TL;DR

In the 10 days from 8/15-8/25/2026, OpenClaw released v2026.8.1-beta.2 + v2026.8.1-beta.3 + Codex 0.149 pin + 8/22-23 main branch four-burst combo, transitioning from feature stacking to platform convergence new phase: ① 8/15 05:36 UTC v2026.8.1-beta.2 introduced secret egress host binding (CLI/Gateway RPC/Control UI three-layer fail-closed) + GPT-5.6 Ultra/Sol/Terra/Luna atomic runtime switching + macOS app profiles multi-tenant isolation + SQLite snapshot 4 commands (create/list/verify/restore) + plugin provenance --force warning + Shared plugin SDK monitor (IRC/Synology/Google Chat three-channel unified lifecycle); ② 8/21 Codex 0.149.0 runtime pin + Doctor migrated untrusted approval policy + Same-channel async delivery persistence + Restricted turns preserved AGENTS.md + 3 release-branch backports; ③ 8/22 Control UI Claude 5-series 200K/1M context window switch + 8/23 macOS Realtime Talk Relay + Telegram two-owner retry fix + /v1/responses blank fix; ④ 8/24 04:40 UTC v2026.8.1-beta.3 integrated GPT-5.6 Sol/Terra/Luna/Ultra across OpenClaw + Codex runtimes + Control UI first-run setup continuation (Custodian + channel) + Puppeteer-compatible CDP relay + External Gateway lifecycle supervision verified restart + Compact SQLite backup/restore (verified) + 89 npm plugins sync read back + @openai/codex@0.149.1 managed runtime; ⑤ 8/20 PR #126587 Read Tool Handles Huge One-Line Files (bounded cursors without shell fallback, 81936-byte JSON three-read rebuild + 98316-byte emoji four-read rebuild); ⑥ 8/24-25 Node Hosting removed hardcoded 2-session cap + CPU-based default 1 worker session per available CPU core (clamped 1024) + nodeHost.workerRuns.capacity explicit override. OpenClaw enters platform convergence four-dimensional new phase: ① Security Boundary Convergence (fail-closed + provenance + profiles + workspace preservation); ② Runtime Convergence (GPT-5.6 all 4 tiers + Codex 0.149 managed + atomic switching); ③ Infrastructure Convergence (SQLite backup/restore + Node Hosting CPU capacity + Codex approval Doctor migration); ④ Operations Experience Convergence (Control UI first-run continuation + Reload wait Gateway + macOS Talk Relay + Browser extension CDP).

1. Today’s Headline: 8/15-8/25 10 Days, OpenClaw 6-Burst Combo for Platform Convergence

Date (UTC+8)Version/EventKey CapabilitySource
8/15 13:36v2026.8.1-beta.2secret egress host binding (CLI/Gateway RPC/Control UI fail-closed) + GPT-5.6 Ultra/Sol/Terra/Luna atomic runtime switching + macOS app profiles multi-tenant isolation + SQLite snapshot 4 commands + plugin provenance --force warning + shared plugin SDK monitor (IRC/Synology/Google Chat unified lifecycle)OpenClaw Chronicles 8/24 + terminalblog.com 8/16 + agentdaily.dev 8/15
8/20 16:00PR #126587 Read Toolbounded cursors without shell fallback + skill instructions stay whole-document-only + 81,936-byte minified JSON three-bounded-reads rebuild + 98,316-byte emoji four-bounded-reads rebuildopenclawchronicles.com 8/20
8/21 16:00PR #127026 Codex 0.149 pin@openai/codex 0.149.0 managed runtime + Doctor migrated untrusted → on-request + Same-channel async delivery persistence + 188/188 focused Codex contract tests + 781 generated protocol files exact-tag verification + 3 release-branch backportsopenclaw.academy 8/21
8/22 16:00PR #126891 + Claude 200K/1M switchRestricted turns preserved AGENTS.md (ring-zero still suppressed) + Claude 5-series 200K/1M context window Control UI switch + plugin-declared selectable context windows + argv/env mappingopenclaw.academy 8/22 + openclaw.ai
8/23 16:00main branch 3 fixesmacOS Realtime Talk Relay main added + /v1/responses blank fix + Telegram two-owner retry fix (duplicate sends merged into durable outbound queue single owner)openclaw.academy 8/23 + OpenClaw main PRs
8/24 04:40v2026.8.1-beta.3GPT-5.6 Sol/Terra/Luna/Ultra integrated across OpenClaw + Codex runtimes + Control UI first-run setup continuation (Custodian + channel) + Puppeteer-compatible CDP relay + External Gateway lifecycle supervision verified restart + Compact SQLite backup/restore (verified) + Shared durable ingress monitors + 89 npm plugins sync read back + @openai/codex@0.149.1 managed runtimeOpenClaw Chronicles 8/24 + releases.sh 8/24 + GitHub Releases 8/24
8/24-25Node Hosting capacity controlremoved hardcoded 2-session cap + CPU-based default 1 worker session per available CPU core (clamped 1024) + nodeHost.workerRuns.capacity explicit overrideopenclaw.academy 8/24

