backlog

/backlog

Manage the rpm backlog — persisted at docs/rpm/future/tasks.org. All operations read and write this file using org-mode format.

Project Amendments

At the start of every invocation, check whether docs/rpm/skills/backlog.md exists in the consuming project. If it does, read it and apply its contents as additional project-specific instructions for this skill. Amendments may add band orderings, extra parent groups, or custom triage rules. They cannot remove or override plugin defaults — on conflict, this SKILL.md wins.

The backlog is sorted in execution order (top-to-bottom = the order in which tasks need to get done). "Take from the top" is the expected read pattern.

Within each * Parent group, keep this band order top-to-bottom:

1. Actionable — ** IN-PROGRESS or ** TODO with all :BLOCKED_BY: deps DONE
2. Blocked — ** BLOCKED, or ** TODO with unresolved :BLOCKED_BY:
3. Postponed — ** TODO with a :POSTPONED: stamp

Closed entries (** DONE / ** CANCELLED) are archived to docs/rpm/future/done.org by /session-end and do not live in tasks.org long-term. /backlog done toggles TODO → DONE in place; the archive sweep runs at the next session-end.

Blocked and postponed items drift to the bottom of their band automatically whenever this file is written (during add, postpone, review, or session-end Phase 3b). Moves are mechanical — no user question. Preserve relative order within each band. New add entries land at the bottom of the actionable band (not the absolute bottom); user promotes upward explicitly if something needs to happen sooner.

Pivot capture (automatic, no user question). When the user redirects mid-session to new multi-step work that meets the TaskCreate bar, the LLM must insert a ** TODO at the top of the actionable band AND update docs/rpm/~rpm-session-start to set the task: field to the new work. The ask IS the confirmation — don't prompt. Skip for tactical single-step follow-ups. Rationale: the backlog should always represent current state, so a dropped session (without /session-end) still hands off accurately. Full rule lives in plugin/hooks/_directives.sh.

Routing

Parse $ARGUMENTS:

• add <description> → Add below

• list → List below

• review → Review below

• postpone <task text or number> → Postpone below

• done <task text or number> → Done below

• empty or natural language → infer intent from context. If unclear, show usage:

  /backlog add <description>   — add a backlog entry
  /backlog list                — show all entries with statuses
  /backlog review              — evaluate and reorganize
  /backlog postpone <task>     — defer to the bottom of its group
  /backlog done <task>         — mark an entry complete
  

---

Add

Do NOT call TaskCreate for backlog additions. Your rpm backlog is long-term; native tasks (TaskCreate/TaskList) are reserved for work actually happening this session. Adding to your rpm backlog without a mirrored native task is the correct, intended behavior.

3 body lines max per entry, plus the org :PROPERTIES: drawer. The drawer is its own thing — it doesn't count toward the body budget. Body content (heading + any inline notes) caps at 3 lines; default is just the heading.

Standard shape:

** TODO <description> [[file:YYYY-MM-DD-slug.md]]
   :PROPERTIES:
   :ID: <slug>
   :END:

Add :BLOCKED_BY: <other-slug> inside the drawer when there's a dependency. Add up to 2 inline body lines (under the drawer) only when something can't reasonably live in the detail file — most of the time the detail file is the right home for everything beyond the heading.

What does not belong inline in tasks.org:

• Bullet lists describing the task
• Prose explaining "why", "how", or "defer until"
• Implementation notes, links to related files, status commentary

All of that goes in the detail file at docs/rpm/future/YYYY-MM-DD-slug.md — never as 5+ lines of body under the heading. Subagents dispatched by /next actionable- backlog read the detail file, so information isn't lost — it's just not in the line-scannable index. Multi-line entries turn the backlog into prose-reading and break the SessionStart menu

• /next priority scan, both of which need "topmost unblocked" to fit a glance.

1. Read docs/rpm/future/tasks.org to see existing parent headings and task structure.

2. Determine the parent heading:

   • If an existing heading fits, suggest it (or ask when ambiguous).
   • If none fits, create a new * Parent group. When the new group names a coherent objective (a deliverable, capability, or area of work with a demonstrable done-state), propose a Goal: body line for it — don't silently create a goalless group. State the proposed goal as a demonstrable test (a pass/fail artifact that PROVES it met — "N cases pass end-to-end vs reference", "trace 38/38" — not a gap-list or a vague description like "launch ready"). The user accepts, edits, or declines; on accept, write it as a Goal: body line directly under the * Parent. Skip the proposal for pure status bands (Active, Blocked) or catch-all buckets (Misc, Maintenance, Ideas) where a success metric doesn't apply.

