@mir-stream/openclaw-prompt-compat

An OpenClaw compatibility plugin that changes one identity sentence in fingerprinted OpenClaw system prompts:

You are a personal assistant running inside OpenClaw.
→
You are a personal assistant running within OpenClaw.

The replacement sentence is configurable; the sentence and structure it matches are not. The plugin is provider- and model-independent and has no runtime dependencies.

Configuration

The plugin exposes one setting, identitySentence: the sentence written in place of the matched OpenClaw identity. The plugin makes no assumptions about the role — use any sentence that fits your deployment.

Config path:

plugins.entries.openclaw-prompt-compat.config.identitySentence

Set it non-interactively

Install 0.2.0 (or later) first — the setting is rejected by the strict config schema until the version that declares it is installed. Then write the value and restart the Gateway:

openclaw config set \
  "plugins.entries.openclaw-prompt-compat.config.identitySentence" \
  "You are the assistant for the Acme support desk."

openclaw gateway restart

The example sentence is illustrative only; substitute your own.

In a setup script, prefer batch mode — unlike value mode, it validates the write against the schema, so an unknown key or an uninstalled version fails loudly instead of writing silently:

openclaw config set --batch-file ./identity.batch.json

[
  {
    "path": "plugins.entries.openclaw-prompt-compat.config.identitySentence",
    "value": "You are the assistant for the Acme support desk."
  }
]

Confirm the stored value with openclaw config get "plugins.entries.openclaw-prompt-compat.config.identitySentence".

Equivalent config-file form

{
  "plugins": {
    "entries": {
      "openclaw-prompt-compat": {
        "config": {
          "identitySentence": "You are the assistant for the Acme support desk."
        }
      }
    }
  }
}

Behavior

• Default: You are a personal assistant running within OpenClaw. An install with no config behaves exactly as it did before this setting existed.
• Maximum length: 2000 characters, after trimming surrounding whitespace.
• Multiple lines are allowed. The prompt continues with ## Tooling on its own line, so extra lines do not disturb the surrounding structure.
• A non-string, empty, whitespace-only, or over-long value logs a warning and falls back to the default rather than silently doing nothing. A value that repeats the original identity sentence is used as given, with a warning that it defeats the rewrite.
• Changing the value takes effect only after a Gateway restart — the rewrite is registered once at plugin startup.

How it is scoped

The plugin registers one input-only text replacement with api.registerTextTransforms. The non-global regular expression changes only the exact identity sentence at the absolute string start, and only when that string carries this ordered structural fingerprint:

1. The exact identity sentence at the absolute string start
2. ## Tooling immediately after the identity sentence
3. ## Safety later in the string
4. ## Workspace after that, on a line of its own — the ## Workspace Files (injected) line it prefixes does not satisfy it
5. ## Workspace Files (injected) after that
6. OpenClaw's internal <!-- OPENCLAW_CACHE_BOUNDARY --> marker later still
7. ## Runtime after the cache boundary

Apart from the identity sentence itself, every anchor is a section heading or an internal marker. Prompt prose is deliberately excluded: descriptive sentences are rewritten from release to release, while these headings and their order have held.

The core prompt must start at the absolute string start. Arbitrary or hook-prepended system context before the core prompt, an inline copy, changed capitalization, whitespace, headings, marker order, or an incomplete prompt does not match. promptMode: "none" does not carry the required fingerprint and is not changed.

The absolute anchor prevents quoted identity or scaffold text in prepended context from borrowing the real prompt's later markers. Atomic marker segments prevent repeated near-matches or marker chains from causing quadratic or combinatorial rescans. Marker gaps may be empty, including a cache boundary immediately followed by ## Runtime.

Compatibility

• Package: @mir-stream/[email protected]
• Plugin id: openclaw-prompt-compat
• OpenClaw host: >=2026.7.1
• OpenClaw plugin API: >=2026.7.1
• Node.js: >=22.22.3 <23 || >=24.15.0 <25 || >=25.9.0

The fingerprint uses only structural headings and markers — no prompt prose. ## Tooling, ## Safety, ## Workspace, ## Workspace Files (injected), and ## Runtime were present and in order in all 11 sampled OpenClaw releases from 2026.2.26 through 2026.7.2-beta.4; the <!-- OPENCLAW_CACHE_BOUNDARY --> marker was introduced in 2026.4.15 and has not changed since. Rendered prompts from 2026.4.15 through 2026.7.2-beta.4 all match.

That match range is not a support range. The supported host floor stays >=2026.7.1; the fingerprint simply happens to match further back.

The fingerprint remains a structural dependency and fails closed: if the required structure is absent or reordered, this plugin makes no replacement.

Install

Install and pin the published version:

openclaw plugins install \
  "npm:@mir-stream/[email protected]" \
  --pin

Restart the Gateway if the running deployment does not reload plugins automatically:

openclaw gateway restart

Disable or uninstall

Disabling is the preferred operational rollback because it preserves the installed package:

openclaw plugins disable openclaw-prompt-compat
openclaw gateway restart

Setup automation should preserve an explicitly disabled entry. Avoid unconditionally reinstalling with plugins install --force, which can enable the plugin again.

To remove the managed package and its install record:

openclaw plugins uninstall openclaw-prompt-compat

An external setup process may reinstall an absent plugin. If setup automation manages this package, give it a persistent opt-out before uninstalling.

Security limitation

registerTextTransforms is a global input-text compatibility API, not a system-prompt-only API. OpenClaw can apply registered input replacements to system prompts, user and assistant messages, tool results, conversation history, and string values inside tool-call arguments.

The structural fingerprint makes ordinary collisions—including a standalone identity sentence or a generic identity-plus-Tooling quote—very unlikely. It is not a security boundary. A user, tool result, or history entry that deliberately copies the exact identity sentence at its start and then all of the ordered structural headings and markers can also match and have its first identity sentence changed. Strict system-prompt-only behavior requires a dedicated upstream OpenClaw transform surface.

Because identitySentence is user-authored text, that text is what gets written wherever the fingerprint matches — not only in system prompts. Treat a configured sentence as content that may surface in any matching string, and keep it free of secrets.

Development

npm ci
npm test
npm run typecheck
npm run build
npm run pack

npm run pack performs an npm package dry-run and lists the files that would be published.