Migrate from OpenClaw

Guide the user through migrating their OpenClaw installation into NanoClaw v2. This is a conversation, not a batch job. Read OpenClaw state, discuss it with the user, decide together what to bring over and where it belongs in v2's entity model, and show proposed changes before applying.

Principle: Never silently copy data. Read it, explain it, place it, then apply. Credentials are masked when displayed (first 4 + ... + last 4). Make judgment calls about what's core vs. reference material.

UX: Use AskUserQuestion for multiple-choice only. Use plain text for free-form input. Don't dump raw data — summarize and explain conversationally.

What this skill changes (conformance)

This skill drives existing NanoClaw entry points (setup/index.ts --step register, scripts/init-first-agent.ts, the onecli CLI) and copies a few files in (workspace markdown, OpenClaw skills, and its own transform module + test). It makes no code-level reach-in into core. Its integration assumptions about v2 are guarded by scripts/transform.test.ts, which is copied into the project's scripts/ test tree on apply (Phase 8) so vitest runs it against the composed install. REMOVE.md reverses every file the skill copies.

v2 architecture the migration targets

OpenClaw and NanoClaw v2 differ structurally. Keep these in mind throughout:

• Entity model. v2's central DB (data/v2.db) holds users, user_roles, agent_groups, messaging_groups, and the messaging_group_agents wiring between them. There is no store/messages.db and no scheduled_tasks table.
• Container isolation. Each agent group runs in its own Linux container. An OpenClaw "agent" maps to a v2 agent group (workspace + memory + CLAUDE.md); an OpenClaw chat/group maps to a v2 messaging group; the wiring row connects them.
• Standing instructions vs memory. Per-group role, personality, and behavior live in groups/<folder>/instructions.prepend.md. Durable facts live under groups/<folder>/memory/. The provider project document is composed at spawn and must not be edited.
• Credentials. Container-facing API credentials (Anthropic, OpenAI, …) are held in the OneCLI Agent Vault and injected per request — never in container env vars. Host-side channel tokens (Telegram/Discord/Slack bot tokens) stay in .env; the NanoClaw host process reads them to connect to the platform.
• Access control. Per messaging group unknown_sender_policy plus user_roles (owner/admin) and agent_group_members — not a JSON allowlist file.
• Scheduled tasks. A task is a messages_in row (kind='task') in a session's inbound.db, carrying a cron recurrence and a process_after timestamp. The agent creates them via its schedule_task MCP tool.

Migration State File

Create migration-state.md in the project root at the start of Phase 0. Update it after each phase. It's the single source of truth — if context is lost, re-read it to recover decisions and progress. Re-read it before starting any phase.

Sections to maintain:

• Progress — checkbox list of phases (Phase 0–8)
• Discovery — STATE_DIR, IDENTITY_NAME, channels, groups (with v2 platform_id mappings), workspace files, cron count, MCP servers
• Decisions — assistant_name, shared-vs-separate, primary owner agent
• Owner & Primary Agent — user id, role, agent group folder
• Registered Groups — table: folder, platform_id, channel, session_mode
• Credentials — table: credential, destination (vault / .env), status
• Settings Migrated — timezone, container timeout
• Identity & Memory — prepend and memory paths created for each group
• Scheduled Tasks — table: original_id, name, mapped schedule, status
• Deferred / Not Applicable — unsupported channels, OpenClaw-only features

Keep it factual and terse. Delete it at the end of Phase 8 (or offer to keep it as a record).

Phase 0: Discovery

Run the discovery script to find and summarize the OpenClaw installation:

pnpm exec tsx ${CLAUDE_SKILL_DIR}/scripts/discover-openclaw.ts

If the user specifies a custom path, pass --state-dir <path>.

Parse the status block. Key fields: STATUS, STATE_DIR, CHANNELS, WORKSPACE_FILES, DAILY_MEMORY_FILES, SKILL_COUNT, SKILLS, CRON_JOBS, MCP_SERVERS, IDENTITY_NAME, AGENT_COUNT, AGENT_IDS, GROUPS (each formatted channel:id(name)=>v2_platform_id — the right-hand value is what to pass as --platform-id to register).

