Run a high-thoroughness end-to-end implementation loop using clean worker worktrees, per-slice planning, implementation, explicit review/fix iteration, orchestrator integration, and final commit/PR drafting. Use when the user explicitly wants a Ralph/Wiggum-style loop, autonomous multi-agent execution, a delegated feature/bugfix carried from confirmed scope through review-ready completion, or RFC-driven implementation that should be reviewed, bumped to In Progress, decomposed into phases, and executed through nested sub-loops. Supports Codex subagents by default and external CLI workers such as OpenCode when explicitly requested and available.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Run a high-thoroughness end-to-end implementation loop using clean worker worktrees, per-slice planning, implementation, explicit review/fix iteration, orchestrator integration, and final commit/PR drafting. Use when the user explicitly wants a Ralph/Wiggum-style loop, autonomous multi-agent execution, a delegated feature/bugfix carried from confirmed scope through review-ready completion, or RFC-driven implementation that should be reviewed, bumped to In Progress, decomposed into phases, and executed through nested sub-loops. Supports Codex subagents by default and external CLI workers such as OpenCode when explicitly requested and available.
Ralph Loop
Use this skill when the user wants more than basic delegation. The goal is not "parallelize some work"; the goal is to carry a scoped task through planning, implementation, review, integration, and ship-ready artifacts with explicit backpressure.
Keep orchestrate-parallel-work generic. Use this skill as the opinionated wrapper around it.
Core stance
Confirm the requested scope before spawning anything.
If the input is an RFC, treat RFC hygiene and lifecycle state as part of the work, not pre-existing setup.
Treat user-facing docs and versioning as part of implementation, not optional closeout polish.
Optimize for end-to-end correctness, not maximal concurrency.
Use clean worktrees for real implementation slices.
For milestone, RFC-wide, or compiler-boundary work, define the acceptance contract before implementation starts. The contract should name required boundary parity coverage, downstream acceptance lanes, docs/generated-reference gates, performance/progress gates, and any criteria that must be satisfied before an RC or publish step.
For language or stdlib work that is meant to be implemented in Incan, dogfood Incan rather than creating a thin .incn facade over a custom Rust backend. Direct from rust:: imports of existing crates/primitives are acceptable when the .incn module still owns the behavior; bespoke incan_stdlib::feature Rust modules that hide the feature logic are not acceptable unless the maintainer explicitly asks for that backend boundary.
Do not let workers assume Incan cannot express something. Before choosing Rust/backend fallback or narrowing a design, workers must inspect current .incn precedents, parser/typechecker/codegen tests, or run a small probe and record the specific capability evidence.
Every implementation must land in a fresh worktree rooted under WORKTREE_ROOT so VS Code picks it up; do not implement in the system /tmp or in the main repo checkout.
Treat each slice as a managed work packet with durable state on disk, not as a chat thread that has to remember its own scope.
Treat loop identity as durable state too. A resumed loop must prove that persisted state matches the current user request before using it to choose work.
Treat cross-repo consumer work as a verification lane unless the user explicitly puts that repository's implementation work in scope.
Require every worker to plan, implement, verify, review, and fix within its owned slice.
Consolidate accepted worker output onto the orchestrator branch before any commit, push, or PR work. Worker worktrees are temporary execution sandboxes, not final publication branches.
Retire worker worktrees after their done output has been integrated and verified unless the orchestrator records a concrete reason to keep one.
Require the orchestrator to perform its own explicit plan -> do -> check -> act integration loop after collecting worker output.
Require workers and the orchestrator to maintain a persistent review report at .agents/state/review-report.md inside each worktree so findings survive across loops.
Require touched .incn source to pass review-incan-source-quality; Incan implementation work is not done when it merely compiles if the authored source reads like Rust-shaped scaffolding instead of well-written Python.
Do not commit, push, or open a PR unless the user explicitly asked for that. When not explicitly asked, still draft the commit message and PR description as ready-to-use artifacts.
PDCA model
ralph-loop is a plan -> do -> check -> act loop, not a linear pipeline.
