---
title: "Earlier release highlights"
manual: "LLM Foundation"
version: "main"
permalink: "https://docs.typo3.org/permalink/netresearch/nr-llm:changelog-history@main"
source: "Changelog/History.rst"
rendered: "2026-11-12T15:41:35+00:00"
---

# Earlier release highlights {#changelog-history}

These are selected historical highlights, not a complete release ledger.

## Version 0.43.22 (2026-11-09) {#version-0-43-22}

The social image and file reference writers work in the editor's draft
workspace, AI-written pages and content elements can carry the b13/ai-label
marker, and the task endpoints refuse inactive tasks and, for editors, tasks
with table input.

-   `set_page_social_image` and `replace_file_reference` work in the
    editor's TYPO3 workspace: in a run started in a draft workspace they give
    the page or content element its version, add the new file reference in that
    workspace and mark the replaced one for deletion there, translations
    included. Live stays as it is until the workspace is published; the
    approval card names the workspace and the answer to the model says the
    change is a draft. `replace_file_reference` refuses an element moved in
    the draft workspace until the move is published or discarded (ADR-274).
-   The take-back of a failed write in a draft workspace discards only the rows
    the call made and keeps the editor's earlier drafts, also further down the
    record's inline children; the answer names each row that is not as it was
    before (ADR-274). One exception is named rather than undone: a failed
    `set_page_social_image` replacement can leave a translation's earlier
    image draft marked for deletion, and the answer says so.
-   Optional b13/ai-label integration for AI-written pages and content
    elements, with acting-user and workspace attribution, origin classification
    and review resets. The nr-llm marker remains available independently.
-   The task execute form and the input-refresh endpoint refuse an inactive
    task with "Task is not active". A backend user with the `tasks_use` grant
    who is no administrator can no longer run a task with the `table` input
    type or refresh its input; both endpoints answer "Task not found".
-   GPT-6 model ids get the reasoning profile only as the exact family name or
    with a `-YYYY-MM-DD` snapshot suffix. Other `gpt-6` ids, such as
    `gpt-6-sol-latest`, get the conservative unknown profile; configure the
    exact family name or a dated snapshot to keep the reasoning profile.
-   Usage statistics distinguish unreported from zero token counts
    (`isPromptTokensReported()`, `isCompletionTokensReported()`), cached
    answers are screened against the current output policy, and boolean tool
    arguments must be native JSON booleans: string and numeric spellings
    trigger argument correction.

## Version 0.43.21 (2026-11-08) {#version-0-43-21}

The file list offers the AI actions on files and folders, a new tool sets an
image's categories from a configured category tree, and agent runs apply the
retention of the provider that actually answered.

-   The file list offers the editor actions: **AI actions** in the context menu
    of a file opens the Editor Action Center for that file, and on a folder it
    lists the actions on files, each leading to the batch page with that folder
    as its scope. A file or folder outside the storage allow-list or the
    editor's file mounts gets no entry (ADR-267).
-   The writing tool `set_file_categories` and its editor action **Set
    categories from the image** choose only existing categories below the root
    named in the extension setting `vision.categories.rootUid` and inside the
    editor's category mounts. Without a root the action is not offered and the
    tool refuses every call (ADR-268).
-   Agent runs apply the retention ceilings of the provider actually used,
    including fallback and failed attempts, and persisted agent calls bypass
    provider-response idempotency caching (ADR-270). Run the TYPO3 database
    schema update before starting workers.
-   Measured quality routing now requires `routing.qualitySet` and
    `routing.qualityFingerprint`; re-evaluate the intended set and copy the
    fingerprint printed by `nrllm:eval:run`. Without them, quality is unknown
    and the other routing signals remain available.
-   The documentation resolves 115 audit findings and documents the Responsible
    AI capabilities, limitations and evidence procedures.

## Version 0.43.20 (2026-11-06) {#version-0-43-20}

The setup wizard comes first under AI > Setup, the error hints module asks
the configuration the administrator chose, in the backend language, and the
assistant no longer reports a write as refused because TYPO3 complained about
another record of the same save.

-   The setup wizard is the first entry under AI > Setup, followed by providers,
    models and configurations; use-case packs come last, because a pack's
    configuration needs a provider and a model. On TYPO3 14.3, while no
    provider exists, the AI > Setup overview opens with a notice and a button
    to the setup wizard. TYPO3 13.4 has no such overview: AI > Setup opens the
    last used entry, or the setup wizard on the first visit.
-   Error hints module: each administrator chooses which active LLM
    configuration is asked, from a list that marks what each configuration may
    receive by its trust zone. The choice is kept in the backend user's module
    data and falls back to the default configuration when the chosen one is
    deleted or deactivated. A refusal for the trust zone says what to do and
    links to Configurations and Providers.
-   Error hints module: refusals and the model's answer follow the
    administrator's backend language instead of always being English; the
    console command stays English.
-   A write the assistant made is no longer reported as refused because TYPO3
    logged an error about another record of the same save (for example from a
    DataHandler hook that writes another page). Every writer reads the record
    back first: what took is reported as written, with TYPO3's messages
    attached; a write of which nothing is left is reported as failed; a partial
    result is reported as written in part, to be checked. The attaching writers
    also check that the new reference points at the file asked for.

## Version 0.43.19 (2026-11-06) {#version-0-43-19}

The Analytics module's daily chart adds up to its cost, request and token
tiles again, and large counts fit their tiles.

-   The daily trend kept only the last bucket of a day that holds rows under
    more than one `request_date` (local midnight next to a seed written at the
    database server's midnight, or rows from before a timezone change). Every
    bucket now counts towards its day, for every range of the module.
-   The request and token tiles of the Analytics module and of the overview
    shorten counts from one million on in the backend user's locale
    ("17,5 Mio.", "17.5M"). The full number stays on hover and for screen
    readers.
-   The charts take their legend, axis titles and screen-reader table headers
    from the translations, and format their numbers in the backend language.

## Version 0.43.18 (2026-11-06) {#version-0-43-18}

The assistant's writers work in the editor's TYPO3 workspace. In a run started
in a draft workspace they store their change there, and live stays as it is
until the workspace is published; the approval card names the workspace.

-   The field writers `update_page_metadata`,
    `set_file_reference_alternative_text`, `set_file_alternative_text`,
    `update_fal_asset_meta` and `publish_record` store a workspace version
    ([ADR-262](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-262@main)). The redirect writers, which cannot keep a
    draft, are refused before the approval card in a draft workspace.
-   The creating writers `create_content_element_draft`,
    `create_page_draft`, `create_record_draft`,
    `attach_file_to_content_element` and `attach_file_to_record` create
    records new in the workspace, also on pages and records that exist there
    only ([ADR-264](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-264@main)). `set_page_social_image` and
    `replace_file_reference` still write in live only and are refused before
    the card in a draft workspace.
-   The structural writers `move_content_element`, `move_page`,
    `copy_record`, `delete_record` and `create_translation_draft` move,
    copy, mark for deletion and translate in the workspace
    ([ADR-266](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-266@main)). Commands TYPO3 would carry out differently there
    are refused before the card.
-   The fill of an empty field after a save and background batches started in
    a draft workspace record the workspace and write there
    ([ADR-265](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-265@main)); a fill whose draft was discarded, published or
    sent to review before the worker ran leaves the field empty.
    `nrllm:editor-action:batch` runs in live.

## Version 0.43.17 (2026-11-05) {#version-0-43-17}

A configuration can exclude single tools ([ADR-261](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-261@main)): the new
field **Excluded tools** withholds the selected tools from agent runs
with that configuration, on top of the allowed tool groups and every other
gate, when the tools are offered and again when a call is executed. It never
enables a tool, and a tool registered later is not excluded until it is
selected. A run waiting for an approval or input keeps the exclusions it
started under. The list of tools a chat is not offered reports such a tool with
the reason `configurationExcluded`; the Tools module and the
**Governance** tab show which configuration excludes which tool.