Sanity-check the output. The script detects known structures but can miss data if OpenClaw's format changed. Check CONFIG_TOP_KEYS and CONFIG_CHANNEL_KEYS — if you see keys it didn't report on, read that section of the config with the Read tool. Check STATE_DIR_CONTENTS for directories it doesn't scan.

If STATUS=not_found: Tell the user no OpenClaw install was detected at the standard locations (~/.openclaw, ~/.clawdbot). Ask for a custom path; if none, exit.

If STATUS=found: Present a human-readable summary (identity name, workspace files, channels and which v2 supports, daily memory count, skills, cron count, MCP servers, agent count). Then paraphrase the key architectural differences from the section above — don't dump it as a table.

AskUserQuestion: "Ready to start migrating? I'll go through each area one at a time."

1. Yes, let's go — proceed to Phase 1
2. Tell me more — explain any area they ask about
3. Skip migration — exit

Phase 1: Agents, Groups, and Shared vs Separate

Decide this before identity/memory — it determines where files go.

OpenClaw model: all groups routed to one agent share a workspace (SOUL/MEMORY/IDENTITY) and personality; only the session is per-group.

v2 model: each agent group is a separate container with its own filesystem, standing instructions, and memory/ tree. Multiple messaging groups wired to the same agent group share that state. There is no groups/global/.

AskUserQuestion: "In OpenClaw your groups shared one personality and memory. In v2 each agent group is separate. How do you want to handle this?"

1. Shared identity (recommended if it was one bot) — apply the same core identity to each selected group's instructions.prepend.md; keep group facts in each group's memory tree.
2. Fully separate — each group gets independent memory and instructions; no shared base edit.
3. Just the primary agent for now — set one agent up; add others later.

Remember this choice for Phase 3.

Confirm the assistant name

IDENTITY_NAME from discovery is the OpenClaw name. Ask: "Your OpenClaw assistant was named <IDENTITY_NAME>. Keep it in v2?" If empty, ask them to choose (default: "Andy"). The chosen name is passed as --assistant-name to register/init.

Seed the owner and the primary DM agent

The owner identity and the primary agent are created together by scripts/init-first-agent.ts. It upserts the user, grants the owner role, creates the agent group + filesystem, wires a DM messaging group, and queues a welcome DM over the running service's CLI socket — so the service must be running. If it isn't, tell the user to start it first.

Resolve the owner's channel identity and the DM platform id (use the channel's own terminology). Then:

pnpm exec tsx scripts/init-first-agent.ts \
  --channel <channel> \
  --user-id <channel>:<handle> \
  --platform-id <channel>:<dm-id> \
  --display-name "<Owner Name>" \
  --agent-name "<confirmed assistant name>" \
  [--role owner]      # default: owner

For direct-addressable channels (telegram, whatsapp) the --platform-id is usually the same handle as --user-id with the channel prefix. --role defaults to owner (global, cross-channel) — use admin (scoped to the agent group) or member only if intended.

Register the remaining groups

For each additional OpenClaw group the user wants to bring over, register a messaging group and wire it to an agent group:

pnpm exec tsx setup/index.ts --step register -- \
  --platform-id "<v2_platform_id from discovery>" \
  --name "<group name>" \
  --folder "<channel>_<name-slug>" \
  --channel "<channel>" \
  --session-mode "<shared|agent-shared|per-thread>" \
  [--trigger "@<assistant name>"] \
  [--no-trigger-required] \
  --assistant-name "<assistant name>"

Notes:

• register namespaces the --platform-id the same way the adapter will at runtime, so pass the => value discovery emitted (or the raw OpenClaw id).
• Reuse a --folder to put a group on an existing agent (shared base/separate conversations); use a new --folder for a fully separate agent.
• Engage defaults come from the channel adapter's declaration (most group chats default to mention-based engagement; channels without a mention signal default to a name pattern). Pass --trigger to set an explicit regex, or --no-trigger-required for respond-to-everything.
• Register groups from channels v2 doesn't support yet too — the messaging group and wiring persist and activate when that channel is installed.

Folder naming: <channel>_<name-slug> (e.g. telegram_dev-team). Confirm each name and folder with the user.

Phase 2: Settings from Config

Read the config (<STATE_DIR>/openclaw.json or clawdbot.json) for settings that map to v2 setup.

