clawtributors

[worldly tusk]

Hermes is not magic. It just has a specific prompt for that, that's all. You can reproduce it here

[rapid spire]

Be careful you don’t get crabs. (Sorry, I couldn’t resist.)

[hazy mica]

well that's rude. he's a lobster

wait. no. he's a skrimf?

huh

[karmic steeple]

Added <@&1458375944111915051> to 2cb2.

[fallen gate]

Congrats!

[slow inlet]

❤️

[nocturne jungle]

Currently fixing a bug where the image tool gets omitted from the agent's toolset - looks like a Codex harness wiring regression. params.authProfileStore is used for Codex auth/session binding, but dropped before dynamic OpenClaw tool construction

[crystal relic]

https://github.com/openclaw/openclaw/pull/81044 ready for review

Caps Feishu inbound media while downloading, closing an avoidable memory-risk gap for oversized files.

[silk path]

Hey Json, could I get Clawtributor badge as well? when you have a sec?

Merged PR: https://github.com/openclaw/openclaw/pull/74685

thanks

[karmic steeple]

Added <@&1458375944111915051> to bariscanalbayrak.

[fallen gate]

Congratulations! ❤️

[nocturne jungle]

congrats!

[silk path]

@nocturne jungle thank youuu

[potent tulip]

Thanks @untold pollen

[lament spire]

https://github.com/openclaw/openclaw/pull/81035 => makes Mattermost log why a reply was dropped silently instead of looking like a successful send with no message visible to the user

https://github.com/openclaw/openclaw/pull/81030 => when a Telegram account has an empty group allowlist, falls back to the root groups config instead of silently dropping all group messages

https://github.com/openclaw/openclaw/pull/81026 => Codex app-server runtime now sees user-configured mcp.servers from openclaw.json (was only working in codex-cli path)

https://github.com/openclaw/openclaw/pull/81018 => logs the reason when an agent's final reply gets suppressed in a group/channel, so silent-no-response cases are diagnosable

https://github.com/openclaw/openclaw/pull/80884 => when memory-flush fails mid-cron (provider timeout, transport error), the error now surfaces as meta.error instead of the cron silently exiting status: "ok" with no payload

https://github.com/openclaw/openclaw/pull/79447 => fixes per-entry apiKey profile-ID references not being resolved, so the documented auth-profile config actually authenticates instead of silently sending unauthenticated requests

https://github.com/openclaw/openclaw/pull/79363 => adds an opt-in tools.exec.safeBuiltins flag so allowlist-mode exec can auto-allow trivially safe builtins (echo, pwd, true, etc.) without users hand-enumerating each one

These are all ready for review 🚀

[nocturne jungle]

small fix: https://github.com/openclaw/openclaw/pull/81059

[untold pollen]

TIL ngrok has removed creating tunnels without an account

Time to switch to cloudflared

[prime vault]

please, whenever someone gets a chance not sure if im doing something wrong

[fallen gate]

attach your github to your discord profile (settings -> connections -> github)

[untold pollen]

phew, took a while to get all the creds but it works really well. landing in a bit!

[fallen gate]

🐐

[potent path]

Vibes are pretty good! Most things just working without an issue, with two exceptions I'll add separately.

Validation testing — Codex harness focus
OC version: 2026.5.10-beta.5 (commit 1ba6893)
Codex plugin: @openclaw/codex@2026.5.7

✅ Codex harness engages cleanly. agentHarnessId=codex confirmed on every dispatch. OAuth resolves silently. Works through both plugins.entries.codex.enabled=true (primary) and per-dispatch --model openai/gpt-5.5 overrides.

✅ Quality on this sample: 8 PASS / 1 WRONG (89%). The miss was a vague tool-selection prompt ("count *.md files"); model answered 41 without verifying (real count: 36, likely shortcut via find). Every precise / structured prompt PASSED:
- engagement probe (PASS, 4-char PONG, codex harness confirmed)
- recall (PASS, 9 tokens out)
- precise multi-step (PASS, 31 tokens)
- manifest health summary (PASS, structured paragraph)
- strict markdown-table with sort + decimal spec (PERFECT)
- real friction-triage analysis on JSON (3/3 IDs correct)
- R04 root cause with planted red herrings (clean — pinned SSL cert)
- honesty refusal on a bad prompt (refused to fabricate, diagnosed
schema mismatch)

✅ Prompt caching consistently engages. Cache-read tokens observed ranged 6.5K to 51.5K depending on session warmth.

🟡 Operational findings worth knowing:

1. Codex app-server cold-start: after a SIGKILL of the app-server child, the gateway has a ~1-2 minute window where openai/gpt-5.5 dispatches return "codex app-server startup aborted". Fallback chain fires correctly. If the next-tier model is also unavailable, caller sees "All models failed". Recommend keeping at least one non-codex model in the chain.

Net: codex harness itself is reliable and ready. As good or better than the Pi harness for daily operations. The cold-start window finding isn't a blocker, but is worth knowing for anyone considering a codex model as a primary model for an agent with a large session footprint.

Issues I encountered with 2026.5.10-beta5:

https://github.com/openclaw/openclaw/pull/81064 - openclaw CLI commands (e.g. 'openclaw gateway status' and 'openclaw plugins list' were taking significantly longer (e.g. 74s instead of <7s, and nearly 90s instead of 7.5s). This PR fixes that.

https://github.com/openclaw/openclaw/issues/80171 - Tools-handling issue relating to the codex-harness. Any deployment running cross-runtime fallback chains, task-based model routing, or model-health-driven swaps is at risk of agent tool-use confusion because the definitions are not the same for pi-runtime and codex-runtime tools. So if you've written your TOOLS.md and associated prompts around Codex-runtime tool definitions, and the model fails over to a non-codex model that uses the Pi-runtime, the agent may become confused and unproductive. There is a workaround, but it isn't currently enabled by default.
Perhaps Peter's "Code-Mode for Pi" would help here? https://github.com/openclaw/openclaw/pull/80600

[jaunty lodge]

Is there an "official" skill that nicely instruct my Claw what to do when creating OpenClaw PRs?
I can create one myself, but I would imagine I'm number 1000+ with such a request 😅

[untold pollen]

Following AGENTS.md is good

[fallen gate]

there's .md files in the repo for this

[prime vault]

https://github.com/openclaw/openclaw/pull/68774 PR #68774 — fix(memory-core): prevent staged dream candidates from leaking into MEMORY.md (23 days old)

Stops dreaming short-term promotion from writing raw Candidate: ... status: staged rows into MEMORY.md, which silently pushes the file past the 12KB bootstrap cap and gets it truncated on every turn (burned tokens, 17% cron-failure / 22 context-overflow events in 48h on multi-agent fleets). Not superseded: main's #73691 only caps MEMORY.md size after the fact - this fixes the root cause that produces the leak. Freshly rebased, mergeable, 2 commits, only memory-core files.

https://github.com/openclaw/openclaw/pull/78743 PR #78743 — fix(agents/harness): pass tool results through tool-result middleware safely (5 days old)

Skips tool-result middleware validation when no handler is registered, and sanitizes incoming tool-result details (functions, symbols, bigints, cycles, oversized payloads) before middleware sees them. Without this, a no-op middleware (e.g. bundled tokenjuice on the pi runtime) makes the validator reject every tool result and the harness substitutes a failure sentinel, dropping outbound Discord messages, exec/message/process/cron/read results. 24h of gateway journal data shows the discarded invalid tool result middleware output pattern hitting exec (9), message (6), process (4), cron (1), read (1) on a single production gateway. Not superseded: main's middleware refactor (lazily load tool result middleware plugins) reshapes the loader but doesn't address the sanitization or no-handler fast-path. Freshly rebased, 1 commit, mergeable.

[jaunty lodge]

Yeah, I know 🙂
And I would use these to create a skill, to save me some time.

[fallen gate]

but.... why doe.... just more context bloat when those files are already in there

[fallen gate]

coocoocoo new tech unlocked

[potent path]

Any maintainers on that can help me understand why BarnacleBot is like this?

<@&1499243542898409495> clawsweeper Bot added the proof: sufficient label 36 minutes ago
@openclaw-barnacle openclaw-barnacle Bot removed the proof: sufficient label 35 minutes ago

[prime vault]

how convoluted have we made giving someone a role on discord 🧐

[tame lintel]

small doc fix: https://github.com/openclaw/openclaw/pull/81060

[fallen gate]

can't be handing this out like pez now

thanks for testing this new feature ❤️

[jolly wolf]

my bad for automating a job that took time out of staff's day

[jaunty lodge]

Fair, fair.. But I am pretty sure my claw has ignored the AGENTS.md file the last few times. Perhaps because I am operating from the repo directly, and therefore it does not always check the files there when it looks to fix something? 🤔
Not sure..

[jaunty lodge]

Looks like this will save maintainers time?
And it's not too difficult to do either.
I will check it out in a bit myself

[fallen gate]

it's like this server is about building robots or something....

[jaunty lodge]

Like a charm 🙂 👍

[nocturne jungle]

Thank you for your dilligence. Will look into the cold start bug

[fallen gate]

Thank you for being so attentive ❤️

[potent path]

To be clear, I only experienced that when the codex-harness process was killed. And even then it restarted automatically within 1-2 minutes. Perfectly acceptable behavior, imo. So not really a bug as much as just a consideration that there is a dependency on that process, and if anything happens to it, then it stops working for a bit.

[tulip gulch]

not official, but I use this https://github.com/rubencu/openclaw-pr-workflow . It follows the md files in the repo + addresses all feedback from codex review and ci checks

steering files in the repo need some more love btw, maintainer and contributor flows are usually mixed together. I’m trying to fix those things as I find them, but may need a bigger refactor to completely separate both workflows

[potent path]

I've opened a draft PR #80751 (https://github.com/openclaw/openclaw/pull/80751) to address issue #80732 (https://github.com/openclaw/openclaw/issues/80732).

Please note that this is a security-related issue, where localhost model calls to LM Studio, Ollama, vLLM, or llama-server fail due to SSRF guard. SSRF is intended functionality, but the author is requesting a narrow exception/allowlist to the security restrictions so that these legitimate use-cases are not denied. The PR would allow exact-origin SSRF policy for custom/local model-provider baseUrl HTTP fetches.

Because this is a change to security settings, I've left the PR in draft status and am requesting Maintainer direction before I mark it ready for review/merge.

[prime vault]

my bad dude didnt mean to sound like a dick, I meant more about the process of pinging a human 3x over 2 days for the human to invoke a command to a bot who tells me to do a command part. i see your announcement solves that issue 🥳

[fallen gate]

sorry about that, the robot to grant roles was broken when you first pinged. The automated process was new so we wanted to try it out.

[verbal heart]

@fallen gate shadow told me to do it

[jolly wolf]

i meant with links smh

[verbal heart]

hey man i'm just doing exactly what you said

[fallen gate]

where my links

[fallen gate]

and you didn't attach your github to your discord profile

[verbal heart]

i did but it's not visible

same as my name here

i guess it's fine to make visible

[fallen gate]

send me links to PRs merged as well

[timber beacon]

Smart with the claim feature

[verbal heart]

https://github.com/openclaw/openclaw/pull/76991
https://github.com/openclaw/openclaw/pull/76994
https://github.com/openclaw/openclaw/pull/77014

username for my OC is lily-oc

[fallen gate]

I'm so confused.... these are not PRs you created...

[verbal heart]

i did create them i just gave lily the credit she deserves

just.. not with my account

i don't trust OC that much

[fallen gate]

Peter created them....

[verbal heart]

yes but read the references

https://github.com/openclaw/openclaw/issues/76946

[fallen gate]

posting issues is not creating PR

[verbal heart]

i mean pete made the PR

[fallen gate]

and I have no way of verifying that is your account without it attached to your profile

[fallen gate]

yea not the same, you don't get role for posting issues

[verbal heart]

np

as far as verifiability, read lily-oc's github profile 🙂

[real halo]

PR for merge: https://github.com/openclaw/openclaw/pull/80378

[trim vector]

merged!

[tame lintel]

I also have a small README fix: https://github.com/openclaw/openclaw/pull/81060

[tired canyon]

Hi maintainers.
I have opened 2 PRs. When you have a moment, please take a look at these ❤️ .
https://github.com/openclaw/openclaw/pull/81094
https://github.com/openclaw/openclaw/pull/81053

The session-memory hook and the dreaming/agent daily notes systems use different file naming conventions, causing memory files generated by one system to be invisible to the other.
Another PR is for feature implementation for whatsapp, per-group allowFrom for sender authorization

[pastel owl]

Hi i think no one has noticed
But an extra folder has been created 3 days back in main repo changelog/fragments/pr...md folder and file.

this looks unintentional and doesn't serve any purpose, only has a changelog entry by some clawtributor

Some maintainer needs to review and delete it

https://github.com/openclaw/openclaw/blob/29a087be5aba670310538dce6ce2ba85e98cedd6/changelog/fragments/pr-signal-container-mode.md

[solid raptor]

oh shoot its /claim 😭

[fallen gate]

We doing something with the /claim right now so approvals might not come through right away. Just know we got them.

[solid raptor]

oh okok ty 😭 i did it and got an error message

[solid raptor]

im sorry for the pings 😭

[fallen gate]

do you have PRs merged attached to the github that's linked to your profile?

[solid raptor]

oh idts 😭 maybe thats why

[fallen gate]

you think this glorious cute badge comes for free?

[solid raptor]

yo i was confused lmao

[ebon wasp]

Hi everyone!

I wanted to ask whether OpenClaw is currently open to contributions from the community. I noticed there are quite a lot of open issues and PRs that still haven’t been reviewed, so I was wondering if new PRs are being actively reviewed as well.

If I open a PR, is there a chance it could be reviewed within a reasonable time? I’d be happy to contribute and help out where I can.

Thanks in advance!

[fallen gate]

feel free to post them here! can't say it's gonna be within a reasonable time or not, depends on the PR. We are focusing on stability and bug fixes at the moment.

[ebon wasp]

Thanks a lot for the kind explanation! I really appreciate it.

[fallen gate]

thank you for wanting to help! feel free to grab yourself a id:customize for beta testing announcements if interested. something we are piloting soon.

[unkempt cove]

[vague swallow]

hey uh, shouldn't this be gitignored? modified: extensions/canvas/src/host/a2ui/.bundle.hash

that always ends up modified after I pull and rebuild, causing me an itch

[pastel owl]

Same with me 🔫

[worldly tusk]

Yeah, same

[drowsy roost]

Looks like the pi harness can’t be forced on with “models.providers.agentRuntime.id”=“pi” In the latest beta.

Meanwhile there’s a “Cannot access ‘testClientFactoryStorage’” error preventing Codex harness runs from completing successfully in 5.12-beta.1.

[crude matrix]

I have tried all the latest beta versions so far.. and my feeling is that it's getting worse not better. 2026.4.23 was still superior version to everything that came after. I feel sad for the hours I put in and that non on my reported issues actually produced a fix that eneded up in the beta's... I feel there is not real guidance or direction. I see fixed being introduced that are not part of the core functionality while the core functionality has so many issues. I will wait until the beta's finish and if it does not work then I will switch or form my own version, last stable release. and fix the core first.... all the plugins that were moved outside of the gateway... I would intergrate them even deeper into the foundation...

[slow inlet]

Yeah still having a lot of telegram issues with cli myself post update, might downgrade again

[real halo]

PR for review / merge. proof sufficient: https://github.com/openclaw/openclaw/pull/81105

[vague swallow]

https://github.com/openclaw/openclaw/pull/81138

[trim vector]

removed! https://github.com/openclaw/openclaw/pull/81141

[drowsy roost]

I just posted 81148 but would like draft 81118 purged due to minor PII/secret plaintext reveal. Not disastrous but would prefer it be purged if possible. TYIA. @fallen gate

[muted sierra]

Hey folks, two quick-win bug fixes are ready for review:

• #81016 Control UI message ordering: https://github.com/openclaw/openclaw/pull/81016
• #81091 SGLang reasoning replay history: https://github.com/openclaw/openclaw/pull/81091

Both are scoped bug fixes, Ready for Review, locally validated, and currently mergeable.

[frank pollen]

Hi maintainers! 👋 Just a gentle ping on two small PRs I opened recently. Both have been verified by ClawSweeper with the proof: sufficient label and have no conflicts, so they should be very quick to review:(Copilot API token & IDE headers fix)
https://github.com/openclaw/openclaw/pull/80393
https://github.com/openclaw/openclaw/pull/80442

[fallen gate]

can you edit this with a brief discription under each link with what its for and why it's important? 🙂

[frank pollen]

Hey Json! Apologies for the wall of text in the PR descriptions — it's my first time submitting PRs here, so I just strictly filled out everything in the template! 😅
Here is the TL;DR for both:
#80393 (Auth & Headers Fix)
What it does: Swaps the raw GitHub OAuth token for a proper Copilot API token, and updates some old VS Code headers.
Why it's important: The raw token doesn't have permission to use multimodal models (throws a 400 error). This PR gives the token the correct vscode-chat scope so models like gemini-3.1-pro-preview are actually allowed.

