engineering-retro（工程复古）

Engineering Retrospective

Generate a structured, git-based engineering retrospective for a configurable time window. This is a read-only analysis — no files are modified except the optional JSON snapshot.

Arguments

/engineering-retro [TIME_WINDOW] [PATH_SCOPE]

• TIME_WINDOW (optional): 24h, 7d (default), 14d, 30d
• PATH_SCOPE (optional): restrict analysis to a subdirectory (monorepo support), e.g. services/api

Examples:

• /engineering-retro — last 7 days, full repo
• /engineering-retro 30d — last 30 days, full repo
• /engineering-retro 14d services/api — last 14 days, scoped to services/api/

Execution Steps

Step 1: Environment Detection

Detect runtime context before any analysis:

# Default branch
DEFAULT_BRANCH=$(git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's@^refs/remotes/origin/@@')
if [ -z "$DEFAULT_BRANCH" ]; then
  DEFAULT_BRANCH=$(git remote show origin 2>/dev/null | grep 'HEAD branch' | awk '{print $NF}')
fi

# System timezone
TZ_NAME=$(date +%Z)

# Time window — convert argument to --since format
# 24h → "24 hours ago", 7d → "7 days ago", 14d → "14 days ago", 30d → "30 days ago"

If DEFAULT_BRANCH detection fails, abort with an error — do not guess.

Step 2: Gather Raw Git Data

Collect commits within the time window on the detected default branch:

# All commits in window (with optional path scope)
git log origin/$DEFAULT_BRANCH --since="$SINCE" --format="%H|%aI|%aN|%s" -- $PATH_SCOPE

# Diff stats for the window
git log origin/$DEFAULT_BRANCH --since="$SINCE" --numstat --format="%H" -- $PATH_SCOPE

Capture: commit hash, author date (ISO), author name, subject line, files changed, insertions, deletions.

Step 3: Compute Aggregate Metrics

From the raw data, compute:

• Total commits in window
• Unique contributors (distinct author names)
• Files changed (unique file paths across all commits)
• Lines added (sum of insertions)
• Lines removed (sum of deletions)
• Net delta (added - removed)
• Avg commit size (total lines changed / total commits)

Step 4: Time Distribution

Analyze commit timestamps (converted to system timezone $TZ_NAME):

• Commits by day of week: Mon-Sun histogram
• Commits by hour: 0-23 histogram
• Peak day: day with most commits
• Peak hours: hours with most activity

Present as a compact text histogram.

Step 5: Session Analysis

Group commits into work sessions using a >2 hour gap as a session boundary:

1. Sort commits by author and timestamp
2. For each author, iterate chronologically — if gap between consecutive commits exceeds 2 hours, start a new session
3. Compute per-session: duration (first commit to last commit), commit count
4. Aggregate: total sessions, average session length, longest session, average commits per session

Sessions with a single commit get a default duration of 0 (point-in-time).

Step 6: Commit Type Classification

Classify each commit using conventional commit prefixes from the subject line:

For commits without conventional prefixes, apply diff heuristics:

• Primarily new files added → feature
• Primarily deletions → refactor
• Test files only → test
• Config/CI files only → chore
• Documentation files only → docs
• Otherwise → uncategorized

Report counts and percentages per category.

Step 7: Hotspot Analysis

Identify the top 10 most-modified files by number of commits touching them:

git log origin/$DEFAULT_BRANCH --since="$SINCE" --name-only --format="" -- $PATH_SCOPE | sort | uniq -c | sort -rn | head -20

Flag any file modified in >50% of total commits as a hotspot. Hotspots indicate:

• Active area of development (expected during feature work)
• Potential coupling issues (if unrelated commits keep touching the same file)
• Possible need for decomposition (if the file is large)

Step 8: PR Analysis

If the remote is GitHub (check git remote get-url origin for github.com):

# Merged PRs in window
gh pr list --state merged --base $DEFAULT_BRANCH --search "merged:>=$SINCE_DATE" --json number,title,author,mergedAt,additions,deletions,changedFiles,reviews

Compute:

• Total merged PRs
• Size distribution: S (<50 lines), M (50-200), L (200-500), XL (>500)
• Review turnaround: time from PR creation to first review (median, p90)
• Merge turnaround: time from PR creation to merge (median, p90)

If not a GitHub remote or gh is unavailable, skip this step and note it in the output.

