Composing Context

Composing turn-local context as a synthetic user-message preface

This skill documents the synthetic user-message preface pattern used by Friday's chat supervisor and FSM type: llm actions to surface turn-local retrieved content (artifacts, narrative memory, temporal facts, future on-demand retrieval) to the model.

Why a synthetic preface and not the system prompt

Per-turn data does not belong in the system prompt. Mixing turn-local bytes into the system prompt invalidates the cacheable prefix on every turn, defeats Anthropic's 4-breakpoint cache layout (blocks 1/2/3 + top-level), and pulls org-private bytes into a cache scope that may be shared across tenants.

The synthetic preface lives in the first user message of the conversation, wrapped in <retrieved_content> envelopes. The system prompt stays byte-stable across weeks (block 1), workspace life (block 2), and session (block 3); turn-local variation rides in the messages array where it doesn't poison the cacheable prefix.

The helper: composePreface

import {
  composePreface,
  type PrefaceEntry,
} from "@atlas/core/agent-context/compose-preface";

const entries: PrefaceEntry[] = [
  {
    source: "user-authored",
    origin: "memory:workspace-stores",
    body: memoryBlocks.join("\n\n"),
    fetched_at: new Date().toISOString(),
  },
  {
    source: "user-authored",
    origin: "artifacts:session",
    body: artifactBlocks.join("\n\n"),
    fetched_at: new Date().toISOString(),
  },
];

const preface = composePreface(entries);
// "" when entries is empty — concatenate without conditional checks.

Field semantics:

The helper is positionally indifferent — the caller decides whether to prepend the result as a synthetic user message, append it, or place it elsewhere. The string is the deliverable; positioning is the caller's choice.

The body is defanged before rendering: any literal </retrieved_content> close tag inside the body is rewritten to <\/retrieved_content> so a hostile payload cannot escape the envelope and land in the instructions frame. This mirrors wrapRetrieved in @atlas/llm.

Three call patterns

1. Pre-load: chat supervisor builds Block 4 from current state

The chat supervisor pulls memory + artifacts + temporal facts at the start of each turn, composes the preface, and prepends it as a synthetic user message at message position 0. The preface is ephemeral — it is not persisted to ChatStorage, so it does not accumulate in the conversation across turns.

See packages/system/agents/workspace-chat/workspace-chat.agent.ts (block4Entries / block4Preface / messagesWithBlock4Preface).

2. FSM type: llm action: same shape, prepended to action prompt

The FSM action runtime mirrors the chat pattern: pull narrative memory (via composeMemoryBlocks) and session-bound artifacts (via composeArtifactEntries), compose with composePreface, and prepend to the action's contextPrompt. The structured-entry shape from composeArtifactEntries is designed to drop straight into composePreface without intermediate string mangling.

See packages/fsm-engine/fsm-engine.ts (the case "llm" action prompt assembly).

3. On-demand mid-turn retrieval (forward-looking)

A future tool (e.g. web_fetch, parse_artifact, mid-turn semantic search) can hand the freshly-retrieved entry list to composePreface and inject it inline using whatever shape the LLM provider supports for mid-stream context — the helper is positionally indifferent and shape-agnostic, so the new caller drops in without reshaping.

Rules for callers

1. Build a list of PrefaceEntry records, not pre-rendered XML strings. The helper owns the rendering. Callers that already have pre-rendered envelopes (e.g. composeArtifactBlocks's legacy return shape) wrap them in a single PrefaceEntry rather than reformatting each one — see Block 4 in the chat supervisor for the canonical example.

2. Set source to a trust tier or content key, not a generic label. The model's <retrieved_content_hygiene> rule keys off provenance to decide trust (user-authored / system-config are trusted; external is not). Use the conventions in @atlas/llm/retrieved-content.ts:ProvenanceSource when the source is a trust tier.

3. Set origin to a parseable host:scope identifier. The model does not parse origin literally, but operators inspecting captured transcripts do. Conventions: workspace:<wsId>/session:<sId>, memory:<store>, web:<url>, fsm:<jobId>:input.

4. Always set fetched_at when the body's freshness matters. Omit only for static content (e.g. system-config). Use the same timestamp for all entries assembled in one turn so operators can correlate them.

5. Do not nest <retrieved_content> envelopes deliberately, except in the legacy chat-Block-4 wrap-of-wrapped pattern that exists for per-section grouping. The defang protects against accidental nesting; deliberate nesting creates confusing transcripts.

6. Never persist the preface as part of a stored UI message. It is an ephemeral runtime artifact rebuilt from current state each turn. Persisting it accumulates stale snapshots in the conversation and defeats the point of having retrieval-gated content.