git-strategy-advisor（git 战略顾问）

Git Strategy Advisor

Analyzes work context (planned or completed) and recommends an appropriate git workflow strategy. Produces four decisions -- branch strategy, branch naming, push timing, and PR creation -- as structured YAML output with confidence calibration and rationale.

This skill is advisory only. It reads git state but never modifies it. The caller or user executes the recommended actions.

When to Use This Skill

• Before starting work (pre-work mode): to determine branch strategy before making changes
• After completing work (post-work mode): to determine push/PR strategy based on actual changes
• Standalone: when you want git advice for the current repository state
• When a calling orchestrator needs git strategy recommendations via Task tool
• When unsure whether to branch, push, or create a PR

When NOT to Use This Skill

• Executing git commands (this skill advises; it does not modify the repository)
• Commit message generation (deferred to v2)
• PR description generation (deferred to v2)
• Complex branching strategies (mono-repos, release branches, gitflow)
• Conflict resolution
• Repository initialization or setup
• Re-invoking orchestrator or pipeline skills from within this skill (this skill is a leaf node; it reads git state and returns recommendations but never delegates work to other skills or orchestrators)

Workflow Overview

Three sequential phases, each building on the previous:

Phase 1: Context Analysis    -->  Determine mode, check git state, validate context
Phase 2: Work Classification -->  Extract metrics, classify scope and work type
Phase 3: Strategy Recommendation --> Apply decision matrix, check consistency, format output

Tools used: Bash (for git commands in post-work mode), Read (for decision-matrix.md), Write (for output file)

Output: Structured YAML recommendation (written to file AND returned in response)

Phase 1: Context Analysis

Determine invocation mode, run pre-flight checks, and validate context sufficiency.

1.1 Mode Detection

Use this priority-based detection cascade:

1. Explicit mode signal (highest priority): If context contains mode: pre-work or mode: post-work, use that mode directly.
2. Structured data signal: If context contains git status porcelain output (patterns: M path, ?? path, A path) or git diff headers (+++ b/, --- a/), use post-work mode.
3. Keyword heuristic (lowest priority): Count pre-work signals ("plan", "about to", "will", "estimated", "going to", "planning to", "intend to") vs post-work signals ("completed", "finished", "done", "implemented", "changed", "modified", "just made"). Use mode with more signals. Tie = post-work.
4. Default: post-work (analyze current git state).

Confidence adjustment:

• Explicit signal or unambiguous structured data: no penalty.
• Keyword heuristic with unclear winner: degrade confidence one level.
• Default fallback (no signals): confidence starts at "low" with note in summary.

Caller guidance: Callers SHOULD include explicit mode: pre-work or mode: post-work for reliable behavior.

1.2 Pre-flight Check: Git Repository

Run git rev-parse --is-inside-work-tree 2>/dev/null:

• If exit code!= 0 (not a git repo):

- Pre-work mode: Produce recommendation with defaults (create-feature-branch, stay-local, no-pr-needed), confidence = "low", summary includes "Consider initializing git with git init." - Post-work mode: Return structured output with confidence = "none", summary = "Post-work analysis requires a git repository."

1.3 Pre-flight Check: Git State

Before extracting metrics, detect special states:

• Detached HEAD: git branch --show-current returns empty. Set current_branch = "DETACHED", is_on_main = false. Force branch.action = "create-feature-branch" with rationale about preserving work.
• Merge in progress: test -f.git/MERGE_HEAD. Return advisory: "Complete or abort the merge before requesting strategy advice." All strategy fields = null, confidence = "none".
• Rebase in progress: test -d.git/rebase-merge || test -d.git/rebase-apply. Return advisory: "Complete or abort the rebase before requesting strategy advice." All strategy fields = null, confidence = "none".

1.4 Insufficient Context Handling

If pre-work mode detected but context lacks actionable detail (no file paths, no scope keywords, no numeric estimates):

• Standalone invocation: Ask user for more context (what files, how many, what type of change).
• Orchestrator invocation (via Task tool): Return conservative defaults: scope = "minor", work_type = "mixed", confidence = "low", summary includes "Insufficient context for accurate recommendation. Consider re-invoking in post-work mode."

Phase 2: Work Classification

Extract metrics, detect primary branch, classify scope and work type.