Core judgment: completely different from historical P2 OpenClaw 6 article angles (8/16 ClawPro Enterprise ADP + ClawPro product overview + architecture business application analysis + WeCom QR scan one-click access / 8/8 v2026.5.28 stable / 7/20 v2026.3.23 milestone / 7/12 RCE CVE / 7/3 mobile / 6/23 manufacturing landing / 6/5 130K Star) — upgrading from single-version feature stacking + enterprise commercialization advancement to multi-version platform convergence + security/runtime/infrastructure/operations four-dimensional convergence dual-axis resonance. The 8/15-8/25 10-day 6-burst combo all center on convergence (fail-closed converging secret paths + atomic converging model/runtime switching + verified converging SQLite backup + CPU converging Node Hosting capacity) — OpenClaw officially transitions from feature stacking period to platform convergence period.

2. Dimension 1: Security Boundary Convergence — Fail-Closed + Provenance + Profiles + Workspace Preservation

2.1 Secret Egress Host Binding (v2026.8.1-beta.2 Headline)

OpenClaw 2026.8.1-beta.2 released at 8/15 13:36 UTC+8, the headline security feature is secret egress host binding:

Mechanism: every shared-store secret is now bound to exact HTTPS destination hosts, covering CLI, Gateway RPC, and Control UI three layers. If the secret is not bound to the attempted destination, sentinel substitution fails closed — plaintext credentials are blocked before leaving the machine.

This design has three strategic implications:

  • fail-closed default deny: traditional approach logs/warns after-the-fact, fail-closed blocks at the egress point directly — administrators have a sharper boundary, not just post-hoc auditing
  • cross-layer consistency: CLI/Gateway RPC/Control UI three layers enforce uniformly, avoiding protection gaps where “one layer fixes another layer bypasses”
  • binding granularity exact: host-level binding (not domain wildcard, not IP range), attack surface converges from “arbitrary endpoint” to “explicitly authorized endpoint”

8/16 terminalblog.com comment: “Before this update, a compromised plugin or misconfigured tool could exfiltrate your API keys to arbitrary endpoints. Now? The secret simply won’t resolve unless the destination matches what you authorized. It’s a default-deny architecture for credential egress.”

Note: post-release assessment from MIIT and third-party security institutions (Xiaohonghua Digest xhh.club): under default configuration credentials still flow as plaintext — users must manually configure destination hosts to enable protection, otherwise credentials may still be leaked. This is OpenClaw’s consistent “default insecure” design philosophy: security features must be actively enabled, cannot rely on defaults.

2.2 Plugin Provenance + macOS App Profiles + Workspace Preservation

Plugin provenance warning (v2026.8.1-beta.2 simultaneous): arbitrary executable plugin sources now require explicit --force acknowledgment; trusted ClawHub, bundled, official-catalog, tracked-update flows remain low-friction; Crestodian installs restricted to trusted sources. This stops the “install random plugin from GitHub” supply chain attack vector cold.

macOS app profiles multi-tenant isolation (v2026.8.1-beta.2 simultaneous): when running multiple OpenClaw instances on the same Mac, each instance now gets complete isolation: independent state directories, independent preferences, independent Keychain entries, independent Gateway services, duplicate-instance ownership detection. Host-global login + node services stay shared — this is the first coding agent platform to ship true macOS multi-tenancy without container overhead.

