Work Summary

Work Summary

Generate a date-ranged Markdown report of coding-agent work from the session history stores that agents such as Claude Code, Codex, opencode, and agy keep on this machine. The outcome is an evidence-based digest of what was asked and what was done — a compact summary by default, a detailed report on request.

Session contract

• Stay read-only toward every history store: never modify, move, or delete session files or databases.
• Keep history content local: it may appear in the report and this conversation, but never send it to an external service. Never stage or commit a generated report.
• Report only recorded facts, with counts and quotes taken from the stores. A range without activity yields an honest "no recorded activity" report, never filler.
• Match the user's language for report prose; keep the template's English section headings as-is so the report shape stays checkable, keep dates in ISO form, and do all bucketing in the user's local timezone.

1. Resolve the range, depth, and scope

Parse the request; default to today at summary depth across all projects.

Current week, month, quarter, and year ranges start at the local calendar boundary and end now. Previous periods cover the complete local calendar period. Weekly labels and directories use the ISO week-year that matches the Monday-start week.

• Depth becomes detailed when the user asks for detail ("상세", "자세히", "detailed", "full report"); a detailed report adds the timeline and request-log sections.
• Scope narrows to one project when the user says so ("이 프로젝트만", "only this repo"); match it against each record's project or cwd field.

This step is complete when the range has concrete start and end instants in the user's local timezone and the depth and scope are fixed.

2. Mine the agent history stores

Read references/agent-history-stores.md completely before touching any store. Then, for every store present on this machine:

1. Query the cheap prompt-level index first (history.jsonl files, SQLite session tables) for in-range prompts, sessions, and projects.
2. Open full session transcripts only for evidence the report needs — assistant outcomes, files touched, commands run — and only for in-range sessions. Discover those sessions from the union of index hits and each store's cheap pre-filter (such as transcript-file mtime); the index stays authoritative for session and prompt counts, transcripts for outcomes.
3. Filter by each record's own timestamp field. Epoch units and timezones differ per store (milliseconds vs seconds, UTC records vs local-time paths); the reference pins each one — normalize before comparing.
4. Drop non-work noise: Claude Code records flagged isMeta or isSidechain, Codex developer-role messages, and agy conversations flagged is_internal.

Silently skip stores that are absent. A store that exists but cannot be read is listed as skipped in the report's Sources line instead of failing the run; a store mined without in-range records is listed as "no in-range activity", not as skipped.

This step is complete when every existing store is either mined for the range or recorded as skipped with a reason.

Natural Korean Prose

When the report is in Korean, write reader-facing text as plain, concrete Korean prose. Use direct sentences; avoid translation-like rhythm, vague AI filler, and internal method vocabulary. Preserve every mined fact, count, date, project path, code identifier, command, quote, link, and source label.

This selector is self-contained. If the runtime already exposes human-friendly-writing, it is an optional final pass over only the newly authored Korean draft before delivery or file writing. Do not install, fetch, or require it. If it is absent or unreadable, continue silently with the rules above and complete the report. After an optional pass, run the normal validation and recount totals. The pass may change wording only, never mined evidence, English section headings, dates, or report structure.

3. Compose the Markdown report

Build the report in this shape; omit the two detailed-only sections at summary depth:

# Work Summary — <range label> (<start> ~ <end>, <UTC±HH:MM>)

- Sources: <stores with activity> (no in-range activity: <none | stores>; skipped: <none | store — reason>)
- Sessions: <n> · Prompts: <n> · Projects: <n>

## Overview

<3–6 sentences: the main threads of work and their outcomes>

## By project

### <project path>

- Asked: <condensed representative requests>
- Done: <outcomes evidenced by transcripts>
- Agents: <store names with session counts>

## By agent

| Agent | Sessions | Prompts | Projects |
| --- | --- | --- | --- |

## Timeline (detailed only)

### <YYYY-MM-DD>

- <HH:MM> · <agent> · <project> — <request → outcome>

## Requests log (detailed only)

| Time | Agent | Project | Request |
| --- | --- | --- | --- |

• Condense long prompts to their intent; never paraphrase into claims the records do not support.
• An empty range keeps the header and Sources lines accurate and states "no recorded activity in this range." under ## Overview.

This step is complete when every section is backed by mined records and the totals match the collected counts.

4. Deliver the report

Reply with the full Markdown report. Write a file only when the user asks for one. If the user gives an explicit output path, use it. Otherwise create the matching directory under the current project and classify the path from the requested range syntax, not from coincidental start and end dates:

Suggest — never apply — a .work-summaries/ ignore entry in repositories that track it. Never stage or commit a report; it can quote prompts from unrelated private projects.