[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M759C3AKPFXNW312D49JWQ3F":3,"skill-same:01M759C3AKPFXNW312D49JWQ3F":59},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":26,"commitSha":27,"skillDirectory":28,"forksCount":14,"repositoryHtmlUrl":29,"repositoryHomepageUrl":6,"ownerHtmlUrl":30,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":31,"install":32,"resources":44},"01M759C3AKPFXNW312D49JWQ3F","4-connect",null,"DDD loop 2 — connect: Domain Message Flow Modelling per use case — typed messages, coupling smells, boundary deltas handed back to decompose. Writes docs\u002Fdomain\u002Fmessage-flows\u002F.","baodq97\u002Fgovkit","baodq97","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F50482002?v=4","01M759BJ9XC68CW9PQCFK22H1X","Bao Do","01M759BJAY1W696G8A8N841875",0,"MIT","https:\u002F\u002Fgithub.com\u002Fbaodq97\u002Fgovkit\u002Fblob\u002Fmain\u002Fdocs\u002Fresearch\u002Fddd-skill-evals\u002Fskill-snapshot-pre-context-pack\u002F4-connect\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: 4-connect\ndescription: >\n  DDD loop 2 — connect: Domain Message Flow Modelling per use case — typed messages, coupling smells, boundary deltas handed back to decompose. Writes docs\u002Fdomain\u002Fmessage-flows\u002F.\ndisable-model-invocation: true\n---\n\n# Domain Connect\n\n> *\"It is necessary to challenge the initial design by applying concrete use-cases to uncover\n> hidden complexity.\"* — ddd-crew, Connect\n\nA context map tells you which contexts exist and that they are related. It cannot tell you whether\nthe split *works*, because it never moves. A **message flow** moves: one real business scenario,\nmessage by message, in order, across the boundaries you just drew. That motion is what exposes the\ncoupling — a boundary looks clean until you watch four synchronous hops cross it to satisfy one\ncustomer request.\n\n**What this skill produces:** a small set of flows for the use cases that matter, plus the findings\nthose flows forced into the open — and, where a flow proves a boundary wrong, a **proposed** change\nhanded back to `3-decompose`. Proposed, not applied: a boundary redrawn by the skill that\ndiscovered the problem skips the reconciliation, seam, and delta-merge discipline that decomposition\nowns.\n\n## Inputs\n\n- `docs\u002Fdomain\u002F` — contexts, relationships, and the domain events each context emits. This is the\n  design under test.\n- `docs\u002Fdomain\u002Fdiscovery\u002F` — the event timeline, if discovery ran. Flows built from a discovered\n  timeline are grounded; flows built from context names alone are speculation.\n- The **use cases** to trace. If nobody names them, propose candidates and get agreement before\n  drawing — see step 1.\n\nNothing under `docs\u002Fdomain\u002F`? Say so plainly: there is no design to challenge yet, and\n`3-decompose` runs first. Modelling flows between contexts you are inventing as you go\nproduces a diagram that validates itself.\n\n## Reference files (read as needed)\n\n- `references\u002Fmessage-flow-notation.md` — the notation: message types and what each one costs in\n  coupling, the three parts of a message (name, contents, order), separate vs combined format, the\n  5-to-9 message rule and why exceeding it is a finding, temporal semantics (`within` \u002F `after` \u002F\n  `every`). Read before drawing the first flow.\n- `references\u002Fcoupling-heuristics.md` — the smell catalogue: what each smell looks like on a flow,\n  what it usually means, and the boundary move that resolves it. Read at step 3.\n\n## Who to involve\n\n- people who design, build and test software\n- people who have domain knowledge\n\nFlows are where developers and domain experts disagree productively: the developer knows what the\nsystem does, the expert knows what the business does, and the gap between those two is the finding.\n\n## Process\n\n### 1. Choose the use cases — the three backbone scenarios\n\nTrace **three**, not the whole backlog, and pick them by role rather than by taste:\n\n| # | Scenario | Why this one |\n|---|---|---|\n| 1 | **the happy path** | the design's own story — if this one is ugly, nothing else will be better |\n| 2 | **the path with money on it** | the scenario the business is actually paid for; coupling here has a price |\n| 3 | **the failure path** | rejection, cancellation, refusal. Models are built happy-path-first, so this is where the missing messages live |\n\nAdd a fourth only for a known hotspot — something discovery flagged, or that the team already argues\nabout.