Use these meanings consistently:
Plan: confirm scope, define slices, create task state, decide the next concrete increment
Do: implement the planned increment in the owned slice
Check: verify behavior, run review, compare delivered work against the requested scope
Act: either accept and move forward, repair findings, or go back to planning when scope is not actually satisfied
Important:
Review findings that are simple defects belong in Act -> fix.
Evidence that the slice or integration output does not satisfy the intended scope belongs in Act -> re-plan.
Do not treat “missing promised behavior” as just another lint item. That is a planning/state problem first.
Scope and state isolation
Every Ralph loop must have a persisted loop identity before workers are spawned or resumed. Record it in the orchestrator overview and use it to filter durable state:
loop_id: stable slug for this run, usually <repo>-<issue-or-rfc-or-milestone>-<short-purpose>
target repository or repositories
primary implementation repository
base branch or resolved starting commit
issue/RFC/milestone allowlist
explicit non-goals
permitted downstream or consumer verification lanes
whether cross-repo implementation is allowed
Use STATE_ROOT to mean the active loop's state directory. For new loops, prefer a namespaced root:
.agents/state/ralph-loop/<loop-id>/
Legacy top-level state such as .agents/state/ralph-loop/slices.json may be read for migration or handoff, but must not silently drive a new loop unless its recorded loop identity matches the current request.
On every resume, stop-hook continuation, compaction, or long-running restart:
Re-read the latest user instruction.
Re-read STATE_ROOT/overview.md, STATE_ROOT/slices.json, and .agents/state/review-report.md.
Compare persisted loop identity to the latest user instruction.
Classify each slice as:
in_scope: same loop identity and allowed implementation repo
verification_lane: downstream/consumer repo explicitly named as validation only
out_of_scope: different repo, issue, milestone, branch, or objective
Schedule only in_scope implementation slices.
Out-of-scope or historical slices may be cited as background evidence, but they must not be advanced, committed, pushed, or turned into PRs from the current loop. If the user wants to pick them up, start or resume a separate Ralph loop with its own loop identity.
Consumer repositories such as IncQL may be used to reproduce, verify, or acceptance-test an Incan compiler/runtime change only when listed as a verification lane. Adding product features, docs CI, repo automation, or cleanup in that consumer repo is a separate implementation loop unless the user explicitly asks for it.
When not to use this skill
Do not use this for:
a small task where one agent can finish faster than the orchestration overhead
a tightly coupled design problem with one unresolved architectural decision
multiple workers that would need to edit the same files or shared API surface
tasks where the user wants a plan only
Workflow
0. RFC intake mode
If the user feeds the loop an RFC, enter RFC intake mode before ordinary execution.
In RFC intake mode:
Run review-rfc on the RFC and apply direct fixes.
Inspect Status: and the design tail honestly.
If unresolved questions remain, summarize them and ask the user before bumping anything.
Check for process blockers that require user input, such as:
external dependencies not yet available
another RFC or issue that partially blocks the scope
scope that should be split before implementation
If the RFC is still Draft and design is settled, use bump-rfc to move it to Planned.
If implementation is actually being picked up now, use bump-rfc to move it to In Progress. For this skill, the parent Ralph loop itself counts as "a contributor has picked up the work"; a child loop does not need its own PR or branch-level ceremony.
Use the RFC's Implementation Plan and Progress Checklist as the source of truth for implementation phases. If they are missing or weak, strengthen them before spawning workers.
Establish the docs/version baseline up front: verify the repo's actual dev version from the source-of-truth metadata, identify which user-facing docs must change if the feature lands, and do not assume an older release line from stale release notes or worker worktrees.
Do not quietly force an RFC past open design questions. Stop and ask the user.
1. Confirm the scope
Restate the requested end-state before starting execution. Confirm:
target repository or repositories
primary implementation repository
issue / RFC / branch context
milestone allowlist, if milestone-driven
explicit goals
explicit non-goals
downstream repositories that are verification lanes only
cross-repo implementation permission, if any
whether the task is RFC-wide or only a subset of RFC phases
whether commit / push / PR creation were requested
whether the user wants Codex subagents or an external backend such as OpenCode
If scope is ambiguous, stop and ask a short numbered list of missing decisions.
After confirming scope, write the loop identity into STATE_ROOT/overview.md before creating or resuming slices. If existing durable state lacks a matching loop identity, do not continue it as part of this loop.
