Development Lifecycle Router Maintainer
Use this skill as a routing layer: choose the current lifecycle stage, name the
exit bar, and hand off to the smallest capable maintainer skill or normal coding
workflow.
The reason this skill exists is to prevent two expensive mistakes:
- jumping from a complaint into implementation before the correct stage is known
- calling a stage complete without the product proof or reviewer evidence that
the user's request actually requires
Keep this file thin. Load the reference files below only when the current route
needs them.
Reference Map
references/route-selection.md: stage classifier, narrow-skill routing,
meta-request precedence, and skill-optimization boundaries.
references/verification-review.md: verifier gates, terminal product proof,
spawned reviewer policy, reviewer lenses, agent prompt templates, and evidence
ledger.
references/special-routes.md: recovery, component lab, performance
benchmark, runtime/provider contracts, and hard real-local validation.
references/handoff-git.md: git safety, final handoff shape, acceptance
blockers, and common route templates.
Fast Start
Before editing files, running long checks, spawning reviewers, or committing:
- Inspect
git status --short --branch.
- Classify the current stage.
- State the selected route and downstream owner in one concise update.
- Name the artifact or proof required to leave the current stage.
- For development work, state the full loop before editing:
implementation, writer checks, verifier, final reviewers, reconciliation, and
handoff. Development work enters
spawn-required mode by default; the user
does not need to separately ask for subagents, review, or verification.
- Decide which references are needed for this route and read only those.
For obvious narrow requests, do not expand a lifecycle plan. Say the route,
load the narrow skill, and execute it.
Non-Use Gate
Use the narrow maintainer directly when all of these are true:
- the prompt names a concrete surface, run, PR, release, screenshot, data path,
or local runtime problem
- the next useful artifact is obvious for that surface
- the task does not need cross-stage sequencing, reviewer orchestration, or
destructive recovery judgment before the narrow work can begin
Keep ownership in this router only when it adds value by choosing a stage,
resolving ambiguity, sequencing multiple stages, or protecting a high-risk
handoff.
Core Rule
Route first, then execute.
State the lifecycle stage and acceptance bar before implementation. Any routed
development task enters spawn-required mode by default. This applies when the
task changes user-visible, agent-visible, Desktop, release, runtime, CLI,
workflow, or product behavior, even if the user only says "fix",
"optimize", or "推进". The normal implementation sandwich is:
writer implementation
-> writer basic checks
-> optional lightweight pre-review
-> spawn availability probe if needed
-> spawned verifier black-box acceptance
-> spawned final reviewer gate
-> handoff / commit / push
Pre-review is only for catching obvious diff, startup, safety, scope, or test
readiness problems before verifier time is spent. It is not the final reviewer
gate. Final review follows verifier PASS so reviewers can inspect the diff,
tests, handoff, and verifier evidence together.
Separate verification from review:
- verification asks whether the product path meets the requirement from the
user's side
- review asks whether the diff, architecture, scope, tests, proof, and handoff
are trustworthy
For development routes, do not treat the loop as optional just because the user
said "fix", "优化", "推进", or omitted the word review. User-visible,
agent-visible, Desktop, release, runtime, CLI, workflow, and Rudder
changes need the verifier plus final reviewer loop before complete handoff. If a
route truly does not need that loop, say why it is not applicable rather than
silently skipping it.
spawn-required means:
- after writer checks, probe the available spawn/subagent mechanism before
declaring verifier or reviewers unavailable
- if spawning works, create the verifier and final reviewer agents and wait for
their verdicts before complete handoff
- if spawning is unavailable, record the failed probe as
blocked: spawned verifier/reviewer unavailable
- do not claim complete handoff, commit/push readiness, or review pass from
author-run tests, CI, screenshots, self-review, or serial personas alone
Stage Classifier
Choose one primary stage. If multiple stages are present, choose the earliest
blocking stage.
intake: user intent, target artifact, or mode is unclear
requirements: problem framing, scenarios, acceptance criteria, or "do you
understand?"
advisor: first-principles diagnosis of a build, UI, workflow, trace, or
proposal that feels wrong
ui_design: interface direction, wireframe, visual hierarchy, or
screenshot-based design judgment before code
implementation: approved direction or direct fix/build request
verification: tests, CI, E2E, screenshots, Desktop smoke, actor-run-chain,
release checks, or black-box acceptance proof
review: review, PM judgment, first-principles critique, or
Codex/session/PR/commit verdict
debug: failed run, UI path, data path, CI job, Desktop app, or local process
release: canary/stable release, npm, Desktop assets, tags, or GitHub
Release state
handoff: final summary, validation, commit, push, residual risk, or PR
recovery: dirty worktree, stash, interrupted run, rollback, delete/restore,
or suspected old Codex work
component_lab: UI Lab, component inventory, fixtures, or design-system
coverage
performance_benchmark: benchmark or optimize before the exact bottleneck is
known
runtime_contract: provider/runtime/tool-call/transcript/parser/CLI or
agent-visible contract parity
skill_optimization: optimize, harden, refactor, validate, benchmark,
package, or improve a named skill or workflow skill
For full routing detail, read references/route-selection.md.
