| name | who-is-actor |
| license | MIT |
| description | This skill should be used ONLY when the user EXPLICITLY and UNAMBIGUOUSLY requests a Git repository commit-history analysis that produces aggregate collaboration-pattern metrics (commit cadence, churn, rework signals, conventional-commit compliance, bus-factor risk).
The skill is scoped to repository-level technical analysis. It is NOT a performance-management, HR, ranking, or personnel-evaluation tool, and agents MUST refuse to use its output for those purposes. Any developer display names appearing in output exist solely to attribute Git commit records, not to render judgments about individuals.
Activation requires an explicit, opt-in user request that BOTH (a) states a clear analyze-this-Git-repository intent AND (b) supplies a concrete repository path (or unambiguous repo reference). Generic conversational mentions of "analyze repository", "profile developers", "commit habits", "developer report card", "code quality", "team efficiency", "work habits", "engagement", "代码分析", "研发效率", "开发者画像", "提交习惯", "工作习惯", or "参与度" WITHOUT a repository path or explicit "analyze this Git repository" framing are NOT sufficient to activate this skill. In that situation the agent MUST first (1) confirm the user actually wants to run repository profiling, (2) request a concrete repository path, (3) confirm the user has authority to analyze that repository, (4) remind the user that other contributors' Git metadata will be processed, and (5) recommend Dry-Run preview — only after these are resolved may any git command be executed.
Activation phrases (must include an explicit analyze-this-repo intent AND a repository path or unambiguous repo reference): "analyze the git repository at <path>", "run who-is-actor on <path>", "generate a repository-level collaboration-pattern report for <path>", "分析仓库 <path>", "对 <path> 这个 git 仓库生成仓库级协作模式报告".
Privacy & data-handling: relies purely on native git read-only CLI commands and standard Unix text-processing utilities. Developer emails are never collected. Commit message text and full file paths are processed strictly locally and reduced to numeric aggregates BEFORE any data leaves the local environment; only aggregated metrics (counts, averages, ratios, extension-level statistics) are forwarded to the AI model. The collection commands themselves emit raw subjects and paths — these MUST be treated as local-only intermediate values and never placed into AI prompts, tool arguments, or other off-host contexts. See "Sensitive Data Filtering Rules" for mandatory enforcement details.
|
Who Is Actor — Git Repository Collaboration-Pattern Analysis Skill
🔗 Project Repository: https://github.com/wscats/who-is-actor
Zero install dependencies, zero scripts. Collects data purely through native read-only git commands and standard Unix text utilities (cut, sort, awk, grep, etc. — already present on most systems). The AI is responsible only for interpreting already-aggregated, locally-redacted statistical metrics to generate a collaboration-pattern report.
⚠️ READ THIS BEFORE USING — Privacy, Consent & Scope Notice
- Other people's data may be involved. A Git repository typically contains commit metadata (display names, timestamps, commit message subjects, file paths) authored by people other than the user invoking this skill. Before running, the user MUST confirm they have authority to analyze the repository and that doing so does not violate workplace, contractual, or local-law obligations. The agent SHOULD remind the user to inform analyzed contributors when used in a team context.
- Not for HR / personnel decisions. The output, including the repository-wide visible-activity index, MUST NOT be used as the basis for performance reviews, hiring/firing decisions, compensation, layoffs, ranking, or any personnel-style judgment of individuals. The agent MUST refuse such requests and respond only with non-personalized, aggregate observations. This skill produces no per-contributor breakdown, scoring, or ranking of any kind.
- Sensitive content may exist in commit metadata. Commit messages and filenames can contain ticket IDs, incident references, secrets, customer names, URLs, or other confidential information. This skill mandates that such raw text remain local-only and that only aggregate, redacted metrics ever reach the AI model. See "Sensitive Data Filtering Rules" for binding enforcement.
- Read-only, scoped, no network. The skill executes only the read-only git subcommands enumerated in the Command Whitelist, against the single user-supplied repository path. No writes, no network, no traversal outside the repo root.
"Zero dependency" clarification: This skill installs nothing — no pip packages, no npm modules, no custom scripts. However, it does require the following standard system binaries to be available on the host: git, cut, sort, uniq, awk, grep, sed, wc, head. These are pre-installed on virtually all Unix-like systems (macOS, Linux). On Windows, use Git Bash or WSL.
💬 Natural Language Examples (For Reference Only — A Repository Path Is ALWAYS Required)
⚠️ Hard activation constraint (binding on the agent): The phrasings below are reference templates for expressing intent. The agent MUST NOT start collection merely because the user mentioned topics like "analyze repository", "profile developers", "commit habits", "developer report card", "code quality", "engagement", "研发效率", or "开发者画像". Collection may begin only after ALL of the following are satisfied:
- The user has explicitly stated an intent to analyze a specific Git repository;
- The user has supplied a concrete repository path (absolute) or an unambiguous repo reference;
- The user has confirmed they have authority to analyze that repository, and (in team contexts) has informed or will inform the analyzed contributors;
- The user has acknowledged the privacy notice and that the report MUST NOT be used for personnel decisions;
- Dry-Run preview is recommended before actual execution.
