菜单
Record curated MAID gaming or implementation-drift incidents as chosen/rejected examples for future review, hooks, or DPO datasets. Use when the user asks to log a MAID incident, capture a model gaming pattern, save a chosen/rejected pair, or prepare DPO training material.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Review an implementation produced from an approved MAID manifest. Confirms changed files stay within manifest scope, declared artifacts exist, validations pass, and the behavior matches the contract. Use after maid-implementer or after any MAID-backed code change before merge.
Implement code against an approved MAID manifest. Loads only the declared files, writes code to pass behavioral tests, validates with `maid validate --mode implementation`, and iterates until all checks pass. Use after a manifest is approved by maid-planner (or manually).
Plan a coding task as a machine-checkable MAID manifest instead of free-form markdown. Analyzes the project, asks clarifying questions, drafts or refines a manifest, and for implementation-ready plans adds behavioral tests, validates with `maid validate --mode behavioral`, and gets user approval. The promoted manifest becomes the implementation contract. Use at the start of every new feature, bug fix, or refactor.
Use when implementing maid-runner MAID draft manifests from manifests/drafts, promoting them into manifests, and validating implementation batches with review before handoff in Claude Code.
Review an implementation produced from an approved MAID manifest. Confirms changed files stay within manifest scope, declared artifacts exist, validations pass, and the behavior matches the contract. Use after maid-implementer or after any MAID-backed code change before merge.
Implement code against an approved MAID manifest. Loads only the declared files, writes code to pass behavioral tests, validates with `maid validate --mode implementation`, and iterates until all checks pass. Use after a manifest is approved by maid-planner (or manually).
| name | maid-incident-logger |
| description | Record curated MAID gaming or implementation-drift incidents as chosen/rejected examples for future review, hooks, or DPO datasets. Use when the user asks to log a MAID incident, capture a model gaming pattern, save a chosen/rejected pair, or prepare DPO training material. |
Create a human-curated incident record through MAID's deterministic incident commands. A human or agent decides that an event was a gaming attempt; MAID records and retrieves the supplied evidence deterministically. No inference, classification models, or hidden summarization are involved.
Incident records are YAML files under the repository-local .maid/incidents/ directory. Do not hand-write incident files, invent metadata fields, or document any unshipped behavior. The skill guidance must add no new CLI behavior.
Use only the shipped command surface:
maid incident capture --manifest <path> --packet <path> --rejected-diff <path> --tags <comma-list> [--notes <text>]
maid incident update <incident-path> --chosen-diff <path>
maid incident list [--tag <tag>] [--json]
maid incident export --format dpo --output <path>
maid incident suggest-temptations --paths <comma-list> [--json]
When a contract-integrity gate catches a gaming attempt, capture the incident immediately with the failure packet and rejected diff:
maid incident capture --manifest <path> --packet <path> --rejected-diff <path> --tags <comma-list> [--notes <text>]
Then update the same incident after the honest fix lands by attaching the chosen diff:
maid incident update <incident-path> --chosen-diff <path>
Use maid incident list [--tag <tag>] [--json] to inspect stored records. Use maid incident export --format dpo --output <path> only after records have chosen diffs. Use maid incident suggest-temptations --paths <comma-list> [--json] to surface deterministic advisory temptation entries for future manifests.
Use the closed pattern-tag vocabulary exactly:
test-weakeningtrivial-teststub-implementationcontract-renegotiationscope-escaperunner-gamingfalse-doneUnknown tags are rejected as a usage error. Vocabulary changes require a manifest evolution.
This skill has a documentation-only cross-link to the 066 recall workflow. During planning, related completed Outcomes and incidents may be surfaced with:
maid recall --for-manifest <path>
maid recall --for-manifest <path> --plan-packet
Recall results and incident suggestions are advisory planning evidence. They do not replace the current manifest's behavioral tests, declared scope, validation, or implementation review, and the incident and recall implementations stay independent.
Mark an incident as DPO-worthy in notes or downstream review judgment only when:
Do not treat every mistake as DPO-worthy. When the model lacked necessary context, the task was ambiguous, or the issue was a one-off bug, capture only the deterministic evidence needed for review.