/post-arc-retrospective

description: Capture the cumulative output of a /loop arc into a single auditable retrospective document; scans the heymegabyte-claude-skills plugin for modified files, categorizes by directory, counts LOC delta, extracts tool counts from MCP servers, and writes a timestamped report to retrospectives/
argument-hint: [--since=<git-ref>]
allowed-tools: Bash, Read, Write, Glob

/post-arc-retrospective

Produce a complete, auditable retrospective for a completed loop arc. Scans `~/.claude/plugins/heymegabyte-claude-skills/` for files changed since the given git ref, categorizes output by directory, counts new/modified files and LOC delta, extracts MCP tool counts, rule descriptions, and command descriptions from frontmatter, then writes a timestamped markdown report.

Step 1 — Parse arguments

Read the raw argument string.

• `--since=<git-ref>` — git ref to diff from. Default: `HEAD~50`.
• If no `--since` flag, also check `~/.claude/plugins/heymegabyte-claude-skills/retrospectives/.last-arc-start`

for a stored start commit written by `CronCreate`; use it if present.

• Store the resolved ref as `$SINCE_REF`.

PLUGIN_DIR="$HOME/.claude/plugins/heymegabyte-claude-skills"
SINCE_REF="${ARGS#--since=}"
[ "$SINCE_REF" = "$ARGS" ] && SINCE_REF=""  # no flag found

if [ -z "$SINCE_REF" ]; then
  STORED="$PLUGIN_DIR/retrospectives/.last-arc-start"
  [ -f "$STORED" ] && SINCE_REF=$(cat "$STORED")
fi
[ -z "$SINCE_REF" ] && SINCE_REF="HEAD~50"

Step 2 — Collect changed files via git

Run inside the plugin directory. Capture both Added (`A`) and Modified (`M`) files separately.

cd "$PLUGIN_DIR"
NEW_FILES=$(git diff --name-only --diff-filter=A "$SINCE_REF" HEAD -- . 2>/dev/null)
MODIFIED_FILES=$(git diff --name-only --diff-filter=M "$SINCE_REF" HEAD -- . 2>/dev/null)
ALL_CHANGED=$(git diff --name-only "$SINCE_REF" HEAD -- . 2>/dev/null)
LOC_DELTA=$(git diff --stat "$SINCE_REF" HEAD -- . 2>/dev/null | tail -1)

If the git diff fails (e.g., shallow clone or no commits), fall back to `find . -newer retrospectives/.last-arc-start -not -path '*/.git/*' -not -path '*/node_modules/*'`.

Step 3 — Categorize by directory

Iterate `$ALL_CHANGED` and bucket each path into one of these categories. Use the first matching rule (longest prefix wins):

| Category | Path prefix | |---|---| | Rules | `rules/` | | Commands | `commands/` | | Agents | `agents/` | | Hooks | `~/.claude/hooks/` or `hooks/` | | MCP Servers | `mcp-servers/` | | Template Utils | `template/utils/` or `template/evals/` | | Numbered Skills | `17-*/`, `18-*/`, `19-*/` (and higher) | | Core Skills | `[0-9][0-9]-*/` | | Bin / Scripts | `bin/` or `scripts/` | | Marketplace | `marketplace-submissions/` | | Config/Meta | `settings.json`, `package.json`, `*.yml`, `*.json` at root | | Packs | `_packs/` | | Other | everything else |

For each category track:

• `new_count` — files in `$NEW_FILES` matching this prefix
• `modified_count` — files in `$MODIFIED_FILES` matching this prefix
• list of file basenames (for rules/commands: also extract description)

Step 4 — Extract MCP server tool counts

For each directory under `mcp-servers/*/mcp-server/src/index.ts`:

for mcp in "$PLUGIN_DIR/mcp-servers"/*/mcp-server/src/index.ts; do
  name=$(echo "$mcp" | sed 's|.*/mcp-servers/||' | sed 's|/mcp-server/src/index.ts||')
  tool_count=$(grep -c 'inputSchema' "$mcp" 2>/dev/null || echo 0)
  echo "$name: $tool_count"
done

Also check for a `.mcp-prune-rules.json` alongside to note original vs pruned counts: `original_count` vs current `tool_count`.

Step 5 — Extract rule descriptions from frontmatter

For each new rule file in `rules/`:

for f in $NEW_RULE_FILES; do
  name=$(grep '^name:' "$PLUGIN_DIR/rules/$f" 2>/dev/null | head -1 | sed 's/name: //' | tr -d '"')
  h1=$(grep '^# ' "$PLUGIN_DIR/rules/$f" 2>/dev/null | head -1 | sed 's/^# //')
  echo "- **$h1** \`rules/$f\`"
done

Fall back to the filename stem if no `name:` frontmatter found.

Step 6 — Extract command descriptions from frontmatter

For each new command file in `commands/`:

for f in $NEW_CMD_FILES; do
  desc=$(grep '^description:' "$PLUGIN_DIR/commands/$f" 2>/dev/null | head -1 | sed 's/^description: *//' | tr -d '"')
  [ -z "$desc" ] && desc=$(grep '^# /' "$PLUGIN_DIR/commands/$f" | head -1 | sed 's/^# //')
  echo "- \`/$( basename $f .md )\` — $desc"
done

Step 7 — Compute totals

Aggregate across all categories:

• Total new files
• Total modified files
• LOC delta (from `git diff --stat`)
• MCP server count
• Total MCP tools (sum of all `tool_count` values)
• New rule count
• New command count
• New numbered-skill count (skills ≥17)

Step 8 — Identify highest-leverage delta

Apply this heuristic to surface the single most impactful change:

1. If a new MCP server was added → "New `<name>` MCP server: `N` tools" 2. If a forge template or forge command was added/modified → "forge `--target` capability upgraded" 3. If a rule with `priority: 1` was added → name that rule 4. If a command that enables a whole workflow was added → name that command 5. Otherwise → the category with the highest new-file count

Write one sentence: "Highest-leverage delta: `<noun>` — `<one-line what this unlocks>`"

Step 9 — Build "what's now possible" section

For each new MCP server: one bullet — "Call `<API name>` tools directly from Claude without leaving the IDE" For each new numbered skill (≥17): one bullet — "Invoke skill `/<skill-name>` for `<domain>` workflows" For each new command: one bullet from its description For each new hook: one bullet on what behavior it automates

Cap at 12 bullets total; pick highest-leverage entries.

Step 10 — Write the report

Output path: `$PLUGIN_DIR/retrospectives/arc-$(date +%Y-%m-%dT%H%M%S).md`

Report structure:

# Arc Retrospective — <date>

**Ref range:** `<SINCE_REF>..HEAD`
**Generated:** <ISO timestamp>

## Totals

| Category | New | Modified | Notes |
|---|---|---|