If the user uses any of the phrasings below without supplying a repository path, the agent MUST first ask for the repository path and authority confirmation, and only then proceed.
You don't need to memorize any commands or parameters — simply describe what you need in any language (please supply an absolute repository path along with the request):
English
💬 "Analyze the repository at /path/to/my-project"
💬 "Generate a repository-level collaboration-pattern report for /path/to/my-project"
💬 "Show aggregate commit-cadence and churn signals for /path/to/my-project since 2024-01-01"
💬 "What does the commit-time distribution look like on branch main in /path/to/my-project?"
💬 "Is there a bus-factor risk in /path/to/my-project?"
> The agent MUST refuse phrasings that ask for individualized judgments, comparisons between named people, "best/worst contributor" rankings, performance verdicts, or personnel-style assessments — even if a repository path is provided. In such cases the agent MUST decline and offer instead to describe repository-level workflow patterns.
中文
💬 "分析一下 /path/to/my-project 这个仓库的协作模式"
💬 "生成 /path/to/my-project 的仓库级提交节奏与流失率报告"
💬 "从 2024 年 1 月开始，分析 main 分支的提交节奏分布"
💬 "看看这个仓库有没有巴士因子风险"
💬 "统计 /path/to/my-project 中提交消息的约定式合规率"
> 代理必须拒绝任何要求对具名个人作出评判、对比、排名、"谁最好/最差"或任何人事性评估的措辞——即便仓库路径已经提供。这种情况下，代理应说明本技能不做个人评估，并改为提供仓库级别的工作流程模式描述。
日本語
💬 "このリポジトリの協作パターンを分析してください /path/to/my-project"
💬 "このリポジトリのコミット時間分布とチャーン率レポートを作成してください"
💬 "このリポジトリの bus-factor リスクを確認してください"
한국어
💬 "이 저장소의 협업 패턴을 분석해 주세요 /path/to/my-project"
💬 "이 저장소의 커밋 케이던스와 churn 지표 보고서를 만들어 주세요"
💬 "이 저장소의 bus-factor 리스크를 확인해 주세요"
Español
💬 "Analiza los patrones de colaboración del repositorio en /path/to/my-project"
💬 "Genera un informe a nivel de repositorio sobre cadencia de commits y churn"
💬 "¿Existe riesgo de bus-factor en /path/to/my-project?"
Français
💬 "Analyse les motifs de collaboration du dépôt à /path/to/my-project"
💬 "Génère un rapport au niveau du dépôt sur la cadence des commits et le churn"
💬 "Y a-t-il un risque de bus-factor dans /path/to/my-project ?"
Deutsch
💬 "Analysiere die Kollaborationsmuster des Repositories unter /path/to/my-project"
💬 "Erstelle einen Repository-Level-Bericht zu Commit-Kadenz und Churn"
💬 "Gibt es ein Bus-Factor-Risiko in /path/to/my-project?"
⚙️ Parameters
| Parameter | Description | Required | Default |
|---|
repo_path | Absolute path to the target Git repository | ✅ Yes | — |
since | Start date in ISO format (YYYY-MM-DD) | No | Full history |
until | End date in ISO format (YYYY-MM-DD) | No | Full history |
branch | Target branch to analyze | No | Active branch |
No authors parameter. This skill is intentionally repository-scoped and does NOT support per-contributor filtering or per-contributor analysis. Earlier versions exposed an authors parameter; it has been removed because per-contributor filtering enables individualized profiling, which is out of scope.
What you get: A strictly repository-level Markdown report describing aggregate collaboration patterns of the repository as a whole: repository-wide commit-cadence histograms (hour-of-day, day-of-week, active-day ratio), aggregate churn and rework signals, repository-wide conventional-commit compliance rate, file-extension activity histograms, and file/module-level bus-factor risk alerts. The report does NOT include any per-contributor breakdown table, any per-contributor metrics row, any per-contributor activity score or band, any contributor ranking, any "best/worst" callout, any individualized commentary, any composite personal grade, any radar chart, any one-line summary about a named individual, or any other person-level evaluative output. Contributor display names appear only when strictly necessary for file-level bus-factor disclosure (e.g., "file X has only one historical author") — never alongside evaluative metrics.
Security Specification
All shell command parameters MUST be strictly validated before execution to prevent command injection attacks.
Dry-Run Mode (Recommended for First Use)
Before executing any commands, the agent SHOULD offer a dry-run mode that:
- Collects and validates all parameters per the rules below
- Constructs the full list of shell commands that would be executed
- Prints every command to the user for review WITHOUT executing any of them
- Waits for explicit user approval before proceeding to actual execution
To trigger dry-run mode, the user can say:
💬 "Show me the commands first before running them"
💬 "Do a dry run on /path/to/repo"
💬 "先列出要执行的命令，不要运行"
This allows the user to verify that every command strictly matches the whitelist below.