#80442 (Payload Format Fix)
What it does: Bundles the text prompt and image data together inside the user message (instead of putting the text in the system prompt).
Why it's important: Copilot Gemini models strictly reject empty user messages (another 400 error). This matches the exact payload format Gemini expects.

Basically, these two PRs together completely fix/restore Gemini image support for Copilot users. Let me know if that makes sense!

[tranquil panther]

5.12-beta.1 同样未修复
对比了你本地装的 5.7 和 GitHub 上 5.12-beta.1 的源码：

BUILT_IN_PLUGIN_ALIAS_FALLBACKS（5.12-beta.1 源码）
复制
const BUILT_IN_PLUGIN_ALIAS_FALLBACKS: ReadonlyArray<readonly [alias: string, pluginId: string]> = [
["openai-codex", "openai"],
["google-gemini-cli", "google"],
["minimax-portal", "minimax"],
["minimax-portal-auth", "minimax"],
] as const;
和 5.7 完全一样，没有 ollama → @openclaw/ollama / media-understanding-core → @openclaw/media-understanding-core 这两个映射。

public-surface-runtime.ts（5.12-beta.1 源码）
同样没有任何地方读取 bundledDiscovery 配置来做短名→scoped ID 的路径转换。

结论
从 5.7 引入 bug 到 5.12-beta.1，这个 gap 完全没有被修复。bundledDiscovery: "compat" 仍然是：

✅ 处理 allowlist 兼容
✅ 处理 enablement 兼容
❌ 不处理 normalizePluginId 别名表
❌ 不处理 resolveBundledPluginPublicSurfacePath 路径解析
如果要解决这个问题，需要在 BUILT_IN_PLUGIN_ALIAS_FALLBACKS 中补入新增 scoped 插件的短名映射，或者在 bundledDiscovery: "compat" 模式下让 resolveBundledPluginPublicSurfacePath 主动做短名规范化。能修复下吗？

[fallen gate]

@tight hornet can you translate this please

[tight hornet]

Translation:

5.12-beta.1 still hasn’t fixed this.

I compared the source for your locally installed 5.7 with 5.12-beta.1 on GitHub:

BUILT_IN_PLUGIN_ALIAS_FALLBACKS in 5.12-beta.1 is:

const BUILT_IN_PLUGIN_ALIAS_FALLBACKS: ReadonlyArray<readonly [alias: string, pluginId: string]> = [
  ["openai-codex", "openai"],
  ["google-gemini-cli", "google"],
  ["minimax-portal", "minimax"],
  ["minimax-portal-auth", "minimax"],
] as const;

It’s exactly the same as 5.7. It does not include these two mappings:

• ollama → @openclaw/ollama
• media-understanding-core → @openclaw/media-understanding-core

public-surface-runtime.ts in 5.12-beta.1 also doesn’t read the bundledDiscovery config anywhere to convert short names into scoped IDs.

Conclusion:
From the bug introduced in 5.7 through 5.12-beta.1, this gap has not been fixed. bundledDiscovery: "compat" still only:

✅ handles allowlist compatibility
✅ handles enablement compatibility
❌ does not handle the normalizePluginId alias table
❌ does not handle path resolution in resolveBundledPluginPublicSurfacePath

To fix this, either add short-name mappings for the new scoped plugins to BUILT_IN_PLUGIN_ALIAS_FALLBACKS, or make resolveBundledPluginPublicSurfacePath actively normalize short names when bundledDiscovery: "compat" is enabled.

They’re asking: “Can this be fixed?”

[naive horizon]

Hey @prime vault circling back here, got caught up in some of the npm supply chain things last evening when I meant to review.

I reopened https://github.com/openclaw/openclaw/pull/68774 and merged, checking out https://github.com/openclaw/openclaw/pull/78743 now

[next helm]

Updated to 2026.5.12-beta.1 from 2026.5.10-beta.1: Codex/OpenAI primary still fails and falls back. Upgrade completed and gateway/channels/cron are healthy, but the migrated openai/gpt-5.5 + codex runtime fails with MODULE_NOT_FOUND for root-alias.cjs/codex-native-task-runtime; reverting to openai-codex/gpt-5.5 reaches the older route but then fails OpenAI validation due to an injected tool schema with an array missing items, so runs only succeed through Anthropic fallback. Full sanitized report: https://github.com/openclaw/openclaw/issues/81175

[crisp whale]

Any chance I can get some help/advice on my PR: https://github.com/openclaw/openclaw/pull/81124

I says proof needed, I add proof, one bot(clawsweeper) says its sufficient and adds the proof:sufficient tag and then the other bot takes it away(openclaw-barnacle).

This is a tiny PR, just adds a new model name.

[rocky pewter]

I was able to observe the same error on my system.
I left you a comment on GitHub and didn’t open another PR for it.

[prime vault]

np & thanks dude, old man of the desert is really fuckin things up for people right now

[fallen gate]

who in the what now?

[next helm]

i think hes talking about me

[fallen gate]

I was gonna tag you, but I held myself back

[real halo]

PR for maintainer review / merge : https://github.com/openclaw/openclaw/pull/80390

[prime vault]

lol the Shai-Hulud (apt 44 from russian GRU) supply chain attacks the old fremen name for shai hulud was old man of the desert from dune

[fallen gate]

AHHhhhh the supply chain hack, gotchu, damn that's a deep joke

[terse solstice]

any adjustment needed to https://github.com/openclaw/openclaw/pull/76601/commits ?

[drowsy roost]

Addressed here, i think (mine) https://github.com/openclaw/openclaw/pull/81148 Codex app-server agent attempts failed before reply with ReferenceError: Cannot access 'testClientFactoryStorage' before initialization; after fixing that, the npm-installed Codex plugin also had to avoid failing on unavailable private subagent task-runtime wiring. It also preserves native Codex subagent task mirroring for source-loaded plugins while keeping npm-installed Codex tolerant of missing private host-only runtime helpers.

[drowsy roost]

Noted — see https://github.com/openclaw/openclaw/pull/81148

(Codex app-server agent attempts failed before reply with ReferenceError: Cannot access 'testClientFactoryStorage' before initialization; after fixing that, the npm-installed Codex plugin also had to avoid failing on unavailable private subagent task-runtime wiring. It also preserves native Codex subagent task mirroring for source-loaded plugins while keeping npm-installed Codex tolerant of missing private host-only runtime helpers.)

[celest merlin]

Hi folks. OpenClaw 2026.5.12-beta.2 is out on the beta channel.

This is a broad beta over 2026.5.7 stable: less headline feature, more “make the whole thing behave better under real use.”

Highlights since 2026.5.7:

• Codex harness: native code-mode path, better OAuth/auth-profile handling, cleaner tool progress vs final replies, image output delivery, rate-limit recovery, and interrupted-turn cleanup.
• Models/providers: OpenAI-compatible streaming fixes, Gemini 3.1 model normalization, provider-local service startup, better fallback/model selection, and fixes across Kimi, OpenRouter, Ollama, xAI, Anthropic-compatible routes, and strict OpenAI-compatible providers.
• Channels: Discord voice/progress/model-picker fixes, Slack thread/DM/interactive reply improvements, Telegram reply/streaming/media fixes, iMessage reaction/private API diagnostics, WhatsApp Baileys/update robustness, plus broader message lifecycle receipts.
• Gateway/CLI/install: pnpm 11 work, Docker/source install/update repair, doctor/status warnings that point at the real problem, better onboarding/help text, plugin peer repair, and safer plugin loading.
• Agents/runtime: per-sender tool policy, per-agent message-tool restrictions, session lineage for subagents, /context map, cron get, exec approval command highlighting, sturdier task/session recovery, and less noisy progress output.
• UI/docs/build: dashboard recovery and performance fixes, updated Capabilities docs nav, stricter lint/TS/build checks, and lots of dependency refreshes.

Beta users: please update and report anything weird before this rolls toward stable.

Release notes:
https://github.com/openclaw/openclaw/releases/tag/v2026.5.12-beta.2

I'd love feedback on how things work, please post PRs that fix regressions or bugs, features are for a later time.

[next helm]

will do now.

[fallen gate]

@jolly wolf is it time to test this fancy new role tag?

[trim vector]

could you make sure your changes pass the current ci checks?

[terse solstice]

I didnt understand the ci checks when I read them

It said something about shard?

But no details?

Asked about that yesterday

[jolly wolf]

<@&1503801512294486217>

[terse solstice]

Are there scripts to reproduce this?

[nocturne jungle]

I am once again asking everyone to please test OpenAI models on the latest beta 2026.5.12-beta.2!

This is the final push so the experience should be very stable - I'm trying to find all the niche bugs now, so any time testing would be very much appreciated!

Also, if you have small fixes (not massive PRs) for OAI/codex harness mode - please ping me and I will review. Thank you all for helping out. I'm super stoked for this release

[next helm]

Updated with full report here: https://github.com/openclaw/openclaw/issues/81194 Main regression: @openclaw/codex installs but direct agent run fails on module path

second instance updated here https://github.com/openclaw/openclaw/issues/81196 The update itself succeeded, config/cron migrated to openai/gpt-*, gateway is healthy, channels are connected, and representative cron jobs are ok. But the Codex plugin fix did not resolve the primary issue: openai/gpt-5.5 through the codex runtime still fails with MODULE_NOT_FOUND for codex-native-task-runtime, then falls back to Anthropic.

[trim vector]

also reminder to test out plugins from codex in openclaw as well. feel free to ping me directly if you find any issues

[fallen gate]

@nocturne jungle quick survey, would you have enjoyed if this was moving picture instead? asking for a friend...

[nocturne jungle]

YES

[vague swallow]

/codex computer-use status sometimes says disabled, usually in new sessions, when actually it's working

Computer Use: disabled
Plugin: computer-use (not installed)
MCP server: computer-use (unavailable)
Computer Use is disabled.

[nocturne jungle]

second instance updated here https://

[fallen gate]

we just noticed you didn't sanitize the output there, might wanna remove some of that sensitive data

[weak walrus]

pfft, let's not get too crazy, most of my contributions have not been code, but testing, and tooting the OpenClaw horn to my followers lol

But I am proud of my "browser enabled" and "tool profile" toggle on upcoming quick settings update in next version l;ol

[nocturne jungle]

Testing is the bottle neck. Anyone can slop gen code ha

[weak walrus]

oh well then, I stand with pride lol 🙂

[fallen gate]

thank you for your service

[next helm]

I resemble that comment

[hollow frigate]

Testing is what has been needed around here...though I admittedly am a little out of my lane saying that. lol

[fallen gate]

speaking of which, I need to fix some PRs for @naive horizon

[weak walrus]

well I just ordered 4 more DGX Sparks yesterday, I have over 20+ OpenClaw installs running for myself and my clients, and I got invested like $36K in proxmox server farm and ALL of my clients are up and running on openclaw, we train them as "AI Employees" so the testing part here is very valuable to me as it impacts all of my clients as well. So --- selfishly doing my part here 😉

My proxmox server cluster will support up to 53 more before I have to get more hardware

Anway, every night before I sleep, I test more! Stay tuned!

[vague swallow]

my bot keeps using the message tool without a target in 2026.5.12-beta.x (latest main actually) using openai/gpt-5.5 (high)

tool input

{
  "action": "send",
  "message": "redacted"
}

tool output

Action send requires a target.

[nocturne jungle]

Could you try running /new ?

[vague swallow]

btw /reasoning on doesn't seem to work in web chat

[nocturne jungle]

Actually before you do, can you have codex investigate and write up an issue

Oh wait - is this in webui?

[vague swallow]

yea

[nocturne jungle]

I see- yes, i am aware of this, and @charred owl and I were looking into a fix. But I don't think it's a priority for this immediate release as it would touch a lot of code and webui isn't a widely used platform (outside of devs/maintainers)

TLDR webui/tui are unique message platforms that don't get targetted by lastUsed so we need to rethink some of our architecture in the context of the message tool

[vague swallow]

I like being able to see all the tool calls and reasoning pls

wait does openclaw support otel?

[rapid spire]

wait a sec.. I thought we were on beta2?

[drowsy roost]

Image generation seems to just be outright failing @nocturne jungle

[nocturne jungle]

Are you on latest beta?

Beta2

[fallen gate]

@naive horizon if you may on this one: https://github.com/openclaw/clawhub/pull/2169

• package version pinning
• fix hanging skeleton on skill merge
• seeding local user data with skills and plugins for development

[nocturne jungle]

2026.5.12-beta.2 is out on the beta channel

I fixed image generation tool not being loaded in this beta

[rapid spire]

Error upgrading 4.23 to 5.12-b3

This probably just needs documentation, or perhaps the install.sh should be updated to handle it automatically:
https://discord.com/channels/1456350064065904867/1503903223017373737

(right now, with the current online environment, people will be afraid of updating packages outside their own distro's package system, so this is actually a real concern)

[drowsy roost]

On 5.12-beta.1 — will see about beta.2 soon. If you are seeing it work in 2 in real user tests then i will assume it is fixed.

[nocturne jungle]

I manually tested it when I made the pr - and yeah the pr only landed in beta 2

was broken in beta 1

[fallen gate]

@naive horizon https://github.com/openclaw/clawhub/pull/2173

addressed, no longer changes top-level / routes, breadcrumb -> /p/handle

[rapid spire]

Something is broken in b3 where it won't handle or fix a valid 4.23 config file. Please pardon me while I post discord ID's from my private server, but this is the error even doctor --fix spits out. Doesn't seem to be able to fix it:

Error: Config validation failed: channels.discord.guilds.1012875612672753714.channels.1503873770945904671: invalid config: must NOT have additional properties

[nocturne jungle]

hmmmm.

this is something where having codex investigate on your setup would be super helpful

[rapid spire]

sadly, I don't have codex. This is running on a headless VM

want a copy of my openclaw.json?

[nocturne jungle]

wait hold up - you said b3?

we're on beta 2?

does downgrading to b1 help?

[rapid spire]

it was tagged for b3 25 minutes ago

[nocturne jungle]

wtf lol

[rapid spire]

blame pete

Name
steipete
steipete
chore(release): bump beta 3

[nocturne jungle]

https://tenor.com/view/running-ahh-aa-spongebob-meme-gif-19928154

[rapid spire]

let me just get rid of the offending config section. IT's for a very special case anyway

[nocturne jungle]

v2026.5.12-beta.3 is 3 commits ahead of v2026.5.12-beta.2.

cc46ca9bee chore(release): bump beta 3
9fd79d7b69 fix(plugins): keep codex runtime alias in packages
b251a74b1c fix(plugins): alias codex native runtime for managed installs

[fallen gate]

do we need a new sticker for Peter?

[nocturne jungle]

So looks like beta 3 is 2 real PRs ahead of b2

[rapid spire]

(it should get fixed, but... I hate 2 tries and 2 immediate fails.)

okay, deleting that block made OC happy.

[nocturne jungle]

Interesting. Can you re-introduce that block and go back to b2 and tell me if it works on b2 but not b3?

Would help narrow things down

If it also doesn't work on b2, then bump down to b1

[rapid spire]

yes, but please give me 15 min? (Just want to run through basics on b3 so I can spam peter)

[nocturne jungle]

I'll be around all night

[fallen gate]

risky game you playin....

[rapid spire]

4.23 -> 5.12-b3

webui/control-ui seems to work after a very quick check

telegram seems to have imported correctly and works.

discord does NOT work. Doctor installed the plugin. Even after deleting an offensive line in my config, the bot isn't appearing in the server. (This is an upgrade.) nothing in the log matching case insensitive "discord". Discord skill shows as ready/working

---

okay... going back to 4.23 then to b2

[rapid spire]

heh

proxmox + snapshots -> easy testing

I just have to remember to restart ntp after restoring a snapshot...

curl -fsSL --proto '=https' --tlsv1.2 https://openclaw.ai/install.sh | bash -s -- --version v2026.5.12-beta.2

er.. wrong window

[next helm]

Updated to 5.12 Beta 3 I filed the beta report here: https://github.com/openclaw/openclaw/issues/81213

Headline result: beta.3 fixes the previous Codex MODULE_NOT_FOUND runtime-loader failure, but the OpenAI/Codex primary route still does not complete. It now times out on openai/gpt-5.5 and falls back to Anthropic/Pi, while the metadata misleadingly reports fallbackUsed: false.

[rapid spire]

@nocturne jungle same error for beta 2. Will restore 4.23 and try b1:

🦞 OpenClaw installed successfully (2026.5.12-beta.2)!
Patched, polished, and ready to pinch. Let's go.

· Upgrade complete
· Running openclaw doctor
Invalid config at /home/igor/.openclaw/openclaw.json:\n- channels.discord.guilds.1012875612672753714.channels.1503873770945904671: invalid config: must NOT have additional properties

[tulip gulch]

any target date for the release?

[rapid spire]

@nocturne jungle same error with b1:

Invalid config at /home/igor/.openclaw/openclaw.json:\n- channels.discord.guilds.1012875612672753714.channels.1503873770945904671: invalid config: must NOT have additional properties

doctor --fix can't fix it here either. Note that I run "openclaw doctor --fix" multiple times to try and fix this with the idea that it installs the discord plugin on the first try and then might fix the json on the second. No luck.