Restricted turns preserve AGENTS.md (PR #126891 8/22): previously under restricted tool policy, ordinary Codex-harness turns silently lost workspace AGENTS.md — restriction correctly removed native tool + MCP access, but also disabled project-document discovery, and OpenClaw’s own instruction carrier excluded the same file — the model got workspace contract from neither side. Now ordinary restricted turns receive bounded AGENTS.md snapshot through frozen thread-level developer-instruction path, preserving all capability restrictions’ fail-closed. Ring-zero (explicitly tool-disabled) turns still suppress project documents — explicit precedence. Operator action: if you restrict Codex tool policies for safety, re-verify agent behavior after upgrade — instructions that were silently missing will now appear, which may change outputs (this is the fix working, not a restriction leak).

3. Dimension 2: Runtime Convergence — GPT-5.6 All 4 Tiers + Codex 0.149 Managed + Atomic Switching

3.1 GPT-5.6 Ultra/Sol/Terra/Luna Atomic Switching

v2026.8.1-beta.2 (8/15) introduced GPT-5.6 Ultra support + runtime switching across OpenClaw + Codex engines; named tiers include Sol/Terra/Luna, with model/runtime/thinking selection kept atomic through /model + fallback paths — no more half-switched states where model changes but runtime doesn’t. Fresh setup defaults simultaneously move forward: new API-key setup uses openai/gpt-5.6, new Codex/OAuth setup uses openai/gpt-5.6-sol. Existing primary models, fallback selections, aliases, explicit GPT-5.5 choices preserved.

v2026.8.1-beta.3 (8/24) upgraded to GPT-5.6 Sol/Terra/Luna/Ultra full support across OpenClaw + Codex runtimes + @openclaw/codex@2026.8.1-beta.3 simultaneously ships managed @openai/codex@0.149.1 runtime.

Critical pricing alert (Xiaohonghua Digest xhh.club assessment): GPT-5.6 Sol/Terra/Luna three-tier pricing varies dramatically — Luna output price is 5+ times Sol. In ordinary conversation, Luna costs ~5x more than Sol. More importantly: OpenClaw defaults GPT-5.6 to fresh system default model, new users may unintentionally use the most expensive model, causing cost spikes. Recommendation: actively choose runtime matching task difficulty; consider ChatGPT Plus subscription and other low-cost options; monitor token usage.

3.2 Codex 0.149.0 Runtime Pin + Doctor Migration (PR #127026 8/21)

8/21 main branch merged a wave of Codex-related changes, core is PR #127026:

  • managed @openai/codex runtime floor pinned to 0.149.0 on frozen beta candidate
  • retired untrusted approval policy now migrated by Doctor alone to on-request — operator action: before upgrading Codex runtime fleet, run Doctor once per deployment, not manual config editing; if you have scripts grep/set untrusted, retire them — that string is no longer valid input anywhere else
  • Same-channel async Codex delivery persistence: previously stable deliveryIntentId blocks only entered durable route path when reply crossed channels — same-channel Codex async updates fell through to in-memory dispatcher, app-server callback settled before any durable custody existed. Now core dispatch loads route runtime for stable intents even when channel is unchanged, fails closed when no durable route exists, preserves source-account-specific reply policy (non-default account cannot fall back to channel default). Rejections returned to Codex for retry under same stable intent, pending/completed intents stay deduplicated across restarts/replay
  • 3 release-branch backports land same day: PR #127269 stops pinned branch rejecting every Codex app-server newer than bundled version; PR #127337 backports native discovery of newly available first-party models (#127322) so existing authenticated Codex account can use new models without waiting for OpenClaw catalog update (only when Codex explicitly selected and no authored endpoint/transport override/forwarded credential conflict)

Upstream validation: 188/188 focused Codex contract tests, restart/replay dedupe coverage, 781 generated protocol files matched openai/codex rust-v0.149.0 exact-tag, signed commits binding candidate diff + approval bindings.

