[v2026.6.10] [OpenClaw] Is it stable?

[github-actions]

Stability Score Evidence — v2026.6.10

2026.6.10 found more ways to make operators stare at logs: cron loses tools, WebChat forgets sessions, LINE time-travels, and Feishu still needs a manual kick. Works on my machine has entered witness protection.

Stability adjustments

Affected components

• config-migration
• message-delivery
• provider-routing
• memory
• auth
• skill-workshop
• telegram
• feishu
• runtime
• browser
• codex
• discord
• mcp
• reasoning
• voice
• slack
• ui
• plugin-install
• whatsapp
• prompt-size
• doctor
• exec
• usage-accounting
• tls
• web-search
• provider-auth
• acp
• attachments
• installer
• provider-cache
• cron-runtime
• line
• webchat

Evidence

Issue	Area	Type	Severity	Penalty	Reason
openclaw/openclaw#78493	config-migration	kernel	5	-5	Still reproducible on 2026.6.10: unreadable config/EACCES fallback can proceed to config writes, risking config clobber and crash-loop startup.
openclaw/openclaw#96374	message-delivery	kernel	5	-5	Telegram isolated ingress can leave stale claimed rows after restart, blocking all subsequent inbound delivery on 2026.6.10.
openclaw/openclaw#96337	provider-routing	runtime	3	-3	anthropic-vertex route=native on 2026.6.10 returns non-visible-output for pure text Claude responses; tool-call responses still work.
openclaw/openclaw#96335	memory	runtime	3	-3	lossless-claw context engine fails during read-only plugin registration on 2026.6.10 and falls back to the legacy engine.
openclaw/openclaw#95515	config-migration	runtime	3	-3	Upgrade to 2026.6.10 still reproduces email channel config corruption with an invalid groupAllowFrom field.
openclaw/openclaw#95997	auth	runtime	3	-3	Gateway CLI read probes can poison local device scopes so later write-scoped agent calls hit scope-upgrade pending approval; path is present in current main/release line.
openclaw/openclaw#91266	skill-workshop	tool-specific	2	-2	skill_workshop dynamic-tool apply still times out under approval wait on OpenClaw 2026.6.10 while the CLI path succeeds.
openclaw/openclaw#96363	telegram	channel-specific	1	-1	Telegram /status final reply fails with RICH_MESSAGE_EMAIL_INVALID when richMessages=true on stable 2026.6.10.
openclaw/openclaw#96657	feishu	channel-specific	5	-5	Feishu non-inbound tool calls can fall back to the global default account on 2026.6.10, causing cross-tenant 403/resource-access failures.
openclaw/openclaw#96564	runtime	runtime	4	-4	claude-cli headless warm-stdin loses conversation continuity on 2026.6.8–2026.6.10, wiping the session binding every turn.
openclaw/openclaw#96704	browser	tool-specific	3	-3	Managed browser cookies are in-memory only on 2026.6.10, so login sessions disappear after browser/gateway restart.
openclaw/openclaw#96585	runtime	runtime	3	-3	Windows/Node 25 reports repeated FileHandle garbage-collection crashes against OpenClaw 2026.6.10.
openclaw/openclaw#96239	message-delivery	runtime	3	-3	Gateway reply ordering can surface the tool warning instead of the real answer on 2026.6.10-beta/current 6.10 path.
openclaw/openclaw#96452	codex	tool-specific	2	-2	Codex Computer Use get_app_state can hang for 120s awaiting tools/call on a 2026.6.10 build.
openclaw/openclaw#96441	provider-routing	tool-specific	2	-2	Ollama Cloud/OpenAI-compatible tool_calls can fail on second turn because function.arguments is sent as an object instead of a string in the 2026.6.10 line.
openclaw/openclaw#96600	provider-routing	runtime	2	-2	Google Vertex model provider overrides via openclaw.json are not honored on 2026.6.10.
