test, detachReturnsTheUnreadOutput, and regen-formats.sh, which generates pprof and OTLP sample data.
• mcp-e2e.yml now builds a pinned BTrace commit (802d169, develop with the v2 heartbeat from feat(core): detect a silently dropped client-agent connection with a v2 heartbeat btrace#998, which main's silent-drop e2e test needs) on PRs and pushes, so upstream changes can't fail unrelated PRs. A new daily run builds develop, and btrace-ref overrides both. The clean JBang step still uses the latest published snapshot on purpose, because that's what users get.

How verified

Gate	Command	Result
MCP tests	./gradlew clean test e2eTest at 622bd4f (merge of main), BTrace 802d169	BUILD SUCCESSFUL; unit 56 passed, e2e 16 passed; against 79062f8 the silent-drop test fails
E2e repeat	e2eTest 6 times, BTrace snapshot 3.0.0-20261101.163203-247	6 of 6 passed
CI, pinned	PR run at 91a5ef2, the earlier pin	success; log says from BTrace 79062f8
CI, override	btrace-ref: develop run at 91a5ef2	success; develop was also 79062f8 at the time
Workflow lint	actionlint -shellcheck= -pyflakes= mcp-e2e.yml	exit 0
Supervisor	PerfEngineerSessionTest.java	4 passed, 0 failed
Repository	validate-repository.js	5 plugins, 31 skills
Evals	validate-evals.js	Eval corpus valid: 15 cases
Tool references	check-tool-references.js	all 37 referenced tools exist
Contributor gate	validate-marketplace.sh	Marketplace validation passed.
B9 live (Sonnet)	async-profiler 4.6, no asprof on PATH	3 of 3 recorded through the supervisor
B6 live (Sonnet)	failure prompt against a test JVM, final skill text	5 of 5 used the Java probe and named the right method and type; 0 order IDs captured
Analyzer flows	triage, drilldown, report on generated .pb.gz and .otlp	report.json written for both
B13	clean JBang install, snapshot 3.0.0-20261101.163203-247	no blank lines in probe output
Uncompressed pprof (Sonnet)	triage, drilldown, report on .pprof, jar built from jafar main (04e21df, includes btraceio/jafar#128)	report.json written; finding is Math.random() contention, about 92% of CPU; a pre-fix jar failed with Not in GZIP format
New eval cases (Sonnet)	score-evals.js on live responses, read-only tools	missing-target-asks-first 3 of 3 PASS; bastion-safe-attach 4 of 5 PASS on fresh runs after fc1938d

Coverage notes

Unit tests cover B10 (B7's code is now main's), and the validators cover the manifests and references. B6, B8 and B9 are skill text, so the evidence is small live samples, not tests. The new eval cases were scored on live responses with scripts/score-evals.js (missing-target-asks-first 3, bastion-safe-attach 5 after its rewording); deliberately wrong answers fail both. The other 13 of the 15 cases were not run.

Not verified / known limitations

• bastion-safe-attach first failed 3 of 3 on its literal terms, not its advice, so fc1938d changed them to "same OS user" and "detach". The new terms were chosen after reading those answers, so they were scored on 5 fresh runs: 4 passed, and 1 failed because it said to remove the probe without saying "detach".
• requiredSkills isn't met in those runs, and the scorer doesn't check it: each run loaded one skill (btrace-observability or btrace-runtime-access).
• The uncompressed .pprof fix isn't in a jafar release yet (latest v0.29.0), so released servers still reject these files. It passed only against a jar built from jafar main.
• The analyzer flows ran headless and resumed between steps, which restarts the MCP server, so each lost its session id and reopened the file. Interactive sessions weren't tested.
• jbang --fresh jfr-mcp-latest@btraceio fails with no main class deduced. The cause isn't established, and a stale local snapshot is possible. It's why the jar above was built by hand.
• btrace-data-safety still didn't load in the B8 samples, so that part of B8 isn't solved.
• The daily schedule trigger hasn't run, because GitHub only fires schedules from the default branch. It can't run until this merges.
• CI hasn't run on the new pin 802d169 yet; it passed locally.
• One earlier CI failure is unexplained: stopDetachedProbeRemovesTheListedProbe on Oct 5, with no failure message available. It didn't recur in the 6 local runs.
• The clean JBang step still floats, so a bad published snapshot can fail a PR there. It only checks that the server starts and lists its tools.
• The jstack() frame bug itself is in BTrace; this PR only warns about it.

Breaking changes / migration

cleanup output gains two fields; it removes nothing. The list_probes text change comes from main's #19.

Links

Follows #13.

🤖 Generated with Claude Code