Giancarmine SalucciandClaude Opus 4.8
81b33fff26
The skill no longer modifies code. It emits exactly two artifacts under simplify/: findings.json (machine-actionable ledger) + a ranked report. Every finding now REQUIRES title/action/verify (locked in ledger.schema.json) so each is independently executable by the calling agent. Removed: apply/report modes, Track A autofix execution, baseline gate, verify-each-change loop, the include-tests switch (tests excluded by default). Reframed SKILL.md, README, PATTERNS (deterministic/judgment), DETECTION, OPTIMIZATION (judge-only model tiering), PORTABILITY, SVELTE, CONSOLIDATION accordingly. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
3.0 KiB
Token & model optimization
Make the model do the least possible work, and only the hard part. These are layered; apply as many as your runtime supports. All are optional accelerators — the skill is correct without any of them.
1. Push work out of the model
- Detection is the bundled binaries — zero model tokens.
- Mechanical findings just record a deterministic autofix command in their
action— the model never reads their code. - Goal: the model only judges
difficulty: semantic | hardcross-file candidates.
2. Span-scoped reads
Never read a whole file. From the ledger you already have
file + start_line/end_line; read only that range plus a few context lines (~95% fewer
read tokens on large files). The codebase enters context only as flagged spans.
3. Cheap-model steps
Use a small/fast model (Haiku / GPT-5-mini / Gemini Flash) for:
- ledger assembly (Step 3) — mechanical JSON→JSON join with the manifest;
- classification / triage — mechanical vs semantic, confidence;
- verification reading — interpreting build/test output (pass/fail + which finding).
4. Model tiering for the JUDGE (capability-detected)
The model is used only to judge cross-file candidates (Pass B), never to edit.
- Cluster + judge on a cheap model.
hard/ multi-file / low-confidence → escalate to a frontier model only then. Confidence-based cascading typically cuts cost ~50–85% at ~95% quality.- How to route, by runtime:
- Claude Code — sub-agents with a per-agent
model:override. - pi — provider-agnostic; pick the model per sub-agent/step.
- Copilot / others without per-step model choice — run on the invoked model and use
confidenceto decide which findings to include vs leave out (low-confidence clusters simply aren't reported).
- Claude Code — sub-agents with a per-agent
5. Optional sub-agent fan-out
On runtimes with sub-agents (Claude Code, pi), run one judge per cluster in parallel isolated contexts; each returns a compact JSON verdict to the parent. Keeps the orchestrator context clean. Falls back to a sequential judge loop everywhere else.
6. Cache-friendly prefix
Keep static context (the pattern catalog, the manifest, the ledger summary) front-loaded and stable so the runtime's prompt cache hits; append per-finding dynamic context (the span) after it. Don't reorder the stable prefix between findings.
7. Resumability = the cheapest rerun
findings.json status is the source of truth. On rerun, load it and process only
pending items — an interrupted or repeated run skips finished work (~80–90% saving).
Don't re-scan-and-overwrite a ledger that still has pending work.
Rough budget intuition
detect = free (binaries). assembly = cheap model, one pass over compact JSON. judging = one cheap-model verdict per cross-file cluster (mechanical/structural findings need no judging — free). So cost scales with the number of cross-file clusters to judge, not with repo size. The skill performs no edits, so there is no apply/verify token cost.