Motivation

Outside of apps/cloud, the server apps have no configured logger (Effect's
default pretty printer), no log or metric export, and no tracer, so the
roughly 145 existing Effect.withSpan call sites in core packages export
nothing. MCP protocol traffic is only visible through the EXECUTOR_MCP_DEBUG
console hook. This makes self-hosted and local deployments harder to operate and
debug.

Closes #1231.

Changes

• New packages/core/observability package exporting observabilityLayer({ serviceName, endpoint, headers, logLevel }): a structured JSON console logger plus OTLP/HTTP export of traces, logs, and metrics. Built on effect/unstable/observability (fetch-based), so the same layer runs on Bun and Cloudflare workerd. No new dependencies; export is disabled when no endpoint is configured.
• All four server apps (local, host-selfhost, host-cloudflare, cloud) boot the layer from their own environment source. Cloud keeps its existing trace pipeline and error-capture correlation and gains only logs and metrics.
• Structured mcp.* log events across the MCP host: tool start/end with outcome and duration, internal errors with a correlation id, session lifecycle, auth outcomes, and elicitation at debug level. Two metrics, mcp.tool.calls and mcp.tool.duration_ms, tagged by tool and outcome. Code payloads are not logged, only mcp.execute.code_length.
• MCP tool-call spans carry OTel SpanKind.CONSUMER, so span classifiers treat each call as a unit of work instead of discarding a root internal span (the stdio transport case).
• Console logs are JSON lines on stderr, since the stdio transport owns stdout as its JSON-RPC channel, and each line carries trace_id/span_id for correlation with exported traces.
• The default HTTP access-log message (the constant "Sent HTTP response" from HttpMiddleware.logger) is replaced with a composed request summary (GET /api/things 200 in 12ms) via a shared httpAccessLogger middleware, keeping the same http.* annotations and adding http.duration_ms.
• Docs: package role in AGENTS.md, configuration in RUNNING.md. EXECUTOR_MCP_DEBUG is unchanged.

Configuration

On the Workers apps these are bindings rather than process env.

Testing and verification

• format:check, lint, typecheck, and test pass.
• New unit tests: logger output shape and level filtering, header parsing,
  disabled-when-unconfigured behavior, OTLP layer teardown; MCP host tests
  assert mcp.tool.start/mcp.tool.end fire with outcome, duration, and
  execution id, and that defects produce the correlation-id error log.
• Payload check against a local OTLP collector stub: spans arrive on
  /v1/traces, trace-correlated log records on /v1/logs, both metric series
  on /v1/metrics.
• Deployed to a self-hosted OTLP backend from this branch (apps/host-selfhost
  and apps/local). Screenshots below. With the endpoint unset, there is no
  network activity and existing e2e tests (including the cloud
  error-correlation contract test) pass unchanged.

MCP tool call classified as a task via SpanKind.CONSUMER, with tool
attributes and child spans:

Endpoint grouped by parametrized route with database child spans:

Disclosure

This change was developed with an LLM coding assistant (Claude).