Bug Description

AGENTS.md states a hard policy on prompt caching:

Prompt Caching Must Not Break

Hermes-Agent ensures caching remains valid throughout a conversation. Do NOT implement changes that would:

• Alter past context mid-conversation
• Change toolsets mid-conversation
• Reload memories or rebuild system prompts mid-conversation

Cache-breaking forces dramatically higher costs. The ONLY time we alter context is during context compression.

The Honcho memory provider violates this directly.

Per the Prompt Assembly architecture doc, Layer 3 of the cached system prompt is the "Honcho static block (when active)". Per the Honcho Memory integration doc, that same block is:

• Base context — session summary, user representation, user peer card, AI self-representation, AI identity card. "Refreshed on contextCadence."
• Dialectic supplement — LLM-synthesized reasoning about the user's current state. "Refreshed on dialecticCadence."

Every refresh rewrites Layer 3 of the cached prompt mid-conversation, invalidating the entire KV cache on every backend that does prefix caching:

• Local: llama.cpp, vLLM --enable-prefix-caching, MLX with PagedAttention
• Cloud: Anthropic / OpenAI / Bedrock with explicit cache control

Causal attention means mid-prefix mutations invalidate everything after the first divergence — every token downstream of the Honcho block must be reprocessed from scratch. On a ~20K-token session that is a multi-minute reprocess on every cadence tick.

The same Prompt Assembly doc already classifies "later-turn Honcho recall injected into the current-turn user message" as an API-call-time-only layer (explicitly not persisted to the cached system prompt) — exactly the correct pattern. The auto-injected base / dialectic block just hasn't been moved over.

Steps to Reproduce

1. Configure Hermes with recallMode: \"hybrid\" (or \"context\") and any finite contextCadence / dialecticCadence.
2. Point Hermes at a backend with prefix caching enabled (e.g. local vllm --enable-prefix-caching, llama.cpp, MLX via vLLM-MLX).
3. Send the first user message — first-turn full prefill is expected.
4. Send N−1 more messages where N = min(contextCadence, dialecticCadence). Prefix cache hits as expected.
5. Send one more message. Observe: full prefill from scratch, not just the new tail tokens.

Cross-check: run the same backend + model under a client that does not mutate the system prompt mid-session (e.g. opencode). Prefix cache hits on every turn, confirming the backend is not at fault.

Expected Behavior

Per the stated policy: the cached system prompt is byte-identical for the duration of a session (compression excepted).

Auto-injected Honcho context rides on the current user message, same path as the existing `_ext_prefetch_cache` injection (`run_agent.py` ~L6737–6740), and is therefore ephemeral and cache-safe.

Actual Behavior

Every `contextCadence` / `dialecticCadence` tick rewrites Layer 3 of the cached system prompt.

Observed on Gemma 4 31B 8-bit (MLX) served via a `vllm-mlx` fork with PagedAttention prefix caching:

With `contextCadence: 1` every user turn is a full prefill. Bumping to 4 reduces the rate to one in four, but the underlying design is still wrong — on any multi-turn session, some fraction of turns will always incur a full prefill regardless of how it is tuned.

Root Cause

`agent/prompt_builder.py` emits the Honcho static block inside `_build_system_prompt()`, whose result is cached on `self._cached_system_prompt`. The Honcho memory provider's `prefetch_all()` result populates this block and is deliberately refreshed on cadence, contradicting the cached-prompt contract.

The correct injection site already exists: `_ext_prefetch_cache` in `run_agent.py` appends per-turn context to the current user message (same file, ~L6737–6740 — also referenced in #5719). Only a subset of Honcho output currently goes through that path.

Proposed Fix

1. Remove the Honcho static block from Layer 3 of `_build_system_prompt()` in `agent/prompt_builder.py`.
2. Extend `_ext_prefetch_cache` (or an equivalent per-turn slot) in `run_agent.py` to carry the formatted `## Honcho Context` block for the current turn only.
3. Document and enforce the invariant: the cached system prompt is byte-stable for the session; any Honcho content that varies turn-to-turn must ride on the current user message.

This is the same architectural move already proposed in #3353 for runtime metadata, and enforces the policy stated in `AGENTS.md`. It subsumes the `contextCadence` / `dialecticCadence` knobs under the same correctness guarantee instead of requiring users to manually trade off memory freshness against prefill cost.

Environment

Relevant Honcho config (`~/.hermes/honcho.json`):

```json
{
"recallMode": "hybrid",
"contextCadence": 4,
"dialecticCadence": 4,
"dialecticDepth": 1,
"dialecticReasoningLevel": "medium",
"observationMode": "directional",
"writeFrequency": "async",
"sessionStrategy": "per-directory"
}
```

Related Issues

Same architectural class — mid-session mutation of cached prompt state:

• perf(ttft): move runtime metadata out of the cached system prompt #3353 — `perf(ttft): move runtime metadata out of the cached system prompt` — same principle, different varying field; this bug is a much larger, more frequently varying instance of the same anti-pattern.
• perf: system prompt timestamp changes after compression, invalidating prefix cache on local LLM backends #8687 — `perf: system prompt timestamp changes after compression, invalidating prefix cache on local LLM backends` — same class, timestamp specifically.
• [Performance] KV cache invalidation on compression hurts local MoE models — defer unnecessary system prompt rebuilds #4319 — `[Performance] KV cache invalidation on compression hurts local MoE models` — same class, compression trigger.
• [Bug]: KV cache invalidation on new user message due to message format differences between agentic loop and session reload #4555 — `[Bug]: KV cache invalidation on new user message due to message format differences between agentic loop and session reload` — byte-stability principle for messages on reload.
• [Performance] Missing global conversation history state causes 314x slowdown on llama.cpp with Qwen3.5-27B #13442 — `[Performance] Missing global conversation history state causes 314x slowdown on llama.cpp with Qwen3.5-27B` — downstream symptom of the same invariant being violated.

Adjacent Honcho issues (for completeness, not duplicates of this bug):

• Bug: Honcho Context block leaking into user-visible messages via vision auto-analysis #5719 — Honcho context block leaking into vision auto-analysis output
• Honcho plugin: race condition in session cache causes silent conclusion/memory loss #5102 — Thread-safety race in `HonchoSessionManager._cache`
• fix(honcho): sequential chat dispatch, session key strategy, and peer-card fallback break end-to-end Honcho UX #4751 — Sequential chat dispatch / session-key / peer-card issues
• Optional lightweight cached startup context for Honcho recallMode=tools #2150 — Feature request for cached startup context in `recallMode=tools`