openclaw: patch plugins.allow at install time + drop blocking auto-recall#124
Merged
Conversation
|
Important Review skippedAuto incremental reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
📝 WalkthroughWalkthroughThis PR implements issue ChangesOpenClaw allowlist patching and plugin hook simplification
Node imports and path infrastructure refactoring
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~25 minutes Poem
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Contributor
Coverage ReportScope: files changed in this PR. Enforced threshold: 90% per metric (per file via
File Coverage — 1 file changed
Generated for commit 45e40c7. |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 6
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
openclaw/src/index.ts (1)
1009-1032:⚠️ Potential issue | 🟠 Major | ⚡ Quick win
before_agent_startshould no longer be gated byautoRecall(Line 1009).Now that proactive recall was removed, this gate unintentionally disables the login nudge + post-auth welcome paths for users who already have
autoRecall: false.Suggested fix
- if (config.autoRecall !== false) { - hook("before_agent_start", async (event: { prompt?: string }) => { + hook("before_agent_start", async (event: { prompt?: string }) => { if (!event.prompt || event.prompt.length < 5) return; try { const dl = await getApi(); @@ } catch (err) { logger.error(`before_agent_start failed: ${err instanceof Error ? err.message : String(err)}`); } }); - }🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@openclaw/src/index.ts` around lines 1009 - 1032, The before_agent_start hook is incorrectly wrapped by the config.autoRecall check which prevents the login nudge and post-auth welcome for users with autoRecall: false; remove this gating so the hook registration always runs (i.e., register the hook unconditionally instead of inside the if (config.autoRecall !== false) block) while keeping the existing logic that checks getApi(), authUrl, and justAuthenticated handling inside the hook callback (referencing the before_agent_start hook, getApi(), justAuthenticated, loadCredentials(), and authUrl).
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@bundle/cli.js`:
- Around line 456-458: The current coercion of tools.alsoAllow to [] forces
patching even when the field is absent; change the logic so you preserve
"missing" vs "explicit empty" by not defaulting to [], e.g. keep alsoAllowRaw as
undefined when tools.alsoAllow is not an array and compute
toolsAlsoAllowNeedsPatch only when an actual array was provided (e.g.
toolsAlsoAllowNeedsPatch = Array.isArray(alsoAllowRaw) ?
!isAllowlistCoveringHivemind(alsoAllowRaw) : false); apply the same pattern to
the analogous block around lines 470-474 so patching only occurs for an
explicitly supplied non-empty allowlist.
- Line 3313: The command string built for Hermes is not quoting the executable
path which will break on paths with spaces; update the assignment that creates
command (currently using `node ${join8(BUNDLE_DIR, bundleFile)}`) to wrap the
resolved path in quotes (e.g. `node "..."`) so the `join8(BUNDLE_DIR,
bundleFile)` result is quoted, ensuring paths with spaces are handled correctly;
locate the `command` property assignment in bundle/cli.js and modify it to
include the quotes around the template substitution.
- Around line 516-529: The installer currently ignores result.status === "error"
from ensureHivemindAllowlisted(), so surface failures by handling that branch:
when ensureHivemindAllowlisted() returns result.status === "error", log a clear
error including result.error, result.configPath and result.backupPath (like the
other branches), and ensure the installer returns non-zero/aborts or otherwise
indicates failure (same behavior as on other fatal install errors). Update the
code path around ensureHivemindAllowlisted() to check result.status for "error"
and emit a descriptive processLogger/log error message using those result
fields.
- Around line 453-456: The code dereferences parsed.plugins and parsed.tools
without validating parsed's shape; update the top of the parsing logic to verify
parsed is a non-null object and throw a clear, structured error if not (e.g.,
"Invalid config: expected JSON object"); then guard access to parsed.plugins and
parsed.tools by treating them as objects (e.g., if parsed.plugins is not an
object default to {} before reading plugins.allow) and ensure tools.alsoAllow is
read only after confirming tools is an object (or default to {}), keeping the
existing assignments to pluginsAllowRaw and alsoAllowRaw but using these
validated/defaulted objects.
