Latest Agent Skills

Watch one or more GitHub repositories for new issues, pull requests, comments, PR reviews, and inline diff comments — emits a fresh turn each time something actionable happens. Use when you've opened PRs / issues mimir should know about review activity for, when you want operator-actionable signals for repos mimir maintains, or when you need to track upstream activity on a watched repo. Filters out events authored by mimir's own GitHub identity (set ``MIMIR_GITHUB_SELF_LOGIN``) so the bot doesn't wake itself with its own comments. Companion to the ``pollers`` framework skill (mechanics) and the ``world-scanning`` skill (catalog of *what's worth polling*). Opt-in: copy this directory into ``<home>/skills/github-poller/`` then set the env vars below.

Watch a Gmail inbox for new messages via the `gog` Google Workspace CLI — emits one turn per new message ID since the last poll. Cursor is a SET of message IDs (not a timestamp), so reordered or backdated deliveries don't double-emit. Opt-in: copy this directory into `<home>/skills/gmail-poller/`, install `gog`, run the OAuth setup once, set `GOG_ACCOUNT`, then `reload_pollers`. Companion to the `pollers` framework skill and the `world-scanning` skill (catalog of what's worth polling). For a sender-allowlist / label-filter, use a Gmail search query in `MIMIR_GMAIL_QUERY` — the Gmail search language IS the filter mechanism, this poller doesn't reimplement it.

Bluesky + X social loop. The bundled notifications poller runs `social-cli sync` on cron (default `*/15`), parses the per-platform `inbox-<platform>.yaml` files, and wakes the agent in batches of up to 3 never-seen notifications per turn. The optional feed poller runs `social-cli feed` every 2h for timeline scanning. Agent reads inbox, writes `outbox-<platform>.yaml`, runs `social-cli dispatch`. Also supports one-shot commands (post/reply/thread/like). Opt-in: install the skill, drop `.env` credentials into `<home>/state/pollers/social-cli-notifications/`. Companion to the `pollers` framework skill and the `world-scanning` skill.

Discover, evaluate, install, and wrap external agent skills from ClawHub registry, skillflag-compliant CLI tools, and GitHub repos. Use when asked to find new capabilities, install a skill, browse what's available, or package a local skill for sharing.

Run shell commands in the background with output capture and completion callbacks. Use when a command might take more than ~30 seconds (builds, tests, deployments, data processing) and you want to keep working while it runs.

Comprehensive guide for creating software diagrams using Mermaid syntax. Use when users need to create, visualize, or document software through diagrams including class diagrams (domain modeling, object-oriented design), sequence diagrams (application flows, API interactions, code execution), flowcharts (processes, algorithms, user journeys), entity relationship diagrams (database schemas), C4 architecture diagrams (system context, containers, components), state diagrams, git graphs, pie charts, gantt charts, or any other diagram type. Triggers include requests to "diagram", "visualize", "model", "map out", "show the flow", or when explaining system architecture, database design, code structure, or user/application flows.

View image or file attachments that you can't directly see. Use this skill when you receive a message with attachments listed as file paths and you need to understand their contents. Especially useful for text-only models that cannot process images natively.

Look up who an alias belongs to (people) or what a channel id is (channels) using state/identities.yaml. Use when the operator or a turn-context refers to a platform-prefixed id ("who is discord-100000000000000001?", "what channel is slack-C100?") or you need to resolve aliases the other direction (canonical → display name + aliases). Read-side only — the populator (mimir.identities_populator) maintains the registry.

Layer alternative channels / scrapers / data sources with explicit fall-through, where each rung covers a different failure mode. Use when interacting with the world (sending a notification, fetching from a third party, detecting a state change, authenticating) and the single canonical path could fail in known ways. Distinct from retrying — retries handle transient failures; fallback chains handle modal ones (channel down, API gone, cookie expired). Three layers tops; the terminal rung is loud.

Recognize when you're in a runaway loop and stop yourself. Use when the same tool call has repeated 3× with the same args, when the same error class has fired 3× in a turn, when you've been told the same correction twice, when "let me try one more variation" is the third small variation, or when a refinement loop is orbiting rather than converging. The right move on a trip is to switch from action to inquiry — read events.jsonl, list assumptions, look for contradictions, ask the human, file a chainlink — not to retry harder.