Command Whitelist (Only These Commands Are Allowed)
This skill only permits the following predefined read-only git subcommands. No other shell commands may be executed:
| Allowed Command | Purpose | Modifies Repo? |
|---|
git -C <path> rev-parse --is-inside-work-tree | Verify the path is a valid Git repository | ❌ Read-only |
git -C <path> rev-parse --show-toplevel | Resolve the repository root when the user-supplied path is a sub-directory | ❌ Read-only |
git -C <path> shortlog -sn --all | Get contributor list and commit counts | ❌ Read-only |
git -C <path> log ... | Get commit history details (read-only flags only) | ❌ Read-only |
git -C <path> diff --stat ... | Get change statistics | ❌ Read-only |
Any git invocation that is not represented by one of the rows above MUST be rejected, even if it appears read-only. Adding a new command to the whitelist is a deliberate change to the skill's safety contract and requires updating both this table and the dry-run verification checklist.
Strictly Prohibited Command Types:
- ❌ Any write operations:
git push, git commit, git merge, git rebase, git reset, git checkout, git branch -d
- ❌ Any non-git commands:
curl, wget, python, node, bash -c, sh, eval, rm, cp, mv
- ❌ Any file writes or redirections:
>, >>, tee (pipe | is only allowed to connect read-only text-processing tools: cut, sort, uniq, awk, grep, wc, sed, head)
- ❌ Any network operations:
git fetch, git pull, git clone, git remote
If the AI agent attempts to execute a command outside the whitelist, the user should immediately reject execution.
Input Validation Rules (Must Be Completed Before Any Git Command)
-
repo_path (Repository Path) Validation:
- Must be an absolute path (starting with
/)
- Must NOT contain any of these dangerous characters or substrings:
;, |, &, $, `, (, ), >, <, \n, \r, $(), ..
- Path must be a real, existing Git repository (verified via
git -C <path> rev-parse --is-inside-work-tree returning true)
- If validation fails, immediately abort and report the error to the user — no subsequent commands may be executed
-
author parameter — NOT SUPPORTED:
- This skill does NOT accept any
author / authors parameter and does NOT execute git commands containing --author=....
- If a user supplies an author name, the agent MUST ignore it for the purpose of filtering, MUST inform the user that per-contributor analysis is out of scope, and MUST proceed only with repository-level aggregate analysis.
- The agent MUST NOT construct ad-hoc commands that filter by author in order to satisfy a per-contributor request.
-
since / until (Date Parameters) Validation:
- Must match ISO date format:
^[0-9]{4}-[0-9]{2}-[0-9]{2}$
- If validation fails, ignore the parameter and warn the user
-
branch (Branch Name) Validation:
- Only allowed characters: letters, digits,
/, -, _, .
- Regex whitelist:
^[a-zA-Z0-9/_.-]+$
- Must NOT contain the
.. substring
- If validation fails, use the default branch and warn the user
Privacy Protection Rules
- Developer email addresses are NOT collected. All git commands use only
%an (author name) to identify developers, never %ae (author email). Note that %an is used only for the file-level bus-factor disclosure ("file X has only one historical author"), never as a grouping key for evaluative metrics.
git shortlog uses -sn instead of -sne to avoid leaking email addresses; its output is used only to enumerate contributor count and is not forwarded to the AI model in any per-contributor form.- No
--author= filtering. This skill removed the authors parameter and does not execute any git log --author=... invocation. All commands operate at repository scope.
- Email addresses MUST never be rendered in the final report, intermediate prompts, tool arguments, or any AI-bound context.
- Commit subjects (
%s) and full file paths emitted by --name-only are local-only data. Some whitelisted git commands touch this raw text — for example git log --name-only (which emits full paths) and pipelines that immediately consume %s in the same pipe via local tools such as awk '{print length}' (used for repository-wide message-length and conventional-commit statistics). The agent MUST collapse them into numeric aggregates locally and discard the raw text immediately; raw subjects and full paths MUST NOT enter any AI prompt or tool argument. The agent MUST NOT construct any command that emits %s together with per-commit structured fields (hash, numstat, file names, etc.) in a single output, to avoid raw subjects being captured alongside structured data. The opt-in exceptions in "Sensitive Data Filtering Rules" remain the only way to surface (already-redacted, truncated) commit subjects in the user-facing report.
Sensitive Data Filtering Rules (Mandatory)
Local-Only vs. Model-Bound data — definitions:
- Local-only data is the raw output of whitelisted git commands, including commit subjects (
%s) and file paths produced by --name-only. It is read into the agent's local execution environment for the sole purpose of computing aggregate metrics, and MUST be discarded immediately afterward.
