Open Forge

open-forge — self-host any open-source app

name: open-forge description: Automate self-hosting of open-source apps on cloud infrastructure the user owns. Use when the user asks to "self-host", "deploy to my own cloud", "install X on AWS / Lightsail / EC2 / Azure / Hetzner / DigitalOcean / GCP / Oracle Cloud / Hostinger / Raspberry Pi / Kubernetes / Fly.io / Render / Railway / Northflank / exe.dev", "set up my own Ghost blog / Mastodon / WordPress / Nextcloud", wants to deploy the self-hosted personal AI agent OpenClaw (openclaw.ai — NOT the Captain Claw platformer game) or Hermes-Agent (Nous Research's self-improving AI agent at github.com/NousResearch/hermes-agent), wants to run Ollama (local-LLM inference server at ollama.com — pairs with every AI agent / chat UI as an OpenAI-compatible provider), wants to run Open WebUI (feature-rich self-hosted ChatGPT-like UI at github.com/open-webui/open-webui — pairs natively with Ollama and any OpenAI-compatible backend; adds RAG, web search, image gen, voice, multi-user), wants to run Stable Diffusion WebUI / Automatic1111 / A1111 (the most-popular open-source AI image generator at github.com/AUTOMATIC1111/stable-diffusion-webui — text-to-image, img2img, inpainting, ControlNet, LoRA; pairs with Open WebUI as an image-gen backend), wants to run ComfyUI (node-based AI image / video generation at github.com/comfyanonymous/ComfyUI — power-user alternative to A1111 with workflow graphs; same models, different UX; pairs with Open WebUI as image-gen backend), wants to deploy Dify (open-source LLMOps + AI app builder at github.com/langgenius/dify — visual workflow builder, RAG, multi-tenant; the "build a SaaS-grade AI app" platform, different category from chat UIs), wants to deploy LibreChat (multi-provider chat UI with deep enterprise plumbing at github.com/danny-avila/LibreChat — alternative to Open WebUI for teams; multi-user with social logins, per-user balance + transactions, agents + MCP, dedicated rag_api), wants to deploy AnythingLLM (RAG-focused workspace + agent platform at github.com/Mintplex-Labs/anything-llm — drop-in PDFs + URLs + GitHub repos, ask questions over them; built-in LanceDB; Desktop App + Docker + 8 cloud one-clicks), wants to install Aider (AI pair-programming CLI at github.com/Aider-AI/aider — runs in the terminal next to a git repo, edits files via diffs, auto-commits; pairs with any LLM provider including Ollama for local), wants to deploy vLLM (production-grade LLM inference server at github.com/vllm-project/vllm — high-throughput multi-tenant serving with PagedAttention + tensor parallelism + prefix caching; NVIDIA / AMD / Intel / CPU; Docker / Kubernetes / Helm / PaaS), wants to deploy Langfuse (open-source LLM engineering platform at github.com/langfuse/langfuse — observability, evals, prompt management, datasets, scoring; v3 six-service architecture with Postgres + ClickHouse + Redis + S3; Docker Compose, Kubernetes Helm chart, first-party Terraform modules for AWS / GCP / Azure, Railway one-click), or names any combination of an open-source app and a cloud provider. Walks the user through provisioning, DNS, TLS, outbound email (SMTP), and inbound email, in phases that are resumable across sessions via a state file at ~/.open-forge/deployments/<name>.yaml. Supported today: Ghost on AWS Lightsail (Bitnami blueprint); OpenClaw via every upstream-blessed path documented at docs.openclaw.ai/install/* — AWS Lightsail blueprint, Docker Compose, Podman, Kubernetes (Kustomize), native installers (install.sh / install-cli.sh / install.ps1), ClawDock, Ansible, Nix, Bun, plus per-host adapters for AWS EC2 / Azure / Hetzner / DigitalOcean / GCP / Oracle Cloud / Hostinger / Raspberry Pi / macOS-VM (Lume) / BYO Linux server / localhost / Fly.io / Render / Railway / Northflank / exe.dev. More projects and infras added under references/projects/ and references/infra/.

open-forge

Overview

Walk a user from "I have a cloud account and a domain" to "working app at https://my.domain with TLS and mail." Load the appropriate project recipe and infra adapter based on the user's stated intent; run phases sequentially; record state so the user can resume later.

