Nippo

Nippo

Use this skill when the user wants to turn recent Claude Code, Codex, or opencode work into a report under reports/.

Inputs

• mode: default, daily, brief, reflection, guide, report, review, insight, trend, plan, ledger
• optional days
• optional project filter
• optional source override: claude, codex, opencode, all
• optional period selector: today, yesterday, this-week, last-week, week-before-last, this-month, last-month, month-before-last

Examples:

• $nippo daily
• $nippo daily codex
• $nippo daily claude
• $nippo daily all
• $nippo review 90 codex

Workflow

1. If the current workspace is this repository, prefer cargo run -q -p nippo -- collect ... so the skill uses the checked-out implementation rather than a potentially stale globally installed nippo.
2. Otherwise prefer nippo collect ... when nippo is already installed.
3. If neither is available, stop and tell the user to install nippo.
4. Treat daily as an alias of the default daily report mode.
5. Default to --source auto. Override when the user explicitly asks for claude, codex, opencode, or all. The collector removes deterministic prompt noise by default; do not pass --include-prompt-noise during report generation.
6. For brief, save the summary output directly and stop.
7. For ledger, do NOT call nippo collect. Run nippo ledger (or cargo run -q -p nippo -- ledger) which scans reports/nippo-*.md (daily reports) for the ## Unclear points section, folds them into reports/ledger.yaml, and prints a CONVERGED / DIVERGENCE-SIGNAL / CONTINUE verdict. Relay that verdict back to the user verbatim; do not over-interpret. If the user asks for agent-config candidates, run nippo ledger --export: it writes recurring General Fix Rules to reports/ledger-export.md. Tell the user these are candidates only — a human must curate them before copying anything into CLAUDE.md / AGENTS.md.
8. For plan, do NOT call nippo collect. Find the newest reports/nippo-*.md by the date in its filename and read it. Also read reports/ledger.yaml if present and use unresolved recurring General Fix Rules as candidate material. Read docs/templates/plan-template.md and docs/reflection-theory.md, present only 1-3 experiment candidates, leave the choice / reason / first-step fields blank for the user, save to reports/plan-YYYY-MM-DD.md, and report the path.
9. For other modes, read the collected JSON and the matching template:
   • docs/templates/nippo-template.md
   • docs/templates/reflection-template.md
   • docs/templates/guide-template.md
   • docs/templates/report-template.md
   • docs/templates/review-template.md
   • docs/templates/insight-template.md
   • docs/templates/trend-template.md
   • docs/templates/plan-template.md
10. For reflection, guide, insight, and plan, also read docs/reflection-theory.md.
11. Save daily reports, including daily, to reports/nippo-YYYY-MM-DD.md. Other modes keep reports/{mode}-YYYY-MM-DD.md. Append -Nd when days > 1.
12. In daily mode, treat the freshly collected JSON as the only source of truth. Do not read an existing reports/nippo-YYYY-MM-DD.md as input; overwrite it with the new report.
13. Ground the daily report header and stats directly in meta and stats, use meta.period for the requested date range, and choose project sections from stats.projects_worked_on in order of message_count.
14. After generating a daily report (which emits an ## Unclear points section), suggest the user run /nippo ledger to fold today's stuck points into the cumulative streak signal — but do NOT run it automatically.
15. Finish with the saved path and a concise report summary. Do not suggest collector changes for values already available directly or by simple derivation from the JSON.

Mode Defaults

• default: --period today
• daily: alias of default, --period today
• brief: --period today --format summary
• reflection: --period today
• guide: --period today
• report: --days 7 --stats-only
• review: --days 90 --stats-only
• insight: --days 7
• trend: split the time window into 3 ranges and run 3 summary collections
• plan: no collection. Reads the newest reports/nippo-*.md and reports/ledger.yaml, writes reports/plan-YYYY-MM-DD.md
• ledger: no time window. Reads reports/nippo-*.md, writes reports/ledger.yaml, prints a streak verdict. --export writes recurring rules to reports/ledger-export.md

Rules

• Data collection must go through nippo collect. Do not reimplement parsing in ad-hoc scripts.
• Do not use Python for data collection.
• Use stats as-is. Do not recalculate counters in prose.
• Write reports in Japanese.
• Do not recommend books or URLs (hallucination risk). Give concept names and search keywords instead.
• For reflection and guide, also read the same-day reports/nippo-YYYY-MM-DD.md if it exists.
• If the first token is neither a mode name, a number, a source selector, nor a period selector, treat it as a project filter and run the default daily mode with --project.
• Date boundaries follow the machine's local timezone. --days 1 and daily mean the current local calendar day.
• Treat one token matching claude, codex, opencode, or all as the source selector and pass it through to --source.
• Treat one token matching today, yesterday, this-week, last-week, week-before-last, this-month, last-month, or month-before-last as a period selector and pass it through to --period. In that case do not add the mode's default --days or default --period today; the explicit --period is the sole time-window flag. For daily, use meta.period.to from the collected JSON as the output filename date instead of computing it from the execution date.
• Codex report data comes from history.jsonl, state_5.sqlite, and rollout data referenced by rollout_path. Treat logs_2.sqlite as diagnostics only.
• Codex-derived reports may have sparse assistant/tool metrics. State that explicitly instead of inventing numbers.
• For daily reports, copy meta.source, meta.total_sessions, stats.projects_worked_on, and stats.tool_frequency from the collected JSON instead of inferring them from an older report.
• In every mode that prints a date range, use meta.period.from and meta.period.to; do not recalculate the range from the execution date.
• Cover the activity-heavy projects first. Use stats.projects_worked_on order (= message_count descending) and give the top 3-5 projects their own section before collapsing anything into その他.
• If you show only a subset of decisions, explicitly state 全N件中M件を記載.
• The current Claude Code or Codex session is excluded by the collector. Do not manually subtract another session from meta or stats.
• Sanitize reference links before writing them. Do not copy malformed URL fragments with trailing Japanese text or punctuation.

Improvement Issues

After the report is complete, an Issue may be proposed when report generation repeatedly performs the same deterministic aggregation or formatting, or when one structural problem affects multiple modes. Group symptoms with the same root cause into one coherent Issue instead of proposing one field at a time. Propose at most two Issues per report.

Do not propose an Issue for values already available directly or by simple derivation:

• Total projects: the length of stats.projects_worked_on
• Reporting period: meta.period
• Top tools and main work window: render_helpers
• Prose or summarization that belongs in a report template

Before proposing, run gh issue list --repo nwiizo/nippo --state all and skip anything already covered or solvable by reading existing fields correctly. If a new Issue is warranted, present a title and body that separate verified facts, affected modes, the shared cause, and the processing that the change would eliminate. Create it with gh issue create --repo nwiizo/nippo only after the user explicitly agrees.