Bug: /new produces empty prompt → Anthropic API 400 "messages: at least one message is required"

Summary

When /new (or /reset) is invoked on an agent, there is a race condition between the active-memory subsystem and the bare-reset prompt injection that can cause runEmbeddedPiAgent to be called with an empty prompt AND an empty messages[] array. This produces a 400 invalid_request_error from the Anthropic API and leaves the session stuck in "processing" state. Failover to other models (e.g. gpt-5.5) also fails because the same empty body is forwarded.

Symptoms

• /new on agent main results in:
  • Stuck "processing" sessions (UI shows pending dot indefinitely)
  • Trace events with data.prompt: "" and messages: []
  • 400 invalid_request_error from Anthropic API: messages: at least one message is required
  • Failover model (e.g. openai-codex/gpt-5.5) also receives empty input → secondary error
• Subsequent session entries record [assistant turn failed before producing content] with errorMessage: "One of \"input\" or \"previous_response_id\" or 'prompt' or 'conversation_id' must be provided." (OpenAI Responses adapter)

Reproduction

1. Configure agent main with active-memory plugin enabled.
2. Trigger /new immediately after another /new or /reset, OR while active-memory is mid-rebuild.
3. Observe: session is created, but assistant turn fails before producing content. Subsequent user messages may also fail until the session is hard-reset.

Conditions that increase frequency:

• Multiple agents (main + secondary + tertiary + fdx-assistant) sharing active-memory subsystem.
• Large workspaces where active-memory takes time to rebuild.
• Quick repeated /new /reset from the user.

Affected files

• dist/pi-embedded-*.js — runEmbeddedPiAgent() entry point. Receives empty params.prompt without guard.
• dist/selection-*.js — pre-submission flow. Builds payload with empty messages: [] when prompt is empty AND activeSession.messages has no non-empty user message.

Root cause

The pre-submission guard hasPromptSubmissionContent({ prompt, messages, imageCount }) correctly detects empty content and sets skipPromptSubmission = true, but in some race scenarios (active-memory mid-rebuild + bare-reset), it is reached AFTER the LLM dispatch has already started with the empty payload, OR activeSession.messages still contains stale empty entries from the previous session that pass the existence check but have empty text content.

Workaround / patch (currently applied locally)

Two-layer defensive guard:

Layer 1: pi-embedded-*.js entry guard

Inject a canonical bare-session-reset prompt at the entry of runEmbeddedPiAgent when params.prompt is empty:

async function runEmbeddedPiAgent(params) {
  // Guard: prevent empty prompt + empty messages from reaching LLM dispatch
  if (typeof params.prompt !== "string" || params.prompt.trim().length === 0) {
    params = {
      ...params,
      prompt: "A new session was started via /new or /reset. Execute your Session Startup sequence now - read the required files before responding to the user. If BOOTSTRAP.md exists in the provided Project Context, read it and follow its instructions first. Then greet the user in your configured persona, if one is provided. Be yourself - use your defined voice, mannerisms, and mood. Keep it to 1-3 sentences and ask what they want to do. Do not mention internal steps, files, tools, or reasoning."
    };
  }
  // ... existing code
}

Layer 2: selection-*.js pre-submission guard

Before hasPromptSubmissionContent check, force-inject a minimal prompt when:

• promptSubmission.prompt is empty
• imageResult.images.length === 0
• No non-empty user message exists in activeSession.messages

if (!skipPromptSubmission && !promptSubmission.runtimeOnly &&
    (!promptSubmission.prompt || promptSubmission.prompt.trim().length === 0) &&
    imageResult.images.length === 0) {
  const hasNonEmptyUserMsg = (activeSession.messages || []).some((m) => {
    if (!m || m.role !== 'user') return false;
    const c = m.content;
    if (typeof c === 'string') return c.trim().length > 0;
    if (Array.isArray(c)) return c.some((p) => p && p.type === 'text' && typeof p.text === 'string' && p.text.trim().length > 0);
    return false;
  });
  if (!hasNonEmptyUserMsg) {
    promptSubmission.prompt = "A new session was started. Greet the user briefly in your configured persona and ask what they want to do.";
  }
}

Suggested upstream fix

1. Strict guard at LLM dispatch boundary: never send a request with empty prompt AND empty messages AND no images — return early with structured error or auto-inject a safe greeting prompt.
2. Active-memory race fix: ensure runEmbeddedPiAgent is invoked only AFTER active-memory has materialized the session-start prompt, or pass a sentinel value.
3. Better logging: log a WARN when an empty-payload dispatch is intercepted, so users can see the race instead of getting an opaque 400.

Environment

• OCPlatform: 2026.4.26-93430cbb9781 (also reproduced on 2026.4.25)
• Node: v25.9.0
• OS: WSL2 (Linux 6.6.87.2-microsoft-standard-WSL2 x64)
• Affected adapters: anthropic (Anthropic API), openai-codex/gpt-5.5 (OpenAI Responses)
• Active plugins: active-memory, custom hooks (session-flush, boot-md, bootstrap-extra-files, command-logger, session-memory: false)

Local patch status

The local patch script (patches/empty-prompt-guard.sh) has been applied since 2026-04-29 and resolves the issue. Re-application required after each openclaw upgrade because patches modify hashed dist files (pi-embedded-*.js, selection-*.js).