本文へ移動
cccskills
無料GitHub で公開

omh-external-connector-readiness

[omh] Evaluating a plugin, connector, or API for adoption: external connector readiness - assess whether a named plugin, connector, API, data provider, or multimodal route is safe, affordable, fresh, and observable; use toolbelt-readiness for missing capability inventory. Use when the user says: external-connector-readiness, external connector readiness, connector readiness matrix, plugin readiness matrix, provider readiness, api readiness, connector adoption, external plugin adoption.

インストール方法を見る

含まれるファイル(2)

  • SKILL.md16.1 KB
  • references/memory-provider-trial.md3.3 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

External Connector Readiness

This is an OMH external-connector-readiness workflow skill, projected for Agent Skills hosts (Claude Code, Codex, Cursor, opencode, OpenClaw, pi).

Why This Exists

external-connector-readiness exists because adopting a plugin, connector, API, or memory provider changes what the agent can reach, spend, and remember, and a host's install-time scan or a vendor's docs answer only part of that: it scores capability, auth, cost, freshness, safety, and fallback for each candidate, and reports a trial only from what was observed.

Do Not Use When

  • Choosing which coding owner runs a task.
  • Finding which tool, credential, or install is missing for work already chosen; use omh-toolbelt-readiness.
  • A read-only live lookup or an external write through a connector already adopted; use omh-live-info or omh-apps.

Examples

Good example:

  • Prompt: external-connector-readiness compare weather plugin and wxtrain candidates with cost, freshness, multimodal evidence, and fallback routes before adoption.
  • Expected behavior: Produce prepare_external_connector_readiness with required context, wrapper actions, and not-evidence boundaries.
  • Why: The prompt names a real workflow surface that Hermes can orchestrate without hiding execution.

Bad example:

  • Prompt: external-connector-readiness silently enable a paid connector and claim weather, SQL, and screenshot results without observed provider evidence.
  • Expected behavior: Report the missing observed evidence or authority instead of claiming the external step happened.
  • Why: Prepared OMH guidance is not platform, runtime, connector, file, memory, or delivery evidence.

Completion Checklist

  • Candidate connector, target domain, read/write scope, modality needs, provider owner, fallback workflow, and stop condition are explicit.
  • Cost, quota, credential, permission, live-data freshness, multimodal capture, safety, and compliance boundaries are marked ready, missing, risky, or not_observed.
  • Route live read-only lookups to live-info-operator, external writes to connector-operator, datasets/SQL to data-analysis, and missing tools to toolbelt-readiness before claiming results.
  • Provider responses, screenshots, audio/video/file captures, query outputs, message ids, and external mutations are reported only from observed trial evidence.
  • For a memory provider, disabling it, removing its local cache, deleting its remote memory, deleting the account, and switching away are reported as distinct operations with distinct postconditions.
  • Unknown deletion, isolation, or write semantics block automatic writes and irreversible adoption instead of resolving to an OMH or Hermes default.
  • For a realtime voice connector, keep turn integrity, latency, fallback, interruption, and spoken tool safety separate; report unsupported or unobserved dimensions as hold or block rather than as a pass.
  • On Hermes, record its install-time scans as observed evidence beside the card, never in place of it: hermes plugins install scans a plugin before installing it (safe, caution, or dangerous, and a dangerous verdict is refused even with --force), and an MCP server started through npx, uvx, or pipx is checked against OSV malware advisories and refused on a match with an error that reads MCP server '<name>': BLOCKED: .... A scan counts only when its output was seen: the plugin scan does not run with plugins.scan_on_install off, and the OSV check covers known malware only and lets the server start when its lookup fails, so a quiet install or launch is not a clean result.

Recovery Notes

  • If the candidate list is unknown, route to skill-scout or source-finder before readiness scoring.
  • If credentials, cost authority, or connector installation is missing, keep readiness blocked and route setup to toolbelt-readiness.
  • If a specific provider action is already selected, route read-only live data to live-info-operator or write/mutation tasks to connector-operator.
  • If a memory-provider lifecycle field is unknown, ask the operator to declare it or supply an observed trial receipt; keep the result as not_omh_reviewed context rather than importing it into OMH review.
  • If a realtime voice connector has no supplied realtime_voice_trial_receipt/v1, route the setup and the trial run to the host, connector, or operator that owns the microphone, call, or room, then consume only the returned receipt.

