reporter

name

reporter

description

Generate structured Chinese daily, weekly, or monthly work reports from recent Codex or Claude Code memory files, ask the user to choose the report type and detail level when unspecified, and create or update recurring Codex automations after collecting the report scope, recurrence, and execution time. Use when the user asks to write, summarize, polish, or manually generate a 日报, 周报, 月报, daily report, weekly report, monthly report, work summary, or recurring work-report automation.

Work Report

Generate evidence-based Chinese daily, weekly, and monthly work reports from recent AI memory files. Invoke this Skill explicitly with $reporter when needed.

Guide a bare invocation

When the user invokes only $reporter or /reporter without a report type, detail level, or automation intent, respond with a short Chinese welcome and usage guide before asking them to choose. Use this structure:

您好，我可以帮您生成日报、周报或月报，也可以创建 Codex 周期自动化任务。

您可以直接说：
- “周报，详细”：立即生成近七天详细周报。
- “生成今天的标准日报”：立即生成当天工作日报。
- “使用工作报告生成器创建自动化：每周五 18:00 生成本周详细周报，时区 Asia/Shanghai。”

可选项：
- 报告类型：日报 / 周报 / 月报
- 详细程度：精简 / 标准 / 详细
- 自动化周期：每天 / 工作日 / 每周 / 每月
- 统计范围：当前周期 / 上一完整周期

请告诉我：现在生成报告，还是配置周期自动化？

Keep the guide concise and copy-ready. Do not collect memory or create an automation until the user replies. Do not repeat this guide when the same request already supplies enough information to proceed.

Choose the run mode

• Treat requests such as “生成今天的日报”“生成本周周报” or “生成本月月报” as manual runs. Do not create an automation.
• Treat requests such as “每天自动生成日报” or “每月自动生成月报” as automation setup.
• If the request is a bare invocation, use the welcome guide above. For other genuinely ambiguous requests, ask whether to generate now or configure a recurring schedule and include one relevant copy-ready example.

Collect preferences before execution

For an interactive manual run, collect any missing report type and detail level before reading memory or drafting. Ask one concise combined question when both are missing; offer report types 日报、周报、月报 and detail levels 精简、标准、详细. If only one preference is missing, ask only for that preference. When the user replies with a complete choice such as 周报，详细, begin collection and drafting immediately without asking them to reconfirm. Do not silently choose while the user is available to answer.

Use these remaining defaults unless the user specifies otherwise:

• Date range: use the report-type defaults below; honor an explicit calendar range or phrases such as 昨天、上周、上个月 instead.
• Output: return the report in chat; write or append to a file only when requested.

Use the active timezone to resolve calendar boundaries:

• 日报: the requested calendar day; default to today through the current time.
• 周报: the requested calendar week; default to the rolling seven days ending now. Interpret 本周 and 上周 as calendar weeks.
• 月报: the requested calendar month; default to the current calendar month through the current time. Interpret 上个月 as the complete previous calendar month.

Offer these detail levels:

• 精简: 20–45 Chinese characters per item, one outcome-focused sentence.
• 标准: 40–90 Chinese characters per item, usually covering action, result, and status.
• 详细: 80–150 Chinese characters per item, optionally including key difficulty, blocker, or next step.

If an unattended automation run has no configured report type or level, infer the type from an unambiguous task name or prompt; otherwise use 周报 and 标准 and state both fallbacks in the result. Do not pause an unattended run to ask a question.

Collect recent memory

1. Prefer paths explicitly supplied by the user.
2. Otherwise run scripts/collect_report_context.py from this Skill directory with --format json and a discovery window that covers the selected report: normally --days 1 for a current-day report, --days 7 for a rolling weekly report, or the inclusive calendar-day count up to 31 for one monthly period. Add a small discovery buffer when the exact boundary may be missed by file modification time. Use --source PATH repeatedly for extra files or directories. Prefer JSON for agent consumption; use Markdown only for human inspection. The script only builds a candidate bundle; it does not decide which claims belong in the report.
3. Inspect the returned source paths and warnings. Read a source directly if it was truncated or if exact context is needed.
4. Filter content by dates written inside the memory, session, or work item. File modification time is only a discovery signal and is not proof that every entry belongs to the requested report range.

The collector supports:

• Codex automation memory under ~/.codex/automations/*/memory.md;
• Codex memory indexes and recent rollout summaries under ~/.codex/memories/;
• Claude Code auto memory under ~/.claude/projects/<project-key>/memory/*.md;
• Claude Code global or project-local MEMORY.md, memory.md, CLAUDE.md, .claude/memory/*.md, and .claude/memories/*.md;
• user-provided memory files or bounded directories.

