architect

Understand the architecture you have before deciding what to change. Start with:

Assess architecture and provide an action plan.

The agent first frames the decision, maps the implemented system, and shows an evidence ledger and attention heat map. You correct the map before it drills down, then redirect or accept the proposed hotspots. A standard run finishes with evidence-backed findings, strengths, unknowns, and dependency-aware action waves. Inspection is read-only by default; private knowledge, executable checks, and file writes require separate approval.

Assessment target   billing platform
Intent              hardening / risk reduction
Mode                standard
Current state       API + workers + ledger + payment provider
Evidence coverage   source ✓ tests ✓ CI ✓ operations partial

Map checkpoint — correct a boundary, or say “continue”.

Start with the outcome you need

Say this	What you get
“Assess architecture and provide an action plan.”	A standard current-state assessment, bounded investigations, and traced action waves
“Give me a quick architecture survey; stop before drill-down.”	A correctable map, evidence coverage, attention heat, and recommended investigations
“Do a deep launch-readiness assessment; ask before runtime access.”	A standard assessment extended with separately authorized operational or experimental evidence
“How should we design the replacement?”	A Stage-0 concept, full design doc, and convergence through `architect-design`
“Draw the current deployment topology.”	A self-checked Mermaid diagram through `architect-diagram`
“Review this assessment report.”	An evidence-and-methodology critique through `architect-review` or the cold-context reviewer

The generated architecture-lenses-reference skill is an internal knowledge router. You do not invoke it directly; assessment, design, and review load only the concepts their current question needs.

What the assessment reads and changes

The default pass may read documentation, source, tests, manifests, CI/CD, deployment and infrastructure definitions, schemas, configuration, operations files, and current local Git history. Its optional profiler executes no repository code, excludes protected path classes before evidence creation, shares finite entry/byte/path/time budgets across every phase, and emits signals, not an architecture score.

It does not run builds, tests, migrations, deploys, network calls, private enterprise queries, or experiments without asking. Saving is also optional. Current-state architecture, future designs, and decision records have distinct semantic destinations. In a compatible repository the pack consumes Core’s semantic-surface-resolution.v1; without Core it returns a zero-write portable handoff whose evidence the user may correct before a capable workflow resolves it. Chat-only and explicit personal-workspace saves remain available when the user-scope pack is used on its own.

Install

agentbundle install --pack architect --scope user <catalogue>

Adapters: claude-code, codex, copilot, kiro-ide, kiro-cli, cursor, and gemini. Default scope: user.

Go deeper

Maintainers: design and invariants.