\n\nThe failure path is the one teams skip and the one that pays. A model with no rejection message\nanywhere is not a clean design; it is a design where nobody asked what happens when the answer is no.\n\nA flow for a single-context CRUD scenario teaches nothing — it crosses no boundary. Say which\nscenarios you chose and why, and let people add the one you missed.\n\n**A prerequisite worth stating:** message flows are *not* a discovery technique. They exist to break\na cut that already exists — *\"when you have an initial cut of your architecture… you can begin\ndesign the message flows\"*. If there is no `docs\u002Fdomain\u002F` yet, this step has nothing to refute.\n\n### 2. Build the flow\n\nFor each use case, in order: identify the **sender** (actor, bounded context, or external system),\nthe **message** it sends, and the **recipient**. Each message carries three things — its **name**,\nits **significant contents**, and its **order number**. Type every message:\n\n| Type | What it is | Coupling it creates |\n|---|---|---|\n| **Event** | a fact that already happened, broadcast | lowest — the sender does not know who listens |\n| **Command** | a request that another context do something | medium — the sender knows the receiver and expects it to act |\n| **Query** | the sender needs the receiver's data *now*, with a response | highest — a runtime dependency; the sender cannot proceed without the receiver being up |\n\nGetting the type wrong is not a labelling slip. It is the whole point: a flow drawn entirely with\nundifferentiated arrows hides exactly the coupling this step exists to find.\n\nKeep each diagram to **5–9 messages**. Beyond that, nobody in the room can hold the flow in their\nhead — and in practice a scenario that needs fifteen messages is telling you either that it is two\nscenarios or that the boundaries are wrong. Treat the overflow as a finding, not a layout problem.\n\nWhen a scenario is time-dependent, be exact about which temporal relation holds: *within* an\ninterval, *after* an event, or *every* period. These are three different business rules and they\ngenerate three different designs.\n\n### 3. Read the flow back for coupling\n\nThis is the step that pays for the diagram. Walk the finished flow and check for the smells in\n`references\u002Fcoupling-heuristics.md`. The ones worth naming here:\n\n- **Synchronous query chain across a boundary** — context A cannot act without asking B, then C.\n  Availability multiplies and latency accumulates. Usually the data is on the wrong side, or B\n  should publish a read model A owns a copy of.\n- **Cycle inside one use case** — A → B → A. Two contexts that must talk back and forth to complete\n  one scenario are temporally coupled; often they are one context wearing two names.\n- **Distributed invariant** — one business rule enforced across two contexts within the flow. Either\n  the rule belongs to a single aggregate, or the business must accept eventual consistency *and*\n  name the compensating action. Leaving it unstated is how double-bookings ship.\n- **God context** — one context appears in every flow, usually as a coordinator forwarding messages.\n  Orchestration is legitimate; a context that decides nothing of its own is a hop, not a boundary.\n- **Pass-through** — a context receives a message and forwards it without making a decision.\n- **Chatty pair** — two contexts exchanging many messages in one flow. Count them; a pair that talks\n  more inside a scenario than either talks to anyone else is a candidate merge.\n\nEvery finding gets: the flow it came from, the messages that show it, and what it would take to\nresolve it. A smell with no evidence attached is an opinion.\n\n### 4. Feed findings back — do not redraw\n\n**Two conditions refute the decomposition outright**, and they are the only place in the whole\nmodelling loop where a design is rejected by evidence rather than by opinion:\n\n> **more than 9 messages in one scenario**, or **one context appearing at every step** ⇒ go back and\n> re-cut.\n\nSay it in those words when either fires. It is not a diagram-tidiness note; it is the loop-back\ntrigger, and `3-decompose` becomes stale the moment it fires.\n\nWhere a flow proves a boundary wrong, write it as a **proposed change** with its evidence, and hand\nit to `3-decompose` (update mode) to merge. Where a flow reveals a domain event nobody discovered,\nhand it to `2-discover` to confirm with people — do not promote your own inference to a confirmed\nevent.