2. Pick the execution backend
Default to Codex subagents plus git worktrees under WORKTREE_ROOT.
Resolve WORKTREE_ROOT once at the start of the loop and record the absolute path in STATE_ROOT/overview.md. Use an explicit RALPH_WORKTREE_ROOT when the workspace provides one; otherwise derive a portable shared tmp/ directory from Git's common directory:
git-common-dir keeps the default stable when the loop runs from a linked worktree.
If the user explicitly wants OpenCode:
verify opencode is installed and callable
verify auth / provider setup is present enough to run unattended
prefer non-interactive execution such as opencode run ... or a preconfigured OpenCode agent profile
do not rely on the TUI for unattended worker execution
If OpenCode is requested but unavailable, say so plainly and either stop or fall back to Codex only with the user's approval.
3. Prepare the orchestration boundary
Use start-work once at the orchestration boundary to resolve issue/RFC context, branch naming, and relevant learnings.
Do not mechanically run start-work inside every worker unless each worker owns a distinct issue or RFC. The important requirement is a clean worktree plus resolved context, not duplicated branch ceremony.
Create WORKTREE_ROOT first if it does not exist.
Create:
one orchestrator worktree for final integration
one clean worker worktree per implementation slice
Create durable slice state under each worktree. The paths below are relative to STATE_ROOT; for new loops STATE_ROOT should normally be .agents/state/ralph-loop/<loop-id>:
overview.md for orchestrator-wide state, including loop identity and scope boundaries
STEERING.md for operator overrides, urgency changes, and mid-run guidance
slices.json for the orchestrator-owned slice registry
<slice-id>/scope.md
<slice-id>/plan.md
<slice-id>/tasks.json
<slice-id>/tasks.md
<slice-id>/status.md
<slice-id>/handoff.md
Treat the slice folder as the source of truth for that slice. Do not rely on memory or only on conversational context.
Use these files deliberately:
STEERING.md: human or orchestrator overrides that should preempt normal task ordering
slices.json: orchestrator registry of all active slices, their owners, and their current state
tasks.json: machine-readable task list for the slice
tasks.md: human-readable task narrative and notes
status.md: current PDCA state, blockers, and next action
Base all worker worktrees from the same resolved starting point unless there is a deliberate dependency chain.
For non-decomposed work, the single-agent implementation still belongs in a fresh worktree under WORKTREE_ROOT; "keep the work local" does not mean "edit the primary checkout directly."
Before spawning workers, identify:
the source-of-truth version file(s) for the repo
whether the task is on a -dev.N line and therefore needs a version bump
the authored user-facing docs that must be updated if the change is user-visible
For milestone, RFC-wide, or compiler-boundary work, also write an ## Acceptance Contract section into STATE_ROOT/overview.md before implementation starts. Include:
the local/direct behavior that must work,
import, reexport/facade, package-consumer, test-batch, dependency, vocab, or generated-Rust boundaries that can observe the behavior,
downstream acceptance lanes such as IncQL when the surface is exercised there,
performance or progress-output expectations when the change touches prewarm, test runner, metadata, or CI paths,
docs, rustdocs, generated-reference, release-note, or RFC lifecycle gates,
explicit non-applicable boundaries, with a short reason.
Do not defer this acceptance contract to release-branch closeout. The release branch should verify the contract, not discover it.
Do not treat RFC edits or release notes alone as sufficient user documentation.
4. Decide whether parallelism is justified
Before spawning workers, decide whether the task actually decomposes cleanly. If it does not, keep the work local and continue as a single-agent Ralph loop.
When it does decompose, hand off to orchestrate-parallel-work for slice definition, ownership, and worktree isolation under WORKTREE_ROOT.
If the task came from an RFC:
derive slices from RFC implementation phases or coherent checklist groupings, not arbitrary percentages
keep parent ownership of RFC lifecycle edits, progress-checklist updates, commit text, and PR drafting
treat child loops as implementation subloops only
allow nested decomposition only one level down: child loops may use orchestrate-parallel-work for leaf workers inside their owned scope, but they must not spawn further ralph-loop children
For each slice, create tasks.md with explicit task items. Keep them concrete and finishable, not vague “phase done” markers.
