feat(annotate-mlflow-trace): add trace tagging and feedback skill

[V-3604]

annotate-mlflow-trace lets a coding agent tag and leave feedback on the current session's MLflow trace from the same conversation, so sessions can be labeled for evaluation without opening the UI. A single script (trace_annotate.py) exposes tag, feedback, and list.

Fixes #13

Trace selection (updated for MLflow 3.14)

tag and feedback default to the most recent trace from the current project, resolved by the mlflow.trace.working_directory metadata that mlflow.claude_code writes in 3.14, so a shared experiment does not surface another project's trace. Selection falls back to the most recent trace on older versions, and --trace-id targets a specific trace. This is the agent and CLI path for writing assessments; it complements Review Queues, which is the UI path for the same data.

Decisions on the open questions in #13

1. One skill, not two. tag-trace and feedback-trace share a workflow and a script, so they are a single annotate-mlflow-trace skill with subcommands rather than two directories.
2. skip_skill_traces.py is included here as an optional Stop-hook pre-filter so annotation turns do not log their own trace. A native skip list in stop_hook_handler is probably the better long-term home, and I can move it there if that is preferred. The skill works without the hook; the hook only suppresses the administrative trace.

Notes

• The hook detects the skill turn by the Base directory for this skill: marker Claude Code injects, so it is specific to Claude Code. Other agents can still annotate by --trace-id.
• Unit tests cover tag parsing, value coercion, and trace selection, including project scoping and fallback. Verified on MLflow 3.12 and 3.14.

[adamgurary]

@B-Step62 can we route this to someone to look at please

[V-3604]

This PR and #21 both touch the skills array in .claude-plugin/plugin.json; #21 adds a ./ prefix to those entries. Once it merges I'll rebase and re-add annotate-mlflow-trace as ./annotate-mlflow-trace so it lands in the new format.

[V-3604]

@B-Step62 Hi, I updated this to fit MLflow 3.14. Trace selection now scopes to the current project using the mlflow.trace.working_directory metadata the Claude Code integration writes, with a fallback for older versions, so it complements Review Queues instead of overlapping. The description now answers the two design questions from #13: I went with a single skill, and I left skip_skill_traces.py as an optional hook with a note that a native skip list in stop_hook_handler may be the better home. I can take either direction on the hook.

[B-Step62]

@V-3604 Do we need this script? The native CLI should support the primitive operations. We would like to avoid having many bespoke scripts.

[B-Step62]

If there is a missing primitive, let's add it to the CLI.

[V-3604]

Agreed, dropped the script. mlflow traces set-tag, log-feedback, and search already cover everything it did, including --source-type HUMAN, --rationale, and --source-id.

The one thing I thought might be missing was scoping trace selection to the current project, but that filters on the mlflow.trace.working_directory metadata, so there is nothing to add to the CLI.

SKILL.md now documents the annotation surface directly: the tag and feedback commands, how to find the trace to annotate, tagging alongside feedback so assessments stay searchable, and log-expectation for ground truth. The PR is one SKILL.md plus the three registration lines.