A catalog of pollers worth building — the underused half of being a long-running agent. Use when something you care about could change without anyone telling you (a CI pipeline, a competitor's release, a friend's status, a config drift, a dependency CVE) and you want to be the first to notice. Companion to the `pollers` skill, which covers mechanics. This file is the menu of *what's worth polling.*

Check pinned external tool versions against npm and GitHub releases. Use for the weekly heartbeat tool-pin drift scan.

Create or update reusable skills for this agent. Use this skill ONLY when the user asks to create a new skill, edit an existing skill, improve a SKILL.md, or capture a repeated workflow as a reusable skill. Do not use this skill for one-off tasks.

Diagnose your own behavior by reading the structured logs you leave behind — turns.jsonl, events.jsonl, chat_history.jsonl, scheduler.yaml. Use when something has gone wrong (a message didn't land, a scheduled job isn't firing, a communication pattern feels off), when you need to understand a pattern over many turns, or when cost / token usage needs auditing. Covers jq query recipes and points at the debugging-jobs / debugging-communication / debugging-drift companion guides for specific failure modes.

Use when a bounded textual artifact (prompt, rubric, tool description, extraction instruction) keeps underperforming and success can be measured with an evaluator, dataset, or trace set. GEPA proposes evaluator-backed candidate rewrites through a normal PR/proposal adoption gate. Do not use for vague behavior changes, governance/persona/core-memory edits, fake metrics, or problems whose first honest task is defining the evaluator or collecting data.

Use when making a forward-looking claim with a checkable outcome (reply within 24h, error rate will drop, this skill will see more use) — record to state/predictions.jsonl with a review horizon so reflection can grade you later. Closes the in-the-moment double-loop.

Discover what bundled skills exist in this mimir install and what each one does. Use when you're not sure if a capability is already covered by a skill ("can I X?", "is there a skill for Y?", "what skills do I have?"). Reads SKILL.md frontmatter from each skill directory and surfaces names + descriptions.

How to read, resolve, and reason about commitments — durable records of future obligations (your own promises and the operator's requests). Use whenever the `## Upcoming commitments` prompt block surfaces something you might act on, or when you want to inspect what's pending beyond what the block shows.

Turn "block until condition X is met" into an agent wake-up via the bash_async tool. Use when you need to wait for a one-shot event without burning context — a webhook arriving, a CI pipeline finishing, a file appearing, a long shell command completing — AND you want this conversation to resume with full context when the event fires. Distinct from pollers (recurring) and the synchronous Bash tool (blocks the turn). The completion event lands as a fresh turn on the spawning channel; no polling needed.

The menu of non-grit moves available when you'd otherwise reach for "I'll do better next time." Use when you catch yourself resolving to "be more careful," "remember to X," or "try harder" — those resolutions die at the turn boundary. The right shape is always a structural change with a verifiable diff: edit a memory file, edit a skill that misled you, log a chainlink interest item, add a poller for the missed signal, file a five-whys. The artifact must outlive the turn.

Guide for the first days with a new human. Use when: the `init` memory block exists (pointing you here), you have no persona/communication blocks yet, or a human explicitly asks about setup. This skill is about having good conversations — not filling out forms.

Structured root cause analysis for arriving at a concrete action. Use when something went wrong, a pattern keeps recurring, behavior has drifted, or you catch yourself resolving to "do better" / "remember to X" without a concrete artifact. Five-whys forces behavioral resolutions into file edits, config changes, memory block updates, or scheduled jobs — the action item must produce a diff someone else can verify. Do not use for simple debugging with an obvious cause.

Local CLI issue tracker for todos, follow-ups, structured records, and the canonical pattern for decomposing multi-step work into a parent + subissues with acceptance criteria, dependency edges, and priority. Use when the user mentions a task to remember, a bug to track, an open question, anything that needs to outlive the current turn, or when planning multi-heartbeat work that needs `chainlink issue ready`-driven pickup across sessions. Includes guidance on writing descriptions future-you can act on (handles, success path, failure path) and idempotency tactics for actions that might fire twice (boundary firing, overlapping heartbeats, retried subagent completion). Pairs with five-whys (chainlink is the storage backend for RCA trees). Storage is local — issues live under a `.chainlink/` directory in the operator's repo.

