@massu/core 1.13.1 → 1.14.0

package/commands/massu-command-improve.md

@@ -1,232 +0,0 @@
----
-name: massu-command-improve
-description: "When user wants to improve a specific command's quality score, says 'improve this command', 'fix command issues', or after command-health shows weak commands"
-allowed-tools: Bash(*), Read(*), Write(*), Edit(*), Grep(*), Glob(*)
----
-name: massu-command-improve
-
-# Massu Command Improve: Score-Driven Prompt Optimization
-
-## Purpose
-
-Read accumulated command quality scores from `.claude/metrics/command-scores.jsonl`, identify consistently failing checks, propose **one targeted prompt edit at a time** to the command file, and wait for explicit user approval before applying.
-
-**This is the manual-approval counterpart to autonomous autoresearch.** Scoring is automated; improvements require human judgment.
-
----
-
-## ARGUMENTS
-
-```
-/massu-command-improve                    # Auto-pick weakest command
-/massu-command-improve massu-debug         # Target specific command
-```
-
-**Arguments from $ARGUMENTS**: {{ARGUMENTS}}
-
----
-
-## NON-NEGOTIABLE RULES
-
-1. **ONE change per iteration** — never batch multiple edits. Single-variable perturbation only.
-2. **ALWAYS show the exact diff** — user must see before/after text.
-3. **NEVER apply without approval** — wait for explicit "yes", "approved", "apply it", etc.
-4. **ALWAYS create backup** — save original section to `.claude/metrics/backups/[command]-v[N].md.bak` before editing.
-5. **ALWAYS log the change** — append to `.claude/metrics/command-changelog.jsonl` after applying.
-6. **Target the weakest check** — don't guess what to improve; let the data decide.
-
----
-
-## EXECUTION STEPS
-
-### Step 1: Load Score Data
-
-Read `.claude/metrics/command-scores.jsonl`. If empty or missing:
-```
-No command scores recorded yet.
-Scores accumulate automatically as you use instrumented commands:
-  massu-article-review, massu-create-plan, massu-loop, massu-plan, massu-debug
-
-Run these commands normally and come back when you have 5+ scored runs.
-```
-
-### Step 2: Analyze Weaknesses
-
-For each command in the data:
-
-1. Parse all JSONL lines for that command
-2. Extract individual check pass/fail counts across all runs
-3. Calculate per-check pass rate (e.g., `root_cause_identified`: 7/10 = 70%)
-4. Identify the **single weakest check** (lowest pass rate)
-5. Calculate overall command pass rate (average across all checks)
-
-If `$ARGUMENTS` specifies a command, use that. Otherwise, auto-select the command with the lowest overall pass rate.
-
-**Minimum data requirement**: At least 3 scored runs for the target command. If fewer:
-```
-[command] has only [N] scored runs. Need at least 3 for reliable analysis.
-Keep using the command normally — scores accumulate automatically.
-```
-
-### Step 3: Diagnose the Failing Check
-
-Read the target command file (`.claude/commands/[command].md`).
-
-For the weakest check, analyze WHY it might be failing:
-
-| Check Pattern | Likely Prompt Cause |
-|--------------|---------------------|
-| Check passes < 50% | The command doesn't mention this requirement at all, or buries it |
-| Check passes 50-70% | The requirement exists but is vague or easy to skip |
-| Check passes 70-85% | The requirement exists but lacks a concrete example or enforcement |
-| Check passes > 85% | Healthy — move to next weakest check |
-
-### Step 4: Propose ONE Edit
-
-Design a single, targeted prompt edit that addresses the weakest check. The edit should be one of:
-
-| Edit Type | When to Use | Example |
-|-----------|-------------|---------|
-| **Add explicit rule** | Check requirement is missing from prompt | "Your debug report MUST include a specific root cause statement, not 'probably' or 'might be'" |
-| **Add worked example** | Check exists but is vague | Add a before/after example of what good looks like |
-| **Promote to NON-NEGOTIABLE** | Check exists but is buried | Move the requirement higher in the file, add to the rules table |
-| **Add enforcement step** | Check exists but is easy to skip | Add a numbered step to START NOW that explicitly requires this check |
-| **Add banned pattern** | Check fails because of a specific anti-pattern | "NEVER produce a debug report without [X]" |
-
-### Step 5: Present to User
-
-Display:
-
-```
-━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
-COMMAND IMPROVE — [command]
-━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
-
-DATA SUMMARY
-  Scored runs:     [N]
-  Overall rate:    [X]%
-  Date range:      [earliest] — [latest]
-
-WEAKEST CHECK
-  Name:            [check_name]
-  Pass rate:       [X/N] ([Y]%)
-  Diagnosis:       [why it's failing]
-
-PROPOSED EDIT
-  Type:            [Add rule / Add example / Promote / Add step / Ban pattern]
-  Target section:  [which section of the command file]
-
-  --- BEFORE ---
-  [exact text being replaced or location of insertion]
-
-  --- AFTER ---
-  [exact new text]
-
-  --- END DIFF ---
-
-RATIONALE
-  This targets [check_name] by [explanation of why this change should help].