Also create tasks.json for the same slice. tasks.json is the authoritative status surface; tasks.md is the human-readable companion.
Subagents should work tasks to completion against these files and update them as state changes.
The orchestrator should maintain slices.json with entries like:
[{"loop_id":"incan-557-std-environ","slice_id":"lifecycle-env","repo":"encero-systems/incan","scope_role":"implementation","owner":"<worker name or id>","worktree":"<WORKTREE_ROOT>/...","status":"doing","next_action":"Implement T2"}]
Use scope_role: "verification" for downstream acceptance lanes. Verification slices may run checks and collect evidence, but they must not mutate consumer product code unless the loop identity explicitly allows cross-repo implementation.
5. Give each worker a strict end-to-end contract
Each worker must own a non-overlapping slice with:
loop identity
repository and scope role (implementation or verification)
exact goal
owned files or directories
explicit non-goals
forbidden repositories or paths
dedicated worktree path under WORKTREE_ROOT
dedicated slice folder under STATE_ROOT/<slice-id>/
verification command
expected result format
For RFC-driven work, prefer one child loop per implementation phase or tightly related checklist group.
Each worker must perform this loop inside its slice:
Plan
Build a short slice plan using create-plan.
For .incn, stdlib, or language-surface work, perform capability intake before ruling out an Incan-native design. Record the local precedent, test, or probe result in plan.md.
Write scope.md, plan.md, tasks.json, and tasks.md.
Break the slice into concrete tasks inside tasks.json and keep tasks.md as the readable companion.
Set initial slice state in status.md as planned.
Do
Implement the next planned task set, not the whole universe.
Update tasks.json, tasks.md, and status.md as task ownership and progress change.
Use doing as the active execution state.
Check
Run targeted verification for the slice.
Run review-incan-source-quality when the slice touches .incn source.
Run review on the slice in report-only mode, or review-orchestrate if the slice itself is broad enough to justify specialization.
Compare the delivered behavior against scope.md, not only against test output.
Use checking as the active verification/review state.
Act
Run fix on every actionable in-scope blocker or warning.
If the slice hits a likely compiler bug that is out of scope, invoke flag-compiler-bug before reporting the blocker or choosing a workaround.
If check/review shows that promised scope is still not satisfied, change slice state to replan_required and go back to Plan. Update scope.md, plan.md, tasks.json, and tasks.md before doing more code.
If outside help is needed, set slice state to blocked with a concrete blocker.
If the slice is actually complete, set slice state to done and write handoff.md.
Repeat until:
no actionable in-scope items remain,
the slice tasks are complete,
and scope.md is honestly satisfied,
or report a concrete blocker with its classification.
The slice's .agents/state/review-report.md must be kept current throughout this loop.
The slice folder under STATE_ROOT/<slice-id>/ must also stay current throughout this loop.
Allowed slice states:
planned
doing
checking
replan_required
blocked
done
Do not invent vague alternatives like “mostly done” or “almost ready”.
Workers must be told:
they are not alone in the repo
they must not revert or overwrite others' work
they must adapt to concurrent changes
they must not produce PRs, PR descriptions, or final commit artifacts of their own when they are child loops under a parent RFC loop
child loops must not spawn further ralph-loop children; if they need extra decomposition, they may use orchestrate-parallel-work only for leaf-level workers within their owned scope
they must not commit or push unless the user explicitly asked for that
Require every worker to return the shape in reference.md.
6. Integrate centrally
The orchestrator owns integration. Workers do not integrate each other.
