Boundary Analysis — act101 Agent Skill

Boundary Analysis

Depth: Level 2 (Investigate).

See ../analysis-protocol/references/protocol.md for: artifact directory structure, the investigation loop, depth levels, summary format, token budget rules, and project map structure. Read that document before proceeding.

Phase 1: Parallel Tool Dispatch

Dispatch all available tools in a single parallel batch. Each subagent runs one tool, saves raw JSON to raw/<tool-name>.json, returns a structured summary.

Must-have tools:

If both must-have tools are unavailable, report that and stop.

Interpret analyze_seams through analyze_clusters. Empty seams are actionable only when clustering produced at least two non-hub clusters. If clustering is hub-collapsed or collapsed into one cluster, treat seam output as uninformative and investigate with dampened clusters, split_module, analyze_surface, or simulate.

Extended tools (use if available, skip and note in manifest if not):

Phase 2: Investigation

For each significant finding, form a hypothesis before dispatching a subagent:

Extraction candidate hypothesis example:

Hypothesis N: <candidate> can be extracted as a clean module. Evidence: analyze_seams identified a seam at this boundary; analyze_extraction scored it high. Confirming query: analyze_surface on the candidate files to measure cut cost (edges to sever vs. edges retained internally). Confirms if: Internal edges >> external edges (low cut cost relative to cohesion). Refutes if: External edges ≥ internal edges (high entanglement — extraction would be expensive).

Layer violation hypothesis example:

Hypothesis N: The violation from <source> to <target> is structural (not accidental), indicating architectural erosion. Evidence: analyze_layers flagged an inversion from layer X to layer Y. Confirming query: graph on <source> to trace the full import path. Confirms if: Multiple files in layer X import from layer Y (pattern, not one-off). Refutes if: Single file, single import — likely a quick fix rather than erosion.

Low-cohesion split hypothesis example:

Hypothesis N: <file> contains two distinct concept clusters that should be separate modules. Evidence: analyze_cohesion scored this module low; it is large. Confirming query: Read the analyze_cohesion LCOM4 fields. For any class with lcom4 ≥ 2, its components are the disjoint method/field clusters — the NAMED split boundary, preferred over eyeballing the skeleton. Cross-check the lcom4_summary modeled_kinds for coverage honesty: a grammar absent from modeled_kinds is unjudged, not cohesive. Confirms if: A class reports lcom4 ≥ 2 and its components partition the methods into groups with no shared state — split along those clusters. Refutes if: lcom4 is 1 (methods share common state/fields), or the file's grammar is absent from modeled_kinds so cohesion was never judged.

Interface bloat hypothesis example:

Hypothesis N: <module>'s public API is wider than necessary — some exports are never used externally. Evidence: analyze_interface_bloat flagged this module as over-exposed. Confirming query: references on each exported symbol. Confirms if: Several exports have 0 external callers. Refutes if: All exports have at least one external caller.

Save investigation notes to investigation/hypothesis-N.md.

Report Structure

# Boundary Analysis: <project name>

## Boundary Map
Current module structure with cluster sizes and labels.
Which clusters are cohesive vs. artificially grouped.

## Layer Architecture
Detected or user-specified layers. Direction consistency score.
(Present only if analyze_layers ran successfully.)

## Extraction Candidates
Ranked by extraction score. Per-candidate:
- Files included
- Internal vs. external edges
- API surface width
- Cut cost (edges to sever)
- Recommended extraction steps (specific act MCP operations)

## Layer Violations
(Present only if analyze_layers ran successfully.)
Per-violation: source layer, target layer, specific files involved, imported symbols,
fix recommendation. Distinguish: inversions (higher layer imports from lower) vs.
skips (layer bypasses an intermediate layer).

## Interface Width Assessment
(Present only if analyze_interface_bloat ran successfully.)
Modules with disproportionately wide public APIs.
Per-module: which exports are used externally, which are never called externally,
recommended visibility reduction.

## Orphan Types
(Present only if analyze_orphan_types ran successfully.)
Type definitions living in the wrong module.
Per-orphan: where defined, where used, recommended move destination.

## Cohesion Assessment
(Present only if analyze_cohesion ran successfully.)
Low-cohesion modules with natural split points. Per-module: cohesion score, identified
split boundary, recommended next step.
For each class with `lcom4` ≥ 2, report its `components` (the disjoint method/field
clusters) as the named split boundary — this is the LCOM4 evidence, preferred over an
eyeballed skeleton read. Quote the `lcom4_summary` `modeled_kinds` so coverage is honest:
a grammar absent from `modeled_kinds` was never judged and must not be reported as cohesive.

## Recommended Decomposition Steps
Prioritized, ordered steps to improve module boundaries.
Each step links to a specific act MCP tool call or skill.
Order: cheapest/highest-value extractions first, invasive restructuring last.

Before committing to a cut, simulate it. Express the proposed cut as `simulate` ops —
`split_file{file,groups}` for a class/module split, `move_file{from,to}` for a relocation —
and record the predicted deltas: `cycles.resolved`/`cycles.introduced`, the `coupling`
changes, `chokepoints`, and conformance `violations.cleared`/`violations.introduced` (the
last only when an `[architecture]` contract exists in `.act/config.toml`). `simulate` never
writes disk. Revise the cut if it introduces violations or fails to resolve the target cycle.
(Delta-field names match the architectural-refactoring skill's simulate step.)

To decide whether a suspected pass-through module is a real boundary at all — the
deletion test — simulate `delete_module{file}`. It drops the module and re-wires
transitive bridges (`A→M→B` ⇒ `A→B`), then reports a `deletions` delta. Read it in
this order:

1. `surface_consumers` — external symbols that **call or extend the module's own
   symbols** (the load-bearing core; `top_consumers` names them). A non-zero count
   means the module is load-bearing and earns its keep — those dependencies can
   never be re-homed because the callee/superclass body is deleted. Cite the count
   and a name or two instead of arguing the deletion test in prose.
2. `surface_modeled` — **honesty gate.** `surface_consumers: 0` is a genuine
   conduit signal ONLY when this is `true`. When `false` the call channel was not
   modeled for the module's grammar; the verdict is UNKNOWN, never "clean conduit"
   (field-access consumption is not modeled at all — treat 0 with care).
3. `rewired_edges` / `severed_edges` — the file-import routing side: high
   `rewired_edges` with `severed_edges: 0` and `surface_consumers: 0` (modeled) is
   a clean pass-through whose callers only routed through it.

Project Map Updates

Updates "Module Map", "Layer Architecture", and "Key Boundaries" sections. Appends to the Analysis History table.