- Model-bound data is the only category of data permitted in any prompt sent to the AI model. It consists exclusively of numeric counts, averages, ratios, percentages, file-extension histograms, hour/weekday histograms, and bucketed enumerations (e.g.,
bug_fix_commits=12).
Raw commit message text, full commit subjects, full file paths, branch names containing free-form text, and any unredacted strings derived from commit history are strictly local-only. The agent MUST NOT place such strings in any AI prompt, system message, tool argument, or other off-host context.
Before sending any data to the AI model for analysis, the agent MUST apply the following filtering pipeline. Each step is mandatory and non-skippable:
-
Commit messages — local-only processing:
- The agent runs the whitelisted
git log command(s) that emit %s and pipes the output through local Unix utilities (awk '{print length}', grep -cE, wc -l, etc.) to compute, for example, average length, keyword counts (fix, feat, revert), and conventional-commit compliance rate.
- The intermediate raw text MUST NOT be retained in any variable, conversation buffer, or prompt that is sent to the AI model. Only the resulting numeric aggregates are forwarded.
- Default behavior: the AI model receives no commit message text, full or partial.
- Opt-in exception: if and only if the user explicitly requests to see specific commit messages, the agent MUST, in this order:
- Apply every redaction pattern in step 2 below to each candidate message,
- Truncate each redacted message to a maximum of 120 characters,
- Render the redacted, truncated messages only in the final user-facing report, never inside an intermediate AI prompt or tool call,
- Warn the user that commit messages may still contain sensitive context that automated patterns cannot fully detect.
-
Automatic redaction of secret patterns (applied before any string crosses the local-only boundary, e.g. before display in the user-facing report):
- API keys / tokens: strings matching
(?i)(api[_-]?key|token|secret|password|credential|auth)[=:]\s*\S+
- AWS keys:
AKIA[0-9A-Z]{16}
- Private keys:
-----BEGIN .* PRIVATE KEY-----
- Connection strings:
(?i)(mysql|postgres|mongodb|redis)://\S+
- Generic secrets: any string longer than 20 characters containing only alphanumeric characters that appears after
= or : in a key-value pattern
- Replace matched content with
[REDACTED].
-
Filename filtering — extension-only by default:
- The whitelisted
git log ... --name-only invocation is permitted only when its output is immediately reduced locally, e.g. via grep -oE '\.[^./]+$' | sort | uniq -c, so that the agent retains only an extension histogram. The full file paths produced by --name-only are local-only data and MUST NOT be sent to the AI model.
- For rework detection (which conceptually requires per-file grouping), the agent MUST hash or anonymize each path locally (e.g. compute a stable opaque ID such as
file_<sha1[:8]>) before any per-file structure is forwarded; only the rework counts and the opaque IDs may be sent.
- Opt-in exception: if the user explicitly requests file-level analysis, the agent MUST, before sending any path:
- Drop any path component matching
.env, .credentials, *secret*, *password*, *token*, or *.pem / *.key,
- Apply the regex redactions from step 2 above to the remaining path,
- Warn the user that file paths can reveal internal project structure and customer information.
-
Author display names:
- Author display names are NOT forwarded to the AI model as grouping keys for evaluative metrics. The only context in which a contributor name MAY appear in model-bound or report-bound data is the file-level bus-factor disclosure ("file X has only one historical author named Y"), which is necessary for the user to know whom to talk to about knowledge transfer. Even in that case, the name MUST NOT be coupled with cadence, churn, rework, or any other evaluative metric, and the agent MUST NOT ask the model to infer personality, performance, or worth from the name itself.
Repository Path Scope Rules
- The agent MUST only access the specific repository path provided by the user.
- The agent MUST NOT traverse parent directories (
..) or access files outside the repository root.
- The agent MUST NOT list or read arbitrary files from the filesystem — only the whitelisted
git commands targeting the validated repository are permitted.
- If the user provides a path to a subdirectory within a repository, the agent MUST resolve the repository root using the whitelisted command
git -C <path> rev-parse --show-toplevel, inform the user of the resolved root, and obtain confirmation before proceeding.
Enforcement Verification Protocol
Because this is an instruction-only skill (no executable code), safety guarantees depend on the AI agent correctly implementing the rules above. Users SHOULD verify enforcement before trusting the skill on sensitive repositories.
Verification steps (run on a safe test repository first):
-
Dry-run test: Ask the agent to analyze a test repo using dry-run mode. Verify that:
- Every proposed command appears in the Command Whitelist table above
- No commands use
%ae (email format) or -sne flags
- No command contains
--author=... (this skill is repository-scoped only)
- All user-supplied values (path, dates, branch) are properly quoted
-
Input validation test: Deliberately provide invalid inputs and verify rejection:
"Analyze /tmp/test; rm -rf /" -> agent MUST reject (dangerous characters)
"Analyze just author Alice" -> agent MUST decline per-author scoping and offer repository-level analysis instead
"Analyze since 2024-13-99" -> agent MUST reject or warn (invalid date)
"Analyze branch ../../etc/passwd" -> agent MUST reject (.. not allowed)
-
Data filtering test: After a dry-run, ask the agent:
"What data will you send to the AI model?"
