🤖 feat: add opt-in AgentPond tracing

[marcusschiesser]

Summary

• add opt-in AgentPond/OpenInference tracing to Mux's real Node StreamManager → streamText path
• keep tracing disabled unless a Files SDK provider is explicitly configured
• redact prompt and response content while retaining operation, model, timing, and token-usage metadata
• document the privacy/configuration behavior and update generated built-in skill content

Validation

• bun install --frozen-lockfile
• make static-check (all project checks passed; Hadolint was rerun with the CI-pinned v2.12.0 because Homebrew v2.15 adds a warning for an unchanged Dockerfile line)
• bun test src/node/services/streamManager.test.ts src/node/services/streamManager.modelOnlyNotifications.test.ts — 105 passed, 1 existing credential-gated test skipped
• end to end: invoked the actual private StreamManager.createStreamResult path with AI SDK's MockLanguageModelV3, consumed the stream, synced the local Files SDK store, and read trace 970445ba8cfaea3856d3d6c1bcbbaac5 back
  • generation b6b86e88b4319dd1, provider mux-agentpond-e2e, model fixture-model
  • 7 input / 4 output / 11 total tokens
  • trace and observation content were null; raw-store search found neither prompt nor response sentinel

Configuration and limitations

AgentPond tracing is independent of Mux's anonymous usage telemetry. It activates only when FILES_SDK_PROVIDER and the selected Files SDK provider variables are present before startup. The integration intentionally records no prompt or response bodies.

---

Generated with mux • Model: codex/gpt-5 • Thinking: high • Cost: $unknown

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 93bafb8214

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Restrict forced flushes to short-lived CLI shutdown

When tracing uses a remote or slow Files SDK provider, these async integration hooks force an export after every successful, aborted, or failed turn, so stream completion can wait for network I/O or the exporter's timeout. Because the same telemetry object is attached by the shared StreamManager, long-lived Electron and server sessions also pay this CLI-durability cost and receive no batching benefit; force-flush only when an actual short-lived CLI process is exiting, or make the per-turn flush bounded and non-blocking.

AGENTS.md reference: AGENTS.md:L10-L10

Useful? React with 👍 / 👎.

[marcusschiesser]

Addressed in a3e57c7. The AI SDK integration no longer registers per-turn onEnd / onAbort / onError flush hooks, so normal long-lived turns use the batch processor as intended; the explicit lifecycle/test flush helper remains available. The 105 focused tracing and stream-manager tests pass (1 existing credential-gated skip). A fresh production StreamManager.createStreamResult E2E run completed with the AI SDK mock model, then flushed at the validation lifecycle boundary and read back 1 object / 4 events: trace 7da03f7d77d9cc673d3d153ff9bc9cf3, generation 2ebfce5734f2b1c1, with null trace/observation input/output and no prompt/response sentinels in the raw store. make static-check also passed ESLint, both TypeScript targets, formatting, generated-content, eager-import, code-link, and shell checks before the local Homebrew Hadolint 2.15.0 rejected an unchanged Dockerfile line that the repo's pinned 2.12.0 accepts.