| name | propose-decomposition |
| description | Introspective function. The organism inspects its own codebase via search_tutorials, search_research, search_docs, search_api_reference, get_coverage_gaps, Graphify dependency graphs, and InfraNodus topic analysis, then proposes a polymorphic decomposition plan. NEVER executes cell-divide. Human gate is mandatory. Depth-guarded: max 3 new levels of nesting per cycle. INERT until Phase B. |
| disable-model-invocation | true |
| user-invocable | true |
| source | lattice-internal (Phase 1 Amendment §2.2) |
| vendored_at | null |
| vendor_strategy | lattice-authored |
| local_adaptations | [] |
| lineage | [] |
propose-decomposition — Organism Self-Inspection
HARD CONSTRAINT — Phase 1: This skill is inert. Its sensory tools (search_tutorials, search_research, search_docs, search_api_reference, get_coverage_gaps, Graphify, InfraNodus) come online during Phase B per Amendments 06/07/08 and Phase 0.6/0.7/0.8 substrate landing. Do NOT invoke this skill in Phase 1.
Purpose
Let the organism examine its own structure and propose where the skill genome should divide next. Output is a proposal, never an execution. The human reads the proposal and decides whether to run cell-divide for the suggested target(s).
Inputs
| Argument | Required | Meaning |
|---|
focus_section | optional | If set, restrict introspection to one section (schema, api, frontend, georef-reality, genai, vw-itwin, ddc, ci, docs). If omitted, scans the entire organism. |
max_depth_delta | optional | Max new levels of nesting this proposal may suggest. Hard cap is 3 per the depth guard. |
Behavior (Phase B+)
- Substrate query pass (anti-amnesia pre-flight):
search_tutorials("decomposition patterns", tool_name=focus_section)search_research("polymorphic harness decomposition")search_docs("section ownership rules", tool_name="lattice")get_coverage_gaps(tool_name=focus_section, severity="high")
- Structural pass (Graphify):
- Run
graphify analyze --root . --format json.
- Identify hotspots (high coupling), hubs, bridges, cycles per
.claude/skills/improve-codebase-architecture/SKILL.md.
- Semantic pass (InfraNodus):
- Read the latest snapshot of
analysis/infranodus/desires.graph.json + analysis/infranodus/goals.graph.json (from the substrate).
- Identify gap and gateway topics that suggest where the harness needs more dedicated cells.
- Synthesis — produce a markdown proposal containing:
- One paragraph framing of the organism's current health (composite score from all section harnesses).
- Per candidate division: target path, parent commit SHA, expected lineage depth after division, rationale (which signals from passes 1–3 motivate this), risk callouts, and the exact
cell-divide command the human would invoke to execute it.
- Depth guard check: refuse to emit proposals that would introduce more than
max_depth_delta (default 3) new nesting levels in a single cycle. If more nesting is warranted, recommend the human run a follow-up cycle after the previous level is verified healthy.
- Explicit "NEXT ACTIONS" list with the exact human-review checkpoint.
- Emit to
meta/harness/PLAN/proposals/<UTC-timestamp>-decomposition.md. Do not commit; the human commits after review.
What this skill does NOT do
- It does NOT execute
cell-divide. It only describes what cell-divide would do.
- It does NOT mutate the genome, the substrate, the knowledge tables, or the docs.
- It does NOT propose deletions. Only divisions and promotions.
- It does NOT exceed the depth guard (3 new levels per cycle).
Hard constraint reminder
Per Phase 1 Amendment §2.2 + §5:
- Do not run this skill in Phase 1.
- The sensory layer (the six
@pxt.query tools + Graphify + InfraNodus) is itself in plan-state until Phase B applies migrations 0014–0016 and the bootstrap pipelines run on the M3 Max. Invoking this skill before Phase B completes will return zero results from every sensory call and produce a vacuous proposal.
