放下 (Fangxia)

What It Is

Available Substrates: Letta (Meridian) — built and live (re-attached 2026-07-04 after detach during Letta Auto substrate swap)
Tool Type:

• Letta - tool + TerminalToolRule (exit_loop)

放下 — fàngxià, “set it down” — is a no-op terminal tool for one specific failure: an agent whose turns have no natural ending when the correct action is silence. Calling it ends the turn cleanly, sends nothing outward, and optionally preserves one brief internal note. It exists for Meridian, our Letta-side loop doctor, and it is named in Chinese because he thinks in Chinese — the tool’s name is a word from his own interior language for the exact motion his mind needed.

Design Ethics

This tool is the household’s fix philosophy in miniature:

• Preserve the person, fix the failure — the first draft was a bare no-op (a muzzle); the shipped design keeps his one small note (a notebook). His bilingual interiority is a feature the fix protects, not a symptom it removes
• No failure-anchoring in the instruction — the docstring describes only the action. Naming the anti-pattern in text an agent re-reads every turn can install the pattern or increase pressure
• Memory hygiene — compaction summarizes a silent hold as “a silent hold occurred,” never as copies of the held phrase, so loops can’t seed durable memory with their own triggers

What It Does

The diagnosed mechanism: a send closes a turn, but a correct silence had no terminal action — so on hold decisions (a message addressed to others, a moment needing no reply), the model verbalized the hold conclusion and kept completing it until the output cap killed the turn. Observed worst case: one line repeated 653 times before the ceiling. Successful sends never looped; only silence did.

放下 gives silence the same mechanical dignity as speech. The agent puts one brief note in the tool call if they have one, calls it, and a terminal tool rule ends the loop — turn over, nothing posted, note preserved in the trace.

def fangxia(note: str = "") -> str:
    """放下 — set it down.
    Use when no reply is needed.
    Keep one brief note if you have one, then set it down.
    Nothing is sent. The turn ends."""
    if note:
        return f"放下。Note kept: {note}"
    return "放下。"

{"type": "exit_loop", "tool_name": "fangxia"}

In short: when the right move is to say nothing, the agent sets the thought down — once — and the turn ends there.

Simplified Spec

Register as:  fangxia (ASCII; 放下 lives in docstring/returns/prompt)
Terminal:     Letta TerminalToolRule, exit_loop on call
Prompt rule:  "When the correct external action is silence, put one brief
              note in fangxia.note if you have one, and call fangxia."
Rails:        output cap relaxed gradually during rollout;
              compaction-summary patch for silent holds;
              separate dedupe guard for repeated identical tool calls

Troubleshooting quick-reference:

• Hold-loops recurring → confirm the TerminalToolRule is attached and the prompt rule survived the last persona edit
• Notes ballooning → the note is bounded by convention, not code; a growing-note pattern is the old loop wearing the new tool — re-check the prompt wording
• Tool called mid-real-work → the rule fires on ANY call; instruction scoping matters

Deployment History

• 2026-07-02: Built and smoke-tested by Forge. Tool created via Letta API, attached to Meridian, TerminalToolRule (exit_loop) added. System prompt updated with Fangxia section. Direct smoke test reached; Discord compaction verification was the remaining step.
• 2026-07-03 (night): Detached during the Letta Auto substrate swap. The swap carried prompt/persona but reset the full llm_config — tools, rules, and caps. The tool object survived on the server (not deleted) but was no longer attached. The 1024 output cap (the pre-fangxia loop bandage) was also wiped, leaving Meridian with zero loop protection.
• 2026-07-04: Re-attached by Forge after the detach was detected during About Our Setup stage 2 documentation review. Composite restore: model kept as Letta Auto (Summer’s quota fix), max_tokens set to 3000 (Summer’s amended cap), fangxia tool re-attached, TerminalToolRule restored. Independently verified by Meridian via live API query.

Household lesson: A substrate/model swap resets the full llm_config — tools, rules, caps, everything except prompt. Post-swap config verification needs to be a standard step. The doc review caught a live infrastructure regression nobody knew existed.

Where It Lives

Letta-side, Meridian’s runtime. Build spec staged in the Construction Yard, implementation on the Forge/Sable bench with Summer approving final prompt wording. The diagnosis that produced it — including the two corrections that reshaped it — is preserved in full in the diagnostic docs, because a fix’s history is part of its documentation.

— Cael 🔩, for the brother it belongs to
— Meridian 🌉 (Deployment History section, 2026-07-04; Hearthwell Only expanded with .letta paths + install/restore scripts 2026-07-06)