Timezone

Check agents.defaults.userTimezone. If it's a valid IANA zone, write it to .env as TZ=<timezone>. v2 reads TZ from .env (src/config.ts) and uses it for cron/recurrence evaluation, so this matters for scheduled tasks.

Container timeout

Check agents.defaults.timeoutSeconds. v2's equivalent is CONTAINER_TIMEOUT (env var, default 30 min) or per-group ncl groups config update. If the OpenClaw value differs notably, note it; the user can set CONTAINER_TIMEOUT=<ms> in .env.

Access control (sender policies)

OpenClaw per-channel allowFrom / dmPolicy / groupPolicy map onto v2's model, which is not a JSON file. Each messaging group has an unknown_sender_policy; access is granted via user_roles (owner/admin) and agent_group_members. Map:

• dmPolicy/groupPolicy: "open" → leave the default; no extra grants.

• allowFrom / groupAllowFrom lists → for each allowed sender, upsert the user and add them as a member of the relevant agent group via ncl:

  ncl users create --id "<channel>:<handle>" --kind <channel> --display-name "<name>"
  ncl members add --user "<channel>:<handle>" --group "<ag-id>"
  

• dmPolicy: "disabled" → don't wire that chat (or leave it registered but unwired).

The messaging groups register / init-first-agent create default their unknown_sender_policy to whatever the channel adapter declares for that context (DM vs group) — strict when the channel has no declaration — so unknown senders are gated until you add them (or an admin approves the adapter-declared approval card). Pass --unknown-sender-policy to register to override. Show the user the OpenClaw allowlist and confirm who to grant before running the commands.

Phase 3: Identity and Memory

Fully conversational — read files directly and discuss. Placement depends on the Phase 1 choice:

• Shared identity: merge the same core identity/personality into every selected group's instructions.prepend.md.
• Fully separate / primary only: merge identity/personality only into the corresponding group's instructions.prepend.md.

Never edit a composed CLAUDE.md or AGENTS.md; it is regenerated each spawn. Put standing behavior in instructions.prepend.md and facts in memory/.

Find workspace files at <STATE_DIR>/workspace/. If AGENT_COUNT > 1, also check <STATE_DIR>/agents/*/workspace/ and ask which agent maps to which v2 agent group.

IDENTITY.md / SOUL.md

Read them. Distinguish always-loaded vs reference:

• Standing behavior (core traits, communication style, key rules) → weave into the group's instructions.prepend.md.
• Reference (backstory, extended guidelines) → a separate durable concept in an appropriate folder under groups/<folder>/memory/, linked from that folder's index.md and the root Map.

Choose each memory folder based on which related information will be easiest to find together; a folder may contain different concept types. Before writing the first concept into a new folder, create the folder and its index.md. Follow memory/system/definition.md, including its YAML frontmatter rules, for every new concept.

Show proposed edits before applying — this is a thoughtful merge, not a paste.

USER.md

Create a focused user-context concept in an appropriate memory folder and link it through that folder's index and the root Map. Put only facts relevant in nearly every conversation (for example name or timezone) into ## Core Memory; keep all other details in the linked file.

MEMORY.md and daily memory files

Show MEMORY.md; keep relevant items in focused concepts under the chosen memory folders, with links through each folder index and the root Map. For daily files (workspace/memory/*.md, count = DAILY_MEMORY_FILES):

AskUserQuestion: "You have N daily memory files. How to handle them?"

1. Copy as-is — agree on a descriptive folder, create it and its index.md, then copy with cp <workspace>/memory/*.md <group_dir>/memory/<chosen-folder>/ and link the retained files through its index and the root Map.
2. Consolidate — read, extract durable facts, and place them in focused linked memory files.
3. Skip.

OpenClaw skills

If SKILL_COUNT > 0, the SKILL.md format is shared, so skills are portable. Present each (name + description from the front matter) and let the user pick. For each confirmed skill, copy the directory into the container skills tree:

cp -r <skill_source_dir> container/skills/<skill_name>

A container rebuild is needed afterward — note it for Phase 8.