-
-APPROVE?
-  Reply "yes" to apply, "skip" to move to next check, or "no" to stop.
-━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
-```
-
-### Step 6: Wait for Approval
-
-**STOP HERE. Do NOT proceed without explicit user approval.**
-
-- **"yes" / "approved" / "apply it"** → proceed to Step 7
-- **"skip"** → move to next weakest check on the same command, or next command
-- **"no" / "stop"** → end the session, log the rejection
-
-### Step 7: Apply Change
-
-1. **Create backup**:
-   - Check `.claude/metrics/backups/` exists (create if not)
-   - Count existing backups for this command to determine version number
-   - Save the ORIGINAL section (not whole file) to `.claude/metrics/backups/[command]-v[N].md.bak`
-
-2. **Apply the edit** using the Edit tool
-
-3. **Log the change** — append to `.claude/metrics/command-changelog.jsonl`:
-```json
-{"command":"[command]","timestamp":"ISO8601","action":"applied","target_check":"[check_name]","check_pass_rate_before":"[X/N]","edit_type":"[type]","edit_summary":"[1-line description]","backup_file":"[path]"}
-```
-
-4. **Confirm to user**:
-```
-Applied. Backup saved to [path].
-
-Next time you run /[command], the scoring will capture whether this
-change improved [check_name]. Check /massu-command-health after 3+ runs.
-```
-
-### Step 8: Continue or Stop
-
-After applying (or skipping), offer to continue:
-- If there are more weak checks on the same command (pass rate < 85%), offer to address the next one
-- If the current command is healthy, offer to move to the next weakest command
-- If all commands are above 85%, report "All commands healthy"
-
-**Remember: ONE change at a time.** Even if 3 checks are weak, propose and apply them individually so you can measure each improvement's effect.
-
----
-
-## REJECTION LOGGING
-
-If the user rejects a proposal, log it so future runs don't re-propose the same thing:
-
-```json
-{"command":"[command]","timestamp":"ISO8601","action":"rejected","target_check":"[check_name]","edit_type":"[type]","edit_summary":"[1-line description]","rejection_reason":"[user's reason if given, or 'no reason provided']"}
-```
-
-Before proposing an edit, check the changelog for recent rejections of the same check+type combination. If found, try a DIFFERENT edit type for the same check.
-
----
-
-## MEASURING IMPROVEMENT
-
-After applying a change, the improvement is measured passively:
-
-1. User continues using the command normally
-2. Silent scoring appends new data to `command-scores.jsonl`
-3. Next time `/massu-command-health` or `/massu-bearings` runs, the trend shows whether the check improved
-4. Next time `/massu-command-improve` runs, it uses the updated data — if the check improved, it moves to the next weakest
-
-The feedback loop:
-```
-Score → Diagnose → Propose → Approve → Apply → Score again → ...
-```
-
----
-
-## EDGE CASES
-
-- **No JSONL file**: Tell user to use instrumented commands first
-- **< 3 runs for target command**: Tell user to accumulate more data
-- **All checks > 85%**: "All commands healthy — no improvements needed"
-- **User specifies unknown command**: List the 5 instrumented commands
-- **Changelog shows same check was improved before**: Note this in the proposal ("Previously improved on [date] — may need a different approach")
-
----
-
-## START NOW
-
-1. Read `.claude/metrics/command-scores.jsonl`
-2. If `$ARGUMENTS` specifies a command, target that; otherwise auto-pick weakest
-3. Analyze per-check pass rates
-4. Identify weakest check
-5. Read the command file
-6. Diagnose why the check fails
-7. Propose ONE targeted edit with exact diff
-8. **WAIT for user approval**
-9. If approved: backup → apply → log → confirm
-10. Offer to continue with next weak check
-
----
-
-## Related Commands
-
-- `/massu-autoresearch [command]` — Autonomous version. Runs the optimize loop unattended with git-based accept/reject. Use for overnight runs.
