使用方法
- 复制安装命令,在自己的智能体环境安装前先检查源码包。
- 从原始文档核对所需工具、凭据和支持的宿主。
- 说明输入材料与期望交付物,并对照来源检查输出。
npx skills add https://github.com/ntanh1402/sdlc-skills --skill sdlc-design-feature-tests开始前需了解
- 安装不包含模型订阅、API 额度或访问文件的权限。
- 仓库星数和安装量反映活跃情况,不等于独立质量或安全测试。
按源仓库内容呈现,保留标题、案例、代码、表格、链接以及原文引用的演示图片。
Design the Feature-level tests
This skill is a lifecycle step after the design. It takes a Feature or ChangeRequest whose design is merged, whether its Requirements are a draft or agreed, and writes the GENERAL/ group of its one TestSuite in wiki/<app>/tests/TS-<name>/: end-to-end flows that span user stories, and every other case tied to no single story. Each case is a TC-<id>-<name>.md file. The cases of one user story are written by sdlc-design-story-tests in that story's STORY-<name>/ group. It never writes runnable test code or touches a code repository.
Its draft prefix is tests-: the draft for FEAT-coupons is wiki/tests-FEAT-coupons. It may run at the same time as sdlc-plan-tasks and sdlc-design-story-tests on the same design.
Before you start
Read, in this skill's folder, references/read-protocol.md; each section names the other references/ files to read when it starts.
Find the wiki as the read protocol says, and check that it matches this skill. RELEASE, next to this file, holds one line such as 1.0.0:
cd "$WIKI" && $T copy check --release "<the line in RELEASE>"status | Do |
|---|---|
current | Go on |
older | Stop: "The wiki's tool is older than this skill. Run sdlc-setup-wiki to upgrade the wiki." |
newer | Stop: "The wiki was set up by a newer release. Update your installed sdlc-skills." |
edited | Stop: "The wiki's tool copy was edited by hand. Run sdlc-setup-wiki to restore it." |
Then read the schema pages schema.md, test-suite.md, test-group.md, test-case.md, user-story.md, feature.md, change-request.md, architecture-decision.md and the page of each Design type in scope, in the wiki's .wiki-llm/schema/.
| Request | Skill |
|---|---|
| No approved design yet | sdlc-design-arch |
| The cases of one user story | sdlc-design-story-tests |
| Build tasks | sdlc-plan-tasks |
| User stories | sdlc-write-stories |
| Runnable tests, or running them | Not this skill. sdlc-build-task writes a Task's unit tests with its code |
| Load or performance tests | Not kept in the wiki; no skill designs them |
| Which tests exist, or what is uncovered | sdlc-ask-wiki |
The gates
A gate is a stop for the person's explicit approval. There are three, so the person decides what is being done (Intent), how (Design), and exactly what is saved (Change-set), each before the next cost is paid.
- Approval is an explicit yes to the gate's question. "Sounds good", silence
or "whatever you think" is not approval.
- An edit at a gate reopens that gate and every later one.
- When you cannot ask the person (a run with no one to answer), or the person
asks you to skip questions or approvals ("don't ask me, just write it"), stop before the first gate and start no draft. Return the open decisions and the gates the work still needs. Never assume approval.
- When the person rejects the design, stop. A draft worktree this run started
is discarded; a worktree an earlier run started is left as it is.
- Being asked to run this skill does not approve anything it writes.
- A refusal this file states is a refusal. Do not turn it into a question.
1. Look up: readiness and the inventory
Accept one FEAT-* or CR-* key. Read it on the default branch:
| State | When | Do |
|---|---|---|
READY | A Feature Draft, ReqApproved, Approved or InDev, or a ChangeRequest Proposed, ReqApproved or Approved, whose # Architecture or # Delta is not None | Go on |
NEEDS_DESIGN | # Architecture or # Delta is None | Stop: route to sdlc-design-arch |
NOT_MERGED | Not on the default branch, but a wiki/*-<Key> draft exists | Stop and name the draft |
CLOSED | A Feature that is Released or Deprecated, or a ChangeRequest that is Implemented or Rejected | Stop: say it is closed; a change to a Released Feature goes through a ChangeRequest (sdlc-write-spec); route nowhere else |
Every state but READY is a refusal, not a question.
