| name | waygent |
| description | Translate natural-language Waygent run, status, explain, resume, verify, and apply requests into stable CLI commands. |
Waygent
Use this skill when the user asks to run, inspect, resume, explain, verify, or
apply a Waygent execution from natural language.
Waygent is the product runtime. This skill translates operator intent into the
waygent CLI and then reports the command outcome. It must not implement
scheduling, provider execution, worktree mutation, trust scoring, or direct
AgentLens writes.
Hard boundaries:
- Waygent must not call
skills/kws-codex-plan-executor.
- Waygent must not call
skills/kws-claude-multi-agent-executor.
- KWS executor skills are not Waygent product dependencies.
- New Waygent runs use
platform.*, runway.*, kernel.*, and lens.*
event families.
Invocation boundary:
- If this skill is explicitly invoked and the user asks to implement, execute,
run a plan, or run multi-agent work from a design/plan, treat that as a
Waygent runtime run request.
- Do not use host
spawn_agent, direct file edits, or chat-managed workers as
a substitute for waygent run.
- The Waygent runtime owns worktree creation. If no Waygent run is created, no
Waygent worktree should be expected.
- If the request names both a design file and an implementation plan, pass the
plan with
--plan and the design/spec with --spec.
- If the user says "멀티에이전트", "multi-agent", or similar, include
--execution-mode multi-agent.
waygent run auto-detects the host and picks the default provider when
--provider is not passed: Claude Code (CLAUDECODE=1 or
CLAUDE_CODE_ENTRYPOINT set) → claude; Codex app/CLI (CODEX_APP=1,
CODEX_CLI=1, or CODEX_ENTRYPOINT set) → codex; unknown host → codex
(preserved fallback). Set WAYGENT_HOST=claude|codex to force a host.
Execution mode defaults to multi-agent in every case.
- Explicit
--provider always wins over auto-detection.
waygent demo is the deterministic offline path and only supports the fake
provider. Use waygent run for Codex or Claude provider execution.
- The repo-local fallback command is
bun run waygent -- run ...; it maps to
the same CLI as waygent run ....
Host-agent model policy:
- When the host agent is asked to implement, review, or coordinate Waygent
runtime work from a plan or design, the main coordinating agent should run
with the strongest reasoning the host supports.
- Codex app / Codex CLI host:
- Main coordinating agent: extra-high reasoning.
- Implementation / review / verification subagents: prefer GPT-5.5 with high
reasoning when the host supports explicit subagent model settings.
- Claude Code host:
- Main coordinating agent: Opus with high reasoning.
- Implementation / review / verification subagents: Opus with high reasoning
when the host supports explicit subagent model settings.
- If the user explicitly names a different model in the request (e.g. "use
sonnet", "gpt-5 only", "소넷으로"), honor the user's explicit choice over
these defaults for that invocation.
- If the host cannot change the main agent or subagent model settings, state
that limitation and use the strongest available configuration instead.
- This policy is a host-agent execution preference. It must not turn Waygent
into a dependency on KWS executor skills, authorize host
spawn_agent as the
implementation path, or bypass the Waygent CLI/runtime boundaries above.
Default mappings:
Natural-language mappings are versioned in
references/nl-lexicon.md (waygent.nl_lexicon.v1). Keep explicit CLI
flags and explicit command names higher priority than NL interpretation.
