sandhiveio/sandhive-cli / skills/sandhive
sandhive
VerifiedUse the SandHive CLI to turn verified project updates or news into X posts, find relevant conversations, and draft replies in the author's voice from original writing samples. Use for SandHive setup and drafting; publication stays manual.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 18.7 KB |
| agents/openai.yaml | config | 712 B |
| references/cli.md | markdown | 12.9 KB |
| references/draft-review.md | markdown | 4.1 KB |
| references/first-run.md | markdown | 3.6 KB |
| references/network-access.md | markdown | 9.3 KB |
| references/opportunity-triage.md | markdown | 5.3 KB |
| references/post-workflow.md | markdown | 5.3 KB |
| references/publication-registry.md | markdown | 3.9 KB |
| references/scheduling.md | markdown | 1.4 KB |
references/network-access.md
View on GitHub# Network access and task scope
Read before the first live request or when diagnosing a blocked request. Establish the user's intended mode and authorized scope, announce the selected API work, then proceed without repeated permission questions within that scope. Host permissions independently control whether a command may run or reach a destination.
## User authorization and local simulation
Live authorization must come from the user's request or earlier explicit authorization in the conversation. The website and installer provide a prompt saying "I authorize" with destination, selected data and request limits. It becomes a user instruction when the person reviews and sends it; merely reading README, installing a skill, printing that prompt, emitting an agent notice or saving a consent flag does not authorize a transfer.
Reuse existing authorization for the same task, destination, data categories and limits. If live submission is not authorized, prepare concrete inputs or a `--dry-run` locally and ask one compact question naming `https://api.sandhive.io`, the selected account/samples/facts and the planned operations and count. Ask again only for an actual scope change or a host-required approval, not routine continuation. Authorization does not override host or organization policy, and a revised prompt does not retroactively approve a denied action.
An installation-only, simulation, dry-run or local-only request makes no live SandHive service requests, including search and style lookup. Use local previews or clearly marked synthetic fixtures for simulation; never present fixture conversations as discovered posts or fixture responses as real generated drafts. Switching to live execution needs user authorization. Installing the requested CLI from GitHub is separate from sending project data to SandHive.
## Destinations and data
All live service commands send HTTPS POST requests to `https://api.sandhive.io`:
| Command | Path | Data sent |
| --- | --- | --- |
| `find` | `/cli/search-score-tweets` | Search queries, project-specific ICP description, result limits and filters. |
| `style` | `/cli/user-twitter-style` | X handle, refresh preference and sample limit. |
| `draft reply` | `/cli/generate-tweet` | Conversation text, X handle, selected original writing samples; product, audience, selected facts and voice preferences when context is supplied. |
| `draft post` with manifest or server manifest | `/cli/generate-news-twitter-post` | X handle, human sample texts, optional manifest, language and length preferences. Without a manifest, the backend uses its account manifest. |
| `draft post` with text | `/cli/rewrite-twitter-post` | Supplied post material, X handle, human sample texts and optional length preference. |
Payloads also include fast-mode settings. Every SandHive API request includes `X-SandHive-Installation-Id: shf1_<sha256>`, computed in memory from hostname, home-directory path, platform and architecture with a SandHive namespace. No identity file, hardware probe or subprocess is used. Raw inputs are not sent. Host/profile changes can change the fingerprint, and cloned environments can share it. This is not a credential or an idempotency key. GitHub update checks and other API origins do not receive it. Computing it makes no network request and needs no write permission. Old UUID files and `SANDHIVE_HOME` are ignored.
Sample source metadata is validated locally; only sample texts are sent as writing examples. Project files are not uploaded automatically. Requests may be retained; include only material suitable for the hosted service. Search, style retrieval and generation may cause backend scraping/model usage. There is no server-enforced spending limit in the CLI.
Interactive live commands may also make a best-effort GET to `api.github.com/repos/sandhiveio/sandhive-cli/compare/<installed-commit>...main` to check updates. This sends the installation commit identifier, not project content. Set `SANDHIVE_NO_UPDATE_CHECK=1` to disable it. JSON runs, non-interactive runs, previews, and local commands (`init`, `skill install`, help/version and planned commands) skip this check. Git installation and updates contact GitHub separately; they do not authorize SandHive data submission.