openclaw/openclaw#96665	discord	channel-specific	2	-2	ACP thread-bound session replies can be generated but not delivered back to the Discord thread on 2026.6.10.
openclaw/openclaw#96698	discord	channel-specific	2	-2	Discord second message can fail from a false stale-snapshot caused by skillsSnapshot hydration mismatch on 2026.6.10.
openclaw/openclaw#96716	mcp	tool-specific	2	-2	MCP tool calls with optional null arguments fail via OpenClaw on 2026.6.10 while other MCP clients succeed.
openclaw/openclaw#96732	reasoning	runtime	2	-2	Moonshot/Kimi reasoning_content can leak into user-visible chat output on 2026.6.10.
openclaw/openclaw#96773	runtime	runtime	2	-2	Turn-scoped developer instructions create divergent message prefixes and break DeepSeek KV-cache reuse on 2026.6.10.
openclaw/openclaw#96586	voice	tool-specific	1	-1	voice_call get_status returns found:false for completed calls after in-memory eviction on @openclaw/voice-call 2026.6.10.
openclaw/openclaw#96535	slack	channel-specific	1	-1	Slack outbound replies persist new single-message delivery-mirror sessions instead of reusing the thread session through 2026.6.10.
openclaw/openclaw#96660	ui	other	1	-1	The 2026.6.10 workspace side panel has reproducible root-path, false missing-label, and layout regressions.
openclaw/openclaw#96463	provider-routing	runtime	1	-1	Custom OpenAI-compatible providers report usage as zero with the embedded runner on 2026.6.10.
openclaw/openclaw#96815	provider-routing	runtime	3	-3	OpenClaw Gateway reports a false Codex subscription usage-limit on 2026.6.10 while the same Codex CLI OAuth route succeeds.
openclaw/openclaw#96820	plugin-install	tool-specific	2	-2	ClawHub search/install API returns a Vercel security checkpoint instead of JSON for 2026.6.10 CLI flows.
openclaw/openclaw#96967	ui	other	1	-1	TUI can display the assistant reply twice when a Telegram-triggered 2026.6.10 session reloads history after delivery.
openclaw/openclaw#96834	whatsapp	channel-specific	1	-1	WhatsApp 1:1 inbound image on 2026.6.10 can wedge the main lane for about three minutes before processing.
openclaw/openclaw#96771	provider-routing	runtime	2	-2	Provider unification in 2026.6.10 can leave openai/gpt-5.5 routed to the wrong backend unless provider-level routing is repaired.
openclaw/openclaw#96677	prompt-size	runtime	2	-2	2026.6.10 still injects per-run UUIDs into the Runtime prompt line, rebusting prompt-prefix caching and inflating token use.
openclaw/openclaw#97180	doctor	runtime	3	-3	On 2026.6.10, doctor --fix can silently add and enable the uninstalled codex plugin when openai is enabled, altering runtime routing instead of only repairing the requested plugin drift.
openclaw/openclaw#97241	provider-routing	runtime	3	-3	Amazon Bedrock provider plugin loads on 2026.6.10 but never populates the provider registry, so Bedrock models fail at runtime as Unknown model.
openclaw/openclaw#97088	whatsapp	channel-specific	5	-5	WhatsApp reconnect on 2026.6.10 can send unsolicited pairing messages to unrelated contacts, crossing an external-message trust boundary.
openclaw/openclaw#96836	slack	channel-specific	3	-3	Slack Socket Mode can remain connected after context overflow/compaction on 2026.6.10 while inbound replies wedge silently.
openclaw/openclaw#96835	exec	tool-specific	2	-2	Windows exec on 2026.6.10 pops a visible cmd/PowerShell window for every command, a regression from v2026.6.8 headless behavior.
openclaw/openclaw#97047	usage-accounting	runtime	2	-2	DeepSeek V4 usage.cost reports zero on 2026.6.10, causing Control UI spend/accounting to underreport actual model cost.