Then read the Requirements, the Architecture or Delta, the decisions, every Endpoint, Channel, Subscription, Table and External contract in scope, the quality targets and risks, the PRD's user flows, the user stories in the target's folder, the References they link and those about testing in references/index.md (a test plan the user gave is one), and the existing TestSuite that verifies the target (grep -rl "<Key>" wiki/<app>/tests/*/overview.md) with all its groups. Run $T outdated --feature <Key> (or --change <Key>): a group it lists was designed from an older version of the spec. When the design itself is listed (designBasedOn older than version), say in the first round that the design is behind the spec and recommend sdlc-design-arch first; go on only when the person says so. Run $T validate and keep the warnings about this suite (tests.story-folder, links.unresolved in its files): a STORY-* group whose story is gone is this skill's to fix. List:
- every PRD user flow that spans two or more stories;
- every nonfunctional Requirement and Constraint no story lists, and every
Functional Requirement no story lists yet (it is not covered here: name it for sdlc-write-stories, then sdlc-design-story-tests, under "For other skills");
- checks that apply across endpoints or stories: authentication,
authorization, privacy, audit and abuse boundaries;
- retry, timeout, idempotency, ordering, replay, dead letters and fallback
that span services;
- data integrity, migration, compatibility, regression and clean-up;
- every new, changed or removed contract no story's cases will cover;
- every frontend route or screen with its access rule and its empty, loading,
error and offline states, when a frontend is in scope;
- the performance and security risks the Requirements and decisions state;
- the cases of the existing
GENERAL/group, and the cases of any orphaned
STORY-* group.
An unfinished run. Last, ls "$WIKI"/.worktrees/*/plan.md 2>/dev/null lists the runs that stopped part-way. When one is for this request (its folder is tests-<Key>, or its plan's title names this request), read references/write-protocol.md now and go on as its section 2 says, without asking the rounds again.
2. Questions and the Intent gate
Read references/test-design-rules.md and references/security-checklist.md now: the choices of section 3 use them.
Ask in rounds. A round is a numbered list of every question that can be answered now, each with your recommended answer. The person confirms or corrects, which is faster than composing an answer, and sees your reasoning before any work is done.
Q1 - Which flows should the e2e cases of TS-coupons walk?
(features/FEAT-coupons/overview.md#user-flows)
➡️ Apply a coupon and pay; apply, remove and pay. Both span STORY-coupons-apply and STORY-coupons-remove.
Q2 - Should security-validation cases be written?
➡️ No: no auth or PII rule changes.
Q3 - What else in the wiki must change because of this?
➡️ Nothing: only the suite's GENERAL group and its cases.Which questions to ask, and how many, is your judgement. Three things are fixed:
- The format. Numbered questions, each with a recommendation. A
recommendation that rests on the wiki cites the path. A question whose answer depends on another question in the same round waits for the next round.
- At least one round, even when a brief seems complete. Its questions are
then the assumptions you would otherwise make without saying so.
- The wiki question. Ask what else in the wiki must change because of
this request, with your own answer as the recommendation.
Wait for the answers. Ask another round when the answers open new questions. Search the wiki again whenever an answer brings a system, a term or a claim about current behaviour you have not read. When nothing that changes the output is left undecided, go to the Intent gate: one approval, after the rounds.
A change another skill owns. When an answer needs a wiki change this skill does not own (such as a Requirement or contract that cannot be tested as written), do not make it. List it under "For other skills" in the Run plan and in the summary. When your work depends on it, stop before the Intent gate and name the skill to run first.
The first round names the target and its state, shows the inventory, and asks the person to choose the flows the e2e cases walk and the categories to write (test-case.md lists them). Each case runs as one of two types:
integration: one service and what it calls: datastores, externals,
messages;
e2e: an actor outcome across the deployed flow, through the owning
frontend when there is one.
