fix(edge-otel): bounded DB span names, parameterised db.query.text, http.route + error.type

[dash0-dev]

Why

Edge DB spans were named after the full statement with filter values interpolated, so api emitted 511 distinct span names in one week. 287 of them carried full-text search terms (built from agents' tool-failure text, which the CLI keeps out of its own telemetry), 133 carried memory keys, and 4 carried user ids. This was recommendation #1 in docs/telemetry-quality-review.md.

What changed

• DB CLIENT spans are named by db.query.summary (SELECT memories, CALL memory_write). db.query.text is parameterised ($n, LIMIT included) and renders a mutation's .select() as RETURNING.
• error.message on every span has request URLs redacted, so a failed request no longer copies its URL and the filter values in it into telemetry. A postgrest-js fetch failure (status: 0) on a DB span records fetch failed.
• DB spans add server.address (the hosted Supabase host; omitted for BYOD) and db.response.status_code. Root SERVER spans get http.route. Every recorded error gets a bounded error.type (an Error's name, a PostgREST code, or _OTHER), and an unexplained 5xx falls back to the status code.
• Also fixes the Evals preflight. The latest Claude Code CLI ships default-on built-in plugins, which the harness's empty enabledPlugins couldn't switch off (packages/evals, one commit, can be cherry-picked out).

How to verify

cd packages/mcp-core && npx vitest run and node scripts/ci/deno-check-functions.mjs. Telemetry was verified on this PR's CI integration-smoke traces in Dash0: 57 span names per run, stable across runs, and no quoted literal in any statement.

Notes for reviewers

With http.route set, Dash0 names api operations by route (POST /mcp) instead of by function. Move the two dash0_operation_name="mcp" check rules to faas_name="mcp" before merging, or the absence rule will fire about 2h after deploy (see docs/otel.md). llms.txt and the user docs are unchanged because this is internal telemetry.

[dash0-dev]

✅ No issues found

The delta since 1540895 is docs plus one comment, and it closes the last open thread: catch-block guidance now names errorTypeOf. A full deep re-pass found nothing above the bar; three low-confidence notes are listed as advisory.

_{3 advisory findings below the confidence bar — see Less certain below.}

Checked: 28 of 29 changed files read · 5 possible issues → 0 confirmed → 0 posted

Less certain (3) — advisory, below the confidence bar

• supabase/functions/_shared/telemetry/otel.ts:1087 — suggestion: Span.error coerces with String() because a throw would break the request, but this call does not. A non-2xx JSON body without message makes .replace throw and the span is lost. (confidence 76)
• supabase/functions/_shared/telemetry/otel.ts:1103 — suggestion: A status-0 failure has no code, so error.type becomes PostgrestError though PostgREST never answered. The bounded TypeError: / AbortError: prefix of its message is dropped. (confidence 86)
• supabase/functions/_shared/api/auth.ts:82 — suggestion: This span and mcp/auth.ts:178 keep the statement as their name and set no db.query.summary or server.address, so they fall outside summary-grouped views. (confidence 80)

Review details

Gate	Status	Details
Description vs. code	✅	Matches the diff: summary-named DB spans, $n statements, URL redaction, http.route + error.type, evals fix.
Prior review feedback	✅	Earlier review comments are resolved.
Documentation	✅	otel.md, CLAUDE.md, key-files and docblocks cover naming, placeholders, redaction, http.route, error.type.
Self-review signals	✅	No debug logs, leftover TODOs, or unreviewed stubs.
Code review	✅	The multi-lens review found no blocking issues.

Found

Quality — produced 5 → posted inline 0 · cleared 0 · carried forward 0 · deferred 0 · below-bar 3
Optimality — ran — 0 proposals; one redaction choke point in Span.error plus the pure, import-free span-semconv module is the optimal shape for the stated intent
Standards — ran — 7 governing docs, 137 normative bullets checked against all 29 files; 0 findings (1 candidate contradicted: webhook.ts catch only ever sees Error instances)

Run

full · 9 lines in delta · tier deep · depth checkout
⚠️ holistic, optimality, measurability and confidence lenses ran from the pipeline rule summaries, not their SKILL.md; span-semconv.spec.ts skimmed (index + 1 block) — ~970 of 1,399 changed lines read — no sub-agent dispatch available — the intent finder ran in-context with the other finders at effective thoroughness 0.95, instead of as its own sub-agent
Memories — 35 indexed · 7 used (4 knowledge · 1 hotspot · 2 lessons)

• knowledge knowledge::createTracedClient@supabase/functions/_shared/telemetry/otel.ts — re-verified at a54351c: DB-arm naming, status-0 'fetch failed' and explicit code type unchanged
• knowledge knowledge::errorTypeOf@supabase/functions/_shared/telemetry/span-semconv.ts — re-verified against postgrest-js 2.110.8: result.error is a plain JSON object, so the code fallback holds
• knowledge knowledge::errorTypeFrom@supabase/functions/_shared/telemetry/span-semconv.ts — re-verified at a54351c: an absent type falls back to the Name: prefix, so a code-less DB error records PostgrestError
• knowledge knowledge::redactUrls@supabase/functions/_shared/telemetry/span-semconv.ts — re-verified at a54351c: http(s)-only by design, prior scheme finding stays contradicted
• hotspot hotspot::supabase/functions/_shared/telemetry/otel.ts — 6 confirmed here on this PR; error-path and missing-update checklist for the DB arm
• lesson telemetry-privacy::anchor-a-credential-strip-on-any-scheme-not-https — checked redactUrls' literal scheme; edge I/O is https-only, so no finding
• lesson otel::omit-the-attribute-rather-than-emit-a-placeholder-outside-the-semconv-enum — error.type _OTHER is the convention's own fallback value, not an invented placeholder

^{Nothing to report — measurability (17 paths classified), integrations (not activated), severity, 0 files skipped.}

^{pr-reviewer · commit a54351c · full review · how these findings are produced · updated Oct 2, 2026 3:51pm UTC}