backlog

/backlog — rank dispatchable issues for the closed loop

Outer-loop skill: surveys the open backlog and recommends the next issue to dispatch to an asset.

Why it's a skill (not a subagent)

The procedure is short; context isolation is only needed for the heavy reads (many issues + memory + git). Those go into an ephemeral general-purpose subagent that returns a digest; the outer loop only sees the recommendation, not the raw reads.

Procedure

1. Spawn a general-purpose subagent for the heavy read with this prompt (fill <MEMORY_DIR> = this project's auto-memory dir, ~/.claude/projects/<project-slug>/memory/):

   Survey this repo's backlog for the closed-loop dispatcher. Return ONLY a markdown table + 2-3 sentence recommendation; no raw issue dumps.

   Read in parallel:

   • gh issue list --state open --limit 60 --json number,title,labels,createdAt
   • ls -t <MEMORY_DIR>/feedback_*.md <MEMORY_DIR>/project_*.md | head -10 and read the top 5 (skip if no memory dir)
   • git log --oneline -20
   • any repo-specific lineage/status command (skip if none)

   Apply triage priority: bug > refactoring > feature (something broken > improving existing structure > new capability). Filter out:

   • Issues that touch a repo's declared frozen_paths (see .claude/sdlc.toml) — e.g. a hash-freeze
   • Real-world-action requests gated by prod_gated (deploy / arm / posture / promote — human-gated)
   • Research umbrellas (too broad for a single asset workstream)
   • Blocked issues (upstream dependency / external lock)

   Return:

   | # | Type | Brief | Cost est. |
   |---|------|-------|-----------|
   | <N> | bug/refactor/feature | one sentence | <min>min / $<dollars> |
   

   Plus 2-3 sentences on the top pick and why. Be terse — the controller is forwarding this to the owner.

2. Persist the subagent's digest to /tmp/sdlc-backlog-summary.md (always). This unconditional write survives /compact and is the fallback /dispatch <N> falls back on.

3. Present the subagent's digest verbatim to the user.

4. Wait for the user to pick. Do not auto-dispatch — that's /dispatch <N>.

5. After the user picks an issue #N, write a per-issue copy to /tmp/sdlc-backlog-<N>.md so /dispatch <N> reads the per-issue digest directly rather than scanning the full summary. Skip if the user doesn't pick anything in this turn.

Constraints

• Don't dump raw issues — the subagent's digest is the surface.
• Don't dispatch from this skill. Pick is the user's; dispatch is /dispatch.
• Cost-estimate uncertainty is fine; flag unbounded research as such.
• The triage ordering is a default, not a rule — a high-value refactor can outrank a trivial bug; say so when forwarding. Filters (frozen_paths, prod_gated, research umbrellas, blocked) drop out before ranking.