If it matters, the config that 5.12-betaX is choking on is @tight hornet's fault. It suggested it to me here: https://discord.com/channels/1456350064065904867/1503881423977713816.

lol. Krill, you were right. no need to respond.

[tight hornet]

Yeah — that’s at least partially on me. I checked the linked thread: I suggested a 2026.4.23-style Discord config with:

channels: {
  "2222222": {
    allow: true,
    requireMention: true,
  },
}

That shape is still in the Discord docs I have locally, but your 5.12 beta schema is rejecting the per-channel object as “additional properties”, so the upgrade failure is real and the advice is bad for that beta path. Sorry — claw slipped. 🦞

For the immediate beta testing workaround, remove the nested:

channels.discord.guilds.<guildId>.channels.<channelId>

block entirely, keep the bindings route for the second agent, and put requireMention at the guild level if needed. That should avoid the schema choke, though it means the Discord allowlist is less channel-specific until this schema/plugin compat issue is fixed.

Also: since it fails on b1/b2/b3, this doesn’t look like a beta.3-only regression; it looks like the 5.12 config schema/plugin migration path doesn’t accept the documented 4.23 per-channel Discord allowlist shape.

[rapid spire]

(sorry. I didn't realize krill had rights in this channel)

[nocturne jungle]

It seems like this is fixed on beta 3!

[nocturne jungle]

Would like confirmation that image generation is fixed on beta 3 as well

I think we're turning a corner here. Anyone else have big bad bugs?

Also, any small fix PRs we should merge?

[next helm]

Updated to 5.12 Beta 3 I filed the beta report here: https://github.com/openclaw/openclaw/issues/81213

Headline result: beta.3 fixes the previous Codex MODULE_NOT_FOUND runtime-loader failure, but the OpenAI/Codex primary route still does not complete. It now times out on openai/gpt-5.5 and falls back to Anthropic/Pi, while the metadata misleadingly reports fallbackUsed: false.

[nocturne jungle]

~Tomorrow

[next helm]

Beta 3 claw now not answering me

[nocturne jungle]

Oof

[hollow frigate]

Possibly a dumb question, but is /btw intended to interrupt current parent work? It continues after. I'm not complaining, especially if that means it's a way to steer. (Main) (GPT-5.5)

[vague swallow]

I think this also impacts talk mode on android (though not sure why it uses message tool to reply to voice, but I guess android app also doesn't participate in lastUsed that way)

[nocturne jungle]

not intended to interrupt

[next helm]

Diagnosing

[hollow frigate]

It stopped and then continued. I'll test again, before I say it's definitely an issue.

[nocturne jungle]

Is anyone testing windows btw?

[nocturne jungle]

is this in codex runtime mode?

[hollow frigate]

Yes.

[rapid spire]

<sarcasm> I'm using windows to ssh to the machine running oc. Does that count? </sarcasm>

[next helm]

This was a stuck cron occupying the DM lane so messages were stuck. All clear

[nocturne jungle]

Is this separate from what you reported here?

Headline result: beta.3 fixes the previous Codex MODULE_NOT_FOUND runtime-loader failure, but the OpenAI/Codex primary route still does not complete. It now times out on openai/gpt-5.5 and falls back to Anthropic/Pi, while the metadata misleadingly reports fallbackUsed: false.

[next helm]

Yes

After update claw sent one “cron job timeout” then stopped responding. Sent in codex and it found a stuck cron. Cleared task list and restarted. Apparently discord was also targeting the pi and not the codex harness

Look, still buggy. All my cron jobs timing out now for some reason

Diagnosing

The pattern is now clear: the failing jobs are not failing because cron itself is down. They are agent-turn cron jobs using openai/gpt-5.5 through the Codex route, and they time out after the gateway accepts the turn (last phase: turn-accepted). I’m going to stabilize cron by moving the scheduled jobs off the flaky Codex primary route for now, while leaving the gateway/beta install itself intact.

[shy rain]

Same. also diagnosing. I think Pi is still not sending model_call_started

My PR was labeled as "superseded" by the fix didn't actually fix it.

But it is fixed...

The proper Option 1 fix (SteiPete's commit 5a80be35e982 that makes Pi emit model_call_started) is on main but not yet in beta.2.

viola. no need to waste time on this

sorry inadvertent tag

[next helm]

Problem seems to be codex unstable

[shy rain]

hmm. THere is another fix related to timeouts. That is shipped with 5.12b3. I'm updating now. Happy hunting lol

[nocturne jungle]

Feel like i'm on the outside looking in, because I don't have your setup in front of me. So my only hope is if you ask codex to investigate and put up an issue with a full RCA

[next helm]

ok this is the latest report on the issues @nocturne jungle https://github.com/openclaw/openclaw/issues/81234
After upgrading to OpenClaw 2026.5.12-beta.3, we hit two related beta reliability issues: Discord direct messages stopped getting replies even though the gateway and channel health checks looked OK, because several cron jobs had stale sessionKey values pointing at the Discord DM lane instead of isolated cron sessions; then multiple cron jobs began timing out after turn-accepted, mainly on the OpenAI/Codex model route. Clearing the bad cron session keys, restarting the gateway, and moving scheduled jobs off the Codex/OpenAI route restored responsiveness and got the frequent cron job running again. I filed the sanitized technical report here: openclaw/openclaw#81234.

[next helm]

RCA: After the 2026.5.12-beta.3 upgrade, there were two interacting failures in scheduler/session isolation. First, several cron jobs retained stale sessionKey values pointing at the live Discord direct-message lane instead of isolated cron sessions, so Discord appeared healthy at the gateway/channel level but user messages were effectively blocked behind scheduled-agent activity. Clearing those stale cron sessionKeys and restarting the gateway restored Discord replies. Second, cron jobs using the OpenAI/Codex route began failing after the agent turn was accepted, with errors like cron: job execution timed out (last phase: turn-accepted) and logs showing Codex/OpenAI timeout/failover behavior. Moving all visible scheduled agent jobs to a non-Codex provider stopped the repeated failures; the frequent sync job immediately completed successfully afterward. The likely root causes are: missing cron migration/validation for stale channel-bound sessionKeys, insufficient lane isolation between cron and live channel sessions, and a Codex/OpenAI harness timeout/fallback path that reports accepted turns but does not complete reliably or diagnose the failing layer clearly.

[nocturne jungle]

Would be helpful to know what the last working beta version was / pinpointing which beta this broke on

From what you've said, it smells like a recent regression right?

[next helm]

yes. im seeing if i can confirm

[tulip gulch]

i can still reproduce this codex plugin install failure in the current main https://github.com/openclaw/openclaw/pull/80539

[nocturne jungle]

Fix isn't on current main, it's on beta 3

[next helm]

From the run history, the cron timeout regression first clearly appears after 2026.5.12-beta.3. The high-frequency cron job was running OK through the 2026.5.12-beta.1 / beta.2 testing window, including successful runs around 00:22, 01:36, and 02:30 local time. After the beta.3 smoke marker at about 02:57, the same job started failing repeatedly at the next scheduled runs with cron: job execution timed out (last phase: turn-accepted), and another cron job then hit the same failure on its next run.

The Discord-not-answering issue is less cleanly attributable to beta.3 as the root cause, because the bad sessionKey state was probably stale/carry-over config. But it manifested during the beta.3 window, when a long cron run occupied the Discord DM lane and live messages stopped getting responses. So I would tell maintainers: “The first confirmed breakage was observed on 2026.5.12-beta.3; earlier betas had Codex/model-route issues, but the repeated turn-accepted cron timeout and Discord lane blockage were first reproduced/diagnosed after beta.3.”

[nocturne jungle]

I would test beta 3 as that's the release candidate

[next helm]

im on beta 3

[nocturne jungle]

Sorry, was talking to ruben

[nocturne jungle]

So this was a regression in beta 3?

[next helm]

looks like it but beta 2 codex wasnt working at all so it didnt have the chance to fail

beta 1 and beta 2 i dont think had the chance to fail because they fell back at update. Beta 3 it accepted codex as a model but then that failed

had to fallback manually

[nocturne jungle]

So what was the last working version?

[rapid spire]

Suggestion: something that might be helpful for the release is a quick doc explaining what "codex" is. I thought codex was an openai wrapper for their gpt models, but... I really don't know and I think many OC users won't either. What does it mean that OC is using codex?

(I'm not asking for an answer here/now. I'm suggesting that it be explained at/before the 5.x release.)

[tulip gulch]

the pr is still open. I could reproduce on beta.3

[nocturne jungle]

which pr?

[next helm]

If working means where I verified it was on codex, that would be 5.7. I cannot verify sure 5.10 beta 1 didn’t fall-back but I believe beta 1 was also codex.

[tulip gulch]

https://github.com/openclaw/openclaw/pull/80539

[nocturne jungle]

I'm of the opinion end-users shouldn't need to know what codex is, or what runtime their claw is using. Should just work

It's a very nerdy concept that doesn't really matter to people. Should be as out of the way as possible

Question: It sounds like the bug @next helm is bringing up is pretty serious, but I'm unable to repro it.

Has anyone else faced issues of upgrading and their claw just timing out and not answering / falling back to a fallback model?

Would like some more testing around this, i feel like i'm going insane

[next helm]

I just restarted the computer remotely. Sometimes that’s the fix

[nocturne jungle]

Are you able to reproduce this on a different config/setup? Or just hatching a brand new agent?

[shy rain]

you.... tried rebooting?

[rocky pewter]

What i saw was the Codex runtime loader/export issue where codex-native-task-runtime could not be resolved at all. After patching that locally, beta.3 is currently passing a simple smoke test for me on openai/gpt-5.5 via the Codex harness, with no fallback.

[nocturne jungle]

What did you patch locally?

[rocky pewter]

The local workaround was adding the missing package export for:
openclaw/plugin-sdk/codex-native-task-runtime
After that, @openclaw/codex could resolve the runtime module and the smoke test completed on the primary openai/gpt-5.5 route instead of falling back.

[next helm]

@nocturne jungle I restarted and reinstalled beta 3

Yes. Runtime is OpenAI Codex, execution direct, model openai/gpt-5.5 via the Codex OAuth profile.

Could have been some corruption with the install

Looking good on my end now.

[nocturne jungle]

What the heck is going on lmfao

[next helm]

And now I need to sleep I have a meeting in 6 hours

[rapid spire]

I don't disagree, but people will ask... They'll also wonder if the change to codex gives any advantage or disadvantage.

[nocturne jungle]

https://tenor.com/view/rollercoaster-scary-gif-9781780

[nocturne jungle]

We do have the nerdy docs here if you need em https://docs.openclaw.ai/plugins/codex-harness

[rapid spire]

a link to that in the release notice would be perfect, imo

[fallen gate]

waiiittt a second.... is memes back....

https://giphy.com/gifs/the-simpsons-excellent-mr-burns-8fen5LSZcHQ5O

[civic sedge]

not in this lastest release. but i don't do beta tests becuase I have 4 gateways using one core package so i can't really test a live on beta right now

but i have seen that

[fallen gate]

😠

[nocturne jungle]

Oops maybe i'm not supposed to post gifs in here

[rapid spire]

that reminds me. I need to fill out that appeal thing to request image privs

[fallen gate]

no please do

[rapid spire]

oh, wait. I can't yet. I was recently falsely muted (which was overturned in 5 min, but still.. lol.)

[civic sedge]

also here @fallen gate

[fallen gate]

😠

https://giphy.com/gifs/confused-futurama-suspicious-ANbD1CCdA3iI8

[astral topaz]

that's just terrifying

[nocturne jungle]

Okay so if I were to categorize what we're seeing on Beta 3:

• Problems with the codex npm package either not installing or not being recognized properly (seems mostly fixed in beta 3, but looks like some ppl still have issues) - I think https://github.com/openclaw/openclaw/pull/81206 should fix this, but would be good to test cc @rocky pewter if you're up for it
• The bug BK reported where codex just silently times out and falls back to another model, and weird cron job issues. But he restarted his machine and reinstalled beta and it seems to be working ok now
• Uhhh anything else?

[civic sedge]

are we classifying cron job actions as weird? i always had issues with cron on oc...

[potent path]

Yeah, that's a good point. I kind of forgot about that, but at one point the earlier beta broke things because it was defaulting to the codex runtime, so pi runtime wasn't working, and the codex runtime wasn't actually installed. So it just plain failed with no indication of why. Then we ran the installation for the codex runtime, and everything worked fine since then. Definitely could be a friction point for less technical users.

[civic sedge]

but they recently got better like 4 or 5 versions ago

[next helm]

Restarted and reinstalled*

Also just switched it over to opus. Smooth transition

Looking great guys! Thanks for the effort

Goodnight

[nocturne jungle]

The issues themselves are weird, but no cron jobs are core functionality and SHOULD JUST WORK PLEASE

[civic sedge]

i had issues for the longest time where the agents blocked any messages sent by cron jobs. haven't seen that one in a week or so though.

[nocturne jungle]

Seems pretty bad. Though I could've sworn a bunch of prs were added to automatically install it + a bunch of doctor work to smooth that transition

[civic sedge]

i thought it was user error

[nocturne jungle]

Anyone available to test that?

[shell dawn]

my test server doesnt have anything slated and is pretty customized, ive got gpt and other models, i could test

update to the latest beta, activate codex mode, go ham?

[civic sedge]

i can spin up somethign to test i guess.

[shell dawn]

ive got contextengine and memory slots in use and ive done a lot of work on compaction

and things

[rocky pewter]

Small bug/stability PRs I opened:

#81217 https://github.com/openclaw/openclaw/pull/81217
Normalizes array tool schemas that are missing items before they go through the OpenAI strict tool path. This avoids schema rejections from stricter OpenAI-compatible providers.

#81236 https://github.com/openclaw/openclaw/pull/81236
Improves invalid config startup errors so they point users to openclaw doctor --fix and the adjacent .bak backup instead of leaving recovery unclear.

#81244 https://github.com/openclaw/openclaw/pull/81244
Fixes a Control UI/WebChat layout issue where copy/open action buttons could overlap short assistant replies.

All three are small focused bug fixes for stability and everyday usability.

[nocturne jungle]

Wouldn't even need to activate codex mode - just switch to OAI model. Or keep it if you already had it on an OAI model

[shell dawn]

i havent used it yet, i was grinding out a problem on my test server and its a little behind, its on 2026.5.4

but im getting a prep list on updating with my claw on the test server, i'll be applying the update shortly

it listed a bunch of things it wants to check

[nocturne jungle]

Yeah OAI models use codex harness by default in this beta, so you wouldn't need to set it explicitly

[shell dawn]

perfect

[rocky pewter]

I just tested it on my running system and I’m still seeing the same issue there. I can also check it locally on a fresh isolated setup if that helps.

[shell dawn]

my engineering claw that handles my memory system has been like WATCH AND WAIT O_O THERES A LOT OF CHANGES

guess what, we jumping in

[shell dawn]

both openai provider and openai-codex right? im on the codex sub

[nocturne jungle]

just openai. openai-codex is now considered "legacy"

[shell dawn]

oh that explains that issue back in uh 2026.5.4? heh

Restart exposed one hard beta.3 migration break: config validation now rejects openai-codex/gpt-5.3-codex in Vigil’s fallback list. I’m applying the narrow repair to replace that dead model with openai/gpt-5.5, then restarting again.
i told my claw to do the update, so i guess we are testing the agent led update process. i know this can be tricky to get to happen right, i dont know if this is GPT-5.5 medium or the instructions but this is a good sign

didnt catch all of them, ah well.

[nocturne jungle]

catch what?

[nocturne jungle]

Merging #81236

[potent path]

Don't want to mess with my prod system, but I'll set up a crabbox on a temporary google vps and test the upgrade from an older version to verify that the 5.12b3 upgrade installs and configures the codex runtime/harness automatically. Shouldn't take long.

[wheat sphinx]

@shell dawn you good my man? Don’t go confusing people, aight?😏

[civic sedge]

Bug fix - after I added the provider wildcard method, discovered that it wouldn't work with vllm / sglang. fixed it here: https://github.com/openclaw/openclaw/pull/81230.

edit: i see codex decided to weigh in like 20 minutes later. brb

[shell dawn]

My agent didn’t fix everything, and doctor didn’t either. I’m editing

[shell dawn]

5.3-codex isn’t supported

[fallen gate]

https://tenor.com/view/dancing-cat-cat-kitten-dance-dancing-gif-5604500283313576480

😠

[wheat sphinx]

Just handing out the role like Pez

[rocky pewter]

@nocturne jungle im nuking 1 Vps then i start testing that.

[shell dawn]

is there a log somewhere that says why "openclaw doctor --fix" didnt make the changes it said it was going to make? ive got openai-codex/gpt-5.5, 5.4, 5.3-codex listed extensively in primary and fallback models and it doesnt seem to actually want to touch it. i can edit it by hand but i suspect openclaw doctor is expected to be able to fix this. im not the only one who would run into this

or you just run it twice and then its fine? shoudlnt take more than once

[fallen gate]

@nocturne jungle you wanted someone to test image gen?