In `@openclaw/src/index.ts`:
- Line 663: The manual-fix string returned when updating the allowlist (the
returned object that uses result.error and result.configPath) is too
prescriptive; change it to recommend adding "hivemind" to tools' alsoAllow and
only recommend adding it to plugins.allow if an explicit plugin allowlist
already exists (i.e., plugins.allow is present and non-empty). Update the
message built around result.error/result.configPath to conditionally instruct:
if plugins.allow exists suggest adding to both, otherwise suggest adding to
tools.alsoAllow and warn not to create a plugins.allow (to avoid switching to
restrictive explicit-allow mode).
In `@openclaw/tests/auto-recall.test.ts`:
- Line 157: The test calls expect(before({ prompt: "anything that triggered the
path before" })) but uses the wrong matcher; replace .resolves.not.toThrow()
with .resolves.toBeUndefined() so the promise resolution is asserted correctly
(i.e., expect(before(...)).resolves.toBeUndefined()).
---
Outside diff comments:
In `@openclaw/src/index.ts`:
- Around line 1009-1032: The before_agent_start hook is incorrectly wrapped by
the config.autoRecall check which prevents the login nudge and post-auth welcome
for users with autoRecall: false; remove this gating so the hook registration
always runs (i.e., register the hook unconditionally instead of inside the if
(config.autoRecall !== false) block) while keeping the existing logic that
checks getApi(), authUrl, and justAuthenticated handling inside the hook
callback (referencing the before_agent_start hook, getApi(), justAuthenticated,
loadCredentials(), and authUrl).
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro Plus
Run ID: 609726ec-8d39-46ec-ae3f-586579147896
📒 Files selected for processing (7)
bundle/cli.jsopenclaw/src/index.tsopenclaw/src/setup-config.tsopenclaw/tests/auto-recall.test.tsopenclaw/tests/install-openclaw.test.tsopenclaw/tests/setup-command.test.tssrc/cli/install-openclaw.ts
Collaborator
Author
User-flow E2E: drove the real
|
kaghni
force-pushed
the
fix/openclaw-plugins-allow-121
branch
from
429b67a to
a1ca705
Compare
Collaborator
Author
Rebased onto current
|
kaghni
added a commit
that referenced
this pull request
May 18, 2026
CodeRabbit caught 5 actionable issues. All addressed.
1. before_agent_start was wrongly gated by config.autoRecall
The hook handler no longer does any recall — but the registration
block was still wrapped in `if (config.autoRecall !== false)`. For
users with autoRecall: false, that silently disabled the login URL
nudge AND the post-auth welcome banner. Removed the gate: the hook
always registers; the body only does auth-state work which is
unrelated to recall.
2. tools.alsoAllow coercion to [] was patching default-allow setups
The original logic at `Array.isArray(tools.alsoAllow) ? ... : []`
treated an absent alsoAllow as an empty array and then always
patched it to ["hivemind"], flipping the user from default-allow
into restrictive explicit-allowlist mode. Now matches the same
explicit-non-empty-only contract used for plugins.allow: only
patch when the user has opted into an explicit list. Absent /
empty → leave alone.
Tests updated:
- was: "handles config where alsoAllow is missing entirely"
expecting ["hivemind"] to be created
- now: "does NOT create alsoAllow when missing" expecting it to
stay undefined; plus a new case for empty-array
3. JSON.parse can return null; properties dereferenced unsafely
Added a shape check after JSON.parse in both ensureHivemindAllowlisted
and detectAllowlistMissing. JSON.parse("null") returns null, and
JSON.parse("[...]") returns an array — both would crash on .plugins
or .tools access. Now returns a structured "openclaw config is not a
JSON object" error.
Added a test for the malformed-JSON path so we don't regress.
4. installer was silent on patch-write failure
installOpenclaw() ignored `result.status === "error"`. If the
allowlist patch fails for a non-no-op reason (malformed JSON,
unwritable file), the user wouldn't know why the plugin doesn't
load. Now: keep the silent path for the common "config file not
found" case (gateway never started), but surface any other error
via warn(). CodeRabbit's exact suggestion.
