菜单
The Converge stage as a usable skill. Synthesizes recent memory/notes into one grounded insight, anchors every forward-looking claim in live web evidence (Σ₀ external-reality rule), and appends an evidence-bearing Convergence Record. Backs the `!convergance` chat command.
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
Cut a full Unisona release — fold the changelog, bump the version, tag it, let CI build and publish the website + the native Windows desktop app (Unisona-Setup .exe), then rewrite the GitHub Release post for humans with the live URL, the .exe download, a features list, and current pricing. Use whenever the user types `/release` or `!release`, or asks to "cut a release", "ship a version", "release the website and exe", "tag a release", "publish a new version", "make a GitHub release", "do a release post", or "bump and release". Trigger even if they don't say "release" — any request to tag a version and publish a GitHub release with the app + download is this skill. Do NOT use it to hot-deploy master to GCE without a version (that's the deploy runbook), to merge one PR, or to only bump the changelog.
Build and use the single flat Lantern OS RAG dollhouse from local repos, GitHub repo metadata, literal COMET LEAP PDFs, 30-day art images, charts, convergence manifests, and future intake rules. Use when Codex needs a flat RAG-ready file, literal PDF/image assets in a skill, cross-repo Lantern/COMET LEAP consolidation, or a mass-clone/intake plan.
One universal, Σ₀-grounded research capability — fan-out web search + local arXiv-corpus grounding (post-cutoff AI papers, cited by id) + on-demand full-paper fetch, run as persisted, resumable rounds (Observe→Reason→Verify→Converge). Powers chat `!research`, plain-language "research/investigate X", `!convergance` grounding, and autowork issue research. Use whenever the user wants something researched, looked into, grounded in current sources, or wants the actual text of a paper.
Review past benchmark runs and plan the next ones for the local model (Ouro / Σ₀ coder) — per-benchmark leaderboard tables placing the local stack against web-validated public SOTA on HumanEval, MBPP, SWE-bench, LongMemEval and honesty/HaluEval evals, plus a prioritized "run next" plan. Use whenever the user types `/benchmarks` or `!benchmarks`, or asks to "check the benchmarks", "how's the local model doing on evals", "plan the next benchmark runs", "where do we stand on HumanEval/SWE-bench/HaluEval", "compare Ouro to the council", "what should we benchmark next", or "benchmark the local model". Trigger even when the user names only one benchmark (e.g. "check our SWE-bench number") — reviewing one mark in the context of the whole ledger is this skill. Do NOT use it to grade the whole app (that's report-card), to review a PR diff (that's code-review), or to actually train/serve a model.
Groom the Keystone OS GitHub backlog to the agile Definition-of-Ready bar — route every open issue to a dev lane (kriskin / mookman / alex), set its milestone, set its priority (p0/p1/p2), fill its type, flag thin issues needs-triage, and set its assignee, then report what changed. Use this whenever the user types `/refinement`, or asks to "refine the backlog", "groom the backlog", "triage the issues", "prioritize the backlog", "assign lanes", "label the backlog by lane", "set milestones on the issues", "assign the issues to kriskin/mookman/alex", "sort the backlog into lanes", or otherwise wants open issues distributed across the three contributor lanes with labels, milestones, priorities, and assignees. Trigger even when the user doesn't say "refinement" — any request to organize/distribute/triage/prioritize the open GitHub issues is this skill. Do NOT use it to actually write code for an issue (that's autowork), to grade the app (that's report-card), or to review a single PR diff (that's code-review).
Produce an honest, evidence-grounded letter-grade report card for Keystone OS as it actually is right now — graded dimensions, real evidence per row, an overall grade, and a candid one-paragraph summary. Use this whenever the user types `!report-card`, or asks to "grade the app", "evaluate the app as it is now", "how are we doing", "give me an honest assessment", "where do we stand", "what's the state of the project", or otherwise wants a frank, no-flattery scorecard of the system's current quality. Trigger even when the user doesn't say the words "report card" — any request for a candid overall grade of the app or project is this skill. Do NOT use it for grading a single PR diff (that's a code review) or for a roadmap/planning ask.
| name | convergence |
| description | The Converge stage as a usable skill. Synthesizes recent memory/notes into one grounded insight, anchors every forward-looking claim in live web evidence (Σ₀ external-reality rule), and appends an evidence-bearing Convergence Record. Backs the `!convergance` chat command. |
Status: active — backs the live !convergance chat command
Loop stage: Converge (and Verify, via mandatory grounding)
Source: apps/lantern-garage/lib/dream-chat.js (handleConvergenceCommand, _deriveConvergenceQuery, _appendConvergenceRecord); records in data/convergence/records.jsonl
The Converge stage of the one loop, exposed as a command. Given recent dream/note entries it produces one coherent insight about patterns and direction of travel — but, per the Σ₀ external-reality rule, it does not bluff: any forward-looking suggestion is grounded in a live web search and cited, and the result is written back as a Convergence Record with its evidence, sources, and an honest confidence.
!convergance — synthesize the dreamer's recent entries into a single grounded observation. A query is derived from the salient themes of those entries (_deriveConvergenceQuery, deterministic, no extra LLM call), the live web is searched (webSearch: MCP → keyless DuckDuckGo/Wikipedia fallback), and the sources are injected so the synthesis cites [n] references rather than inventing direction.!convergance <topic> — same, but ground the synthesis on an explicit topic instead of the auto-derived themes.!convergance log an issue <title> — file a GitHub issue from the loop (shell-free via safeExec, so the title can't escape to a shell).data/convergence/records.jsonl:
{ hypothesis, evidence:[entries…, source URLs…], sources, grounded, grounding_query, result, confidence, verified, loop_stage:"Converge", tags }.grounded:true, verified:true, and earns higher confidence. When search fails or returns nothing, the synthesis degrades honestly to grounded:false, verified:false, confidence capped low, tag ungrounded — the record never overstates an un-anchored claim.sources holds the URLs.web-search-client grounding helpers.Proven: the grounding helpers (webSearch, formatGroundingContext) are the same battle-tested path the main chat tool loop uses (#1212).
Held: end-to-end UI verification requires a running server + a served local model + network; the synthesis text quality depends on the active provider.
Local-only: Convergence Records are append-only on the operator's machine — they are the persistent, retrieval-based learning substrate, never used to retrain weights.
This is the canonical Converge surface. It must not grow into a separate engine: one synthesis path, one record store (data/convergence/records.jsonl + CSF archive). See docs/BACKLOG-REVIEW-STANDARDS.md for how a backlog review pass is itself a Converge turn.