The read tools read records as the run's draft workspace shows them
([ADR-263](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-263@main)): `read_records`, `search_records`,
`get_pagetree`, `read_fal_asset_meta`, `search_fal_files`,
`get_fal_references` and `highlight_record` return each record under its
live uid with filters, limits and offsets applied to what that workspace shows,
and `get_record_history` and `get_tsconfig` read in that workspace. In live,
and in a run that did not record its workspace, they read live rows.
`get_record_history` and `get_tsconfig` no longer read another workspace's
draft when the model names a version row's uid or a record new in another
workspace.

Choosing a model in the model form no longer unticks capabilities nobody
denied. Where discovery cannot say what a model lacks — an OpenAI or Gemini
chat model the built-in catalogue does not list yet, every Groq model, an
Ollama model whose capabilities the host did not report — the list is marked
as inferred, and a record that already ticks that kind of model keeps its other
capabilities, such as `tools` and `vision`.

## Version 0.43.16 (2026-11-05) {#version-0-43-16}

An MCP server can be connected through OAuth ([ADR-259](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-259@main)): the
new authentication mode *OAuth connection* discovers the server's
authorization server, uses a pre-registered client or registers one through
Dynamic Client Registration, and runs the authorization code flow with PKCE
from **Connect** in the MCP module. Access and refresh tokens are
Vault secrets and are renewed before they expire; **Disconnect**
deletes them.

The chat changes a content element in the editor's TYPO3 workspace
([ADR-260](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-260@main)): a run started in a draft workspace stores a
workspace version of the element through the DataHandler, live stays as it is
until the workspace is published, and the approval card names the workspace.
`get_page_content` reads the page as that workspace shows it. A run is pinned
to the workspace it was started in and writes nowhere else; a run that does not
record it keeps writing in live only. The other writers still refuse a draft
workspace. `AiActorContext` carries `workspaceId` and
`AfterAiRecordWrittenEvent` carries `workspace`.

## Version 0.43.15 (2026-11-05) {#version-0-43-15}

The model rounds of an agent run are settings ([ADR-101](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-101@main)):
`agent.defaultMaxIterations` (default 5) for a run that asks for no number,
and `agent.maxIterations` (default 20), the most any run may take, also after
an approval or an input, with the value in force when the run resumes. Invalid
values fall back to 5 and 20, and a default above the maximum is the maximum.
The Governance tab and the playground show the values that apply, and the new
`AgentRoundLimitsInterface` lets consuming extensions read them.
`AgentRuntime::MAX_ITERATIONS` is now the fallback of the maximum.

## Version 0.43.14 (2026-11-04) {#version-0-43-14}

Error hints for administrators ([ADR-258](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-258@main)). The backend module
*AI > Operation > Error hints* lists the ten newest errors from the TYPO3 log
files and asks a model, per entry and on request, for the likely cause and the
next step; `nrllm:exception:hint` does the same on the command line. What is
sent depends on the provider's trust zone, argument values and source code are
never sent, and a hint is labelled a guess unless it points at a frame in the
project's own code. `get_last_exception` and `probe_url` now name the class
and throw site of entries TYPO3's handlers log without a trace.

## Version 0.43.13 (2026-11-04) {#version-0-43-13}

Five administrator tools for site chores ([ADR-257](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-257@main)).
`list_redirects` lists the redirects of EXT:redirects; `create_redirect`,
`update_redirect` and `disable_redirect` create, change and switch off one
redirect under the redirects module's rules, and the approval card shows each
field before and after and when visitors get the change. `clear_cache` clears
one cache group or the page caches by tag; its card names a full flush as one.
The four writers ship disabled and pause for approval.

## Version 0.43.12 (2026-11-04) {#version-0-43-12}

Routing can rank by what calls actually took and cost. With the new setting
`routing.callHistory` (off by default), the Balanced, Quality and Economy
modes add each model's mean latency over its last seven days of calls of the
same operation, and rank cost by the measured cost per call instead of the list
price wherever every priced candidate has one ([ADR-256](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-256@main)). A call
through a criteria-mode configuration now records its cost and model, so cost
budgets count those calls from this version on. Provider health is looked up by
adapter type, which can change decisions in a measured mode. Database change:
two columns on `tx_nrllm_telemetry`.

## Version 0.43.11 (2026-11-04) {#version-0-43-11}

A configuration can answer a repeated plain chat or completion request from a
cache: the new field `answer_cache_ttl` (seconds, 0 = off by default, at most
one day) serves the same answer again without a provider call. Agent runs, tool
calls, schema-bound calls and streamed answers are never cached, and an answer
is kept only where the content privacy level is `full` for the provider that
served it ([ADR-249](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-249@main)). Evaluation runs always ask the model.
Database change: one column on `tx_nrllm_configuration`.

## Version 0.43.10 (2026-11-04) {#version-0-43-10}

The text under a record select says what the chosen model, configuration,
task, provider or skill is for, and changes with the choice, also for keyboard
and screen reader users. Next to the select, a button creates a new record and
another opens the module that lists them; each is shown only to users who may
use it ([ADR-255](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-255@main), [Choosing a record from a list](https://docs.typo3.org/permalink/netresearch/nr-llm:administration-record-selects@main)).

## Version 0.43.9 (2026-11-03) {#version-0-43-9}

A consuming extension reads what it spent and how often each of its operations
ran and failed through the public `ExtensionUsageReaderInterface`, which also
says how far back the per-call counts reach ([ADR-248](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-248@main)). A
streamed call records its caller on the usage aggregate as a non-streamed call
does, so its cost is no longer filed as unattributed.

## Version 0.43.8 (2026-11-03) {#version-0-43-8}

A tool call whose arguments do not fit the schema its tool was offered with is
answered with what does not fit, and nothing of that turn runs: no tool, no
approval card, no question, no external call budget ([ADR-246](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-246@main)).
Azure OpenAI provider records send the key in the `api-key` header and reach
the deployment with an `api-version`, or the v1 API when the endpoint ends in
`/openai/v1` ([ADR-247](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-247@main)). Telemetry rows of a configuration in
automatic model selection name the model the selection chose instead of an
empty one, and such a configuration passes that model to the adapter, so its
embeddings no longer fall back to the adapter's default model.

## Version 0.43.7 (2026-11-03) {#version-0-43-7}

The skill select of a configuration and a task filters and labels its skills
on TYPO3 13.4 and 14.3: an orphaned or disabled skill is offered only where it
is already attached, with its state in the label, and a new `always`
attachment no longer offers a process skill. Before, the select offered every
skill unfiltered.

## Version 0.43.6 (2026-11-03) {#version-0-43-6}

The page for running an editor action on several records no longer asks for
typed record numbers or a typed page number or folder identifier: the records,
and the page or folder a background batch starts from, are chosen in TYPO3's
element browser, on TYPO3 13.4 and 14.3. The server checks the choice as
before.

## Version 0.43.5 (2026-11-02) {#version-0-43-5}

A task can be marked for the places that offer it, in the new field
**Offered first in**. The editors' module *AI tasks* lists the tasks
marked for it first, under *Prepared for this module*, and labels the other
groups by category; a task marked only for another extension's place is left
out there. Another extension orders its own task select by a placement with
`TaskPlacementItems::groupByPlacement()` (ADR-245,
[Placements](https://docs.typo3.org/permalink/netresearch/nr-llm:configuration-tasks-placements@main)). Run the database schema update for the
new column `tx_nrllm_task.placements`.

## Version 0.43.4 (2026-11-02) {#version-0-43-4}

A site can name the LLM configuration its AI calls use and the DeepL key its
translations use, as identifiers in two site settings of the new set
`netresearch/nr-llm`; a site that names neither keeps the installation's
(ADR-244, [Site settings](https://docs.typo3.org/permalink/netresearch/nr-llm:administration-site-settings@main)). A caller states its site with
`withSite()`, which also scopes the term lists bound to a site.

## Version 0.43.3 (2026-11-02) {#version-0-43-3}

A provider and a task can tighten the privacy level for what is stored of their
runs, with a `privacy_level` field. It cannot loosen the global level
(ADR-242). The rule set is a tagged snippet plus a term list, with no new record
(ADR-243). Run the database schema update for the new columns.

