Preprompt & Continuity-Summary Improvement Proposal

**Date:** 2026-05-15

**Session:** ses_1d42640c8ffe7J2uRRGcGHXgbe

**Model:** kimi-for-coding/k2p6

**Status:** Draft — updated with source-code analysis

---

Problem Statement

The current preprompt system causes:

1. **Massive duplication** — the same conversation appears 2-3 times in different formats

2. **Stale status markers** carried forward as "Latest" even after resolved

3. **Attention drift** — model cannot distinguish resolved bootstrap context from live message

4. **Context bloat** approaching 200K reload threshold faster than necessary

This directly caused a self-contradiction bug: the model stated a fact, then immediately offered to investigate that same fact.

---

Root Cause Analysis (Source Code)

The injection logic lives in dispatch-acp-HFOhX_U-.js (buildPersistentChatInjectedPrompt, line 966).

Two injection paths exist:

**Path A — Continuous session** (shouldInjectPersistentChatPrompt() returns false):

• Turn metadata

• `## Recent Perpetual Log Excerpt` (last **6** log entries)

• Current user message

**Path B — Reload / new backend session** (shouldInjectPersistentChatPrompt() returns true):

• Turn metadata

• `## Rolling Continuity Summary` (full `continuity-summary.txt`)

- Which itself contains: "Current thread" summary + **30** log entries

• `## Recent Perpetual Log Excerpt` (last **6** log entries **again**)

• Current user message

**Result:** On reload, the model sees the conversation **twice**: 30 entries inside the continuity summary + 6 more in the separate section. This is the duplication.

---

Recommended Fixes

1. Recent Log Excerpt — Context-Dependent Length

**For continuous sessions:**

• **Not needed at all** or **max 2 entries**

• The ACP backend already holds the full conversation in its context window

• The only purpose is catching immediate follow-ups ("so did you check?")

• **Current:** 6 entries. **Proposed:** 0-2 entries.

**For reloads:**

• **This is the compaction** — it's the only conversation history the model receives

• **Need more than 10** — current system injects 30 inside continuity summary + 6 separate

• **Proposed:** 20-25 entries as the single source of truth

• Remove the 30-entry duplicate inside continuity-summary

2. `continuity-summary.txt` — Remove Conversation History

The continuity summary should be **state/memory only**, not conversation transcript.

**Remove from rebuildPersistentChatContinuitySummary() (JS, line 931):**

• `Recent log excerpt:` block (lines 940-943)

• Keep only: `Current thread / pending next action:` (latest user request, assistant status, actionable notes)

**Remove from render_summary() (Python memory script, lines 452-458):**

• `Recent conversation excerpt:` block

**Keep in continuity-summary:**

• Active decisions and guardrails

• Pending threads / open topics

• Timer status

• Model capability notes

3. Stale Status Markers

In summarizePersistentChatCurrentThread() (JS, line 908):

• `Latest assistant status` should be the **immediately preceding turn only**

• Do not carry "ACP_TURN_FAILED" forward once the session has reloaded

• Add a "resolved" tag or drop the field if the error is historical

4. Visual Separators (Low-Hanging Fruit)

In buildPersistentChatInjectedPrompt() (JS, line 966):

• Add explicit markers between bootstrap context and live message

• Current: `[Persistent-chat authoritative state...]` then dumped text

• Proposed: clear `--- BOOTSTRAP CONTEXT (already resolved) ---` vs `--- LIVE TURN ---`

---

Implementation Checklist

JavaScript Runtime (`dispatch-acp-HFOhX_U-.js`)

• [ ] `rebuildPersistentChatContinuitySummary()`: Remove `Recent log excerpt` block

• [ ] `buildPersistentChatInjectedPrompt()`:

- If continuous session: reduce readPersistentChatLogEntries(6) → readPersistentChatLogEntries(2) or skip entirely

- If reload: increase to 20-25 entries

- Add visual separator between bootstrap and live message

• [ ] `summarizePersistentChatCurrentThread()`: Fix stale status forwarding

Python Memory Script (`persistent-chat-memory`)

• [ ] `render_summary()`: Remove `Recent conversation excerpt` section

• [ ] `render_decision_state()`: Compress to one-line format

Gateway Config

• [ ] Verify `PERSISTENT_CHAT_CONTINUITY_PATH` is still read correctly after format change

• [ ] Test reload path: ensure 20-25 log entries are accessible on fresh session

---

File Paths

| Component | Path |

|---|---|

| JS runtime patch | /home/gleb/openclaw-tools/patches/openclaw-runtime/dispatch-acp-HFOhX_U-.js |

| Python memory script | /home/gleb/openclaw-tools/scripts/persistent-chat-memory |

| Live JS (needs re-patch) | /home/openclaw/.local/npm-global/lib/node_modules/openclaw/dist/dispatch-acp-HFOhX_U-.js |

| Continuity summary output | /home/openclaw/workspace/state/persistent-chat/continuity-summary.txt |

| Preprompt debug output | /home/openclaw/workspace/state/persistent-chat/last-injected-preprompt.txt |

---

Rationale

• **Context quality over token price**: The goal is reducing reloads at 200K, not saving API dollars

• **One fact, one location**: Conversation history lives in the per-turn recent log; decisions live in continuity summary

• **Continuous sessions need almost no log injection**: ACP already remembers

• **Reloads need generous log injection**: Because ACP remembers nothing; this is the compaction

• **Stale status is worse than no status**: An old error marked "Latest" causes the model to re-debug resolved issues

---