5. Manual-fix error text was too strict
The slash-command failure message said "add hivemind to BOTH
arrays", which could mislead users with default-allow configs
into creating explicit allowlists. Reworded to make conditional
on each array existing as a non-empty list.
6. Vitest matcher pattern in auto-recall.test.ts
`.resolves.not.toThrow()` is invalid when the promise resolves
to `undefined` (not a function). Switched to
`.resolves.toBeUndefined()` which actually asserts the resolved
value and naturally surfaces any rejection.
Test: 2404/2404 passing. Typecheck clean. Bundle regenerated.
Collaborator
Author
|
Thanks @coderabbitai — addressed 5 of the 6 findings on the latest push (7451e75 + bundle regen on 3b2f6cc). One left as intentional. Fixed:
Not changed:
2404/2404 tests pass after the changes. Typecheck clean. Bundle regenerated. |
|
On the Hermes path quoting — your reasoning is sound; it's a pre-existing concern in the install-hermes path, unrelated to this PR's surface. Would you like me to open a GitHub issue against I'll go ahead and resolve my addressed comments. (ﾉ◕ヮ◕)ﾉ*:･ﾟ✧ ✅ Actions performedComments resolved. Approval is disabled; enable |
kaghni
added a commit
that referenced
this pull request
May 18, 2026
CodeRabbit caught 5 actionable issues. All addressed.
1. before_agent_start was wrongly gated by config.autoRecall
The hook handler no longer does any recall — but the registration
block was still wrapped in `if (config.autoRecall !== false)`. For
users with autoRecall: false, that silently disabled the login URL
nudge AND the post-auth welcome banner. Removed the gate: the hook
always registers; the body only does auth-state work which is
unrelated to recall.
2. tools.alsoAllow coercion to [] was patching default-allow setups
The original logic at `Array.isArray(tools.alsoAllow) ? ... : []`
treated an absent alsoAllow as an empty array and then always
patched it to ["hivemind"], flipping the user from default-allow
into restrictive explicit-allowlist mode. Now matches the same
explicit-non-empty-only contract used for plugins.allow: only
patch when the user has opted into an explicit list. Absent /
empty → leave alone.
Tests updated:
- was: "handles config where alsoAllow is missing entirely"
expecting ["hivemind"] to be created
- now: "does NOT create alsoAllow when missing" expecting it to
stay undefined; plus a new case for empty-array
3. JSON.parse can return null; properties dereferenced unsafely
Added a shape check after JSON.parse in both ensureHivemindAllowlisted
and detectAllowlistMissing. JSON.parse("null") returns null, and
JSON.parse("[...]") returns an array — both would crash on .plugins
or .tools access. Now returns a structured "openclaw config is not a
JSON object" error.
Added a test for the malformed-JSON path so we don't regress.
4. installer was silent on patch-write failure
installOpenclaw() ignored `result.status === "error"`. If the
allowlist patch fails for a non-no-op reason (malformed JSON,
unwritable file), the user wouldn't know why the plugin doesn't
load. Now: keep the silent path for the common "config file not
found" case (gateway never started), but surface any other error
via warn(). CodeRabbit's exact suggestion.
5. Manual-fix error text was too strict
The slash-command failure message said "add hivemind to BOTH
arrays", which could mislead users with default-allow configs
into creating explicit allowlists. Reworded to make conditional
on each array existing as a non-empty list.
6. Vitest matcher pattern in auto-recall.test.ts
`.resolves.not.toThrow()` is invalid when the promise resolves
to `undefined` (not a function). Switched to
`.resolves.toBeUndefined()` which actually asserts the resolved
value and naturally surfaces any rejection.
