gsp-dev

Repository: jubscodes/get-shit-pretty

Step 1: Gather state

Run these in parallel:

gh api repos/jubscodes/get-shit-pretty/milestones --jq '.[] | {title, number, open_issues, closed_issues, description}'

gh issue list --state open --json number,title,labels,milestone,assignees,createdAt,updatedAt --jq '.[] | {number, title, labels: [.labels[].name], milestone: .milestone.title}'

git log main..HEAD --oneline

git status --short

Step 2: Parse focus

$ARGUMENTS determines what to show:

• all or empty — full analysis + recommendation
• milestone — milestone progress only
• blockers — issues blocking other issues
• ready — issues ready to start (no dependencies, no blockers)

Step 3: Analyze issues

For each open issue in the active milestone, classify:

3a: Dependency graph

Read issue bodies to extract Related / Blocks / Depends on references. Build a lightweight dependency picture:

• Which issues block others?
• Which issues are leaf nodes (no dependencies)?

3b: Branch coverage

Cross-reference open issues against commits on the current branch (git log main..HEAD). Classify each issue as:

• Done on branch — commits clearly address this issue
• In progress — partial commits exist
• Not started — no related commits

3c: Effort signal

Estimate relative size from issue body:

• S — single file change, config, or docs
• M — multiple files, one workstream
• L — cross-cutting, new architecture pattern

3d: Impact signal

Rate based on what the issue unblocks:

• High — blocks 2+ other issues or is required for milestone completion
• Medium — blocks 1 issue or fills a gap
• Low — standalone improvement

Step 4: Recommend next task

Apply this priority stack:

1. Blockers first — issues that block the most other work
2. Nearly done — issues with partial branch coverage (finish what's started)
3. High impact + small effort — quick wins that unblock others
4. Sequential dependencies — if A blocks B, recommend A

Pick the top 1-3 recommendations. For each, explain:

• Why this one next (what it unblocks)
• What's already done (branch commits)
• What remains (concrete tasks)

Step 5: Display

Output with clean terminal formatting:

  /gsp:dev
  ═══════════════════════════════════════

  ─── Milestone ────────────────────────

  v0.5.0 — Agentic design engineering
  ████████░░░░░░░░░░░░ 20% (2/10)

  ─── Issues ───────────────────────────

  Done on branch:
    ✅ #12  Cross-runtime installer fixes
    ✅ #19  Consolidate prompts/ vs agents/

  Ready to start:
    ○  #26  Validate plugin manifest             S  blocks: —
    ○  #23  /gsp:style skill                     L  blocks: #18, #24
    ○  #24  Composable skills pattern             M  blocks: —

  Blocked:
    ◌  #18  Quick mode                           M  needs: #23
    ◌  #21  designprompts.dev integration         M  needs: #23
    ◌  #25  CHANGELOG + migration guide           S  needs: all

  Not in milestone:
    ○  #7   Theme-aware terminal output
    ○  #4   Onboarding experience

  ─── Recommendation ───────────────────

  → #23 /gsp:style skill (L)
    Why: blocks #18 (quick mode) and #21 (designprompts.dev).
    First composable skill — proves the pattern for #24.
    Start: create references/styles/ preset format + SKILL.md

  → then #26 Validate plugin manifest (S)
    Why: quick win, verifies architecture work already on branch.
    All code exists — just needs manual testing.

  → then #24 Composable skills pattern (M)
    Why: defines the agent→skill contract.
    Can run in parallel with #23 implementation.

Step 6: Offer to start

After displaying recommendations, ask:

"Want to start working on one of these?" with the top recommendations as options.

If the user picks one, display the issue body summary and suggest the first concrete step.

Important notes

• Read-only by default — don't modify issues, just analyze
• Use gh CLI — don't hardcode issue data, always fetch fresh
• Branch-aware — recommendations account for what's already done
• Keep it short — the recommendation section is the value, not the analysis
• Don't over-fetch — read issue bodies only when needed for dependency analysis, not for every issue