## Version 0.43.2 (2026-11-02) {#version-0-43-2}

In an agent run that offers tools, as the AI chat does, a term list checks only
what a writing tool would store. Chat text streams and stays free, and a search
for a forbidden term runs (ADR-238). A reasoning model of OpenAI gets an output
limit of at least 16,000 tokens unless the call switches thinking off, and an
answer that ends without text at the limit raises `OutputBudgetExhaustedException` (ADR-241).

## Version 0.43.1 (2026-11-01) {#version-0-43-1}

The text of a tool round reaches an observer while the model writes it, on OpenAI's
Responses transport (ADR-240). A chat can show an answer before the round has ended.
A run of text without whitespace now waits in the live redaction until it ends, up
to 4 KB, so a JWT is masked whole (ADR-088).

## Version 0.43.0 (2026-11-01) {#version-0-43-0}

Term lists, a mark on what an AI wrote, and where the time of a model call goes.
Update the database schema: `pages` and `tt_content` gain
`tx_nrllm_ai_written`, and the table `tx_nrllm_term_list` is new. Until the
update ran, new pages and content elements cannot be saved. A consumer pinned to
`^0.42` needs `^0.43`. Details in the repository `CHANGELOG.md`.

-   Term lists: forbidden and required terms in a record stop generated text, and
    the editor reads which term and which list did it (ADR-238).
-   A page or content element an agent run or a fill run wrote carries
    `tx_nrllm_ai_written` (ADR-239).
-   The step of a model call in the Verlauf shows the cached part of the prompt, the
    wall time of the HTTP call and the time the provider reports.

## Version 0.42.0 (2026-11-01) {#version-0-42-0}

Background work. Update the database schema: `tx_nrllm_agentrun` gains
`origin`, `answer` and `outcome`. A consumer pinned to `^0.41` needs
`^0.42`. Details in the repository `CHANGELOG.md`.

-   `nrllm:task:enqueue` queues an AI Task, `nrllm:agent:work` works the queue
    from cron or the scheduler without a Messenger consumer, and the Agent Runs
    module lists these runs under *Background runs* (ADR-234).
-   `AgentRuntimeInterface::defer()` stores a queued run and never executes it
    in the caller. A class outside nr-llm that implements the interface has to add
    it.
-   `worker.technicalBeUserUid` names the backend user whose grants on a vault
    secret count on the command line. With `allowCliAccess` off and without this
    setting, every queued run fails on its first model call.
-   The editor actions for the alternative text and the file metadata send the
    image to the model (ADR-235). They are no longer offered on a model that
    declares capabilities without `vision`.
-   An editor action runs over a folder or a page tree in the background, from the
    batch page and from `nrllm:editor-action:batch`. Each record becomes a run
    that waits for its own approval (ADR-236).
-   An empty field can be filled by AI after a save, from `fieldFill.fields`
    (ADR-237).

## Version 0.41.18 (2026-10-31) {#version-0-41-18}

Behaviour change in guided processes; no database schema or extension
setting change. Details in the repository `CHANGELOG.md`.

-   In a run holding a process pin, a turn that pauses for input runs the
    reads beside the question before the pause, as a turn that pauses on a
    proposal already does (ADR-214, amendment of 2026-10-31). A process step
    can report its progress and highlight in the same response as its
    question, without a model round of their own. A remote tool, a write and
    a second input call beside the question do not run. Outside process runs
    nothing changes.

## Version 0.41.17 (2026-10-30) {#version-0-41-17}

Bug fix; no database schema or extension setting change. Details in the
repository `CHANGELOG.md`.

-   A write proposal shows a length only against a configured range
    (ADR-214): the approval card of `update_page_metadata` and
    `update_content_element` no longer shows a bare character count, such
    as "Länge 23 Zeichen" for a page title, for a field without a range in
    `tools.structuredPreview.ranges`. With the shipped default only
    `pages.description` carries one.
-   The fourth parameter of `FieldMeasurer::measure()`,
    `$countWithoutRange`, is optional, deprecated and ignored. It stays in
    place so that an existing call keeps passing `$richText` fifth; new
    calls omit it and pass `richText:` by name.

## Version 0.41.16 (2026-10-30) {#version-0-41-16}

Bug fix; no database schema change. Details in the repository
`CHANGELOG.md`.

-   Gemini structured output keeps properties whose type is one type plus
    `null`, such as `["string", "null"]`: they are sent as that type
    with `nullable: true`. An object whose properties were all nullable
    previously reached Gemini without properties, and Gemini rejected the
    request with HTTP 400.
-   An object without any expressible property, and an array without
    expressible items, are left out of the response schema instead of
    being sent empty.

## Version 0.41.15 (2026-10-29) {#version-0-41-15}

Bug fix; no database schema change. The unused extension settings
`image.dalle.defaultModel` and `image.dalle.defaultSize` are removed.
Details in the repository `CHANGELOG.md`.

-   Image generation without a model no longer calls the retired
    `dall-e-3`: `DallEImageService::generate()` and
    `generateMultiple()` take the model registry's default image model
    and fall back to `gpt-image-2`. An explicit model still wins.
    `ImageGenerationOptions` defaults `model` to `null`.
-   `ImageGenerationOptions::hd()` is sent as quality `high` to
    gpt-image models; `standard` is not sent, so the API default
    `auto` applies.
-   The image test page reports a request the provider refuses as a
    refused request with its status code (502) instead of "not
    configured". The exception carries the status through
    `getStatusCode()`.

## Version 0.41.14 (2026-10-29) {#version-0-41-14}

No database schema or extension setting change. Details in the
repository `CHANGELOG.md`.

-   Optional revision of a DeepL translation by a language model
    (ADR-233): `TranslationOptions::withPolish()` lets
    `translateWithTranslator()` and its batch variant revise the result
    for fluency, terminology and tone. The revision is an ordinary request
    (a polished translation uses two); when it changes tags, placeholders
    or glossary terms, adds text, fails or is refused by the budget, the
    first translation is returned. Off by default.

## Version 0.41.13 (2026-10-28) {#version-0-41-13}

No database schema or extension setting change. Details in the
repository `CHANGELOG.md`.

-   A streamed answer reaches the caller while the provider is still
    writing it. The streaming adapters send with nr-vault's
    `sendStreaming()` (nr-vault 1.1) instead of the blocking
    `sendRequest()`, which delivered every chunk at the end. A provider
    without an API key, such as a local Ollama, streams too, after the same
    host check, and each such call writes an nr-vault audit row.
-   A dropped connection before the first chunk still lets the streaming
    fallback try the next configuration. A 401 or 429 whose error body is
    cut off keeps its exception class.
-   With nr-vault 0.16 or without `curl_multi_*` the request is sent as
    before.

## Version 0.41.12 (2026-10-28) {#version-0-41-12}

No database schema or extension setting change. Details in the
repository `CHANGELOG.md`.

-   New developer page [Report usage from your own provider calls](https://docs.typo3.org/permalink/netresearch/nr-llm:developer-own-provider-usage@main): how an extension
    that calls a provider with its own code checks the editor's budget and
    reports its usage, so its calls count against the budget and appear in
    Analytics and the billing report under its extension key.
-   Translations through `translateWithTranslator()` and
    `translateBatchWithTranslator()` keep the caller named with
    `withCallerSource()`; before, DeepL and the LLM translator recorded
    them unattributed.
-   `nrllm:demo:seed` records seeded completion usage as `complete`.

## Version 0.41.11 (2026-10-28) {#version-0-41-11}

No database schema or extension setting change. Details in the
repository `CHANGELOG.md`.

-   An editor's own budget status is public API: `UserBudgetStatusService`,
    the value objects `UserBudgetStatus`, `BudgetWindowStatus` and
    `BudgetMeter`, and `LocalizedNumberFormatter` are now `@api`, so
    another extension can show the usage and limits the AI Tasks module
    shows (ADR-232).
-   The billing report names its entitlement count "Entitled to AI Tasks".
    Users whose access comes only from an extension with its own permission
    checks, such as an editor assistant or a chat, are not part of it.

