v0.10.0

Minor Changes

• 6837e8e: Self-monitoring separates caller errors from server errors. Only server errors (New Relic unreachable, rate limiting or timing out, an unexpected answer, a refused stored key, internal errors) set span status ERROR and error.type, and rpc.server.duration now also carries rpc.grpc.status_code for them, so the New Relic service's Error rate counts them (it stayed 0 before). Caller errors (invalid input, limits, quotas, unknown query ids, tools or accounts not granted) leave the span status unset, so they no longer open Errors Inbox groups; they are still counted by status, error.code and the new fault attribute (caller/server). Alerts on span status ERROR now see server errors only.

Patch Changes

• 59e4792: Self-monitoring access logs name server/discover (the MCP 2026-07-28 discovery request) instead of mcpMethod = other.

• db45930: Bound redaction and NRQL masking work on long input: every secret pattern now does work linear in the input, and
  output cells, trailers, column names, error messages and the audit NRQL column redact only the text they keep (with a
  margin so a secret crossing the cap is still recognised). An unterminated NRQL string ending in a backslash is now
  masked too.

• 074ed5e: A malformed RB_RATE_LIMIT_* value no longer appears in the startup error, so a secret pasted into the wrong variable never reaches the log.

• bb78276: Hardening from the pre-launch review:

  • Event-type allow and deny lists now also cover queries without a FROM clause. With a list set, run_nrql refuses SHOW EVENT TYPES and other FROM-less statements or subqueries (RB_LIMIT_QUERY); use list_event_types, which respects the lists.
  • COMPARE WITH inside a subquery is refused with a clear error (New Relic does not support it there).
  • Requests to NerdGraph (which carry the User API key) and OTLP exports never follow redirects: a 3xx answer fails the request. This also makes OTLP export work on Cloudflare Workers, which rejects the redirect: 'error' mode the exporter used.
  • Behaviour change: "Don't ask again" on the OAuth consent page is now offered only to apps identified by their client ID metadata document (such as claude.ai). Dynamically registered apps, whose identity is not verified, ask every time. Approvals remembered for them earlier are no longer used or listed.

• 8e93886: OAuth client limits are fairer. Metadata-document clients that completed a token exchange (or hold a grant) can always fetch their document again; records of clients that never completed an exchange are kept for 5 minutes and, when 200 are stored, the oldest one makes room instead of new clients being refused. Dynamic client registration now counts toward the deployment-wide cap only once the registration succeeds, and every attempt that passes the registration policy counts toward the per-address cap, accepted or not.

• 6a4e007: trace_error with error_group_id finds the group's traced occurrence again (0.9.1 verification, NEW-13). It filtered TransactionError on error.group.guid, which those events do not have, so it always said no traced occurrence was found. It now looks the group up in Errors Inbox and picks the latest traced error with the group's error class and message on its entity, falling back to class and entity (with a note) when the message does not match exactly. An unknown group id, or a group that does not come from APM TransactionError events (OpenTelemetry spans, browser errors), is a clear RB_INPUT_INVALID with a hint. When no occurrence is found but other errors carry trace ids, the note says tracing works instead of implying it is off. The call now makes up to 5 requests with error_group_id.

• aead7e5: trace_error fills errors[].httpStatus for every agent (0.9.1 verification, NEW-14). It read only httpResponseCode, which the PHP agent sends as a string ("500"), so the status was null. It now takes the first valid HTTP status (100-599) from http.statusCode, response.statusCode, response.status or httpResponseCode, numeric or a 3-digit string.

