Service Decomposition: $ARGUMENTS

$ARGUMENTS names the candidate boundary ("billing", "the notifications module") or an existing service to audit; empty = survey the whole system, extraction candidates and existing splits both.

Principle

A service boundary is not a code-quality tool — it is an organizational and operational tool that happens to involve code. Splitting a process buys independent deploys, independent scaling, and failure isolation; it pays for them with network partial-failure, the loss of cross-boundary transactions, contract versioning, and an operational floor per service. Messy code is never a driver: a mess cut in two is two messes joined by a network.

The governing rule is extract the seam before the service. Every good service boundary must first exist and hold as a clean in-process module boundary — own interface, own data access, no reach-ins. If the boundary can't be drawn inside one process, the network will not draw it for you. This is core.md's door rule: extraction is a heavy, nearly-one-way door, but modularization is how you change the door — a proven in-process seam makes later extraction cheap and is worth having even if you never extract. So the default verdict is always the modular monolith, and extraction must argue its way past it.

This skill decides and records — it does not restructure. Execution routes out (Phase 4).

Phase 0 — Load profile

Tooling preflight — one call, before step 1. Some tools this skill relies on are deferred by the harness: the session lists them by name only and loads their schemas on demand, so calling one before it is fetched fails. Listing a tool in allowed-tools does not un-defer it. Issue a single ToolSearch up front covering the whole run — select:WebFetch,WebSearch (Phase 1's live feasibility checks on infra and platform limits) — instead of one round-trip per discovery. A name already loaded costs nothing to include; a schema discovered missing mid-run costs a turn.

1. Read .claude/PROJECT.md — Architecture (layers, module map, dependency direction), Integrations (existing infra: queues, gateways, deploy platform), Commands. If missing or still TEMPLATE, fall back to the root CLAUDE.md (always in context) when it carries those facts — proceed on it, noting you're running without a kit profile. Only if neither has them, STOP and tell the user to run /bootstrap first.
2. Read CONTEXT.md / CONTEXT-MAP.md if present — bounded contexts are the natural candidate boundaries; name candidates in their terms.
3. Read docs/adr/ (or the ADR location in PROJECT.md) — a prior decision about this boundary means start at /revisit, not here: reopening it requires a broken assumption, not a second opinion.
4. Read skills/codebase-design/SKILL.md for the seam/interface vocabulary — every boundary statement uses it — and WHEN-TO-CUT.md for the tests a boundary must pass (axis of change, writer census, next feature).

Phase 1 — Evidence, not vibes

Collect per candidate boundary. Fan out arch-tracer agents for the coupling map on anything beyond a small tree; pass .claude/schemas/finding.schema.json for mergeable output. Ground every claim in path:line or a command's output.

Code evidence (agents can find this):

• Runtime coupling — synchronous call chains crossing the candidate boundary, shared tables/schemas, shared transactions, shared caches, shared in-memory state. Each one is either severed before extraction or becomes a distributed failure mode after it.
• Data ownership — can each side own its data outright? A table written by both sides is the single strongest STAY/MODULARIZE-first signal.
• Connection budget — services × pool size × expected instances against the store's ceiling; waits surface as request timeouts and tripped breakers, not connection errors.
• Change coupling — git log --format=%H --name-only mined for commits touching both sides of the boundary: code that ships together wants to live together. Independent change histories support a split; interleaved ones refute it.
• Scaling asymmetry — evidence (from /perf reports, profiles, infra dashboards named in Integrations) that one part needs materially different resources or scale.

Organizational evidence (agents cannot see the org — ASK, never guess): Use AskUserQuestion for: team count and ownership map (Conway's law: the service topology you can sustain is the team topology you have), deploy cadence and who blocks whom, on-call/ops capacity, and the platform's operational maturity (CI/CD per service, observability, service discovery). One team ≈ one deployable is the reference class; a one-team org extracting microservices is buying tax with no driver.

External evidence (research, when the answer isn't in the repo): Platform capabilities, event-bus/gateway options, reference-class outcomes for this kind of split — verify live via WebSearch/WebFetch; route a load-bearing unknown ("can our platform even run N services?") to /spike, and an infra component choice (broker, gateway) to /select-tech. Never decide on remembered facts about platforms or tools.

Phase 2 — Weigh drivers against the tax

Real extraction drivers (each needs Phase-1 evidence to count):

Fake drivers — name them when you see them, they never justify extraction: "the code is a mess" (→ /arch-health), "microservices are best practice", "it'll be easier to understand" (locality drops across a network), "we might need to scale someday" (speculative scale = STAY; extract when measured), résumé-driven architecture, "it's too big" / "it's too small" (size is not a finding — co-change is). Against every driver that counts, name the boring alternative — a bigger machine, reassigned ownership, a deploy pipeline replacing manual testing and long-lived branches — say why it is insufficient, and name the measure that will show in three months whether the split worked.

Vetoes — no driver outweighs these; when one fires the verdict is STAY and the report says which. Atomicity: two updates the business needs both-or-neither stay in one transaction and one service — a saga buys "both eventually, or an observable compensation". Consistent read: a decision needing one consistent view of both sides must read it from one store. Operated by the customer: someone else's per-service ops floor is not yours to spend — cap at MODULARIZE.

The distributed tax — the extraction side must accept ALL of these, listed explicitly: network calls that partially fail (timeouts, retries, idempotency per resilience rule), no cross-service transactions (sagas / eventual consistency where a BEGIN used to do; durable execution is a third option, priced in class limits not vendor ones — deterministic replay, and workflow code versioned while old runs are still in flight), versioned contracts between the parts (Hyrum's law now applies internally), per-service observability + deploy + on-call floor, harder local dev, and data duplication where joins used to be — plus the agent tax, which nobody prices in: one user-visible flow now spans N repos, N deploys, and N telemetry stacks, so no single agent session can trace it end-to-end with Read and Grep. Comprehension shifts from code-reading (cheap, always available) to distributed tracing (needs the observability floor already built). The monolith's most under-priced asset is that its entire truth fits one context window.

Which axis? A cut runs along abstraction (layers inside one deployable), subdomain (services) or instances (replicas, shards): say in one line why the other two do not answer this pressure, and while the domain is still being learned cut along abstraction only.

Verdict per candidate boundary:

• STAY — no evidenced driver. The monolith is the correct architecture; say so plainly. A monolith serving its load with one team is a success, not a smell.
• MODULARIZE — drivers are real or plausible but Phase-1 found reach-ins, shared data, or interleaved change history. Make the boundary real in-process first: own interface, own data access, composition-root wiring. This is the default when in doubt — it captures most of the benefit, keeps transactions, and leaves the door open.
• EXTRACT — evidenced drivers outweigh the accepted tax and the in-process seam already exists and holds (no reach-ins, data ownable, contracts nameable). Both conditions, not either. Name the remainder, not just the piece cut out: one sentence, no "and", for what the other side does. A remainder that survives only as "the rest of X" means the cut is in the wrong place — move it, cut into three, or leave it whole.
• MERGE — the boundary already exists and is wrong: co-change, lockstep deploys, a shared transaction, workflow, hot domain code or unownable data. Name the blast radius bought back.

Phase 3 — Grill and record

Take the top-ranked verdict (user picks via AskUserQuestion if it's a genuine toss-up) and /grill it: the exact interface at the seam, who owns which data, what happens on partial failure, the migration's first reversible step. Then record the verdict — including STAY — as an ADR via /domain-model, with the drivers, the accepted tax, and the assumptions it rests on stated explicitly (that's what a future /revisit will test).

Phase 4 — Route to the executor

Output

Write the report below to the Plans location from PROJECT.md (e.g. <plans>/<YYYY-MM-DD>-decompose.md; git policy per PROJECT.md → Artifact git policy), then present it:

## Decomposition — <scope>

### Verdicts
| Boundary | Verdict | Drivers (evidenced) | Blockers / tax accepted | Route |

### Evidence
- <boundary>: coupling `path:line`…, change-coupling stats, org facts (from user)

### Recommended next
- ⭐ <boundary> — <verdict> — <one-line why> → <route>

### ADR
- <recorded / to record via /domain-model>

Hard rules

• Decide and record only — never restructure here; execution goes through Phase-4 routes.
• Default is the modular monolith. Extraction must clear both bars: evidenced drivers AND a proven in-process seam. Never extract along a seam that doesn't hold in-process.
• Org facts come from the user. Team topology, ops maturity, deploy pain — ask; never infer from code.
• Fake drivers are findings. Name them in the report instead of acting on them.
• STAY is a first-class verdict. Record it as an ADR so the question isn't relitigated.
• External facts verified live, never from memory.

Cross-reference

• /codebase-design — the seam/interface vocabulary; MODULARIZE work is designed in it.
• /arch-health — finds the rot that often masquerades as an extraction driver.
• /revisit — start there when a prior ADR already covers this boundary.
• /rollout — the staged execution of an EXTRACT or MERGE verdict, and the behaviour freeze that keeps its stages reversible.
• /domain-model — bounded contexts in; the verdict ADR out.
• /select-tech, /spike — infra component choice / feasibility unknowns.