\n\nThis is the loop working as intended. The modelling process is iterative; connect is where decompose\ngets its corrections.\n\n### 5. Emit\n\nWrite one file per flow under `docs\u002Fdomain\u002Fmessage-flows\u002F`, plus a `README.md` index that lists the\nflows and collects every finding in one place. `status: draft`, `owner: TBD`.\n\n## Output shape\n\n````markdown\n---\nid: DOMAIN-FLOW-0001\ntitle: \u003CUse case> — domain message flow\nstatus: draft\nowner: TBD\ndate: \u003Cdate>\ncontexts: [Allocation, Logistics, Billing]\n---\n\n## Scenario\n\u003C!-- one paragraph, in business language: who wants what, and what \"done\" means -->\n\n## Flow\n\n```mermaid\nsequenceDiagram\n  actor Planner as Depot Planner\n  participant Logistics\n  participant Allocation\n  Planner->>Logistics: 1. ScheduleTransfer [cmd] {unitId, fromDepot, toDepot, window}\n  Logistics->>Allocation: 2. IsUnitAvailable? [qry] {unitId, window}\n```\n\n| # | From | Message | Type | Contents | To | When |\n|---|---|---|---|---|---|---|\n| 1 | Depot Planner | `ScheduleTransfer` | command | unitId, fromDepot, toDepot, window | Logistics | — |\n| 2 | Logistics | `IsUnitAvailable?` | query | unitId, window **→** available, freeFrom | Allocation | — |\n| 3 | Logistics | `TransferLapsed` | event | unitId | Allocation | **after** 30 min of no confirmation |\n\n**A query carries its response in the same row, after a `→`.** The notation draws a query and its\nanswer as one unit precisely because the sender is *blocked* in between — splitting them into two\nrows doubles the diagram and hides the only interesting thing about a query. Everything left of the\narrow is what the sender sends; everything right of it is what it waits for. A query with nothing\nafter the `→` has not been thought through: someone has to say what comes back, because the shape of\nthe answer is what the caller is coupled to.\n\n**The `When` column is for time-driven messages only**, and it exists because the semantics decide\nthe design: **within** 5 minutes, **after** 5 minutes and **every** 5 minutes are three different\nsystems. Leave it `—` for a message that follows immediately from the one above. Put the trigger\nhere rather than in the scenario prose — a temporal rule buried in a paragraph is invisible on the\ndiagram, and it is usually the rule that produces the failure path nobody drew.\n\n## Findings\n| # | Smell | Evidence (messages) | What it suggests | Proposed change |\n|---|---|---|---|---|\n\n## Open questions\n\u003C!-- one line each: the question, and who could answer it -->\n````\n\nThe `README.md` index carries the use-case list, why each was chosen, and the consolidated findings\ntable with a status column (`proposed` \u002F `accepted` \u002F `declined`) so a later reader can tell which\nfindings were acted on.\n\n## Hard rules\n\n- **Length budget: ≤ 90 lines per flow file.** A flow longer than that is usually two scenarios.\n  A budget caps prose, not findings: over it, cut rationale a reader can infer and anything\n  restated from an upstream artifact — never open questions, provenance, or a stated absence.\n- **Never invent a message.** A flow is only as real as the events, commands and queries that were\n  discovered or modelled. Inferring `PaymentRefunded` because a refund \"must exist somewhere\"\n  produces a diagram that validates a design against fiction. If the flow has a gap, the gap is an\n  open question with a name attached.\n- **Domain messages, not transport.** No HTTP verbs, no queue names, no retry policy, no\n  serialization format. The moment a flow becomes an integration design, it stops being able to\n  challenge the boundary — every coupling problem starts looking like a technology choice.\n- **Type every message.** An untyped arrow hides the difference between a broadcast fact and a\n  blocking read, which is the difference the whole step exists to surface.\n- **A temporal rule belongs in the `When` column, never only in the scenario paragraph.** If the\n  scenario says \"within fifteen minutes\", some message is the one that must happen within fifteen\n  minutes — name it and put the rule on that row. Stated only in prose, the rule cannot be drawn,\n  cannot be checked, and is invisible to everyone reading the diagram instead of the paragraph.\n  `ddd_check` reports this as `temporal-rule-in-prose`.\n- **A query states what comes back.** `Contents` for a query reads `sent fields → returned fields`.\n  Without the `→` half, the row records a question nobody answered.