feat: add OpenTelemetry error reporting

[matheus1lva]

Summary

Report Worker errors as OpenTelemetry log records over OTLP/HTTP. Cloudflare Workers can't run the OpenTelemetry Node SDK (its OTLP exporter is Node-coupled), so the OTLP/HTTP log payload is built and sent directly via fetch — still standard OTLP. Vendor-neutral: point OTEL_EXPORTER_OTLP_ENDPOINT or OTEL_EXPORTER_OTLP_LOGS_ENDPOINT (and the matching OTEL_EXPORTER_OTLP_*_HEADERS for auth) at any OTLP backend. No-op until configured.

How to review

• src/observability.ts: captureError(ctx, env, error) POSTs an OTLP LogsData payload (ERROR severity, exception.* attributes). Signal-specific logs endpoints are used as-is; the generic endpoint gets /v1/logs appended. Headers are percent-decoded per the OTLP env-var spec. timeUnixNano is BigInt(Date.now()) * 1_000_000n. Uses ctx.waitUntil so the export completes after the response is returned.
• src/index.ts: unexpected-error and ApiError with status ≥ 500 call captureError. OTEL_* fields are on the Env type.
• src/edge-cache.ts: background caches.default.put failures are routed through captureError.

Test plan

• Manual: set OTEL_EXPORTER_OTLP_ENDPOINT (+ headers), trigger the 500 branch, confirm the log record reaches the backend.
• Manual: leave the endpoint unset, confirm nothing is exported and responses are unchanged.
• Automated: bun run test and bun run typecheck.

Risk / impact

Low. With the OTLP endpoint unset (the default), reporting is a no-op and behavior is unchanged. No funds, auth, or migration paths touched.

[murderteeth]

Summary

Adds vendor-neutral OpenTelemetry error reporting to the Cloudflare Worker. Since the OTel Node SDK can't run on Workers, src/observability.ts hand-builds an OTLP/HTTP JSON log record and POSTs it via ctx.waitUntil (off the latency path). Fires on unexpected errors and on ApiError with status ≥ 500; no-op until OTEL_EXPORTER_OTLP_ENDPOINT (or _LOGS_ENDPOINT) is set. Verified locally: typecheck passes, 43/43 tests pass, wrangler deploy --dry-run bundles cleanly. Payload carries only error.name/message/stack — no request data or secrets.

Dependencies

• package.json unchanged — but bun.lock gained an orphan @opentelemetry/api@1.9.1 entry (see Issues). The package itself is benign (official OTel org, no install scripts), but nothing imports it.

Issues

• bun.lock:171 — Orphan @opentelemetry/api@1.9.1 lockfile entry with no package.json counterpart; residue from the abandoned SDK approach, and it does get installed (resolves vitest's optional peer).
  • Change: revert the bun.lock hunk so the lockfile matches the manifest.
  • Done when: bun install --frozen-lockfile succeeds and @opentelemetry/api no longer appears in bun.lock or node_modules.

Suggestions

• src/index.ts:92 / src/edge-cache.ts:79 — ctx.waitUntil(edgeCache().put(...)) has no .catch, so background cache-write failures become unhandled rejections that never reach the OTLP backend. This PR is the natural place to route them through captureError.
• test/observability.test.ts:59 — the "reports rejected exports" test stubs a resolved 401; the actual fetch-rejection .catch branch (src/observability.ts:71) is untested, and the documented no-op path (endpoint unset → zero fetches) has no test either.
• src/observability.ts:39 — String(Date.now() * 1_000_000) exceeds MAX_SAFE_INTEGER; String(BigInt(Date.now()) * 1_000_000n) is exact.
• src/observability.ts:16 — parseHeaders splits on , without percent-decoding, diverging from the OTLP env-var spec; a header value containing a comma would be split into a bogus second header.

Verdict

APPROVE

---

How This Was Reviewed

This review was conducted using the review-pr skill.

[matheus1lva]

8668b12:

• reverted the orphan @opentelemetry/api lockfile entry (bun install --frozen-lockfile is clean; package not in node_modules)
• cache waitUntil failures go through captureError
• timeUnixNano is BigInt(Date.now()) * 1_000_000n
• parseHeaders percent-decodes values
• tests cover fetch rejection, unset-endpoint no-op, 401 export, and the header/timestamp cases

[murderteeth]

Summary

Re-review of 8668b12 (delta: one commit since ee2411a). Prior review items verified: orphan @opentelemetry/api lockfile entry removed (frozen install clean), edge-cache waitUntil failures now routed through captureError, real fetch-rejection and no-op tests added without weakening the old ones, timeUnixNano now exact via BigInt. Typecheck, tests (46/46), and wrangler deploy --dry-run all pass. One new blocking defect introduced by the parseHeaders rewrite.

Issues

• src/observability.ts:22 — The new decodeURIComponent calls in parseHeaders are unguarded, so a malformed OTEL_EXPORTER_OTLP_HEADERS value containing a raw un-encoded % (e.g. authorization=Api-Key 50%off) throws URIError synchronously inside captureError. Verified by execution. At src/index.ts:111/123 the throw escapes the catch block, turning every 5xx into a Cloudflare 1101 worker exception instead of the JSON error response; at src/edge-cache.ts:88 it rejects the waitUntil promise, re-introducing the uncaught background rejection this revision just fixed.
  • Change: wrap the two decodes in a try/catch inside parseHeaders only — fall back to the raw (undecoded) key/value or skip the malformed pair.
  • Keep: the comma-split + percent-decode behavior for well-formed values must remain (a %2C in a value still decodes to ,); the existing observability tests must still pass.
  • Done when: a test sets OTEL_EXPORTER_OTLP_HEADERS to a value with a raw % and asserts captureError does not throw, alongside the existing decode tests passing.

(This is a new regression introduced when applying previous review feedback. Lets prioritize a spec for automated pr reviews and cite this as a reference case.)

Verdict

REQUEST_CHANGES

---

How This Was Reviewed

This review was conducted using the review-pr skill.

[murderteeth]

Previous round's item is resolved — decodeHeaderPart guards the decode, %2C still decodes, raw-% test added. Branch on its own is green (typecheck clean, 47/47).

Issues

• Resolve conflicts with main

Verdict

REQUEST_CHANGES

---

How This Was Reviewed

This review was conducted using the review-pr skill.

[murderteeth]

Summary

PR #13 adds vendor-neutral OTLP/HTTP error reporting for Worker failures. The latest revision resolves the previous malformed-header issue and cleanly incorporates current main. No blocking issues found.

Issues

None.

Suggestions

None.

Verdict

APPROVE

---

How This Was Reviewed

• Reviewed the complete diff against current main
• Verified prior review findings and conflict resolution
• Confirmed no new dependencies
• bun run typecheck passed
• bun run test passed: 101/101 tests
• wrangler deploy --dry-run passed
• GitHub CI checks are green
• Visual verification was not applicable because there are no UI changes

This review was conducted using the review-pr skill.

[matheus1lva]

on hold until doppler - so we can get better deployments