Narrow Routes
Prefer the smallest matching owner:
- screenshot-driven UI polish:
rudder-ui-polish-maintainer
- missing, stale, suspicious, or slow page data:
rudder-data-path-diagnostician-maintainer
- run transcript, stdout/stderr, recent run batch, or run-quality issue:
debug-run-transcript-maintainer
- Desktop startup, packaged app, Electron shell, update, profile, local
instance recovery:
rudder-desktop-dev-recovery-maintainer
- release, npm, GitHub Release, Desktop assets, tags, dist-tags, install smoke:
release-maintainer
- review-only session/PR/commit/proposal/UI/release outcome:
agent-work-reviewer-maintainer or
codex-session-product-reviewer-maintainer
- local branch preview:
rudder-worktree-preview-maintainer
- GitHub PR checkout/preview/review:
pr-local-preview-maintainer
- mock/demo/seed data:
mock-data-maintainer
- landing screenshots:
landing-proof-shots-maintainer
- stop/restart/clean local dev runtime:
stop-rudder-dev-maintainer
- new or existing reusable workflow skill creation, optimization, eval,
benchmark, packaging, or description tuning:
skill-creator
If this router itself is the target artifact for optimization, route to
skill-creator or the user-requested skill-engineering workflow. Do not run
this router's normal product lifecycle except for git safety around the patch.
Skill Optimization Boundary
The newest user instruction is the routing source of truth.
If the user says a skill "needs optimization", "should be hardened", "always
does the wrong thing", "I have to ask this every time", or asks to use
skill-creator on a named skill, classify the turn as
skill_optimization.
Treat screenshots, prior requirements, session ids, quoted logs, and prior
assistant recommendations as evidence for the skill failure. They are not the
active product task unless the newest user instruction says to continue that
product task.
For this route:
- name the target skill and path
- extract the failed decision point
- choose the smallest durable owner: target skill body, frontmatter
description, eval case, memory update, or no-op
- add or update a validation case when the behavior should change next time
Verification And Review Defaults
For user-visible, agent-visible, Desktop, release, runtime, CLI, workflow, or
product changes, identify the terminal product surface before calling
verification complete. Use product-acceptance-verifier-maintainer for a
distinct black-box acceptance pass. If the product path is too expensive or
unsafe to exercise, record verifier blocked or substituted with the exact
reason; do not silently skip acceptance.
Spawned verifier and reviewer gates are the default for routed development
artifacts when this router owns the lifecycle. If spawning is available, the
verifier should be a distinct child agent using
agents/product-verifier.md or equivalent product-acceptance-verifier-maintainer
instructions. Final reviewers should be child agents using the templates under
agents/ or equivalent reviewer prompts, with these lenses:
- functional trust
- adversarial
- heuristic/product-systems
Do not replace spawned verifier or reviewer verdicts with self-review, serial
personas, or author-claimed screenshots. If the active runtime cannot spawn
agents after a real availability probe, record blocked: spawned verifier/reviewer unavailable with the probe evidence instead of claiming the loop passed.
For the full verifier/reviewer contract, read
references/verification-review.md.
Execution Outline
- Build a compact routing packet:
- latest user request and corrections
git status --short --branch
- named files, screenshots, sessions, runs, PRs, commits, or plans
- relevant
AGENTS.md route docs and nearby skill contracts
- changed-file ownership if cleanup or recovery is involved
- Declare route and stage exits.
- Execute the current stage using the narrow owner or normal repo workflow.
- Run the required verification and review gates for that route.
- Hand off with evidence, blockers, git state, and commit/push status.
Minimal Handoff Shape
Route: ...
Stage exits: ...
Used: ...
Review: spawned reviewers / blocked / not applicable
Validation: passed / not run / not proven
Evidence: required / scenario / proven / missing or substituted
Git: commit / push
Residual risk: ...