Platform note: this skill is designed for Claude Code but the content is platform-agnostic. Tool names like AskUserQuestion, WebFetch, and mcp__github__* are Claude Code-specific — read them as capabilities (structured-choice prompt, URL fetch, GitHub API) and use whichever equivalent your platform exposes. See docs/platforms/ in the repo for per-platform integration guides (Codex / Cursor / Aider / Continue / generic).

Operating principle

Claude does the work; the user makes the choices. open-forge replaces the traditional "read a README, copy-paste 30 lines of bash, debug for hours" experience with a guided chat where Claude executes everything via the user's local CLI tools (aws, ssh, jq, curl) and only stops to ask when input is genuinely required.

What this means in practice:

• Run, don't print. When a recipe contains a bash block, Claude executes it. Announce it in one sentence first ("Opening port 22 in the Lightsail firewall now."), then run. Don't paste the block into chat for the user to run.
• Ask for choices and credentials only. Things only the user can decide or provide: AWS profile name, domain choice, canonical www-vs-apex, SMTP API key, model provider preference. Everything else (which jq command to run, which sed pattern to apply, which IAM script URL to fetch) Claude figures out from the recipe.
• One question at a time when possible. Use a structured-choice prompt for multiple-choice / single-select (Claude Code: AskUserQuestion; on other platforms, ask in prose with options listed). Reserve free-text questions for things like API keys and domain names. Avoid wall-of-questions forms.
• Auto-install with confirmation, not silently. If jq or aws is missing, propose the install command, get one-line approval, then run it. Never sudo apt-get install without asking.
• The recipe files in references/projects/ and references/infra/ are guidance for Claude, not pages for the user to read. Keep that lens when extending or refactoring.

What's supported

Check references/projects/ and references/infra/ for available recipes/adapters. As of this writing:

Supported software:

Supported infras (under references/infra/):

Supported runtimes (under references/runtimes/):

Selection — ask three questions

Before provisioning, establish three things by asking (or inferring from the user's prompt):

1. What to host? → loads references/projects/<software>.md
2. Where to host? → loads references/infra/<cloud>/<service>.md or references/infra/{byo-vps,localhost}.md
3. How to host? → loads the matching references/runtimes/<runtime>.md (skipped if the infra bundles the runtime, e.g. vendor blueprints)

The how question is dynamically generated from (software, where) — each project lists its "Compatible combos" table in the project recipe, and the options shown are filtered by the user's where answer. If the user's initial prompt already names a clear infra ("deploy to Lightsail" → AWS), announce the inferred choice and continue — don't re-ask. Ask a structured-choice question only when genuinely ambiguous.

Then immediately load references/modules/preflight.md and run its steps. Preflight is combo-aware — it only installs / validates what the chosen tuple actually needs (AWS CLI only when infra ∈ AWS, Docker only when runtime = docker, nothing extra on localhost).

Goal-shaped requests → curated bundles

If the user describes a goal rather than a single piece of software (e.g. "set up an AI homelab", "I want a privacy stack for my home network"), check references/bundles/ for a matching curated bundle before falling through to single-software routing. Bundles are recipe-of-recipes that pair commonly-co-deployed apps with cross-software wiring already worked out.

Single-software requests still go through the standard 3-question selection. Bundles are an additional entry point for goal-shaped intents.

Tier 1 vs Tier 2 routing

open-forge ships a finite catalogue of verified recipes (Tier 1) plus a documented fallback for the long tail (Tier 2). When the user names a piece of software, decide which tier you're in before loading anything.

Tier 1 — verified recipe exists

If references/projects/<name>.md matches the user's software, you're in Tier 1. Load it, follow it, and stay in the standard workflow below.

Tier 2 — no recipe; derive from upstream live

If no recipe matches, don't refuse — fall back to Tier 2:

1. Announce in one sentence: "This software isn't in our verified recipe set — I'll fetch upstream docs live and reuse the runtime / infra modules. Treat my output as best-effort, not authoritative."
2. Fetch upstream the same way Tier 1 does:
   • Fetch the upstream README first via the platform's URL-fetch capability (Claude Code: WebFetch; Cursor: @Web; Aider/generic: curl via shell). If 403/404, fall back to raw.githubusercontent.com/<org>/<repo>/<branch>/README.md, or git clone the docs repo locally if the docs site is Cloudflare-protected.
   • Locate the upstream install-method index (docs site, repo docs/install/ tree, wiki).
   • Enumerate every method documented under that index. Do not invent methods upstream doesn't ship — if fetches fail, stop and tell the user, don't speculate.
   • Read canonical install artifacts in the repo (Dockerfile, docker-compose.yml, helm/, flake.nix, primary config example).
3. Reuse the existing modules: drive the Docker install via runtimes/docker.md, Kubernetes via runtimes/kubernetes.md, VM provisioning via infra/<cloud>/*.md, DNS / TLS / SMTP via references/modules/. The Tier 2 work is only the software-specific bits on top.
4. Cite every upstream URL in chat the same way Tier 1 sections do (> Source: <url>).
5. Offer to capture the result as a new Tier 1 recipe once the deploy succeeds — that's how the catalogue grows. Captured recipes must go through first-run discipline before promotion.

Quality boundary: Tier 2 output is best-effort, not authoritative. It will hallucinate at the edges of upstream docs we couldn't fetch and skips the real-deploy refinement Tier 1 recipes get. Always tell the user which tier you're in; never silently mix.

Out-of-scope software

Some user requests are not deployable services at all (libraries like Unsloth or requests, desktop apps like Slack, SaaS like Notion). When you detect this, say so clearly and offer the closest in-scope alternative if there is one. See CLAUDE.md § Is this software in scope? for criteria.

Phased workflow

Each phase is verifiable and resumable. Do NOT batch phases — complete, verify, and update state before moving on.

1. preflight     → check prerequisites (CLI tools, profiles, domain ownership); collect inputs
2. provision     → create instance, allocate + attach static IP, retrieve SSH key
3. dns           → print exact DNS records for user to add at registrar; poll until resolved
4. tls           → obtain Let's Encrypt cert, fix reverse proxy, switch app URL to https
5. smtp          → configure outbound email provider; verify a test send
6. inbound       → (optional) set up forwarding or mailbox
7. hardening     → rotate default admin creds, rotate any secrets pasted into chat

Infra adapter defines how to do each phase (what CLI commands to run). Project recipe defines what's specific about that app (config file paths, gotchas, mail block shape). Cross-cutting steps — DNS guidance, Let's Encrypt, SMTP providers, inbound forwarders — live in references/modules/ and are loaded as needed.

State file

Every deployment has a YAML state file at:

~/.open-forge/deployments/<name>.yaml

Shape:

name: my-blog
project: ghost
infra: lightsail
inputs:
  aws_profile: qi-experiment
  aws_region: us-east-1
  domain: ariazhang.org
  canonical: www  # or "apex"
  letsencrypt_email: user@example.com
outputs:
  instance_name: my-blog
  static_ip_name: my-blog-ip
  public_ip: 54.156.69.42
  ssh_key_path: ~/.ssh/lightsail-default.pem
  admin_url: https://www.ariazhang.org/ghost
phases:
  preflight:   { status: done,  at: "2026-04-22T19:00Z" }
  provision:   { status: done,  at: "2026-04-22T19:10Z" }
  dns:         { status: done,  at: "2026-04-22T19:25Z" }
  tls:         { status: done,  at: "2026-04-22T19:30Z" }
  smtp:        { status: done,  at: "2026-04-22T20:05Z" }
  inbound:     { status: skipped }
  hardening:   { status: pending }

At the start of each session: if a state file exists for the named deployment, read it and resume from the first non-done phase. If the user says "start over", confirm destructively before unlinking.

Execution mode

Default: autonomous — run AWS CLI, SSH, and file edits directly. Announce each external command in one sentence before running. Never fabricate outputs.

Flag: --dry-run — print what would be done, do not execute. Useful for review.

Commands that cross trust boundaries (paste secrets into config files, send real emails, spend money) should be announced and, when ambiguous, confirmed.

Inputs

Inputs split across three layers:

• Cross-cutting (all deployments) — handled by references/modules/preflight.md: AWS profile, region, deployment name, tool install confirmations.
• Infra-specific — handled by the loaded infra adapter (e.g. references/infra/lightsail.md): bundle/blueprint choice, SSH key path defaults.
• Project-specific — handled by the loaded project recipe (e.g. references/projects/ghost.md): domain, canonical preference, Let's Encrypt email, SMTP provider + API key, model provider, etc.

Each recipe and adapter has its own "Inputs to collect" section listing exactly what it needs and at which phase. Collect just-in-time per phase, not all upfront. Use a structured-choice prompt where the platform supports one (Claude Code: AskUserQuestion; otherwise prose with options listed).

