WeeklyCLAWEp. 14
Welcome + structure
Welcome to Weekly Claw
Andy's out. The show goes on.
Quick welcome first. We'll set the room, then cover the outside OpenClaw news, the changelog and developer experience, and the community block.
- 1. Welcome, housekeeping, and how today will run.
- 2. OpenClaw in the news: Anthropic, Google, SMB automation, and security pressure.
- 3. Changelog and dev experience: what changed, what got better, what still hurts.
- 4. Events, welcomes, and community rules if there is time.
WeeklyCLAWEp. 14
OpenClaw in the news
OpenClaw is now the reference design
Copied, metered, packaged, attacked.
This week's outside signal was bigger than one article. Anthropic created a billing lane for third-party agents. Google is reportedly building an OpenClaw-shaped assistant. SMB press is packaging OpenClaw as business automation. Security writers are treating skills like a supply-chain problem.
That is what mainstream looks like: adoption, competition, pricing pressure, and risk all arrive together.
2026-05-15WeeklyClawEp. 14
May 15, 2026
Anthropic news
Anthropic news
Anthropic reopened the door
But it put a meter on agent usage
The important bit is not just that Claude subscriptions can touch third-party agents again. Anthropic is separating normal chat use from agent-runtime use and attaching dedicated Agent SDK credits to that lane.
Claude subscriptions are no longer treated like unlimited agent runtime. That is the news.
- VentureBeat framed it as Anthropic reinstating OpenClaw and other third-party agent usage on Claude subscriptions, but with capped Agent SDK credits.
- That validates the OpenClaw shape: tools, memory, approvals, and long-running loops are now a priced infrastructure lane.
WeeklyClawEp. 14
May 15, 2026
Theme 1
Theme 1
Codex moved closer to the center
Auth, app-server, code mode, and session continuity
The real 2026.5.12 story is Codex no longer feeling like a bolt-on runtime. It is becoming one of the default ways OpenClaw thinks, codes, resumes, and authenticates.
- OpenAI setup now starts with ChatGPT/Codex login by default, while API key setup stays explicit.
- Codex app-server runs picked up OAuth/profile reuse, source replies in WebChat and TUI, MCP server projection, and safer auth-refresh fallback.
- OpenClaw can bind to node-backed Codex CLI sessions, which matters when work is already alive on another machine.
- Code Mode showed up in community chat as the obvious next step: "code mode is sick and we get it out of the box in openclaw."
WeeklyClawEp. 14
May 15, 2026
Theme 2
Theme 2
Runtime quality became the release theme
Less spooky state, fewer silent failures
The 2026.5.14 beta reads like a cleanup crew for every weird thing that happens after agents run for real users: stale auth shadows, plugin load failures, compacted sessions, malformed provider JSON, and channel-specific ghosts.
- Per-agent bootstrap profiles let operators tune context injection and bootstrap budgets without rewriting global defaults.
- Gateway method registry now preserves scope metadata and keeps plugin-owned RPC methods from colliding with core handlers.
- Status and doctor now surface stale auth shadows, failed channel plugins, pending update restarts, and LaunchAgent updater leftovers.
- Release validation grew Docker journey lanes for onboarding, plugin installs, upgrades, media, memory, and marketplace flows.
WeeklyClawEp. 14
May 15, 2026
Theme 3
Theme 3
Channels got practical fixes
The product is where messages land
A lot of the work was unglamorous and absolutely necessary. If an agent cannot speak clearly in the channel where the user lives, the rest is theater.
- WhatsApp gained status reactions for queued, thinking, tool, done, error, and compacting lifecycle states.
- Discord fixed token-secret startup reporting, stale draft cleanup, thread-name sends, and voice-message Ogg transcoding.
- Telegram isolated polling, HTML fallback text, direct request timeouts, and group media handling all got repaired.
- Slack now avoids surprise link unfurls, preserves finalized draft previews after tool warnings, and routes plugin modal submissions.