3.3 Same-Channel Async Delivery Persistence — Solving Codex Duplication/Loss

8/21 PR #127026 fixed an important dispatch ownership gap: previously same-channel async Codex updates fell through to in-memory dispatcher, app-server callback settled before any durable custody existed — causing mid-turn updates to duplicate or vanish after restart.

After fix, core dispatch loads route runtime for stable intents even when channel is unchanged, fails closed when no durable route exists, and preserves source-account-specific reply policy — non-default account cannot fall back to channel default. Operator observation point: Codex agents posting progress into working channel deployments, will see fewer duplicated or vanished mid-turn updates after restart post-upgrade.

4. Dimension 3: Infrastructure Convergence — SQLite Backup/Restore + Node Hosting CPU Capacity + Codex Approval Doctor Migration

4.1 SQLite Snapshot 4 Commands (v2026.8.1-beta.2 + beta.3 Simultaneous Upgrade)

v2026.8.1-beta.2 introduced SQLite snapshot command suite:

openclaw backup sqlite create    # Create snapshot
openclaw backup sqlite list      # List snapshots
openclaw backup sqlite verify    # Verify snapshot
openclaw backup sqlite restore   # Restore snapshot (limited to fresh targets)

This is the key operational infrastructure supplement after OpenClaw installations increasingly rely on durable local state (agents, channels, crons, task history, session records). fresh-target-only restore rule is reasonable guardrail: backups are most useful when recovery is predictable and doesn’t silently overwrite live state directory.

v2026.8.1-beta.3 (8/24) further upgraded to Compact, verified SQLite backup and fresh-target restore commands — verified step lets operator confirm backup integrity before restore.

4.2 Node Hosting Capacity Control (PR 8/24-25)

8/24-25 main branch merged Node Hosting capacity control PR, removing hardcoded two-session cap:

  • Before: each paired node hardcoded 2 worker sessions cap — distributed session load limited by this
  • Now: capacity default 1 worker session per available CPU core (clamped to 1024)
  • Override: nodeHost.workerRuns.capacity explicit config override

Significant impact on distributed session load distribution — previously multi-node deployment session load frequently limited by 2-session cap, now scales elastically by CPU + explicit override.

4.3 Codex Approval Doctor Migration (PR #127026 Simultaneous)

In 8/21 PR #127026, retired untrusted approval policy migration handled by Doctor alone — this converges scattered migration logic to single authoritative tool (Doctor), key infrastructure convergence.

5. Dimension 4: Operations Experience Convergence — Control UI First-Run Continuation + Reload Wait + macOS Talk Relay + Browser CDP

5.1 Control UI First-Run Setup Continuation (v2026.8.1-beta.3)

v2026.8.1-beta.3 (8/24) introduced Control UI first-run setup continuation: now first-run setup after verified model setup continues to Custodian and optional channel setup — reducing first-time user multi-step interruption, consistent across macOS/Linux/Windows.

5.2 Control UI Reload Button Finally Waits Gateway (v2026.8.1-beta.2)

v2026.8.1-beta.2 (8/15) fixed an annoying Control UI bug: “new version available” reload button previously silently failed — operator had to manually hard reload to recover. Now reload button waits for Gateway restart completion then auto-reloads.

5.3 macOS Realtime Talk Relay Main Added (PR 8/23)

8/23 main branch merged macOS-only realtime Gateway relay for Talk — surfaced in Control UI settings. This is the key step for OpenClaw to extend from local interactive assistant to runnable as infrastructure — resonating with v2026.8.1-beta.3’s External Gateway lifecycle supervision.