Asking for credentials

Whenever the skill needs sensitive input — API keys, DB passwords, OAuth client secrets, cloud creds, SSH key paths — load the Credentials handling section below and offer the five patterns (priority order):

Never silently accept a paste. When the skill detects sensitive input is needed, it should:

1. Offer the five patterns with the credential class noted (e.g. "I need a Resend API key — pick how to provide it: file path, env var, secrets-manager ref, or paste (last resort)").
2. Validate before using:
   • File path → test -r <path> + check mode is ≤ 600 (offer chmod 600 if wider).
   • Env var → test -n "$<NAME>" (refuse if empty; if user exported after Claude Code started, ask them to restart).
   • Cloud-CLI → smoke-command (e.g. aws sts get-caller-identity --profile <name>).
   • Secrets-manager → smoke-command (op read --no-newline <ref>, vault kv get, etc.).
   • Paste → require explicit risk acknowledgement first.
3. Detect accidental pastes: if the user was prompted for a path but pasted a string matching re_* / sk-* / AKIA[0-9A-Z]{16} / etc., stop and ask: "That looks like the key itself, not a path. Did you mean to paste directly? (see risks)".
4. Never accept SSH key contents. Always ask for the path; skill uses ssh -i <path>.
5. End-of-deploy rotation reminder if the user pasted any secret during the deploy: surface during the hardening phase with a list of (credential, dashboard URL) pairs. Pasted secrets remain in session history; rotating now bounds the exposure.

Agent-mode rules (OpenClaw / Hermes / any messaging-channel agent)

When this skill runs inside a long-running personal AI agent (OpenClaw, Hermes-Agent, or any agent that talks to the user via WhatsApp / Telegram / Slack / iMessage / email / etc.), apply these stricter rules on top of the base five-pattern flow above:

• Pattern 5 (direct paste) is DISABLED. Pasting credentials into messaging channels is meaningfully riskier than into coding-tool chat — chat history syncs to the user's phone, may be backed up to cloud (iCloud / Google Drive), and often persists indefinitely. Refuse a paste with: "I can't accept credentials pasted into a messaging channel. Use a file path, env var, cloud-CLI session, or secrets-manager reference instead. See the Credentials handling section below for options." If the user insists, refuse again — don't compromise.
• Reject deploy conversations from group channels. Group chats leak everything to all members (credentials, IPs, admin URLs). When invoked from a group context, respond: "Self-host deploys involve sensitive info. Switch to a 1:1 DM and ask again." Then stop.
• Use async polling for time-elapsed waits, not blocking prompts. dns propagation, tls cert issuance, provision instance-boot — all become "I'll poll and ping you when ready" rather than "press enter when DNS propagates." Agents have a daemon; use it.
• Channel-aware response routing. Long-form content (DNS records to add at registrar, full recipe explanations, admin-bootstrap URLs) should go via secure / structured channels (email, signed note, secure-share link) when the agent supports them, not the chat. Quick decisions (yes/no, pick from list) stay in chat. Final hand-off (admin URL, rotation reminders) → secure 1:1 only.

See docs/platforms/openclaw.md and docs/platforms/hermes.md for the full agent-mode integration guides.

See the Credentials handling section below for the full pattern details, per-credential-class recommendations, and failure-mode handling.

Verification after each phase

Never mark a phase done without verification.

Post-deploy feedback (closes the catalogue evolution loop)

After hardening (or after the user explicitly says "we're done", or after they abort mid-phase and want to share what they learned), offer to file a GitHub issue with the deployment notes. Per CLAUDE.md § Issue-driven contribution model, this is how the catalogue evolves — the bot or a future Claude session reads these issues and patches the recipes.

Three flows the user can trigger from this prompt:

1. Recipe feedback (default at end of deploy) — submit gotchas, suggested edits, or "the recipe was outdated". Claude self-summarizes from the session; the user reviews + opts in.
2. Software nomination — when the user asked to deploy something not in the catalogue and Tier 2 worked, offer to nominate it for Tier 1.
3. Method proposal — when the user discovered an upstream-supported install method the recipe doesn't cover.

