v0.16.1 — Token-optimized sub-agent execution for implement-backlog
Changed
/architect:implement-backlog — token-optimized sub-agent execution. The skill now runs as a thin sonnet orchestrator (was opus) that delegates heavy steps to model-tiered sub-agents, per a new "Sub-Agent Execution & Model Assignment" section:
| Step | Delegated work | Model |
|---|---|---|
| 1 | Shared-context pack derivation (parallel, one agent per file) | sonnet |
| 3 | Epic / siblings / design-report digest (Explore) | haiku |
| 4 | Mini-plan drafting against Epic-wide contracts | opus |
| 5 | Implementation per coherent unit (opus escalation only for judgment-heavy design) | sonnet |
| 6 | Epic-consistency verdict & whole-Epic roll-up review | opus |
| 7 | Progress comments + impl-log mirror drafting | haiku |
Two cost rules are now explicit:
- Context protection — the orchestrator holds compact digests, never full report/source bodies.
- Cheapest-capable tier — opus is reserved for planning and consistency judgment only.
AGENTS.md model tiers and the CLAUDE.md command reference are synced, including guidance for runtimes without model switching (Codex/Omnigent) to preserve the delegation structure at the session model.
See CHANGELOG.md / CHANGELOG_ja.md for details.
🤖 Generated with Claude Code