A developer's unit tests are written with the code (sdlc-build-task). Load tests are not kept in the wiki: a latency, throughput or capacity Requirement is listed as not covered here.
The same round asks about the seams the cases test at, the exclusions, and the wiki question; the usual answer to that one is "nothing: only the suite's GENERAL/ group and its cases".
- A category the evidence requires (
references/test-design-rules.md,
"Required categories") may be declined only when the user explicitly accepts the residual risk; record that acceptance in the group's log.md.
- A step names the real table, topic, endpoint, cache key, bucket or index,
from the design file's tableName, channelName, method and path, keyPattern, bucket or indexName. When a case needs one the design does not give, stop before the Intent gate, list the missing names, and route to sdlc-design-arch. Never write the wiki key or a made-up name instead.
- A contract or Requirement that cannot be tested as written is another
skill's change: sdlc-write-spec.
- A question about the Requirements, the acceptance criteria or the design
that the person cannot answer now becomes an open question of the group; the group then stays Draft.
- When the Requirements are still a draft (
Draft,Proposed), say so: the
cases may have to change when the spec does.
Intent gate. Restate the target, the inventory, the flows, the categories, the seams, the exclusions, the open questions and the residual risk the person accepted. Ask for approval.
3. What goes into the design
Read references/case-format.md now. For each chosen category, partition inputs and branches with references/test-design-rules.md. "All cases" means complete coverage of the requirement, contract and risk partitions not left to a story, not every combination of inputs.
One suite per target: TS-<name>, the target's key without its prefix (FEAT-coupons → TS-coupons, CR-coupon-stacking → TS-coupon-stacking). If it exists, add to it. If the name is already taken by the suite of another Feature or ChangeRequest, stop and ask.
This skill writes only the GENERAL/ group: an e2e case for each PRD user flow that spans two or more stories, a check that applies across endpoints or stories, a case for a nonfunctional Requirement or Constraint no story lists. A Functional Requirement no story lists gets no case here: its cases belong to the story sdlc-write-stories will write for it. A case that proves one story's acceptance criteria belongs in that story's group, which sdlc-design-story-tests writes; name the story and route there. A case's id is the highest id in the suite + 1, across all its groups; ids of deleted cases are never reused.
An orphaned story group. A STORY-* group whose story was split, merged or deleted is a warning. Move each of its cases to GENERAL/, keeping its id and dropping the link to the gone story from # Covers, log each move in GENERAL/log.md, then delete the empty group folder and list its deleted files in the Run plan's ## Files; or, when a case now proves another story, name that story for sdlc-design-story-tests. Show each move at the Design gate.
A group designed from an older version. Compare the spec at the version in the GENERAL/ group's basedOn with the current one (git log -p on the overview.md). Revise the cases the change affects, add the cases it needs, and delete the cases of removed Requirements as the last rule of test-case.md says; leave the other cases as they are. Show each revised case in full at the Design gate.
Each case is a file TC-<id>-<behaviour words>.md in GENERAL/, written with references/case-format.md.
Open questions. An answer the person gave to a question under the group's # Open questions is applied to the cases it affects; the question is then removed and the answer logged in the group's log.md.
4. Run plan and the Design gate
Read references/write-protocol.md now.
Start or resume the draft tests-<Key> as write-protocol.md section 2 and section 3, "At section 4", say. The input paths are the Feature's or ChangeRequest's overview.md and, for a ChangeRequest, the target Feature's overview.md.
Prepare the test design. Build the matrix from every inventory row to its cases; rows without their own wiki concept (a failure branch, a risk) are added by hand. The design is:
- the group with its scope, seams, exclusions and open questions;
- each case: key, test type, category, risk, purpose, what it covers;
- the coverage matrix, with every uncovered row and why;
- the totals by Requirement, Design concept, risk, type and category;
- assumptions and residual risks.
Write the Run plan (write-protocol.md section 4). Its ## Files are the suite's overview.md when it is new, the group's overview.md and log.md, and one line per case file with what it covers. When the user gave a test plan that the wiki does not hold yet, add its Reference folder references/REF-<name>-test-plan/ (overview.md, log.md and the plan as a content file); the Design gate is the person's approval of it.
