菜单
Pi session, trace, metrics, routing, workflow-friction, workflow-telemetry, usage, or local JSONL analysis with DuckDB. Use for aggregating or correlating Pi runtime logs. Not for adding telemetry, generic SQL/database design, or non-Pi logs.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Pi extensions and tool policy: extensions/*.ts, hooks, registerTool, promptGuidelines, registerCommand, footer/status UI, tool_result, session hooks, or subprocesses. Not for slash-command placement; use pi-command.
README, CHANGELOG, docs/, RFCs, ADRs, guides, or Markdown structure. Not for prose cleanup or archival work.
Prose cleanup for filler, hype, vague claims, repetition, detection tells, or uncited specifics. Not for Markdown architecture or archival work.
Requirements, user stories, acceptance/verification criteria, or testable outcomes. Not for PRD drafting (/prd) or Pi /goal prompts.
Use for /prd-it or to draft, refine, or review a Product Requirements Document. Not for acceptance criteria, planning, or Pi /goal prompts.
Forgejo Actions: .forgejo/workflows, forgejo-runner, labels, secrets, workflow_dispatch, CI/CD, or deployment. Not for Git remotes, repository setup, or Forgejo APIs; use git-workflow.
| name | pi-log-analytics |
| description | Pi session, trace, metrics, routing, workflow-friction, workflow-telemetry, usage, or local JSONL analysis with DuckDB. Use for aggregating or correlating Pi runtime logs. Not for adding telemetry, generic SQL/database design, or non-Pi logs. |
Use this skill to investigate existing Pi runtime data with the deterministic helper in pi/analytics/. Use logging-observability to design or add telemetry, analysis-workflow for debugging without telemetry, and database for generic database work.
JSONL is the append-only source of truth. DuckDB views and any DuckDB or Parquet files are disposable local analysis artifacts.
uv run --project pi/analytics python pi/analytics/pi_log_query.py catalog
uv run --project pi/analytics python pi/analytics/pi_log_query.py views
SELECT with a result cap:uv run --project pi/analytics python pi/analytics/pi_log_query.py query \
"SELECT event, count(*) AS n FROM metric_events GROUP BY event ORDER BY n DESC" \
--limit 50
uv run --project pi/analytics python pi/analytics/pi_log_query.py validate routing_classifier_events
--ignore-malformed only after validation when an incomplete exploratory result is acceptable. Report the omitted row count.session_entries.message, session_entries.content, and trace_events.payload as potentially sensitive. Do not select or print raw content unless the task requires it.session_entries as the canonical corpus. history_entries can overlap it; never union both without explicit session-level deduplication.route_decision_id. Use routing_decisions_joined for legacy occurrence matching..tmp/pi-log-analytics/. Never commit DuckDB, Parquet, JSONL, or copied runtime data.read_ndjson_auto() for heterogeneous trace payloads.--ignore-malformed without first validating and reporting the omitted records.