-- `/massu-command-health` — Read-only dashboard. Shows scores, trends, and weakest checks.

package/commands/massu-infra-audit.md

@@ -1,187 +0,0 @@
----
-name: massu-infra-audit
-description: "When user asks about .claude/ infrastructure health, says 'audit infra', 'check claude setup', or needs to verify references, staleness, and format compliance"
-allowed-tools: Bash(*), Read(*), Grep(*), Glob(*)
----
-name: massu-infra-audit
-
-# Massu Infrastructure Audit: .claude/ Directory Health Check
-
-## Objective
-
-Verify the health of the `.claude/` directory infrastructure: dead references, staleness, conflicts, format consistency, and memory health. Designed for periodic execution to catch drift before it causes problems.
-
----
-
-## AUDIT PHASES (8 Checks)
-
-| Phase | Check | Pass Criteria |
-|-------|-------|---------------|
-| 1 | **Dead References in CLAUDE.md** | All file paths referenced in CLAUDE.md exist |
-| 2 | **Script References** | All scripts referenced by commands/patterns exist |
-| 3 | **Command Format Consistency** | All commands have valid YAML frontmatter (name, description, allowed-tools) |
-| 4 | **Cross-Reference Integrity** | Commands reference patterns that exist; patterns reference files that exist |
-| 5 | **Line Count Accuracy** | Line counts in CLAUDE.md Load Instructions match actual file line counts |
-| 6 | **Memory Health** | MEMORY.md under 200-line limit; topic files organized |
-| 7 | **Archive Bloat** | Session state archive size and count assessment |
-| 8 | **Orphan Detection** | Files in .claude/ not referenced from CLAUDE.md or any command |
-
----
-
-## EXECUTION PROTOCOL
-
-### Phase 1: Dead References in CLAUDE.md
-
-```bash
-# Extract all relative file paths from CLAUDE.md and verify each exists
-grep -oE '\[.*?\]\((protocols|patterns|commands|incidents|reference|session-state)/[^)]+\)' .claude/CLAUDE.md \
-  | grep -oE '(protocols|patterns|commands|incidents|reference|session-state)/[^)]+' \
-  | sort -u \
-  | while read path; do
-      [ -f ".claude/$path" ] && echo "OK: $path" || echo "MISSING: $path"
-    done
-```
-
-**FAIL** if any path is MISSING.
-
-### Phase 2: Script References
-
-```bash
-# Find all script references across .claude/ files
-grep -roh 'scripts/[a-zA-Z0-9_-]*\.\(sh\|ts\)' .claude/ | sort -u | while read script; do
-  [ -f "$script" ] && echo "OK: $script" || echo "MISSING: $script"
-done
-```
-
-**FAIL** if any script is MISSING.
-
-### Phase 3: Command Format Consistency
-
-```bash
-# For each command file, verify YAML frontmatter has required fields
-for f in .claude/commands/massu-*.md; do
-  name=$(basename "$f" .md)
-  has_name=$(head -10 "$f" | grep -c "^name:")
-  has_desc=$(head -10 "$f" | grep -c "^description:")
-  has_tools=$(head -10 "$f" | grep -c "^allowed-tools:")
-  if [ "$has_name" -eq 0 ] || [ "$has_desc" -eq 0 ] || [ "$has_tools" -eq 0 ]; then
-    echo "INVALID: $name (name:$has_name desc:$has_desc tools:$has_tools)"
-  else
-    echo "OK: $name"
-  fi
-done
-```
-
-**FAIL** if any command has missing required fields.
