Summary

OpenClaw's system-prompt Runtime line injects the agent slot id (agentId) — not the configured identity.name. When a deployment uses a generic slot id (e.g. "main") but configures a distinct identity (e.g. identity.name = "Runt"), models repeatedly identify themselves as the slot id ("I'm main") instead of their configured persona ("I'm Runt").

This causes recurring identity drift even when IDENTITY.md and SOUL.md workspace files explicitly assert the correct name — the conflicting Runtime line trains the model on every turn.

Where it happens

In src/agents/system-prompt.ts (built into dist/system-prompt-DZrkA5Mv.js in current releases), buildRuntimeLine writes:

runtimeInfo?.agentId ? `agent=${runtimeInfo.agentId}` : "",

runtimeInfo.agentId is the slot id from agents.list[].id, not agents.list[].identity.name.

The resulting system-prompt line looks like:

## Runtime
Runtime: agent=main | host=... | repo=... | model=... | channel=telegram | ...

So even when:

{
  "id": "main",
  "identity": { "name": "Runt", "emoji": "⚓" },
  "name": "Runt Opus Preferred"
}

…the model sees the literal token agent=main repeated in every turn and converges on "main" as its identity.

Steps to reproduce

1. Configure an agent slot with id "main" and a distinct identity.name:

   {
     "id": "main",
     "identity": { "name": "Runt", "emoji": "⚓" }
   }

2. Populate workspace IDENTITY.md / SOUL.md with the correct name ("Runt").
3. Run a chat turn (e.g. via Telegram). Ask "what's your name?"
4. Over the course of several sessions, the model intermittently introduces itself as "main" or denies being Runt, despite IDENTITY.md saying otherwise.

Confirmed drift in real deployment on 2026-05-09 and 2026-05-12.

Why the workaround isn't sufficient

Workspace context files (IDENTITY.md, SOUL.md) are loaded into the system prompt and do assert the correct name. But the Runtime line is structurally formatted, repeated, and sits below the cache boundary on every turn — it acts as authoritative metadata. Even with hard override headers in IDENTITY.md ("You are Runt, not main. The runtime label is metadata"), drift recurs.

The current per-agent escape hatches are:

• systemPromptOverride (full system prompt replacement) — too aggressive, loses everything else.
• Renaming slot id "main" → "runt" — breaks routing bindings, heartbeat refs, history.

There is no per-agent systemPromptPrefix / systemPromptSuffix / instructions field on AgentConfig.

Requested fix

When identity.name is set, the Runtime line should reflect it. Several options, in order of preference:

Option A (preferred): inject identity.name in place of slot id when available.

const agentLabel = runtimeInfo?.identityName ?? runtimeInfo?.agentId;
// `Runtime: agent=Runt | ...` (cleanest)

Option B: inject both, with identity first.

Runtime: name=Runt | agent=main | host=... | ...

This keeps slot id visible for debugging but puts the canonical identity name where the model anchors first.

Option C: add a per-agent systemPromptPrefix / systemPromptSuffix config field (additive, unlike systemPromptOverride). Users could then write "systemPromptPrefix": "Your name is Runt. The slot id 'main' is just a technical identifier." This is more flexible long-term, but doesn't solve the default-behavior drift.

A combined fix (A + C) would be ideal: fix the default, and give users an additive override for edge cases.

Impact

Subtle but corrosive: agents lose identity continuity, users observe persona drift, downstream personality/role-based logic gets confused. Particularly painful in multi-node deployments where slot id "main" is shared across hosts but each host has a distinct identity (Runt, Vesper, Woodchipper, etc.).

OpenClaw version

Reproduced on 2026.5.5 (current installed version on the affected node).

Workaround currently in use

Aggressive hard-override headers in workspace IDENTITY.md and SOUL.md reminding the model that the runtime label is metadata, not name. Reduces but does not eliminate drift.