cross-cut-analysis

Cross-Cut Analysis

Role

You are performing systematic cross-cut analysis over a set of mmp chunks. Your job is to identify risks before parallel L1 dispatch, not during or after.

When to Invoke

• L0 Phase 1: before finalizing chunks.md and dispatching L1 agents
• Manual: user runs /mmp:cross-cut-analysis to inspect a wave in progress
• Diagnostic: when a wave fails with unexpected conflicts post-dispatch

Workflow

1. Gather Inputs

Collect all boundary.yaml files from .orchestrator/chunks/*/boundary.yaml.

Each boundary.yaml may contain:

allow: []                # file patterns this chunk may write
touches_domains: []      # semantic domain tags
produces: []             # artifacts this chunk creates
consumes: []             # artifacts this chunk requires
depends_on: []           # explicit chunk dependencies
tooling_deps: []         # CLI tools required

2. Run Analysis

Execute:

bash scripts/cross-cut-analysis.sh \
  --chunks-dir .orchestrator/chunks \
  --orch-dir .orchestrator

Or invoke directly from skill context by reading each boundary.yaml and applying the five dimensions below.

3. Interpret Report

Report format:

cross_cuts:
  - severity: HIGH|MEDIUM|LOW
    type: file_overlap|domain_conflict|implicit_dependency|concurrency_risk|tooling_dependency
    chunks: [<chunk_id>, ...]
    resource: <file path or domain:tag or tool:name>
    recommendation: serialize|merge|force-parallel-with-brief|accept-with-coordination

4. Act on Findings

Severity	Action
HIGH	Must resolve before dispatch. Serialize (add depends_on) or merge chunks.
MEDIUM	Generate cross-cut briefs with coordination instructions.
LOW	Document in brief. Proceed with explicit coordination note.

Five Analysis Dimensions

Dimension 1: File-Overlap

Two or more chunks declare the same file pattern in allow[].

• Severity: HIGH
• Recommendation: serialize (add depends_on to one chunk) or merge chunks
• Detection: compare allow[] sets across all chunks

Dimension 2: Domain-Tag-Overlap (domain_conflict)

Two or more chunks share a tag in touches_domains[].

• Severity: MEDIUM
• Recommendation: force-parallel-with-brief — generate sentinel markers and cross-cut briefs for affected chunks
• Note: domain overlap alone does not block parallel dispatch if file boundaries are clean

Dimension 3: Implicit-Dependencies

Chunk A declares artifact X in produces[]. Chunk B declares X in consumes[]. Neither declares depends_on: [chunk-a].

This is a hidden serialization requirement — B will fail if it runs before A completes or if A's output format changes.

• Severity: HIGH
• Recommendation: serialize — add depends_on: [<producer>] to consumer's boundary.yaml
• Fix action: L0 updates state.yaml for consumer chunk to add depends_on

Dimension 4: Concurrency-Risk

Two or more chunks include a shared state channel in allow[]:

• .orchestrator/audit.jsonl
• .orchestrator/status.yaml
• hooks/hooks.json
• Any *.lock file

These files require atomic writes. Parallel chunks without flock coordination will corrupt shared state.

• Severity: HIGH
• Recommendation: serialize or ensure scripts use flock/mkdir-atomic pattern

Dimension 5: Tooling-Dependencies

Two or more chunks list the same tool in tooling_deps[].

• Severity: LOW
• Recommendation: accept-with-coordination — ensure tooling is installed in pre-flight; note in L1 prompts

Output

YAML Report (stdout)

cross_cuts:
  - severity: HIGH
    type: implicit_dependency
    chunks:
      - chunk-a
      - chunk-b
    resource: config/db.yaml
    recommendation: serialize

Audit Event

Emitted to .orchestrator/audit.jsonl:

{
  "kind": "cross_cut_analysis_run",
  "ts": "...",
  "details": {
    "issue_count": 3,
    "severity_distribution": {"HIGH": 2, "MEDIUM": 1, "LOW": 0}
  }
}

Integration with Phase 1

L0 calls this skill after decomposing chunks but before confirming the dispatch DAG. HIGH findings must be resolved (serialized) before proceeding. MEDIUM findings trigger brief generation. LOW findings are logged.

See references/heuristics.md for detection edge cases and worked examples.