- "최근 승인된 플랜 실행해줘" ->
waygent run --latest
- "design.md plan.md 멀티에이전트로 구현해줘" ->
waygent run --plan plan.md --spec design.md --execution-mode multi-agent
- "최고 품질로 실행해줘" ->
waygent run --plan plan.md --spec design.md --profile max-quality
- "비용 절약 모드로 실행해줘" ->
waygent run --plan plan.md --profile cost-saver
- "이 run id로 다시 실행해줘" ->
waygent run --plan plan.md --run <run_id>
- "상태 보여줘" ->
waygent status --last
- "이벤트 보여줘" ->
waygent events --run <run_id> --json
- "자세히 검사해줘" ->
waygent inspect --run <run_id> --json
- "왜 막혔어?" ->
waygent explain --last
- "재개해줘" ->
waygent resume --last
- "검증 다시 실행해줘" ->
waygent verify --last
- "이 태스크만 검증 다시 실행해줘" ->
waygent verify --last --task <task_id>
- "리뷰 실행해줘" ->
waygent review --last
- "이 태스크 리뷰해줘" ->
waygent review --last --task <task_id>
- "검증 통과한 것만 적용해줘" ->
waygent apply --run <run_id>
- "고아 런 정리해줘" ->
waygent orphans (delete via
waygent orphans --delete <id> --yes)
- "stale run 정리해줘" ->
waygent orphans --stale
- "stale run 막힌 상태로 표시해줘" ->
waygent orphans --mark-blocked <run_id>
- "실행 가능한 waygent-task 만들어줘" ->
waygent scaffold-plan --id <task_id> --title <title> --claim <path:mode> --risk <low|medium|high> --verify <command>
- "design 검증해줘" ->
waygent lint-design --path design.md
- "plan 검증해줘" ->
waygent lint-plan --path plan.md
--profile presets:
max-quality: main=opus/high, subagents=opus/high.
balanced: main=opus/high, subagents=sonnet/medium.
cost-saver: main=haiku/medium, subagents=sonnet/medium.
The preset model names above are Claude-shaped. In a Codex run, max-quality
is provider-aware: main=gpt-5.5/xhigh, implement/review/verify/repair
subagents=gpt-5.5/high, with full plan preflight, manifest spec slicing,
builtin hooks, and required method evidence unless explicitly overridden by
supported CLI flags.
Explicit --main-model, --main-reasoning, --subagent-model, and
--subagent-reasoning override any preset.
Plan-author guidance — verify_isolation (optional):
Each task may declare verify_isolation: "isolated" | "fast" | "auto". The
default is "auto". Use "isolated" when the verify command must observe
the worker's cross-package changes (most integration tests). Use "fast"
to opt out of automatic escalation when you are certain the diff is
self-contained. See docs/operations/verification.md for failure surface
and kill switches.
Closeout loop:
- After a Waygent run, apply, resume, or implementation-producing command
changes code or docs, inspect
git status --short --branch --untracked-files=all
before declaring the work complete.
- Run the smallest verification gate that proves the changed surface. Use
git diff --check for docs-only changes, add the offline Waygent gates for
runtime changes, and add console or native gates only when those surfaces
changed.
- If verification mutates tracked files after staging, restage the generated
changes and rerun
git diff --check before any commit or final
completion claim.
- Report unresolved dirty state explicitly, separating pre-existing user
changes from the current Waygent work.
Stop rules:
- If the plan is missing or
--latest is ambiguous, ask for the plan path.
- If scaffold inputs do not include explicit file claims, risk, and
verification commands, ask for those fields instead of inferring
apply-capable write scope from prose.
- If apply reports
dirty_source_checkout, report the blocker and do not retry.
- If verification fails, use
waygent explain --last before waygent verify
or waygent resume.
- If a run reports
review_evidence_missing, run or recommend waygent review
instead of rerunning verification.
- If
resume does not report apply_verified_checkpoint, do not run apply;
inspect or explain the run first.
- If apply reports
checkpoint_manifest_missing, checkpoint_patch_missing,
or checkpoint_digest_mismatch, report the blocker and do not retry from
chat.
- If a run reports
intake_decision_required, explain the specific blocker
and ask only the short question from the operator decision. Do not rewrite
the plan from chat unless the user approves the change.
- If a run reports
invariant_violation_predispatch, prescriptive_drift,
policy_ack_missing, or stale_test_candidates_missing, the normalized
design contract is the source of truth. Reference the normalized_design
artifact when explaining the blocker.
- If intake recovery reports
recovered, proceed with the run result and
mention the normalized plan and recovery report artifact refs when useful.