Menü
Two-phase read-only code-smells analysis: regression checks + new systemic patterns
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Basierend auf der SOC-Berufsklassifikation
Reconstruct Dazzle/HM judgement from stems before inventing structure
Reconstruct Dazzle/HM judgement from stems before inventing structure
Run all quality checks on modified files — the on-demand pre-ship quality gate
Commit, verify, tag, and push with the repo's pre-flight gate suite (lint, type, drift gates, docs build)
Use when authoring or revising trial.toml scenarios for dazzle qa trial — puts an LLM in the shoes of a real business user evaluating a Dazzle app. Use when the user asks to "set up qa trials", "write a trial scenario", "evaluate this app as a user", or mentions qualitative/trial/business-user testing. Also use when a trial report is sparse or unhelpful — the scenario itself is usually the root cause.
Use when writing or editing Dazzle .dsl files — syntax rules, the mistakes agents actually make, and pointers to the drift-gated references. Complements (never replaces) CLAUDE.md's DSL Quick Reference.
| name | smells |
| description | Two-phase read-only code-smells analysis: regression checks + new systemic patterns |
Run a two-phase code smells analysis: regression checks against established rules, then a scan for new systemic patterns. This is a read-only analysis — do NOT make any code changes.
This command runs as a Workflow (.claude/workflows/smells.js): the regression checks, the three pattern-category finders, and the decay-harness finder fan out as parallel agents, each returning schema-validated findings. This main loop then writes the report + log from the validated result and presents the summary. (Replaced the hand-rolled "dispatch 4 background subagents" prose — the Workflow gives deterministic fan-out + structured output.)
Structural decay is now measured precisely and gated live in CI by the framework structural-fitness harness — so /smells no longer re-derives it by grep heuristics, it consumes it:
dazzle fitness code (churn×complexity hotspot queue → dev_docs/framework-hotspots.md), the complexity ratchet (tests/unit/test_complexity_ratchet.py + tests/unit/fixtures/complexity_baseline.json — gates new CC>15 / MI-rank drops), and the import-linter layer contracts ([tool.importlinter] — gates new core↛back/ui, ui↛back, back↛sqlite). The decay-harness finder surfaces the standing baselined debt these gates hold flat: top hotspots, the MI-rank-C set, the highest-CC functions, ratchet status, and the import allow-list size. Enforcement of new structural decay belongs to those gates, not to a smells recommendation./smells owns what the harness can't see. Its real value-add is the semantic smells radon/import-linter are blind to: silent failures / swallowed exceptions, inconsistent error strategy, near-duplicate blocks, Any/# type: ignore masking, mutable globals/hidden singletons, and compat shims. The complexity-globals finder is now seeded by the hotspot queue — it spends agent judgment naming the structural smell inside the top decayed files rather than re-counting lines.Backward compatibility is NOT a requirement at this stage. The project has one major user who is fully engaged with the dev process. When recommending fixes:
(The workflow's finder prompts already encode this policy + the scope below.)
Focus on src/dazzle/ — the merged tree (back/, ui/, render/ all live under it since the #1055 package merge; #1056 was the follow-on mypy burndown). Ignore tests/, examples/, and auto-generated files.
Invoke the Workflow tool with name: "smells". It returns:
{
regressions: [{id, check, status: PASS|FAIL|TRACK, details}],
patterns: [{pattern, category, instances, root_cause, canonical_fix, done_criteria, enforcement}],
decay: {hotspots:[{rank,file,score,churn,mi_rank}], c_rank_files:[…], high_cc_functions:[{file,function,cc}],
ratchet_status, import_contracts:{status,allowlist_size,entries:[…]}, priority_targets:[…], notes},
regressed: <count of FAIL regressions>
}
The finders inherit the session model (no model override) per the Subagent Model Policy in CLAUDE.md — pattern recognition is judgment work.
From the returned data, write agent/smells-report.md:
## Code Smells Report — [date]
### Regression Check Results
| # | Check | Status | Details |
|---|-------|--------|---------|
(one row per `regressions` entry)
### New Patterns Found
| Pattern | Category | Instances | Root Cause | Fix |
|---------|----------|-----------|------------|-----|
(one row per `patterns` entry, ordered by severity × instance count)
### Structural Decay (live harness)
From `decay`. The CI ratchet + import contracts gate *new* decay; this is the standing baseline.
- **Ratchet:** [`decay.ratchet_status`] **Import contracts:** [`decay.import_contracts.status`], allow-list size [`decay.import_contracts.allowlist_size`]
- **Priority refactor targets** (high-churn × MI-rank-C): [`decay.priority_targets`]
| Rank | Hotspot file | Score | Churn | MI |
|------|--------------|-------|-------|----|
(top rows from `decay.hotspots`)
Highest-CC functions: (top 3–5 from `decay.high_cc_functions` as `file:function (cc N)`). Note: [`decay.notes`].
### Recommended Next Actions
1. [highest-priority *semantic* pattern — the harness already owns structural enforcement]
2. [second]
3. [the single best refactor target from `decay.priority_targets`, if a finder named a concrete smell inside it]
### Comparison with Previous Round
- Regressions: X checks regressed
- New patterns: Y found
- Resolved since last round: [diff against the previous agent/smells-report.md if present]
- **Decay delta:** files that *climbed* the hotspot ranking vs the previous report's table (rising debt — the leading indicator); any change in `import_contracts.allowlist_size` (must not grow)
Append a timestamped summary to agent/smells-log.md (create if missing):
## Smells Run — [date]
- Regressions: X/Y checks passed
- New patterns: Z found
- Top concern: [highest priority pattern]
- Decay: ratchet [clean|N violations], import contracts [kept|broken], allow-list [size]; top target [priority_targets[0]]
- Commit: [current HEAD sha]
Present the regression table and the top patterns. Surface any FAIL regressions first — those are the ones that broke an established rule (note: lint-imports broken in 1.9 or a non-clean ratchet_status means a structural gate regressed — treat as a FAIL-class signal). Then give the one-line decay summary (ratchet/contracts status + the single best refactor target). Do NOT make any code changes; this is read-only analysis.