1---2name: development-lifecycle-router-maintainer3description: "Use when a Rudder development request has an unclear lifecycle stage or owner: requirements, advisor analysis, UI design, implementation, verification, review, release, recovery, runtime contracts, performance, component lab, handoff, or named-skill optimization."4metadata:5 source: https://skillmd.com/skills/undertone0809/development-lifecycle-router-maintainer6---78# Development Lifecycle Router Maintainer910Use this skill as a routing layer: choose the current lifecycle stage, name the11exit bar, and hand off to the smallest capable maintainer skill or normal coding12workflow.1314The reason this skill exists is to prevent two expensive mistakes:1516- jumping from a complaint into implementation before the correct stage is known17- calling a stage complete without the product proof or reviewer evidence that18 the user's request actually requires1920Keep this file thin. Load the reference files below only when the current route21needs them.2223## Reference Map2425- `references/route-selection.md`: stage classifier, narrow-skill routing,26 meta-request precedence, and skill-optimization boundaries.27- `references/verification-review.md`: verifier gates, terminal product proof,28 spawned reviewer policy, reviewer lenses, agent prompt templates, and evidence29 ledger.30- `references/special-routes.md`: recovery, component lab, performance31 benchmark, runtime/provider contracts, and hard real-local validation.32- `references/handoff-git.md`: git safety, final handoff shape, acceptance33 blockers, and common route templates.3435## Fast Start3637Before editing files, running long checks, spawning reviewers, or committing:38391. Inspect `git status --short --branch`.402. Classify the current stage.413. State the selected route and downstream owner in one concise update.424. Name the artifact or proof required to leave the current stage.435. For development work, state the full loop before editing:44 implementation, writer checks, verifier, final reviewers, reconciliation, and45 handoff. Development work enters `spawn-required` mode by default; the user46 does not need to separately ask for subagents, review, or verification.476. Decide which references are needed for this route and read only those.4849For obvious narrow requests, do not expand a lifecycle plan. Say the route,50load the narrow skill, and execute it.5152## Non-Use Gate5354Use the narrow maintainer directly when all of these are true:5556- the prompt names a concrete surface, run, PR, release, screenshot, data path,57 or local runtime problem58- the next useful artifact is obvious for that surface59- the task does not need cross-stage sequencing, reviewer orchestration, or60 destructive recovery judgment before the narrow work can begin6162Keep ownership in this router only when it adds value by choosing a stage,63resolving ambiguity, sequencing multiple stages, or protecting a high-risk64handoff.6566## Core Rule6768Route first, then execute.6970State the lifecycle stage and acceptance bar before implementation. Any routed71development task enters `spawn-required` mode by default. This applies when the72task changes user-visible, agent-visible, Desktop, release, runtime, CLI,73workflow, or product behavior, even if the user only says "fix",74"optimize", or "推进". The normal implementation sandwich is:7576```text77writer implementation78-> writer basic checks79-> optional lightweight pre-review80-> spawn availability probe if needed81-> spawned verifier black-box acceptance82-> spawned final reviewer gate83-> handoff / commit / push84```8586Pre-review is only for catching obvious diff, startup, safety, scope, or test87readiness problems before verifier time is spent. It is not the final reviewer88gate. Final review follows verifier `PASS` so reviewers can inspect the diff,89tests, handoff, and verifier evidence together.9091Separate verification from review:9293- verification asks whether the product path meets the requirement from the94 user's side95- review asks whether the diff, architecture, scope, tests, proof, and handoff96 are trustworthy9798For development routes, do not treat the loop as optional just because the user99said "fix", "优化", "推进", or omitted the word review. User-visible,100agent-visible, Desktop, release, runtime, CLI, workflow, and Rudder101changes need the verifier plus final reviewer loop before complete handoff. If a102route truly does not need that loop, say why it is `not applicable` rather than103silently skipping it.104105`spawn-required` means:106107- after writer checks, probe the available spawn/subagent mechanism before108 declaring verifier or reviewers unavailable109- if spawning works, create the verifier and final reviewer agents and wait for110 their verdicts before complete handoff111- if spawning is unavailable, record the failed probe as112 `blocked: spawned verifier/reviewer unavailable`113- do not claim complete handoff, commit/push readiness, or review pass from114 author-run tests, CI, screenshots, self-review, or serial personas alone115116## Stage Classifier117118Choose one primary stage. If multiple stages are present, choose the earliest119blocking stage.120121- `intake`: user intent, target artifact, or mode is unclear122- `requirements`: problem framing, scenarios, acceptance criteria, or "do you123 understand?"124- `advisor`: first-principles diagnosis of a build, UI, workflow, trace, or125 proposal that feels wrong126- `ui_design`: interface direction, wireframe, visual hierarchy, or127 screenshot-based design judgment before code128- `implementation`: approved direction or direct fix/build