## Version 0.41.10 (2026-10-28) {#version-0-41-10}

No database schema change. A new extension setting
`billing.markupPercent` (whole percent, default 0) sets the markup of
the billing report. Details in the repository `CHANGELOG.md`.

-   A monthly billing report in the Analytics module: for a calendar month,
    the estimated cost with the markup, the active editors with their cost,
    the editors entitled to the AI features as of today, and the results by
    service type, calling extension and task, with a CSV export of the same
    month.

## Version 0.41.9 (2026-10-28) {#version-0-41-9}

Bug fix; no database schema or extension setting change. Details in the
repository `CHANGELOG.md`.

-   Backend views print token counts and USD amounts in the backend user's
    language: a German reader sees "5.138.706" and "11,40 $" instead of
    "5,138,706" and "$11.40". This covers the AI Tasks budget panel, the
    run estimate, the Analytics and overview modules, the usage columns of
    the list views, the AI actions batch estimate and the agent runs.
    Amounts stay in USD.

## Version 0.41.8 (2026-10-28) {#version-0-41-8}

No database schema or extension setting change. A new AJAX route
`nrllm_aitask_budget` serves the AI Tasks budget panel. Details in the
repository `CHANGELOG.md`.

-   The AI Tasks module shows editors their own usage today and this month
    against their personal budget, with a progress bar per limit, and says
    when no personal limit is set. The run form adds a cost estimate for
    one run of the task and states whether the budget lets the next run
    start, naming the limit that would refuse it.
-   Landing pages: canonical and language alternate metadata outside the
    document head is rejected, and the bilingual title panels follow the
    shared page family.

## Version 0.41.7 (2026-10-25) {#version-0-41-7}

Bug fix; no database schema or extension setting change.
`ReplaceFileReferenceTool` takes a fourth constructor argument, which
the container wires. Details in the repository `CHANGELOG.md`.

-   The writing tools `set_file_reference_alternative_text`,
    `attach_file_to_record` and `replace_file_reference` check every
    record of another table that holds an inline child, up the chain,
    before they write. An editor can no longer change a file reference on
    an inline child whose outer record is locked, outside their grant or on
    a page they may not edit. A relation of a table to itself, such as a
    grid container holding content elements, is not followed, because the
    page module edits those children on their own (ADR-230).

## Version 0.41.6 (2026-10-25) {#version-0-41-6}

New feature; no database schema or extension setting change. Details in
the repository `CHANGELOG.md`.

-   The new writing tool `set_file_reference_alternative_text` sets the
    alternative text of one use of an image (`sys_file_reference`), in
    place and in the reference's language. It ships disabled, works in the
    live workspace only and asks for approval (ADR-230).

## Version 0.41.5 (2026-10-25) {#version-0-41-5}

Database schema change: run the database analyzer, because
`tx_nrllm_skill` gains the column `page`. Existing rows keep today's
behaviour, and no skill approval is invalidated. Two additive public API
changes. Details in the repository `CHANGELOG.md`.

-   A skill can declare in its frontmatter whether it works on a page:
    `page: required`, `optional` or `none`. `Skill::getPageNeed()`
    returns it as `SkillPageNeed`. The field is not part of the version
    digest.
-   `ApproverNotPermittedException` and
    `SubmitterNotPermittedException` say why they refused:
    `getCause()` and `getDenialReason()`. A surface can tell a
    switched-off tool from missing permissions.
-   The decision and reranker sidecar images install filelock 4.2.4 and
    setuptools 84.2.1 for open advisories.
-   The public product site: mobile layout, language switch and a header
    shared with the other Netresearch pages.

## Version 0.41.4 (2026-10-24) {#version-0-41-4}

Bug fix; no database schema or extension setting change. One additive
public API change: `ApprovalDenialReason` gains the case `KEEP`.
Details in the repository `CHANGELOG.md`.

-   A guided process can tell "keep the current content" from "skip this
    point". The denial result for `KEEP` says the point is settled and
    must not be called skipped.

## Version 0.41.3 (2026-10-22) {#version-0-41-3}

Bug fixes; no database schema or extension setting change. One additive
public API change: `SuspendedRunState` takes an optional
`maxIterations` as its last constructor parameter. Details in the
repository `CHANGELOG.md`.

-   A resumed agent run keeps the round cap it was started with. Before,
    every segment after an approval or an answered input ran on the loop
    default of 5. A run suspended before this release resumes on the
    default as before.
-   `set_page_social_image` with `replace` removes a copy of the old
    reference that a translation left on the default-language page, also
    on TYPO3 14.3.10.

## Version 0.41.2 (2026-10-22) {#version-0-41-2}

Bug fixes; no public API, database schema or extension setting change.
Details in the repository `CHANGELOG.md`.

-   An invoked process skill is composed as its own section, "Active
    process", ahead of the approved skills, and governs the order of steps
    and the structure of answers; approved skills keep applying to wording,
    terminology and safety. Before, an approved skill attached always could
    override the process, so a guided tour ended its first turn without the
    tool calls it prescribes.
-   Skill import refuses a GitHub repository tree that GitHub reports as
    truncated and keeps the last known-good skills, and a malformed GitHub
    response keeps its typed failure when the diagnostic logger throws.
-   The decision and reranker sidecar images build again on Python 3.14,
    lock their dependencies with uv and install certifi 2026.10.19.

## Version 0.41.1 (2026-10-21) {#version-0-41-1}

Bug fixes and test coverage; no public API or extension setting change. Update
the database schema: skill sources gain two columns for the synchronization
lease. Skill sources, skills and the skill audit table must share one database
connection. Details in the repository `CHANGELOG.md`.

-   A process skill attached in the field "Skills loaded on demand" is listed,
    can be started and continues; before, only skills loaded always counted
    (ADR-214).
-   Gemini tool conversations keep native thought signatures and echo native
    function-call IDs through parallel calls and stored approval resumes
    (ADR-222, ADR-224).
-   MCP imports skip tools whose schema they cannot keep intact instead of
    importing a weakened schema.
-   A failing preview read check no longer blanks the approval inbox, and
    optional diagnostic loggers that throw no longer replace the original
    outcome in agent recovery, governance, MCP import, skill tokens and the
    Ollama model picker.
-   Skill metadata closes only at a standalone `---` delimiter.
-   Skill synchronization holds a persisted lease and publishes skill, orphan
    and audit changes in one transaction (ADR-221).
-   Gemini responses keep every visible text part and separate native thinking
    (ADR-223); MCP imports report tool names another server already holds.

## Version 0.41.0 (2026-10-20) {#version-0-41-0}

Approved skill versions, guided processes and skills loaded on demand
(ADR-214, ADR-228), with a security fix for the read tools. Callers of the
PHP API are unaffected; a class that implements `AgentRuntimeInterface`,
`ToolCallPolicyInterface` or `ToolLoopServiceInterface` has to follow
their new members. Update the database schema and run the upgrade wizard
"Write the version digest onto existing skills". Details in the repository
`CHANGELOG.md`.

-   An administrator approves one version of a skill; an approved version from
    a trusted source is composed into the system message as an instruction,
    every other skill stays fenced. Skills can be written in the backend and
    loaded on demand through the new `load_skill` builtin.
-   A guided process is started, continued and ended through a skill
    invocation and decided in the chat, one proposal per turn. New builtins
    `ask_choice` and `highlight_record` ship disabled.
-   Approval cards carry a write's record, fields and proposed values as
    structured data, and a write states whether it did everything it planned.
-   Invocation rules, optional actor-bound MCP authentication, retrieval
    experiment provenance, verified evaluation generators and worker
    operations status (ADR-215, ADR-216, ADR-217, ADR-219, ADR-220).
-   Security: the read tools check web mounts against the run's acting user,
    not the backend user of the request (#1022).

## Version 0.40.0 (2026-10-08) {#version-0-40-0}

Every built-in approval preview is in the acting user's language. No breaking
change to the PHP API; details in the repository `CHANGELOG.md`.

