Architecture Rescue

Architecture Rescue

Use this skill when the user wants architecture cleanup that is broader than a single diff: repo-wide, subsystem-wide, or cross-module shape correction.

This skill is for analysis and proposal first. By default it is report-only.

What this is (and is not)

• This skill: architecture rescue and hygiene across seams, ownership, interfaces, and module depth.
• Not diff-scoped cleanup: this skill is for system shape, seams, and ownership, not for small behavior-preserving cleanup on a concrete diff.
• Not normal planning: planning slices approved feature work into phases/stories/beads; architecture-rescue diagnoses architecture friction and proposes rescue candidates before planning.

Default mode

Default to report-only unless the user explicitly asks for planning/execution follow-through.

In report-only mode:

• do not create beads
• do not invoke pulse:planning
• do not invoke pulse:executing
• do not edit architecture artifacts unless the user asks

Inputs to read first

For the target scope (whole repo or named subsystem), read:

1. project operating docs (CLAUDE.md, AGENTS.md, and nearby module docs)
2. relevant history/<feature>/CONTEXT.md decisions when present
3. .pulse/memory/* entries that apply to architecture constraints, incidents, or non-regression rules
4. pulse:gitnexus and GitNexus context if available

If GitNexus is configured, prefer it for topology, process flow, and impact/breadth evidence before manual grep.

Architecture lens

Use consistent terms from references/LANGUAGE.md.

Look specifically for:

1. Shallow modules: interface complexity close to implementation complexity
2. Bloated interfaces: callers need to know too many invariants/orderings/error modes
3. Leaky seams: cross-module coupling forces parallel edits or duplicated policy
4. Ownership drift: responsibilities spread across multiple modules with unclear source of truth
5. Deepening opportunities: ways to concentrate behavior behind smaller, clearer interfaces

Use the deletion test: if deleting a module merely moves complexity to callers, it was likely earning leverage; if it removes no real complexity, it was likely shallow or pass-through.

Process

1) Scope and map

• Confirm scope: whole repo, bounded subsystem, or named execution flow.
• Build a quick architecture map: modules, seams, owners, and flow handoffs.

2) Find rescue candidates

Generate a short list (3-7) of high-signal candidates with concrete evidence.

For each candidate capture:

• Area/files
• Current friction
• Rescue move
• Expected leverage/locality gain
• Risk/coordination cost

3) Produce a report

Return a ranked architecture rescue report:

1. Executive summary (what hurts most and why)
2. Candidate list (ranked)
3. Suggested rescue sequence (what to do first/second/later)
4. Open questions that must be answered before execution

Do not generate implementation tasks unless asked.

Optional follow-through modes (explicit opt-in only)

Use only when the user explicitly requests them.

• phase-shaping: hand off top approved candidate(s) to pulse:planning
• execution-ready: after validating gate approval, proceed via pulse:swarming/pulse:executing

Before leaving report-only mode, restate the mode change and requested scope.

Output contract (report-only)

Provide concise output with:

• scope analyzed
• top rescue opportunities
• rationale in leverage/locality terms
• dependencies and risks
• recommended first move

Keep recommendations specific, evidence-backed, and bounded to the requested scope.

References

• LANGUAGE.md
• DEEPENING.md
• INTERFACE-DESIGN.md