From bender
Orchestrates the full roadmap-to-committed-code flow. Iterates through milestones — implementing, testing, reviewing, and committing each — with human approval gates.
How this skill is triggered — by the user, by Claude, or both
Slash command
/bender:pipeline [--auto] [roadmap file path] (optional — --auto skips human review gate per milestone; roadmap auto-discovers if not specified)[--auto] [roadmap file path] (optional — --auto skips human review gate per milestone; roadmap auto-discovers if not specified)The summary Claude sees in its skill listing — used to decide when to auto-load this skill
**ROADMAP IN, COMMITTED CODE OUT. EVERY MILESTONE PASSES CI AND REVIEW BEFORE COMMITTING.**
ROADMAP IN, COMMITTED CODE OUT. EVERY MILESTONE PASSES CI AND REVIEW BEFORE COMMITTING.
Do not skip steps, do not reorder phases, do not commit without human approval — unless --auto mode is active.
Parse mode flag. If $ARGUMENTS contains --auto, enable auto mode (skip Phase 4 human review gate). Strip the flag before interpreting the rest as a file path.
Locate the roadmap. If $ARGUMENTS provides a file path, read that file. Otherwise auto-discover in this order:
ROADMAP.mdMILESTONES.mddocs/ROADMAP.mddocs/MILESTONES.md.github/ROADMAP.mdIf no roadmap is found after checking all locations, stop immediately. Tell the user no roadmap was found and suggest running /bender:roadmap first.
Read recent git history. Run git log --oneline -15 to understand what has been completed recently.
Check for work in progress. Run git status to detect uncommitted changes or in-progress work.
Check project conventions. Read CLAUDE.md if it exists, then README.md, to understand project patterns, constraints, and file maintenance protocols.
Discover CI command. Use this fallback chain (stop at first match):
bin/ci exists → use itCLAUDE.md specifies a CI/test command → use itpackage.json with scripts.test → npm test (or yarn test/pnpm test/bun test based on lockfile)Makefile with ci or test target → make ci or make testCargo.toml → cargo testGemfile → bundle exec rspecgo.mod → go test ./...pyproject.toml or setup.py → pytestParse dependency graph. Scan milestones for IDs (M<n>: prefix) and {depends: ...} markers:
{depends:} markers → milestones without markers are independent[x] items satisfy dependenciesParse the roadmap for incomplete - [ ] items.
Evaluate the dependency graph to find milestones whose dependencies are all satisfied.
If 2+ milestones are independent and ready, ask the user one question: run them in parallel worktrees or sequentially?
pipeline/<milestone-id>). Max 3 parallel worktrees at once; batch the rest.If only one milestone is ready, select it and proceed.
Identify affected files, apply the narrow interpretation rule, write a 3–5 line summary (what, approach, files, out-of-scope). No pause for approval — pipeline trusts roadmap specificity.
Follow standard bender practices: TDD, verification before completion claims, systematic debugging if anything breaks. No scope creep — each milestone implements exactly what the roadmap says.
Execute the discovered CI command.
/bender:code-reviewIf --auto mode is active: skip this phase entirely. Log the summary below to the conversation for traceability, then proceed directly to Phase 5.
Otherwise, this is the only pause point in the pipeline. Present a summary:
git diff --stat outputWait for user response:
Fidelity check. Re-read the milestone item verbatim from the roadmap. List what was implemented. Confirm the two match. If they don't → flag the delta to the user before committing.
Commit. Stage all changes and create a single conventional commit: type(scope): description. Do not push unless the user asks.
Mark done. Update the roadmap: change - [ ] to - [x] for the completed milestone.
Doc sync. Read CLAUDE.md for any file maintenance protocol. Update referenced docs as needed. If no protocol is found, only update the roadmap checkbox. All doc changes go in the same commit (amend the previous commit).
Return to Phase 2 to select the next ready milestone.
If parallel worktrees were used:
--no-ff)When all milestones are complete: present a final summary:
--auto mode is active, in which case CI + code review must still pass.type(scope): description.Guides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.
npx claudepluginhub bitfootco/bender --plugin bender