• 4acc0a0: Golden-metric trend notes say why there is no trend (previous window unread, no data in the previous window, too few
  points) instead of always claiming the halves were compared; changeRatio null for a zero baseline is documented.

v0.9.1

Patch Changes

• c897f62: get_attributes now applies name_contains before the 1,000-attribute cap (NEW-08). On event types with more than 1,000 attributes, such as Log, a filter whose matches sat past the cut returned an empty list with truncated: true. totalAttributes is the real count before filter and cap, and truncated means the filtered list shown was cut.
• 614ff91: An explicit baseline: "same_window_7d_ago" whose control falls outside the connection's lookback limit now falls back to same_window_1d_ago in analyze_transactions and analyze_deployment_impact, instead of running no control and reporting used: "previous_window" (NEW-11). baseline.used and reason name the requested baseline and what ran, and a note stays. The limit is never relaxed.
• edfff5f: list_recent_issues no longer falls back silently (0.9.0 verification, NEW-09). When group_by counts only the page, groupsNote (and a text note) says why and how many issues the counts cover (window too large for the extra requests, request budget, or a failed request). A limit above the API page size (50 rows) is reported as limitNote, e.g. "limit 200 lowered to 50 (API page size)", instead of returning fewer rows without a word.
• 2f7b3bb: run_nrql no longer fails output validation on queries with both FACET and TIMESERIES (P8-44). NerdGraph returns totalResult and otherResult as an array of time buckets for that shape; the output schema and text now accept either an object or an array.
• 2330bb5: Self-monitoring logs are easier to use (P8-50). A standalone GET stream or subscriptions/listen stream that reaches the request deadline now logs mcp stream closed at deadline at info instead of an mcp request timed out warning; a real request that never finishes still warns. Access-log (http request) records now export the integer HTTP status, method and the matched route pattern (fixed sets; never raw paths), so 4xx/5xx can be counted in New Relic. The observability guide now lists the attribute names as New Relic shows them (tool, ms, code, without log.). Deadline and access-log records also export the MCP method (fixed list, else other) and, for tools/call, the tool name.
• 86c809a: Self-monitoring gives the New Relic service entity throughput and response time: relic-bridge now exports the OpenTelemetry rpc.server.duration metric (New Relic builds the golden metrics from it), adds telemetry.sdk.* and a random per-process service.instance.id to the resource, and names spans tools/call <tool> so Errors Inbox groups are distinct per tool.
• b5ab36e: trace_error now says how many errors carry a trace id when it finds no spans ("none of N errors carry a trace id" or "even though K of N errors carry a trace id") instead of always claiming some do (NEW-10). The count is folded into the existing tracing check query.
• 0c719e6: whoami returns limits.callBudget again, as a deprecated alias with the same value as limits.perCallBudget (NEW-12). Version 0.9.0 renamed callBudget to perCallBudget and removed callsUsed and callsRemaining without a deprecation path. The alias will be removed in the minor after next, so switch to perCallBudget. callsUsed and callsRemaining stay removed: the budget is per tool call, not a running total. The whoami description now says what perCallBudget means.

v0.9.0

Minor Changes

• 44faed0: The call budget is now unmistakably per tool call (NEW-02, BUG-28). whoami reports limits.perCallBudget (the most New Relic calls one tool call may make) instead of callBudget, callsUsed and callsRemaining, which only described whoami's own call and looked like a broken running total. Its text points to rate limits and grant quotas for limits across calls, and the docs say the same. These output fields were removed: reconnect your MCP client after upgrading so it reads the new tool schema.

Patch Changes

• 9a86880: Tool output schemas no longer forbid extra properties, so a client that connected before an upgrade no longer rejects results that gained fields (for example whoami after 0.7.0 to 0.8.0, NEW-01). Input validation is unchanged. Reconnect your MCP client after upgrading relic-bridge to see new tools and parameters: a stateless server cannot signal a changed tool list.

• 41a8ecd: partial now means involuntary truncation only (0.8.0 verification run, 2026-10-13, NEW-05). Golden metrics left out by your own max_metrics (and related entities left out by max_dependencies) no longer set partial: true or a partial status in get_golden_metrics, entity_health, dependency_health and generate_alert_insights_report; they are listed as skipped with a reason (max_metrics, budget, error or no_data). A cut forced by the request budget or an error still sets partial.

  list_recent_issues with group_by now counts the whole requested window, not only the returned page (NEW-07). The window is read in up to 4 extra pages, counted in the declared cost; if it is larger or the budget runs out, the counts cover the page and the new groupsScope field says page (a note says why). groupsScope is always present.

