Frontline Hotspot
Frontline Hotspot

Seven Weeks of Silence, Half of All PRs Ever Merged: OpenClaw 2.0 Is a Foundation Rewrite, So Back Up Before You Upgrade

OpenClaw 2.0 shipped on 2026-08-31 (tag v2026.8.1; GitHub API measures 388,423 stars, TypeScript, license field NOASSERTION but actually MIT). After a near-seven-week silence it merged 16,000+ PRs — about half the project's entire history — with 933 contributors: a foundation rewrite, not a point release. Two threads matter most: the browser becomes a first-class citizen and the Gateway consolidates credentials and permissions; credential governance gets private credential requests (masked prompts, values never enter chat or model context) and an opt-in proxy (protected-secret substitution only at approved destinations). Two breaking changes — the OpenProse plugin and /prose command removed, and codex/* plus openai-codex/* routes unified into openai/* (conflicts not auto-resolved); sessions move to SQLite with bounded rollback; plugin SDK deprecation deadline 2026-09-01. The README's four-way comparison table is already stale (2.0 only just introduced Shared Cloud Sessions and Active Memory), so do not read it as current reality.

Published September 1, 20268 min read
<!-- openclaw-2-0-release-hotspot | hotspot | Seven Weeks of Silence, Half of All PRs Ever Merged: OpenClaw 2.0 Is a Foundation Rewrite, So Back Up Before You Upgrade -->

OpenClaw's last stable release was almost seven weeks ago. Put that against its own cadence and it reads as an anomaly: in the preceding 230 days, the project shipped 106 releases, roughly one every two days. When a project that treats releases as breathing goes quiet for the better part of seven weeks, the community reaches for one of two explanations: something broke, or something big is coming.

At 03:30:51 UTC on August 31, 2026, the answer landed: tag v2026.8.1, release name "OpenClaw 2026.8.1", shipped simultaneously for Mac, Linux, and Windows. The official blog post is refreshingly honest about how it happened. It is titled OpenClaw 2.0, Accidentally. By their account, the original scope was just two things: simplify installation, and rebuild the browser experience as a first-class surface. Then the cleanup snowballed, the deeper it went, until the change was large enough that only a 2.0 would do.

The numbers back that up. This release involved 933 contributors, 569 of them first-time contributors, and merged more than 16,000 pull requests, roughly 50% of all PRs the project has ever merged. In other words, one release matched the accumulated merge volume of the preceding eight months. This is not feature stacking. This is a rewrite. Chinese-language coverage corroborates the timeline: The Paper (reporter Hu Hanyan) published at 22:46 on August 31, 2026, with follow-ups from Zhanzhang Baike and National Business Daily.

Our site already carries a full profile of the OpenClaw project itself, which covers what it is and how thick its ecosystem runs. This piece covers what happened on August 31, 2026, and what that event means for deployments you already have running.

Seven Weeks of Silence Was Not Stalling. It Was a Rebuild

Open-source projects go quiet in two ways. The first is abandonment: issues unanswered, PRs uncollected, nothing but bot-driven dependency bumps in the commit log. The second is a closed-door rewrite of the mainline, where commit frequency actually rises and the only thing missing is a stable tag. OpenClaw's silence was the second kind, and there is a hard metric to prove it: you do not accumulate 16,000 merged PRs during a period in which nobody was working.

The number worth sitting with is that 50%. When half of a repository's lifetime merged PRs land in a single release, you are not adding features, you are replacing a foundation. Features ship incrementally; foundations cannot. If you want to move session state from scattered files into SQLite, or move credential custody from "each client keeps its own copy" to "the Gateway holds it," you have to complete the path switch in one shot, because there is no shippable intermediate state. That also explains why the team accepted breaking a cadence built across 106 releases.

It is also the honest reason this is a 2.0 and not a 1.x. A major version bump is an explicit signal: seamlessness is not guaranteed. The next two sections spell out where exactly the cost lands.

Installation: Verify Before You Save

The part closest to ordinary users is the install flow. First-time setup now starts from what you already have on your machine: an existing ChatGPT or Claude subscription, an API key in hand, or a local model already running via Ollama or LM Studio. A large amount of configuration was trimmed or moved out of initial setup entirely, and what remains can be handled later by talking to Claw. Installation stops being "answer twenty questions before you can use it" and becomes "use it now, configure as you go."