\n- **Propose boundary changes; never apply them.** `3-decompose` owns the model — it holds the\n  reconciliation rules, the stable ids, the human edits and the extraction seam. A boundary quietly\n  redrawn here is a change nobody reviewed.\n- **Findings carry evidence.** Message numbers, not adjectives. \"Feels coupled\" is not a finding;\n  \"messages 2–5 are synchronous queries crossing three contexts before the planner gets an answer\"\n  is.\n- A flow that turns out clean is a real result. Say so — a use case that crosses four contexts with\n  four events and no queries is evidence the split is working, and it is worth recording as such.\n\n## Worked example\n\n**Input:** the equipment-rental model — contexts `Allocation`, `Logistics`, `Billing`,\n`Notifications` — plus a discovery hotspot about units being double-booked across depots.\n\n**Use case chosen:** *\"Priority depot transfer, booked and billed\"* — the paid add-on, and the\nscenario touching the most contexts.\n\n| # | From | Message | Type | To |\n|---|---|---|---|---|\n| 1 | Depot Planner | `ScheduleTransfer` | command | Logistics |\n| 2 | Logistics | `IsUnitAvailable?` | query | Allocation |\n| 3 | Allocation | *(response)* | query | Logistics |\n| 4 | Logistics | `ReserveUnit` | command | Allocation |\n| 5 | Allocation | `EquipmentAllocated` | event | — |\n| 6 | Billing | `TransferFeeCharged` | event | — |\n\n**Findings the flow forced out:**\n\n| Smell | Evidence | What it suggests |\n|---|---|---|\n| Synchronous query chain | 2–4: Logistics asks, then commands, on the same data | check-then-act across a boundary — between 2 and 4 another planner can reserve the same unit |\n| Distributed invariant | the no-double-booking rule is enforced by Logistics' check but owned by Allocation's data | the rule belongs to **one** aggregate; Allocation should own `ReserveUnit` end-to-end and answer with success or rejection |\n\n**Proposed change handed back to `3-decompose`:** collapse the check-then-act pair into a\nsingle `ReserveUnit` command that Allocation either accepts or rejects, and record the invariant as\nAllocation's. This is the same double-booking hotspot discovery surfaced — the flow is what turned\nit from a worry into a located defect with two message numbers on it.\n\nNote what the example does **not** do: it does not move `Billing` into `Logistics` because the two\nappear adjacent, and it does not silently edit `docs\u002Fdomain\u002Fallocation\u002Fmodel.yaml`. It writes the\nfinding, names the change, and leaves the model to the skill that owns it.\n","# Domain Connect\n\n> *\"It is necessary to challenge the initial design by applying concrete use-cases to uncover\n> hidden complexity.\"* — ddd-crew, Connect\n\nA context map tells you which contexts exist and that they are related. It cannot tell you whether\nthe split *works*, because it never moves. A **message flow** moves: one real business scenario,\nmessage by message, in order, across the boundaries you just drew. That motion is what exposes the\ncoupling — a boundary looks clean until you watch four synchronous hops cross it to satisfy one\ncustomer request.\n\n**What this skill produces:** a small set of flows for the use cases that matter, plus the findings\nthose flows forced into the open — and, where a flow proves a boundary wrong, a **proposed** change\nhanded back to `3-decompose`. Proposed, not applied: a boundary redrawn by the skill that\ndiscovered the problem skips the reconciliation, seam, and delta-merge discipline that decomposition\nowns.\n\n## Inputs\n\n- `docs\u002Fdomain\u002F` — contexts, relationships, and the domain events each context emits. This is the\n  design under test.\n- `docs\u002Fdomain\u002Fdiscovery\u002F` — the event timeline, if discovery ran. Flows built from a discovered\n  timeline are grounded; flows built from context names alone are speculation.\n- The **use cases** to trace. If nobody names them, propose candidates and get agreement before\n  drawing — see step 1.\n\nNothing under `docs\u002Fdomain\u002F`? Say so plainly: there is no design to challenge yet, and\n`3-decompose` runs first. Modelling flows between contexts you are inventing as you go\nproduces a diagram that validates itself.\n\n## Reference files (read as needed)\n\n- `references\u002Fmessage-flow-notation.md` — the notation: message types and what each one costs in\n  coupling, the three parts of a message (name, contents, order), separate vs combined format, the\n  5-to-9 message rule and why exceeding it is a finding, temporal semantics (`within` \u002F `after` \u002F\n  `every`). Read before drawing the first flow.