• 5bdea2f: - list_top_transactions with sort_by: slowest now skips transactions with fewer than min_requests requests (default 10), so one-request transactions no longer top the ranking (NEW-03).

  • list_event_types lists global event types only in globalEventTypes, no longer also in eventTypes (NEW-04).
  • The automatic seasonal baseline of analyze_transactions and analyze_deployment_impact prefers the same windows 7 days earlier for windows of 24 h or more or across a weekend, when the maxWindowDays limit allows it; otherwise it uses 1 day and says which maxWindowDays a weekly control needs (NEW-06).

• fd46afe: trace_error now says span data is missing whenever the account or entity has no Span events, even if errors carry a trace id (BUG-12), and notes when its tracing check could not run.

v0.8.0

Minor Changes

• 99c8916: Name lookups can be narrowed with optional domain and type (for example domain: "APM") on every tool that takes name, so an app that also has a same-named browser entity no longer fails with an ambiguity error. The ambiguity error now suggests domain. APM-only tools (transactions, deployment impact, threads, GC) keep resolving a bare name to the APM entity.
• 5942246: Deployment and transaction verdicts are more reliable. analyze_transactions and analyze_deployment_impact now call a change in error rate when |z| >= 3 and at least 5 errors more or fewer than expected (a strong z-score under half a percentage point is no longer "similar"), and every note says which rule decided. New optional baseline input (previous_window, same_window_1d_ago, same_window_7d_ago): when the throughput shifted a lot between the windows (for example the overnight trough) the same windows one day earlier are compared as a control (2 extra queries), and a change that also happened then is a warning instead of regression_suspected.
• e0ebb62: maxWindowDays now limits how far back a query reads as well as how wide it is: SINCE 60 days ago UNTIL 59 days ago (and the same with dates, epoch seconds or milliseconds, COMPARE WITH offsets, or subqueries) is refused with RB_LIMIT_QUERY, in run_nrql and in every tool that reads a time window. Exactly maxWindowDays days is still allowed.

Patch Changes

• 75d5976: Fix list_change_events in account scope returning no APM deployments: the account-wide change feed does not carry deployments recorded through the APM deployment API, so the tool now also reads the deployments of up to 50 of the account's APM applications and merges them (limits are reported in a new notes field). Events no longer come back with type: null; the type is taken from the category when NerdGraph leaves it empty.

• 5a2c88d: Error contracts and polish from the live test run: get_query_result with an unknown or expired id now fails with the new RB_QUERY_NOT_FOUND code and a "rerun run_nrql" hint; a run_nrql query with several ;-separated statements is refused with "one NRQL statement per call"; limit and time-range errors show their details as plain text instead of wrapping them as untrusted New Relic data; whoami reports callsUsed and callsRemaining and states that the call budget is per tool call; list_event_types flags global types (Public_*, NrUsage, ...) that cannot be mixed with account data (globalEventTypes); get_attributes takes name_contains to narrow large attribute lists.

• f169613: Golden metrics: the value and trend now ignore the time bucket that is still filling (no more "throughput 0" or a false drop), the trend compares equal halves of the complete buckets, timeseries: false really returns one aggregate without TIMESERIES, signals read "down 39%" instead of "down -39%", and query adjustments such as "LIMIT MAX replaced" are stated once instead of on every metric.

• e265186: Golden-metric trends (get_golden_metrics, entity_health) now compare the window with the previous equal window, like analyze_transactions, instead of the two halves of one window (live test run, 2026-10-12, BUG-06). This adds one query per metric, counted in the request cost; beyond the window limit the old halves method is used and a note says so.

• 5a6343b: list_recent_issues takes group_by (condition or entity) and counts the returned page per group, so repeated alerts for one condition or host collapse into a short summary.