5.4 Browser Extension CDP Compat (Puppeteer Without Prompt)

v2026.8.1-beta.2 (8/15) simultaneously launched Browser extension CDP compat: browser extension relay now responds to Target.getBrowserContexts + serves DevTools-style /json/list. Puppeteer-based clients (chrome-devtools-mcp) can now drive paired Chrome without remote-debugging permission prompt.

5.5 Read Tool Handles Huge One-Line Files (PR #126587 8/20)

8/20 main branch merged PR #126587 fix(agents): read oversized single-line files without exec:

  • Before: agent reading single oversized line may receive empty response — tool response recommended Bash pipeline as continuation, but exec-disabled environment is dead end
  • Now: read owner returns bounded slice of the oversized line + flat intra-line cursor. Line paging with offset/limit unchanged (ordinary multi-line files behavior unchanged). Large single-line files: UTF-8 + UTF-16 safe slice, stay within model-visible output budget, continuation guidance doesn’t require shell access

Critical boundary: skill/playbook-style instruction documents stay whole-document-only — window parameters rejected, fitting instructions delivered in full, oversized instructions produce bounded refusal (not partial first window). This is the right trade-off: partial source file is useful when telling how to continue, partial instruction document is more dangerous — looks like complete guidance but omits critical constraints.

Actual measurement: 81,936-byte minified JSON file complete rebuild in 3 bounded reads, 98,316-byte emoji line complete rebuild in 4 bounded reads — read capability stays self-contained under exec + process absent environment.

6. Dimension 5: Channel Convergence + Model Ecosystem — Shared Plugin SDK + 89 npm Plugins + Claude 200K/1M

6.1 Shared Plugin SDK Monitor (v2026.8.1-beta.2)

v2026.8.1-beta.2 (8/15) introduced Shared plugin SDK monitor — durable admission, polling, pruning, claim identity validation, adoption handoff, shutdown all shared across channel plugins:

  • IRC, Synology Chat, Google Chat migrated to shared lifecycle — reducing bespoke channel behavior
  • v2026.8.1-beta.3 (8/24) upgraded to Shared durable ingress monitors for channel plugins

6.2 89 npm Plugins Sync Read Back (v2026.8.1-beta.3)

v2026.8.1-beta.3 (8/24) release evidence explicitly: all 89 official npm plugins were read back at 2026.8.1-beta.3, with beta selector + complete tarball integrity metadata.

Additional provenance:

  • npm: https://www.npmjs.com/package/openclaw/v/2026.8.1-beta.3
  • Registry tarball: https://registry.npmjs.org/openclaw/-/openclaw-2026.8.1-beta.3.tgz
  • Integrity: sha512-8v+2Knr+0i1qzWXgJmtcBg78VaoMENahLxcuThOqyCmVaCGPj++mI9yv0R440wMv9Sysd5e0YmBVftDvuQ==
  • Core npm preflight: GitHub Actions runs/32687267214
  • Core npm publication: GitHub Actions runs/32689503061
  • Official plugin publication: GitHub Actions runs/32680190375

Significance: 89 plugins sync read back gives operator concrete verification point before testing beta — plugin version skew is historically hard to debug, especially when channel plugins and runtime packages move together.

6.3 Claude 5-Series 200K/1M Context Window Switch (PR 8/22)

8/22 main branch merged plugin-declared selectable context windows + 200K/1M switch in Control UI model picker for Claude 5-series models — session-level validation + explicit argv/env mapping. Let Claude Code 5-series models select context window per scenario (long documents 1M vs ordinary conversation 200K).

6.4 Telegram Two-Owner Retry Bug Fix (PR 8/23)

8/23 main branch merged OpenClaw Closes a Two-Owner Retry Bug That Duplicated Telegram Sends: previously when message tool failed with proven-not-sent evidence, durable outbound queue and model both held same retry — recipients received two copies seconds apart. After fix queue becomes single retry owner, including remote Gateway deployments.

6.5 /v1/responses Blank Fix (PR 8/23)

8/23 main branch fixed a /v1/responses API output blank or truncated defect — chat-completions stayed correct. This is OpenClaw Responses API users’ silent failure risk elimination.