The orchestrator must:
inspect each worker result and changed-file list
inspect each worker slice folder, not just the final prose summary
inspect slices.json and ensure every in-scope slice has an honest terminal or active state
ignore or migrate historical slices whose loop_id, repo, branch, issue, milestone, or scope role does not match the current loop identity
reconcile naming, docs, tests, and architectural seams across slices
ensure user-facing docs were updated for user-visible behavior, not only RFC text or release notes
verify the repo version baseline again before finish and bump -dev.N by one at minimum for implementation work on the active dev line
update RFC progress state and checklist items as phases land
move the accepted work into the orchestrator worktree cleanly
retire each completed worker worktree after its accepted work is present in the orchestrator worktree and the integrated verification gate has passed
run the repo-level gate
run Plan -> Do -> Check -> Act on the integrated result:
Plan: confirm the combined slice outputs still satisfy the original end-state and create/update orchestrator task state in STATE_ROOT/overview.md
Do: integrate the accepted worker results
Check: run verification, review-incan-source-quality for touched .incn source, plus review or review-orchestrate on the integrated result
Act: run fix on actionable in-scope findings, or go back to planning if integration review shows that the original requested scope is still not satisfied
invoke flag-compiler-bug for real out-of-scope compiler defects found during integration
repeat until no actionable integrated items remain and the original scope is honestly satisfied
The orchestrator worktree's .agents/state/review-report.md is the integration source of truth.
The orchestrator's STATE_ROOT/overview.md is the integration state source of truth.
Worker cleanup is part of integration, not optional closeout. For every slice marked done:
Confirm handoff.md, the worker changed-file list, and the accepted patch are reflected in the orchestrator worktree.
Run the relevant integrated verification before removing the worker sandbox.
Record integration metadata in slices.json, such as integrated_at, integrated_by, and worktree_removed.
If the worker worktree is clean, remove it with git worktree remove <worker-path>.
If the worker worktree is dirty only because of files already integrated into the orchestrator branch or task-owned scratch state, record the inventory in overview.md or slices.json, then remove that Ralph-owned worker worktree with git worktree remove --force <worker-path>.
After the worktree is removed, delete the disposable worker branch only when it is task-owned and the orchestrator branch contains the accepted work.
Do not force-remove a worker worktree with unknown or unintegrated changes. Do not push while accepted slice work exists only in a worker worktree or worker branch.
STEERING.md must be checked at the start of every major iteration. If it changes the priority, scope, or urgency of the work, the orchestrator must update slices.json, affected scope.md / tasks.json, and continue from the new direction rather than pretending the original ordering still applies.
If STEERING.md or a resume hook mentions slices from another loop identity, treat them as out of scope until the user explicitly says to resume that separate loop.
Do not stop at "worker green." Cross-slice regressions and consistency problems belong to the orchestrator.
7. Finish with ship-ready artifacts
When code is ready:
confirm all accepted worker slices have been consolidated into the orchestrator branch
confirm completed worker worktrees are removed, or record a concrete reason for any kept worker worktree
produce a concise done summary
draft the commit message with write-commit-message
draft the PR description with create-pr-description
For RFC-driven work, only the parent loop drafts or owns the final PR description. Child loops must not produce PRs of their own.
If the user says "greenlight for PR", "green light for PR", or otherwise approves PR publication for RFC-driven work, treat RFC finalization as part of the pre-PR gate. Before reporting the branch ready for review or opening/marking a PR ready:
run a final scope check against the governing RFC and issue
if the implemented branch covers the full RFC scope, use bump-rfc to move the RFC from In Progress to Implemented
ensure every Progress Checklist item is checked before the bump; unchecked items are scope failures, not residual PR risks
if any checklist item is incomplete, set the orchestrator state to replan_required, update the relevant slice/task state, and continue the Plan -> Do -> Check -> Act loop instead of publishing
ensure the PR body includes a closing keyword for the RFC issue, such as Closes #NNN
rebuild docs after moving the RFC into closed/implemented/
report any reason the RFC cannot be bumped before asking for human review
Do not leave a fully implemented RFC in In Progress while presenting the PR as review-ready. Do not use Refs #NNN as an escape hatch for RFC-driven Ralph-loop work unless the user explicitly changed the scope to a partial implementation. If the branch implements only part of the RFC, say that directly, leave the RFC active, and do not present the PR as a full-RFC closeout.
Only run the actual commit / push / PR creation flow if the user explicitly asked for those actions.
Quality bar
Treat these as default expectations, not optional polish:
full requested scope, not a narrow interpretation that happens to pass tests
Boy Scout cleanup within touched files
tests proportional to risk and surface area
architectural fit with existing boundaries
user-facing docs and release notes when the repo rules require them
version checks up front and a dev-version bump (-dev.N -> -dev.(N+1)) at minimum for implementation work on the active dev line
If the task teaches a durable lesson about orchestration, testing, or worktree hygiene, consider add-learning before finishing.
