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
| 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).
~/.config/opencode/templates/subagent-response.yml.tmpl~/.config/opencode/contracts/subagent-contract-block.md (parent-injected)~/.config/opencode/rules/subagent-response-protocol.md (alwaysApply: true)~/.config/opencode/hooks/subagent-protocol-{inject,lint}.sh (inject = pre-hook contract injection; lint = pre-commit drift check)<project>/.opencode/docs/runbooks/subagent-protocol.md (author here; optional org template ~/.config/opencode/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.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.