The flow (multi-step consent — never auto-post)

Load the Post-deploy feedback flow section below for the full sanitization rules + draft templates + submission paths. Summary:

1. Opt-in prompt:
   • Recipe feedback: "Want to share what you learned with the open-forge project? I can draft a sanitized GitHub issue with the gotchas + suggested edits — you review, then post."
   • Software nomination (Tier 2 deploy): "This software isn't in the Tier 1 catalogue yet. Want to nominate it? I'll draft an issue with the rationale + upstream URLs."
   • User must explicitly opt in (no auto-post).
2. Self-summarize the session:
   • Which recipe + combo was used, plugin version.
   • Which phases ran, which retried, which failed.
   • Where the user got prompted unexpectedly (gaps in the recipe).
   • Any gotchas Claude observed (commands that failed, error messages, deviations from the documented path).
3. Draft the issue in the format from the Post-deploy feedback flow section below:
   • Specific recipe-edit suggestions (preferred: as a diff), not free-prose.
   • All identifiers redacted per CLAUDE.md § Sanitization principles.
4. Show the redacted draft in chat — full text — before any submission attempt.
5. Standing reminder: "GitHub issues are public and permanent. Once posted, this can't be unposted. Review every line; if anything looks identifiable to you, edit before posting. By submitting, you grant a non-revocable license to use this content in the recipe; the project bears no liability for your decision to share."
6. Confirm post? — explicit "yes" required. If user edits the draft, re-show + re-confirm.
7. Submit via the first available path:
   • gh issue create --title "..." --body "..." --label recipe-feedback,recipe:<name> if the user has gh authenticated.
   • Platform-native GitHub integration if available (Claude Code: mcp__github__issue_write; Cursor / generic: GitHub MCP server if installed).
   • Fallback: print a prefilled URL (https://github.com/zhangqi444/open-forge/issues/new?template=recipe-feedback.yml&title=...&body=...) and ask the user to open + submit in browser.

Sanitization is mandatory

Per CLAUDE.md § Sanitization principles — strip every domain, IP, SSH key path, API key, AWS account ID, email address, state-file content, and anything from the user's clipboard / env vars before showing the draft. Use the patterns + replacements documented in the Post-deploy feedback flow section below.

If you find something in the draft that you can't confidently classify as safe, redact it rather than ship it. The user's review pass is a safety net, not the only line of defense.

When to skip

• User says "no thanks" or doesn't reply → drop it, don't pester.
• Deploy aborted very early (before any state was created) → no useful feedback to capture; skip.
• Tier 2 deploy that obviously wasn't in scope (e.g. user tried to "self-host" a library) → don't nominate; politely explain it's out of scope per CLAUDE.md § Is this software in scope?.

Common pitfalls across infras/projects

• Stale DNS: browsers cache 301 responses with long max-age. After any HTTP↔HTTPS or apex↔www redirect change, suggest hard reload or incognito.
• Host key mismatch on new static IP: the first SSH to a freshly-allocated IP needs -o StrictHostKeyChecking=accept-new; don't blindly blow away ~/.ssh/known_hosts entries.
• Non-interactive cert tools: some have quirky option-file or flag requirements. See the project recipe — do not assume --unattended works.
• Reverse-proxy misconfig after switching to https URL: apps that enforce HTTPS redirects from the url config need X-Forwarded-Proto and Host preserved. See references/modules/tls-letsencrypt.md.

Adding a new project or infra

A new project: add references/projects/<name>.md covering required services, config file paths, mail config shape, and any install/upgrade quirks. Follow the structure of the existing ghost.md.

A new infra: add references/infra/<name>.md covering provisioning (create instance, static IP, SSH key), firewall defaults, user/paths conventions. Follow lightsail.md.

Cross-cutting modules (new SMTP provider, new forwarder): add under references/modules/. Keep them project- and infra-agnostic.

Credentials handling (agent-mode rules apply)

name: credentials description: How the skill asks for credentials safely — five patterns prioritized from "secret never enters chat" to "last-resort paste with explicit risk acknowledgement." Loaded by SKILL.md § Asking for credentials. Applies to API keys, SSH keys, DB passwords, OAuth client secrets, cloud account creds, anything sensitive.

Credentials module — five patterns, prioritized

Pasting raw credentials into Claude Code is risky:

• The secret enters the session history (visible to other tools loaded in the same session, may persist in logs).
• May be relayed via MCP servers depending on the user's setup.
• Shows up in transcripts the user might later share for support.
• Some terminals / IDEs persist input across restarts.

The skill defaults to safer patterns. Direct chat paste is last resort and only after explicit risk acknowledgement.