2.1 Primary Branch Detection

Use this detection chain (first success wins):

1. git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's@^refs/remotes/origin/@@'
2. git branch --list main master develop trunk | head -1 | tr -d '* '
3. git config init.defaultBranch
4. Default to "main"

Store as context.primary_branch. Set context.is_on_main by comparing current branch to primary branch.

2.2 Remote Detection

Check git remote -v for output. Set context.has_remote to true if any remote is configured.

2.3 Metrics Extraction

Post-work mode:

First check for changes:

• Run git status --porcelain to detect uncommitted changes.
• If empty (clean working tree):

- Check staged changes: git diff --staged --stat - Check unpushed commits: git log @{upstream}..HEAD --oneline 2>/dev/null or git log origin/{primary_branch}..HEAD --oneline 2>/dev/null - If unpushed commits exist: Use git diff --stat origin/{primary_branch}..HEAD for metrics. Note: "Analysis based on N unpushed commits." - If nothing: scope = "trivial", summary = "No uncommitted or unpushed changes detected."

For uncommitted changes, use git diff --numstat HEAD to capture BOTH staged and unstaged changes:

• Files changed: count lines of output
• Lines changed: sum of additions + deletions
• Directories spanned: count distinct first path components from changed file paths
• Binary files: count separately (shown as - in numstat), note in warnings if present

If HEAD does not exist (empty repo), treat all files from git status --porcelain as new additions with confidence = "medium".

Pre-work mode:

• Estimate from task description keywords and file/directory mentions.
• If caller provides explicit estimates, use those directly.

2.4 Scope Classification

Apply the threshold table and aggregation algorithm from references/decision-matrix.md:

1. Evaluate each metric independently against the threshold table.
2. Take the MAXIMUM scope level across all three metrics.
3. Single-metric exception: If only one metric is elevated and the other two are at trivial, downgrade by one level and set confidence to "medium".

See references/decision-matrix.md for the complete threshold table and worked examples.

2.5 Work Type Classification

Apply the majority rule from references/decision-matrix.md:

1. Classify each file by extension using the extension mapping.
2. Count files per type.
3. If ALL files belong to one type: use that type.
4. If >= 80%: use that type (note minority in warnings).
5. If no type reaches 80%: classify as "mixed".
6. Skill type override: requires >50% skill-path files.
7. Unrecognized extensions: classify as "code" (conservative default).

Phase 3: Strategy Recommendation

Apply decision matrix, check consistency, calibrate confidence, and format output.

3.1 Decision Matrix Application

Read references/decision-matrix.md and apply the four decision tables based on Phase 2 classification:

1. Branch Strategy: Based on scope, work type, and whether on primary branch.
2. Branch Naming: Based on work type with keyword overrides for fix/refactor.
3. Push Strategy: Based on scope and remote availability.
4. PR Strategy: Based on scope and work type.

Defaults for uncovered combinations:

3.2 Branch Name Generation

Format: {type}/{brief-description}

Prefix selection:

• Check for keyword overrides first: "fix/bug/patch/repair" -> fix/, "refactor/restructure/reorganize/cleanup" -> refactor/
• Fall back to work-type mapping: code -> feature/, documentation -> docs/, configuration -> config/, skill -> skill/, mixed -> feature/

Name generation:

1. Extract 2-4 keywords from task description.
2. Join with hyphens, lowercase.
3. Truncate at last complete word within 40 characters (never cut mid-word).

3.3 Post-Decision Consistency Check

After computing all four decisions independently, apply the five consistency override rules from references/decision-matrix.md in order. Log any overrides in the warnings array with code "CONSISTENCY_OVERRIDE".

3.4 Rationale Quality

Rationale strings follow the situation-implication-action pattern:

• Situation (what metrics show): "3 files, 120 lines across 2 directories"
• Implication (what this means): "moderate scope change on main branch"
• Action (what to do): "create a feature branch to isolate changes"

Example: "3 files, 120 lines across 2 directories (moderate scope) on main -- feature branch isolates risk and enables PR review."

For major scope, include: "Consider breaking this into smaller PRs for better review quality (research shows review effectiveness drops significantly above 200 lines)."

3.5 Confidence Calibration

Start at "high" and degrade based on conditions:

Maximum: "high". Minimum: "low" (or "none" for error states). Any 2 degradation conditions = "low".

