[ephraimduncan]

During a telemetry.dev ingest outage, pi showed error blocks over its TUI. One prompt with two tool calls wrote 19 reports to stderr, and pi did not keep them. After this change, no telemetry.dev text reaches the pi terminal, and the outage is one entry in ~/.pi/agent/telemetry-dev.log. omp and opencode 1.x had the same problem and get the same correction.

[telemetry.dev] [TypeError: fetch failed] {
  [cause]: Error: connect ECONNREFUSED 127.0.0.1:4999
      at TCPConnectWrap.afterConnect [as oncomplete] (node:net:2083:16) {
    errno: -61,

One report for an outage

Each failed batch went to onError and to the console. flush() reported the same failure a second time, because the span processor rejects its flush when an export fails. The exporters of one client and the client flush now share one failure state, from createExportFailures in @telemetry-dev/otel.

stateDiagram-v2
    [*] --> Healthy
    Healthy --> Reported: export or flush fails / onError(error)
    Reported --> Reported: export or flush fails / no report
    Reported --> Healthy: a request reaches the ingest

Loading

An export without data sends no request, so it does not end the series. TelemetrySpanProcessor shares the state between its trace and metric exporters. The change applies to all SDK users. It keeps the onError contract that the SDK can merge repeated failures into one report.

Errors go to the log of the host

The pi, omp, and opencode integrations now default to logLevel: "silent". Their default onError writes to the log that each host keeps:

Host	Default onError
pi	Appends to telemetry-dev.log in the pi agent directory, with mode 0600. When the file is more than 5 MB, it moves to telemetry-dev.log.1, as pi does with mcp.log. pi has no logger for extensions.
omp	pi.logger.error("telemetry.dev error", ...). omp keeps this file logger off the console because console output corrupts its TUI.
opencode	client.app.log with the service telemetry-dev. opencode 1.18.9 draws plugin console output over its TUI. opencode 2.x runs plugins in opencode serve --service, where stdout is /dev/null.

An explicit onError or logLevel replaces the default. Outside these integrations, the SDK default stays warn on the console. The cursor integration does not change. Its daemon starts with stdio: "ignore", so its output never reaches a terminal.

sdk-logs 0.218 and the next release

The TypeError: Cannot read properties of null (reading 'exportCompleted') in pi comes from @opentelemetry/sdk-logs 0.218. After it, the pi process sends no more logs. #80 moved the catalog to ^0.223.0, but npm still has @telemetry-dev/otel 0.1.6 and @telemetry-dev/sdk 0.1.9 with ^0.218.0. The open release PR does not change the otel version, because #80 changed no file in packages/otel.

This PR changes packages/otel, so the next release publishes otel too. With otel at 0.1.7, pnpm pack writes these ranges:

@telemetry-dev/otel 0.1.7   sdk-logs ^0.223.0, otlp-transformer ^0.223.0
@telemetry-dev/sdk          sdk-logs ^0.223.0, api-logs ^0.223.0, @telemetry-dev/otel ^0.1.7

^0.1.7 excludes otel 0.1.6, and ^0.223.0 excludes 0.218.x. The sdk also imports createExportFailures, which otel 0.1.6 does not export. A new test in logs.test.ts keeps a scheduled log export open during a flush. With the 0.218 read, the test fails.

Tests

Command	Packages	Result
pnpm run check (vp check and the OpenAI v6 type check)	all	pass
vp test run --changed	otel, sdk, pi, omp, opencode	77, 135, 22, 20, and 24 passed
pnpm run test	all TypeScript packages	pass

New and changed tests: transport.test.ts "does not retry non-retryable statuses and reports a series of failures once", otel-byo.test.ts "an outage that stops traces and metrics reaches onError once", client.test.ts "an ingest outage reaches onError once until an export gets through again", and logs.test.ts "logs keep reaching /v1/logs after a flush starts during a scheduled export". pi.test.ts, omp.test.ts, and opencode.test.ts each have one "errors go to ..." test. Each new assertion failed against a deliberate error in the owner code, for example a report for each batch, the 0.218 _flushAll read, or the warn log level.

pi session check

pi 1.4.0 loaded the worktree build with pi -e sdk/packages/pi/dist/register.mjs. Its imports resolved to the worktree packages. A local fake ingest answered each request after 700 ms. For the outage, TELEMETRY_DEV_BASE_URL pointed to a port with no server. "Before" is SDK main without this PR.

Check	Before	After
Outage, one prompt with two tool calls: [telemetry.dev] reports on stderr	19	0
Outage: screen captures with [telemetry.dev] text	4 of 7	0 of 7
Outage: entries in ~/.pi/agent/telemetry-dev.log	no file	1
Outage through 3 prompts, Escape, /new, /reload, and exit: entries	not run	1
Fake ingest: /v1/logs requests after the 5-tool prompt, /new, and /reload	7, 2, 2	7, 2, 2
Fake ingest: stderr bytes for the full run	0	0

Related: #85