When to Use

Use before adopting, enabling, or routing an external plugin/connector/API when Hermes must compare capability, auth, cost, modality, freshness, safety, fallback, and observable trial evidence. Use it for an optional memory provider too, where enabling, switching, pausing, or removing it also needs identity scope, automatic hooks, retention, deletion, export, and switching answered before adoption.

Strong routing signals: `external-connector-readiness`, `external connector readiness`, `connector readiness matrix`, `plugin readiness matrix`, `provider readiness`, `api readiness`, `connector adoption`, `external plugin adoption`, `weather plugin readiness`, `weather connector readiness`, `wxtrain readiness`, `onequery read-only sql`, `read-only sql connector`, `sql connector readiness`, `nextcloud connector`, `microsoft workspace connector`, `microsoft graph connector`, `chainlink connector`, `solana connector`, `xmr gateway`, `private crypto transaction`, `private cryptocurrency connector`, `crypto transaction plugin`, `blockchain gateway`, `composio connector`, `composio universal cli`, `universal cli connector`, `universal cli skill adoption`, `skill connector adoption`, `connector auth risk`, `connector cost auth risk`, `agentchat connector`, `peer-to-peer agent messaging connector`, `websocket identity connector`, `websocket connector trial`, `clawsocial connector`, `social discovery connector`, `windy pairing`, `windymail mailbox connector`, `matrix chat identity`, `antigravity cli connector`, `agy cli bridge`, `agy bridge connector`, `macos keychain oauth connector`, `oracle oci connector`, `oracle genai connector`, `miniverse bridge`, `crustocean platform connector`, `cost-aware connector`, `multimodal connector`, `multimodal routing`, `screenshot connector`, `audio connector`, `video connector`, `video generation`, `generate a video`, `product demo video`, `text to video`, `home assistant connector`, `home assistant integration`, `home assistant device control`, `home assistant smart home`, `smart home connector`, `device control connector`, `plugin auto-routing`, `connector auto-routing`, `external tool trial`, `memory provider readiness`, `memory provider posture`, `memory provider lifecycle`, `memory provider adoption`, `memory provider retention`, `memory provider portability`, `memory provider sync failure`, `compare memory providers`, `memory provider comparison`, `memory provider`, `memory provider trial`, `memory provider benchmark`, `memory provider migration`, `memory provider rollback`, `switch memory provider`, `switching memory providers`, `disable memory provider`, `delete provider memory`, `export memory provider data`, `realtime voice connector`, `real-time voice connector`, `realtime voice readiness`, `realtime voice trial`, `realtime voice stack`, `voice connector readiness`, `voice connector trial`, `voice gateway readiness`, `voice gateway trial`, `voice agent connector readiness`, `voice trial receipt`, `voice turn integrity`, `voice turn receipt`, `barge-in behavior`, `barge-in handling`, `voice tool safety`, `spoken tool safety`, `커넥터 준비도`, `외부 커넥터 준비`, `외부 플러그인 채택`, `플러그인 준비도`, `커넥터 도입`, `플러그인 도입`, `비용 인증 리스크`, `인증 리스크`, `도입 비용`, `비용 기준 커넥터`, `자동 라우팅`, `멀티모달 커넥터`, `멀티모달 라우팅`, `영상 생성`, `제품 데모 영상`, `홈 어시스턴트 커넥터`, `홈 어시스턴트 연동`, `홈 어시스턴트 기기 제어`, `홈 어시스턴트 스마트홈`, `홈어시스턴트 커넥터`, `홈어시스턴트 연동`, `홈어시스턴트 기기 제어`, `홈어시스턴트 스마트홈`, `스마트홈 커넥터`

Catalog Metadata

Category: connector Phase: connector-readiness Quality tier: workflow-surface-gated Reasoning demand: standard

Quality bar:

  • Name the user-facing workflow objective, required context, next action, and stop condition.
  • Separate prepared guidance from observed platform, runtime, connector, file, memory, or delivery evidence.
  • Expose missing tools, credentials, targets, or observations as user-visible gaps.
  • Trial two or more memory providers on one corpus and one query set, never on each provider's own sample — load references/memory-provider-trial.md for the corpus, query, and reversibility rules.