-   The approval cards of the remaining fifteen previewing tools and the lines
    the tool loop puts on a card use the acting user's language (ADR-213,
    #1016). Runs suspended before the upgrade bounce once at resume.
-   New public class `ApprovalPreviewHeadings` names the first line of every
    built-in preview for approval cards in other extensions.
-   TCA labels on approval cards resolve TYPO3 14 translation domain
    references.
-   `attach_file_to_content_element` refuses a translated content element.

## Version 0.39.0 (2026-10-07) {#version-0-39-0}

Typed decisions (ADR-211), hidden-record reading, a tool to attach a file to
any record's file field, and localised approval previews. Breaking changes to
the PHP API; full details and the migration in the repository
`CHANGELOG.md`.

-   **Breaking:** `completeStructured()` and
    `completeStructuredForConfiguration()` return a
    `StructuredCompletionResponse`; read `->data` where the array was read.
    Implementations of `CompletionServiceInterface` and
    `LlmServiceManagerInterface` gain the optional `?ModelResolution`
    parameter, the latter also `decideForConfiguration()`.
-   **Breaking:** model prices are `float` and stored as `decimal(12,2)`
    (database compare).
-   **Breaking:** the `llm_judge` evaluation grader is replaced by
    `--grader decision`.
-   `DecisionServiceInterface`, the TypeSafe provider and a local decision
    sidecar answer typed yes/no, choice and score questions.
-   `read_records` and `search_records` take `include_hidden` (#1009).
-   New writing tool `attach_file_to_record` (ADR-212, #1011).
-   Approval previews of `create_page_draft`, `move_page` and
    `delete_record` use the acting user's language (ADR-213, #1012).

Versions 0.21.0 to 0.38.2 are listed in the repository `CHANGELOG.md`.

## Version 0.20.0 (2026-07-16) {#version-0-20-0}

Closes out the operator-config audit: every backend-visible provider, model
and configuration setting now either works or is gone. Three breaking
changes; full details in the repository `CHANGELOG.md`.

-   **Breaking:** the provider *API timeout* is now applied as a
    total-response timeout (previously write-only — requests ran unbounded).
    Default moves 30 → 120 seconds; the `nrLlm_providerApiTimeout120`
    upgrade wizard migrates rows persisted at the old default; timed-out
    requests are not retried (#384).
-   **Breaking:** *Max retries* counts retries after the initial attempt —
    `0` sends exactly one request instead of none (#387).
-   **Breaking:** the never-functional PromptTemplate stack is removed
    (ADR-069); drop the orphaned `tx_nrllm_prompttemplate` table via the
    database analyzer (#399). `BudgetServiceInterface::check()` gained an
    optional `?LlmConfiguration` parameter (#389).
-   Per-configuration daily limits (requests / tokens / cost) are enforced
    alongside per-user budgets — most restrictive wins (#389).
-   Model *Max output tokens* and *Dimensions* act as call defaults when the
    caller and configuration leave them unset (#390).
-   Provider *Organization ID* is sent as `OpenAI-Organization`;
    `options.customHeaders` is applied on all request paths (#388).

## Version 0.19.1 (2026-07-15, first shipped with 0.20.0) {#version-0-19-1}

Prepared but never tagged; its fixes first ship in 0.20.0.

-   Criteria-mode configurations work with every `*ForConfiguration()`
    call — the model is resolved through `ModelSelectionService` instead of
    the (null) direct relation (#372).
-   Embedding cache tags built from dotted preset identifiers are sanitized
    (#372).

## Version 0.19.0 (2026-07-14) {#version-0-19-0}

-   Per-user usage attribution for the specialized speech and image services
    (ADR-057): transcription, synthesis and image generation record the
    calling backend user instead of the ambient bucket (#362).
-   Configuration identifiers with dots no longer crash cached calls — cache
    keys and tags are sanitized (#365); `LlmTranslator` chat rows carry the
    per-user attribution (#361); backend model-test button fixed (#363).

## Version 0.18.0 (2026-07-14) {#version-0-18-0}

-   Configuration presets module UI (ADR-056 follow-up): the Configurations
    backend module surfaces pending presets with one-click apply/update.

## Version 0.17.0 (2026-07-13) {#version-0-17-0}

-   Configuration presets: consuming extensions declare the
    `LlmConfiguration` records they need via the
    `nr_llm.configuration_preset` registration (ADR-056).

## Version 0.16.1 (2026-07-10) {#version-0-16-1}

Compatibility fixes for strict-mode MySQL/MariaDB, surfaced by the new
MariaDB CI leg: the setup wizard's AJAX persist path now enforces the
column length limits and the TCA identifier contract (overlong values
previously produced a 500 instead of being truncated), generated
identifier suffixes are collision-free within a batch, and
decimal-backed values such as `temperature` are rounded to their
column scale so every DBMS round-trips identical values.

For the complete, itemised list see the canonical
[CHANGELOG.md](https://github.com/netresearch/t3x-nr-llm/blob/main/CHANGELOG.md).

## Version 0.16.0 (2026-07-10) {#version-0-16-0}

This release adds **RAG site-search tools**: the new `rag` tool group
lets agent runs answer questions about the website's own public
content with cited evidence. `site_rag_query` returns curated sources
(`source_id`, title, URL, match excerpt) from a priority cascade over
whichever search index is installed — EXT:solr (via its HTTP select
API), ke_search, indexed_search — with an always-available
pages/tt_content database fallback that matches natural-language
questions word-wise; `site_fetch_source` reads a source's full
indexed text. Index-level filtering is strictly public-only (what the
anonymous visitor could read), and every evidence package names the
answering backend. See [ADR-049](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-049@main).

For the complete, itemised list see the canonical
[CHANGELOG.md](https://github.com/netresearch/t3x-nr-llm/blob/main/CHANGELOG.md).

## Version 0.15.0 (2026-07-09) {#version-0-15-0}

This release turns the Tool Playground into a **glass-box run inspector** —
every request, response, tool execution and the final answer stream live
into an ordered step list with token, cost and timing detail, plus a
dry-run mode that shows the exact prompt without calling the model — and
grows the built-in tool set to **39 tools in 7 groups**: record history
("who changed this?"), URL resolution, TCA/TypoScript validation,
error-analysis (last exception, source read/search, URL probe), FAL file
tools and system diagnostics. Groups can be toggled centrally, restricted
per configuration and selected per run, cascading fail-closed.

**Breaking:** `ToolInterface` gained `getGroup(): string` — every tool
implementer must declare its group (third-party tools: use the extension
key).

For the complete, itemised list see the canonical
[CHANGELOG.md](https://github.com/netresearch/t3x-nr-llm/blob/main/CHANGELOG.md).

## Version 0.14.1 (2026-07-05) {#version-0-14-1}

A patch release fixing tool calling with parameterless tools. A tool that
takes no arguments emitted its (empty) JSON-Schema `properties` — and its
empty replayed call `arguments` — as a JSON array `[]` instead of an
object `{}`, which strict providers such as Ollama reject with an HTTP 400.
The bounded agent loop and the Tool Playground now work when a parameterless
tool (environment, PHP info, or backend user/group introspection) is offered.

The Skills, Tools and Playground admin documentation was also refreshed to
match the shipped backend (module section count, the full built-in tool
catalogue, two-tier tool authorization, and updated screenshots).

For the complete, itemised list see the canonical
[CHANGELOG.md](https://github.com/netresearch/t3x-nr-llm/blob/main/CHANGELOG.md).

## Version 0.14.0 (2026-07-04) {#version-0-14-0}

This release adds a **Skills and Tools** system: extensions and editors can
ingest `SKILL.md` files from GitHub (SHA-pinned, admin-reviewed) and attach
them to tasks and configurations, and a function-calling **tool runtime** lets a
model run an agent loop over an admin-curated, permission-gated set of tools —
with an interactive tool playground in the backend. It also lands a broad
security and accessibility hardening pass (SSRF/CSRF fixes, API keys moved out of
URLs, RBAC on tool execution, EN/DE translations, WCAG text alternatives), and
the CI now actually runs the functional and backend end-to-end suites so they
gate merges.

**Breaking:** custom `ToolInterface` implementations must now declare a
`requiresAdmin(): bool` method — `true` for tools that expose system,
host, or cross-user data, `false` for tools that self-enforce the acting
user's TYPO3 permissions. Without it the tool fails at runtime (ADR-038).

Together AI, Fireworks AI and Perplexity are now first-class OpenAI-compatible
providers, and provider endpoints entered in the wizard or the record editor are
canonicalized on save so they no longer break when saved without an API version
path.

For the complete, itemised list see the canonical
[CHANGELOG.md](https://github.com/netresearch/t3x-nr-llm/blob/main/CHANGELOG.md).

## Version 0.13.0 (2026-06-26) {#version-0-13-0}

Provider selection is now database-driven end to end. **Breaking:** the
extension-configuration `defaultProvider` fallback is removed — select a
provider per call (the options object's `provider` field) or mark a
Configuration active and default in the backend module; otherwise
`getProvider(null)` throws (ADR-034). The dead `plugin.tx_nrllm` TypoScript
constants/setup were removed and the "no provider specified" error now carries
actionable backend-module guidance (#254, #255).

For the complete, itemised list see the canonical
[CHANGELOG.md](https://github.com/netresearch/t3x-nr-llm/blob/main/CHANGELOG.md).

## Version 0.12.0 (2026-06-11) {#version-0-12-0}

Specialized services (image, text-to-speech, transcription) gain full usage and
cost tracking, join the model registry with `image`/`text_to_speech`/
`transcription` capabilities, and resolve their model and system prompt from
Configuration records (ADR-032, ADR-033). Adds a prompt-snippet library
(`tx_nrllm_promptsnippet`, ADR-031), per-request timeouts on the secure HTTP
client, and arbitrary gpt-image sizes. **Requires nr-vault ^0.10.0.**

For the complete, itemised list see the canonical
[CHANGELOG.md](https://github.com/netresearch/t3x-nr-llm/blob/main/CHANGELOG.md).

## Version 0.11.1 (2026-06-10) {#version-0-11-1}

Security and robustness fixes from the extension-wide code review: the setup
wizard dispatches through nr-vault's SSRF-guarded secure HTTP client, the
provider adapters surface streaming errors as typed, credential-sanitized
exceptions, TTS text splitting is multibyte-safe, and FAL/Whisper
configuration parsing is hardened.

For the complete, itemised list see the canonical
[CHANGELOG.md](https://github.com/netresearch/t3x-nr-llm/blob/main/CHANGELOG.md).

## Version 0.11.0 (2026-06-10) {#version-0-11-0}

The backend module's default Configuration is now the single source of truth
for generic completion: `chat()`, `complete()` and `streamChat()` resolve
the active default database-backed configuration (provider adapter, model and
vault-backed credentials) when no provider is pinned, with per-call options
overriding the stored defaults. The extension-configuration `defaultProvider`
becomes a fallback for installations without a usable default configuration.

For the complete, itemised list see the canonical
[CHANGELOG.md](https://github.com/netresearch/t3x-nr-llm/blob/main/CHANGELOG.md).

### Added {#added}

-   **Default-configuration routing for generic completion.** Calls without a
    pinned provider route through the module-managed default
    `LlmConfiguration`; per-call `ChatOptions` override its stored
    defaults. `chatWithConfiguration()` / `completeWithConfiguration()` /
    `streamChatWithConfiguration()` accept an `$optionOverrides` array.

### Changed {#changed}

-   The extension-configuration `defaultProvider` is consulted only when no
    usable default configuration exists. Defaults without a model, or with
    backend-group access restrictions, are skipped — group-restricted
    configurations are never auto-applied without a backend-user context.

## Version 0.10.0 (2026-06-09) {#version-0-10-0}

The specialized AI services (DALL-E, FAL, Whisper, TTS, DeepL) now authenticate
through nr-vault's audited secure HTTP client instead of plaintext API keys,
bringing them in line with the database-backed providers (ADR-012, ADR-030).

For the complete, itemised list see the canonical
[CHANGELOG.md](https://github.com/netresearch/t3x-nr-llm/blob/main/CHANGELOG.md).

### Changed {#changed-1}

-   **Specialized services authenticate through nr-vault.** Each service stores
    an nr-vault secret *identifier* and authenticates via
    `$vault->http()->withAuthentication(...)`; the secret is resolved,
    injected, audited, and memory-scrubbed inside the vault and never surfaces
    in this extension. FAL (`Authorization: Key …`) and DeepL
    (`Authorization: DeepL-Auth-Key …`) use the nr-vault 0.8.0 `prefix`
    option. DeepL's Free/Pro routing stays automatic via a one-time, scrubbed
    `:fx` suffix check.

### Removed {#removed}

-   **Plaintext API keys for the specialized services.** Configuration keys are
    now nr-vault identifiers (`providers.openai.apiKeyIdentifier`,
    `image.fal.apiKeyIdentifier`, `translators.deepl.apiKeyIdentifier`).
    Requires `netresearch/nr-vault ^0.8.0`.

## Version 0.9.0 (2026-06-08) {#version-0-9-0}

This release migrates image generation to OpenAI's **gpt-image-\\**\* model
family (DALL·E-3 was retired by OpenAI), makes chat JSON mode actually request
JSON, and corrects the empty base-URL handling of the specialized services.

For the complete, itemised list see the canonical
[CHANGELOG.md](https://github.com/netresearch/t3x-nr-llm/blob/main/CHANGELOG.md).

### Added {#added-1}

-   **gpt-image-\\\* image generation.** `ImageGenerationOptions` accepts the
    `gpt-image-*` family by prefix and validates its size set
    (`1024x1024` / `1536x1024` / `1024x1536` / `auto`);
    `DallEImageService` maps the family to a shared capability profile and
    sends a minimal payload (gpt-image rejects `response_format` /
    `style` / `quality`), reading the returned `b64_json`.

### Fixed {#fixed}

-   **Chat JSON mode.** `OpenAiProvider` now maps `response_format=json` to
    OpenAI's `{"type":"json_object"}` so `CompletionService::completeJson()`
    receives valid JSON instead of prose.
-   **Empty base URL.** An empty ext_conf `baseUrl` for the DALL·E, FAL and
    TTS services now falls back to the provider default instead of being used as
    a scheme-less request URL.

## Version 0.8.0 (2026-06-02) {#version-0-8-0}

This release adds usage analytics and turns on real cost tracking, and
completes the provider **middleware pipeline** that now powers fallback,
pre-flight budget enforcement, usage accounting, and response caching
around every provider call. It also migrates the domain API to typed
value objects.

For the complete, itemised list see the canonical
[CHANGELOG.md](https://github.com/netresearch/t3x-nr-llm/blob/main/CHANGELOG.md).

### Added {#added-2}

-   **Usage Analytics dashboard.** A new *Admin Tools → LLM → Analytics*
    submodule with cost and request trends, breakdowns by provider, model,
    and service, KPI tiles, and per-user usage against each user's monthly
    budget. The Providers, Models, Configurations, and Tasks list views also
    gained per-row *Cost / Requests / Tokens (last 30 days)* columns. See
    [Usage analytics](https://docs.typo3.org/permalink/netresearch/nr-llm:administration-analytics@main).
-   **Real cost tracking.** Usage is now priced from the configured model
    rates (prompt/completion token split), so the *AI cost this month* widget
    and the dashboard show real figures instead of `0`. The
    `tx_nrllm_service_usage` table gained model and token-split columns
    plus per-task attribution.
-   **Automatic budget pre-flight.** Completion, embedding, translation, and
    vision requests are checked against the configured budget before the call
    is made.

### Changed {#changed-2}

-   The domain API moved to typed value objects (chat messages, tool specs,
    vision content, capability sets, provider options). The legacy
    string/array options accessors are deprecated in favour of the typed
    equivalents.
-   Requires `netresearch/nr-vault` `^0.6.0`.

### Breaking {#breaking}

-   The legacy `Model::CAPABILITY_*` class constants have been **removed**
    in favour of the `ModelCapability` backed enum (for example
    `ModelCapability::CHAT->value`). They had been deprecated since the enum
    was introduced.

## Version 0.7.0 (2026-04-22) {#version-0-7-0}

### Added {#added-3}

-   **Provider fallback chain.** `LlmConfiguration` can now list
    other configuration identifiers to retry against when the primary
    fails with a retryable error (connection / HTTP 5xx / 429 rate-
    limit). Non-retryable errors (4xx other than 429, configuration
    problems, unsupported feature) bubble up unchanged. Streaming is
    intentionally excluded from fallback because chunks cannot be
    replayed against a different provider. See [ADR-021: Provider Fallback Chain](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-021@main) and
    [Fallback chain](https://docs.typo3.org/permalink/netresearch/nr-llm:developer-fallback-chain@main).
-   **Attribute-based provider registration.** New
    `#[AsLlmProvider(priority: N)]` attribute. Providers bearing
    the attribute are automatically tagged and made public by
    `ProviderCompilerPass` at container compile time; no
    `services.yaml` edit required. Legacy yaml tagging still works
    for third-party providers and takes precedence when both
    mechanisms are present. See [ADR-022: Attribute-Based Provider Registration](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-022@main) and
    [Registering a provider](https://docs.typo3.org/permalink/netresearch/nr-llm:developer-provider-registration@main).
-   **Per-capability BE group permissions.** Every
    `ModelCapability` enum value is now a native TYPO3
    `customPermOptions` entry under the `nrllm` namespace. BE
    group editors see a checkbox per capability (chat, completion,
    embeddings, vision, streaming, tools, json_mode, audio). New
    `CapabilityPermissionService` resolves checks against the
    current BE user with admin short-circuit and CLI / frontend
    bypass. See [ADR-023: Native Backend Capability Permissions](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-023@main). These registrations and the check service
    were later removed ([ADR-117: Withdraw the backend capability permissions](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-117@main)); current backend access uses
    [Backend user permissions](https://docs.typo3.org/permalink/netresearch/nr-llm:administration-permissions@main).
-   **Dashboard widgets.** Two TYPO3 dashboard widgets sourced from
    `tx_nrllm_service_usage`: *AI cost this month*
    (`NumberWithIconWidget`) and *AI requests by provider (7d)*
    (`BarChartWidget`). Loaded conditionally from
    [`Configuration/Services.php`](https://docs.typo3.org/m/typo3/reference-coreapi/main/en-us/ExtensionArchitecture/FileStructure/Configuration/ServicesYaml.html#file-extension-configuration-services-php) only when
    [`typo3/cms-dashboard`](https://packagist.org/packages/typo3/cms-dashboard) is installed. See [ADR-024: Dashboard Widgets](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-024@main).
-   **Per-user AI budgets.** New `tx_nrllm_user_budget` table
    with six independent ceilings (requests / tokens / cost × daily /
    monthly). New `BudgetService::check()` aggregates usage on
    demand from `tx_nrllm_service_usage` — one DB roundtrip for
    both windows via conditional `SUM()`. Orthogonal to the
    existing per-configuration daily limits: both checks must pass.
    See [ADR-025: Per-User AI Budgets](https://docs.typo3.org/permalink/netresearch/nr-llm:adr-025@main) and [Per-user AI budgets](https://docs.typo3.org/permalink/netresearch/nr-llm:administration-user-budgets@main).

### Changed {#changed-3}

-   CI: mutation testing runs only on `push`, `merge_group` and
    `schedule` events. PR CI gets the fuzz suite + unit / functional
    / PHPStan / rector / code style; the  15 min mutation job is
    deferred because its per-PR signal is hard for authors to action
    locally.
-   CI: `.semgrepignore` added to exclude `Tests/`,
    `Build/Scripts/` and vendor directories from Opengrep SAST.
    Previously failing on legitimate `unlink()` fixture cleanup.
-   CI: fuzz workflow now invoked with
    `fuzz-testsuite: fuzzy` matching the phpunit.xml suite name.

## Version 0.6.0 (2026-03-24) {#version-0-6-0}

### Added {#added-4}

-   `DocumentCapableInterface`: providers can now advertise
    PDF/document support; `ChatCapabilitiesInterface` exposes
    this via `getProviderCapabilities()`.
-   Multimodal content arrays in `chatCompletion`: pass images,
    PDFs, and text blocks as structured content arrays alongside
    regular string messages.
-   Tool message conversion: `tool_result` blocks are now mapped
    correctly when assembling provider payloads.

### Changed {#changed-4}

-   Migrated CI infrastructure to `netresearch/typo3-ci-workflows`
    shared workflows (PHP tests, docs, E2E).
-   Replaced GrumPHP with CaptainHook for pre-commit hooks.

### Fixed {#fixed-1}

-   PHPStan baseline regenerated; `ignoreErrors` patterns
    broadened for deprecation and array function rules to handle
    phpstan-typo3 v2/v3 parameter name differences.
-   E2E tests stabilised: heading verification added, module
    overview landing page assertions updated.

## Version 0.5.0 (2026-03-09) {#version-0-5-0}

### Added {#added-5}

-   AI-powered full-chain task wizard: describe a task in
    plain language, AI generates task + configuration +
    model recommendation in one step.
-   AI-powered configuration wizard: generate
    configurations with system prompts, parameters,
    and model selection.
-   Custom TCA `ModelIdElement`: input field with
    "Fetch Models" button that populates from provider
    API, auto-fills capabilities and pricing.
-   `ModelConstraintsWizard`: field wizard that loads
    parameter constraint bounds per model.
-   Dashboard improvements: side-by-side wizard
    callouts, fixed headline from "LLM Providers"
    to "LLM Integration".
-   Task execution UI: collapsible prompt details, improved result display.
-   Enhanced model discovery: better Anthropic, Google, DeepSeek,
    Mistral support.
-   TER publish workflow.
-   Documentation: wizards guide with screenshots, tasks section,
    updated configuration reference.

### Changed {#changed-5}

-   Renamed `SafeCastTrait` extracted from duplicated helpers
    in `TaskController` and `WizardGeneratorService`.
-   SQL injection defense: regex whitelist validation for
    table/column names in `FetchRecordsRequest`
    and `LoadRecordDataRequest`.

### Fixed {#fixed-2}

-   Restored `method_exists()` guards for
    `setShortcutContext()` (TYPO3 v13 compatibility).
-   PHPUnit 12: replaced `createStub` with `createMock`
    to fix deprecation warnings.

## Version 0.4.8 (2026-03-07) {#version-0-4-8}

### Changed {#changed-6}

-   Rewritten introduction with value-oriented positioning.
-   Restructured README around value proposition and audience segments.
-   Updated package metadata with value-oriented descriptions.
-   Added integration guide for extension developers.

## Version 0.4.7 (2026-03-07) {#version-0-4-7}

### Added {#added-6}

-   Help page in the LLM backend module.
-   Setup wizard links on empty-state list pages.

### Fixed {#fixed-3}

-   Use canonical endpoint URLs for known providers in setup wizard.
-   Remove container class from backend module templates.

## Version 0.4.6 (2026-03-06) {#version-0-4-6}

### Fixed {#fixed-4}

-   Add Fluid-compatible `getHasApiKey()` getter for
    `{provider.hasApiKey}` in templates.

## Version 0.4.5 (2026-03-06) {#version-0-4-5}

### Fixed {#fixed-5}

-   Use `GET /v1/models` for Anthropic connection test.

## Version 0.4.4 (2026-03-06) {#version-0-4-4}

### Fixed {#fixed-6}

-   Use table-specific connection and simplify column checks.
-   Wrap test cleanup in try/finally and assert labelField.

## Version 0.4.3 (2026-03-06) {#version-0-4-3}

### Fixed {#fixed-7}

-   Handle tables without `uid` column in TCA utilities.
-   Remove hardcoded temperature from chat completions.

## Version 0.4.2 (2026-03-06) {#version-0-4-2}

### Fixed {#fixed-8}

-   Add `rootLevel` to provider, configuration, and model TCA definitions.

## Version 0.4.1 (2026-03-06) {#version-0-4-1}

### Fixed {#fixed-9}

-   Use `max_completion_tokens` instead of `max_tokens`
    for OpenAI chat completions.

## Version 0.4.0 (2026-03-06) {#version-0-4-0}

### Breaking {#breaking-1}

-   Prevent plaintext API key storage via setup wizard;
    keys now require vault encryption.

### Fixed {#fixed-10}

-   Cast ExtensionConfiguration timeout values to integer.

### Changed {#changed-7}

-   Use Symfony `Uuid::v7()` instead of manual UUID generation.

## Version 0.3.2 (2026-03-04) {#version-0-3-2}

### Added {#added-7}

-   Extract thinking blocks from LLM responses (`<think>` tag support).

### Fixed {#fixed-11}

-   Preserve newlines in `extractThinkingBlocks`.
-   Restrict CI push trigger to main branch only.
-   Add `merge_group` trigger to CI workflow.

## Version 0.3.1 (2026-03-02) {#version-0-3-1}

### Fixed {#fixed-12}

-   Add Overview submodule for TYPO3 v13 module overview compatibility.

## Version 0.3.0 (2026-03-01) {#version-0-3-0}

### Added {#added-8}

-   Expose `chatWithConfiguration` and
    `streamChatWithConfiguration` on
    `LlmServiceManagerInterface`.

### Fixed {#fixed-13}

-   Use integer values for `f:be.infobox` state attribute
    for TYPO3 v13 compatibility.
-   Explicitly enable fuzz and mutation tests.

## Version 0.2.2 (2026-03-01) {#version-0-2-2}

### Fixed {#fixed-14}

-   Use `tools` parent for TYPO3 v13 module compatibility.

### Changed {#changed-8}

-   Consolidate caller workflows into 4 grouped files.
-   Fix documentation issues found by analysis.

## Version 0.2.1 (2026-02-28) {#version-0-2-1}

### Changed {#changed-9}

-   Require `netresearch/nr-vault` ^0.4.0 for API key encryption.

## Version 0.2.0 (2026-02-28) {#version-0-2-0}

### Added {#added-9}

-   PHP 8.2+ and TYPO3 v13.4+ compatibility.
-   TYPO3 v13.4 ddev install command.
-   Coverage uploads and fuzz/mutation CI workflow.
-   Unit tests for enums, WizardResult DTO, providers,
    services, and specialized classes.
-   Coverage tests for PromptTemplateService and TranslationService.

### Changed {#changed-10}

-   Moved `phpunit.xml` and `phpstan-baseline.neon` into `Build/` directory.
-   Expanded CI matrix to PHP 8.2-8.5 and TYPO3 v13.4/v14.
-   Replaced TYPO3 v14-only APIs with v13-compatible equivalents.
-   Narrowed testing-framework to ^9.0 for PHPUnit 12 compatibility.
-   Removed dead ProviderRegistry class and orphaned phpstan baseline file.
-   Removed 55 dead translation keys.
-   Harmonized composer script naming to `ci:test:php:*` convention.
-   Migrated CI to centralized workflows.
-   Added SPDX copyright and license headers.
-   Replaced generic emails with GitHub references.

### Fixed {#fixed-15}

-   Resolved CI failures for PHP 8.2 and TYPO3 v13 compatibility.
-   Resolved PHPStan failures for dual TYPO3 v13/v14 support.
-   Fixed PHPUnit deprecation warnings.
-   Used `CoversNothing` for excluded exception and enum test classes.
-   Localized user-facing hardcoded strings in controllers.
-   Disabled functional tests in CI (environment-specific).
-   Fixed direct `php-cs-fixer` call in `ci:test:php:cgl` script.

## Version 0.1.2 (2026-01-11) {#version-0-1-2}

### Fixed {#fixed-16}

-   Fixed CI: use correct org secret name for TER token.
-   Simplified TER upload workflow.

## Version 0.1.1 (2026-01-11) {#version-0-1-1}

### Fixed {#fixed-17}

-   Fixed CI: create zip archive for TER upload.

## Version 0.1.0 (2026-01-11) {#version-0-1-0}

Initial release of the TYPO3 LLM Extension.

### Added {#added-10}

**Core Features**

-   Multi-provider support (OpenAI, Anthropic Claude,
    Google Gemini, Ollama, OpenRouter, Mistral, Groq).
-   Unified API via `LlmServiceManager`.
-   Provider abstraction layer with capability interfaces.
-   Typed response objects (`CompletionResponse`, `EmbeddingResponse`).
-   Three-tier configuration architecture (Providers, Models, Configurations).
-   Encrypted API key storage using sodium_crypto_secretbox.

**Feature Services**

-   `CompletionService`: Text completion with format
    control (JSON, Markdown).
-   `EmbeddingService`: Vector generation with caching
    and similarity calculations.
-   `VisionService`: Image analysis with alt-text,
    title, description generation.
-   `TranslationService`: Translation with formality
    control and glossary support.
-   `PromptTemplateService`: Centralized prompt
    management with database-driven templates.

**Specialized Services**

-   Image generation (DALL-E).
-   Text-to-speech (TTS) and speech transcription (Whisper).
-   DeepL translation integration.

**Provider Capabilities**

-   Chat completions across all providers.
-   Embeddings (OpenAI, Gemini).
-   Vision/image analysis (all providers).
-   Streaming responses (all providers).
-   Tool/function calling (all providers).

**Infrastructure**

-   TYPO3 caching framework integration.
-   Backend module for provider management and testing.
-   Prompt template management with versioning and performance tracking.
-   Comprehensive exception hierarchy.
-   Type-safe enums and DTOs for domain constants.

**Developer Experience**

-   Option objects with factory presets (`ChatOptions`).
-   Full backwards compatibility with array options.
-   Extensive PHPDoc documentation.
-   Type-safe method signatures.

**Security**

-   Enterprise readiness security workflows and supply chain controls.
-   SLSA Level 3 provenance, Cosign signatures, and SBOM generation.
-   OpenSSF Scorecard and Best Practices compliance.

**Testing**

-   Comprehensive unit and integration tests.
-   E2E testing with Playwright.
-   Property-based (fuzz) testing support.

## Upgrade Guides {#upgrade-guides}

### Upgrading a pre-1.0 installation {#upgrading-a-pre-1-0-installation}

Do not remove and reinstall the package as a normal upgrade. Back up the
installation, review the releases between the installed and target version,
and update its Composer constraint deliberately. See
[From previous versions](https://docs.typo3.org/permalink/netresearch/nr-llm:installation-upgrading-previous@main) for package setup, schema checks and
upgrade wizards. Review configuration and integrations before reopening
editor access.

The current package line is 0.43. For that line, for example:

**Select the 0.43 release line**

```bash
composer require netresearch/nr-llm:^0.43 --with-all-dependencies
```

This does not select a future 0.44 release. Choose its constraint explicitly
when following that release's migration notes.

## Breaking changes policy {#breaking-changes}

During 0.x, minor releases may contain incompatible changes. Read the
breaking-change and migration notes for every crossed minor line. Patch
releases are intended for compatible fixes; the public `@api` contract is
narrower than every class present in the package.

The notice period that starts at 1.0 and the treatment of consumer interfaces
are described in [API stability](https://docs.typo3.org/permalink/netresearch/nr-llm:api-stability@main) and [Deprecation and removal policy](https://docs.typo3.org/permalink/netresearch/nr-llm:api-deprecation@main). Those pages
are the policy reference; this release overview adds no broader guarantee.

## Deprecation policy {#deprecation-policy}

See [Deprecation and removal policy](https://docs.typo3.org/permalink/netresearch/nr-llm:api-deprecation@main) for the current notice period, replacement inventory
and the distinction between 0.x and the policy from 1.0.