gdd-watch-authorities

name	gdd-watch-authorities
description	Fetches the design-authority feed whitelist, diffs against .design/authority-snapshot.json, and writes .design/authority-report.md (consumed by /gdd:reflect). Authority monitoring only - no trend-watching.
argument-hint	[--refresh] [--since <date>] [--feed <name>] [--schedule <weekly|daily|monthly>]
tools	Read, Write, Task, Bash

/gdd:watch-authorities

Runs design-authority-watcher on demand. Fetches the curated design-authority feed whitelist, diffs against the prior snapshot, classifies new entries into five buckets, and writes .design/authority-report.md. Phase 11's reflector picks up the report automatically when you next run /gdd:reflect.

Authority-monitoring only. Not trend-watching. See reference/authority-feeds.md §"Rejected kinds" for what this skill will never fetch.

Steps

1. Parse args. Extract optional flags: --refresh, --since <date>, --feed <name>, --schedule <cadence>. Anything that doesn't match one of these is an error - print Unknown flag: <arg>. Valid flags: --refresh --since <date> --feed <name> --schedule <weekly|daily|monthly>. and STOP.

   Mutual exclusion rules:

   • --schedule is handled entirely by this skill - it does not combine with the other three. If --schedule is present alongside any of --refresh | --since | --feed, print --schedule cannot combine with other flags. Schedule registration runs this skill with no flags at the configured cadence. and STOP.
   • --refresh and --since are mutually exclusive - print --refresh and --since are mutually exclusive. --refresh re-seeds the snapshot silently; --since surfaces a backlog from a boundary date. Pick one. and STOP.

2. Handle --schedule <cadence> branch (early-return).

   If --schedule is set:

   • Validate cadence ∈ {weekly, daily, monthly}; else print Unknown cadence: <value>. Use one of: weekly, daily, monthly. and STOP.

   • Probe for the scheduled-tasks MCP via ToolSearch:

     ToolSearch({ query: "scheduled-tasks", max_results: 3 })
     

   • If the probe returns an empty result set: print scheduled-tasks MCP not connected. Install it with: claude mcp add scheduled-tasks ... then retry with --schedule. - this is a documented fallback (not an error). Terminate with ## WATCH COMPLETE and exit 0.

   • If the probe returns one or more scheduled-tasks tools: register the cron. Discover the MCP's registration tool name at runtime from the ToolSearch result and follow its schema. Target command: /gdd:watch-authorities with NO flags (the cron invokes the default diff-and-report behavior). Cadence → cron expression mapping:

     • weekly → 0 9 * * 1 (Mondays 09:00 local)
     • daily → 0 9 * * * (every day 09:00 local)
     • monthly → 0 9 1 * * (1st of each month 09:00 local)

   • After registration: print Scheduled /gdd:watch-authorities to run <cadence>. and terminate with ## WATCH COMPLETE.

3. Validate --since <date> (if present).

   Accept ISO8601 (YYYY-MM-DD or YYYY-MM-DDTHH:MM:SSZ). Sanity-check via Bash date -d "<value>" +%s (GNU) or the POSIX equivalent python3 -c "from datetime import datetime; datetime.fromisoformat('<value>'.replace('Z','+00:00'))". On parse failure: print Invalid --since value: <value>. Use ISO8601 (YYYY-MM-DD or YYYY-MM-DDTHH:MM:SSZ). and STOP.

   If the parsed date is earlier than 2020-01-01, ask: Very old --since value: <value>. Did you mean something more recent? Proceed? [y/N]. On anything other than y/Y, STOP.

4. Spawn the watcher.

   Build the Task(subagent_type="design-authority-watcher", ...) prompt. The prompt supplies the agent's required-reading block (watcher step 0), echoes the invocation flags verbatim (watcher Flags section), and instructs the agent to follow its own fetch/diff/classify/write loop:

   Task("design-authority-watcher", """
   <required_reading>
   @reference/authority-feeds.md
   @.design/authority-snapshot.json
   @.design/STATE.md
   </required_reading>
   
   Invocation flags: <joined flag list or "none">
   
   Fetch the feeds listed in reference/authority-feeds.md, diff against .design/authority-snapshot.json,
   classify new entries per the D-17 decision table, write .design/authority-snapshot.json and
   .design/authority-report.md.
   
   Terminate with ## WATCH COMPLETE.
   """)
   

   <joined flag list> is the subset of --refresh | --since <date> | --feed <name> actually passed - e.g., --refresh, --since 2026-03-01, --feed wai-aria-apg, --refresh --feed radix-ui-releases, or literally none when no flags were supplied.

5. Print summary.

   After the agent returns:

   • If STATE.md gained a <blocker type="contract-violation"> on this run (snapshot version mismatch, hash-format violation, or over-200 entries per feed), surface the blocker verbatim and stop - do not print the default "review and reflect" line.
   • Otherwise print the agent's one-line stdout summary (normal mode: Surfaced N entries across M feeds. K skipped. See .design/authority-report.md.; first-run / refresh mode: Seeded snapshot for N feeds — next run will surface new entries.) followed by: Review and reflect: /gdd:reflect.

6. Terminate with ## WATCH COMPLETE.

Do Not

• Do not modify agents/design-authority-watcher.md.
• Do not modify agents/design-reflector.md - Phase 13.2 does not touch the reflector agent (CONTEXT.md D-25).
• Do not write to .design/authority-snapshot.json or .design/authority-report.md directly - those are the agent's writes.
• Do not fetch URLs outside reference/authority-feeds.md. The whitelist is the allow-list.