---
title: "Platform Updates — July 24–31, 2026"
slug: "platform-update-2026-07-31"
date: "2026-07-31"
tag: "Application"
summary: "10 production image releases (1.109.0 → 1.119.0) plus the iblai-web-frontend packages: per-account usage checkpoints and OpenClaw worker cost attribution in the Data Manager, infinite-scroll projects and SDK-sourced agent settings in OS, nonce-based CSP enforced by default across the SPAs and web-utils v2, a v2 iblai-js and Auth SPA, field-level RBAC write filtering, the v1 application feature, and all-interface network binding in the CLI."
author: "ibl.ai Engineering"
linkedin: |
  Ten production image releases in eight days, across nine services — plus six frontend packages. A few of the things that shipped:

  → Usage checkpointing and cost attribution. A new seed_usage_checkpoints command sets each account's usage checkpoint before a cutoff date, so billing covers only post-cutoff usage — and checkpoints are only ever raised, never lowered. Worker-side agent cost is now persisted and attributed per platform.

  → Reconciliation that can't be held hostage. It no longer opens a WebSocket to each managed worker, so one unreachable worker can't delay or fail the run.

  → Model flexibility where it actually bites. Topic generation now adapts to models that reject explicit temperature values, and platforms running Anthropic models no longer need an OpenAI key sitting around to generate topics. Session titles handle structured content-block responses.

  → Content Security Policy enforced by default across six SPAs — OS, LMS, Auth, Course, HQ, and Video — nonce-based, shipped as web-utils v2, with enforcement decoupled from NODE_ENV.

  → Field-level RBAC write filtering, and the v1 application feature ("Apply") landing across web-containers with translations and a human-support tab.

  → Infinite-scroll project lists, agent settings sourced from the SDK, grading in embedded/iframe mode, and a desktop-app IPC deadlock resolved by marshaling Tauri events to the main thread.

  → Every SPA in the CLI now binds to all interfaces, so the port is reachable on the host regardless of which IP Docker's auto-set HOSTNAME resolved to.

  This is one week. Our customers run this stack on their own infrastructure, with the full source code — every release above is something they own rather than something they wait for.

  Full changelog: https://ibl.ai/updates/platform-update-2026-07-31

  #iblai #AgenticAI #EnterpriseAI #Engineering #ShipLog
---

Releases from July 24–31, 2026 — covering `iblai-prod-images` `1.109.0` through `1.119.0` and the `iblai-web-frontend` packages.

| service | image / pin | version span |
|---|---|---|
| dm | `iblai-dm-pro` | `4.316.1-ai` → `4.324.0-ai` |
| os | `iblai-os-spa` | `0.102.2` → `0.108.9` |
| lms | `iblai-lms-spa` | `0.51.0` → `0.54.3` |
| edx | `iblai-edx-pro` | `sumac.2.42.0` → `sumac.2.45.0` |
| auth | `iblai-auth-spa` | `1.19.0` → `2.1.1` |
| cli | `ibl-cli` | `5.52.0` → `5.53.1` |
| course | `iblai-course-spa` | `1.0.24` → `1.0.26` |
| hq | `iblai-hq-spa` | `2.1.26` → `2.1.28` |
| video | `iblai-video-spa` | `1.0.18` → `1.0.25` |
| web-utils | `@iblai/web-utils` | `1.13.5` → `2.1.6` |
| web-containers | `@iblai/web-containers` | `1.15.5` → `1.15.8` |
| iblai-js | `@iblai/iblai-js` | `1.26.6` → `2.2.7` |
| agent-ai | `@iblai/agent-ai` | `2.7.0` → `2.8.0` |
| data-layer | `@iblai/data-layer` | `1.9.6` → `1.11.1` |
| mcp | `@iblai/mcp` | `1.8.2` → `1.8.4` |

---

## iblai-dm-pro (`4.316.1-ai` → `4.324.0-ai`)

**Global configuration uniqueness.** A partial unique constraint (`unique_global_config_key`) now enforces one global row per configuration key where `platform IS NULL`, with a data migration that de-duplicates existing rows before applying the constraint.

**Session title generation with structured model responses.** `generate_session_title` now handles models (e.g. Anthropic/Claude) that return `AIMessage.content` as a list of content blocks, flattening via the `extract_text` helper before processing.

**Topic generation model flexibility.** `get_proposed_topic_from_llm` adapts to models that reject explicit temperature values (reasoning models, certain Azure deployments) by catching the `BadRequestError`, resetting to the default temperature, and retrying. Topic generation also uses the mentor's own Anthropic model when the provider is `anthropic`, removing the previous requirement for an OpenAI key on Anthropic-backed platforms.

**Documents list performance.** The documents list endpoint (`documents/pathways/{pathway}/`) now batches `MediaFilePermission` lookups into a single query and `bulk_update`, replacing per-document queries for S3-backed documents.

**Health-check mentor Langfuse opt-in is async-safe.** The `ENABLE_LANGFUSE_FOR_FAKE_MENTORS` flag is read via `sync_to_async` at all async call sites, sourced purely from dynamic config (defaults `False`), and short-circuits on `mentor.unique_id` before any DB access.

**Webcrawler periodic retraining.** URL-based documents (webcrawler and other URL sources) now correctly trigger periodic retraining on their configured `retrain_interval_days`. The periodic task logs candidate counts and each document scheduled for retraining.

**Usage reconciliation and checkpointing.** The cumulative cost rollup reads correctly against the current ClickHouse view structure. A new `seed_usage_checkpoints` management command sets `CreditAccount.usage_checkpoint_usd` to recorded usage before a cutoff date (default `2026-07-23`), ensuring billing covers only post-cutoff usage. Options: `--apply` writes checkpoints and records a zero-amount `checkpoint_backfill` transaction per account; `--reconcile` settles post-cutoff usage immediately; `--username` / `--platform` scope a targeted run. Checkpoints are only raised, never lowered.

**OpenClaw worker cost attribution.** Worker-side usage is now persisted and emitted to Langfuse for billing. Single-tenant hosts bill all worker cost to that platform via `ClawInstance.bill_all_usage` (default on, superuser-only); shared hosts leave cost unattributed. Reconciliation no longer opens a WebSocket to each managed worker, so an unreachable worker cannot delay or fail it. Kill switches: `CLAW_USAGE_INGEST_ENABLED` and `CLAW_USAGE_LANGFUSE_EMIT_ENABLED`.

### Deployment

- Run migrations. Migration `0339_claw_usage_persistence` must complete before the new image goes live — it adds `bill_all_usage` to `ibl_ai_mentor_clawinstance`, which is read on request paths. Sets `lock_timeout = 5s`; retry the deploy if it aborts under contention.
- Rolling deploy is safe once `0339` has run: a node on the previous image can still create workers because the column has a database-level default.
- Register the fake-LLM model for health-check cost tracking: `python manage.py register_langfuse_health_model`
- Usage ingest and Langfuse emission stay inert until periodic tasks are created (separate operator step).

## iblai-os-spa (`0.102.2` → `0.108.9`)

**Projects list pagination.** The projects sidebar now supports infinite-scroll pagination.

**Navbar routing.** "New Chat" correctly routes home when off a chat route, and the analytics page is no longer treated as a prompt-gallery route.

**Agent settings from SDK.** The in-repo agent settings tab is replaced by the SDK's `AgentSettingsTab` component, sourced from `@iblai/iblai-js`.

**Grading in embedded mode.** The iframe integration handles `MENTOR:ENABLE_GRADING` messages from the parent frame and reads `DOCUMENTFILTER`, `EDX`, and `ENABLE_GRADING` payloads from `event.data`. EDX metadata is dispatched via `chatActions.setMetadata`.

**Desktop app stability.** Tauri event emits are marshaled to the main thread, resolving an IPC deadlock in the desktop app.

**Welcome message layout.** Long embed welcome messages scroll correctly, and the default welcome message width matches the chat input.

**Header logo fallback.** The ibl.ai logo is used as the default header when no custom logo is configured.

**Content Security Policy.** CSP is now enforced by default (report-only in dev), with `*.iblai.org` and `*.iblai.tech` whitelisted. The enforcement decision is decoupled from `NODE_ENV`.

**RBAC permissions.** Mentor access RBAC permissions are corrected.

## iblai-lms-spa (`0.51.0` → `0.54.3`)

**Content Security Policy.** Nonce-based CSP middleware is now active via `@iblai/iblai-js/security`, with `*.iblai.org` / `*.iblai.tech` whitelisted and enforcement on by default.

## iblai-edx-pro (`sumac.2.42.0` → `sumac.2.45.0`)

**Grading-enabled field and API field renames.** A `gradingenabled` field is added, and `course_id` / `block_id` fields are renamed. This is a breaking change (PR [#175](https://github.com/iblai/iblai-edx-pro/pull/175)).

## iblai-auth-spa (`1.19.0` → `2.1.1`)

**Major version bump to v2.** The auth SPA moves to the v2 SDK release line. The `env.js` script is nonced for strict-dynamic CSP compliance.

## ibl-cli (`5.52.0` → `5.53.1`)

**SPA network binding.** The `vibe`, `courseai`, `videoai`, and `hq` SPAs now bind to all interfaces (`HOSTNAME: "0.0.0.0"`). Each runs the Next.js standalone server and attaches to both `ibl_local_default` and `internal-network`; without the override, the server bound whichever single IP Docker's auto-set `HOSTNAME` resolved, which could leave the SPA port unreachable on the host. `mentor` and `os` already carried this fix.

## iblai-course-spa (`1.0.24` → `1.0.26`), iblai-hq-spa (`2.1.26` → `2.1.28`), iblai-video-spa (`1.0.18` → `1.0.25`)

**CSP middleware.** All three SPAs now include nonce-based Content Security Policy middleware.

---

## @iblai/web-utils (`1.13.5` → `2.1.6`)

**Content Security Policy (breaking — v2.0).** CSP is now enforced by default in production (report-only in dev). Nonce-based, using `@iblai/iblai-js/security`. Allowlisted domains: `*.iblai.org`, `*.iblai.tech`. Enforcement is decoupled from `NODE_ENV`.

**Todo list extraction from chat.** `write_todos` tool calls are now recovered from chat history. `tool_call.end` frames are captured, and tool calls are upserted by ID instead of appended. Named chat socket frame types.

**Field-level RBAC write filtering.** New helpers for field-level RBAC write filtering.

**edX JWT forwarding.** Preserved token forwarded as `edx_jwt_token` on native redirect.

**Tenant-switch race fix.** Loader-hang race closed in `performAuthCheck` during tenant switches, with mid-check deferral coverage.

## @iblai/web-containers (`1.15.5` → `1.15.8`)

**Application feature v1.** New "Apply" UI with translations, completing the application feature across web-containers.

**Human support tab.** New tab for human support in the agent interface.

**Agent settings convergence.** Settings tab split into per-sub-tab components, converged onto OS mentor settings layout, with i18n alignment to the mentor monolith.

**Datasets tab enhancements.** Info box, RBAC gate, controlled props, theming, and controlled page/search in `useDatasetsWithPagination`. Training toggle uses `CustomSwitch`. RBAC role and policy save buttons disable while saving.

## @iblai/iblai-js (`1.26.6` → `2.2.7`)

**Major version bump to v2.** Tracks the `web-utils` v2 CSP enforcement release line. Auth SPA nonces the `env.js` script for strict-dynamic CSP.

**ACP integration.** Complete ACP integration in the auth package.

**Microsoft auth icon.** Microsoft icon added for microsoft/azure SSO providers.

## @iblai/agent-ai (`2.7.0` → `2.8.0`)

**edX JWT via request-jwt.** Fetches edX JWT via `request-jwt` with refresh-div fallback. edX grading context gated behind `enablegrading` flag.

## @iblai/data-layer (`1.9.6` → `1.11.1`)

**Studio course creation.** New endpoints integration for studio course creation.

## @iblai/mcp (`1.8.2` → `1.8.4`)

Tracks upstream `web-containers` and `data-layer` releases. No standalone feature changes.