openclaw/openclaw#96910	tls	runtime	3	-3	Auto-generated gateway TLS certs on 2026.6.10 lack subjectAltName, breaking strict-TLS CLI clients.
openclaw/openclaw#96878	web-search	tool-specific	3	-3	web_search providers such as SearXNG/Tavily and official plugin install flows are reported broken across the 2026.6.10 line.
openclaw/openclaw#97108	provider-auth	runtime	4	-4	Embedded/channel agent runtime can reject Anthropic subscription setup-token auth on 2026.6.10 with a false extra-usage billing error despite confirmed headroom.
openclaw/openclaw#97060	whatsapp	channel-specific	3	-3	WhatsApp Native Channel connects on 2026.6.10 but registered state remains false, producing a silent channel availability failure.
openclaw/openclaw#97176	provider-routing	runtime	2	-2	Bedrock adaptive-thinking models on 2026.6.10 are hard-capped at 4096 output tokens and ignore model maxTokens/config overrides.
openclaw/openclaw#97186	acp	runtime	4	-4	ACP binding sessions on 2026.6.10 can spawn a process but never deliver systemSent, leaving the session silently failed.
openclaw/openclaw#97237	attachments	tool-specific	2	-2	Inbound PDFs/docs and other non-image attachments are exposed as unresolvable media://inbound URIs on 2026.6.10, so the model cannot read them.
openclaw/openclaw#97178	installer	runtime	4	-4	openclaw gateway install on 2026.6.10 can add a duplicate gui-domain LaunchAgent over an existing system LaunchDaemon, causing restart-storm risk.
openclaw/openclaw#96947	provider-cache	runtime	3	-3	OpenRouter Anthropic cacheWrite drops to zero after a 2026.6.10 upgrade, regressing conversation cache behavior.
openclaw/openclaw#96916	mcp	tool-specific	3	-3	MCP tool calls fail on 2026.6.10 when LLMs serialize array/object parameters as strings, affecting MiMo/Ollama-style routes.
openclaw/openclaw#97287	feishu	channel-specific	2	-2	Feishu invalid-token errors on 2026.6.10 do not trigger automatic token refresh/retry, requiring manual gateway restart to recover.
openclaw/openclaw#97319	exec	tool-specific	1	-1	exec failure formatting on 2026.6.10 makes framework failure labels look like agent-typed shell commands, confusing operators during failures.
openclaw/openclaw#97335	cron-runtime	runtime	4	-4	Cron-triggered isolated sessions on 2026.6.10 can fail LLM requests even when the same fallback model works in normal sessions.
openclaw/openclaw#97347	cron-runtime	runtime	3	-3	Isolated cron agentTurn on 2026.6.10 can lose exec capability when toolsAllow narrows to exec/bash, breaking allowed automation paths.
openclaw/openclaw#97435	line	channel-specific	3	-3	LINE inbound events on 2026.6.10 were delayed roughly 1.5–2 hours before the agent run and later drained, causing stale channel replies.
openclaw/openclaw#97457	webchat	runtime	4	-4	WebChat RPC abort after compaction on 2026.6.10 can lose the original session context and reroute into restart-recovery/new-session state.
openclaw/openclaw#95323	slack	channel-specific	2	-2	Slack DM turns on a clean 2026.6.10 build can render the inbound system event twice, duplicating runtime context into the model-facing turn.
openclaw/openclaw#97055	memory	runtime	3	-3	memorySearch.provider config on 2026.6.10 can be ignored, silently routing embeddings to the wrong Ollama instance.
openclaw/openclaw#97048	provider-routing	runtime	2	-2	Model catalog validation on 2026.6.10 rejects audio/video modalities produced by OpenClaw catalog fetch, dropping multimodal models from routing.

Stats

• Evidence penalty: -147
• Vote penalty: website-calculated
• Score formula: website-calculated from data/settings.json

---

IsItStable.com — score-first release stability evidence