WeeklyClawEp. 14
May 15, 2026
Theme 4
Theme 4
Provider and model edges got less weird
MiMo, OpenRouter, Azure, Kimi, JSON, and safe origins
The week had a pattern: make more model surfaces work, then make their weird failures legible instead of leaking raw parser pain into operators.
- Xiaomi MiMo reasoning-only finals now become visible answers, including compatible proxy routes.
- OpenRouter reasoning replay gets normalized without breaking valid pass-back.
- Azure OpenAI Responses defaults to the current preview route when versions are unset.
- Malformed JSON from web search, media, provider APIs, Teams, Google Chat, Firecrawl, Tavily, Perplexity, xAI, and more now returns owned errors.
WeeklyClawEp. 14
May 15, 2026
Theme 5
Theme 5
Startup and plugin paths stayed on a diet
Cold paths are product work
OpenClaw keeps learning the boring lesson that fast status, fast startup, and lazy loading are developer experience features.
- Canvas HTTP host, media resolver, CLI, and runtime modules now lazy-load instead of taxing Gateway startup.
- Protocol validators lazy-compile on first use instead of burning cold-start CPU on every schema.
- Resolved skills are cached across warm turns by redacted effective config.
- Plugin metadata snapshots stay fresh without forcing repeated full discovery across large plugin sets.
WeeklyClawEp. 14
May 15, 2026
Security + ops
Security + ops
More fail-closed defaults
Less ambient trust
Security work this week was concrete: narrower origins, stricter paths, less implicit widening, and better repairs when installs drift.
- Custom provider baseUrl trust is scoped by exact origin, so LAN and tailnet routes work without broad private-network access.
- Provider catalog entry paths and node platform IDs are constrained.
- World-equivalent Windows ACL principals stay classified as critical risks.
- CLI/update can finalize externally swapped runtimes and run doctor repair before reporting completion.
WeeklyClawEp. 14
May 15, 2026
Concrete bits
Concrete bits
Names, commands, config keys
The engineer-proof slide
If the room asks what actually shipped, use these. They are specific enough to survive contact with engineers.
- agents.defaults.runRetries and agents.list[].runRetries for embedded Pi runner retries.
- acp.fallbacks for backup ACP runtimes before first output.
- channels list --all for bundled and catalog channel visibility.
- OPENCLAW_HEAVY_CHECK_LOCK_SCOPE=worktree for independent heavy-check locks.
- memorySearch inputType, queryInputType, and documentInputType for asymmetric embedding endpoints.
- channels.slack.unfurlLinks controls bot link preview behavior.
WeeklyClawEp. 14
May 15, 2026
Community + X
Community + X
Fast got measurable
The official account led with speed, not vibes
OpenClaw posted that the latest release is about 3.5x faster, backed by end-to-end RTT tests every 6 hours over real message channels. That matters because speed claims usually rot unless they are measured continuously.
The latest OpenClaw release is ~3.5x faster.
- Signal: Telegram bot-to-bot RTT tests, every published npm release, every 6 hours, running on Blacksmith CI.
- Good framing: the team is not just optimizing. It is catching slowdowns before users feel them.
WeeklyClawEp. 14
May 15, 2026
Maintainer signal
Maintainer signal
Codex review became part of the loop
Peter showed the sharper workflow
Peter posted about building media storage into Discrawl, letting Codex say it was done, then running his Codex review skill. The useful bit is the second pass, not the first patch.
codex said it is done, then I used my codex review skill...
- This matched clawtributors chat: "codex review in a loop is the bomb."
- The lesson: agent coding without agent review is just faster confidence. The loop is the product.
WeeklyClawEp. 14
May 15, 2026
DX cold open
DX cold open
Contributors were doing real work, not sightseeing
The community was in the trenches
The clawtributors export was not hype chatter. It was PR asks, beta testing, model confusion, tool visibility, Discord truncation, plugin risk, and review pressure.
- 3,160 messages over the last 7 days.
- 189 beta mentions, 164 review mentions, 267 test mentions, 150 tool mentions.
- People were not asking only what shipped. They were asking whether the fixes actually fixed things.