Design gate. Show the design in full, never only a summary, then the Run plan. Ask for approval. Approval with an uncovered in-scope row is allowed only when the user explicitly accepts that gap as residual risk.
5. Write
In the draft worktree, under wiki/<app>/tests/:
- a new suite folder
TS-<name>/withoverview.md(type: TestSuite,
# Verifies linking the Feature or ChangeRequest) when there is none;
- the group
TS-<name>/GENERAL/withoverview.md(type: TestGroup,
basedOn: { version: <the target's version>, at: <its versionDate> }, # Scope saying which flows and checks it covers) and log.md;
status: Draftwhile# Open questionshas items, otherwiseApproved;# Open questions, a numbered list, each question linking the
Requirement, story, Design concept or decision it is about;
- one
TC-<id>-<name>.mdper new case inGENERAL/, withtestTypeand
testCategory;
# Referenceson the group that links the test plan it was written from,
and any other Reference a tester needs, with a note on what to take from it;
- an entry in the group's
log.md, for example
* **Test design**: Added 4 cases for FEAT-coupons. Written by an AI model with sdlc-design-feature-tests.
Every step and validation must make sense to a tester who did not design the system and does not read the wiki. Do not write index.md or checklist.csv: draft finish writes both. Do not change a STORY-* group except to empty an orphaned one.
6. Self review, finish and commit
Follow write-protocol.md section 3, "At section 6", with the self review of its section 5. In the self review, have the review sub-agent check by reading that:
- every case has its test type, category, risk, concrete preconditions,
complete steps, clean-up and links;
- each Action starts with what to do and its target, copies the values it
needs, and points at no other step or data; preconditions, steps and clean-up use real names, never a wiki key;
- each Expected result is one checkpoint per table, topic, cache entry, page
or group of API fields, - <field>: <value> lines joined with <br>, and each Validation names the fields it checks in bold;
- ids are unique in the suite; no case proves only one story's criteria;
- no case is written twice, no existing case was changed except those a
newer spec version affects, and no runnable code was written;
- each step that calls an API or publishes to a channel has its request and
expected response;
- the group's
basedOnnames the current version of its target, and it is
Draft exactly when it has open questions;
GENERAL/has ane2ecase for each PRD user flow that spans two or more
stories;
- each accepted risk or gap is recorded in the group's
log.md.
Then run $T coverage tests --feature <FEAT> (or --change <CR>) in the worktree, and $T validate: no warning about this suite may be left except in a STORY-* group that sdlc-design-story-tests owns.
Commit message: tests(<Key>): design <n> Feature-level cases. In the summary, "Not done" names each gap the user accepted, each open question, and each Requirement or story that no case of any group covers yet, and "Next" names sdlc-design-story-tests for each story whose group is missing or behind, and says the suite's checklist.csv lists its cases.
The summary
End the run with at most four lines:
Done: wiki/tests-FEAT-coupons, 5 files, pushed. Open the pull request.
Not done: no case covers STORY-coupons-remove yet
For other skills: REQ-coupons-expiry lacks an exact timezone rule (sdlc-write-spec)
Next: sdlc-design-story-tests STORY-coupons-remove; TS-coupons/checklist.csv lists the cases.Leave out "Not done" and "For other skills" when they have nothing to say. Do not list what the self review fixed, and do not repeat the test design: the person approved the content at the gates, so the last message only says where it is and what comes next.
Done when
- The user chose the flows and categories, and any declined required
category has an accepted residual risk.
- Every in-scope flow, cross-story check, contract, branch and risk is
covered, or the gap was accepted.
- Every case has concrete preconditions, self-contained steps with real
names, checkpoint expected results, bold validations, clean-up and links, and a unique id; no case is duplicated; no runnable code was written.
- The self review found nothing it could not fix, or the person accepted what
was left at the Change-set gate.
draft finishwasok, the change set was approved, anddraft commit
reported the branch.