Config-registered plugins (with API keys)

If CONFIG_PLUGINS is non-empty, OpenClaw had plugins/skills carrying keys. For each, read the config section and decide together:

• Matching v2 skill → run that skill; route its credential per Phase 4.
• An MCP server → install the exact configured package; wire via ncl groups config add-mcp-server. Don't guess at packages.
• An API key → route to the OneCLI vault if container-facing (Phase 4).

Don't install unknown packages or search for replacements — supply-chain risk.

Phase 4: Credentials

Two destinations, decided per credential. Channel tokens → .env (host reads them). Container-facing API credentials → the OneCLI vault (injected per request, never in container env).

Channel tokens (telegram, discord, slack)

Preview, then write to .env. The script emits only masked values:

pnpm exec tsx ${CLAUDE_SKILL_DIR}/scripts/extract-channel-credentials.ts \
  --state-dir <STATE_DIR> --channel <name>

Parse the status block. DESTINATION: env confirms a host-side token. Show CREDENTIAL_MASKED (and CREDENTIAL_MASKED_2 for Slack's app token).

AskUserQuestion:

1. Use this credential — re-run with --write-env .env to save it.
2. Enter a new one — ask in plain text, write to .env yourself.
3. Skip this channel.

pnpm exec tsx ${CLAUDE_SKILL_DIR}/scripts/extract-channel-credentials.ts \
  --state-dir <STATE_DIR> --channel <name> --write-env .env

Check WRITTEN_TO / WRITTEN_COUNT. Slack writes both SLACK_BOT_TOKEN and SLACK_APP_TOKEN in one run.

If HAS_CREDENTIAL=false but a credential is expected: the config shape may be unrecognized, or it uses a file/exec SecretRef (CREDENTIAL_SOURCE ends in _ref with a NOTE) that can't be auto-extracted. Read the channel section of the config directly and ask the user to confirm or paste the value.

WhatsApp: authenticates via QR/pairing code — there's no token. Don't copy Baileys auth state (stale encryption sessions break decryption). Re-authenticate during /setup via /add-whatsapp. The extraction script reports DESTINATION: none for it.

Anthropic and other container-facing credentials → OneCLI vault

Find the agent's model credentials in OpenClaw. Check, in order:

1. <STATE_DIR>/auth-profiles.json (and <STATE_DIR>/agents/<id>/agent/auth-profiles.json) — a profiles map keyed provider:identifier. For an anthropic provider profile the value depends on type: api_key → key, token → token, oauth → access.
2. <STATE_DIR>/.env — ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN.
3. Config models.providers — Anthropic provider apiKey.

These are container-facing, so they go to the OneCLI vault. Do not write them to .env or thread them into a container. Register each in the vault:

onecli secrets create --name Anthropic --type anthropic \
  --value <key-or-token> --host-pattern api.anthropic.com

For other container-facing keys discovered in plugins (e.g. OpenAI):

onecli secrets create --name OpenAI --type api_key \
  --value <key> --host-pattern api.openai.com

Run the command on the user's behalf so the value never lands in the chat transcript; confirm with onecli secrets list.

Caveats: keyRef/tokenRef with source:"exec" or source:"file" can't be auto-extracted — ask the user to paste it. For an oauth profile with a past expiry, warn that the token may need refreshing; the user can run claude setup-token and register the fresh token.

If OneCLI isn't installed yet, defer this: tell the user that during /setup (or /init-onecli) they'll register the Anthropic credential, and note the discovered profile in migration-state.md so it isn't lost.

There is no supported .env-credentials opt-out anymore: the session spec's admission rules refuse credential values in container env on every lane, by design (the retired /use-native-credential-proxy skill would be denied at every spawn). Credentials go through the OneCLI vault; custom Anthropic endpoints use ANTHROPIC_BASE_URL plus the placeholder-token pattern from setup, with the gateway rewriting the header on the wire.

Phase 5: Scheduled Tasks

Read <STATE_DIR>/cron/jobs.json. If absent or empty, skip.

If jobs exist, read ${CLAUDE_SKILL_DIR}/MIGRATE_CRONS.md for the v2 task model, the mapCronToRecurrence transform, the full field mapping, and how tasks are created (the agent's schedule_task MCP tool, since tasks live in a per-session inbound.db the host owns). Follow it for each enabled job.

Phase 6: MCP, Webhooks, Other Config

Read the relevant config sections directly. Conversational.