WeeklyClawEp. 14
May 15, 2026
Review pressure
Review pressure
The queue became visible
PRs are now part of product experience
A healthy project creates a new problem: contributors ship, then need routing. When PRs dominate the room, review speed becomes developer experience.
- "Guys, send PRs my way."
- Multiple ready-for-review asks came in around gateway startup, admin HTTP RPC, cron failures, Telegram HTML responses, and plugin fixes.
- A contributor floated two-review signoff per PR. The exact mechanism matters less than the signal: review bandwidth is a bottleneck.
WeeklyClawEp. 14
May 15, 2026
Beta cadence
Beta cadence
The community asked for slower truth
Fast releases need proof, not just velocity
One contributor asked whether betas are just checking if things break, or checking whether fixes actually fix what they were supposed to fix. That is the right question.
- "Am I the only one that feels like the betas are too quick?"
- The useful next loop is release notes tied to test asks, affected surfaces, and who can verify them.
- Speed is only impressive if the validation path keeps up.
WeeklyClawEp. 14
May 15, 2026
Tool visibility
Tool visibility
Users still hate silent work
Verbose is a workaround, not a product state
The most human complaint this week: a bot can spend ten minutes doing thirty tool calls and the user sees nothing. That feels broken even when the backend is busy.
- Discord reply delays, hidden tool output, generated media handoff misses, and visibleReplies confusion all showed up in one morning.
- ToolProgress helped, but users should not need War and Peace JSON to trust that the agent is alive.
- The product read: progress visibility is part of reply quality.
WeeklyClawEp. 14
May 15, 2026
Code Mode
Code Mode
Code Mode became community language
The harness surface is becoming normal
Code Mode was not buried in release notes. People were posting about it, asking whether ClawSweeper had it, and treating it as the obvious OpenClaw coding path.
- PR #80600 became shorthand for Code Mode for all.
- Community phrasing: "code mode is sick and we get it out of the box in openclaw."
- The next question is not whether agents can edit code. It is whether routing, auth, and review stay sane around that power.
WeeklyClawEp. 14
May 15, 2026
Plugin risk
Plugin risk
Extensibility still needs guardrails
People want power without footguns
Plugin and provider chatter had a split personality: excitement about governance/runtime plugins, and anxiety about installing too much into a live agent.
- A contributor shared a runtime governance plugin for action validation.
- Another user felt overwhelmed after installing many plugins, memory plugins, voice plugins, and Deepgram.
- Peter was blunt: plugins are an additional risk surface, start with shipped ones and add carefully.
WeeklyClawEp. 14
May 15, 2026
Docs + auth
Docs + auth
Docs lag creates archaeology
Back-channel truth is not enough
A sharp docs point came through: users rely on docs as source of truth, but some supported paths had already shifted in practice.
- Discord config docs still described old valid fields.
- Codex/OpenAI auth naming confused users who had openai, openai-codex, and codex routes in the same install.
- Doctor warnings for per-agent auth shadows are the right kind of UX: turn hidden state into a named problem.
WeeklyClawEp. 14
May 15, 2026
Validation infra
Validation infra
CPU became strategy
The bottleneck moved downstream
Between Blacksmith RTT tests, Docker release lanes, Codex review, Crabbox-style test boxes, and community beta testing, the meta-story is validation capacity.
- Agentic coding raises the change rate.
- CI, tests, review, and real-channel checks have to absorb that rate.
- The best OpenClaw work this week made validation repeatable instead of heroic.
WeeklyClawEp. 14
May 15, 2026
Close
Close
30-second presenter close
Read this verbatim
This week, OpenClaw got faster and less tolerant of silent failure. The headline release was Codex becoming more native: better auth, better session continuity, app-server fixes, Code Mode, and review loops. Under that, the team kept tightening the real surfaces: Discord, Telegram, Slack, WhatsApp, providers, plugins, startup, update repair, and release validation. The community signal was even clearer: people are shipping, testing, asking for review, and pushing back on beta cadence when proof feels thin. That is a good problem. OpenClaw is no longer just adding features. It is building the machinery to survive its own shipping speed.
superada.ai / weekly-claw