• 7fb8c15: analyze_deployment_impact and analyze_transactions now apply the same |z| >= 3 rule to average and p95 latency as to error rate and throughput: the window query adds the standard deviation of duration, and a latency move beyond the 15% band and 5 ms only counts as worse or better when it is statistically significant. Noisy low-traffic windows no longer give false latency regressions, and each note names the rule and the z value (BUG-07).

• a4a85e9: Fix list_top_transactions failing on every call: the generated NRQL used ORDER BY after FACET, which NRDB rejects. The ranking aggregate (p95, error count or request count per sort_by) now leads the SELECT list, which is how NRDB orders facets.

• 1a65af5: Log tool fixes from the live test run (2026-10-12): analyze_entity_logs message patterns no longer fail with a too-large response on busy entities (messages are cut at New Relic, at most 20 patterns with one example each, and a too-large sample is repeated smaller instead of dropping the section); list_recent_logs and analyze_entity_logs share one level vocabulary that also accepts the Monolog/syslog names (warning, notice, critical, alert, emergency); time_range: "last 7 days" is accepted with a 7-day limit (whole-second rounding no longer counts against the window); list_recent_logs paging keeps one fixed since/until, never returns lines outside it and no longer skips or repeats lines with equal timestamps; lines with an empty message show their log or logged_message field, or a short attribute summary.

• deca849: Fix list_recent_logs failing with a response-too-large error on accounts with very wide Log events: it now reads a fixed set of columns (timestamp, level and its aliases, message, log, logged_message, messageId and a few context attributes) with long text cut at New Relic, instead of every attribute. Lines without a message still fall back to log, logged_message, then a short summary built from that small context set.

• d8b68ab: Fix the relationship tools. get_related_entities with types no longer returns an empty list for hosts: the kind is matched on each neighbour's real domain and type across several pages, and edges from the entity to itself are left out with a note. dependency_health now skips key transactions, service levels, alert objects, workloads and dashboards, reads several relationship pages, and measures hosts and called services first. decode_entity_guid reports infra host guids as type HOST (inferred) instead of NA.

• f94ccbe: Report and issue fields fixes:

  • generate_user_impact_report fills front end names from the app events and reports sessionsWithErrors as unknown (null, with a note) instead of 0 when JavaScript errors carry no session id; the js_errors_up threshold (rate 1.5x, 10+ errors) is documented.
  • list_error_groups and entity_health fill lastSeenAt from the in-window occurrences when the group has none.
  • trace_error says distributed tracing appears to be disabled when no traced occurrence is found and the entity or account has no Span or traced error data.
  • search_incidents parses JSON-encoded titles into titles[] and a joined title.
  • list_recent_issues adds a groups summary and a text header when several issues on the page share a condition.

• c1fcb9d: list_error_groups, list_recent_issues and generate_alert_insights_report now return short permalinks (entity redirect, issue page with account) instead of long launcher URLs that were cut off with "…" and made the output much larger. Links are left out under minimal_output.

• 642d8ac: Fix list_synthetic_monitors failing on every call: the monitor status is now read from monitorSummary { status }, because NerdGraph has no status field on the monitor outline type.

v0.7.0

First published release. It contains everything built in Phases 1-7: local stdio, self-hosted HTTP, Cloudflare
Worker, sign-in and admin panel, OAuth 2.1 for MCP clients, the extended tool set and the power features below.
Versions 0.1.0-0.6.x were never published; the 0.1.0 notes are kept below for reference.

Minor Changes

• Claude Code plugin (P7-07): MCP config, skills and commands for using relic-bridge from Claude Code.

• Config import and export (P7-06): move projects, grants and quotas between servers as one validated file, without secrets.

• Doctor command (P7-05, P7-17): checks configuration, keys, accounts and client configs, and says how to fix what it finds.

• Per-grant usage quotas (P7-12): daily and monthly caps on New Relic calls and inspected events, with a soft warning and RB_LIMIT_QUOTA past the cap.

• MCP Tasks (P7-13, P7-21, P7-22, P7-23, P7-29): long composite reports run as tasks with polling, cancel, notifications and compare-and-swap task records.

