qyl-mcp-server 17.0.2 → 17.0.3

package/dist/skills/qyl-investigate/SKILL.md

@@ -27,7 +27,7 @@ saved queries and failure status in that project, never the telemetry.
 | The user wants to know | Start with | Then |
 | --- | --- | --- |
 | What is wrong or what changed (open-ended) | `triage` | `get_failure_context` for the top failure, `list_releases` when a release coincides |
-| Why this request, trace or session failed | `get_failure_context` with the `trace_id`, `session_id` or `failure_id` | `compare_traces` against a passing execution, `get_trace` for the span tree |
+| Why this request, trace or session failed | `get_failure_context` with exactly one of `trace_id`, `session_id`, `failure_id` or `ci_run_id` | `compare_traces` against a passing execution, `get_trace` for the span tree |
 | Which errors recur and their status | `list_failures` | `update_failure_status` when the user records a decision |
 | Why something is slow | `aggregate_telemetry` grouped by operation with `duration_p95` | `get_trace` with `selection: critical_path` on a slow `trace_id` |
 | Whether a fix held | `verify_fix` with the `failure_id` and `since` (when the fix shipped) or `release` | repeat with the returned `verify_id` once `needs` is met |
@@ -54,6 +54,8 @@ saved queries and failure status in that project, never the telemetry.
   spans and 100 logs, `aggregate_telemetry` 20 groups, `get_trace` 100
   spans. `has_more` and `next_cursor` continue the same query; `coverage`
   says how much was examined, so report truncation instead of guessing.
+  A `cursor.stale` error means results ahead of the cursor changed after it
+  was issued: run the query again without the cursor.
 - **Trace detail.** For a failure, call `get_trace` with
   `selection: errors_with_ancestors`: the error spans plus their path to the
   root. Widen to `all` only when that path is not enough, and raise
@@ -67,12 +69,16 @@ saved queries and failure status in that project, never the telemetry.
 - **Filters.** `matchers` take `key`, `op` (`eq`, `neq`, `in`, `prefix`,
   `contains`, `exists`, `gt`, `gte`, `lt`, `lte`) and `value` or `values`;
   `discover_attributes` lists which keys and values exist before you filter.
+  `eq`, `neq` and `in` compare typed values: the number 1 does not match the
+  string "1", so pass a value in the type `discover_attributes` reports.
 
 ## Reading a verdict
 
 - `verify_fix.verdict` is `passed`, `failed`, `pending` or `inconclusive`.
   `pending` means the `needs` (requests and minutes) are not met yet: say
-  so and offer to check again later with the `verify_id`.
+  so and offer to check again later with the `verify_id`. `inconclusive`
+  means the evidence cannot decide: no requests after the change, or a scan
+  that did not hold the whole window after it.
 - `triage` ranks failures by `change`: `new`, `worsening`, `ongoing`,
   `recovering`, `regressed`. A `coinciding_release` is a lead, not a cause.
 - `get_failure_context.observed` is evidence (root-cause span, error spans,
@@ -81,6 +87,10 @@ saved queries and failure status in that project, never the telemetry.
   could not establish. Report the unknowns with the findings.
 - `coverage.scan_complete: false` or `truncated: true` means the answer
   rests on part of the data. Say which part.
+- `check_telemetry_health.metric_conflicts` lists metric timestamps that
+  hold contradicting measurements. `query_metric` refuses a window that
+  holds one with a `conflict` error: narrow the window or choose another
+  `series_id`.
 
 ## Record and reuse (only when asked)
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "qyl-mcp-server",
-  "version": "17.0.2",
+  "version": "17.0.3",
   "type": "module",
   "description": "qyl telemetry MCP server: analysis tools for a live collector; Bun 1.4 runtime",
   "repository": {