One detail deserves isolated praise: the wizard issues a real completion request and only saves the route once that request passes. This kills the most common false positive in agent tooling. You paste an Ollama address, the UI says connected, and you only discover on the first real call that the service is down, the port is wrong, or the model name is misspelled. Having a local model installed is not the same as having a usable connection. Moving that verification ahead of the save step beats any post-hoc debugging.

What is more notable is the restraint shown around credentials. The team draws an explicit line between discovering credentials and transferring them: the wizard can reuse credential references that already exist, but it will not silently copy plaintext keys into another account, another machine, or a remote worker. When you import OpenClaw memory, credentials and skills are excluded. This looks less convenient, and it is the right call. The moment an agent tool starts auto-moving secrets, the blast radius stops being bounded by one machine. It is the same principle we keep returning to in our comparison of agent credential and permission models: the fewer places a credential is copied to, the better, and ideally zero.

The Browser Becomes a First-Class Citizen

The second original goal was the browser. It has been rebuilt as a first-class experience: you open it and you are in a conversation, able to keep configuring, return to work in progress, and follow along as things run. On the surface this is a UI change. Underneath it commits the project to a specific architectural judgment: there should be many entry points, and exactly one place where state lives.

That connects directly to the Gateway. The Gateway is the service that holds ordinary session state, model credentials, permissions, and persistent work; the browser, the CLI, paired devices, and remote workers are simply different ways of operating that Gateway. Once that split holds, capabilities follow naturally: you can pick up a laptop session on your phone, hand ongoing work to a remote worker, and watch a CLI-initiated job advance from the browser.

One caution: more entry points does not mean better privacy. Incognito mode behaves like this: transcripts stay in memory until restart, but it does not disable providers or tools. Do not mistake it for a privacy mode. Requests that were going to reach a model still reach a model.

Eight Release Highlights, Three That Matter

The release body lists eight. Reciting all eight is not useful; here is the overview, then the three worth your attention.

ChangePRsWhy it matters
Conversation history search#105057 / #105635 / #105585 (Thanks @hercial61)Search visible conversation text by exact word or phrase, reopen surrounding messages from a hit
Sessions beyond your Gateway#123280 / #127752 / #131744 / #132374Sessions run on paired devices or cloud workers, with the workspace migrating along
Progress cards survive refresh#125125 / #125438 / #125442 / #125444 / #121549 / #121840 / #121815 / #121813Follow subagent activity and cumulative edits across web and native chat
Structured question answering#109922 / #110372 / #130262 (Thanks @obviyus)Answer via cards, message buttons, or plain text, with a free-text fallback and an explicit Skip path
Interactive results and dashboards#101840 / #108983 / #110960 / #127315 (Thanks @100yenadmin, @hannesrudolph)Widgets inside chat, pin to session dashboards, grant specific actions or web sources, export rendered views as images
Private credential requests#129670 / #123216 / #132122Masked prompts for credentials; values never enter chat or model context
Approve recurring work once#129526 / #131602Grant an automation permission for one precise operation, reviewable and revocable later
Audio and video improvementsMedia stays attached across uploads, generated replies, playback, and refresh

Three of the eight carry the real weight, and they point in the same direction.

Private credential requests is the piece the agent ecosystem has been missing most. The agent asks you for a credential through a masked prompt; the value never enters chat and never enters model context. An opt-in proxy confines protected-secret substitution to destination addresses you have approved. The weight sits in "never enters model context." In most agent frameworks today, a key the user types ends up somewhere in the context, and then gets copied again by logs, tracing, and replay tooling. Keeping it out of context is a design-level improvement, not a cosmetic one.

Approve recurring work once resolves the oldest dilemma in automation: either you confirm every single time and hate it, or you grant blanket permission and get nervous. The approach here is to grant a precise operation once, keep it reviewable and revocable, and require re-approval whenever the job or the operation changes. Granularity sits on the operation rather than the session, which is the correct place for it.

Sessions beyond your Gateway is the clearest expression of the positioning shift. A session can run on a paired device or a cloud worker, the session workspace migrates with it, and warm machines and project seeds get reused. The mechanism for shared cloud sessions is specific: the Gateway supplies the conversation and the reconciled workspace to the remote worker while proxying model requests so provider credentials stay on the Gateway side; if the worker disappears, a new one can be started from persisted state. That description deserves a second read. Remote execution is no longer "hand your keys to someone else and let them run it" but "someone else runs it, the keys stay with you." Anyone working on multi-agent cluster orchestration will immediately recognize what this unblocks. Our piece on OpenHuman's memory and agent cluster orchestration tackles the same class of problem from a different angle and is worth reading side by side.