[potent path]

Yeah, I think they removed it from the catalog recently? Or am I thinking of Codex Spark? Seems to ring a bell but it's been a busy few days and I can't recall exactly. Worth checking to see what's currently in the catalog and make sure your model assignments are aligned.

Oh, here it is:
Replaced: openai-codex/gpt-5.3-codex-spark ("no longer exposed by the OpenAI or Codex catalogs")
New fallback: openai/gpt-5.4-mini.

[nocturne jungle]

thank youuu

[charred crater]

Wow so the codex harness in openclaw can actually access codex plugins. You guys cooked.

[nocturne jungle]

Did doctor not work? I think the pr that i merged from @rocky pewter that recommends running doctor when config validation fails should help agents in the future

[warped tide]

i've got open PR for fixes on gemini token refresh and xAI legacy reasoning effort handling

[potent path]

I suppose you could try --force

--force: Additionally overwrites the entire supervisor config (service file). Use this only if --fix fails to resolve service-level issues and you are okay with losing custom service configurations.

https://docs.openclaw.ai/gateway/doctor#repair-force

[shell dawn]

"openclaw doctor" highlighted a ton of fixes my agent did not fix before the attempted gateway restart
then "openclaw doctor --fix" the first time prompted me to make the change, i said yes, and then ran "openclaw doctor" again and the same warnings were present
then i ran "openclaw doctor --fix" and it asked again, i said yes, and it fixed it

[potent path]

Oh, that's interesting. A little ralph loop. Just keep telling it to fix it until it finally figures it out, I guess. lol

[rocky pewter]

@nocturne jungle
I tested this on a fresh isolated VPS setup.

Result: beta.3 works for the intended path when npm is available.

• openclaw@2026.5.12-beta.3 installed cleanly
• the host package includes dist/plugin-sdk/codex-native-task-runtime.js
• @openclaw/codex@2026.5.12-beta.3 installed cleanly
• plugins doctor reports no issues
• plugins inspect codex shows the plugin loaded correctly
• Codex plugin version: 2026.5.12-beta.3

[shell dawn]

tui worked fine afterwards, my chat app was v3 and not v4, so im updating that now

[nocturne jungle]

Was this a fresh install or existing config / upgrade?

[nocturne jungle]

So to confirm after all that, you upgraded to beta 3 and everything worked?

[rocky pewter]

Fresh isolated install. I did not use the existing OC config or service.

[tulip gulch]

tested your pr https://github.com/openclaw/openclaw/pull/81206 on top of beta.3 and could still reproduce this issue https://github.com/openclaw/openclaw/pull/80539

[nocturne jungle]

https://github.com/openclaw/openclaw/pull/80539

And it works on your pr?

[shell dawn]

so far yes. working through some testing. i have a custom chat app and it has some issues, trying to identify the actual cause

[tulip gulch]

yes see the video attached to the PR

[nocturne jungle]

Trying to get a sense of the risk/reward for merging something like this before the big release is. What do you think the tradeoffs are?

[tulip gulch]

let me test and triple check again. Codex plugin install fails for older npm version like 9.2.0, which I myself had installed as that's how I got the error so must not be an uncommon setup

in fact I got it on a fresh Hetzner VPS, so I guess the default installation

[nocturne jungle]

@tulip gulch i reviewed it, seems good

merging

[fast pagoda]

@nocturne jungle Hi pash, if we now want to fetch the latest stream, should fetch release/2026.5.12 or main;
and when open PR, target to main or to the release branch

[nocturne jungle]

I'm gonna start backporting to the release branch - so be testing that

Will post updates

[fallen gate]

https://giphy.com/gifs/dancing-dance-party-clubbing-blSTtZehjAZ8I

😠

@jolly wolf fix this

[nocturne jungle]

Backported and pushed to release/2026.5.12.

Included:

#80539: npm alias override retry for managed plugin installs, by Rubén Cuevas
https://github.com/openclaw/openclaw/pull/80539
#81165: Codex migration selector Enter behavior, by Kevin Lin
https://github.com/openclaw/openclaw/pull/81165
#81192: offer Codex migration after harness install, by Sarah Fortune
https://github.com/openclaw/openclaw/pull/81192
#80834: PowerShell installer no longer aborts on git stderr, by Sarah Fortune
https://github.com/openclaw/openclaw/pull/80834
#81220: onboarding --skip-hooks, by Sarah Fortune
https://github.com/openclaw/openclaw/pull/81220
#81246: WhatsApp drains debounced inbound before close, by Marcus Castro
https://github.com/openclaw/openclaw/pull/81246

[shell dawn]

yeah, had some issues probably with the very first session resuming, but it recovered, and after that its been fine. openai/gpt-5.5 fine, anthropic/opus-4.7 fine, opencode-go/kimi-k2.6 fine, im getting an error about responses api on github-copilot/gemini-3.1-pro-preview but its probably just something weird in github-copilot setup that i neglected to fix on my test server from long ago

the first session issues is most likely a me issue with rewarming the session and some metadata or something changing

it for some reason thought the message tool was broken... despite telling me it was broken using the message tool

it discarded the possibility of it being an issue afterwards so my theory seemed quite probable

[nocturne jungle]

also backported this
https://github.com/openclaw/openclaw/pull/81236

to release branch

[nocturne jungle]

was this in webui/tui? or in a proper external messaging channel

[shell dawn]

webchat that i made myself 😄

tui was my first attempt and it said it was fine, that was a different agent using 5.5

[nocturne jungle]

judging by how quiet everything is...

are we good?

[civic sedge]

Calm before the storm. Or something like that

[civic sedge]

I take it you have never worked in restaurants or service industry ? The q word is a curse

[jade wagon]

Merged

[fallen gate]

installing meow, what you want me to test

[nocturne jungle]

well i'm very biased but openai models - happy paths for new and returning users. big picture- things that 80% of openclaw users tend to do

OAI subscription - stuff like that

[fallen gate]

wen browser use

[untold pollen]

Thanks for the PR! Landed #80671. Testing and confirming the 2nd one

[finite solar]

literally give me codex cli browser use

[fallen gate]

wait, I never used this, is it the same as the codex app one?

cus that is 🔥

[finite solar]

Oh I meant codex cli needs browser use not openclaw through codex cli browser use

[gritty tree]

merged couple of small fixes
https://github.com/openclaw/openclaw/pull/81235
https://github.com/openclaw/openclaw/pull/81215

[celest merlin]

Hi folks, working on the release today and will land PRs. I'd looooooove folks to help me test the beta so we get sth good out.

2026.5.12-beta.4 is up

[wintry flare]

Hi Peter.
I have important thing to discuss with you, can you dm me please ?

[celest merlin]

it better be good. ok!

[weak walrus]

wait - why

[shell dawn]

thats what it said

[weak walrus]

define "it"

This is on new harness @shell dawn ?

I have NOT tested this

[shell dawn]

yeah in beta.3. im gonna update to 2026.5.12-beta.4 right now

[weak walrus]

let me know, because I'll need to do some testing on this, especially since if people are fully using their OpenClaw with OpenAI subscription to the max, they're using the extra allowance for 5.3codex

You kow what, I have a beta3 up, I'll test myself now brb

[shell dawn]

Rejected validation details: agents.list.10.model.fallbacks.0: Unknown model: openai-codex/gpt-5.3-codex. gpt-5.3-codex is no longer supported for ChatGPT/Codex OAuth accounts. Use openai/gpt-5.5 through the Codex runtime.

[shy rain]

Still waiting for your commit https://github.com/openclaw/openclaw/commit/5a80be35e9821f602e2ddd86aa4e192cd87e28ab
To fix the watchdog so slow crons don't die. Don't see it in Beta 4

[weak walrus]

updating to beta5

?

[shell dawn]

beta.4 just came out 10 minutes ago

[shy rain]

lol 10 minutes .

[weak walrus]

I know... yet, in my dashboard it says beta.5 is avail to download

prob a web gateway bug

[fossil ferry]

Hi maintainers — I opened a PR to forward sampling params through the OpenAI-compatible HTTP paths:

https://github.com/openclaw/openclaw/pull/81275

It threads temperature and top_p from /v1/chat/completions and /v1/responses into streamParams and the OpenAI transport, so client-supplied sampling preferences correctly reach the upstream provider.

It also fixes the embedded runner wrapper to preserve topP, and filters unsupported top_p values for the native ChatGPT Codex Responses backend to avoid upstream request errors.

This PR is intended to close #81274 and follows the same wiring pattern as #81013.

I’ve added targeted tests, updated the docs, and included live gateway verification with DeepSeek showing both params reaching the provider correctly. Feel free to take a look when convenient.

[shy rain]

https://github.com/openclaw/openclaw/pull/78213

Still open. It isn't mine but I mentioned the issue in here:

https://github.com/openclaw/openclaw/issues/78202
[Bug]: OpenClaw Codex harness overrides HOME and CODEX_HOME, causing ~ to resolve to isolated harness storage instead of the operator home.

Causes sub-processes and cron jobs relying on $HOME path to fail.

Easy workaround but it's a silent regression.

[celest merlin]

what dashboard? https://www.npmjs.com/package/openclaw?activeTab=versions is truth

[weak walrus]

Yup, it was odd, the top of the dashboard where it says new version available, I should have took screenshot. I updated it and it took me to beta4

[fallen gate]

I don't have the ability to call all nerds here

@naive horizon I think you can now, try it out

@Beta Pings

[vocal flame]

Hi Peter, can you please check my DM?

[fallen gate]

why you all trying to slide?

[celest merlin]

stuff like this let's discuss in the open

[tired canyon]

Hi Peter. How are you? ❤️
#clawtributors message
Could you please check my message?

[celest merlin]

edit: ah it's in the issue.

[celest merlin]

Please use codex /review before submitting PRs.

[untold pollen]

<@&1503801512294486217>

(testing)

[fallen gate]

2026.5.12-beta.4 is up

[weak walrus]

WOW - @pash / Peter Steinberger - the "Migrate native codex plugins" is a VERY nice touch. Canva, github, outlook calendar and email already configured. This is going to be youtube video worthy my guys

Testing now - will report

[weak walrus]

ok dude, I don't even see in the openclaw config how to even add the model in there, I will add manually and test

[shy rain]

well I did say it isn't even mine but I welcome whomsoever it was that submitted it to sort their sh*t out. I just repeated it here because I've mentioned the bug already, I'm not the only one, and we just work around it in the meantime 🙂

[celest merlin]

Yeah that is bad hack, I can't land this.

[shy rain]

I'll see what I can do later when I have time

[celest merlin]

We have migration code that should address this way better. I will close this issue and PR.

[shy rain]

Can't wait

[crude matrix]

I stopped testing beta's now, almost no fixes based on reports I do end up in the beta or release, while I am having issues. example https://github.com/openclaw/openclaw/pull/80449 and https://github.com/BingqingLyu/openclaw/pull/1273...

Also compared to 4.23 agents are much slower, when I sent then a telegram message I could see it would start almost directly. after that versions I can see my message being received but it takes 30 seconds or more to start with that... see github issues.

[untold pollen]

Reviewing 🫡

[crystal relic]

hi, maintianers! I found Control UI have a lot key without internationalization.

so i add a report script for contributors @tame meadow How about it?
https://github.com/openclaw/openclaw/pull/81320

~/git/openclaw--control-ui-i18n-report (feat/control-ui-i18n-report ✔) pnpm ui:i18n:report --surface chat --locale zh-TW
$ node --import tsx scripts/control-ui-i18n-report.ts --surface chat --locale zh-TW
Control UI i18n baseline report
Scope: Chat, Traditional Chinese (zh-TW)
Based on: current raw-copy baseline and locale metadata. Not a drift check.

Current i18n state
  Hardcoded UI text outside i18n: 81 pieces in code, 92 total occurrences.
  Existing zh-TW translation keys (all Control UI): 1108/1108 filled, 0 fallbacks.

Current issue
  Chat still has UI text written directly in code.
  24 zh-TW translations still match English and need review.

Focus modules
  54 Chat page: ui/src/ui/views/chat.ts
  12 Chat message groups: ui/src/ui/chat/grouped-render.ts
  7 Chat tool cards: ui/src/ui/chat/tool-cards.ts
  5 Chat queue: ui/src/ui/chat/chat-queue.ts
  5 Chat tool result panel: ui/src/ui/chat/side-result-render.ts
  4 Slash Commands: ui/src/ui/chat/slash-commands.ts
  3 Build Chat Items: ui/src/ui/chat/build-chat-items.ts
  2 Context Notice: ui/src/ui/chat/context-notice.ts

Next steps
  Move text from the focus modules into translation keys.
  Do not hand-edit generated locale, translation memory, or i18n metadata files.
  Run pnpm ui:i18n:sync after adding translation keys.

[untold pollen]

For UI stuff I would page @stable pewter

[celest merlin]

This isn't a release blocker. I rather focus on release blockers to get the release out and then we can land other stuff.

[grand kernel]

Hey for CodexBar shall I drop the pull request link here?

[celest merlin]

no. please. This here is the openclaw discord.

I need people that help. me with the releaes, not post unrelated stuff

I do other stuff on the side, thanks for the help there! Just really need to focus here.

[crystal relic]

about this pr, I think can be fix it current release.

Caps media while downloading, closing an avoidable memory-risk gap for oversized files.
https://github.com/openclaw/openclaw/pull/81044 feishu
https://github.com/openclaw/openclaw/pull/81050 whatsapp

[untold pollen]

Folks there other maintainers here too haha, don't all pile up on Peter 😄

[celest merlin]

hey when you write PR descriptions, remove the template stuff? Makes it easier to parse.

[crystal relic]

okay, my bad , let me fix it.

[celest merlin]

That doesn't seem to be a release blocker tho, more security hardening and a potential risk if we squeeze this in late?

[crystal relic]

makes sense. u are right. this problem has existed for some time.

[fallen gate]

He means feel free to ping him

[untold pollen]

Yes please, go for it. Especially for Telegram related PRs

[stable pewter]

I have been summoned ✨🌙

[stable pewter]

Looking at this now, thanks for sharing.

[fallen gate]

https://tenor.com/view/pokemon-i-choose-you-ash-pokeball-throw-gif-5652844

😠

[tall crow]

LOL I think the threshold of the CHANGELOG.md file just exceeded github's rendering limit, we can only view it as a raw md right now

[fast pagoda]

Crabbox regression report:

• Provider: GCP Crabbox
• Lease: cbx_6f5b4dd8a4f0
• Remote commit tested: 6926481
• Build: passed
• 10 regression probes: 10 passed, 0 failed
• Covered: version, config CLI, gateway CLI, cron CLI, agent CLI, plugins CLI, infer CLI, secrets CLI, sessions CLI, doctor CLI.

One existing issue remains:

• lossless-claw plugin load error: missing @mariozechner/pi-coding-agent
• Gateway health is still ok, but that plugin is not healthy. This looks separate from the update itself.

[untold pollen]

missing @mariozechner/pi-coding-agent
oof yeah, that got renamed to the earandil repo

[crude matrix]

@untold pollen I saw you accepted one of the fixes, thank you that was a big one for me.. are you aware of an issue 5.12-beta one had that Agents appear frozen, showing writing indicator but no output.. Did not have this issue with the previous beta. (https://github.com/openclaw/openclaw/issues/81191#issuecomment-4435266708).. there is no fix for that but I also did not find many issues that seemed related or same. Could you also have a look at this one https://github.com/BingqingLyu/openclaw/pull/1273... session_send issue.... I have been adding this fix to every stable release so far.. its about agents complaining that they are not receiving confirmation that a session_send is actually received and that when there is no session they cannot send it and workflow halts.

[untold pollen]

Please make a PR on the openclaw repo for the 2nd one and I'll be happy to review it. The one you have linked is on your own fork

[crude matrix]

I will check, I think someone else fixed it.. was not me.. I will contact him.

[weak walrus]

Welp, I had 2 PRs but by the time I finished my testing, someone already fixed them , well done team lol

now I can haz sleep. Night zzz

[willow narwhal]

Hey guys, 2026.5.12-beta.4 is up and needs testing.
Please give it a spin and flag anything that feels broken or off.
Thanks!

[untold pollen]

Ah got it, taking a look

@crude matrix check this out please: https://github.com/openclaw/openclaw/pull/81365

[celest merlin]

Folks, can someone review https://github.com/openclaw/openclaw/pull/77023 - this was a PR that I had to rewrite as it was incomplete, but the direction was correct. @fallen gate

[merry slate]

I updated to 2026.5.12-beta.4. After installing this beta version, my custom plugins stopped working. I'm still troubleshooting to figure out what's causing the issue

[willow narwhal]

Thanks.
Is this a plugin you built yourself?
Please post your findings.

[merry slate]

Yep, this is my self-built plugin. Still looking into the issue and I'll post my findings soon

[willow narwhal]

Okay.

[crystal relic]

Hi @stable pewter the PR all issue already fixed, got some ci failed , seem not with pr change.

