菜单
Build deterministic, complete 30-day issue-level evidence and rank recurring operating-system improvements for the CEO routine.
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
Decide when CEO self-improvement should stay in additive runtime guidance versus becoming a PR against the Micronaut Agent Company source repository.
"Reference the upstream GitHub gh CLI skill only for explicit human/operator exceptions where a non-plugin GitHub client is authorized. Normal GitHub API operations must use the GitHub Sync plugin tools, including the Hermes MCP-bridged runtime names when present. Do not depend on a propagated GITHUB_TOKEN and do not search the filesystem, plugin config, or other files for a token. Any maintainer-visible non-plugin GitHub write still requires the manual GitHub-flavored Markdown footer: one blank line, `---` on its own line, then `###### \u2728 This message was AI-generated using <exact model id>`."
Search-only Skills marketplace discovery for CEO Training; inspect candidates and prepare evidence-backed approval requests without installing, updating, assigning, or executing any skill.
Shared GitHub, GitHub Sync plugin, footer, PR-linking, and review-thread protocol used by Micronaut company agents.
Diagnose, plan, implement, verify, and review Micronaut-specific GraalVM native-image work across Native Build Tools, reachability metadata, micronaut-graal, micronaut-build GraalVM support, generated AOT/native assets, and CI reports.
Shared definition of done for Micronaut planning, implementation, QA, security review, code review, and PR handoff.
| name | ceo-issue-history |
| description | Build deterministic, complete 30-day issue-level evidence and rank recurring operating-system improvements for the CEO routine. |
Use this CEO-only skill before interpreting monthly self-improvement history. Do not replace the collector with an ad hoc issue sample or raw run-log review.
Choose an explicit UTC asOf from the routine's scheduled boundary. Load the assigned issue-history evidence skill, locate its imported runtime directory, and execute the bundled collector:
node <ceo-issue-history-skill-directory>/scripts/issue-history-evidence.mjs --as-of <ISO-8601-UTC>
The script uses PAPERCLIP_API_URL, PAPERCLIP_API_KEY, and PAPERCLIP_COMPANY_ID. It enumerates the canonical company issue and agent inventories, reads durable issue resources with bounded concurrency, and reads company-wide heartbeat activity from GET /api/companies/:companyId/activity?entityType=heartbeat_run&limit=500. Run-entity events are deduplicated by event ID against issue resources.
Evidence uses exactly [asOf-30d,asOf). asOf is mandatory: never infer it from the model clock. Exit code 2 and outcome blocked_incomplete_evidence mean coverage was not provably complete. In that state, create no proposal, approval, PR, or other discovery mutation; report the sorted missing-resource ledger.
The output is canonical compact JSON capped at 32,000 UTF-8 bytes, with controlled reason codes, stable sha256: fingerprints, bounded/redacted identifiers and issue/event references, capped rejected details, aggregate canonical-agent coverage plus an inventory fingerprint, coverage metadata, and no raw evidence text, secrets, or logs. Only schema-valid prior decisions with a controlled status and valid timestamp can suppress or defer a candidate.
Eligibility is objective:
cross_issue_recurrence: at least two distinct issues and three events;concentrated_loop: at least three failed, blocked, or changes-requested events on one issue across at least two run IDs, three GitHub Sync churn events on one issue, or two stale-recovery-recursion refusals on one issue;duplicate_incident: at least two liveness escalations on one issue;recovery_fanout: at least three recovery actions on one issue;critical_one_off: one concrete governance, security, data-loss, unapproved external-write control failure, or one issue where a structured GitHub Sync regression is followed by a harness liveness escalation.GitHub Sync attribution uses exact allowlisted structured provenance namespaces (paperclip-github-plugin or github-sync in pluginKey, source plugin fields, or contextSource) as well as exact plugin-origin namespaces. Unrelated strings that merely contain github do not qualify, and caller-supplied github_sync_churn labels cannot bypass the provenance check. Duplicate event IDs count once. Prior active fingerprints are suppressed. Implemented, rejected, or no-change decisions need a fresh post-decision threshold. Ranking is severity, distinct issue count, event count, recency, then fingerprint; the cap of three is applied after deduplication.
no_change with complete coverage is a successful verified no-op. Do not manufacture routine work. For ranked_candidates, inspect only the cited issue evidence, verify the owner/target and exact change, then follow company-package-evolution for governance and implementation. Keep direct queue corrections and active productivity-review decisions separate from new package proposals.
Terminal run token and USD-cost fields are parsed independently. Missing metrics and affected totals remain unknown, never zero, while known subtotals are labeled. Never patch Paperclip source, core, or imported-package files.
The evidence pass is the discovery lane. Broad runtime-skill reconciliation and managed-repository instruction audit are separate bounded maintenance lanes; durable PR follow-through belongs to Writer or Engineer, never CEO. Load references/maintenance-lanes.md only when running this routine. Do not count maintenance outcomes as new ranked candidates.