\n- `references\u002Fcoupling-heuristics.md` — the smell catalogue: what each smell looks like on a flow,\n  what it usually means, and the boundary move that resolves it. Read at step 3.\n\n## Who to involve\n\n- people who design, build and test software\n- people who have domain knowledge\n\nFlows are where developers and domain experts disagree productively: the developer knows what the\nsystem does, the expert knows what the business does, and the gap between those two is the finding.\n\n## Process\n\n### 1. Choose the use cases — the three backbone scenarios\n\nTrace **three**, not the whole backlog, and pick them by role rather than by taste:\n\n| # | Scenario | Why this one |\n|---|---|---|\n| 1 | **the happy path** | the design's own story — if this one is ugly, nothing else will be better |\n| 2 | **the path with money on it** | the scenario the business is actually paid for; coupling here has a price |\n| 3 | **the failure path** | rejection, cancellation, refusal. Models are built happy-path-first, so this is where the missing messages live |\n\nAdd a fourth only for a known hotspot — something discovery flagged, or that the team already argues\nabout.\n\nThe failure path is the one teams skip and the one that pays. A model with no rejection message\nanywhere is not a clean design; it is a design where nobody asked what happens when the answer is no.\n\nA flow for a single-context CRUD scenario teaches nothing — it crosses no boundary. Say which\nscenarios you chose and why, and let people add the one you missed.\n\n**A prerequisite worth stating:** message flows are *not* a discovery technique. They exist to break\na cut that already exists — *\"when you have an initial cut of your architecture… you can begin\ndesign the message flows\"*. If there is no `docs\u002Fdomain\u002F` yet, this step has nothing to refute.\n\n### 2. Build the flow\n\nFor each use case, in order: identify the **sender** (actor, bounded context, or external system),\nthe **message** it sends, and the **recipient**. Each message carries three things — its **name**,\nits **significant contents**, and its **order number**. Type every message:\n\n| Type | What it is | Coupling it creates |\n|---|---|---|\n| **Event** | a fact that already happened, broadcast | lowest — the sender does not know who listens |\n| **Command** | a request that another context do something | medium — the sender knows the receiver and expects it to act |\n| **Query** | the sender needs the receiver's data *now*, with a response | highest — a runtime dependency; the sender cannot proceed without the receiver being up |\n\nGetting the type wrong is not a labelling slip. It is the whole point: a flow drawn entirely with\nundifferentiated arrows hides exactly the coupling this step exists to find.\n\nKeep each diagram to **5–9 messages**. Beyond that, nobody in the room can hold the flow in their\nhead — and in practice a scenario that needs fifteen messages is telling you either that it is two\nscenarios or that the boundaries are wrong. Treat the overflow as a finding, not a layout problem.\n\nWhen a scenario is time-dependent, be exact about which temporal relation holds: *within* an\ninterval, *after* an event, or *every* period. These are three different business rules and they\ngenerate three different designs.\n\n### 3. Read the flow back for coupling\n\nThis is the step that pays for the diagram. Walk the finished flow and check for the smells in\n`references\u002Fcoupling-heuristics.md`. The ones worth naming here:\n\n- **Synchronous query chain across a boundary** — context A cannot act without asking B, then C.\n  Availability multiplies and latency accumulates. Usually the data is on the wrong side, or B\n  should publish a read model A owns a copy of.\n- **Cycle inside one use case** — A → B → A. Two contexts that must talk back and forth to complete\n  one scenario are temporally coupled; often they are one context wearing two names.\n- **Distributed invariant** — one business rule enforced across two contexts within the flow. Either\n  the rule belongs to a single aggregate, or the business must accept eventual consistency *and*\n  name the compensating action. Leaving it unstated is how double-bookings ship.\n- **God context** — one context appears in every flow, usually as a coordinator forwarding messages.