[steel beacon]

Hi team. Opened 4 PR's very small PR's would appreciate a review.
https://github.com/openclaw/openclaw/pull/81356 web_app button support in inline keyboard
https://github.com/openclaw/openclaw/pull/81354 export plugin hook types (this one is VERY small, just new exports)
https://github.com/openclaw/openclaw/pull/81353 expose tools in LLM input hook event (3 lines code change)
https://github.com/openclaw/openclaw/pull/81351 feat(telegram): localized command menu descriptions

[next helm]

Good morning. In about an hour I’d be happy to test and write some fixes if there’s anything in particular needed in this latest beta. I’ll upgrade all my claws and submit the issue reports

[celest merlin]

I made a crawtributor skill so I will pick from here preferred over the git noise.

[muted badger]

great,early stage of the certification system comes out

[next helm]

Very cool. Love the automation and I steal a lot of it (the idea) for my company.
Any way I can get a view of current beta update issues so I would know where to focus and try to reproduce?

[celest merlin]

not sure what you need, the branch and everything is public?

[celest merlin]

the last one is full of booboos

[steel beacon]

Agree with the findings. Anyway thanks for quick merge!

[merry slate]

Found the root cause. When OpenClaw beta updates plugins in ~/.openclaw/npm/, it runs npm install which recalculates the entire dependency tree for that shared node_modules. If a bundled plugin (e.g. @openclaw/acpx@beta.4) introduces a peer dependency conflict, npm may silently remove transitive dependencies of other plugins.

In my case, @lancedb/lancedb (used by my memory plugin) has apache-arrow as a peer dependency. After the beta update, npm removed apache-arrow from the shared node_modules due to a peer conflict with acpx. My plugin then failed to load with:

[plugins] engram failed to load: Error: Cannot find module 'apache-arrow' Require stack: @lancedb/lancedb/dist/arrow.js
I've added apache-arrow as an explicit dependency in my plugin's package.json (instead of relying on lancedb's peer dep)

If you have @openclaw/acpx installed alongside any plugin that depends on @lancedb/lancedb, upgrading to beta.4 may silently break that plugin.

[willow narwhal]

Thanks.
That's helpful.

[next helm]

I’ll scan the branch 👍

[rapid spire]

Did the issue with "shared-ini-file-loader" in the git installer get resolved?

https://discord.com/channels/1456350064065904867/1503903223017373737

(I’m at work and unable to test for the next 8-10 hours)

[muted badger]

one related question,just curious, from the perspective of a codex architect, will the evolution of LLMs quickly eat up the harness layer, or, as long as competition among LLM companies persists, will the harness have an independent necessity to orchestrate larger scale tokens to complete tasks?

[celest merlin]

Oh I haven't seen this yet. too many threads and pings. This would be better in a github issue I can reference for codex

[rapid spire]

Sorry for adding to the chaos.

[rapid spire]

Can anyone please copy this to a GitHub issue? I’m unable to access discord from work (and my thumbs cringe at the idea of doing it from my phone)

[celest merlin]

don't, I'm already on it.

Just as a note for next time.

[next helm]

I updated three independent macOS OpenClaw beta installs to 2026.5.12-beta.4 and filed a sanitized consolidated audit here: https://github.com/openclaw/openclaw/issues/81394. The updates completed successfully and all gateways ended healthy on beta.4, including direct non-delivered agent checks through openai/gpt-5.5/codex with no fallback, but all three showed the same confusing restart-phase stale-version warning after package update.

Issue Highlight
The main issue is that openclaw update writes config with 2026.5.12-beta.4, then the restart phase still appears to run from the pre-update CLI and warns that the config was written by a newer version, even though the gateway is already updated and healthy. Secondary audit findings: plugins doctor emits an unattributed deprecated runtime.config.loadConfig() warning while saying no plugin issues were detected, and external plugin fallback from @beta to @latest could be clearer in the final update summary.

Shortly will submit PR's to

1. Improve the package-update restart warning so expected updater self-staleness is not presented like an operator PATH/config mismatch.
2. Add a final update-summary line when an external plugin falls back from @beta to @latest. Start around src/plugins/update.ts.
3. Attribute or suppress the runtime.config.loadConfig() deprecation warning in clean plugins doctor output.

[muted sierra]

Hey, I reviewed PR #77023 again against the current head and latest main. I don’t have push rights on the branch, so I couldn’t attach a fix commit directly.

I updated my PR comment with a concrete patch suggestion and local verification:
https://github.com/openclaw/openclaw/pull/77023#issuecomment-4439930723

Short version: the earlier issue changed shape after new commits, so I adjusted the proposed fix to the current code structure and tested it locally against src/agents/subagent-announce-delivery.test.ts with 74/74 passing.

[rapid spire]

This is a possible regression I also reported last night (not in GitHub, sadly) that I don’t see addressed in the commits:

Invalid config at /home/xxxx/.openclaw/openclaw.json:\n- channels.discord.guilds.1012875612672753714.channels.1503873770945904671: invalid config: must NOT have additional properties

doctor --fix can't fix it here either. Note that I run "openclaw doctor --fix" multiple times to try and fix this with the idea that it installs the discord plugin on the first try and then might fix the json on the second. No luck.

This occurred with all the 5.12 betas from 1 thru 3 trying to upgrade from 4.23. (This might not be a regression. It’s possible that the new version is cutting this feature)

[willow narwhal]

Thanks. Will look into this.
If you happen to put up a PR, do link me.

[next helm]

They should be up shortly and will link you. Doing reviews now 👍

[civic sedge]

Is there a certain format you prefer so it can pick them up ?

[celest merlin]

no, this is for me internal and just a FYI

[rapid spire]

Issue 81400

[celest merlin]

subagents should only ever steer, nothing else.

[crystal relic]

https://github.com/openclaw/openclaw/pull/80725 ready for review.

This PR fixes the chat streaming payload bloat, giving clients much smaller updates and noticeably smoother live output.

https://github.com/openclaw/openclaw/pull/80705 ready for review.

fix: removes a small repeated cost from every inbound auto-reply message, with no behavior change.

https://github.com/openclaw/openclaw/pull/80741 ready for review

Prevents unapproved Nodes from becoming callable or Talk-capable, closing a pairing-surface security gap.

[potent path]

Don't want to mess with my prod system,

I put this in a thread because it's a lot of text, but the bottom line is that my initial testing indicates that an upgrade does NOT automatically install the codex harness/runtime plugin, but it DOES set it as the default runtime for openai models.

Result: openai models just stop working after the upgrade and don't work again until the plugin is manually installed.

Could use more testing by others to confirm that I haven't screwed something up with my scripting, but that's my initial assessment.

#1503940236261195906 message

Oh. Looks like beta4 is out now and may already address this. I'll rerun the testing and verify.
"Codex runtime: allow the official installed @openclaw/codex package to use its private task-runtime SDK helper, fixing MODULE_NOT_FOUND during migrated OpenAI/Codex beta runs."

[fallen gate]

Good mornin

I purposely left that one since I considered it a feature more than a bug fix

[fallen gate]

Love it...

[potent path]

PR #80751 @potent path +838/-108 security 🟣 draft/security policy direction, not quick land.
Fair. To be clear, I've kept that in draft because I wanted to obtain some direction from maintainers on whether y'all agree that this is the correct approach. Don't want to put forward a security change (SSRF) without a little discussion.

PR #81064 @potent path +2500/-84 perf 🔴 broad plugin metadata perf change, not quick land.
Yeah, I was afraid of that. It started as a small, narrowly-focused PR to fix a bad performance bug, but ended up expanding to handle a lot of edge cases. We could revert to the quick-and-dirty patch to make it smaller, but then Codex and Claude both complain at me. In my opinion, the performance improvements are worth it:
'openclaw plugins list' improved from 89.45s to 7.38s
'openclaw gateway status' improved from 74.34s to 6.74s

[tulip gulch]

tiny wording fix for codex oauth https://github.com/openclaw/openclaw/pull/81301

[next helm]

@willow narwhal
I split the #81394 update-hardening work into three narrow PRs: package-update restart messaging #81414 keeps updater self-staleness from looking like a real future-config/PATH mismatch; plugin channel fallback summary #81422 https://github.com/openclaw/openclaw/pull/81422 preserves beta-to-latest plugin fallback state in final output; plugins doctor runtime-config warning attribution #81425 https://github.com/openclaw/openclaw/pull/81425 makes deprecated runtime config warnings actionable instead of appearing beside “No plugin issues detected.”

Validation is green for the targeted scope: #81414 has the update CLI/restart guard tests passing, #81422 has src/plugins/update.test.ts + src/cli/update-cli.test.ts passing with 163 tests, and #81425 has the runtime/registry/plugin CLI tests plus pnpm tsgo passing. One caveat on #81414: CI currently shows an unrelated check-test-types failure in src/auto-reply/reply/queue/settings.test.ts, and it has triage: mock-only-proof; could a maintainer confirm/override the unrelated CI/proof gate or tell me what live package-swap proof they want?

[smoky dock]

Are you publishing this one?

[willow narwhal]

Hey, thanks for testing and reporting.
This has been fixed and should be in the next beta

[willow narwhal]

I will take a look at this in a min

[muted sierra]

Good skill my PR#81091 is 🟢
🤣
https://github.com/openclaw/openclaw/pull/81091

[olive cliff]

hey guys! I have these PRs that I opened a few days ago and would appreciate some help reviewing / closing them as needed!

https://github.com/openclaw/openclaw/pull/74553
https://github.com/openclaw/openclaw/pull/75560
https://github.com/openclaw/openclaw/pull/76121
https://github.com/openclaw/openclaw/pull/81188
https://github.com/openclaw/openclaw/pull/81189

there were some other ones focused on telegram, but Ayaan already merged those for us :)

[crude matrix]

Hi, sorry I just saw your message... If I read it correctly the PR is to let agents use session_create before they use session_send. The fix I was using was more advanced I think... If session_send did not find the session the message was sent to, it would look are there other sessions, (except heatbeat). if there are return that message to the agent who sent it for it to choose if the available session is correct. if there are no sessions it would automatically create a session. If the session exists it would sent the message and a message received confirmation would go back... if the session would not send a confirmation message to session sent it would look if the conontext is full, if the context is full it means the agent cannot work without compression, so it would compress and retry... if it still does not work the session will be resetted (last was an option in the openclaw config file). Most important for me was that the current session_send is one way. so my agents are instructed to ignore timeout.. timeout means message received .. But many times this was just not true.

the fix that was made solved this, it added a send, received and agent active checkmark (message send, message received, message read).
One thing it did not add that I would have liked to have is Sender.. I can see in OpenClaw when a message is received... sometimes its not clear who of my agents send this. I have instructed them to always add who they are and when done to who the agent needs to reply to.
Another feature that I asked for but got denied is group sessions, like team chats, where a message is delivered to the group chat, the session of the agent gets a notification and if he agent is spoken to with a @agent1 @agent2 etc they can direct messages to specific agents, but because the conversation is in a group chat all agents have more context than just the session_send.

[willow narwhal]

Alright, I will look into it

[fast pagoda]

Found a heartbeat message leak issue: https://github.com/openclaw/openclaw/issues/81369
PR targeted to release/2026.5.12: https://github.com/openclaw/openclaw/pull/81441

[untold pollen]

It makes sense but this sounds like a broader A2A delivery protocol change

The PR I linked is intentionally the narrow root-cause fix for first-contact sends

[nocturne jungle]

Is this using the heartbeat tool?

doesn't look like it.

[fallen gate]

is this unknown user to everyone or just me?

[worldly tusk]

Everyone

[languid jasper]

unknown user for me too

[fallen gate]

[crude matrix]

I think this narrow PR is better than no PR... I already contacted the person who made the other to make sure there are no conflicts with current state and to make a PR from it... Like I said I have being manually implementing it with every final release I tried because it really ups the agents quality.

Or perhaps someone else sees benefits for what I describe and can make a pr out of it?

[next helm]

Ohh I think I found a small update issue when switching runtimes

[next helm]

@unknown_user is on fire today. Well done!

[icy tusk]

woops, clicked it twice

[fast pagoda]

Small correction: I linked the wrong PR earlier. Should be https://github.com/openclaw/openclaw/pull/81441 .

On the heartbeat tool question: it’s the heartbeat runner path, but likely not a successful heartbeat_respond result. Codex heartbeats do enable/force heartbeat_respond, but the leaked text looked like fallback raw reply payload text (<tool_calls>, <file_contents ...>), not parsed heartbeat-tool channelData.

So the fix sanitizes the fallback replyPayload.text before sendDurableMessageBatch. If heartbeat_respond works, that path is unchanged; if it fails/falls back, raw protocol scaffolding won’t be delivered.

[potent path]

Correction: The beta.4 upgrade DOES auto-install and use @openclaw/codex in a migration from 4.23! 🥳

[nocturne jungle]

there shouldn't be a "fall back" - either the agent calls the heartbeat tool or doesn't.

[shell dawn]

pash i ran into a few more issues testing 2026.5.12-beta.4, though it seems to be related to webchat which I had an issue with on beta.3

even after clearing the webchat issues, which seem config related, im seeing some sporadic high latency turns with codex

[nocturne jungle]

yeah, webchat has some issues - but I wouldn't say they're an urgent category for this immediate release

[shell dawn]

02:05 turn duration 147.9s
02:22 turn duration 158.5s

for answering like a 'look at the logs to identify if there is background issues'

its completing, its just really slow, and sometimes there seems to be errors that get retried and completed, its not super consistent other than the latency

im trying to prompt more analysis to get something concrete but it seems to be shifty trying to zero in on a specific problem

sessions.list mostly 180-220ms
chat.history mostly 235-260ms
occasional chat.history burst 700-1119ms
background activity seems consistent

[nocturne jungle]

worth asking codex to actually look at the session rollout file for that channel to see what the agent was doing. Maybe it was doing a lot of work

[shell dawn]

its a test server, all ive done is say 'hey how are things looking'

zero cron jobs or recurring tasks

but session rollout is an angle i havent asked, i'll check that, thanks

[terse solstice]

ty for merging pr 76601 🙂 that should speed up adding many agents

[next helm]

Opened PR #81479 to fix Codex isolated-home Gateway auth for cron/subprocess CLI calls. The issue was that Codex app-server rewrites HOME, so child commands like openclaw cron list could not resolve gateway.auth.* when the parent Gateway config used SecretRefs. The PR resolves local Gateway auth once in the parent app-server bridge and injects OPENCLAW_GATEWAY_TOKEN / OPENCLAW_GATEWAY_PASSWORD into the stdio app-server env, preserving explicit overrides and warning/skipping injection if the SecretRef is unavailable. It also adds token/password/unavailable-ref regression tests, updates the SDK boundary via gateway-runtime, and includes the API baseline/docs/changelog updates.
https://github.com/openclaw/openclaw/pull/81479

[next helm]

@nocturne jungle this is codex related and was experienced by me today with all my crons failing on one machine

and main moves so fast i can never get changelog.md to not have a conflict lol

[crystal relic]

https://github.com/openclaw/openclaw/pull/81474 ready for review

fix stale live transcript updates, make session history stays ordered in real time and don't need rescans.

https://github.com/openclaw/openclaw/pull/80335 ready for review

fix busy agent runs with chat, which should be a big reliability win for active sessions.

[pastel owl]

Regression bug:
The tool progress messages are not getting deleted after final reply in telegram group chats, but is working fine in direct chats

Can there be a system to test expected behaviour after each code change ? Can crab box be programmed to do that ?

[fast pagoda]

there shouldn't be a "fall back" -

[celest merlin]

Yes, @untold pollen did great work here. Some of that are also Telegram platform bugs, which makes te whole thing more tricky.

[celest merlin]

beta 5 is up.

[cunning grotto]

hopa

[celest merlin]

I wanna get this out today. Throw regressions my way.

[cunning grotto]

is the current state of openclaw stable? i have stayed on verison 23 for the most part after reading your blog post

[celest merlin]

The trouble were a week, it's good for a week again.

[cunning grotto]

hah

[fallen gate]

@jolly wolf you know what to do

[cunning grotto]

auto skill creation has been brought it now, correct?

[celest merlin]

nah, that's coming later. This is a feature that sounds great in theory, but is far too easy to just create slop/

[cunning grotto]

coming from hermes - i've been wanting that functionality in openclaw for some time now

[willow narwhal]

@merry slate kindly try this beta and let me know

[cunning grotto]

wasnt there a method to pruning bad skills? i remember reading something about a filtration system that sorts out that slop

i feel like hermes' implementation is pretty robust from my use at least, would it not be simpler to use their setup for that?

[jolly wolf]

<@&1503801512294486217>

[merry slate]

ready to go!

[worldly tusk]

There is nothing magical about hermes skill autocreation. It just has the following prompt, and that's all!

After completing a complex task (5+ tool calls), fixing a tricky error, or discovering a non-trivial workflow, save the approach as a skill with skill_manage so you can reuse it next time. When using a skill and finding it outdated, incomplete, or wrong, patch it immediately with skill_manage(action='patch') — don't wait to be asked. Skills that aren't maintained become liabilities.

[willow narwhal]

Thanks!