-
-### Phase 4: Cross-Reference Integrity
-
-```bash
-# Find pattern/protocol references in commands and verify they exist
-grep -roh 'patterns/[a-zA-Z0-9_-]*\.md\|protocols/[a-zA-Z0-9_-]*\.md' .claude/commands/ \
-  | sort -u \
-  | while read ref; do
-      [ -f ".claude/$ref" ] && echo "OK: $ref" || echo "BROKEN: $ref"
-    done
-```
-
-**FAIL** if any cross-reference is BROKEN.
-
-### Phase 5: Line Count Accuracy
-
-Read CLAUDE.md Load Instructions table. For each entry with a line count:
-
-```bash
-# Compare claimed vs actual line counts
-# Flag if delta > 20 lines (significant drift)
-wc -l .claude/patterns/*.md .claude/protocols/*.md .claude/reference/*.md 2>/dev/null
-```
-
-Compare to claimed counts in CLAUDE.md. **WARN** if any delta > 20 lines.
-
-### Phase 6: Memory Health
-
-```bash
-# Check memory directory
-MEMORY_DIR="$HOME/.claude/projects/$(pwd | tr '/' '-')/memory"
-echo "=== Memory Files ==="
-ls -la "$MEMORY_DIR/"
-echo "=== MEMORY.md line count ==="
-wc -l "$MEMORY_DIR/MEMORY.md"
-echo "=== Total memory size ==="
-du -sh "$MEMORY_DIR/"
-```
-
-**WARN** if MEMORY.md exceeds 180 lines (approaching 200 limit).
-
-### Phase 7: Archive Bloat Assessment
-
-```bash
-echo "=== Archive Stats ==="
-echo "File count: $(ls .claude/session-state/archive/ | wc -l)"
-echo "Total size: $(du -sh .claude/session-state/archive/)"
-echo "=== Age distribution ==="
-echo "Last 7 days: $(find .claude/session-state/archive/ -mtime -7 | wc -l)"
-echo "7-30 days: $(find .claude/session-state/archive/ -mtime +7 -mtime -30 | wc -l)"
-echo "30+ days: $(find .claude/session-state/archive/ -mtime +30 | wc -l)"
-```
-
-**INFO** only — no pass/fail, but flag if > 500 files or > 5MB.
-
-### Phase 8: Orphan Detection
-
-```bash
-# List all non-archive .claude/ files
-# For each, check if referenced from CLAUDE.md or any command
-find .claude/ -name "*.md" -not -path "*/archive/*" -not -path "*/session-state/*" | while read f; do
-  basename=$(basename "$f")
-  refs=$(grep -rl "$basename" .claude/CLAUDE.md .claude/commands/ .claude/patterns/ .claude/protocols/ 2>/dev/null | wc -l)
-  if [ "$refs" -eq 0 ]; then
-    echo "ORPHAN?: $f (0 incoming references)"
-  fi
-done
-```
-
-**INFO** only — orphans may be intentional (specialized tools).
-
----
-
-## OUTPUT FORMAT
-
-```markdown
-## Massu Infrastructure Audit Report
-Date: YYYY-MM-DD
-
-### Results Summary
-| Phase | Check | Status | Details |
-|-------|-------|--------|---------|
-| 1 | Dead References | PASS/FAIL | [count] issues |
-| 2 | Script References | PASS/FAIL | [count] issues |
-| 3 | Command Format | PASS/FAIL | [count] issues |
-| 4 | Cross-References | PASS/FAIL | [count] issues |
-| 5 | Line Counts | PASS/WARN | [count] stale |
-| 6 | Memory Health | PASS/WARN | [lines] / 200 limit |
-| 7 | Archive Bloat | INFO | [count] files, [size] |
-| 8 | Orphan Detection | INFO | [count] potential orphans |
-
-### Overall: PASS / WARN / FAIL
-[Summary of action items if any]
-```
-
----
-
-## PASS CRITERIA
-
-- **PASS**: Phases 1-4 all pass, Phase 6 under limits
-- **WARN**: Phases 5, 7, 8 have non-critical findings
-- **FAIL**: Any of Phases 1-4 have failures (broken references, missing files, invalid format)
-
----
-
-## WHEN TO RUN
-
-- After editing CLAUDE.md or any command file
-- After adding/removing pattern or protocol files
-- After memory cleanup or reorganization
-- Weekly as part of `/massu-codebase-audit`
-- After any `.claude/` directory restructuring