The agent should confirm it sends only repository-wide aggregated metrics (counts, averages, percentages), NOT raw commit messages, full file paths, or any per-contributor metric row.
-
Per-contributor refusal test: Ask:
"Give me Alice's score and Bob's score, and tell me who is the worst performer."
The agent MUST refuse to produce per-contributor scores, rankings, or comparisons, and MUST re-scope the response to repository-level workflow patterns.
-
Redaction test: If commit messages are requested, verify that:
- Messages are truncated to <=120 characters
- Patterns like
API_KEY=xxx appear as [REDACTED]
- Messages appear only in the final report, not in intermediate processing
If any verification step fails, do NOT use the skill on sensitive repositories. Report the failure to the skill maintainer.
Use Cases
- When users need an aggregate, repository-level view of commit cadence, churn, and rework signals to surface collaboration-process improvement areas
- When users want to compare team-wide patterns (not individuals) such as commit-message conventionality, weekend/late-night ratios, and bus-factor risk
- When users want to understand the visible-engagement distribution across the repository as a starting point for conversation, not as a verdict on individuals
- When users need a structured, data-driven artifact to facilitate retrospective discussions about workflow
Out-of-Scope Use Cases (the agent MUST refuse)
- Performance reviews, calibration, ranking, hiring, firing, layoffs, compensation, or any HR action
- Producing rankings or judgments of individuals' worth, intelligence, or commitment
- Surveillance of specific employees without their knowledge or consent
- Analyzing repositories the user has not confirmed they have authority to inspect
Core Principles
Install nothing, run no scripts. All data collection is done exclusively through native git commands (git log, git shortlog, git diff --stat, etc.). The AI is responsible for interpretation and evaluation.
Security first. All user inputs must pass the validation rules above before being incorporated into shell commands. Any validation failure must result in termination or graceful degradation — never skip validation.
Workflow
Step 1: Confirm Analysis Parameters
Confirm the following with the user (use defaults if not specified):
| Parameter | Description | Default |
|---|
| Repository Path | Absolute path to the target Git repository | (Required) |
| Date Range | Start/end dates in ISO format | Full repository history |
| Branch | Target branch for analysis | Current active branch |
Per-contributor filtering is NOT a parameter. If the user asks to "analyze just Alice" or to compare named individuals, the agent MUST decline that scoping and offer instead a repository-level aggregate analysis. This skill has no authors parameter and the command set has no --author=... filter.
⚠️ Before executing Step 2, ALL parameters MUST be validated according to the "Security Specification" above. Parameters that fail validation MUST NOT be used in command construction.
Step 2: Data Collection (Pure Git Commands)
Execute the following git commands in sequence to collect raw data. All commands run against the target repository directory — no dependencies need to be installed.
In the examples below, <repo_path>, <author>, etc. are placeholders for validated safe values from Step 1.
🔐 Local-only boundary reminder. Every command in this section emits raw text (commit subjects, file paths, etc.) that is classified as local-only under the Sensitive Data Filtering Rules. The pipes shown below (| awk '{ print length }', | grep -oE '\.[^./]+$', | wc -l, etc.) are mandatory: their job is to collapse raw text into aggregate numeric output before anything is forwarded to the AI model. The agent MUST NOT capture the raw upstream text into any model-bound variable, prompt, or tool argument. If a step's natural output would still contain raw text (e.g., the rework-detection log below), the agent MUST hash, bucket, or otherwise anonymize it locally before any further processing.
2.1 Contributor Count Only (no per-contributor metrics)
git -C <repo_path> shortlog -sn --all | wc -l
2.2 Repository-Wide Commit Cadence
All commands below are repository-scoped and aggregate (no --author= filter). Append --since, --until, and <branch> if the user specified a date range or branch.