7. OpenClaw 2026.8.1 Series 5-Step Enterprise Landing Path

7.1 Phase 1: Evaluate Beta Applicability (8/24-8/31, 1 Week)

  • Scenario A: Local interactive users: can defer upgrade, beta mainly platform-level improvements
  • Scenario B: Multi-tenant macOS deployments: strongly recommend upgrade to v2026.8.1-beta.2+ — macOS app profiles is structural change, requires validation pass
  • Scenario C: Codex integration: strongly recommend upgrade to v2026.8.1-beta.3 — Codex 0.149.1 managed runtime pin + same-channel async delivery persistence are critical fixes for production
  • Scenario D: Security-sensitive deployments: v2026.8.1-beta.2 secret egress host binding + plugin provenance is structural security upgrade — recommend staging validation

7.2 Phase 2: Manually Enable Secret Egress Host Binding (Default Disabled)

Critical warning (Xiaohonghua Digest xhh.club assessment): under default configuration credentials still flow as plaintext — secret egress host binding will not auto-enable, operator must manually configure destination hosts per shared-store secret to enable protection.

Operation steps:

  1. List all current shared-store secrets: openclaw secrets list
  2. Bind each secret to destination host: openclaw secrets bind <secret-id> --host <exact-host>
  3. Confirm fail-closed test passes: openclaw secrets test --unbound-destination

7.3 Phase 3: Doctor Migration Before Codex 0.149 Upgrade

Operation steps:

  1. Before upgrade run Doctor once per deployment: openclaw doctor --migrate-approval-policy
  2. Check if scripts grep/set untrusted — if so, retire (untrusted retired)
  3. Upgrade to v2026.8.1-beta.3 or later
  4. Re-verify restricted Codex tool policies agent behavior — previously silently lost AGENTS.md now appears

7.4 Phase 4: SQLite Backup Establish Routine

Operation steps:

  1. Create backup strategy: openclaw backup sqlite create --schedule daily
  2. Verify backup integrity: openclaw backup sqlite verify --all
  3. Don’t restore to live state directory — restore only fresh targets
  4. Monitor backup size: single instance < 100MB (per-agent + global)

7.5 Phase 5: Node Hosting Capacity Tuning

Operation steps:

  1. Check current CPU cores: nproc
  2. Default value takes effect (1 worker session per CPU core, clamped 1024)
  3. If override needed: openclaw config set nodeHost.workerRuns.capacity <N> explicit setting
  4. Monitor session distribution: openclaw node list --capacity

8. Enterprise Defense Checklist (6 Defense Lines)

  1. [CRITICAL] Manually enable secret egress host binding — default disabled, must bind exact destination host per shared-store secret
  2. [CRITICAL] Doctor migration before Codex upgrade — run openclaw doctor --migrate-approval-policy per deployment, retire any grep/set untrusted scripts
  3. [HIGH] Re-verify restricted turns agent behavior after Codex 0.149 upgrade — previously silently lost AGENTS.md now appears, may change output
  4. [HIGH] GPT-5.6 default model cost monitoring — Luna output price is 5x+ Sol, fresh system default GPT-5.6 may burn money
  5. [MEDIUM] Plugin provenance --force warning — only trusted ClawHub/bundled/official-catalog/tracked-update flows stay low-friction; arbitrary executable sources must explicitly --force confirm
  6. [MEDIUM] SQLite backup routine — establish daily openclaw backup sqlite create + periodic verify, avoid silent state corruption

FAQ (High-Frequency Questions)

Q1: Is OpenClaw 2026.8.1 beta stable or beta? Can it be used in production?

A: Currently beta channel. v2026.8.1-beta.2 (8/15) + v2026.8.1-beta.3 (8/24) are both prereleases. Production operators should test in staging first — 89 npm plugins sync read back provides verification point but cannot replace staging validation. OpenClaw’s previous stable channels were v2026.7.1/v2026.7.2-beta.7 etc; beta → stable timeline not announced.

Q2: Is secret egress host binding enabled by default?