Two Breaking Changes and a September 1 Deadline

Benefits covered, now the cost. This release carries two breaking changes, and neither is the "you might see an error after upgrading" variety.

First, OpenProse migration (breaking). The bundled OpenProse plugin and the /prose command are gone. You need to run openclaw doctor --fix to clean up stale configuration and follow the upstream Agent Skill migration path. The good news is that existing .prose source files are preserved (#128494). The files survive, but the default entry point does not, so any script or automation depending on /prose breaks outright.

Second, OpenAI route migration (breaking). Running openclaw doctor --fix moves codex/* and openai-codex/* model references, provider config, stored sessions, and automation routes over to openai/*, while preserving Codex runtime intent. Note the two bolded items: this is not a config-file edit. Model references recorded inside your historical sessions and the routes on your automation jobs get rewritten too. Where conflicts arise, an operator has to resolve them manually, which means this path is not fully automatic.

Third, an imminent deprecation dated 2026-09-01. External plugins need to migrate ahead of the SDK removal gate, moving configuration from plugin-sdk-config-runtime-subpath to api.pluginConfig. That deadline falls on release day, so anyone maintaining an external plugin gets no buffer.

Here is the official upgrade guidance verbatim:

"If the automatic update fails, use a local coding harness to help complete the update, diagnose any migration errors, and verify that the Gateway starts correctly. Back up your configuration and state before making changes."

That the project says this in its own release notes signals how wide the failure surface is. For step-by-step procedure, rollback criteria, and the common errors, we put together a separate OpenClaw 2.0 upgrade and migration SOP. Read it before you upgrade.

Sessions Move to SQLite: Downgrades Are Bounded

If this article leaves you with one sentence, let it be this one: the upgrade validates, archives, and migrates file-backed session records and transcripts into SQLite, and downgrades are bounded.

"Bounded" means something precise. An older, file-backed version can restore archived records that existed before the migration, but it cannot see sessions created after the move to SQLite. So you upgrade to 2.0, use it for three days, decide it is not for you, and roll back to 1.x: those three days of new sessions are unreadable by the old version. They are not necessarily lost, but the old version in your hands cannot reach them.

Rolling back the entire state directory is worse. That blunt approach has a wider blast radius than sessions alone: approvals records revert, delivery and dedup records revert, and some ratcheting channel credentials have to be re-associated. Rollback is not free. It takes your granted permissions and your already-delivered-message dedup state with it, and duplicate delivery is the kind of thing that causes real damage in automated workflows.

That leaves exactly one practical recommendation, supported by both the vendor and common sense: back up first, then upgrade. And the thing you back up is not just the config file, it is the entire state directory.

Governance, Ecosystem, and the Numbers to Discount

A few facts from the periphery that matter, with sourcing labeled.

The first-party repository numbers come from a direct GitHub API check on September 1, 2026: openclaw/openclaw shows 388,423 stars and 81,539 forks, written in TypeScript, created November 24, 2025, still receiving pushes today, not archived. Against the 384,631 stars we recorded on this site on July 31, that is roughly 3,800 more in a month.

Licensing needs its own paragraph. GitHub's SPDX detector returns NOASSERTION for this repository, which looks alarming but is a false positive: the LICENSE file governs, and the license is MIT. Teams planning commercial integration should not be misled by an API field; read the LICENSE text.

On governance: founder Peter Steinberger joined OpenAI this year, but the OpenClaw Foundation remains independent. His words: "OpenAI hired me, not OpenClaw," and "The OpenClaw Foundation is independent, with sponsors rather than owners." Together those two sentences are about as clear a boundary statement as an open-source project can make while sitting inside a large company's gravitational field.

Two external signals on the ecosystem side: Nvidia's Jensen Huang said at GTC that "every company in the world today needs to have an OpenClaw strategy," and at Microsoft Build, Pavan Davuluri confirmed that OpenClaw runs natively inside Microsoft Execution Containers on Windows. The first is sentiment; the second is genuine platform-level adoption.

And now the part that has to be discounted: the claims of 4.5 million new users per week and "fastest-growing repository in GitHub history" trace only to SaaS Sentinel's reporting. We found no first-party source, so treat them as second-hand community-temperature indicators, not audited metrics. Growth numbers are the easiest thing in open source to amplify through a chain of reposts, and every extra link in that chain widens the error bars. Stars, forks, and contributor counts are checkable directly against the API, and those are the numbers to trust.