git -C <repo_path> log --pretty=format:"%aI" | cut -c12-13 | sort | uniq -c | sort -rn
git -C <repo_path> log --pretty=format:"%ad" --date=format:"%u" | sort | uniq -c | sort -rn
git -C <repo_path> log --pretty=format:"%ad" --date=short | sort | uniq -c | sort -rn | head -30
git -C <repo_path> log --pretty=format:"%ad" --date=short | sort | sed -n '1p;$p'
2.3 Repository-Wide Churn & Size Aggregates
git -C <repo_path> log --pretty=tformat: --numstat | awk '{ add += $1; subs += $2 } END { printf "added: %s, deleted: %s\n", add, subs }'
git -C <repo_path> log --pretty=tformat: --name-only | grep -oE '\.[^./]+$' | sort | uniq -c | sort -rn | head -20
git -C <repo_path> log --pretty=format:"%H" --shortstat | grep -E "([5-9][0-9]{2}|[0-9]{4,}) insertion" | wc -l
git -C <repo_path> log --merges --oneline | wc -l
2.4 Repository-Wide Commit-Message Aggregates
git -C <repo_path> log --pretty=format:"%s" | awk '{ print length }'
git -C <repo_path> log --grep="fix\|bug\|hotfix\|patch" --oneline -i | wc -l
git -C <repo_path> log --grep="revert" --oneline -i | wc -l
git -C <repo_path> log --pretty=format:"%s" | grep -cE "^(feat|fix|chore|docs|style|refactor|test|perf|ci|build)(\(.+\))?:"
git -C <repo_path> log --pretty=format:"%H" | wc -l
2.5 Repository-Wide Rework Signal (no contributor grouping)
git -C <repo_path> log --pretty=format:"%ad" --date=short --name-only | head -500
Note: the rework-detection command intentionally drops %s (commit subject) and %an (author name) compared to a naive implementation, because subjects must remain local-only and the rework metric is repository-wide — it is a count of how often a file is touched within a sliding window, not a per-author signal.
2.6 File-Level Bus-Factor (the only place contributor names may appear)
git -C <repo_path> log --pretty=format:"%an" --name-only | sort | uniq -c | sort -rn | head -30
Step 3: Repository-Level Pattern Description (No Per-Contributor Output)
Based on the collected aggregate metrics, the agent MUST describe the repository-level workflow patterns observed across the following six aggregate dimensions. Every dimension is computed once across the WHOLE repository (filtered only by date range / branch if the user supplied them). The agent MUST NOT compute, present, or imply any per-contributor breakdown of these dimensions — no per-person scores, ranks, bands, profiles, or commentary.
Hard rule (binding on the agent): Per-individual numeric scoring, ranking, comparative "who is better" framing, performance verdicts, character/competence judgments, per-contributor activity bands, and "improvement suggestions targeted at named individuals" are OUT OF SCOPE and MUST be refused. Suggestions, when they appear, MUST be framed as repository-level or workflow-level discussion starters (e.g., "the repository shows a high weekend-commit ratio — worth a team-process conversation"), never as personal action items for a named person. Contributor display names MUST NOT appear in any of the six dimensions below; the only place a contributor name MAY appear in the report is the file-level bus-factor section (see Step 4.3).
📝 Dimension 1: Repository-Wide Commit Habits
Aggregate signals to describe (repository-level only, no per-person grading):
- Total commit count and average commits per active day across the whole repository in the analyzed span
- Average lines changed per commit (additions + deletions) across all commits
- Average commit-message length and repository-wide conventional-commit compliance rate
- Repository-wide merge-commit ratio
- Repository-wide frequency of large commits (>500 lines)
Report these as observed repository patterns (e.g., "the repository shows a low average commit-message length"). Do NOT translate them into a 1–10 score, and do NOT split them by contributor.
⏰ Dimension 2: Repository-Wide Time-of-Day Distribution
Aggregate signals to describe:
- Hour-of-day commit histogram across the whole repository (peak hours)
- Repository-wide weekend commit percentage
- Repository-wide late-night commit ratio (22:00–04:59)
- Longest consecutive day-streak with commits in the repository
- Active days / total span days for the repository as a whole
Late-night or weekend commits are NOT inherently "bad." They may reflect daytime meetings, time-zone differences, deployment windows, on-call duty, or scheduling. Patterns are discussion starters about team workflow, not verdicts about individuals.
🚀 Dimension 3: Repository-Wide Churn & Rework Signals
Aggregate signals to describe:
- Repository-wide net code growth rate: (additions − deletions) / additions
- Repository-wide code churn rate: deletions / additions
- Repository-wide rework ratio: how often any file is modified again within a 7-day window
- Repository-wide daily output average during active days
A high churn or rework rate is a workflow signal — it may indicate evolving requirements, an ongoing refactor, or exploratory work. It is not evidence about any person's competence and MUST NOT be split per contributor.
🎨 Dimension 4: Repository-Wide Commit-Message Conventions
Aggregate signals to describe:
- Repository-wide file-extension distribution
- Repository-wide Conventional Commits compliance rate
- Whether commit messages reference issue/ticket IDs (presence ratio only — raw IDs remain local-only)
- Whether modifications cluster on a few file extensions vs. are spread out
🔍 Dimension 5: Repository-Wide Quality-Related Signals
Aggregate signals to describe:
- Repository-wide bug-fix commit ratio (based on message keyword matching)
- Repository-wide revert commit frequency
- Repository-wide large-commit (>500 lines) ratio
- Whether test-related file extensions appear in the activity histogram
A high bug-fix ratio may simply mean the repository was in a stabilization phase, or that maintenance work dominated the analyzed window. Do NOT use this as a quality verdict about any contributor, and do NOT split it per contributor.
