Tagged Prompt Snippet Library

Context

Consuming extensions — first nr_repurpose — assemble prompts from recurring building blocks: a persona, a tone of voice, a target audience, an image style, a layout instruction. Editors want to manage these fragments centrally, once, instead of re-typing them into every extension's own configuration.

The existing PromptTemplate entity does not fit this need. It is a heavyweight complete prompt: it binds a feature, carries model parameters (temperature, max tokens, top-p), supports versioning with parent/variant relations, and tracks usage performance. A persona like "You are Nova, a friendly expert." has none of these concerns — it is a fragment that only becomes a prompt when a consumer composes it with its own instructions. Forcing fragments into PromptTemplate would either bloat every fragment record with irrelevant model fields or fork the template semantics depending on a "fragment" flag.

A second question is how consumers select fragments. A fixed category enum (like Task categories) would require an nr-llm release every time a consuming extension introduces a new fragment kind, which contradicts the goal of nr-llm being a shared foundation that consumers extend without touching it.

Decision

Introduce a separate, lightweight PromptSnippet entity (table tx_nrllm_promptsnippet) next to — not on top of — PromptTemplate:

1. Fragments, not templates. A snippet is identifier + name + description + fragment text. No model parameters, no versioning, no performance tracking. PromptTemplate stays untouched.

2. Free-form CSV tags instead of a category enum. Snippets carry a comma-separated tags field. Consumers query PromptSnippetRepository::findActiveByTag(), which matches tags as exact, case-insensitive tokens — style never matches lifestyle. The tag vocabulary is a convention between editors and consumers (established so far: audience, tone_of_voice, persona, layout, style), documented in the TCA field description and the administration guide. New fragment kinds need no nr-llm release.

3. JSON metadata side-channel. An optional metadata JSON object carries consumer-specific settings (e.g. {"voice": "nova"} on persona snippets so speech features can pick a matching TTS voice). getMetadataArray() returns [] for empty or invalid JSON — bad editor input must never break a consumer.

4. Composition stays in nr-llm. PromptSnippetComposer renders an ordered label-to-snippet map into labeled prompt blocks (LABEL: + fragment text, blank-line separated), so all consumers produce uniformly structured prompt sections.

5. Editing via FormEngine. The backend module gets a "Snippets" list following the established Providers/Models/Tasks pattern; create/edit links into FormEngine, no custom forms.

Amendment (2026-08-09): a configuration selects snippets by tag

Until this amendment the library had no reader in a production prompt. The only consumers were RunAugmentation::$forcedSnippets and the codec that rehydrates it, and RunAugmentation is constructed nowhere but the tool playground — so outside the playground the whole snippet system was inert, and the "consuming extension" of point 2 above was the only way a snippet ever reached a model.

ADR-139 names a tagged snippet as the supported way to attach editorial context to a request. That needs a selection an operator can make. tx_nrllm_configuration.snippet_tags is it: a CSV of tags whose itemsProcFunc lists the tags the snippet records actually carry, so the vocabulary stays consumer-owned and a new fragment kind still needs no nr_llm release.

The selection is composed into the effective system prompt, not into extra system messages. ConfigurationSnippetResolver appends the labeled blocks to the system_prompt that ConfigurationCallPlanner::callOptions() has merged, behind the configuration's own prompt. One insertion point reaches chat, completion, streaming and the agent loop, because every configuration-driven entry point on LlmServiceManager builds its options there.

The alternative — rendering each snippet as its own leading system message, the shape the playground uses — was rejected. It only works in the playground because the loop bakes the configuration's system prompt ahead of the snippet messages first. Anywhere else a snippet system message would be the first system message in the list, which is exactly the condition under which MessageShaper::applySystemPrompt() leaves the list alone: the configuration's own system prompt would be dropped from the run, silently and with no error. The characterisation tests added with ADR-139 pin that behaviour.

These details follow from the code:

• Dedup is by snippet identifier. PromptSnippetRepository::findActiveByTag() loads all active snippets and filters in PHP, so a snippet carrying two selected tags comes back from both lookups and would otherwise be composed twice.

• An unknown tag is empty, not an error — the free-tag model has no referential integrity by design, so a typo degrades to "no snippets".

• The tool playground reads the same composed value. Its bake site in ToolLoopService::assemble() goes through the same resolver, so a previewed transcript is the transcript a live run sends. A forced snippet the configuration already selects by tag is composed once, not twice: the resolver's identifier dedup never sees the forced list, so the bake site skips a forced block its composed prompt already contains.

• Hiding a snippet takes it out of every configuration. The repository ignores enable fields on purpose — the backend module lists hidden records — so ConfigurationSnippetResolver is where a hidden record is dropped. is_active remains the operational switch; hidden is the editorial one, and both now keep a snippet out of a production prompt. This is the one place in the extension where hidden decides a runtime outcome, because it is the one place where an editor's list-module action would otherwise keep shipping text to a provider.

• The composed block counts against the context window. The prompt is prepended after ContextWindowManagerInterface::fit() has run, so both callers (ToolLoopService and ConversationService) hand the composed prompt to fit() instead of letting it re-read LlmConfiguration::getSystemPrompt() — otherwise the budget of ADR-107 is short by exactly the snippet block.

What this does not change: a caller-supplied system message still suppresses the configuration's system prompt — and with it the snippet block — because per-call precedence is decided before this composition is read. That is the pre-existing rule, not a new one.

Consequences

• Editors manage personas, tones, audiences, styles, and layouts once, centrally; every consuming extension reads the same library.

• The free-tag model keeps nr-llm release-independent from consumer vocabulary — at the cost of no referential integrity: a typo in a tag silently yields an empty query result. The documented convention and the tag badges in the list view mitigate this.

• Token matching is implemented over the CSV field in PHP, not SQL LIKE, guaranteeing exact-token semantics on every database platform. The snippet library is small (tens of records), so loading active snippets for tag filtering is not a performance concern.

• Two prompt-related entities now coexist. The split is intentional (template = complete prompt, snippet = fragment) and documented here, in the administration guide, and in both entities' PHPDoc.

• Since the 2026-08-09 amendment an operator can attach snippets to a configuration without writing any code, and every request made with that configuration carries them — including requests from consuming extensions that know nothing about snippets.