Agent Presentation

---
name: agent-presentation
description: Shared, runtime-free taxonomy for presenting OwlMeans Viable agent output. Use when classifying LLM thinking/history messages, adding semantic structured-output cards, or changing their server-to-browser event contract.
---
<!-- AUTO-GENERATED — do not edit. Regenerate via sync-agent-meta. -->

# Agent output presentation

**Install:** `"@owlmeans/viable-common": "^0.0.46"` in `dependencies`

`@owlmeans/viable-common`'s `agent/presentation.ts` owns the browser-safe vocabulary for agent
output. Keep the server and every client on `agentPresentationHelper.classifyAgentMessage()`; consumers may refine a
provisional classification as streamed text becomes parseable, but must not invent a competing
taxonomy.

## Contract

- The top-level categories are code, unified diff, structured output, Markdown and plain text.
  Code and diffs carry one of React TSX, CSS, TypeScript, JSON or other; structured output carries
  a semantic kind; Markdown carries a document kind; every presentation carries a specialist role.
- Classify from the combined source: agent, helper, action, explicit output type, text and parsed
  value. Helper/agent attribution selects the specialist family; content shape alone is not enough.
- A transport may send `presentation` on thinking start and stop events. It is optional for rolling
  compatibility; the final server hint is based on complete output and the browser refines it while
  it streams. Keep legacy `outputType` intact.
- Unwrap a single LangChain tool call's `args` before semantic classification
  (`agentPresentationHelper.agentToolCallArguments`). Known schemas render as cards; unknown objects and arrays remain
  generic structured output, never raw JSON.
- `structuredKindOf`'s shape checks are ORDER-SENSITIVE where one payload could match more than
  one: a scaffold plan carries its own `stories` array, so `ScaffoldPlan` (`identity` + `guestHome`
  + `areas` + `stories`) is checked before the bare `StoryPlan` (`stories` alone) check catches it
  first. `RuntimeDecision` matches on `jobs` + `agents` + `kv` alone — NOT `actor`/`worker`, which
  the live model call never sends (those two are added only once a decision is persisted onto a
  `StoryDesign`); requiring them missed every runtime-decision call the model actually makes.
- `outputType === 'tool_calls'` does not automatically mean `AgentStructuredKind.ToolCalls`: a
  SINGLE pinned-schema call (`choose-files-for-fix`, `plan-scaffold`, …) still renders as its own
  schema card. It reads as `ToolCalls` when any of these holds: more than one call arrived in the
  same turn, the action belongs to an unconstrained tool-calling loop (`agentLoopActions` —
  `coding-agent-ask`, `arbitrary-modification`, `fix-agent`, `architect`, `declaration-lookup`,
  `source-extract`), or the single call's shape only reaches a `Generic*` kind. Without the
  action check, a free-flight `read_sources({ files })` call reads as an `ArtifactSelection` card
  instead of a tool call, because its args happen to shape-match that schema.
- `agentPresentationHelper.isAgentMessageHidden()` owns utility-run suppression. The source extractor's `source-extract`
  range-selection calls are internal context reduction and never render in thinking or history.

## Consumer rules

- Treat Markdown as a compact draft, not as a full document. Keep raw HTML disabled.
- Use one token renderer for code and diff code portions; infer a diff language from its changed path.
- Plain text remains plain. Empty live output is an activity state, not a fabricated model response.
- Labels for known structured fields belong to the consuming app's i18n resources. Open model keys
  should be humanized rather than passed through an unbounded translation namespace.
- This package classifies WHICH kind and category a message is; it says nothing about how a
  consumer renders one. `product-viable`'s manager-web renders every `AgentStructuredKind` as a
  flat typed result card and captures per-tool-call outcomes as a platform-side thinking event
  (`AgentThinkingType.ToolResult`, in `sources/common`, not this package) — see that repo's
  `agent-output-cards` skill before changing either side of this contract.

## External docs

- https://github.com/remarkjs/react-markdown/blob/main/readme.md — safe-by-default Markdown with custom React components.
- https://github.com/remarkjs/remark-gfm — GFM tables, task lists, autolinks and strikethrough plugin.
- https://github.com/FormidableLabs/prism-react-renderer — React token rendering for compact source views.
- https://fontsource.org/fonts/jetbrains-mono/install — self-hosted variable JetBrains Mono for Vite apps.