一键导入
rdm-revise
Act on document reviews requesting changes — work each comment and drive the review to addressed
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Act on document reviews requesting changes — work each comment and drive the review to addressed
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Rate each phase's difficulty and assign a model tier from its body
Review the plan for an rdm roadmap, phase, or task before implementation begins
Implement an rdm roadmap phase or work on an rdm task — plan, execute, then finalize into needs-review for rdm-review
Create an rdm roadmap with phases for a topic
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
Drive one named rdm roadmap from not-started to reviewed autonomously — pick the next actionable phase, estimate it, dispatch it on its model tier, interpret the outcome, and advance — batching decisions and blockers instead of interrupting
| name | rdm-revise |
| description | Act on document reviews requesting changes — work each comment and drive the review to addressed |
| allowed-tools | ["Read","Bash","Glob","Grep"] |
Act on submitted document reviews with verdict request-changes: work through each comment, apply the requested edits through rdm update commands, and drive the review to addressed. This is not rdm-review (which reviews implementations) — this skill acts on document reviews of roadmaps, phases, and tasks.
$ARGUMENTS may name a specific review id; when empty, work every review in the queue.
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.
./target/debug/rdm review requests --format json --project rdm
This lists submitted reviews with verdict request-changes. If $ARGUMENTS names a review id, work only that review; otherwise work each review in turn.
./target/debug/rdm review show <review-id> --format json --project rdm
Read the review's summary (body) before touching any comment: it carries the reviewer's overall intent, and each comment is an instance of that intent, not an isolated request. The JSON also carries, per comment, the stored anchor (a tagged union on anchor_type) and a computed resolution.
For each comment with status: open, branch on the anchor:
anchor_type: text-quote, resolution.state: resolved — the quoted span was located. resolution.range_start/range_end index the body version named by resolution.body: original means the document as of the review's created_commit (re-read it with --at <created_commit> if needed), current means the document at HEAD. Make the targeted edit at that span.anchor_type: text-quote, resolution.state: drifted — the document changed since the review; the range indexes the created_commit body, and resolution.quote is the text the reviewer saw. If you can confidently map the intent onto the current text, edit it and note the drift in your reply. If not, do not guess — go to step 6.resolution.state: unresolved, or an unrecognized anchor_type — treat as whole-document feedback: read the target document (or the comment's doc phase, for roadmap reviews) in full, apply the change in context, and note in your reply that no anchor was resolved.Apply the change through the matching update command (never edit plan files directly). Pass the full revised body with --body — update does not read stdin, so a heredoc is silently ignored. Capture the multiline body into a shell variable with a quoted heredoc (which keeps backticks, $, and punctuation literal), then pass --body "$body":
body=$(cat <<'EOF'
...full revised body...
EOF
)
./target/debug/rdm phase update <stem-or-number> --roadmap <slug> --body "$body" --no-edit --project rdm
# or: ./target/debug/rdm task update <slug> --body "$body" --no-edit --project rdm
# or: ./target/debug/rdm roadmap update <slug> --body "$body" --no-edit --project rdm
--body is authoritative: rdm takes the value verbatim and never touches stdin.
Each mutation only stages the change — it does not commit. Land it immediately, before moving to the next comment:
./target/debug/rdm commit -m "docs(plan): address review comment"
Then capture the SHA of that commit immediately, before any review update (which creates its own commit):
git -C "$RDM_ROOT" rev-parse HEAD
Commit per comment, not per batch: rdm commit lands every currently staged change as one commit, so if you edited more than one comment's target before committing, that single SHA would no longer identify which comment it resolved. Commit after each comment's edit, before starting the next, to keep a 1:1 comment→commit provenance trail — do not batch multiple comments' edits for efficiency here.
./target/debug/rdm review update <review-id> --comment <n> --status addressed --applied-commit <sha> --reply "What changed, and whether the anchor resolved (note drift if any)." --project rdm
Always pass --applied-commit explicitly with the SHA captured in step 4 — never a later HEAD, which may be a review-file-only commit.
Set a reply asking for clarification, leave the comment open (no --status), and continue with the remaining comments:
./target/debug/rdm review update <review-id> --comment <n> --reply "The quoted text changed since the review — did you mean X or Y?" --project rdm
When a comment should not be acted on, say why and close it without a commit:
./target/debug/rdm review update <review-id> --comment <n> --status wont-fix --reply "Reasoning for not making this change." --project rdm
When no comments remain open:
./target/debug/rdm review update <review-id> --state addressed --project rdm
If any comment was left open for clarification, leave the review submitted (the transition would be refused anyway) and report the open comment ids and the questions you asked.
addressed or wont-fix) must carry a reply explaining the decision.addressed without an actual applied change and its commit SHA.anchor_type values are expected (written by newer rdm versions) — they are whole-document comments, not errors.