Honor CLAUDE_CONFIG_DIR when it is set. Use --claude-root PATH for a custom Claude Code data directory and --project-root PATH for project-local files. For a project-specific report, also pass --claude-project PROJECT_PATH_OR_KEY; repeat it to include multiple Claude projects. Omit it only when the report should aggregate all Claude projects.

Claude Code auto memory commonly uses an index file such as memory/MEMORY.md plus sibling topic files with YAML frontmatter. Treat both as Markdown evidence. Treat CLAUDE.md as background context only; it must not independently support a report item. Use file modification time only to discover recent candidates; use dates inside the topic content to decide whether a work item belongs in the report. If Claude files are found but all are older than the requested window, report that clearly instead of claiming incompatibility.

Do not recursively scan an entire home directory or workspace. Do not expose credentials, tokens, private keys, or unrelated personal data found in memory.

Build the report

Read references/format-spec.md before drafting.

1. Establish the report type, exact date range, and timezone. Display explicit dates in the output; do not rely only on relative labels such as 本周.
2. Extract work items supported by memory in that range.
3. Merge duplicate mentions of the same task across sources.
4. Group items by stable project, product, or work domain. Avoid one-item micro-sections when two items naturally belong together.
5. Rewrite command logs and conversational fragments as outcome-oriented work statements.
6. Preserve technical names such as TokenService, CoreDNS, model names, APIs, and deployment modes when they help identify the work.
7. Preserve supported progress percentages, throughput, resource specifications, failure rates, and other decision-useful measurements. Never invent or extrapolate numbers.
8. Mark status accurately: completed, in progress, blocked, or pending verification. Never turn an attempted or failing task into a completed result.
9. Use the compact, paste-ready layout in references/format-spec.md by default. Match an existing document or user-provided sample when requested.
10. Generate the final report directly. Do not include the source inventory, hidden reasoning, or raw memory excerpts unless requested.

If evidence is insufficient, say which dates or sources are missing and produce only the supported portion. Never invent work to make the report look complete.

Write to an existing report

Only edit a report file when the user asks. Preserve earlier daily, weekly, or monthly sections and the file's established heading/list style. Append a new report-type and date block when absent. If the same type and date block already exists, show or apply a scoped replacement for that block rather than duplicating it.

Configure Codex automation

When the user requests scheduling:

If the user only asks what automation is supported, explain the available fields and provide this copy-ready example before collecting values: 使用工作报告生成器创建自动化：每周五 18:00 生成本周详细周报，时区 Asia/Shanghai。

1. Ask whether to create a new automation or update an existing one when the request does not make that clear.
2. Ask which report to generate: 日报、周报、月报. Keep report type separate from automation recurrence; for example, a user may generate a weekly report every Monday.
3. Ask for the report coverage policy when it is not implied: 截至触发时的当前周期 or 上一完整周期. State the resulting date semantics before creation.
4. Ask for the repeat cycle. Support at least 每天, 工作日, 每周, and 每月; accept a custom cycle only when the automation interface supports it.
5. Ask for the fields required by that cycle:
   • 每天: execution time;
   • 工作日: execution time;
   • 每周: one or more weekdays and execution time;
   • 每月: day of month and execution time.
6. Ask for timezone and report detail level. Also collect explicit memory paths and an output path when the user needs them. If timezone is omitted, use the active local timezone and state it before creation.
7. Summarize the human-readable configuration and obtain confirmation before creating a new recurring task. Do not expose a raw recurrence rule.
8. Inspect $CODEX_HOME/automations/*/automation.toml for a matching report type, name, or prompt. Prefer updating the matching automation over creating a duplicate.
9. Call the Codex automation_update tool; do not hand-edit automation.toml. Use the Codex list_projects tool to resolve the target project when a project ID is required.
10. Make the automation prompt self-contained. Include the $reporter invocation, report type, exact date-range policy, detail level, timezone, Codex and Claude Code memory roots or explicit paths, output destination when requested, and the instruction to avoid fabrication.
11. In an unattended run, do not ask follow-up questions. Use configured values, apply the documented fallbacks, and report missing evidence explicitly.
12. After the tool succeeds, confirm the task name, enabled state, report type, date-range policy, human-readable repeat cycle, execution time, timezone, memory scope, detail level, and output destination.

Creating this Skill does not itself create a recurring automation. Only create one after the user requests scheduling and supplies the required schedule.

Validation

Read references/script-template.md before modifying or validating the collector. Run the structure and compilation checks plus a normal Codex-memory path, a Claude project-scoped path, a Markdown-fence path, and an error path.