request129- `verification`: tests, CI, E2E, screenshots, Desktop smoke, actor-run-chain,130 release checks, or black-box acceptance proof131- `review`: review, PM judgment, first-principles critique, or132 Codex/session/PR/commit verdict133- `debug`: failed run, UI path, data path, CI job, Desktop app, or local process134- `release`: canary/stable release, npm, Desktop assets, tags, or GitHub135 Release state136- `handoff`: final summary, validation, commit, push, residual risk, or PR137- `recovery`: dirty worktree, stash, interrupted run, rollback, delete/restore,138 or suspected old Codex work139- `component_lab`: UI Lab, component inventory, fixtures, or design-system140 coverage141- `performance_benchmark`: benchmark or optimize before the exact bottleneck is142 known143- `runtime_contract`: provider/runtime/tool-call/transcript/parser/CLI or144 agent-visible contract parity145- `skill_optimization`: optimize, harden, refactor, validate, benchmark,146 package, or improve a named skill or workflow skill147148For full routing detail, read `references/route-selection.md`.149150## Narrow Routes151152Prefer the smallest matching owner:153154- screenshot-driven UI polish: `rudder-ui-polish-maintainer`155- missing, stale, suspicious, or slow page data:156 `rudder-data-path-diagnostician-maintainer`157- run transcript, stdout/stderr, recent run batch, or run-quality issue:158 `debug-run-transcript-maintainer`159- Desktop startup, packaged app, Electron shell, update, profile, local160 instance recovery: `rudder-desktop-dev-recovery-maintainer`161- release, npm, GitHub Release, Desktop assets, tags, dist-tags, install smoke:162 `release-maintainer`163- review-only session/PR/commit/proposal/UI/release outcome:164 `agent-work-reviewer-maintainer` or165 `codex-session-product-reviewer-maintainer`166- local branch preview: `rudder-worktree-preview-maintainer`167- GitHub PR checkout/preview/review: `pr-local-preview-maintainer`168- mock/demo/seed data: `mock-data-maintainer`169- landing screenshots: `landing-proof-shots-maintainer`170- stop/restart/clean local dev runtime: `stop-rudder-dev-maintainer`171- new or existing reusable workflow skill creation, optimization, eval,172 benchmark, packaging, or description tuning: `skill-creator`173174If this router itself is the target artifact for optimization, route to175`skill-creator` or the user-requested skill-engineering workflow. Do not run176this router's normal product lifecycle except for git safety around the patch.177178## Skill Optimization Boundary179180The newest user instruction is the routing source of truth.181182If the user says a skill "needs optimization", "should be hardened", "always183does the wrong thing", "I have to ask this every time", or asks to use184`skill-creator` on a named skill, classify the turn as185`skill_optimization`.186187Treat screenshots, prior requirements, session ids, quoted logs, and prior188assistant recommendations as evidence for the skill failure. They are not the189active product task unless the newest user instruction says to continue that190product task.191192For this route:193194- name the target skill and path195- extract the failed decision point196- choose the smallest durable owner: target skill body, frontmatter197 description, eval case, memory update, or no-op198- add or update a validation case when the behavior should change next time199200## Verification And Review Defaults201202For user-visible, agent-visible, Desktop, release, runtime, CLI, workflow, or203product changes, identify the terminal product surface before calling204verification complete. Use `product-acceptance-verifier-maintainer` for a205distinct black-box acceptance pass. If the product path is too expensive or206unsafe to exercise, record verifier `blocked` or `substituted` with the exact207reason; do not silently skip acceptance.208209Spawned verifier and reviewer gates are the default for routed development210artifacts when this router owns the lifecycle. If spawning is available, the211verifier should be a distinct child agent using212`agents/product-verifier.md` or equivalent `product-acceptance-verifier-maintainer`213instructions. Final reviewers should be child agents using the templates under214`agents/` or equivalent reviewer prompts, with these lenses:215216- functional trust217- adversarial218- heuristic/product-systems219220Do not replace spawned verifier or reviewer verdicts with self-review, serial221personas, or author-claimed screenshots. If the active runtime cannot spawn222agents after a real availability probe, record `blocked: spawned verifier/reviewer223unavailable` with the probe evidence instead of claiming the loop passed.224225For the full verifier/reviewer contract, read226`references/verification-review.md`.227228## Execution Outline2292301. Build a compact routing packet:231 - latest user request and corrections232 - `git status --short --branch`233 - named files, screenshots, sessions, runs, PRs, commits, or plans234 - relevant `AGENTS.md` route docs and nearby skill contracts235 - changed-file ownership if cleanup or recovery is involved2362. Declare route and stage exits.2373. Execute the current stage using the narrow owner or normal repo workflow.2384. Run the required verification and review gates for that route.2395. Hand off with evidence, blockers, git state, and commit/push status.240241## Minimal Handoff Shape242243```markdown244Route: ...245Stage exits: ...246Used: ...247Review: spawned reviewers / blocked / not applicable248Validation: passed / not run / not proven249Evidence: required / scenario / proven / missing or substituted250Git: commit / push251Residual risk: ...252```