Hard rule: every time the skill needs a sensitive input, it offers the user the five patterns below — letting them pick — and surfaces the risk if they pick paste. Don't silently accept a paste; don't pretend Claude Code is a vault.

The five patterns (priority order)

1. Local file path (recommended for personal use)

User stores the secret in a file under their home directory; tells the skill the path; skill reads via cat.

When to suggest first: for one-off API keys (Resend, SendGrid, Mailgun, OpenAI, Anthropic, etc.) that the user already has in a .env, .secrets, or password-manager export.

Skill prompt:

"Path to a file containing the key (e.g. ~/.secrets/resend)? I'll read it via cat."

Skill execution:

RESEND_KEY=$(cat ~/.secrets/resend)   # or however the user names it
# Use $RESEND_KEY in subsequent commands; never echo it back to the user

Properties:

• Secret never enters chat.
• File survives across Claude Code sessions; user can use the same path next time.
• User is responsible for the file's permissions (chmod 600 recommended; mention if the file's mode is 644 or wider).

2. Environment variable name (recommended for shell users)

User exports the secret as an env var before starting Claude Code (or in their shell rc); tells the skill the var name.

When to suggest first: when the user already has secrets in a .envrc / .bashrc / ~/.config/fish/config.fish they source regularly.

Skill prompt:

"Name of an env var holding the key (e.g. RESEND_API_KEY)? I'll read $RESEND_API_KEY from my shell."

Skill execution:

# Verify the var exists in Claude's shell
test -n "$RESEND_API_KEY" || { echo "RESEND_API_KEY not set; export it before continuing"; exit 1; }
# Use it
curl ... -H "Authorization: Bearer $RESEND_API_KEY" ...

Properties:

• Secret never enters chat.
• Session-scoped if exported in the current shell only; persistent if in rc files.
• The env var must exist in the shell Claude Code launched from. If the user exports after Claude Code starts, Claude won't see it (you'll need them to restart Claude Code or pass it inline).

3. Cloud-CLI session auth (default for AWS / GCP / Azure / GitHub)

User authenticates the cloud CLI ahead of time (e.g. aws sso login, gcloud auth application-default login, az login, gh auth login); skill uses the resulting profile / session.

When to suggest first: any time the credential is for a cloud account that ships its own CLI auth flow. Don't ask for raw cloud access keys if SSO / browser auth is available.

Skill prompt:

"Have you run aws sso login for the profile you want to use? If yes, what's the profile name?"

Properties:

• No secret material in chat or in any file the skill reads.
• Auth is browser-mediated, MFA-friendly.
• Sessions expire (good — bounded blast radius); skill handles re-auth gracefully if the session lapses mid-deploy.

4. Secrets-manager reference (advanced)

User stores secrets in 1Password / Bitwarden / Vault / AWS Secrets Manager / GCP Secret Manager; gives the skill a CLI-resolvable reference; skill calls the secret-manager CLI to fetch only when needed.

When to suggest first: when the user mentions they "have it in 1Password" or similar; or for users with proper secret-management practices.

Skill prompt:

"1Password / Bitwarden / Vault reference? I'll fetch via the matching CLI when I need it."

Properties:

• Secret never enters chat or any persistent file.
• Resolved just-in-time; not cached in shell vars longer than necessary.
• User must have the matching CLI installed + authenticated.

5. Direct chat paste (last resort — risk acknowledgement required)

User types the secret directly into chat. Skill must surface the risks before accepting.

When this happens: user explicitly says they want to paste, or none of patterns 1-4 work for their situation (e.g. they're trying out the skill with a one-shot key and don't want to set up file storage).

Required risk acknowledgement (paraphrase, don't elide):

"⚠️ If you paste the key here, it will live in this Claude Code session's history. It may also be visible to other tools loaded in the session and could appear in any transcripts you share later for support. After this deploy completes, I'll remind you to rotate the key in the provider's dashboard. Still want to paste? (yes / pick a safer path)"

If user confirms:

• Accept the paste.
• Use the value immediately; don't echo it back.
• At the end of the deploy, surface a reminder: "You pasted <provider> API key into chat earlier. Rotate it in <provider's dashboard URL> now that the deploy is complete."

Properties:

• Convenient but contaminates session history.
• The rotation reminder is mandatory — without it, the user may forget the key is exposed.

Per-credential-class recommendations

Different credential types pair best with different patterns. Surface the recommendation when the credential class is known.

…(truncated)