lidge-jun/codexclaw / plugins/codexclaw/skills/dev
cxc-dev
VerifiedUse for coding, PR delivery, scaffolding and QA. Triggers: develop, fix, refactor, test, review, docs, browse, stacked PR, 개발, 수정, 검토, 스택 PR.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 12.0 KB |
| agents/openai.yaml | config | 143 B |
| references/async-questions.md | markdown | 6.2 KB |
| references/browser-routing.md | markdown | 3.4 KB |
| references/code-mode-examples.md | markdown | 4.3 KB |
| references/development-practice.md | markdown | 6.3 KB |
| references/hosted-ci-evidence.md | markdown | 3.2 KB |
| references/logging.md | markdown | 2.8 KB |
| references/methodology-overlays.md | markdown | 2.8 KB |
| references/native-execution.md | markdown | 6.3 KB |
| references/peer-collaboration.md | markdown | 10.6 KB |
| references/product/crud-product-development.md | markdown | 2.7 KB |
| references/reader-documents.md | markdown | 6.7 KB |
| references/skill-catalog.md | markdown | 14.1 KB |
| references/skill-ownership.md | markdown | 5.0 KB |
| references/stacked-prs.md | markdown | 16.0 KB |
| references/static-analysis-gate.md | markdown | 2.6 KB |
| references/static-analysis.md | markdown | 3.3 KB |
references/async-questions.md
View on GitHub# Async user questions
Use this when useful missing input arises during ongoing work. The current host's
tool schema and instructions decide availability and permitted use; model names
(Astra, Grok, Sol, etc.) do not substitute for checking the callable tool list.
Ask only when the answer changes the work and is not already available in context.
Main agents may leave useful async questions throughout ordinary work and active
goals, whenever the tool is exposed and the host permits it. No Interview entry
or phase change is needed. SessionStart announces this policy. PostCompact only
queues a workspace/session-scoped hint; the next root UserPromptSubmit consumes
it once on an event that accepts model context. No prompt means no same-turn/Stop
recovery guarantee. These hooks neither enable a missing tool nor force a question.
## Select the exposed tool
- When `request_user_input_async` is exposed and allowed, prefer it for a question
during work. Leave it for the user to see without expecting a reply; continue
authorized work. Use its actual namespace.
- If it is absent or rejected, inspect the available schema/policy and use an allowed
plain-text note/question where permitted, without claiming a question card was sent.
Do not switch optional mid-work questions to blocking `request_user_input`. Reserve
that tool for its existing authorized workflow (such as Interview). Never invent an async tool,
change feature flags, or retry a forbidden request through another channel.
- Blocking `request_user_input` has different fields and limits. Never copy its
`header`, `id`, `question` or option-object payload into the async tool.
The observed async schema accepts `questions[]` with `title` and optional string
`options[]`. Check the live schema before copying this example:
```json
{"questions":[{"title":"Which export format should this report use?","options":["CSV (Recommended)","JSON"]}]}
```
Ask a concise, self-contained question naming the decision it changes. Put the
recommended option first when choices help; omit options for free text. Do not add
an Other choice or free-text placeholder when the UI already supplies free text.
Group only independent questions; a later question whose meaning depends on an
earlier answer must wait. Do not inherit the blocking tool's three-question limit.
## Send, continue, incorporate
1. Send once and treat immediate return as submitted/pending, not answered. A
preselected option has not been submitted. Do not end the turn just because
the tool returned, repeatedly resend the question, or invent an answer-poll tool.
2. Continue concrete work that does not depend on the answer: inspect inputs,
calculate shared results, review existing code, or prepare common sections.
Keep the affected decision and outstanding question in the existing work record
when it must survive a turn or compaction; create no separate tracking system.
3. Do not schedule a wait, poll for a reply, or make progress depend on an optional
answer. Continue independent work and follow any host-required reply opportunity
before committing an optional choice; then use a stated, reasonable assumption
within the authorized scope. A missing reply never blocks completion of optional
work. Required input or approval remains a real dependency: silence and
preselection are never consent; only the dependent action stays pending.
For a bound goalplan, after the host confirms a question was submitted, run `cxc loop ask --session <id> --id <decision-id> --question <exact text> [--recommendation <text>] [--option <text>]... [--work-phase <id>]... [--criterion <criterion-id>]...` (list the offered options, recommended first by convention; the recommendation must be one of them) to keep it in the existing plan. Record each dependent work phase, and with `--criterion` each unmet criterion that cannot be met until the reply arrives; otherwise Stop keeps the goal running. Check `cxc loop show` before sending a similar question after compaction. When the user answers, run `cxc loop decide --session <id> --id <decision-id> --answer <reply>` and resume only newly runnable phases. `ask` does not deliver a question; CLI success is no proof of host submission. Link a phase only while its action truly requires the reply.
4. A later user message supplies the answer. Match it to the pending decision,
update assumptions and affected work, and retain the original objective unless
the user changes it. An ambiguous reply leaves the unresolved part pending.
When no independent work remains and required input is missing, yield with the
dependency clearly stated; report partial status rather than claiming completion.
5. Keep leaving useful new questions as new uncertainties arise during the work,
even when earlier optional questions are unanswered. Do not repeat the same
unresolved question, send reminders, or create questions just to demonstrate the tool.
For the export example, compute common report rows while format is pending. If
the user chooses JSON, serialize those rows as JSON. If an optional default was
already used, revise the affected artifact when feasible. A required publication
approval leaves publication pending regardless of time spent preparing the report.
Interview retains its existing blocking question flow; do not migrate it to async.
## Goal and Interview boundaries
Follow the actual host's question/approval rules and the current goal's authority.
An active goal alone does not prohibit optional async questions or make their
answers a completion dependency. Do not bypass an actual host prohibition or
turn async questions into a blocking Interview. Subagents return question candidates to the main agent;
they do not question the user independently.
The persisted CodexClaw Interview currently captures synchronous `request_user_input`
question IDs and returned answers. Async submission and later messages are not
automatically captured by that path. See the [Interview owner](../../interview/SKILL.md)
before using questions as readiness evidence; never mark a pending async question as
answered, manufacture ledger events, or assume the exact-name synchronous goal guard
covers async variants. This reference is agent guidance, not new runtime enforcement.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| finishing-a-development-branchobra/superpowersUse when implementation is complete, all tests pass, and you need to decide how to integrate the work | Stars: 321k | Updated | Claude CodeCodex | Verified | Install |
| systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixes | Stars: 321k | Updated | Claude CodeCodex | Verified | Install |
| test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation code | Stars: 321k | Updated | Claude CodeCodex | Verified | Install |
| skill-creatoranthropics/skillsCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy. | Stars: 197k | Updated | Claude CodeCodex | Verified | Install |