• OS keychain support (P7-11, P7-30): keep New Relic keys in macOS Keychain, Windows Credential Manager or libsecret; helper programs run by absolute path.

• OpenTelemetry export (P7-08, P7-18): traces, metrics and allowlisted logs to your own OTLP collector; filesystem paths are never exported (P7-30).

• Paging and limits are the same on every list tool (P6-24). Breaking, listed with the P6-22 renames: list_change_events and
  list_recent_logs page with cursor / nextCursor (newest first, by time) instead of truncated; get_related_entities
  pages and no longer returns truncated; list_alert_policies, list_alert_conditions, list_recent_issues,
  search_incidents and list_error_groups take limit (default 25 to 50, max 200) and the cursors they return are
  composite and bound to the tool, account and filters (a NerdGraph cursor passed back as is is refused); list_accounts,
  list_entity_types and list_event_types take limit and cursor (list_accounts no longer returns truncated).
  On time-windowed lists a cursor keeps the window of the first call. list_top_transactions and get_attributes stay
  bounded and say so in their descriptions. The tools/list budget in the offline eval grew from 52 000 to 53 500
  characters for the new limit and cursor arguments.

• Breaking (pre-1.0): one name per concept in tool arguments and structured output (P6-22). Tool names are unchanged,
  so grants keep working; clients and scripts that pass arguments by the old names must update them. No aliases are kept
  (nothing was released with the old names). The convention is in docs/dev/tool-conventions.md.

  Arguments:

  • guid → entity_guid on get_entity, get_related_entities, get_golden_metrics, decode_entity_guid,
    get_dashboard, dependency_health, entity_health.
  • since → time_range on get_golden_metrics, dependency_health, list_event_types, get_attributes.
  • name (substring filter) → name_contains on search_entities, list_alert_policies, list_alert_conditions,
    list_dashboards, list_synthetic_monitors. name now always means an entity name resolved to entity_guid.
  • run_nrql: timeout_s → timeout_seconds.
  • list_recent_issues: policy_id is a digit string, like list_alert_conditions and list_alert_policies output.
  • generate_user_impact_report: compare → compare_previous.

  Structured output:

  • The account of a result is accountId (was account); lists of account ids are accountIds (run_nrql,
    get_query_result, list_entity_types, resolve_entity); whoami / list_accounts return defaultAccountId and
    unreachableAccountIds.
  • entity_health returns entityGuid (was guid).
  • Lists cut without paging return truncated (was maybeMore on list_change_events, list_recent_logs,
    list_top_transactions; hasMore on get_related_entities).
  • list_error_groups: total (was totalCount).
  • search_entities, list_dashboards, list_synthetic_monitors also return hiddenOutsideScope.
  • get_golden_metrics, dependency_health, list_event_types, get_attributes return resolved since and until
    (ISO 8601) instead of the window expression.
  • run_nrql / get_query_result: retryAfterSeconds (was retryAfterS).
  • APM transaction stats: avgDurationSeconds, p50DurationSeconds, p95DurationSeconds, p99DurationSeconds;
    Kafka rates: bytesInPerSecond, bytesOutPerSecond, messagesInPerSecond.
  • Paged tools print nextCursor: <value> in their text (was next_cursor:), matching structuredContent.

• New tool get_ui_link (P7-03): offline links that open an entity, dashboard, NRQL query (query builder) or alert issue
  in the New Relic UI, with the US or EU host of the connection. Built only from validated ids and never from chart URLs.

• Opt-in write tools (P7-09, P7-19, P7-20, P7-25, P7-26): a few small changes in New Relic (record a change, acknowledge or close an alert issue, create a dashboard). Off by default at every layer, per grant, audited before each mutation, with confirmation or human approval.

Patch Changes

• Phase 6 review nits: entityQuery.and() parenthesises parts containing OR, generate_user_impact_report no longer mislabels
  front ends when an entity name is empty, and section summaries say "skipped (budget or grant, see the section notes)".