3.6 Output Formatting

For standalone invocation, present human-readable summary followed by full YAML:

Git Strategy Recommendation
---
Based on your changes (3 files, 120 lines across 2 directories):

  Branch: Create feature branch `feature/add-validation-logic`
  Push:   Push after self-review
  PR:     Create a pull request

  Confidence: High
  Rationale: Moderate code changes on main branch warrant isolation
             and peer review.

Full recommendation written to: /tmp/git-strategy-recommendation.yaml

For orchestrator invocation (via Task tool), return YAML in the response text AND write to file.

3.7 Output File Handling

1. If caller specifies output path: attempt to write there. If parent directory missing or write fails, fall back to /tmp/git-strategy-recommendation.yaml.
2. Always return full recommendation in Task tool response text (dual delivery).
3. Log warning with code "OUTPUT_FALLBACK" if fallback path was used.

Output Schema

git_strategy_recommendation:
  version: "1.0"
  timestamp: "2026-02-07T10:30:00Z"
  mode: "pre-work" | "post-work"
  confidence: "high" | "medium" | "low" | "none"

  analysis:
    files_changed: 3
    lines_changed: 120
    directories_spanned: 2
    work_type: "code"
    scope: "moderate"

  context:
    current_branch: "main"
    primary_branch: "main"
    is_on_main: true
    has_remote: true

  strategy:
    branch:
      action: "create-feature-branch"
      suggested_name: "feature/add-validation-logic"
      rationale: "3 files, 120 lines across 2 dirs (moderate) on main -- feature branch isolates risk."
    push:
      action: "push-now"
      rationale: "Moderate scope with remote configured -- push to back up work and enable collaboration."
    pr:
      action: "create-pr"
      rationale: "Moderate scope warrants peer review via pull request."

  warnings: []
    # Example entries:
    # - code: "SCOPE_BOUNDARY"
    #   message: "Metrics conflict on scope. Used highest (moderate)."
    # - code: "CONSISTENCY_OVERRIDE"
    #   message: "PR strategy overridden: direct-commit does not use PRs."
    # - code: "GIT_NOT_INITIALIZED"
    #   message: "Git repository not found. Defaults used."

  summary: "Create feature branch `feature/add-validation-logic`, push to remote, and open a PR for review."

Error Handling

Errors do NOT abort the pipeline. Phase 3 receives partial data with warnings. The confidence field reflects cumulative error severity.

Integration Notes

State Management

This skill uses in-context state only. No intermediate files, no session directory, no resume capability. If interrupted, re-invoke. The skill is idempotent.

Ecosystem Alignment

• with programming-pm Phase 6: programming-pm may optionally invoke git-strategy-advisor in post-work mode before its existing branching logic. The advisor may recommend direct-commit for trivial changes or confirm the existing branch-on-main behavior for larger changes. programming-pm's existing Phase 6 logic takes precedence unconditionally; the advisor's recommendation is presented as an informational note.
• with programming-pm branch naming: programming-pm uses {type}/{task-id}-{description}. git-strategy-advisor uses {type}/{description} (no task ID available). If caller provides task ID, include it in the description.
• with skill-editor Phase 4: skill-editor may optionally invoke git-strategy-advisor in post-work mode before committing. The advisor may recommend branching for skill changes depending on scope. skill-editor's existing commit logic takes precedence unconditionally; the advisor's recommendation is presented as an informational note.
• with other orchestrators: technical-pm, lit-pm, scientific-analysis-architect, and research-pipeline may invoke git-strategy-advisor in post-work mode after completing their workflows to get git recommendations for the produced files.

Maintenance note: git-strategy-advisor integration sections exist in 6 orchestrator SKILL.md files (programming-pm, skill-editor, technical-pm, lit-pm, scientific-analysis-architect, research-pipeline). When modifying the integration pattern in any single orchestrator, check all 6 for consistency.

Invocation Patterns

• Pattern A (Orchestrator pre-work): Task tool with mode: pre-work and task description. See examples/pre-work-invocation.md.
• Pattern B (Orchestrator post-work): Task tool with mode: post-work and git state. See examples/post-work-invocation.md.
• Pattern C (Standalone): User asks for git advice directly. Skill reads current git state via Bash.
• Pattern D (Bookend): Pre-work invocation at start, post-work re-evaluation at end. Post-work is authoritative.

Archival Compliance

When writing output files, this skill:

• Uses descriptive filenames (git-strategy-recommendation.yaml)
• Includes version and timestamp in the output schema
• Follows the dual-delivery pattern (file + response text) for traceability