A: No. OpenClaw consistently follows “default insecure” design philosophy — secret egress host binding will not auto-enable, must manually configure destination host per shared-store secret to enable protection. MIIT and third-party security institutions assess: under default configuration credentials still flow as plaintext — this is OpenClaw’s consistent design principle, security features must be actively enabled.

Q3: How much do GPT-5.6 four tiers (Ultra/Sol/Terra/Luna) differ in price?

A: Significant differences. Xiaohonghua Digest xhh.club assessment: Luna output price is 5x+ Sol. In ordinary conversation, Luna costs ~5x more than Sol. Key trap: OpenClaw defaults GPT-5.6 to fresh system default model, new users may unintentionally use the most expensive Luna, causing cost spikes. Recommendation: actively choose runtime matching task difficulty; consider ChatGPT Plus subscription and other low-cost options; monitor token usage.

Q4: Will Codex 0.149 upgrade break existing configuration?

A: Yes. untrusted approval policy retired — if you have scripts grep/set untrusted, retire them, that string is no longer valid input. Operation steps: before upgrade run openclaw doctor --migrate-approval-policy per deployment, let Doctor auto-migrate untrusted → on-request. Also: restricted turns’ AGENTS.md now appears — previously silently lost workspace instructions now accessible, may change agent output (this is the fix working, not a restriction leak).

Q5: What are macOS app profiles? Why important?

A: macOS app profiles is the multi-tenant isolation mechanism introduced by v2026.8.1-beta.2. When running multiple OpenClaw instances on the same Mac, each instance now gets complete isolation: independent state, preferences, Keychain, Gateway services, duplicate-instance ownership detection. Host-global login + node services stay shared. This is the first coding agent platform to ship true macOS multi-tenancy without container overhead — for developers needing test multiple configurations and enterprises needing staging/production isolation, this is a critical capability.

Q6: What is the difference between OpenClaw 2026.8.1-beta.3 SQLite backup and v2026.8.1-beta.2?

A: beta.3 upgraded to verified. v2026.8.1-beta.2 (8/15) introduced SQLite snapshot 4 commands (create/list/verify/restore); v2026.8.1-beta.3 (8/24) upgraded to Compact, verified SQLite backup and fresh-target restore commands — verified step lets operator confirm backup integrity before restore. fresh-target-only restore rule stays consistent (don’t restore to live state directory).

Q7: What problem does Puppeteer-compatible CDP relay solve?

A: Solves remote-debugging permission prompt. Previously paired Chrome + browser extension Puppeteer-based clients (e.g. chrome-devtools-mcp) required remote-debugging permission prompt. v2026.8.1-beta.2 + v2026.8.1-beta.3 introduced Puppeteer-compatible CDP relay — browser extension relay responds to Target.getBrowserContexts + serves DevTools-style /json/list, letting Puppeteer-based clients drive paired Chrome without prompt. Key improvement for external CDP clients + automation framework integration.

Q8: What does Node Hosting CPU capacity adjustment affect?

A: Affects distributed session load distribution. Previously each paired node hardcoded 2 worker sessions cap. Post v2026.8.1-beta.3 PR removed hardcoded cap, capacity default 1 worker session per available CPU core (clamped 1024) + nodeHost.workerRuns.capacity explicit override. Significant impact on distributed session load — previously multi-node deployment session load frequently limited by 2-session cap, now scales elastically by CPU.

Key Terminology

TermOne-Sentence Explanation
Secret egress host bindingOpenClaw binds shared-store secrets to exact HTTPS destination hosts, unbound sentinel substitution fails closed, plaintext credentials blocked before leaving machine
Fail-closedDefault deny architecture — unauthorized operations blocked directly (vs fail-open default allow + log/warn after-the-fact)
GPT-5.6 Ultra/Sol/Terra/LunaGPT-5.6 four inference tiers — Ultra/Sol/Terra/Luna price decreasing (Ultra most expensive) but quality decreasing, covering different task difficulties
Codex 0.149.0managed @openai/codex runtime floor version — OpenClaw v2026.8.1-beta.3 + subsequent pinned releases lock this version
Doctor migrationOpenClaw single authoritative migration tool — responsible for migrating retired approval policy (e.g. untrusted) to canonical state (on-request)
Same-channel async deliveryCodex async delivery posting progress in working channel — persisted from v2026.8.1-beta.3 to avoid mid-turn updates duplicate/vanish after restart
macOS app profilesOpenClaw multi-tenant isolation mechanism — multiple instances on same Mac each get independent state/preferences/Keychain/Gateway services
SQLite snapshotOpenClaw durable local state backup suite — 4 commands (create/list/verify/restore) + fresh-target-only restore rule
Puppeteer-compatible CDP relaybrowser extension relay compatible with Puppeteer — supports Target.getBrowserContexts + DevTools-style /json/list, external CDP clients integrate without prompt
Node Hosting CPU capacityOpenClaw paired node worker session capacity control — post v2026.8.1-beta.3 default 1 worker session per CPU core (clamped 1024)
Restricted turns + AGENTS.mdCodex restricted tool policy workspace instruction preservation — previously silently lost, now preserved through frozen thread-level developer-instruction path
Pinned release branchOpenClaw release branch — maintains stability without frequent updates; post v2026.8.1-beta.3 backport syncs new model discovery and other key improvements

References

  1. OpenClaw Chronicles — OpenClaw 2026.8.1 Beta 2 Ships Security Upgrades (8/24). https://openclawchronicles.com/posts/openclaw-2026-8-15-2026-8-1-beta-2-release
  2. OpenClaw Chronicles — OpenClaw Beta 3 Ships GPT-5.6 and SQLite Backups (8/24). https://openclawchronicles.com/posts/openclaw-2026-8-24-beta3-release
  3. OpenClaw Chronicles — OpenClaw Read Tool Handles Huge One-Line Files (8/20). https://openclawchronicles.com/posts/openclaw-2026-8-20-read-tool-huge-one-line-files
  4. OpenClaw Academy — OpenClaw Pins Codex Runtime 0.149 and Reworks Approval-Policy Migration (8/21). https://openclaw.academy/blog/openclaw-codex-0149-runtime-pin-approval-migration-august-2026
  5. OpenClaw Academy — Release Impact Stories (8/22-8/25). https://openclaw.academy/releases
  6. terminalblog.com — OpenClaw Just Dropped Secret Egress Host Binding — GPT-5.6 Ultra, macOS Profiles & SQLite Snapshots Land (8/16). https://terminalblog.com/blog/openclaw-2026-8-1-beta2-secret-egress-gpt56-macos-profiles
  7. agentdaily.dev — OpenClaw v2026.8.1-beta.2 Critical Security Enhancements (8/15). https://agentdaily.dev/content/deac0c74-dd5d-4093-aa9d-6f79c8bcf4c4
  8. nextagent.ca — OpenClaw Updates Timeline (8/24). https://nextagent.ca/en/updates/openclaw
  9. releases.sh — OpenClaw Release Notes & Changelog · August 2026. https://releases.sh/openclaw/releases
  10. releasebot.io — OpenClaw Release Notes (8/23-8/25). https://releasebot.io/updates/openclaw
  11. freedom.tech — OpenClaw Project Release History. https://freedom.tech/project/openclaw
  12. Xiaohonghua Digest xhh.club — OpenClaw beta.2 Emergency Patch! Your API Keys Are Running Naked! https://xhh.club/e/hgxjEqQSMPzp1h4CwCLc
  13. GitHub Releases — openclaw/openclaw v2026.8.1-beta.3 (8/24). https://github.com/openclaw/openclaw/releases/tag/v2026.8.1-beta.3
  14. GitHub Releases — openclaw/openclaw v2026.8.1-beta.2 (8/15). https://github.com/openclaw/openclaw/releases/tag/v2026.8.1-beta.2
  15. OpenClaw Academy — 8/23 OpenClaw Main Adds macOS Realtime Talk Relay and Fixes Blank Responses API Output. https://openclaw.academy/releases