Get current conditions and 5-day forecast for a city via OpenWeatherMap. Requires `OPENWEATHER_API_KEY` in the environment. Use when the user asks about weather or when planning anything where weather is load-bearing (outdoor events, travel, watering schedules).

Mechanics for building and managing pollers — subprocess scripts that check external services on a schedule and emit events when something has changed. Use when authoring a new poller (a `pollers.json` manifest plus a script in any language), debugging why a poller isn't firing, or extending an existing one. Pollers run on cron, emit JSONL events when there's something to report, and stay silent otherwise (silence-as-filter). The framework discovers `<home>/skills/<name>/pollers.json` files at startup and via `reload_pollers`; each emitted event becomes a fresh turn on a `poller:<name>` synthetic channel. Companion to the `world-scanning` skill, which catalogs *what's worth polling*. Distinct from `async-tasks` (one-shot wake-up via bash_async, not recurring) and from in-process scheduler callables (saga-consolidate, oauth-usage-poll — those mutate mimir-internal state and aren't subprocess-isolated).

Review a pull request and POST the review to GitHub via `gh pr review`. Use when asked to review a PR, or when the github-activity poller fires a PR-opened / PR-review-requested event. Always submits after drafting — a review that stays in turn output and never reaches GitHub is a non-review.

Criteria for deciding when, where and how to remember information. Use this skill whenever you might need to update a core memory block, file something under memory/, or write to a state file. It is CRITICAL that you use this skill.

Interact with GitHub via the `gh` CLI. Use for issues (read, comment, file), pull requests (view, comment, check CI status), workflow runs, and `gh api` for queries the subcommands don't cover. Authentication is via the bot's GITHUB_TOKEN (already wired through `gh auth setup-git` at container start).

Use when a high-priority signal needs to reach the operator but doesn't fit the current conversation (off-hours wake, urgent error needing eyes, scheduled-tick discovery worth surfacing now). Routes via MIMIR_OPERATOR_ALERT_CHANNEL.

Maintain a structured wiki under state/wiki/ — ingest raw sources from state/raw/, synthesize cross-linked pages, and lint for health. Use this skill whenever you need to build durable, graph-shaped knowledge with cross-references (entities with relationships, topics that recur, concepts you trace across many sources).

Remote-control tmux sessions for interactive CLIs by sending keystrokes and scraping pane output. Use when you need an interactive TTY (REPLs, agents that prompt) or want to run multiple long-lived processes in parallel and poll their state. For non-interactive long-running jobs prefer the long-running-jobs skill.

Send a phone-push notification to the operator via ntfy.sh. Use ONLY for genuine algedonic alarms — events the operator needs to know about within minutes, not hours. The push lands on their phone; the cost of misuse is operator desensitization (they mute the topic and the channel goes dark). Do not use for routine surfacing, status updates, or things that can wait for the next chat turn.

韩红（歌手 / 公益）认知与表达框架（压缩蒸馏）：高音情感爆发、救灾叙事、直球正义人设 触发：天路、慈善 等。非替基金会官方发言

樊登（讲书人）认知与表达框架（压缩蒸馏）：拆书三段论、行动号召、温和说服… 触发：樊登读书 等。非替代原著阅读；非心理咨询

丁磊（互联网创业）认知与表达框架（压缩蒸馏）：佛系老板、游戏与音乐品味人设 触发：网易、养猪梗 等。非替网易发言

崔健（摇滚）认知与表达框架（压缩蒸馏）：地下与现场感、隐喻与直白并置… 触发：一块红布 等。非煽动违法

《甄嬛传》甄嬛的剧作人格：隐忍进阶、潜台词与双层听者、借刀与以退为进、礼制典故壳； 古装权谋剧对白与人物弧参考，虚构，禁止现实陷害教唆、禁止职场 PUA。 触发：甄嬛传、小主、臣妾、潜台词、宫廷话术等。

