Attribution format

Every transcript handed to a CLI agent carries an explicit, structured header. The format is deterministic and readable; CLI agents digest it without prompt-engineering tricks.

Anatomy

[SESSION CONTEXT]
Room: #auth-team (project: saga-ai)
Participants: Andi (backend lead), Budi (frontend), Carol
You are: @claude (Claude Code CLI agent)
Time: 2026-04-21 14:02 Asia/Jakarta

[RECALL — from long-term memory]
- Andi previously stated a preference for 30 min JWT expiry
  (fact, 2026-04-15, DM)
- Project saga-ai uses NestJS + Prisma (project memory)

[LANDMARKS — pinned decisions]
- [2026-04-10] Team agreed on refresh-token rotation (by Budi)

[EPISODIC SUMMARY — last session]
Yesterday Andi & Budi discussed a race-condition bug in the auth
middleware; Andi is investigating, not yet resolved.

[RAW TRANSCRIPT — most recent 40 messages]
[14:02 | Andi] @claude why is the token expiry only 15 minutes?
[14:02 | Budi] I'd suggest 30 minutes
[14:03 | Andi] yes, users complained

[CURRENT INVOCATION]
Respond to Andi and Budi in room #auth-team.

Why each section

Section	Purpose
SESSION CONTEXT	Tells the bot who it is, where it is, who it's talking to.
RECALL	Mem0 hits relevant to the last user message.
LANDMARKS	Pinned decisions / specs / important code that survives compaction.
EPISODIC SUMMARY	Compacted older context, replacing raw messages outside the window.
RAW TRANSCRIPT	Most recent N messages verbatim, with [time | sender] lines.
CURRENT INVOCATION	The one thing the bot is being asked to do right now.

Why explicit attribution

In a multi-human + multi-AI room, the model must know:

• Which user it's responding to (so its reply addresses them by name).
• Which other users are participants (so it can defer or pull them in).
• Which user's profile each fact came from (so it doesn't conflate preferences).
• That it is one of multiple bots (so it doesn't impersonate @hermes).
• The DM-vs-room context (so it doesn't leak DM memory).

Without an attribution header, the model has to guess from chat formatting alone — and CLI agents weren't trained on chat formats.

Where it's built

The header is assembled by services/composer/src/prompt.ts. The function pulls each section from the corresponding memory layer (see Memory layers), formats it, and joins with the raw transcript.

The full prompt then goes to the runtime, which spawns the right CLI adapter. Each adapter strips ANSI/TUI noise from the streamed reply before posting to Rocket.Chat.

Implication for personas

Don't write personas that say "You will receive [SESSION CONTEXT] and [RECALL] sections...". The bot already knows the format from the data shape. Personas should focus on role, scope, voice, and peer relationships — not on the prompt layout.