Should You Upgrade, and How

Our judgment is direct: yes, upgrade, but not this afternoon.

The case for upgrading is that the direction is right. Pulling session state, credentials, and permissions into the Gateway buys collaboration, handoff, and governability: sessions can be picked up across devices, remote workers can continue work without credentials leaving the Gateway, and credential requests can stay out of model context. That is a threshold any tool must cross on the way from "toy for individual enthusiasts" to "production infrastructure for teams," and OpenClaw crossed it seriously.

The case against rushing is equally clear. Two breaking changes will genuinely sever /prose dependencies and codex/* routes, the SDK config deadline for external plugins lands on 2026-09-01, and the downgrade path is bounded, meaning sessions created after the SQLite move are invisible to older versions. Upgrading production under that combination would be irresponsible.

The cadence we would recommend has three steps. First, take a full backup of the state directory and confirm it restores. Second, upgrade one non-critical machine, run the skills and automations you depend on daily, and watch specifically for failures around /prose and external plugins. Third, only then push to your primary machines. When migration conflicts appear, do not guess; follow the official advice and let a local coding harness help with diagnosis. The complete procedure and rollback criteria live in the OpenClaw 2.0 upgrade and migration SOP, and if you care more about how permission models are designed than about this specific release, the comparison of agent credential and permission models goes deeper on the trade-offs.

One closing assessment. Sixteen thousand pull requests, 933 contributors, 569 of them first-time: those numbers do not say a project is popular, they say a project got something right about making strangers willing to submit their first line of code. As for the "biggest update ever" framing, readers can judge that themselves. What is worth remembering is the plain advice: back up before you upgrade, because downgrades are bounded.


Sources

  • Official blog, "OpenClaw 2.0, Accidentally": https://openclaw.ai/blog/openclaw-2-accidentally/
  • GitHub release v2026.8.1 (2026-08-31T03:30:51Z, Mac / Linux / Windows)
  • GitHub API check (2026-09-01): 388,423 stars / 81,539 forks / TypeScript / created 2025-11-24 / not archived
  • The Paper (reporter Hu Hanyan), published 2026-08-31 22:46; Zhanzhang Baike; National Business Daily
  • Architecture and migration details: WinBuzzer's English coverage of the Gateway, the SQLite migration, and downgrade boundaries
  • Growth figures (4.5M new users weekly, fastest-growing GitHub repository): relayed by SaaS Sentinel, second-hand, no first-party report found

This article is AI-assisted and human-edited. Last updated: 2026-09-01

Related

Frontline Hotspot

ChatGPT Images 2.5: Half the Latency, Real Consistency

OpenAI launched ChatGPT Images 2.5 on 2026-09-09: up to 50% lower latency than 2.0, better preservation of reference-photo subjects and multi-turn edit consistency; ChatGPT adds sketch mode, templates, image comments and prompt sharing; the API ships two models, Flare and Sunburst. This piece breaks down each upgrade, argues the real leap is latency plus consistency rather than raw image quality, reads the two-model split as capability tiering and pricing segmentation (analysis, not official wording), and weighs the long-term lock-in cost of closed APIs.

Sep 9, 20269 min read
Frontline Hotspot

Nvidia's $13B Hugging Face Deal: What It Means for Open Source

Reported 2026-09-04 (Cailianspress and others): NVIDIA announced the acquisition of Hugging Face for about \$13B — \$11.9B to investors and \$1B for employee equity retention — one of the largest deals in NVIDIA's history. Jensen Huang committed to keeping HF an open platform without forcing NVIDIA compute. This piece breaks down the deal structure, why a compute hegemon would buy the open-source ecosystem's front door, how much developers should trust the promise ("not forced" is not the same as "not default"), and the hosting-platform implications.

Sep 8, 20269 min read
Frontline Hotspot

Gemini 3.8 Drops: Flash and the Security-First Flash Cyber

Google released Gemini 3.8 on 2026-09-02 (US) / 09-03 (Beijing) as two models: the general Flash for long-horizon engineering and agents, and the security-focused Flash Cyber for autonomous vulnerability discovery and automated patching, available only to defenders via the Fairwind Program. Official numbers: HLE-Verified 54.9%, CWE-Bench pass@1 47.2%, cross-language vuln discovery >70%, 2.6x Chrome patches, critical vulns found in <2 hours; intro pricing \$0.75/\$3.75 per million tokens.

Sep 5, 20269 min read