Miller v0.5.4

Miller v0.5.4 Release Notes

• Audience: Agents and users relying on Miller telemetry, dashboard diagnostics, and compact retry guidance to understand how tools are being used across workspaces.
• Release shape: Patch release for query-safe tool telemetry dimensions, better empty-result/error classification, and compact guidance for common search issues.
• Pinned extractor: julie-extract 2.5.1.

What Is New

• Tool calls now share a visible correlation id. Every normal MCP tool call records the same cid in the telemetry ledger and JSONL log line, making it easier to connect dashboard rows, exported telemetry, and local logs.
• Telemetry captures query-safe tool shape. Read tools now record bounded metadata such as operation/mode, format, limit bucket, input-shape flags, route/backend, and workspace selector presence without storing raw query text, target paths, snippets, or source content.
• Empty results are easier to diagnose. Search, content, patterns, inspect, trace, impact, context, and edit calls now classify common no-result cases with empty_reason metadata.
• Errors get stable categories. Tool failure telemetry now includes error_category, so future analysis can group failures without depending on raw exception text.
• Search compact output gives better retry hints. Empty search output now suggests text/content modes and comment/string-region searches when symbol search has no useful hit.
• Invalid source-region searches show examples. search(regions=...) validation errors now include compact examples for comment, doc_comment, and string_literal.
• Tool usage baseline is documented. The repository now includes the June 16 telemetry baseline and follow-up plan in docs/findings/2026-06-16-tool-usage-telemetry-baseline.md.

Verification

• Local verification before release:
  • dotnet build Miller.slnx -c Release (0 warnings / 0 errors)
  • scripts/test.sh (2267 passed)
  • scripts/test.sh scale (28 passed against the real julie-extract 2.5.1)
  • scripts/test-plugin.sh (18 passed)
  • git diff --check
• Live release verification should be recorded separately after publication in docs/findings/.