\n  Orchestration is legitimate; a context that decides nothing of its own is a hop, not a boundary.\n- **Pass-through** — a context receives a message and forwards it without making a decision.\n- **Chatty pair** — two contexts exchanging many messages in one flow. Count them; a pair that talks\n  more inside a scenario than either talks to anyone else is a candidate merge.\n\nEvery finding gets: the flow it came from, the messages that show it, and what it would take to\nresolve it. A smell with no evidence attached is an opinion.\n\n### 4. Feed findings back — do not redraw\n\n**Two conditions refute the decomposition outright**, and they are the only place in the whole\nmodelling loop where a design is rejected by evidence rather than by opinion:\n\n> **more than 9 messages in one scenario**, or **one context appearing at every step** ⇒ go back and\n> re-cut.\n\nSay it in those words when either fires. It is not a diagram-tidiness note; it is the loop-back\ntrigger, and `3-decompose` becomes stale the moment it fires.\n\nWhere a flow proves a boundary wrong, write it as a **proposed change** with its evidence, and hand\nit to `3-decompose` (update mode) to merge. Where a flow reveals a domain event nobody discovered,\nhand it to `2-discover` to confirm with people — do not promote your own inference to a confirmed\nevent.\n\nThis is the loop working as intended. The modelling process is iterative; connect is where decompose\ngets its corrections.\n\n### 5. Emit\n\nWrite one file per flow under `docs\u002Fdomain\u002Fmessage-flows\u002F`, plus a `README.md` index that lists the\nflows and collects every finding in one place. `status: draft`, `owner: TBD`.\n\n## Output shape\n\n````markdown\n---\nid: DOMAIN-FLOW-0001\ntitle: \u003CUse case> — domain message flow\nstatus: draft\nowner: TBD\ndate: \u003Cdate>\ncontexts: [Allocation, Logistics, Billing]\n---\n\n## Scenario\n\u003C!-- one paragraph, in business language: who wants what, and what \"done\" means -->\n\n## Flow\n\n```mermaid\nsequenceDiagram\n  actor Planner as Depot Planner\n  participant Logistics\n  participant Allocation\n  Planner->>Logistics: 1. ScheduleTransfer [cmd] {unitId, fromDepot, toDepot, window}\n  Logistics->>Allocation: 2. IsUnitAvailable? [qry] {unitId, window}\n```\n\n| # | From | Message | Type | Contents | To | When |\n|---|---|---|---|---|---|---|\n| 1 | Depot Planner | `ScheduleTransfer` | command | unitId, fromDepot, toDepot, window | Logistics | — |\n| 2 | Logistics | `IsUnitAvailable?` | query | unitId, window **→** available, freeFrom | Allocation | — |\n| 3 | Logistics | `TransferLapsed` | event | unitId | Allocation | **after** 30 min of no confirmation |\n\n**A query carries its response in the same row, after a `→`.** The notation draws a query and its\nanswer as one unit precisely because the sender is *blocked* in between — splitting them into two\nrows doubles the diagram and hides the only interesting thing about a query. Everything left of the\narrow is what the sender sends; everything right of it is what it waits for. A query with nothing\nafter the `→` has not been thought through: someone has to say what comes back, because the shape of\nthe answer is what the caller is coupled to.\n\n**The `When` column is for time-driven messages only**, and it exists because the semantics decide\nthe design: **within** 5 minutes, **after** 5 minutes and **every** 5 minutes are three different\nsystems. Leave it `—` for a message that follows immediately from the one above. Put the trigger\nhere rather than in the scenario prose — a temporal rule buried in a paragraph is invisible on the\ndiagram, and it is usually the rule that produces the failure path nobody drew.\n\n## Findings\n| # | Smell | Evidence (messages) | What it suggests | Proposed change |\n|---|---|---|---|---|\n\n## Open questions\n\u003C!-- one line each: the question, and who could answer it -->\n````\n\nThe `README.md` index carries the use-case list, why each was chosen, and the consolidated findings\ntable with a status column (`proposed` \u002F `accepted` \u002F `declined`) so a later reader can tell which\nfindings were acted on.\n\n## Hard rules\n\n- **Length budget: ≤ 90 lines per flow file.** A flow longer than that is usually two scenarios.\n  A budget caps prose, not findings: over it, cut rationale a reader can infer and anything\n  restated from an upstream artifact — never open questions, provenance, or a stated absence.\n- **Never invent a message.