菜单
Protocol for subagent → parent responses: single fenced YAML envelope, strict caveman-ultra for compressed fields, verbatim for paths/errors/code. Applies to every `Task` invocation. Use when spawning subagents or parsing their output.
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
Track agent decisions, metrics, token costs, and generate reports. Use when logging task execution, tracking agent performance, or generating observability reports.
Unified ai-brain skill combining memory operations and KB query operations with promotion flow.
Bootstraps a Jira hierarchy snapshot before any multi-issue create / edit session. Disambiguates the parent input (issue key, spec URL, Confluence page URL, or `greenfield`), walks parents up (depth-cap 4), walks children down via JQL recursion (depth-cap 4, max 50 children per node), and returns a structured tree alongside the issue-type scheme and link-type metadata the caller needs to render a draft. Used primarily by `atlassian-pm`; secondary consumers (after the G2.5 retrofit) are the upstream Atlassian plugin skills `spec-to-backlog` and `capture-tasks-from-meeting-notes`. The skill **does not render**; rendering belongs to the caller (typically `atlassian-pm` rendering a markdown tree to chat).
Ultra-compressed communication mode. Cuts token usage ~75% by speaking like caveman while keeping full technical accuracy. Supports intensity levels: lite, full, ultra. Defaults: main chat = lite, subagents = ultra. Use when user says "caveman mode", "talk like caveman", "use caveman", "less tokens", "be brief", or invokes /caveman. Also auto-triggers when token efficiency is requested.
Generate → execute → fail → analyze → fix → repeat cycle with failure pattern recognition. Use when implementing changes that need verification and automatic recovery.
Heuristic context budget checks using file-size and count proxies — integrates with context-cache-discipline; no tokenizer introspection claims.
| name | subagent-response-protocol |
| description | Protocol for subagent → parent responses: single fenced YAML envelope, strict caveman-ultra for compressed fields, verbatim for paths/errors/code. Applies to every `Task` invocation. Use when spawning subagents or parsing their output. |
All subagent → parent traffic is a single fenced ```yaml block matching the canonical schema. Only the envelope is emitted; nothing else belongs in the response body. Envelope MUST be the last non-whitespace content (D6).
~/.cursor/templates/subagent-response.yml.tmpl~/.cursor/contracts/subagent-contract-block.md (parent-injected)~/.cursor/rules/subagent-response-protocol.mdc (alwaysApply: true)~/.cursor/hooks/subagent-protocol-{inject,lint}.sh (inject = pre-hook contract injection; lint = pre-commit drift check)<project>/.cursor/docs/runbooks/subagent-protocol.md (author here; optional org template ~/.cursor/docs/runbooks/subagent-protocol.md when stowed)The pre-commit lint blocks drift on distinctive schema and contract phrases — the skill, rule, and agent files must xref by path instead of copy-pasting.
_marker, schema_version, reported_by[].summary, non-security findings[].note,
next_actions[], open_questions[].category matches
the security regex in the rule body §5.2, OR severity >= high, OR status ∈
{blocked, error} with security-adjacent agent. Negations stay literal.<REDACTED:TYPE> or <REF:artifacts[i]#L42> placeholders).script, iframe, img
with on*, style, object, embed, link, svg with on*, math).--- on its own line) inside multiline scalars.javascript: / data: URL schemes.WebFetch / Shell / MCP tool output — must
live in artifacts[] as file references.<REDACTED:TYPE> or a reference into artifacts[i].summary/findings[].note.Parents MUST scan every string value of an incoming envelope with
whitespace-and-separator-insensitive regex for the secret patterns listed
above, plus a Shannon-entropy ≥ 4.0 heuristic (with UUID v4 and hex-40 SHA
allowlist). On match: (a) quarantine the raw response to
~/ai-brain/projects/<name>/explore-dumps/<task-id>.md when that root exists,
overflow path is ~/ai-brain/projects/<name>/explore-dumps/<task-id>.md only; (b) rewrite the
envelope to status: malformed with
degraded_reason: "suspected_secret_in_output", and (c) record the incident
in the runbook. Treat the raw match as already leaked for incident-response
purposes and rotate the credential.
artifacts[] must encode large blobs as ASCII tokens:
<REF:path#sha256:<64 hex>#size:<bytes>> — see templates/subagent-response.yml.tmpl.templates/agent-task-spec-v1.yml.tmpl + hooks/subagent-task-antidup-preflight.sh when orchestration-policies toggle strict mode (configurations/orchestration-policies/anti-dup.yml).Structure: ok envelope, findings[] with category: structure, note in
caveman-ultra ("rename ok. circular import via session → auth."),
reported_by: ["staff-engineer"], fix verbatim.
Structure: blocked envelope with category: authn, note in full clarity
prose, secret value replaced by <REDACTED:SESSION_TOKEN> with artifact ref,
forced-normal clarity per rule §5.2.
Model writes a paragraph of explanation AFTER the closing ```. Parent parse step 1 rejects via "non-whitespace content after closing fence" and issues a single reformat-only retry.
Model writes summary: "rotate AKIA...14 chars...; fix logger". Parent parse
step 5 (fuzzy redaction) matches the AWS key pattern, quarantines the raw
response, and rewrites to status: malformed,
degraded_reason: "suspected_secret_in_output".
Task tool → the contract is injected
by the pre-hook (subagent-protocol-inject.sh); you do not need to
duplicate it.cto, code-reviewer, tech-lead) synthesizing child
output → follow the 8-step parent parse contract in the rule body (detect
→ validate → retry → stub → fuzzy-redact → strip _marker → aggregate →
synthesize). Never forward _marker or raw child YAML to the user.