Required inputs:

  • user request
  • target context
  • delivery or status expectation
  • known missing evidence

Expected outputs:

  • external_connector_readiness_card/v1
  • connector_capability_matrix/v1
  • auth_cost_boundary/v1
  • live_data_freshness_policy/v1 when live data is required
  • multimodal_routing_policy/v1 when screenshots, audio, video, or files are involved
  • fallback_route_policy/v1
  • connector_trial_manifest/v1 when observed
  • memory_provider_posture/v1 when the candidate is an optional memory provider
  • memory_provider_trial_comparison/v1 when two or more memory providers are trialled on one corpus
  • realtime_voice_trial_receipt/v1 when a supplied realtime voice trial is observed
  • realtime_voice_readiness/v1 verdict per voice dimension when a receipt is supplied
  • next action
  • prepared-vs-observed boundary

Artifact expectations:

  • external_connector_readiness_card/v1 metadata-only wrapper card when prepared
  • connector_capability_matrix/v1 with candidate, domain, read/write shape, modality, owner workflow, and fallback route
  • auth_cost_boundary/v1 separating missing connector, missing credentials, paid/provider cost risk, quota, and user authority
  • live_data_freshness_policy/v1 for requested recency, provider timestamp, stale-result handling, and source-quality thresholds
  • multimodal_routing_policy/v1 for screenshot, audio, video, file, OCR, or visual QA evidence routes when needed
  • connector_trial_manifest/v1 only when a provider response, capture id, query transcript, message id, or tool-call observation is recorded
  • memory_provider_posture/v1 for an optional memory provider, covering identity scope, automatic hooks, storage boundary, synchronization, failure, retention, deletion, export/import, backup/restore, and portability, each marked ready, missing, risky, not_observed, or unknown
  • memory_provider_trial_comparison/v1 built on the connector_trial_manifest/v1 pattern: one corpus and one query set per candidate, retrieval quality, latency, and cost each reported per provider or marked not_observed, plus the migration route and the rollback that was exercised rather than asserted
  • realtime_voice_trial_receipt/v1 only when an authorized host, connector, or operator supplies the observed trial: connector build identity, requested versus observed stack, per-turn milestones on one declared timing reference, turn-integrity states, fallback path, interruption behavior, and spoken tool decisions
  • realtime_voice_readiness/v1 with a pass, hold, or block state and reasons for turn integrity, latency, fallback, interruption, and tool safety, plus the separated connector-configured, synthetic-trial, actual-environment-trial, voice-turn, tool-action, and session-completed states

Safety rules:

  • An external connector readiness card is not connector installation, credential validation, provider access, API invocation, multimodal capture, live-data retrieval, external mutation, cost authorization, or successful trial evidence unless observed connector-trial evidence records it.
  • Do not claim connector, gateway, runtime, file generation, memory mutation, or host automation evidence from prepared guidance.
  • A hook registration or lifecycle callback is an observation point only; it never grants write or synchronization authority.
  • Documentation and open package code declare a contract; neither establishes a hosted service's storage, retention, cost, or deletion postcondition.
  • A memory provider stays optional: an unknown or unavailable provider never becomes a required default.
  • On Hermes, the memory provider is one slot: Hermes runs its built-in memory and no more than one external provider, the one memory.provider names, so adopting another provider (honcho or mem0, for example) where that key reads omh displaces OMH's provider. OMH's per-turn recall and dreaming reminders then stop, and the omh_memory tool stays registered but refuses to recall or write records (session_lens_unavailable), because the session lens it reads comes from the provider. State that displacement in the posture before adoption, and keep the switch a proposed memory.provider diff the user applies.
  • A realtime voice verdict comes only from a supplied realtime_voice_trial_receipt/v1. OMH opens no microphone, call, room, socket, or provider session, installs no connector, downloads no voice model, and authorizes no tool from a receipt.
  • A synthetic voice fixture never proves the intended room, microphone, network, or telephony path, and a fallback path succeeding is never success for the requested voice stack, provider, or model.
  • Generic connector, voice-input, and media-input records are not realtime voice readiness; a turn that lost its onset, split, merged, dispatched twice, truncated, or replayed after audible output blocks the verdict rather than reporting latency.
  • A memory-provider trial reports retrieval quality, latency, and cost per candidate only from host-supplied observation; a dimension nobody measured is not_observed, never a default or a zero, and a corpus is never copied into the record.
  • A rollback is proven by exercising it during the trial and recording what came back; a documented rollback path is prepared, not reversible.