** A flow is only as real as the events, commands and queries that were\n  discovered or modelled. Inferring `PaymentRefunded` because a refund \"must exist somewhere\"\n  produces a diagram that validates a design against fiction. If the flow has a gap, the gap is an\n  open question with a name attached.\n- **Domain messages, not transport.** No HTTP verbs, no queue names, no retry policy, no\n  serialization format. The moment a flow becomes an integration design, it stops being able to\n  challenge the boundary — every coupling problem starts looking like a technology choice.\n- **Type every message.** An untyped arrow hides the difference between a broadcast fact and a\n  blocking read, which is the difference the whole step exists to surface.\n- **A temporal rule belongs in the `When` column, never only in the scenario paragraph.** If the\n  scenario says \"within fifteen minutes\", some message is the one that must happen within fifteen\n  minutes — name it and put the rule on that row. Stated only in prose, the rule cannot be drawn,\n  cannot be checked, and is invisible to everyone reading the diagram instead of the paragraph.\n  `ddd_check` reports this as `temporal-rule-in-prose`.\n- **A query states what comes back.** `Contents` for a query reads `sent fields → returned fields`.\n  Without the `→` half, the row records a question nobody answered.\n- **Propose boundary changes; never apply them.** `3-decompose` owns the model — it holds the\n  reconciliation rules, the stable ids, the human edits and the extraction seam. A boundary quietly\n  redrawn here is a change nobody reviewed.\n- **Findings carry evidence.** Message numbers, not adjectives. \"Feels coupled\" is not a finding;\n  \"messages 2–5 are synchronous queries crossing three contexts before the planner gets an answer\"\n  is.\n- A flow that turns out clean is a real result. Say so — a use case that crosses four contexts with\n  four events and no queries is evidence the split is working, and it is worth recording as such.\n\n## Worked example\n\n**Input:** the equipment-rental model — contexts `Allocation`, `Logistics`, `Billing`,\n`Notifications` — plus a discovery hotspot about units being double-booked across depots.\n\n**Use case chosen:** *\"Priority depot transfer, booked and billed\"* — the paid add-on, and the\nscenario touching the most contexts.\n\n| # | From | Message | Type | To |\n|---|---|---|---|---|\n| 1 | Depot Planner | `ScheduleTransfer` | command | Logistics |\n| 2 | Logistics | `IsUnitAvailable?` | query | Allocation |\n| 3 | Allocation | *(response)* | query | Logistics |\n| 4 | Logistics | `ReserveUnit` | command | Allocation |\n| 5 | Allocation | `EquipmentAllocated` | event | — |\n| 6 | Billing | `TransferFeeCharged` | event | — |\n\n**Findings the flow forced out:**\n\n| Smell | Evidence | What it suggests |\n|---|---|---|\n| Synchronous query chain | 2–4: Logistics asks, then commands, on the same data | check-then-act across a boundary — between 2 and 4 another planner can reserve the same unit |\n| Distributed invariant | the no-double-booking rule is enforced by Logistics' check but owned by Allocation's data | the rule belongs to **one** aggregate; Allocation should own `ReserveUnit` end-to-end and answer with success or rejection |\n\n**Proposed change handed back to `3-decompose`:** collapse the check-then-act pair into a\nsingle `ReserveUnit` command that Allocation either accepts or rejects, and record the invariant as\nAllocation's. This is the same double-booking hotspot discovery surfaced — the flow is what turned\nit from a worry into a located defect with two message numbers on it.\n\nNote what the example does **not** do: it does not move `Billing` into `Logistics` because the two\nappear adjacent, and it does not silently edit `docs\u002Fdomain\u002Fallocation\u002Fmodel.yaml`. It writes the\nfinding, names the change, and leaves the model to the skill that owns it.\n",{"name":5,"description":24,"disable-model-invocation":25},"DDD loop 2 — connect: Domain Message Flow Modelling per use case — typed messages, coupling smells, boundary deltas handed back to decompose. Writes docs\u002Fdomain\u002Fmessage-flows\u002F.