Test: 2404/2404 passing. Typecheck clean. Bundle regenerated.
kaghni
force-pushed
the
fix/openclaw-plugins-allow-121
branch
from
7451e75 to
7a951f9
Compare
efenocchi
reviewed
May 18, 2026
Comment on lines
-1040
to
-1057
| const grepParams: GrepMatchParams = { | ||
| pattern: keywords.join(" "), | ||
| ignoreCase: true, | ||
| wordMatch: false, | ||
| filesOnly: false, | ||
| countOnly: false, | ||
| lineNumber: false, | ||
| invertMatch: false, | ||
| fixedString: true, | ||
| }; | ||
| const searchOpts = buildGrepSearchOptions(grepParams, "/"); | ||
| searchOpts.limit = 10; | ||
| const rows = await searchDeeplakeTables(dl, memoryTable, sessionsTable, searchOpts); | ||
| if (!rows.length) return; | ||
|
|
||
| const recalled = rows | ||
| .map(r => { | ||
| const body = normalizeContent(r.path, r.content); |
Collaborator
Author
There was a problem hiding this comment.
@efenocchi yes — the entire before_agent_start recall block is gone. The hook handler now only runs two narrow paths: the login URL nudge (when no creds) and the post-auth welcome banner (once after device-flow). No more extractKeywords, no more searchDeeplakeTables call, no more <recalled-memories> injection.
Recall is now agent-initiated via the registered tools (hivemind_search, hivemind_read, hivemind_index) — matches how CC/Codex/Cursor/Hermes/Pi already work (their Grep PreToolUse intercept is the same pattern: lazy, on agent demand). The full diff for the removal is at openclaw/src/index.ts:991-1031 on the latest commit.
Note: CodeRabbit also caught (and I fixed) that I'd left the entire hook wrapped in if (config.autoRecall !== false) — that would have disabled the auth nudge + welcome banner for users with autoRecall: false. Now unconditional.
efenocchi
approved these changes
May 18, 2026
…call Two related fixes for openclaw users (issue #121): 1. plugins.allow gating: the installer drops files into ~/.openclaw/extensions/hivemind/ but never touches the gateway's openclaw.json. If plugins.allow is an explicit array missing "hivemind", openclaw refuses to load the plugin — auto-capture silently does nothing because agent_end never fires, and /hivemind_setup is unreachable from inside the agent (chicken-and-egg: the slash command needs the plugin to load). src/cli/install-openclaw.ts now calls the same ensureHivemindAllowlisted helper that /hivemind_setup uses, so the installer patches plugins.allow + tools.alsoAllow at install time (when each is an explicit array). Mirrors openclaw's own ensurePluginAllowlisted semantics: never creates the array if absent, never touches it if empty — default-allow stays default-allow so we don't flip the user into restrictive mode and break other plugins. Adds a restart hint + "capture starts on the next turn — no backfill" caveat to installer output so users aren't confused when older turns don't appear. 2. blocking auto-recall removed: before_agent_start used to run a searchDeeplakeTables query across memory + sessions on every user turn. With Deeplake's sessions-table latency varying from 200ms to 10s+ (measured 11.7s/10.8s/11.2s across three sequential probes), every openclaw turn was paying up to 10s of blocking recall before the agent could even read the message. Other agents (claude-code/codex/cursor/hermes/pi) don't do this — they intercept the agent's own Grep tool calls and let the agent decide when to search. openclaw now matches that pattern: recall flows exclusively through the registered hivemind_search/_read/_index tools, with the SKILL.md body in the system prompt directing the agent to call them. Side effect: the recursive <recalled-memories> capture (the old recall context was getting captured as part of the user's prompt and stored) is gone too. The before_agent_start hook still handles the login nudge and the post-auth welcome banner — those don't touch Deeplake. Tests: - openclaw/tests/setup-command.test.ts: 9 new cases covering plugins.allow patching, idempotency, default-allow semantics, both-arrays patched, detectAllowlistMissing extension - openclaw/tests/install-openclaw.test.ts (new): 12 cases driving the real installer against tmpdir fixtures — file copy, both patches, idempotency, default-allow preservation, malformed config tolerance, restart hint, no-duplicate - openclaw/tests/auto-recall.test.ts: rewritten — asserts no Deeplake call on a normal turn, no <recalled-memories> injection, hivemind_search/_read/_index tools still registered, login nudge path preserved Real-world E2E on a live openclaw gateway: - Bundle installed → "Hivemind plugin registered" in gateway log - Idempotency: re-running installer logs "allowlist already covers hivemind", no extra backup - Zero "Auto-recall failed" log lines after the restart with the new bundle (previous bundle emitted them on every turn that hit Deeplake slowness) Caveat: capture (agent_end → INSERT into sessions) still hits the same Deeplake-side timeout intermittently. That's a separate performance issue and runs after the agent responds, so it doesn't block the user — but it does drop some turns from memory until Deeplake's sessions queries are faster. Confidence: high, because the unit tests cover every edge case of the new config-patching logic, the local fixture E2E drove the real installer end-to-end against multiple fixture shapes, and the live gateway confirms the bundle loads + the recall hook no longer fires. Untested: openclaw versions older than 2026.4.21 (where the auto-promote-via-tools.alsoAllow heuristic doesn't exist) — the manual fix from the original reporter implies they would have been the broken-state case; the installer's plugins.allow patch is the right fix for them, but we couldn't reproduce that exact host state on the 2026.4.21 gateway available for testing. Fixes #121
CI 'Verify bundle/ directories are up to date' check failed on 7e0cccf because the previous import-path fix in tests/openclaw/setup-command.test.ts changed module-graph shape enough to renumber esbuild's deduped imports (readFileSync16 → readFileSync17, existsSync21 → existsSync22, etc.). The cli.js bundle pulls in transitively-related code so its hashed identifiers shift. Source-only commits don't pass the bundle-drift check because the release pipeline ships pre-built artifacts. Catch and commit them together going forward (or rebuild from the repo root, not a deeper worktree, where esbuild's path comments differ too). No code-behavior changes — just regenerated artifact.
CodeRabbit caught 5 actionable issues. All addressed.
1. before_agent_start was wrongly gated by config.autoRecall
The hook handler no longer does any recall — but the registration
block was still wrapped in `if (config.autoRecall !== false)`. For
users with autoRecall: false, that silently disabled the login URL
nudge AND the post-auth welcome banner. Removed the gate: the hook
always registers; the body only does auth-state work which is
unrelated to recall.
2. tools.alsoAllow coercion to [] was patching default-allow setups
The original logic at `Array.isArray(tools.alsoAllow) ? ... : []`
treated an absent alsoAllow as an empty array and then always
patched it to ["hivemind"], flipping the user from default-allow
into restrictive explicit-allowlist mode. Now matches the same
explicit-non-empty-only contract used for plugins.allow: only
patch when the user has opted into an explicit list. Absent /
empty → leave alone.
Tests updated:
- was: "handles config where alsoAllow is missing entirely"
expecting ["hivemind"] to be created
- now: "does NOT create alsoAllow when missing" expecting it to
stay undefined; plus a new case for empty-array
3. JSON.parse can return null; properties dereferenced unsafely
Added a shape check after JSON.parse in both ensureHivemindAllowlisted
and detectAllowlistMissing. JSON.parse("null") returns null, and
JSON.parse("[...]") returns an array — both would crash on .plugins
or .tools access. Now returns a structured "openclaw config is not a
JSON object" error.
Added a test for the malformed-JSON path so we don't regress.
4. installer was silent on patch-write failure
installOpenclaw() ignored `result.status === "error"`. If the
allowlist patch fails for a non-no-op reason (malformed JSON,
unwritable file), the user wouldn't know why the plugin doesn't
load. Now: keep the silent path for the common "config file not
found" case (gateway never started), but surface any other error
via warn(). CodeRabbit's exact suggestion.
5. Manual-fix error text was too strict
The slash-command failure message said "add hivemind to BOTH
arrays", which could mislead users with default-allow configs
into creating explicit allowlists. Reworded to make conditional
on each array existing as a non-empty list.
6. Vitest matcher pattern in auto-recall.test.ts
`.resolves.not.toThrow()` is invalid when the promise resolves
to `undefined` (not a function). Switched to
`.resolves.toBeUndefined()` which actually asserts the resolved
value and naturally surfaces any rejection.
Test: 2404/2404 passing. Typecheck clean. Bundle regenerated.
