Compounding & Quality¶

These 7 workflows close the loop between building and learning. They turn solved problems into searchable knowledge, enforce multi-persona review, stress-test scope before committing, and ship code through a structured pipeline.

New in v2.0.0.

/compound¶

Captures a recently solved problem or learned pattern while context is fresh.

/compound                   # full mode: research, overlap detection, dedup
/compound --lightweight     # single-pass, fewer tokens

What it does:

1. Asks you to choose Full or Lightweight mode
2. Classifies the solution by track (bug or knowledge) and category
3. Researches .planning/solutions/ for overlapping prior art
4. Writes a structured document with YAML frontmatter to .planning/solutions/[category]/
5. Commits and offers @agentic-learning space to schedule review

Output: .planning/solutions/[category]/[filename].md — searchable by /plan-phase (Step 2b) and /knowledge-base.

When to use:

• After /debug resolves a bug — capture the problem, root cause, and fix
• After /verify-work passes — capture notable patterns from the phase
• After any "aha moment" in a session — compound it before context fades

Agent: solution-writer (inline persona or learnship-solution-writer subagent).

Learning checkpoint: either-or · reflect

/review¶

Multi-persona code review through six lenses.

/review                     # interactive mode (default)
/review --report            # report-only, no conversation
/review --autofix           # apply fixes automatically

The 6 review lenses:

Output: Severity-ranked findings (P0–P3) with confidence scores (0.0–1.0). Only lenses relevant to the diff are activated (e.g., no security lens for a CSS-only change).

Agent: code-reviewer (inline persona or learnship-code-reviewer subagent).

Config: review.auto_after_verify — when true, automatically runs after /verify-work passes.

Learning checkpoint: learn · either-or

/challenge¶

Product and engineering challenge gate — is this worth building?

/challenge "[proposal description]"
/challenge                  # reads from MILESTONE-CONTEXT.md or ROADMAP.md

What it does:

1. Gathers context from PROJECT.md, ROADMAP.md, DECISIONS.md
2. Applies two lenses in parallel (or sequentially):
3. Product lens: Does this solve a real user problem? Is the scope right?
4. Engineering lens: Is this the simplest architecture? What are we coupling to?
5. Asks 3–5 forcing questions per lens
6. Synthesizes a verdict: proceed, rethink, or reduce scope
7. Records the decision to DECISIONS.md

When to use: Before committing to a milestone, large feature, or major refactor. 15 minutes of challenge prevents weeks of rework.

Agent: challenger (inline persona or learnship-challenger subagent).

Learning checkpoint: either-or · brainstorm

/ship¶

End-to-end ship pipeline: test → lint → commit → push → PR.

/ship                       # full pipeline
/ship --skip-tests          # skip test step
/ship --dry-run             # show what would happen without executing

What it does:

1. Pre-flight: Detects test runner, linter, git status
2. Test: Runs detected test suite (skip with --skip-tests or ship.auto_test: false)
3. Lint: Runs detected linter if present
4. Stage: git add changed files
5. Commit: Conventional commit format (feat:, fix:, chore:, etc.) when ship.conventional_commits: true
6. Push: git push to current branch
7. PR: Creates pull request with auto-generated description when ship.pr_template: true
8. Confirm: Shows summary and suggests /compound for notable patterns

Config options:

Learning checkpoint: reflect

/ideate¶

Codebase-grounded divergent ideation — discover what's worth building next.

/ideate                     # scan codebase and generate ideas
/ideate "[focus area]"      # scope to a specific area

What it does:

1. Scans codebase for TODOs, test gaps, hotspots, and friction points
2. Generates 15–25 ideas across four thinking frames:
3. User pain: What frustrates users right now?
4. Inversion: What if we did the opposite of the current approach?
5. Assumption-breaking: What are we assuming that might be wrong?
6. Leverage: Where would a small change have outsized impact?
7. Deduplicates and adversarial-filters weak ideas
8. Presents top 5–7 ranked survivors with effort estimates

When to use: Between milestones (after /complete-milestone, before /discuss-milestone), or when you feel stuck on what to build next. Requires an existing project — for pre-project ideation, use @agentic-learning brainstorm [idea] instead.

Agent: ideation-agent (inline persona or learnship-ideation-agent subagent).

Learning checkpoint: brainstorm · either-or

/guard¶

Safety mode for sensitive phases — warns before destructive commands and locks file scope.

/guard auth/ payments/ config/    # protect these directories
/guard --off                       # deactivate safety mode

What it does:

1. Determines scope from arguments or asks interactively
2. Creates .planning/guard-state.md with protected paths and rules
3. While active:
4. Warns before any destructive command (rm, DROP, truncate, etc.)
5. Warns before editing files outside the guarded scope
6. Adds a 🛡️ GUARD prefix to the session banner
7. Persists across sessions via guard-state.md
8. Deactivate with /guard --off

When to use: Working on auth, payments, database migrations, or any area where accidental changes could be catastrophic.

Learning checkpoint: learn

/sync-docs¶

Detects stale documentation after code changes.

/sync-docs                  # scan for drift
/sync-docs --autofix        # fix simple cases automatically

What it does:

1. Identifies documentation files (README, docs/, API docs, inline JSDoc/docstrings)
2. Compares against recent git changes
3. Scans for: renamed references, outdated paths, stale examples, broken links
4. Reports findings by severity (high/medium/low)
5. Auto-fixes simple cases with --autofix (renamed references, updated paths)

When to use: Before /complete-milestone (it's suggested automatically), after large refactors, or periodically during long milestones.

Learning checkpoint: learn

The extended phase loop¶

v2.0 extends the phase loop with three new steps after verification:

flowchart LR
    DP["/discuss-phase"] --> PP["/plan-phase"] --> EP["/execute-phase"] --> VW["/verify-work"]
    VW --> RV["/review"]
    RV --> SH["/ship"]
    SH --> CP["/compound"]
    CP -->|"next phase"| DP
    VW -->|"all done"| AM["/audit-milestone"]

# Full v2.0 phase lifecycle
/discuss-phase N
/plan-phase N
/execute-phase N
/verify-work N
/review              # multi-persona code review
/ship                # test → lint → commit → push → PR
/compound            # capture what you learned
/discuss-phase N+1   # next phase

The last three steps are recommended after every phase. They surface naturally through done-banner suggestions after /verify-work passes.

v2.1 extends this with optional per-phase workflows:

/verify-work N       # manual UAT
/secure-phase N      # v2.1: STRIDE security verification
/review              # multi-persona code review
/ship                # test → commit → push → PR
/compound            # capture what you learned
/extract-learnings N # v2.1: structured meta-knowledge extraction

See also: Security & Documentation · Session & Learnings · Recovery

/ideate (v2.1 enhanced)¶

v2.1 adds --explore mode for Socratic exploration alongside the existing scan mode:

/ideate                          # scan mode: codebase hotspots → ranked ideas
/ideate [focus]                  # focused scan on a specific area
/ideate --explore [topic]        # v2.1: Socratic exploration with mid-conversation research

In --explore mode, the agent asks one question at a time, offers mid-conversation research when factual questions arise, and crystallizes outputs to notes, todos, decisions, or phase proposals.

Agent personas¶

17 agent personas, each available as inline (sequential with <persona_context> blocks), @./agents/ file reference, and dispatch (parallel subagent via Task()). On Windsurf, personas are installed as model_decision rules for native conditional adoption.