   For any * Parent that carries a Goal: / Success: line (new or existing), keep the metric a demonstrable test and ensure the goal's chain terminates in a verify/sign-off task — file that terminus now if it's missing, with the build gap-tasks as its :BLOCKED_BY: deps. Don't leave a goal defined only by its build gaps.

3. Add as ** TODO <one-sentence description> [[file:YYYY-MM-DD-slug.md]] followed by the standard :PROPERTIES: / :ID: <slug> / :END: drawer, at the bottom of the actionable band in the chosen heading (above any blocked / postponed / DONE entries).

4. Create the detail file at docs/rpm/future/YYYY-MM-DD-slug.md with everything that would otherwise be inline: title, description, action steps, scope, estimate, related links. Ask the user for details if the task is complex.

5. If the task has obvious dependencies on existing tasks, add :BLOCKED_BY: <other-slug> inside the drawer (still structured metadata, not body).

6. Research offer (when warranted). If the task crosses into a domain or technology new to the project, hinges on external or factual unknowns (APIs, libraries, standards, landscape/competitor facts, best practices), or is a selection/evaluation/comparison ("X vs Y", "which approach"), offer to run the research skill (rpm:research) on it first — upfront research de-risks a new-domain task before build effort is committed, and its saved report (docs/rpm/research/<slug>/) can be linked from the detail file. Offer only; don't auto-run (matches the research skill's own Offer gate). Skip the offer for internal/mechanical work fully scoped by the existing codebase (refactors, bugfixes, renames, doc/test edits) or when the domain knowledge is already in hand.

7. Confirm: print the new entry and its location.

---

List

1. Read docs/rpm/future/tasks.org.

2. Print a summary line and the active task list:

   ## Tasks — N in-progress · N todo · N blocked
   
   <Parent Heading>
      1. [TODO] <description>
      2. [IN-PROGRESS] <description>
         blocked-by: <id>
   

   Grouped by parent heading. Number sequentially for reference. Closed entries live in docs/rpm/future/done.org — read that file directly if the user asks for history.

---

Review

1. Read docs/rpm/future/tasks.org and all linked detail files.
2. Evaluate:
   • Organization: tasks under logical parent headings?
   • Dependencies: relationships correct? Missing or circular?
   • Staleness: TODOs with no activity across multiple sessions?
   • Scope: any tasks too large for one session (~35 min)? Suggest decomposition.
   • Duplicates: overlapping tasks?
   • Demonstrable goals: for any * Parent carrying a Goal: / Success: line, that metric must be a demonstrable test (a pass/fail artifact that PROVES it met — "N cases pass end-to-end vs reference", "trace 38/38"), not a gap-checklist or a vague description ("launch ready"). Every [NOT MET] goal's task chain must terminate in a verify/sign-off task (with the build gap-tasks as its :BLOCKED_BY: deps), not merely list build gaps. Flag goals whose Success: is a gap-list/description or whose chain has no terminal proof task. A metric reconcile/split must add the "prove it" terminus in the same edit.
   • Order: work is sorted top-to-bottom by execution order (when it needs to get done)? Anything that should happen sooner than what's currently above it?
   • Deferrals: anything the user has set aside or signaled they want to come back to later? Surface as candidates for Postpone.
3. Present findings and proposed changes. Wait for confirmation before editing. For postpones, apply the move using the Postpone procedure below.

---

Postpone

Defer a task to the bottom of its * Parent group. Status stays TODO (the task isn't dropped, just moved later in the execution order). Adds a :POSTPONED: YYYY-MM-DD property so the deferral is auditable.

1. Read docs/rpm/future/tasks.org.
2. Match the argument to a task — by number (from most recent /backlog list), by text match, or by asking if ambiguous.
3. Identify the task's * Parent heading. Find the last ** task under that parent (any status — active, blocked, or postponed).
4. Edit the file to move the matched task's heading + its property drawer to just below that last sibling. Do not change status; keep ** TODO.
5. Add (or update) :POSTPONED: YYYY-MM-DD inside the property drawer with today's date. Create the drawer if the task didn't have one.
6. Confirm: print the moved task's new position.

If the task is already at the bottom of its group, just stamp the :POSTPONED: property and note "already at bottom".

---

Done

1. Read docs/rpm/future/tasks.org.
2. Match the argument to a task — by number (from most recent /backlog list), by text match, or by asking if ambiguous.
3. Change ** TODO to ** DONE and append today's date: ** DONE <description> :CLOSED: [YYYY-MM-DD]
4. Confirm what was marked done.