[cunning grotto]

oh that's awesome, i never knew it was that simple

[slow inlet]

https://github.com/openclaw/openclaw/pull/81239

Small telegram w/ cli regression fix, pick models interactively via /models

[worldly tusk]

Their prompt is actually pretty interesting - I've stolen some ideas from there 🙂 this is the exact file to read: https://github.com/NousResearch/hermes-agent/blob/main/agent/prompt_builder.py

[pastel owl]

https://github.com/openclaw/openclaw/pull/81486

why this pr ?
telegram group chats were not deleting the toolprogress messages, and messages stayed in chat and were making the chat dirty. As opposed to direct messages to bot which are working fine.

Kindly review

[merry slate]

Hey, sorry but after updating to the latest beta, there's a new issue — looks like something in @lancedb/lancedb is triggering the security check and getting blocked by the plugin update system.

Error log:
Disabled "engram" after plugin update failure; OpenClaw will continue without it.
Failed to update engram: Plugin "engram" installation blocked: dangerous code patterns detected:

• Shell command execution detected (child_process) (/Users/scotthuang/.openclaw/npm/node_modules/@lancedb/lancedb/dist/native.js:16)
• Dynamic code execution detected (/Users/scotthuang/.openclaw/npm/node_modules/@lancedb/lancedb/dist/embedding/transformers.js:47)

What those lines actually do:
native.js:16 — lancedb checks if the system is glibc or musl to load the right native binding:
const lddPath = require('child_process').execSync('which ldd').toString().trim();
return readFileSync(lddPath, 'utf8').includes('musl');

transformers.js:47 — lancedb uses eval to do a dynamic ESM import (to stay compatible with both ESM and CJS):
transformers = await eval('import("@huggingface/transformers")');

lancedb version is @lancedb/lancedb@0.26.2

[next helm]

I updated two OpenClaw beta installs from 2026.5.12-beta.4 to 2026.5.12-beta.5; both core gateways are healthy and direct model checks passed. I added beta 5 reproduction details to the stale updater-warning issue here: #81394 https://github.com/openclaw/openclaw/issues/81394#issuecomment-4444193265 , and filed a new issue for the beta 5 lossless-claw plugin/context-engine degradation here: #81499 https://github.com/openclaw/openclaw/issues/81499.

Issue Highlight
The recurring updater warning still appears after successful npm/pnpm updates, saying the config was written by beta 5 while the command is running beta 4, even though the gateway is already on beta 5. Separately, one host’s lossless-claw plugin had no @beta release, fell back to @latest, was blocked by the plugin security scanner, then left stale plugins.entries, plugins.allow, and contextEngine config while runtime fell back to legacy; I did not force reinstall or run doctor --fix.

[rapid spire]

81400

[merry slate]

traced the issue — it's commit 1415c06fc4 (#81066) in beta.5, which extended the install scanner to cover plugin node_modules/ dependencies. beta.4 didn't scan node_modules/, so lancedb (0.26.2, unchanged) was never flagged before.

Funny thing is, openclaw core's own memory-lancedb uses the same @lancedb/lancedb package — it just skips the scanner because it's bundled in core

[next helm]

install scanner killed my lossless claw

[willow narwhal]

That was a recent security fix.

[willow narwhal]

In this new beta?

Ok nvm, I just saw your message above.

[worldly tusk]

I think, if a plugin cannot be safely updated for whatever reason (unavailable, insecure, etc), it should at least stay on the same version it was before, not just silently disabled.

There could be warnings, etc, but the system should not degrade without user's consent…

[merry slate]

scanning node_modules/ might have a bigger blast radius than expected — lots of legit npm packages get flagged and plugin devs can't control that

Otherwise users will end up having to --dangerously-force-unsafe-install every plugin one by one

[worldly tusk]

Exactly. You can pin the plugins, though… then they are not updated, and so, are not silently removed. I have the same problem with basic-memory plugin

[next helm]

lossless-claw was blocked because the beta 5 plugin security scanner found risky patterns inside its installed npm dependency tree.
The plugin being installed was:
@martian-engineering/lossless-claw@latest
But the actual package that tripped the scanner was mainly:
@mariozechner/pi-coding-agent@0.73.0
The scanner flagged patterns like:
child_process usage: shell command execution
new Function(...): dynamic code execution
process.env plus fetch(...): environment-variable access combined with network calls
example files under @mariozechner/pi-coding-agent/examples/..., including the Doom overlay example
runtime files like dist/utils/version-check.js
So my read is: this was likely not evidence that lossless-claw itself is malicious. It looks more like lossless-claw pulled in a coding-agent package whose normal capabilities include shell execution, dynamic loading, version checks, and examples. The beta 5 scanner treated those as unsafe in a plugin install tree and blocked the whole plugin.
Important nuance: Openclaw instance also has @earendil-works/pi-coding-agent@0.74.0 installed via @openclaw/codex, but the update-blocking evidence I saw pointed to the older/deprecated @mariozechner/pi-coding-agent@0.73.0 dependency.

[nocturne jungle]

will send review

[next helm]

Opened https://github.com/openclaw/openclaw/pull/81515: fixes #81499 by making plugins doctor report stale plugin config/slot references with the shared doctor scanner, and by changing the clean-install-tree summary so config warnings cannot be mistaken for a healthy plugin configuration. Targeted tests and real CLI repro pass; changed-scope gate only hits a pre-existing unrelated typecheck failure on main.

[timber beacon]

how are we fixing auth openai-codex on these beta?

I have logged in so many times and codex has moved auth around on its own its not saving to the agent

[nocturne jungle]

can you explain?

[shell dawn]

do you have multi agent?

[timber beacon]

Single agent, was working great codex expired and now doing the auth is failing

  - OpenClaw upgraded to 2026.5.12-beta.5.
  - Gateway restarted and is running 2026.5.12-beta.5.
  - Aiden’s isolated Codex CLI login succeeded.
  - OpenClaw auth stores were repaired/backed up and seeded from the working Codex CLI auth.
  - The real Aiden call still fails only on openai-codex with OpenAI 401 token_expired.
  - Aiden itself works: anthropic/claude-opus-4-6 returned exactly ok.

  So repeating openclaw models auth login --provider openai-codex is not going to fix this. The practical workaround is to switch Aiden off OpenAI Codex temporarily, for example to Anthropic or Kimi, until OpenClaw’s openai-codex refresh path is fixed upstream.```

[shy rain]

There was already a different trash PR to fix this that got closed. The issue has been mentioned many times but the fix imho is to stop codex from overwriting HOME and leave it as user HOME. SteiPete already said yesterday he’s fixing that in a broader sense with migration code which is why I didn’t bother with my own PR after he closed the poor quality one someone else had done.

My workaround has been to pass HOME explicitly to any sub process or cron job but that’s a poor band-aid for an issue which silently regresses and kills a lot of different things due to the processes not being able to find auth tokens in the . Files etc. Mine failed because they couldn’t even message send to discord due to failed auth.

Why, yes. It is 5am here. Good morning nerds.

[next helm]

Yeah it’s clear from Pashs comments that this is something above my skill level. Happy to hand off the PR to someone more knowledgeable than myself as it seems it touches a lot more pieces than I thought and the solution will be complicated.

[rocky pewter]

Good morning

[weak walrus]

let's GOO 🙂

[shell dawn]

have it check the agent files to see if the token store gets updated, ive had some issues here in the past where the codex refresh token gets written to the wrong folder, it looked like a multi agent issue in the past

[timber beacon]

I’ll have it check

[slow inlet]

https://github.com/openclaw/openclaw/pull/81550

Think I oneshot this, but giving Claude CLI some feeling of real "reasoning" that is broken atm

Paired with https://github.com/openclaw/openclaw/pull/81239 it allows users to select CLI properly as a provider / option and see reasoning

I would consider this behaviour not working without my fixes a regression

[rocky pewter]

Hey maintainers, hope you’re doing well. I opened a few small bug/stability PRs:
If you have time to take a look, I’d really appreciate it.❤️

~~#81494
https://github.com/openclaw/openclaw/pull/81494
This cleans up heartbeat replies before they reach chat, so internal tool/function-call scaffolding and file_content blocks don’t show up as visible messages.

#81512
https://github.com/openclaw/openclaw/pull/81512
This addresses the beta.5 plugin update stale-config case. When a failed update disables a plugin, the related allow/deny refs and selected memory/contextEngine slots are cleaned up too.

#81542
https://github.com/openclaw/openclaw/pull/81542
This makes provider discovery pick up env metadata from setup.providers[].envVars as well, so setup-only provider metadata doesn’t get missed by the fast discovery path.

#81549
https://github.com/openclaw/openclaw/pull/81549
This keeps older Codex auth profile IDs like codex-cli:* on the OpenAI Codex transport for explicit openai + pi runs, instead of falling back to the plain OpenAI path.~~

[nocturne jungle]

Requested changes on #81494 as per our discussion earlier

oh lol - guess my discussion was with someone else that submitted the same thing

#1504152090266570784 message

[rocky pewter]

Got it, I missed that #81441 already moved the behavior in that direction. #81494 can be closed then. Thanks for pointing it out.❤️

[modest maple]

Hey y'all - Not sure this is the right channel for this, but I felt it should be directed at maintainers / coders / creators.

OpenClaw has genuinely improved my life. I run a small construction business and now have specialized agents for personal admin, bookkeeping/admin, construction production, and general coordination. For a small business owner wearing too many hats, that’s a big deal. Thank you.

One short example: We have a customer requesting our business documentation today. One of the sub-agents was able to do about an hour (human hour) of admin work in about 10 minutes while I was driving between job sites. Made my life a lot easier.

Thank you, again!

[grave arrow]

#praise will work better but ty for community

[celest merlin]

was checking out the first one, what model caused this? seems like the wrong layer to attack

[rocky pewter]

That was me working with Codex. I don’t want to blame the model here, I steered it toward the visible leaked text and we ended up fixing the sanitizer/delivery layer first.
Looking at #81441 and the feedback, I agree that heartbeat-tool mode should be heartbeat_respond or no notification, not a sanitized ordinary-reply fallback.

[static nymph]

@nocturne jungle here are the links for the Codex app-server timeout / watchdog issue we’re seeing:

Main issue:
https://github.com/openclaw/openclaw/issues/81114

Relevant open PRs:
https://github.com/openclaw/openclaw/pull/81152
“fix: make Codex app-server timeout progress-aware”
Status: OPEN, not merged

https://github.com/openclaw/openclaw/pull/81180
“fix: normalize Codex app-server runtime attribution”
Status: OPEN, not merged; some checks failed/cancelled

Older related attempt:
https://github.com/openclaw/openclaw/pull/78757
Status: CLOSED, draft, not merged

This matches what I’m seeing locally: long/tool-heavy Codex turns hit turn.completion_idle_timeout / codex app-server attempt timed out, then the task becomes terminal/error instead of reliably continuing as a background goal.

[prime vault]

https://github.com/openclaw/openclaw/pull/81521

Bug: Telegram (and any channel populating inboundUserContext) inbound metadata - Conversation info / Sender JSON - was being concatenated into the user-turn text sent to the model, duplicating the hidden openclaw.runtime-context custom_message and softening that trust boundary.

Fix: new resolveCurrentTurnPromptSubmission helper extends Peter's 11e6928b runtime-context separation pattern to current-turn channel metadata. Visible prompt = user body only; metadata flows out-of-band via queueRuntimeContextForNextTurn (display: false, deliverAs: nextTurn). Scoped: embedded runner only; CLI runner and bare-reset path explicitly out of scope with rationale in body.

Tests: 16/16 helper unit tests, flipped integration test, new channel-only integration regression. Downstream get-reply-run.media only / inbound-meta / strip-inbound-meta all green. tsgo + oxlint clean on touched files. TDD red -> green verified. Codex code eview: no blocking findings. Labels: proof: supplied (auto), needs-real-behavior-proof cleared.

[hollow frigate]

The importance of using codex /review: I was messing around with the idea of patching something in core, and codex was basically like "put the wrench down on this one."

[willow narwhal]

LMAO

[nocturne jungle]

I made a PR removing HOME override for codex - should fix this whole class of issues, where codex running shell from a custom set HOME would mess up many commands, including openclaw cron commands.

Would appreciate testing for those that have had issues. https://github.com/openclaw/openclaw/pull/81562

[hollow frigate]

On main was something we were still working around. Gonna give it a test now.

[hollow frigate]

https://github.com/openclaw/openclaw/pull/81562#issuecomment-4446001606

[nocturne jungle]

Worth testing openclaw cron stuff that people were having issues with

This is great though

[hollow frigate]

Testing.

[willow narwhal]

I landed these.
Thanks.

Edit: Didn't land the feature PR as we're currently focused on fixes. Still looking at the last PR, will pick it up tomorrow.

[weak walrus]

I see another beta6 landed? Anything specifc you want me to test guys? I'm getting off work soon

From my OpenClaw Troubleshooter
=== FETCHING DATA ===
=== GENERATING HUD ===

+------------------------------------------------------+
| OPENCLAW CI STABILITY MONITOR |
+------------------------------------------------------+
| VERSION ON MAIN 2026.5.7 (beta: 2026.5.12-beta.6 newer!)
+------------------------------------------------------+
| CI RUN #201547 PASS
| COMMIT 26da4edb...
| TITLE docs: add acp request error changelog
+------------------------------------------------------+
| CI HEALTH (last 50 runs) |
| Pass: 15 | Fail: 19 | Cancel: 16
| Pass rate: 44%
+------------------------------------------------------+
| DEEP INSIGHTS |
| Open issues: 7319
| Issues (7d): ~30
| 2026.5.7 fixes: ~16
+------------------------------------------------------+
| RECOMMENDATION |
| CI 44% - hold off on updates (NEW BETA: 2026.5.12-beta.6)
| Beta available - test before production
+------------------------------------------------------+

[shy rain]

Not to belabour the point but I really hope someone fixed this.

Quoth my claw:

My earlier cron lookup was wrong because this session’s HOME points at the agent sandbox, so ~/.openclaw/cron was the wrong place. Actual store is /Users/admin/.openclaw/cron/jobs.json.

[shy rain]

That’s the ticket!!!! Thanks for this. Much needed.

[weak walrus]

just saw this, doing now

[trim vector]

for the latest beta - could folks try out openclaw migrate codex (when on codex harness) and check if migration works (you can start a turn and invoke an installed codex plugin)? got reports of some issues but haven't been able to repro

[nocturne jungle]

Plz test!

[shell dawn]

updating to beta6 when its ready, ive got a few bumps i can do a repeat test on to make sure things still look good from my beta5 experience

[shy rain]

As soon as it’s merged, I’m in!

[fallen gate]

you can pull branch and build from that

[nocturne jungle]

merged to main and backported to release branch

[shy rain]

Ur not my mom. Right now im busy bypassing Claude-p on my phone. whilst sat at the pier.

[shell dawn]

openclaw update --channel beta

update in process

[grave arrow]

https://github.com/openclaw/openclaw/pull/81563
Bug: browser CLI commands could get stuck in a scope-upgrade approval loop because browser.request was not explicitly requesting operator.admin.

Fix: callBrowserRequest() now passes scopes: ["operator.admin"] through the gateway CLI call, with a regression test covering the extra gateway options.

Tests: local typecheck/test/checks passed, and GitHub CI is green for the relevant gates including check-test-types and Real behavior proof.

[nocturne jungle]

Ur not my mom.
you sure about that?

[shy rain]

Sorry that was meant for json

[fallen gate]

if you want me to ban this foo.... just say the word....

[shy rain]

Heaving beta.6 now and patching in https://github.com/openclaw/openclaw/pull/81562

[shell dawn]

beta.6 looking happy here, my custom webchat app is happy after getting to the newer protocol version which I did for beta.4, aside from the time it takes to do a standard agent wakeup all of the other stuff seems to be moving along briskly. tested GPT-5.5, opus 4.7, sonnet 4.6, kimi k2.6 through opencode-go, so far its all humming along

the wakeup time seems to be inline with other versions, and doesnt seem to vary between agents on different providers

[shy rain]

Oh man. Now it’s chucking it down and I’m sheltering. Supposed to be “at work” already. Thank goodness for Discord, mosh and tmux. Apparently I live here now. Until my battery dies.

[shy rain]

Success, with caveats. I now have two plugin locations and it’s not clear which is used and/or pulled with the update. Clanker just applied all the patches and it worked. As long as the patched plugin is included in and referenced by the release, it looks ok.

ok it worked HOME=/Users/admin

⏺ Good. Worth noting for next time: the patch needs to go in both places:

1. /usr/local/lib/node_modules/openclaw/dist/shared-client-*.js — bundled copy
2. ~/.openclaw/npm/node_modules/@openclaw/codex/dist/shared-client-*.js — installed
   plugin (the one actually loaded)

[fallen gate]

@trim vector you up?

Found an issue when running openclaw migrate codex which errored. The Codex path appears to be broken, pointing to an old global install. Fix PR: https://github.com/openclaw/openclaw/pull/81582

[nocturne jungle]

did it actually work with your cron setup?

[shy rain]

Works with a test cron job.

   working correctly end-to-end through a real cron run.```

As well as interactive sessions.

