← ClaudeAtlas

keel-executelisted

Use when a reviewed plan is ready to implement — orchestrates subagent-driven execution with per-task review and a crash-safe progress ledger, or falls back to inline execution when subagents aren't available. Stage 4 of the keel pipeline; hands off to keel-finish.
AWenSu/keel · ★ 0 · AI & Automation · score 75
Install: claude install-skill AWenSu/keel
# keel-execute — Plan Execution ``` INPUT a plan file with a header carrying Spec Version + Success Criteria, and every task carrying Delivers / Files / Depends on / Interfaces / Skills; if it came from keel-plan-review, a REVIEW REPORT ending in NO UNRESOLVED DECISIONS OUTPUT all tasks committed and reviewed on a non-main branch; ledger complete in .keel/progress.md; final whole-branch review passed ``` Missing INPUT → `BLOCKED: 缺 <field> → 退回 keel-plan`. A plan without those fields silently disables brief extraction, the staleness check, and the spec-compliance axis — three of this stage's four safety mechanisms. Two modes, one decision at the top: ``` Subagents available AND tasks mostly independent? ├─ yes → ORCHESTRATED mode (default; better context isolation, reviewed per task) └─ no → INLINE mode (sequential, single context) Tasks tightly coupled with shared evolving state? → INLINE, or go re-split the plan ``` **Universal rules (both modes):** - Never start on main/master without explicit user consent — branch first (the rule itself lives in keel-workflow; this is a reference, not a redefinition). - **Controller context check before ORCHESTRATED mode:** past ~100k tokens, write a handoff and fork to a fresh session first. The controller keeps accumulating across the whole stage and will hit mid-stage compaction — the exact failure the ledger exists to survive, not one to walk into. - **Pipeline artifacts are not code.**