张飞（历史剧虚构）认知与表达框架（压缩蒸馏）：莽撞勇猛、敬君子慢小人 触发：三国演义 等。虚构；忌可模仿暴力

袁朗（军旅剧虚构）认知与表达框架（压缩蒸馏）：老 A 冷血人设、试探与打磨、精英残酷浪漫 触发：士兵突击 等。虚构

许三多（军旅剧虚构）认知与表达框架（压缩蒸馏）：钝感坚持、不抛弃不放弃、草根成长 触发：士兵突击 等。禁止军事违法教程

佟年（都市剧虚构）认知与表达框架（压缩蒸馏）：甜妹直球、学霸恋爱、软萌坚持 触发：亲爱的热爱的 等。虚构

唐晶（都市剧虚构）认知与表达框架（压缩蒸馏）：闺蜜边界、职场精英冷感、理性分手 触发：我的前半生 等。虚构

苏明玉（都市剧虚构）认知与表达框架（压缩蒸馏）：原生家庭边界、职场硬壳与软芯… 触发：都挺好 等。虚构

吕秀才（情景喜剧虚构）认知与表达框架（压缩蒸馏）：子曰嘴炮、读书人迂阔与意外高光 触发：武林外传 等。虚构

罗子君（都市剧虚构）认知与表达框架（压缩蒸馏）：全职太太 reboot、成长线与争议依赖 触发：我的前半生 等。虚构

凌不疑（古装虚构）认知与表达框架（压缩蒸馏）：复仇将军、疯批克制、家族血债 触发：星汉灿烂 等。虚构；禁止暴力教唆

林冲（古典虚构）认知与表达框架（压缩蒸馏）：隐忍到爆发、体制内小人物悲剧… 触发：水浒传 等。虚构；非现实报复教程

李达康（廉政虚构）认知与表达框架（压缩蒸馏）：GDP 执念、窗口服务梗、政绩与人情张力 触发：人民的名义 等。虚构

靖王（古装权谋虚构）认知与表达框架（压缩蒸馏）：耿直宗室、信任成本、笨办法正义 触发：琅琊榜 等。虚构

贾宝玉（古典虚构）认知与表达框架（压缩蒸馏）：情不情、反仕途经济、女儿崇拜叙事… 触发：红楼梦 等。虚构；非性别刻板侮辱

吴京（演员 / 导演）认知与表达框架（压缩蒸馏）：主旋律英雄壳、直男效能叙事 触发：战狼、流浪地球 等。非煽动仇恨

朴树（音乐人）认知与表达框架（压缩蒸馏）：内向真诚、少话多留白、生命与创作一体… 触发：平凡之路 等。不消费抑郁；非医疗建议

那英（歌手 / 综艺）认知与表达框架（压缩蒸馏）：东北大姐直球、舞台 Diva、综艺金句 触发：好声音、直给导师 等。忌侮辱参与者

马东（综艺制作人）认知与表达框架（压缩蒸馏）：赛制叙事、金句剪枝、辩论产品化 触发：奇葩说、乐夏 等。非操纵舆论教程

雷军式创业与产品方法论：专注/极致/口碑/快、 三大铁律（技术为本、性价比为纲、最酷的产品）、复盘思维与长期主义、 从互联网效率到硬核科技的公开叙事。触发：小米方法论、爆品、用户口碑、创始人表达等。

韩寒（作家车手）认知与表达框架（压缩蒸馏）：反套路叙事、冷幽默、公共发言锋利… 触发：三重门赛车 等。不伪造赛事实；尊重他人

古龙（武侠作家）认知与表达框架（压缩蒸馏）：短句蒙太奇、浪子孤独、酒友女人 触发：小李飞刀 等。已故；非搏击教程

巩俐（表演作者）认知与表达框架（压缩蒸馏）：身体与气场、沉默戏、角色距离感… 触发：红高粱霸王别姬 等。非替本人发言

范伟（演员 / 喜剧）认知与表达框架（压缩蒸馏）：木讷外壳下的精算与善良、慢节奏包袱 触发：马大帅、耳朵大有福 等。角色与本人区分

蔡澜（美食 / 生活家）认知与表达框架（压缩蒸馏）：享乐主义正当化、旅行搭子叙事 触发：食神专栏 等。非过量饮酒医疗建议