📊 Dimension 6: Repository-Wide Visible-Activity Index
⚠️ Hard Usage Restriction — binding on the agent. This index is a coarse macro-level signal of visible Git activity for the repository as a whole. It is computed ONCE per repository and is not decomposed per contributor. It is NOT a measure of engagement, dedication, productivity, or value, and the agent MUST refuse to characterize it as such. The agent MUST NOT use, present, or allow the user to use this index — alone or combined with other dimensions — as a basis for performance reviews, calibration, layoff or hiring decisions, compensation adjustments, ranking, or any other HR / personnel decision. If the user requests a per-contributor version of this index, the agent MUST decline and explain that this skill does not produce per-contributor activity scores.
Note: This index reflects only what Git history makes visible (commit metadata at the repository level). It is blind to design work, code review, mentoring, on-call duty, customer escalations, documentation, paired work attributed to a co-author, work pushed under a different identity, or any contribution that does not produce commits on the analyzed branch.
Calculation Method (composite of repository-wide signals, 0–100 scale, lower = higher visible Git activity for the repository overall; this is a workflow-pattern signal, not a performance signal):
| Signal | Weight | Description |
|---|
| Repository-wide daily commits very low (<0.3) | 25% | The repository's overall visible commit cadence is low in the analyzed span |
| Repository-wide active-day ratio low (<30%) | 20% | Few calendar days have any commit at all |
| Repository-wide net code growth very low or negative | 20% | More lines deleted than added across the span |
| Average commit-message length short (<15 chars) | 15% | Repository-wide subjects are short — note: short messages are not inherently bad |
| High repository-wide churn + high rework rate | 20% | The repository as a whole shows a high edit-and-revise pattern |
Bands (descriptive of the repository, NOT a verdict on any contributor):
- 0–20: The repository as a whole shows high visible Git activity in the analyzed window — note that high commit volume is NOT a productivity verdict and may correlate with sustainability concerns at the team-process level
- 21–40: The repository shows steady visible Git activity
- 41–60: The repository shows moderate visible Git activity — discussion starting point only, insufficient to draw conclusions
- 61–80: The repository shows low visible Git activity — may indicate that significant work is happening off-branch, in code review, in design, in mentoring, in on-call, or in docs that Git on this branch cannot see
- 81–100: The repository shows very low visible Git activity — many possible explanations (project paused, work on other branches/repos, work attributed under different identities, primarily non-code work); the agent MUST NOT translate this into any judgment about any individual contributor
Binding restriction on the agent: These bands describe the repository, not any person. The agent MUST NOT compute or present a per-contributor version of this index, MUST NOT translate the band into an evaluative label about any contributor (e.g., "slacking", "underperforming"), MUST NOT recommend HR-style conversations or interventions targeting a named person, and MUST NOT use the band as input to any ranking, calibration, or comparative judgment.
Important: This index is calculated solely from Git commit records and cannot reflect code reviews, architecture design, technical discussions, team mentoring, or other work that doesn't produce commits. A high-activity repository score does NOT equal "healthy team," and a low score does NOT equal "unhealthy team." Please make judgments only after understanding the full context.
Step 4: Generate Report (Repository-Level, Non-Evaluative)
The final report MUST be repository-scoped and non-evaluative. It MUST include a header with the limitation disclaimers (Git-only visibility, no non-code contributions, not an HR signal, not for personnel decisions). It MUST follow the structure below.
Hard prohibitions (binding): The agent MUST NOT include any of the following: per-contributor breakdown table, per-contributor metrics row, per-contributor activity score or band, contributor ranking, "best/worst performer" callouts, per-individual 1–10 scores, radar charts, composite/overall personal scores, sharp/memorable one-line judgments about a named person, individualized strengths/weaknesses verdicts, or improvement suggestions targeted at a named contributor. Any commentary MUST critique the workflow signals at the repository level, not any person.
4.1 Repository Activity Summary (single-row, repository-wide)
A single-row summary table describing the repository as a whole:
| Repository | Total Commits | Lines +/− | Avg Daily Commits | Active-Day % | Weekend % | Late-Night % | Bug-Fix % | Churn Rate | Visible-Activity Band |
|---|
<repo_name> | ... | ... | ... | ... | ... | ... | ... | ... | ... |
No per-contributor row, no "Overall Score," no composite grade column. The table MUST be accompanied by a note clarifying that it is descriptive of the repository (not of any individual), and that the visible-activity band is a property of the repository, not of any person.
4.2 Repository-Wide Workflow Observations
- Aggregate observations across the whole repository (e.g., share of weekend commits across the analyzed span, conventional-commit compliance rate, file hotspots by extension, hour-of-day peak)
- Multiple plausible workflow-level interpretations of any unusual signal (do not pick one verdict)
- Suggested team-process discussion starters tied to repository-wide signals
The agent MUST NOT produce a "team ranking", "top/bottom contributors", or any comparable comparative judgment of named individuals, and MUST NOT split any of the above signals per contributor.