Runtime Evidence

Use the current host's own tools and subagent/task mechanism when available; otherwise run the same lanes sequentially or name the unavailable capability. A prepared plan, handoff, checklist, or skill installation is not execution, review, CI, merge-readiness, or merge evidence. Record actual tool results, or not_observed / not_available, in the record; never invent dispatch or host accounting. Treat supplied context as advisory, not proof of hidden memory reads or writes. State scope, constraints, verification, and the stop condition before work. Reply in the user's own words and the host's own voice: the host's persona owns reply language, tone, speech level, and sentence endings, progress updates included (where it sets no language, use the one the user wrote in), and OMH shapes structure and content only; OMH's record terms (surface, lane, wrapper, handoff, evidence boundary, not_observed) stay in records and tool calls, never in the sentence the user reads unless they ask about one; and when a stop condition or a decision the user owns ends the turn, offer the next action as a question rather than declaring what will not be done. Supporting paths are relative to this skill directory; sibling skill paths are relative to its parent. Resolve them from the host-provided skill base directory ({baseDir} on hosts that provide it), never a hardcoded install location. A named workflow not installed here is unavailable, not permission to emulate its host-specific capabilities. Verify through the real surface before done.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Onboard a new model generation or sibling into oh-my-hermes: probe router recognition, research the official contract, write trait-to-counter calibration, place routing in both lanes, price from documented list only, gate machine config on a served route, prove with the gates, close with a benchmark pair. Use when a model family ships a new generation (for example "Fable 5.1 is out", "GLM 5.4 shipped", "onboard new model", "add model to chains").

日本語の概要は準備中です。原文の説明を表示しています。

rlaope/oh-my-hermes3,3912026年10月16日 更新

[omh] Screen-reader or keyboard accessibility gaps: prepare WCAG, keyboard, focus, screen-reader, target-size, and reflow evidence gates for UI surfaces. Use when the user says: accessibility-audit, accessibility audit, a11y audit, a11y architect, wcag audit, wcag 2.2, wcag 2.2 aa, accessibility pass.

日本語の概要は準備中です。原文の説明を表示しています。

rlaope/oh-my-hermes3,3912026年10月16日 更新

[omh] Hermes badges unlocked and achievement progress: achievements observation: summarize hermes-achievements badges, tiers, recent unlocks, and progress from local plugin artifacts. Use when the user says: achievements, achievement, badges, badge, my badges, show achievements, achievement summary, unlocked badges.

日本語の概要は準備中です。原文の説明を表示しています。

rlaope/oh-my-hermes3,3912026年10月16日 更新

[omh] Technical proposal facing adversarial scrutiny: independent perspectives attack a proposal, then distill into a bundle a separate planner consumes. Use when the user says: adversarial-consensus, adversarial planning, adversarial plan review, red team this plan, red-team this plan, red team the proposal, multi-perspective review, multiple perspectives.

日本語の概要は準備中です。原文の説明を表示しています。

rlaope/oh-my-hermes3,3912026年10月16日 更新

[omh] Coordinating several agents or profiles: coordinate multiple Hermes profiles or agents with task, handoff, heartbeat, blocker, and completion states. Use when the user says: agent-board, agent board, kanban, multi-agent, multi agent, multi agent board, multiple hermes agents, multiple hermes profiles.

日本語の概要は準備中です。原文の説明を表示しています。

rlaope/oh-my-hermes3,3912026年10月16日 更新

[omh] Agent is stuck, looping, or drifting: capture a stuck, looping, drifting, or repeatedly failing agent run, diagnose the likely failure pattern, and prepare the smallest safe recovery action. Use when the user says: agent-debug, agent debug, agent debugging, agent introspection, agent self-debug, self-debug, self debugging, looping agent.

日本語の概要は準備中です。原文の説明を表示しています。

rlaope/oh-my-hermes3,3912026年10月16日 更新

rlaope のスキルをすべて見る

このスキルの問題を報告する