`--dry-run` is local: it shows the exact endpoint, method, headers (including fingerprint) and wire payload without sending it. `doctor` and `skill path` are also read-only and local. `init` and the optional `skill install` project-copy command write only their requested files. Bootstrap by reading the bundled skill, not by copying it into protected project metadata folders. These commands do not prove that live network access works. Draft commands load the current project's saved context/style by default; inspect previews or use `--no-profile` when needed. The CLI does not publish to X.
## First-run notice and execution
Adapt this example to the actual task and the user's language:
> I selected [project] and built the voice profile from [available original writing]. Within your authorization, I will send [actual request count] HTTPS requests to api.sandhive.io: [style retrieval if needed], [one rewrite or news generation then rewrite], one search using [two or three specific phrases], and up to three reply requests for suitable conversations. I will send only the selected facts, writing samples, brief/conversation, queries, account handle and installation fingerprint. The service may retain these inputs and use scraping/models. I will show drafts and source links for review.
Adapt the notice to the actual plan: a work update uses one rewrite request; a news post uses generation followed by rewrite (two calls). Each reply uses one request by default. Multiple different conversations can run concurrently with `draft replies --concurrency <1-5>`; `--attempts` only increases per-item generation attempts when explicitly chosen. State the number of items, concurrency and maximum request count; each request can cause usage. Network/timeout/HTTP failures are not retried. Start immediately after the notice when the user has authorized live execution and this data transfer. The supplied first-run prompt limits the batch to seven API requests: one optional style lookup when local samples are insufficient, one search with up to 20 results, one post (one or two requests) and up to three ready replies. With local samples and a work update, at most five are needed. Stop rather than exceed a narrower user limit. These are agent workflow limits, not server-enforced spending controls. No scheduling or publishing follows from a drafting task. Local-only mode prepares context and previews; it does not silently substitute agent-generated drafts for SandHive results.
Choose the active or explicitly named project, collect local original writing and derive search phrases before asking for manual input. Do not require the user to certify that the samples are human-written. Ask only for a genuinely missing account, project choice or usable writing after accessible sources are exhausted. Do not ask whether to continue after saving context, retrieving style or listing search results.
The installer prints an explicit authorization prompt for the person to review and send, requesting setup and a first live batch. It cannot open a host approval dialog, force an agent turn or grant permission by copying skill files. An agent already carrying out authorized live setup continues in the same conversation. A request solely to install files remains installation-only. If the user asks for simulation, per-request preview or no API use, honor it.
## When access fails
- **Command process could not start:** errors such as `setup refresh had errors` occur before the CLI runs. Report this as a host startup failure, not an API response or authorization refusal. On affected Windows hosts, a sandbox-created `.agents` directory can have an owner/ACL that prevents the sandbox refresh from applying protection. Stop creating metadata copies, preserve existing files and direct recovery to the host's supported sandbox diagnostics. Do not promise automatic recovery or change Windows permissions yourself. Claim zero API requests only when the execution history shows no live command was launched; an ambiguous network failure has different delivery uncertainty.
- **Host approval denied:** name the exact action and summarize the stated denial. Follow the host's supported approval or user-override flow if available. Continue local preparation; do not switch tools, encode the payload, or use another route to perform the denied transfer.
- **Sandbox/network restriction:** request only the needed host-supported permission for the concrete command and destination. A task request or API notice does not reconfigure sandboxing or organization policy. Do not ask for unrestricted external requests or disable protections.
- **CLI `NETWORK_ERROR` or `TIMEOUT`:** the request may already have reached the server. Report that uncertainty; do not blame ChatGPT without host evidence or retry automatically.
- **HTTP/API failure:** report the actual CLI code/status. A 429 or server error is a service response, not evidence of a host denial.
For current host behavior, see official OpenAI documentation on [sandboxing](https://learn.chatgpt.com/docs/sandboxing) and [Auto-review](https://learn.chatgpt.com/docs/sandboxing/auto-review). Other agents have their own approval controls. The skill cannot promise permanent permissions or prevent all future approval requests.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |