At Self Eval

Self-Evaluation Summary

resolve evidence -> group by business line -> deduplicate -> <=8 outcomes -> review

Author and date window

Normalize flags or plain language such as 统计1-3月的产出 and 统计上半年的工作

into (author, from, to). Explicit input wins; fill only missing values:

| User gave | Resolve to |

| --- | --- |

| from and to | use both |

| only from | to = today |

| only to | from = Jan 1 of that year |

| no range | current quarter |

Parse half-years, calendar quarters, month ranges, and rolling periods. A spoken

period without a year means the current year, except months still ahead of today,

which mean the previous year. Query through <to + 1 day> because Git's --until is exclusive.

Always exclude merge commits.

Use an explicit author when supplied; otherwise resolve it independently per repository

with git -C <root> config user.name. Never infer aliases or use the remote login as

the author. List candidates with `git -C <root> log --no-merges --since=<from>

--until=<to+1day> --format=%H%x09%an --branches HEAD <upstream refs>`, taking the

upstream refs from git -C <root> for-each-ref --format=%(upstream) refs/heads/;

passing no ref walks only the current branch, so work on an unmerged branch reads as

no activity, while --all reaches into remote branches nobody tracks. Keep the rows

whose %an equals that name exactly. Never pass --author: it matches the whole

Name <email> header, so anchored patterns silently match nothing. If no commits

match, report that together with the author names the window actually holds; do not

try spelling or language variants. When results exist, mention once that the

user can provide other author names if needed.

Project evidence

Read optional workProjects from ~/.agent-tools/config.jsonc. An entry is a path

or { "path", "prompt" }, where the prompt is free text this skill follows for that

project, such as how to label items or which commits to skip.

• No project named: current Git repository plus configured projects.
• Projects named directly: only those projects.
• "Also include" / 另外包含: add them to the default scope.

Resolve paths to Git roots and deduplicate them. Report invalid paths; a non-Git current

directory does not block other valid projects. Do not clone remote URLs without consent.

Inspect all local branches, current HEAD, user-named branches, and configured upstreams.

Unless the user requests local-only data, refresh only the configured upstream

branches, grouped into one best-effort git fetch --no-tags <remote> <branch...> per

involved remote. Fetch failure is non-fatal. Do not change the working tree or local

branch history. Deduplicate commits by hash.

Begin the result with the resolved author/window and, for multiple repositories, a

plain-language list of the repositories checked. Mention remote-update failures without

Git jargon. Do not show commit counts or a generic source line.

Optional logs

A pasted daily/weekly log or explicit file path provides business context. If neither is

supplied, optionally read dailyLog.output and log.output (an automatically recorded

AI session log; a dated markdown file, or a directory of per-day <date>.md reports,

of which read the dates inside the window; entries under log.projects may add their

own output paths, check those too) from ~/.agent-tools/config.jsonc; the

latter also captures work that never produced commits,

such as troubleshooting or research sessions. Explicit conversation input always wins;

a missing key or file is non-fatal.

Commit-backed items may be summarized directly. Keep log-only work separate for user

confirmation; include it only after confirmation, using only dates stated in the log or

by the user. Never turn undated log text into work inside the selected window.

When the user asks for a summary from a log alone, skip project scanning; when the

window has no Git evidence, fall back to the log. Either way the log becomes the

primary source: say so in the result and skip per-item confirmation, as the whole

draft is the user's own account.

If the user supplies a remote repository URL, ask before cloning it to a temporary

directory and remove it afterward unless asked to keep it. Confirm before scanning every

remote branch.

Group the work

Group related commits into business-line or module outcomes. Fold lint, formatting,

version bumps, repeated syncs, and follow-up fixes into the outcome they supported.

Drop isolated trivial housekeeping. Similar subjects alone are not duplicates; use the

repository, changed paths, and intent, and surface uncertain attribution for review.

Write the summary

Output a numbered list with no more than 8 top-level outcomes. Each item starts with the

business line/module and states a concrete result with 2-4 key points or one concise

sentence. Prefer high-impact delivery, architecture, and key fixes; condense or omit

routine work. Do not distribute the quota evenly if that hides standout results.

Mention a concentrated month only when it adds useful distribution context, never as a

productivity score. Match the user's language. Every confirmed outcome must trace to a

commit or a user-confirmed log item.

> 作者: <author> · 窗口: 2026-01-01 ~ 2026-06-30 · (自动推断, 如不对请指正)
> 项目: <repo1> · <repo2>

1. <业务线A>: <结果>; <关键点>
2. <业务线B>: <结果>

End with one reminder that this is a draft requiring review. Do not write it to a file

unless asked. Add at most one extra hint, and only when evidence is likely incomplete:

• unconfirmed log-only work exists;
• terse commits lack business context and no log was available;
• the conversation or log provides concrete evidence of unscanned projects/branches.