fix(mcp): make BTrace MCP deployment work and test it end to end

[jbachorik]

What & why

Running the BTrace MCP server against a real JVM showed that most operations were broken. The unit tests had missed this because they replace the BTrace client with fakes. This PR fixes the MCP-side defects, adds end-to-end tests that run the server against a live sample JVM (no LLM involved), runs those tests in CI against a BTrace build, and corrects the skill guidance.

Server fixes (5cba3ee)

• Attach failed with No masked btrace.jar found. The BTrace client finds its agent through btrace.jar.path or a classpath entry named btrace.jar, but JBang supplies btrace-<version>.jar. The server now sets btrace.jar.path from the jar it has already located.
• A successful deploy never returned. The client's submit runs the probe command loop until the probe exits. Deploys now submit on a daemon thread, stop buffering output after the status reply, and close the client when deployment times out or fails.
• deploy_script rejected public classes. It named the source BTraceScript_<timestamp>.java, so public classes failed to compile and every other script fell back to BTrace short syntax, which allows no fields and no @Export. The file is now named after the first declared class.

Codex (212707d): the Codex manifest declared only skills, so Codex installs had no MCP server. Codex's plugin MCP format does not expand a plugin-root placeholder in args (checked in codex-rs plugin_config.rs), but it resolves a relative cwd against the plugin root, so the server is launched with cwd: ".". The file sits under .codex-plugin/ so Claude Code does not also load it as a root .mcp.json.

End-to-end tests (0037366, d0802cb): the new ./gradlew e2eTest drives every tool and prompt against a sample app.

• No tool returns probe output after deployment, so the test checks probe effects through @Export counters read with jcmd: instrumentation fires, send_event runs its handler, a detached probe keeps running, and exit_probe stops it.
• By default the server is launched through JBang, as the manifests do. With BTRACE_JAR set, it runs on a plain classpath against that jar.
• The sample app lives outside io.btrace, which BTrace does not instrument, and the agent uses a free port.

CI (cb5b89b): the new MCP e2e workflow builds btraceio/btrace (default branch, or a ref given on manual dispatch) with :btrace-dist:btraceJar and runs e2eTest against the resulting jar. It never depends on the unpublished io.btrace:btrace:3.0.0-SNAPSHOT artifact.

Skills (9cff5dc): agents are told not to call list_probes (see limitations), and the skill no longer promises output inspection.

Rejected alternatives:

• Codex $schema Agent Plugins MCP format: it does support ${PLUGIN_ROOT}, but it isn't needed, and the cwd form is verified.
• Linking the e2e job to a published BTrace artifact: none exists. Building from source also tests against current BTrace.
• Running the server's BTrace client in a separate process to avoid System.exit: deferred. The fix belongs in the BTrace client.

How verified

Gate	Result
./gradlew clean check (mcp)	BUILD SUCCESSFUL, 17 unit tests passed, 80% line-coverage rule passed
./gradlew e2eTest (JBang launch)	9 passed, 1 skipped (list_probes)
BTRACE_JAR=<fresh btrace develop ee8434e build> ./gradlew e2eTest	9 passed, 1 skipped, no JBang in server log
e2e with the jar-path fix disabled	5 failures (the suite catches the defect)
BTRACE_JAR=<directory> / blank BTRACE_JAR	fails with BTrace JAR not found / falls back to JBang
codex mcp get btrace --json (isolated CODEX_HOME, plugin installed from this repo)	cwd resolves to the installed plugin root; that command answers initialize and lists 7 tools
MCP e2e workflow on GitHub (first run, this PR)	BTrace build, jar lookup and e2eTest steps all succeeded
actionlint .github/workflows/mcp-e2e.yml	no findings
Workflow steps reproduced locally (fresh clone, :btrace-dist:btraceJar, jar lookup, e2eTest)	all pass
scripts/validate-marketplace.sh on this branch merged into origin/main	Repository consistency valid: 4 plugins, 22 skills, Marketplace validation passed.

Coverage notes

The e2e suite covers all 7 tools and 3 prompts on one sample JVM, including the session lifecycle: deploy, event, detach and exit. It does not cover a failed probe start, because the BTrace client still calls System.exit(1) there. It does not cover agents choosing these tools; LLM evals stay manual and are not part of CI.

Not verified / known limitations

• Codex has not launched the server inside a real (logged-in) session. Only the resolved launch configuration was checked.
• list_probes still terminates the server. Client.connectAndListProbes in BTrace calls System.exit(0), and submit calls System.exit(1) on failure. The fix belongs in btraceio/btrace; the e2e case is @Disabled and the skill warns against the tool until then.
• The e2e sample app runs on Java 11, which BTrace now warns is deprecated for targets.
• Workflow action pinning differs from validate.yml. The new workflow pins actions to commit SHAs (the same ones BTrace's CI uses), while validate.yml uses major tags.

Breaking changes / migration

None. deploy_script now uses the script's class name as the source file name; scripts that relied on short syntax still work.

Links

Depends on #5 (base of the stack). Retarget to main after #5 merges.

🤖 Generated with Claude Code