\n",true,"main","850634e254ea6a5034bf2eddad5fe95091713231","docs\u002Fresearch\u002Fddd-skill-evals\u002Fskill-snapshot-pre-context-pack\u002F4-connect","https:\u002F\u002Fgithub.com\u002Fbaodq97\u002Fgovkit","https:\u002F\u002Fgithub.com\u002Fbaodq97","2026-09-07 14:26:00",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":25},[34,37,40],{"tool":35,"command":36},"npx","npx skills add baodq97\u002Fgovkit --skill 4-connect",{"tool":38,"command":39},"bunx","bunx skills add baodq97\u002Fgovkit --skill 4-connect",{"tool":41,"command":42},"pnpm","pnpm dlx skills add baodq97\u002Fgovkit --skill 4-connect","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fbaodq97\u002Fgovkit\u002Fblob\u002Fmain\u002Fdocs\u002Fresearch\u002Fddd-skill-evals\u002Fskill-snapshot-pre-context-pack\u002F4-connect\u002FSKILL.md",[45,51,55],{"path":46,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":49,"sourceHtmlUrl":50},"evals\u002Fevals.json","REFERENCE","参考文档",4063,"https:\u002F\u002Fgithub.com\u002Fbaodq97\u002Fgovkit\u002Fblob\u002Fmain\u002Fdocs\u002Fresearch\u002Fddd-skill-evals\u002Fskill-snapshot-pre-context-pack\u002F4-connect\u002Fevals\u002Fevals.json",{"path":52,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":53,"sourceHtmlUrl":54},"references\u002Fcoupling-heuristics.md",7174,"https:\u002F\u002Fgithub.com\u002Fbaodq97\u002Fgovkit\u002Fblob\u002Fmain\u002Fdocs\u002Fresearch\u002Fddd-skill-evals\u002Fskill-snapshot-pre-context-pack\u002F4-connect\u002Freferences\u002Fcoupling-heuristics.md",{"path":56,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":57,"sourceHtmlUrl":58},"references\u002Fmessage-flow-notation.md",7281,"https:\u002F\u002Fgithub.com\u002Fbaodq97\u002Fgovkit\u002Fblob\u002Fmain\u002Fdocs\u002Fresearch\u002Fddd-skill-evals\u002Fskill-snapshot-pre-context-pack\u002F4-connect\u002Freferences\u002Fmessage-flow-notation.md",[60,67,74,81,88,94],{"skillNo":61,"name":62,"nameZh":6,"description":63,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":64,"categories":65,"keywords":66,"publishedAt":19},"01M759HTCHQJVZ4JWDA2JNPJJZ","trigger-shaped","Use when the user says \"plan a sprint retro\"; drafts retro\u002Fnotes.md for the team.","https:\u002F\u002Fgithub.com\u002Fbaodq97\u002Fgovkit\u002Fblob\u002Fmain\u002Fscripts\u002Ffixtures\u002Fskill-lint\u002Ftriggering\u002Fskills\u002Ftrigger-shaped\u002FSKILL.md",[],[],{"skillNo":68,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":71,"categories":72,"keywords":73,"publishedAt":19},"01M759HS3E2T97ZN1ED2Z96WAJ","terse-unguarded","Step 2 — writes docs\u002Fx\u002Foutput.json quietly.","https:\u002F\u002Fgithub.com\u002Fbaodq97\u002Fgovkit\u002Fblob\u002Fmain\u002Fscripts\u002Ffixtures\u002Fskill-lint\u002Ftriggering\u002Fskills\u002Fterse-unguarded\u002FSKILL.md",[],[],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":19},"01M759HPZN28VZHKYFR754CVXT","guarded","Summarizes ledger totals into reports\u002Fmonthly.csv.","https:\u002F\u002Fgithub.com\u002Fbaodq97\u002Fgovkit\u002Fblob\u002Fmain\u002Fscripts\u002Ffixtures\u002Fskill-lint\u002Ftriggering\u002Fskills\u002Fguarded\u002FSKILL.md",[],[],{"skillNo":82,"name":83,"nameZh":6,"description":84,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":85,"categories":86,"keywords":87,"publishedAt":19},"01M759HMFQAZAWFVSS3KWPCCW0","open","Draft a spec. Use when asked to \"draft a spec\" or \"write the RFC\".","https:\u002F\u002Fgithub.com\u002Fbaodq97\u002Fgovkit\u002Fblob\u002Fmain\u002Fscripts\u002Ffixtures\u002Fskill-lint\u002Fdead-hint\u002Fskills\u002Fopen\u002FSKILL.md",[],[],{"skillNo":89,"name":76,"nameZh":6,"description":90,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":91,"categories":92,"keywords":93,"publishedAt":19},"01M759HJ3XX74S70PW9AR8NS66","Deploy the thing.","https:\u002F\u002Fgithub.com\u002Fbaodq97\u002Fgovkit\u002Fblob\u002Fmain\u002Fscripts\u002Ffixtures\u002Fskill-lint\u002Fdead-hint\u002Fskills\u002Fguarded\u002FSKILL.md",[],[],{"skillNo":95,"name":96,"nameZh":6,"description":97,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":98,"categories":99,"keywords":100,"publishedAt":19},"01M759HCWJN9ACRNC1JW76DR1E","working-discipline","Thinking-checkpoint discipline for substantive implementation, debugging, or autonomous work — 21 trigger-gated items (trigger → mandatory question → evidence to produce) that cut agent error and human escalation. Use before an autonomous or long-running run, when asked to \"work carefully\", \"double-check\", \"làm cẩn thận\", \"reduce errors\", \"be more autonomous\", or after looping twice on the same failing fix. NOT a procedure — checkpoints fire on their triggers and demand evidence, with an N\u002FA escape valve so a strong model pays near-zero overhead.","https:\u002F\u002Fgithub.com\u002Fbaodq97\u002Fgovkit\u002Fblob\u002Fmain\u002Fplugins\u002Fswe-flow\u002Fskills\u002Fworking-discipline\u002FSKILL.md",[],[]]