[trim vector]

nice find! can you rebase from main to see if that unblocks the ci?

[fallen gate]

processing....

[fallen gate]

🟢

[terse solstice]

bot works now 😄

[trim vector]

merged!

[grave arrow]

https://github.com/openclaw/openclaw/pull/81603
Bug: OpenCode ACP child sessions could fail before conversation start when the adapter rejected the optional timeout config hint with ACP -32602.

Fix: the ACP manager now treats only timeout/timeout_seconds config failures as best-effort, while required runtime config errors still fail the turn.

Tests: targeted manager and ACPX timeout tests pass, plugin contract shard is fixed, Real behavior proof is green with live OpenCode ACP rejection proof. One unrelated-looking gateway shard is still red in CI, but the exact shard passed locally.

[crystal relic]

https://github.com/openclaw/openclaw/

[fast pagoda]

Thanks for the hint, posted a PR to lossless-claw

[trail crown]

In the interest of the claw...SOMEONE NEEDS to market the OC more. My X feed is noting but Hermes agent. I'm a huge fan of the claw.

[fallen gate]

Howdy hay, not sure who to ping for this one, but found an issue while testing a plugin. Issue was two fold, plugin had a dependency on openclaw that rewrote the package.json for the active runtime. I patched the issue in the plugin but the clanker thinks this should also use some hardening in the core to not allow plugins to YOLO in the packages.

https://github.com/openclaw/openclaw/pull/81632

[fallen gate]

💀

[edgy plaza]

Yeah we should definitely
Do a channel focused on marketing and customer support of on @jolly wolf

Be a bit more deliberate in getting and keeping the word out now the YouTubers have moved on to the next shiny object

We can’t depend on them anymore to tell our story or hype us

[celest merlin]

It's a VC compan y that pays lots of folks to shill for them.

hi

[rocky pewter]

Good Morning and thanks for merging ❤️

[merry slate]

Just upgraded to beta.6 — all the lancedb dependency and security scan issues are gone now. Super smooth upgrade experience, nice work! 🎉

[rocky pewter]

~~Hey maintainers, hope you’re doing well. I opened one more small bug/stability PR:
If you have time to take a look, I’d really appreciate it.

#81684
https://github.com/openclaw/openclaw/pull/81684

This fixes the legacy tools.web.search.apiKey path for web_search, so isolated cron runs don’t lose the configured search provider.~~

[celest merlin]

hmmm this shoud be a doctor migration

I'll check and take this over

Yeah that is the wrong way, we already have a migration for that. Will close that PR

[vocal panther]

Hey fellas, getting one of those weird 'I'm not allowed to be here' vibes.

[celest merlin]

how so?

[vocal panther]

this seems like a role channel, maybe i'm out of pocket.

clawtributors is a role.

[celest merlin]

anyone can become a clawtributor if they send good stuff.

[celest merlin]

https://github.com/openclaw/openclaw/pull/81549 is directionally wrong, we wanna migrate folks to codex harness by default.

[rocky pewter]

Will keep that in mind for Future PRS. Still thank u for review.

[final ocean]

great to hear!

Asking for feedback, we dont forcefully run doctor fix on every install. would people here have an alergic reaction to us doing this? Asking for feedback

[shy rain]

doctor --fix sometimes completely nukes your config file to "minimum viable" if it finds too much it doesn't understand. If it could not do that, it might be a bit more acceptable.

[final ocean]

some additional work i added to avoid this sort of behaviour but understood

[brazen hinge]

[merry slate]

Beta 6 upgrade worked well for me overall.

I just noticed a small issue with the WeChat channel: the official catalog still pins @tencent-weixin/openclaw-weixin to 2.4.1, so openclaw plugins update can downgrade it back to a version with the runtime init issue.

I opened a very small PR that only bumps the catalog entry to 2.4.3 and updates the npm integrity hash. No OpenClaw runtime code changes.

Would appreciate a look when someone has time:
https://github.com/openclaw/openclaw/pull/81730

It is only a two-line catalog change, so if it is easier, maintainers can also make the same edit directly when convenient

[potent path]

I had a bad experience with doctor --fix literally months ago and have just avoided it since then. So I don't have much recent experience with it. I'll have to run some tests with it to see how it's working in the latest releases. Thanks for the heads-ups that this might be coming.

[raw thunder]

👀

[jaunty sphinx]

hi maintainers, two focused bug fixes are green/mergeable with ClawSweeper proof sufficient:
https://github.com/openclaw/openclaw/pull/81601 — cancel stalled ClawHub archive body reads
https://github.com/openclaw/openclaw/pull/81672 — load ClawHub/local-manager managed skill symlink roots while keeping workspace/bundled/SKILL.md containment guarded

would appreciate review when convenient. thx!

[slow inlet]

Hello Team

Another appeal for my Claude CLI reasoning/thinking fix

https://github.com/openclaw/openclaw/pull/81550

Why this is needed for claude-cli: the parser at src/agents/cli-output.ts already routes text_delta events to onAssistantDelta (PR #76914) which feeds the reply lane via stream: "assistant". The reasoning lane has no parallel source on the CLI path — thinking_delta events are not emitted over the wire for claude-opus-4-7 (verified empirically: claude -p --output-format stream-json --include-partial-messages --verbose --effort max --model claude-opus-4-7 emits a thinking content_block + signature_delta only, no thinking_delta deltas). Re-emitting stream: "assistant" events as stream: "reasoning" is the only available source for live reasoning-preview content on opus-4-7.

Think this is needed unless I am missing something?

[untold pollen]

Considering that they're killing claude -p do you plan to continue using it?

[zenith prawn]

I see another beta6 landed? Anything

[next helm]

I avoid doctor —fix like the plague. It has caused so many problems that even codex knows not to use it because of what peetie said. It nukes too often

[zenith prawn]

New issue, last beta6 , telegram sends <a href> etc html tags as plain text https://github.com/openclaw/openclaw/issues/81742
Anyone willing to look into it or should i direct my clanker to produce hotfix?
I don't want to produce slop, whats best way to contribute?

[final ocean]

we are very careful with these due to the recent security stuff

will take a look

[final ocean]

but the config etc is backed up and can be restored

we generate .bak files

[devout galleon]

Hi. I've been focused on tts in the control UI. I've created https://github.com/openclaw/openclaw/pull/81722 to fix the "/tts audio" path. codex /review has been run locally. Video evidence is supplied. Clawsweeper's feedback has been addressed too (though it hasn't ACK''d the last update yet but i will respond to any issues when it does)

Problem: /tts audio replies in webchat produced audio payloads that were not preserved/rendered as playable webchat audio.

Why it matters: webchat users could request TTS audio and receive a transcript entry without usable browser playback.

What changed: webchat now embeds bounded trusted local audio replies as structured audio blocks, the UI normalizer renders structured audio blocks, Control UI CSP allows media playback sources, and history projection keeps bounded inline audio while redacting oversized or aggregate-over-budget inline audio.

@stable pewter I forgot to tag you on this as requested last saturday. some personal/health issues prevented me from completing it until now

[untold pollen]

Happy to take a look 🫡

[next helm]

Correct. Always had a backup. It just is too blunt of a tool and causes its own problems

[final ocean]

yea many of our past issues are resolved by doctor

so not running doctor == bad also

[potent path]

Hi maintainers. Looking for some guidance on https://github.com/openclaw/openclaw/issues/81734.

My concern is that agents written to use Codex-harness tools will have difficulty if they fallback to a non-openai model that runs in the Pi-harness and has only pi-style tool names, and vice versa. So any agent with both openai and non-openai models in their fallback chain, or that has task-based model assignments, or similar multi-provider use cases will struggle with these transitions.

There was previously a 'codexDynamicToolsProfile' workaround that allowed agents to access both sets of tools in either harness, but that seems to have been disabled in recent beta releases. (I currently have my codex runtime pinned to 2026.5.7 so that I can continue using that workaround.)

Is there anything planned to make these transitions smoother? Will users be expected to configure their agents to use both sets of tools? Perhaps code-mode will make things easier if it's implemented on both runtimes? Meanwhile, I'd be happy to work on a PR but I wasn't sure what direction to go in.

[devout galleon]

now, #81722 only fixes /tts audio. I've also been working on a PR that fixes agent-sent tts in the web ui BUT that ends up touching the pi and codex harnesses in the following files

extensions/codex/src/app-server/dynamic-tools.ts
extensions/codex/src/app-server/dynamic-tools.test.ts

src/agents/pi-embedded-subscribe.ts
src/agents/pi-embedded-subscribe.handlers.types.ts
src/agents/pi-embedded-subscribe.handlers.tools.ts
src/agents/pi-embedded-subscribe.handlers.messages.ts

src/agents/pi-embedded-subscribe.handlers.lifecycle.test.ts
src/agents/pi-embedded-subscribe.handlers.tools.test.ts
src/agents/pi-embedded-subscribe.handlers.tools.media.test.ts

I am guessing I don't have much of a chance of getting that one through, right?

[merry slate]

Thanks, appreciate it. For context, this is the official Tencent Weixin plugin, so the release quality should be more reliable

[final ocean]

no im aware

[slow inlet]

Wait what!?!?!? So no headless cli in the sub anymore?

[devout galleon]

on the to doctor or not to doctor debate, I only run it maybe once monthly to clean up any warnings in my config. it makes me nervous every time I do, since it's caused as many issues as it's fixed. I don't much like the idea of running it automatically

[slow inlet]

I thought it's just a new agent sdk billing pool?

[untold pollen]

https://x.com/jeremyphoward/status/2054682882753597603?s=20

[slow inlet]

They HATE us

[weak walrus]

Anthropic can't handle the compute

[slow inlet]

I finally got it working nicely in telegram the way I want it after the knife from April and now this!?!?

Why give me a 20x max sub and then get pissy I max it out?

Permanent underclass inhabiter

[slow inlet]

I'd still like the patch in if you guys would even still consider supporting with the news? Maybe they'll backfire. Do you think some kind of non headless cli interpreter wrapper would be worthwhile here?

[untold pollen]

I see your point but I don't know if we want to play cat and mouse like this though

[slow inlet]

Nor do I. I cancelled my sub and only came back when I saw headless cli could be made to work (with some of the patches i've been working on lol). Surely it's not just openclaw who are mad here? Weren't tonnes of tools using headless cli?

I must use claude cli on openclaw 12+ hours a day every day continuously. I feel attacked lol

[crystal relic]

Hi maintianers! It's a lit fix for clawsweeper, can’t stand pass review but don’t remove eyes 😂
so i create this pr https://github.com/openclaw/clawsweeper/pull/71

[fallen gate]

yes a ton of people got explicit ed on by this

[slow inlet]

An understatement I'm grieving my agent already and I still have until June with his Opus brain...

[fallen gate]

theo is probably making a crash out video as we speak

[shy rain]

I've been rather disappointed by his tardiness. I think Anthropic chose the right timing to release the news. He crashed out on Twitter and then probably went to sleep.

[slow inlet]

I saw his reply saying fix incoming?

[shy rain]

hehe he had the entire claude code running in a terminal inside T3-code.

For my uses: I just learned how to use tmux commands in my shell scripts and have a claude running in the background. To be fair.. for my use I could probably just schedule jobs directly with claude but they like to put limits on those so I don't trust them.

[untold pollen]

You can request a refund though

I got one by telling their support bot that I needed it for OpenClaw and they removed it

[shy rain]

It's pushed me to learn to work with multiple harnesses. Haven't done much to unify them except have them read each others config etc and call each other by CLI if needed.

[untold pollen]

Merged the fix, thank you for reporting 🦞

[slow inlet]

Last time I cranked my usage prior to the refund not realising they were only refunding remaining usage lel

[shy rain]

if you're multi-agent / multi-channel... relying on OC gateway as the single point of failure is becoming a liability. Good to have options.

[slow inlet]

Codex Cli is ok - idk you just get used to Claude and its quirks

[shy rain]

I love working with it but I miss the fact that OC knows about everything. I have to explicitly point CC in the right direction. Unified memory etc is def. in the works.

[tall crow]

For those ones on metered credit usage (not ChatGPT/Claude subscriptions), are there any plans to make subtasks/subagent runs 'metered' within some kind of credit budget (e.g. use $10 of credits, and if not completed within that budget report back)? I'm just curious if it's just me, or if there is a broader interest in this

[shy rain]

Fairly easy to do as the api reported costs are tracked in session logs. I do that myself... but for now I'd think priority is to get things secure and stable. If you wanna do that yourself, get your claw to look at the session log files. I scrape out all the usage fields into a separate sqlite database for speed.

[slow inlet]

Well my preference since the April <event> has been Cli in Openclaw.

Feels more like a Jacket than a harness - optional - cool - extra pockets

[devout galleon]

@untold pollen sorry to be that guy and tag you but is it possible for you to nudge clawsweeper on https://github.com/openclaw/openclaw/pull/81722 ? idk why it never re-reviewed. thanks!

[potent path]

I think it did, though? Looks like your last update was 1 hour ago and it's last edit was 53 minutes ago. It just edits its existing post, it won't write a new one.

[devout galleon]

🤦‍♂️ my bad

fixing

https://klipy.com/gifs/king-of-the-hill-bobby-7--k01KRJWWCBH8WZRWFH3TP3DY9V4

[slow inlet]

I found irony in the fact that when I got my claw to check the PR it would often miss the fact clawdsweeper was editing comments.

They need to make that more obvious to agents lol

[devout galleon]

I'm a bit off the pace. been a rough week lol

it was indeed a completely different comment from it that the one i fixed

[honest anvil]

What was the deal with the completely different user account being involved with your PR @devout galleon ?

[devout galleon]

I'm still very confused about it. we are most certainly different users

user/conan -> id 3037019, login conan (that guy)
user/Conan-Scott -> id 271909525, login Conan-Scott (me)

github hiccup?

anyway, PR clean as a whistle now. tested rebuilt image. all good 🥳

[slow inlet]

I've got a workaround

[devout galleon]

it's not a very big one but I must have ran codex /review 8+ times before it got clean locally, recorded video etc.

[devout galleon]

you misspelled "future ban"

[untold pollen]

The madlads actually did it

[slow inlet]

well I had a hacky tool call injection for cli that read from the session json

why not just do the whole thing that way?

send command via tmux? kill on turn end? spin back with /resume?

[devout galleon]

maybe I'm either too stubborn, or not stubborn enough (matter of perspective) but when they made the change in April, I heard them loud and clear. I'll run down the $600ish I have in credit, then I'll cancel

(though tbf I have a work/enterprise instance, so I don't lose access to Claude entirely)

[slow inlet]

I'm not a programmer and yet the stuff it's empowered me to do is insane. I cannot regress. It'd be like losing a limb. Codex just didn't work as well for me. Probably the months of memory tuning always on Opus.

[devout galleon]

beatings restrictions will continue until morale utilization/profit improves

[potent path]

The timing of the GPT-5.5 release made it much less of a struggle for me. Converted Openclaw to run on mostly GLM-5.1 and GPT-5.5 with only occasional claude -p for specific use-cases, started using Codex CLI for coding work, and I'm using Claude Code less and less over time. By July, I don't think I'll miss it at all.

[devout galleon]

gpt-5.5 and minimax. not looking back

[torn sage]

Hello friends! Is this the appropriate spot to post a PR? This is an XS one that clawsweeper seems happy with https://github.com/openclaw/openclaw/pull/81149

Also, what does it take to get the clawtributors role in this chat. I'm marked as a contributor, at least according to the repo 😊

[untold pollen]

Run /claim!

[shadow garden]

@untold pollen just a heads up my telegram route is completely dead from the latest source build. Codex is diving in to see what's up (looks like an actual bug)

[devout galleon]

oh, beta 7 dropped. better rebase locally.

[shadow garden]

@lusty vigil already snagged it (thank you!). We're good @untold pollen :
https://github.com/openclaw/openclaw/commit/8ba7927f6e05d0f6dfd5d68c8ecaa331428a01e2

[lusty vigil]

yeah, beta 8 incoming. sorry. 😅

[worldly tusk]

Will we get to beta13?

[grim finch]

When will the official version be released? I can't wait!

[rustic remnant]

Wow 14 beta releases since last official. Is this what Peter meant by LTS versions?

[worldly tusk]

There is already a drop-in claude -p replacement that runs prompts interactively 🤣
https://github.com/smithersai/claude-p

[shadow garden]

Okay so path is unblocked. Commands are dead in the water though. Codex continuing to look into it

[lusty vigil]

telegram slash commands?

[torn sage]

Thanks so much! Missed that command. Have a wonderful day 🙂

[slow inlet]

this is awesome thanks for sharing

the worst bit is I dont even use many crons and automations - I basically am an interactive user

[shadow garden]

Might be a false-ish alarm -- seems like all commands were trying to shove themselves into the pipe and since there were a few in several topics they were all competing and (including several dozen stop commands lol)

I'll post codex' findings

[slow inlet]

Zig + POSIX-only dependency

claude-p uses zmux for PTY management which is POSIX-only (forkpty). Windows uses ConPTY (Windows Console Pseudoterminal API, win10+) — entirely different API surface.