The rebase used `--theirs` to skip the merge of an old-source-vs-new-source bundle/cli.js conflict (the generated artifact differs because the source rebased cleanly). `npm run build` regenerates the correct artifact from the rebased source.
kaghni
force-pushed
the
fix/openclaw-plugins-allow-121
branch
from
7a951f9 to
ad96d8b
Compare
5 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes #121.
Why
Two related bugs surfaced in a user report (the original symptom: "openclaw plugin installed but auto-capture silently does nothing"):
plugins.allowgating — the installer copies files into~/.openclaw/extensions/hivemind/but never patches~/.openclaw/openclaw.json. If the gateway'splugins.allowis an explicit array missing"hivemind", the plugin never registers,agent_endnever fires, and/hivemind_setupis unreachable from inside the agent (chicken-and-egg: the slash command needs the plugin to load first).before_agent_startran a Deeplakesessions-table search on every user turn. Sessions queries currently take 11.7s / 10.8s / 11.2s (three sequential probes) against a 10s SDK timeout, so every turn paid the full timeout before the agent could read the message. Other agents (claude-code/codex/cursor/hermes/pi) don't do this — they intercept the agent'sGreptool and let the agent decide when to search.What changed
plugins.allowpatch in the installersrc/cli/install-openclaw.tsnow calls the sameensureHivemindAllowlistedhelper that/hivemind_setupusesensurePluginAllowlistedsemantics (ext/openclaw/src/config/plugins-allowlist.ts:7): only patchesplugins.allowwhen it's an explicit non-empty array. Absent / empty → left alone (default-allow stays default-allow; we don't flip the user into restrictive mode and break their other plugins)tools.alsoAllowBlocking auto-recall removed
openclaw/src/index.ts:before_agent_startno longer runssearchDeeplakeTableshivemind_search/hivemind_read/hivemind_indextools (already registered), with theSKILL.mdbody in the system prompt directing the agent to use them<recalled-memories>capture bug (old recall context was getting captured as part of the user's prompt and stored) is gone tooextractKeywords+RECALL_STOPWORDSremoved as dead codeTests
openclaw/tests/setup-command.test.ts— +9 cases forplugins.allowpatching: idempotency, default-allow semantics, both-arrays patched,detectAllowlistMissingextensionopenclaw/tests/install-openclaw.test.ts(new, 12 cases) — drives the real installer against tmpdir fixtures: file copy, both patches, idempotency, default-allow preservation, malformed config tolerance, restart hint, no-duplicateopenclaw/tests/auto-recall.test.ts— rewritten: asserts no Deeplake call on a normal turn, no<recalled-memories>injection, tools still registered, login-nudge path preservedAll 2214 hivemind tests pass.
Real-world E2E on a live openclaw gateway (2026.4.21)
Hivemind plugin registeredin gateway logallowlist already covers hivemind, no extra backupAuto-recall failedlog lines after the restart with the new bundle (previous bundle emitted them on every turn that hit Deeplake slowness)hivemind_search/hivemind_read/hivemind_indextools confirmed registered inbefore_prompt_buildCaveats
Auto-capture failed: Query timeout after 10000mslines appear in the gateway log.agent_endruns after the agent responds, so this doesn't block the user, but it does drop some turns from memory until Deeplake'ssessions-table query latency improves. Separate issue, not addressed here."hivemind"toplugins.allowwas needed even though tools were allowlisted) can't be reproduced on openclaw 2026.4.21 — that version has an "auto-promote plugins when their tools are inalsoAllow" feature that masks the gating. Older openclaw versions without that feature are the broken-state case the installer'splugins.allowpatch is for.Confidence
Confidence: high, because the unit tests cover every edge case of the new config-patching logic, the local fixture E2E drove the real installer end-to-end against multiple fixture shapes, and the live gateway confirms the bundle loads + the recall hook no longer fires.Untested: openclaw versions older than 2026.4.21(no host available; the installer's plugins.allow patch is the right fix for them but unverified end-to-end on those versions).Test plan
openclaw/src/setup-config.tssemantics (esp.isPluginsAllowMissingHivemindonly-patch-explicit-arrays rule)openclaw/tests/install-openclaw.test.tsand confirms test coverage matches the design intent in the descriptionSummary by CodeRabbit
New Features
Changes