What problem does it solve? Implementation subagents often spin without producing output: they read dozens of files without writing code, conflict with parallel agents on the same files, burn millions of tokens, and force the orchestrator into endless send_message催促 loops. This Skill replaces supervision-style management with contract-based governance so subagent work boundaries, read permissions, write ownership, and exit conditions are engineered upfront. ## Core Features & Use Cases - Contract Snapshots with Negative Space: Every implementation subagent dispatch requires a contract defining GOAL, OWN files, MAY READ whitelist, OUT OF SCOPE, MUST NOT, EXIT criteria, and BLOCKED conditions. - Three-Level Bounded Reading & Action Gates: Level 0 auto-allowed reads, Level 1 budgeted reads with reason codes, Level 2 escalation; subagents must produce a mutation or a structured BLOCKED before budgets exhaust—no "keep researching" state. - Single-Writer Ownership & Structured Result Recycling: File-level conflict checks before parallel dispatch, plus compact DONE/FAILED reports (status, changed files, validation, artifact refs) that keep subagent work history out of the orchestrator context to protect prefix caching. - Audit & Routing Tooling: Ships session-discipline-audit.js and splice-ab-experiment.js for post-hoc session discipline checks, plus workflow_preflight_router.py for logical-role model routing, pre-flight validation, and explicit fallback policy. - Use Case: Before dispatching three parallel coding subagents, generate contract snapshots with disjoint OWN file sets, set tool-call budgets, and require structured BLOCKED reports—then audit the session afterward to verify no催收 loops or cache-miss spikes occurred. ## Quick Start Ask the agent to dispatch an implementation subagent using the subagent-execution-governance contract snapshot template with bounded reads, budgets, and structured BLOCKED reporting.