Step 9: Focus Score

Compute the ratio of focused commits (touching 3 or fewer files) to total commits:

focus_score = commits_touching_le_3_files / total_commits

Interpretation:

• >0.8: highly focused, small incremental changes
• 0.5-0.8: moderate focus, mix of targeted and broad changes
• <0.5: broad changes dominating, may indicate large refactors or low commit discipline

Step 10: Per-Author Breakdown

For each contributor, report:

• Commit count
• Lines added / removed
• Top 3 most-touched files
• Primary commit types (from Step 6)
• Number of sessions and average session length (from Step 5)

Frame this as contributor highlights — recognition of work done, not a ranking or performance metric. Order alphabetically by author name.

Step 11: Week-over-Week Comparison

Check for a prior snapshot in .engineering-retros/:

• Find the most recent *.json file
• If it exists and covers the adjacent prior window, compute deltas:

- Commit count delta (%) - Lines changed delta (%) - Contributor count delta - Focus score delta - Category distribution shift

If no prior snapshot exists, note this is the first retrospective and skip comparison.

Step 12: Save Snapshot

Save a JSON snapshot for future comparisons:

.engineering-retros/<YYYY-MM-DD>.json

Schema:

{
  "date": "YYYY-MM-DD",
  "window": "7d",
  "path_scope": null,
  "branch": "main",
  "timezone": "PST",
  "metrics": {
    "commits": 0,
    "contributors": 0,
    "files_changed": 0,
    "lines_added": 0,
    "lines_removed": 0,
    "net_delta": 0,
    "focus_score": 0.0
  },
  "categories": {},
  "hotspots": [],
  "sessions": {
    "total": 0,
    "avg_length_minutes": 0
  },
  "authors": {},
  "pr_stats": null
}

Create the .engineering-retros/ directory if it does not exist. Ensure .engineering-retros/ is in .gitignore (add it if missing — this is the one permitted file modification).

Step 13: Generate Narrative Summary

Produce the final output in this structure:

Engineering Retrospective — [DATE_RANGE] ([TIMEZONE]) Branch: [DEFAULT_BRANCH] | Scope: [PATH_SCOPE or "full repo"]

Metrics

• Commits: N | Contributors: N | Files changed: N
• Lines: +N / -N (net: +/-N)
• Avg commit size: N lines | Focus score: N.NN

Time Patterns

• Peak day: [DAY] | Peak hours: [RANGE]
• [compact histogram]
• Sessions: N total | Avg length: Nm | Longest: Nm

Work Breakdown

• [category]: N commits (NN%)
• ...

Hotspots

• path/to/file — N commits [HOTSPOT if >50%]
• ...

Contributor Highlights

• [Author]: N commits, +N/-N lines, focused on [top files], primarily [categories]
• ...

PR Summary (if available)

• Merged: N | Size dist: S/M/L/XL | Median review turnaround: Xh

Week-over-Week (if available)

• Commits: +/-N% | Lines: +/-N% | Focus: +/-N.NN

Observations

• [2-4 bullet points identifying patterns, achievements, and areas worth attention]
• Based on data only — no speculation about intent or quality judgments about individuals

Constraints

• Read-only: no code modifications, no branch changes, no git operations that alter state
• No hardcoded timezone: always detect from date +%Z
• No hardcoded branch: always detect dynamically via git symbolic-ref or git remote show
• No individual performance judgments: author breakdown is for recognition, not evaluation
• Path scope respected: all git commands must include -- $PATH_SCOPE when a scope is provided
• Snapshot storage: .engineering-retros/ only, never .context/retros/