feat(coding-agent): wire pi-langfuse agent tracing into MemSWE smoke runner

[LaloLalo1999]

Summary

Adds the pi-langfuse extension (from https://pi.dev/packages/pi-langfuse) to the MemSWE harness so that runs emit complete observability to Langfuse for every action and datum on a run.

The harness already exported its own benchmark/memory/verifier/scoring spans over OTLP (memswe-trace-scaffold.ts), but the agent session itself loaded empty extensions (createMemSwerResourceLoader), so the agent-internal layer — prompts, LLM generations, tool calls, final response, usage, cost — was never captured. This PR closes that gap by loading pi-langfuse into the faux and minimax agent sessions, giving a single Langfuse view that combines:

• Harness spans (existing): benchmark, memory, verifier, scoring + run metrics, over OTLP.
• Agent traces (new): root agent, per-request generation, per-tool tool observations.

Changes

• packages/coding-agent/package.json: pinned pi-langfuse@1.5.4 as a devDependency (harness scripts run via tsx; not shipped in dist/, and excluded from the published npm-shrinkwrap.json).
• scripts/memswe-langfuse-extension.ts (new): isLangfuseAgentTracingConfigured() gate + loadLangfuseExtensions() that loads the extension from its package manifest path (so its @langfuse/* / @opentelemetry/* types stay out of the harness typecheck and a missing optional dep degrades to an inert run).
• scripts/memswe-smoke-runner.ts: thread observabilityEnabled into the agent sessions; load the extension when configured; record langfuse_tracing in agent-result.json; log activation per session. Also imports the already-exported SessionSpec/VerifierSpec types the runner used but never imported.
• README.md: new Observability (Langfuse) subsection documenting both layers and setup.

Activation & safety

• Agent tracing activates only when pi-langfuse finds credentials: the LANGFUSE_PUBLIC_KEY / LANGFUSE_SECRET_KEY env pair, or a saved ~/.pi/agent/pi-langfuse/config.json. --no-otel-trace disables it.
• Without credentials the extension is inert (it returns without prompting in headless mode), so the deterministic faux path is unchanged and default smokes never call Langfuse.
• Langfuse credentials remain scrubbed from verifier/setup subprocess environments (existing SECRET_ENV_NAME_PATTERN), so traces never reach agent- or author-controlled commands.

Verification

• npm run check (biome, pinned-deps, ts-imports, shrinkwrap, tsgo --noEmit, browser-smoke): passes. The published npm-shrinkwrap.json is unchanged (devDependencies are excluded from the shrinkwrap walk).
• Wiring probe: loadLangfuseExtensions() loads the extension with 0 errors and registers its 3 commands and all 19 observability event handlers (before_provider_request, tool_call, tool_result, agent_end, …); isLangfuseAgentTracingConfigured() returns false/true for empty/keyed env.
• Smoke-runner unit tests covering touched logic pass. The full memswe:smoke run and 5 suite tests require the sibling ../memswe benchmark repo, which is not cloned in this environment (pre-existing ENOENT, unrelated to this change).

🤖 Generated with Claude Code

---

Generated by Claude Code

---

^{Need help on this PR? Tag /codesmith with what you need. Autofix is enabled.}

[blocksorg]

Mention Blocks like a regular teammate with your question or request:

@blocks review this pull request
@blocks make the following changes ...
@blocks create an issue from what was mentioned in the following comment ...
@blocks explain the following code ...
@blocks are there any security or performance concerns?

Run @blocks /help for more information.

Workspace settings | Disable this message