The OpenClaw workspace bootstrap block (SOUL.md, IDENTITY.md, USER.md,
TOOLS.md, BOOTSTRAP.md, MEMORY.md, HEARTBEAT.md) was only being merged into
Codex's config.instructions. The Codex app-server runtime overlay
consistently applies the explicit developerInstructions field, so persona
and style guidance present in the workspace was failing to shape Codex
behavior on session resume.
Build the workspace bootstrap block before finalizing developerInstructions
and join it into both:
- the baseline developerInstructions (initial assignment), and
- the context-engine developerInstructions (when context engine is active),
preserving the existing config-engine projection addition.
The existing config.instructions merge stays intact, so the bootstrap now
reaches Codex through both paths and downstream hooks
(resolveAgentHarnessBeforePromptBuildResult) see what Codex will actually
receive. AGENTS.md remains excluded because Codex loads it natively.
Update the existing 'passes OpenClaw bootstrap files through ...' test to
also assert the developerInstructions field carries SOUL.md and the Codex
AGENTS.md substitution note while still excluding the native AGENTS.md
content.
Fixes#77363.
When a user's config has a stale `channels.<id>` entry (e.g. `appId`
or tokens left over from an earlier install) and the plugin is no
longer on disk -- for instance because the externalized npm package
was uninstalled or pruned during an upgrade -- `handleChannelChoice`
used to dead-end with "<channel> plugin not available." and leave
onboard stuck until the user manually deleted the config entry and
re-ran the CLI.
Two discovery paths are affected:
1. The `installedCatalogEntry` branch: when
`loadScopedChannelPlugin` returns null but the catalog entry still
carries `install.npmSpec`, fall back to
`ensureChannelSetupPluginInstalled` with the same entry so onboard
can reinstall the plugin from the official catalog.
2. The bundled-enable `else` branch: with a non-empty
`channels.<id>` record, `isStaticallyChannelConfigured` drops the
channel from `installableCatalogEntries`; if the plugin is also
missing on disk (so it never enters `manifestInstalledIds`), both
discovery buckets come back empty and the channel falls through to
`enableBundledPluginForSetup`. Before delegating to that bundled
path, consult the trusted catalog via
`getTrustedChannelPluginCatalogEntry` and, if an `install.npmSpec`
is available, drive the same catalog install flow used by a fresh
pick of the channel.
Both new fallbacks re-apply the `resolveConfigDisabledHint` guard
that `enableBundledPluginForSetup` has always enforced, so an
operator-disabled channel (`plugins.entries.<id>.enabled === false`
or explicit `channels.<id>.enabled === false`) with a stale config
entry cannot be silently reinstalled or re-enabled through the
catalog path.
Both branches also keep their previous behavior when no catalog npm
spec is available (e.g. purely bundled channels), so this change is
a superset of the old flow rather than a replacement.
Affects all externalized channel plugins listed in the core
package's `files` exclusion (qqbot, bluebubbles, discord, whatsapp,
line, msteams, feishu, googlechat, nostr, zalo, zalouser,
synology-chat, tlon, twitch, and similar).
Remove stale managed-root openclaw manifests, locks, hidden locks, and installed copies before npm plugin installs.
Relink plugin-local openclaw peer symlinks after shared-root npm install, rollback, update, and uninstall mutations so SDK-using plugins keep resolving openclaw/plugin-sdk/*.
Force safe npm commands out of inherited legacy/strict peer-dependency modes.
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
Co-authored-by: Patrick Erichsen <patrick.a.erichsen@gmail.com>