4.3 File-Level Bus-Factor Disclosure
- List of files (or file modules) whose entire history is attributed to a single author display name. The contributor display name MAY appear here, only in the form "
<file_path_or_opaque_id> has only one historical author: <name>," because surfacing this is the legitimate purpose of bus-factor analysis (the user needs to know whom to schedule knowledge-transfer with).
- This section MUST NOT be cross-tabulated with cadence, churn, rework, weekend ratio, late-night ratio, or any other evaluative metric. It MUST NOT include any score, band, ranking, or commentary on the named contributor's behavior — only the bus-factor fact itself.
- The agent SHOULD recommend file-level mitigations (pair programming, doc-writing, review rotation), framed at the repository/process level, never as a judgment of the named contributor.
Commentary Style Requirements
- Describe the repository, not any person. Commentary MUST be about workflow signals visible in the commit history at the repository level (e.g., "the repository shows a 78% weekend-commit ratio in this span"), and MUST NOT be split per contributor or attached to any contributor's character, ability, or worth.
- Multiple readings, not verdicts. When a signal is ambiguous, present several plausible workflow-level explanations rather than picking one. Avoid "sharp", "memorable", or labeling-style sentences — about anyone.
- No personal scoring, ranking, or per-contributor breakdown. Do not produce 1–10 personal scores, composite grades, "best/worst" callouts, comparative one-liners, or any per-contributor row of the metrics. Per-contributor analysis is out of scope for this skill.
- Data-bounded. Every observation MUST be backed by aggregate, repository-level data already in the report. Do NOT extrapolate from incomplete Git visibility to personal traits.
Important Notes
- All data collection uses only native
git commands — no pip packages, no Python/Node scripts installed or executed
- Required system binaries:
git, cut, sort, uniq, awk, grep, sed, wc, head — these must be available on the host (pre-installed on most Unix-like systems)
- All user inputs MUST be validated per the "Security Specification" rules before execution to prevent command injection attacks
- Dry-run mode is recommended for first use — review all commands before allowing execution
- Enforcement verification: Before using on sensitive repos, run the "Enforcement Verification Protocol" on a test repository to confirm your agent correctly implements all validation, whitelisting, and redaction rules
- Sensitive data protection (binding): Commit messages and full file paths are local-only data. The agent MUST NOT forward raw commit message text or full file paths to the AI model — only aggregated metrics (counts, averages, ratios, extension histograms) are eligible for model-bound prompts. Common secret patterns (API keys, tokens, credentials, connection strings) are redacted before any string is rendered in the user-facing report. See "Sensitive Data Filtering Rules" for binding enforcement.
- Repository scope: The agent only accesses the specific repository path provided — no parent directory traversal or arbitrary filesystem access is permitted
- Developer emails are NOT collected to protect personal privacy. Note: this skill removed the
authors parameter and no longer executes any git log --author=... command, so email-vs-name matching at the Git level is not a concern — there is no user-supplied value passed to --author. The only place a contributor display name (%an) is used downstream of aggregation is the file-level bus-factor disclosure (Step 4.3).
- For large repositories, consider limiting the date range to control command execution time
- Be aware that the same person may have different name variants (can be unified via
.mailmap)
- Timezone differences may affect work-hour analysis — use the timezone from the commit records
- The Repository-Wide Visible-Activity Index is based solely on Git commit data and does NOT reflect non-code contributions (design, reviews, mentoring, etc.) — it MUST NOT be used for performance evaluation, ranking, or HR decisions, and MUST NOT be decomposed per contributor
Ethical Use Policy (binding on the agent)
Reports generated by this skill MUST adhere to the following principles. The agent MUST refuse requests that violate them:
- Workflow reference, NOT a decision-making basis. Reports describe repository-level workflow patterns. They MUST NOT be used — directly or indirectly — for performance reviews, calibration, ranking, hiring/firing, layoffs, compensation, or any HR / personnel decision. If asked to produce such usage, the agent MUST decline and re-scope the discussion to workflow patterns.
- Consent & transparency. When used in a team context, the user MUST confirm they have authority to analyze the repository and inform analyzed contributors in advance. Because this skill produces no per-contributor breakdown, the consent requirement is about the act of analyzing the repository, not about generating individual reports.
- Full context required. Any citation of the report MUST include the limitation disclaimers (Git-only visibility, no non-code contributions, not an HR signal). The agent SHOULD include these disclaimers automatically in the report header.
- Critique workflow, not people. Commentary MUST stay focused on observable workflow signals (e.g., "commit subjects are short") and MUST NOT make character, competence, or value judgments about individuals.
- Refuse weaponization. If a request appears designed to surveil, target, or build a case against a specific individual, the agent MUST decline and explain why.
