Improvement Gate#

7-layer mechanical quality gate: any required layer fail = reject/revert.

When to Use#

• 验证已执行的候选是否应保留（gate.py）
• 管理人工审核队列（review.py --list）
• 完成待审批项（review.py --complete）

When NOT to Use#

• 给候选打分 → use improvement-discriminator
• 执行文件变更 → use improvement-executor
• 评估 skill 结构 → use improvement-learner

7-Layer Gate#

LayerGateRequiredPass Condition0SchemaGateYesCandidate has id, category, risk_level, execution_plan1CompileGateYesModified .py files pass py_compile; non-Python files auto-pass2LintGateNo (advisory)No lines >120 chars, no mixed tabs/spaces in diff3RegressionGateYesEvaluator verdict != "reject" (checks evaluator_evidence)4ReviewGateYesDiscriminator recommendation=accept AND panel not DISPUTED AND LLM judge != reject5DoubtGateYesCandidate text has < threshold hedging words (threshold varies by category: docs=2, prompt=4, code=3)6HumanReviewGateNo (advisory)Flags needs_human=true for medium/high risk or prompt/workflow/tests/code categories

Gate execution: layers run sequentially. First required-layer failure stops execution and triggers revert (if file was modified) or reject.

CLI — gate.py#

bash

python3 scripts/gate.py \
  --ranking ranking.json \         # REQUIRED: ranking artifact from discriminator
  --execution execution.json \     # REQUIRED: execution artifact from executor
  --state-root /path/to/state \    # default: lib/state_machine.DEFAULT_STATE_ROOT
  --evaluation eval.json \         # optional: evaluator artifact (forwarded by orchestrator)
  --layers schema,compile,review \ # optional: run only these layers (default: all 7)
  --output receipt.json            # default: auto-generated path

ParamDefaultWhen to change--layersall 7Use schema,compile for fast structural checks only--evaluationNoneOrchestrator passes this automatically when evaluator ran

CLI — review.py (Human Review Queue)#

bash

# List pending human reviews
python3 scripts/review.py --state-root /path/to/state --list

# Complete a review
python3 scripts/review.py --state-root /path/to/state \
  --complete review-cand-01-docs \
  --decision approve \             # approve | reject
  --reason "低风险文档变更，LGTM" \
  --reviewer engineer-name         # default: cli-user

4-Way Decision Logic (after all required layers pass)#

ConditionDecisionActionaccept_for_execution + low-risk docs/reference/guardrail + successkeepFile stays modifiedrecommendation=rejectrevertRestore backup, append to veto logrecommendation=hold OR non-auto-keep-eligiblepending_promoteRestore backup, create review requestexecution.status=unsupportedrejectNo file change, log reason

如果 HumanReviewGate 标记 needs_human=true，即使 keep eligible 也会升级为 pending_promote。

正确: gate 返回 pending_promote → 查审批队列 → 人工批准 $ python3 scripts/review.py --state-root ./state --list → ID: review-cand-01-docs Category: docs Risk: low Since: 2025-01-15T10:00:00Z $ python3 scripts/review.py --state-root ./state --complete review-cand-01-docs --decision approve --reason "confirmed safe" → Review review-cand-01-docs completed: approve

错误: gate 返回 revert 后仍然保留文件变更 → revert 时 gate 自动调用 restore_backup()，原文件已恢复。不要手动跳过。

Output — Gate Receipt#

json

{"decision": "keep", "reason": "low-risk docs candidate executed successfully",
 "gate_layers": {"all_passed": true, "layers_run": 7, "layer_results": [...]},
 "rollback": {"attempted": false}, "next_step": "propose_candidates", "next_owner": "proposer"}

Related Skills#

• improvement-discriminator: ReviewGate checks its panel consensus + LLM verdict
• improvement-executor: Gate validates executor output; reverts via rollback_pointer
• improvement-evaluator: RegressionGate checks evaluator verdict when --evaluation provided
• improvement-orchestrator: Calls gate as stage 5, forwards evaluator artifact
• benchmark-store: Pareto front data consumed by RegressionGate