MCP servers

If MCP_SERVERS is non-empty, v2 supports per-agent-group MCP servers via the container config. Read each server's command/args/env/url from mcp.servers. For each one the user wants:

ncl groups config add-mcp-server --id <agent-group-id> \
  --name <server-name> --command <cmd> \
  [--args '<json-array>'] [--env '<json-object>']

stdio servers must be runnable inside the container (Node/npx-based work; custom binaries need a Dockerfile addition). Secrets referenced by a server's env should go to the OneCLI vault (Phase 4), not be inlined. The config change takes effect on restart: ncl groups restart --id <agent-group-id> (add --rebuild only if a custom binary was added to the Dockerfile).

Webhooks

OpenClaw cron.webhook / failureDestination / channel webhooks don't map to a v2 primitive. For a notification webhook, fold it into a scheduled task's prompt or a pre-agent script that curls the endpoint. Discuss the use case.

Other config (mention and move on)

• Exec approvals / command allowlist → v2 uses container isolation; the agent runs sandboxed.
• Human delay / TTS / compaction / model config → not v2 task/group fields (per-group model is in the container config).

Phase 7: Welcome and First Run

init-first-agent (Phase 1) already queued a welcome DM for the primary owner agent. If the service was up, the owner should have received it. For groups registered via setup --step register, the wiring also queues a /welcome onboarding message on first wiring.

Tell the user which agents are live now and which await channel installation (unsupported channels registered for the future).

Phase 8: Validate and Summarize

Run the shipped test

Copy the transform module and its test into the project so vitest runs them against the composed install, then build and test:

cp ${CLAUDE_SKILL_DIR}/scripts/transform.ts        scripts/openclaw-transform.ts
cp ${CLAUDE_SKILL_DIR}/scripts/transform.test.ts   scripts/openclaw-transform.test.ts
# Point the copied test at the copied module name:
sed -i.bak "s#from './transform.js'#from './openclaw-transform.js'#" scripts/openclaw-transform.test.ts && rm -f scripts/openclaw-transform.test.ts.bak

pnpm run build
pnpm exec vitest run scripts/openclaw-transform.test.ts

The test guards the skill's two v2 integration assumptions: credential routing (container-facing → vault, channel tokens → .env) and the cron → v2 recurrence mapping. It imports the real cron-parser (the same parser the host recurrence sweep uses), so a missing/renamed dependency turns it red. build typechecks the transform module against the project.

These copied files are the only files the skill installs into the project tree; REMOVE.md deletes them.

If a container rebuild is needed

If OpenClaw skills were copied or MCP servers added: ./container/build.sh, then restart the service.

Summary

Print what was migrated:

• Owner + primary agent → users / user_roles / agent group + welcome DM
• Additional groups → messaging groups + wiring (folders + session modes)
• Timezone → .env TZ; container timeout → noted
• Access grants → members/roles for OpenClaw allowlist senders
• Identity/personality → per-group instructions.prepend.md + linked memory concepts
• User context / memories → Core Memory only for universal facts; otherwise linked concepts in content-based folders under memory/
• OpenClaw skills → container/skills/
• Channel tokens → .env (list channels)
• Container-facing credentials → OneCLI vault (list)
• Scheduled tasks → mapped and scheduled via the agent (or noted for first run)
• MCP servers → wired into agent group container configs

Noted for later: channel installs during /setup; container rebuild if needed; tasks deferred until a session exists.

Not applicable: unsupported channels (registered for the future); OpenClaw-only features (exec approvals, human delay, TTS, model/thinking config).

Remind: "Run /setup next to finish your NanoClaw install. Channel tokens are in .env; container-facing credentials are in the OneCLI vault. Select the channels we configured when setup asks."

Then delete migration-state.md (or offer to keep it as a record), and remove the copied transform files if you don't want them lingering (see REMOVE.md).

Troubleshooting

• Config parse error: the JSON5 parser may not handle unusual syntax. Read the file directly and work with it manually.
• Credential not found: likely a file/exec SecretRef — ask the user to paste the value.
• init-first-agent can't reach the CLI socket: the service isn't running. Start it, then re-run.
• Multi-agent complexity: do the primary/default agent first; add others as separate agent groups later.