| name | rdm-backlog |
| description | Run a batched backlog grooming pass over rdm's stale/duplicate/tag-cluster/archivable signals and emit a reviewable, propose-only plan of exact rdm commands — no mutations |
| allowed-tools | ["Read","Bash","Glob","Grep"] |
Run a batched grooming pass over the backlog and emit a reviewable plan. This skill reads rdm backlog report and proposes a set of consolidate / merge / retire / archive actions — each written as the exact rdm command a human would run — plus the open questions it could not answer confidently. It never mutates the plan repo: no create, update, merge, archive, promote, staging, or rdm commit. Deciding what to actually change stays a human call; this skill does the analysis and hands over a ready-to-run batch.
This mirrors how rdm-autopilot batches blockers rather than guessing: surface everything at once, then stop.
IMPORTANT: This is the rdm source repo. Always run cargo build first, then use ./target/debug/rdm — never bare rdm. If you modify any rdm source, cargo build again before running it.
Contract
$ARGUMENTS is all optional:
[--project <name>] — plan repo to groom. If omitted, the standard resolution chain applies (--project flag > RDM_PROJECT env var > default_project in rdm.toml). This dogfood repo uses --project rdm.[--older-than <days>] — passed straight through to rdm backlog report to tune the staleness threshold.[--tag <tag>] — passed straight through to rdm backlog report to scope the scan.
There is no positional argument naming an item to change, because this skill changes nothing. Thread whatever --project/--older-than/--tag you were given through to every read command below.
Non-mutation guarantee
This skill runs exactly one rdm command that touches plan data — rdm backlog report (read-only) — plus optional read-only rdm roadmap list / rdm search lookups during analysis. It never runs create, update, merge, archive, promote, rdm commit, or rdm discard. Every action it proposes is text for a human to run later, never executed here. The allowed-tools above deliberately omit Write/Edit: the emitted grooming plan is this skill's final chat message, not a file it writes.
Steps
cargo build (this is the rdm source repo).- Parse
$ARGUMENTS into optional --project/--older-than/--tag values.
- Fetch the candidates (read-only):
./target/debug/rdm backlog report --format json [--older-than <days>] [--tag <tag>] --project rdm
The JSON has four arrays: stale_tasks (slug, title, status, created, age_days), duplicate_clusters (members: slug/title), tag_clusters (tag, tasks: slug/title), and archivable_roadmaps (roadmap, title, phase_count).
- Empty case: if all four arrays are empty, say plainly "Nothing to groom — the backlog report returned no signals" and stop. Do not fabricate a plan or emit an empty
## Grooming plan / ## Open questions skeleton.
- Otherwise, run the grooming analysis below and emit the plan. If only some arrays are empty, omit that category's subsection entirely rather than printing an empty header.
Grooming analysis
Turn each populated category into proposed actions. Emit the whole thing as one reviewable batch — a flat, ordered list of {command, rationale} pairs grouped by category under clear subheadings, so a human can copy/paste-run them top to bottom (or not). Every proposed mutating command must carry --no-edit where the verb supports it, because the human is expected to run them unattended later.
Autopilot-oriented framing (read first): whenever you propose creating or extending a thematic roadmap, the phase body you propose (the --body/stdin content for promote) must be structured with ## Context / ## Steps / ## Acceptance Criteria headings — the same shape every existing phase body uses. That way, if a human executes the batch, a later rdm-estimate (which needs a body to rate difficulty from) and rdm-autopilot (which needs actionable phases) can pick the roadmap up immediately. The goal of every consolidation you propose is a roadmap that is ready for /rdm-autopilot.
stale_tasks — for each task, decide:
- Retire if it reads as superseded or no longer relevant:
./target/debug/rdm task update <slug> --status wont-fix --reason "<why it is stale / superseded>" --no-edit --project rdm
- Consolidate if it is still valid work that fits a theme, into an existing roadmap (
./target/debug/rdm promote <slug> --into <roadmap> --no-edit --project rdm) or a new one (./target/debug/rdm promote <slug> --roadmap-slug <new-slug> --project rdm). Note: --no-edit/--body apply only to --into; omit them with --roadmap-slug.
- Otherwise → open question (below), not a blind guess.
duplicate_clusters — for each cluster, pick a survivor (state the rule you used: most complete body, or earliest created) and fold the rest in:
./target/debug/rdm task merge <survivor> --from <other1> --from <other2> --no-edit --project rdm
If no survivor is clearly best, → open question, do not guess.tag_clusters — a cluster of related tasks under one tag is a consolidation candidate. First check whether a thematic roadmap already covers it (./target/debug/rdm roadmap list --project rdm / ./target/debug/rdm search <tag> --type roadmap --project rdm). If one exists, propose ./target/debug/rdm promote <slug> --into <existing-roadmap> --no-edit --project rdm per task. If none exists, propose one ./target/debug/rdm promote <first-slug> --roadmap-slug <new-thematic-slug> --project rdm (no --no-edit/--body — those apply only to --into) and then ./target/debug/rdm promote <slug> --into <that-new-slug> --no-edit --project rdm for the rest, in that create-then-fold order. Never propose both --into and --roadmap-slug for the same task — they are mutually exclusive.archivable_roadmaps — each is already all-terminal, so propose:
./target/debug/rdm roadmap archive <roadmap> --project rdm
Rationale: "all phases terminal, not yet archived." Never add --force — these candidates never need it, and --force exists only to override incomplete roadmaps, which is not this skill's job.
Open questions
Always end the plan with an ## Open questions section (include it even when short) listing every candidate the analysis could not confidently resolve — an ambiguous retire-vs-keep, a duplicate cluster with two equally plausible survivors, a tag cluster too thin or mixed-theme to name a roadmap for. State the item(s) and why it is ambiguous, and attach no command. The rule: never propose a merge / retire / archive / consolidate action you are not confident about — file it as an open question instead. Hold destructive-if-wrong actions (retire, merge, archive) to a stricter confidence bar than purely additive ones (proposing a new roadmap), because undoing a wrong wont-fix/merge/archive costs a human more than an over-eager but harmless roadmap proposal.