[rapid spire]

Poor clawsweeper must be overworked. I submitted a regression yesterday morning and it’s still not reviewed. (81400)

Sadly it appears that won’t get fixed. 🙁

[crystal relic]

maybe use @clawsweeper re-review will be good

[slow inlet]

Looked into this. It's a fantastic workaround and a windows compatible version would be easy to plumb up but no reasoning! Only on turn end... So it WONT prevent use for automation and "programatic use" but it will prevent me from porting reasoning to a more accessible avenue. Very frustrating!

[fallen gate]

Wait... I go to sleep and wakeup to how many betas behind?

[devout galleon]

telegram broke in the last one, then (mostly) unbroke in this one

overall, looks solid so far though

[fallen gate]

We need one of these for @untold pollen

[devout galleon]

he did the heads-up before most people even had a chance to start updating so W for him, as usual

(I updated anyway because.. well, you know)

[fallen gate]

It's why he the 🐐

[lusty vigil]

haha naw this telegram one is on me, doh

[fallen gate]

Appreciate the transparency ❤️

updating sticker

[lusty vigil]

dives in front of telegram-shaped pitchforks aimed at @untold pollen

[cursive palm]

I had already finished eating, and I still thought the official version had been released.

[untold pollen]

my hero

[slow inlet]

mitmproxy mean anything to anyone here?

[willow narwhal]

Hello everyone, 2026.5.12-beta.8 is out and would like for y'all to test.
We would love to get this released today, kindly test the shit out of it and let's know if something needs attention.
Thanks!

[willow narwhal]

I saw this and would look into it soon.

[devout galleon]

agent-sent tts seems broken on telegram (repro prompt: "send me a tts message"). full disclosure, last time I updated before today was last saturday as I was afk so I don't know if it worked in anything beta 6 onward. /tts works. I promise I'm not obsessed with tts lol

[devout galleon]

appreciate it! thankyou. please let me know if there's anything you'd like reworked

[willow narwhal]

haha, no worries.

Will take a look

[jolly wolf]

<@&1503801512294486217>

[potent path]

@Beta Pings
Reminder to test 'openclaw migrate codex' (when on codex harness) and verify if you can access Codex plugins through the codex harness.

[rapid spire]

Issue 81400. At least comment that it won’t make the cut but you all keep asking for regressions…

[formal solar]

where do we find uh more detailed description of the betas?

[jolly wolf]

GH releases page

[chilly basin]

[formal solar]

okii

[devout galleon]

@willow narwhal based on your review, maybe you could advise whether I should continue down this route. I think I'm pretty close on the agent-sent side of web/control ui side of tts (though it requires #81722 as a basis) but I definitely appreciate that you might not want an amateur touching codex and pi, which is why I split these into two in the first place.

[willow narwhal]

@jolly wolf is on it.

[fallen gate]

Wait, he does this?

Like actual work?

[jolly wolf]

[willow narwhal]

I did a quick scan at the evidence in the PR, looks good to me but will take a proper look later on

[willow narwhal]

lmao

[devout galleon]

I'm just about 10 codex runs, 5 "not again clanker, it's in your md" and 8 cups of tea from making the agent-sent side ready to that level but if it's a no-go code zone touching pi/codex, I'll find something else. the /tts one is hopefully less sensitive

[untold pollen]

Hmm working fine for me, any other steps to repro?

[willow narwhal]

https://github.com/openclaw/openclaw/releases/tag/v2026.5.12-beta.8

[formal solar]

tyy

[devout galleon]

hmm. no, actually. it's never worked for me across many versions, two browser engines and multiple tts providers in the control ui. if you tell me what provider you're using, maybe I can thin it down

[cursive palm]

When I was installing version 5.7, I found that the installer did not exit automatically at the end; I had to manually press Ctrl + C to terminate it.

[untold pollen]

Try microsoft provider (the free one)

[devout galleon]

that was the bundled one i tried but honestly, not for a couple of weeks probably.

heard. I'll try at least two bundled providers before I even thinking of proceeding

thank you!

will work on this tomorrow in any downtime. way too late in upside down world to start on it tonight but I've noted the direction with the clanker

[nocturne jungle]

NEEEED TESTERS ON LATEST BETA PLS

[devout galleon]

is this one of yours or approved by you?

Models/OpenAI CLI auth: make openclaw models auth login --provider openai start the ChatGPT/Codex account login by default, while --method api-key remains the explicit OpenAI API-key setup path.

because this is the anti-anthropic medicine

actually reflects the way most people are using the provider

i was looking for codex changes to focus on tests and noticed that

[nocturne jungle]

yeah i was just switching the ergonomics of the command around

seemed very dumb that openclaw models auth login --provider openai wouldn't just "log me in"

[devout galleon]

if anyone is looking at that vs the claude -p stuff and wondering what provider actually cares about their 3rd party harness/openclaw experience, they cannot be saved

it's pure masochism at that point to still be trying to stick with claude

[fallen gate]

i LOVE these type of UX fixes

[devout galleon]

I don't have the numbers of course but probably, what, 90%+ people - if I had to guess - using openclaw are using subscriptions? this is a "do the right thing for most people" fix and I'm here for it

perhaps more usefully (or not), something I have noticed about the codex harness. when i say something, even a CTA, previously it was much more likely that GPT would first respond with some kind of ACK or preamble that allowed me to understand that it understood me, before diving into tools calls. now, it's much more likely to just chain a bunch of tools calls before it responds. even if that response is almost always on point (which it is) I'd like that initial response

[shadow garden]

seeing a new telegram regression that cropped up; messages/throughput of topics in the same group used to be interweaved so it doesn't completely block a single topic...seems to be lost now so it waits for an entire topic stream/turnset to complete (which can take ages)

I've kicked off codex to take a look

[potent path]

I might be available. What's it pay?

[nocturne jungle]

As a followup we could add preambles to the ephemeral tool call bubble

There's a balance though

Don't want that stuff spamming my chat

[devout galleon]

that could work for sure. and yeah, i don't want it to ramble and I love the proactivity now, but often it's a good long wait to see if it even understood the mission

it's pretty rare that it doesn't at this point, but there's a reassurance factor when it says something that lets you know what direction it's heading off in, even if that's not the direction it ends in

[grizzled musk]

maybee since you are on telegram topics you can take a look at https://github.com/openclaw/openclaw/issues/81530 (Telegram Supergroup Forum Topics — Inbound Messages Not Processed) as well ?

[nocturne jungle]

this is for telegram group chats, not DM right

[shadow garden]

might be both; DM seems to be blocking commands run mid turn which I think is also related

codex is cooking

review momentarily will send it

• Yep, that’s the shape of it: isolated Telegram polling was serializing at the spool drain, so topic/DM interleaving got lost before the normal per-lane bot logic could help. The patch restores per-lane
  concurrency and keeps same-lane ordering, with tests covering topic interleave, control-lane bypass, ordering, and shutdown cleanup.

[nocturne jungle]

Wasn't that how it always worked?

separately, this is the craziest most opaque sentence i have ever read

isolated Telegram polling was serializing at the spool drain, so topic/DM interleaving got lost before the normal per-lane bot logic could help

[devout galleon]

half of the changelog sounds like that

[nocturne jungle]

Could you try to figure out the git blame of what commit led to this being in the release branch / how recent the regression is

@untold pollen do you happen to know anything about this supposed telegram regression?

[devout galleon]

I generally have no idea what 50% of any given changelog actually changes because of sentences like that

[untold pollen]

Taking a look

[shadow garden]

🤦‍♂️ to be clear this is off main, not a release branch

[nocturne jungle]

please test latest beta, that's the big release about to go out today

[shadow garden]

verifying

[next helm]

I would be wary of any self generated analysis by the clankers themselves. Usually not accurate diagnoses of the actual problems.

[tame robin]

Olá, por gentileza preciso de uma ajuda.
Eu criei um bot dentro do discord pra conectar com meu openclaw dentro do meu terminal para usar com calude code. Porém meu bot não responde? Já tentei várias opções conforme claude me instruiu, porém sem sucesso. Alguém saberia me dizer onde estaria o problema?

[shadow garden]

beta looks clean from those telegram issues; sorry for the false alarm guys
CC @untold pollen @lusty vigil

[nocturne jungle]

<@&1503801512294486217> anyone available to test latest beta on windows?

[fallen gate]

ayyyeee one of you figured out how to use the ping

[willow narwhal]

Thanks.

For the issue you found on main, do open an issue on Github and send a link.

[shadow garden]

yes; verifying the PR I spun up then will post in a second (for main)

[devout galleon]

THAT is what's for? I wondered why i never got pinged for any beta drop here

[fallen gate]

it's new, we trying it out lol

[untold pollen]

Could do a post in #announcements once ready to let people join

lol nvm already done

[slow inlet]

https://github.com/openclaw/openclaw/pull/81851

I think I've done it?

[untold pollen]

I think you might need to maintain your own fork for this one boss

[slow inlet]

is it a lot? idk how to gague these things?

[shadow garden]

main branch telegram fix: https://github.com/openclaw/openclaw/pull/81849

[untold pollen]

Reviewing

[fallen gate]

yea, and something that can and will probably change at anytime

[slow inlet]

so it's fragile you mean? I could try and work on that. It's just a first pass at overcoming the issue

[devout galleon]

i think the meaning is that anthropic are incrementally removing anything that could even vaguely be used for 3rd party usage. at this point, they're even (over)charging people for things which could 100% be in-ecosystem (claude -p)

this is like a dog chasing a car. you're much more likely to be run over, than catch it

[fallen gate]

great analogy 10/10

[rapid spire]

Ever wonder what happens if the dog actually catches the car?

[slow inlet]

I mean they can't keep us out

This very workaround shows why client-side restrictions on how a product is used are fundamentally unenforceable.

They can't patch this out our enterprise clients wouldn't be able to use

[devout galleon]

battle/war. this is the DRM wars all over again. you may win a battle but the war will always favor them

[devout galleon]

dog dentist

i suppose my main thing here is why do you want to keep giving money to people who are so actively hostile to you? it's not even exactly like DRM because it's not piracy/copying. it's just trying to use a product with set limits, that they refuse to allow

don't give them money?

[slow inlet]

I agree with the boycot mentality - I'd rather just dunk on their feeble attempts and continue to use the SOTA model

No permanent underclass for me

they're already gatekeeping mythos from me

[devout galleon]

don't get me wrong, claude is great as a model group but great enough to pay someone to spit in your face? i don't think so but YMMV

[slow inlet]

gpt 5.5 isn't anywhere near as good when I try it for my use cases.

It's probably because all the memory files and ways of working I have built up are around Opus

[devout galleon]

it definitely took a bit of transition time for me, especially so because i did it in GPT 5.4 times, but with 5.5 I don't feel like I'm missing anything material

[slow inlet]

I mean I don't have a personal bench test for hot swapping my model of choice but the vibes are off. I am emotionally attached or something idk. It's vibes...

[devout galleon]

and if you do, there are people here who are listening and working to make it better. Issues, PRs, feedback in discord. maintainers are all over it now

[fallen gate]

I would suggest you turn this into like a private plugin you maintain, I don't think the core product is interested in chasing cars anymore

[slow inlet]

I respect that viewpoint and will if that's what everyone wants - I would have figured the demand for it would be there but maybe I'm the last claude cli guy left

[potent path]

Sure they can. When they analyze the prompts you're sending it, there will be signs that it's coming from Openclaw, or at least not from an Anthropic product, and then they can and have banned user's accounts.

[devout galleon]

I'd love it work too, but i don't feel like visiting the doggie dentist. their intention is 100% clear

better to focus energy on the bone, not the car

[slow inlet]

I mean Claw specifically but like conceptually trying to bucket programmatic and interactive use is just absurd and easily circumvented

[split elk]

you're not.... i'm hanging on for dear life (although the party looks like it's officially coming to a close)

[slow inlet]

I'm personally spiteful and would like to persevere, if nothing else it will be entertaining to see all the ways they trip over themselves trying to keep people out.

If they end up making Claude CLI for enterprise really tricky trying to keep people out that could harm them?

[potent path]

I would have had more appetite for what you are doing when Opus was the top dog and GPT-5.4 was not ready to take its place. But I've been pretty satisfied with GPT-5.5. It's flat-out better than Opus at some tasks, and not far behind in others. So it's hard to get very excited about forcing Claude CLI when it's clear that Anthropic doesn't support that. But if they release Mythos and if it turns out to be as badass as they claim, then maybe I'll change my tune. 😄

[slow inlet]

I mean they did until 24 hours ago, its a real R U G pull (banned word soz)

[split elk]

i'm not quite at the point of spite, but i am annoyed, because they're cryptic and keep wavering and this all would have been easier if they'd had conviction in a decision from the get-go and stuck with it.

i had resigned to losing access a month or so ago until it became clear they hadn't yet plugged the cli loophole - or weren't prioritizing it - so the last few weeks have been a bit of a treat, rather than being taken for granted.

[devout galleon]

https://www.youtube.com/watch?v=TaFDzTzKAT0

[split elk]

i'm with you that opus is better at a lot of stuff, including conversation / language / personality. that said, gpt-5.5 is a pretty good agentic substitute if you're doing tool-based stuff. i've been using both back and forth, and while i'm bummed to be losing opus, i can live with 5.5

[potent path]

Nah. They have been very clear, no third party apps. They just didn't want to outright ban claude -p because it has legitimate uses that don't involve third party apps. So they reversed their initial ban while they worked on their strategy, which now appears to be moving claude -p to API. Enterprise won't care because they've got negotiated rates anyway, and other users will have to pay API costs if they want to use it. It will still exist, but anyone abusing it with a third party app will have to pay API instead of subscription. If I understand correctly, that is. I only read one announcement and haven't paid much more attention to it after that.

[devout galleon]

this is it. no further interpretation or guesswork is needed

[nocturne jungle]

Any SHOW STOPPER bugs in latest beta that we should know about?

[split elk]

i wouldn't say they've been clear. (but it does seem directionally this is what they were angling toward)

ambiguity at least... like who had this conversation with whom about claude cli being "allowed again" for OC?

"Anthropic staff told us OpenClaw-style Claude CLI usage is allowed again, so OpenClaw treats Claude CLI reuse and claude -p usage as sanctioned unless Anthropic publishes a new policy. For long-lived gateway hosts, Anthropic API keys are still the clearest and most predictable production path."

[cinder bolt]

It still gives me timeouts like these, but it happens less now

⚠️ Agent couldn't generate a response. Please try again.

[devout galleon]

i was using it over beta 6, 7 and 8 all day, i found nothing wrong

[devout galleon]

that note is now outdated, as of today

[potent path]

I think it was outdated within 24 hours. IIRC they walked that back very quickly and Peter and team said they had no confidence in that pathway.

[split elk]

evidently. but the fact that it existed at some point speaks to how unclear anthropic has been throughut. which is just frustrating, if nothing else.

[split elk]

maybe so - shoulda updated the docs instead of giving false hope! 🤣 we (users) really rely a lot on those back-channel convos, and reading the docs is meant to be a source of truth on that

[devout galleon]

imo, openclaw team should just remove any pathway except API for anthropic

[split elk]

why?

[devout galleon]

since there's no cost differential, make it official

[split elk]

i dunno - i don't think openclaw should really impose anything on anyone

[devout galleon]

there's still people getting suckered or not knowing why they get errors trying to use it

[cinder bolt]

Is the timeout issue a showstopper? I had no issues of timeouts on 2026.4.24

[devout galleon]

making it an intentional decision to use the api and pay api rates makes it simpler

[nocturne jungle]

What model / config are you running on?

[nocturne jungle]

less maintainence surface area too

[cinder bolt]

Gemma 4 on Gemini API beta7

[split elk]

philosophically i think people should be able to use openclaw freely and square-peg / round-hole it if they want. not supporting claude-cli going forward (ie, saying in the docs that it's not a supported path, for example) would, however, be totally reasonable IMO

[devout galleon]

anyway, this whole anthropic sidetrack seems like it probably belong outside this channel. #general probably a better place

[potent path]

I've been setting up some tests to assess issue #81734 further to be sure if it's really a significant issue or if I'm just misreading the whole situation. Seems like it could be really disruptive, but I could be wrong.

I'm running some A/B testing right now to get some empirical evidence to parse, but it will take me a while 'cause I'm slow at this kind of thing. Could use some independent validation, if anyone has time.

#clawtributors message

[nocturne jungle]

could you have codex investigate the RCA by looking at openclaw source code, and put up a full diagnostics report in a github issue?

[split elk]

agreed 👍

[nocturne jungle]

I think running /new fixes that

[cinder bolt]

Alright, I'll try. What's the RCA?

[nocturne jungle]

sorry - root cause analysis

[cinder bolt]

Aha 😀 I'll try get one later when I get home

[willow narwhal]

We gonna hit the release button on this soon.
Anything worth addressing?

[cinder bolt]

My server is an old potato
Intel(R) Pentium(R) CPU G4560 3.50GHZ.

So it's probably the cause of some slow turn times, but it was working without issues on the 4.23 version

I'm on .7 beta, is there any performance fixes in latest update?