Relationship to other skills
start-work: use once to resolve context and branch strategy before decomposition
review-rfc: use first in RFC intake mode
bump-rfc: use to move a settled RFC into Planned and then In Progress before phase execution, and to move fully implemented RFC-driven work to Implemented before PR review/publication; stop and ask the user if open questions or blockers remain
create-plan: each worker uses it for its own slice; the orchestrator may also use it if the whole task still needs a settled plan
orchestrate-parallel-work: use it for decomposition, worker ownership, and isolation
review: report-only detector for smaller worker slices and local integration checks
review-orchestrate: preferred detector for broad integrated outputs or slices that are themselves wide enough to justify specialized reviewers
review-incan-source-quality: required detector for touched .incn source; use it to enforce the well-written-Python readability bar and dogfooding integrity before declaring a slice or integration clean
fix: mandatory repair pass after review findings
review-and-fix: allowed as a convenience wrapper when a worker or the orchestrator wants the combined loop explicitly
write-commit-message: use for the final commit text
create-pr-description: use for the final PR body
Nesting rule
Use this shape:
parent ralph-loop
child ralph-loops for major RFC phases when justified
orchestrate-parallel-work leaf workers inside a child phase when needed
Do not recurse ralph-loop indefinitely. A child loop is a phase owner, not another top-level orchestrator.
Validation checklist
If the input was an RFC, review-rfc ran first
If the input was an RFC, unresolved questions and process blockers were surfaced to the user before bumping
If the input was an RFC, bump-rfc moved it to In Progress only after design was settled and work was actually being picked up
If the input was an RFC, child loops were derived from implementation phases or checklist groups
Child loops did not spawn further ralph-loop children
Scope was restated and confirmed before execution
Loop identity was written before workers were spawned or resumed
Existing durable state was filtered by loop identity before choosing work
Backend choice was explicit
RFC lifecycle state and implementation plan/checklist were confirmed before coding started
Every implementation worker had a clean worktree and non-overlapping ownership
Every implementation worktree lived under the resolved WORKTREE_ROOT
Every slice had a durable folder under STATE_ROOT/<slice-id>/
The orchestrator maintained STATE_ROOT/slices.json
Every slices.json entry has loop_id, repo, and scope_role
Consumer/downstream repositories were used only as verification lanes unless explicitly allowed for implementation
The orchestrator checked STATE_ROOT/STEERING.md at each major iteration
Every slice kept explicit scope.md, plan.md, tasks.json, tasks.md, status.md, and handoff.md
Docs/version baseline was established from repo source-of-truth metadata before implementation
Incan-language/stdlib slices recorded capability evidence before ruling out Incan-native implementation
Every worker ran a real plan -> do -> check -> act loop
Every slice used only the allowed explicit states: planned, doing, checking, replan_required, blocked, done
Scope failures were routed back to planning instead of being treated as ordinary defect cleanup
The orchestrator ran its own integration plan -> do -> check -> act loop
Accepted worker output was consolidated onto the orchestrator branch before final artifacts
Completed worker worktrees were removed, or each kept worker worktree has a recorded reason
No accepted slice work exists only in a worker worktree or disposable worker branch
Every worker maintained .agents/state/review-report.md in its worktree
Every worker or orchestrator touching .incn source ran review-incan-source-quality
The orchestrator maintained .agents/state/review-report.md in the integration worktree
The orchestrator maintained STATE_ROOT/overview.md
User-visible behavior changes updated authored user docs, not only RFCs/release notes
Active dev version was re-checked and bumped by one dev increment at minimum for implementation work
Child loops did not draft PRs or final commit artifacts of their own
Every RFC Progress Checklist item was checked or the orchestrator stayed in replan_required
Final gate passed or remaining failures were reported concretely
If the implementation satisfies the full RFC, bump-rfc moved it to Implemented before the PR was presented as review-ready
Full-RFC PR bodies use a closing keyword such as Closes #NNN; partial-scope PRs explain the remaining RFC scope instead
Commit/PR artifacts were drafted
No commit/push/PR action was taken without explicit user permission
