> ## Documentation Index
> Fetch the complete documentation index at: https://chatcli.edilsonfreitas.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Conversation Control
> Use /compact and /rewind to manage conversation context: intelligently compact history and rewind to earlier points in the conversation.
ChatCLI provides two powerful mechanisms for conversation history control: **intelligent compaction** to reduce context size without losing critical information, and **rewind** to restore the conversation to an earlier point.
***
## /compact — Context Compaction
The `/compact` command reduces conversation history size so it fits within the model's context window, preserving the most important information.
### Usage Modes
Without arguments, runs the 3-level compaction pipeline:
```bash theme={"system"}
/compact
```
The automatic pipeline follows this order:
1. **Level 1 — Trimming** (near-lossless): strips `` tags, compacts XML, removes duplicates
2. **Level 2 — Structured summarization**: extracts facts (files read, modified, commands executed, decisions) in bullet point format
3. **Level 3 — Emergency truncation**: drops middle messages, keeping system prompt and recent messages
During the pipeline, the terminal shows **live feedback** for each phase (no more silent freeze):
```text theme={"system"}
│ 📦 Compacting history (23 msgs, 4.2 MB → target 2.9 MB)
│ 🧹 Trim: stripping reasoning/dedup (no LLM)…
│ 🧠 Summarizing old messages via LLM (may take 30-90s — ESC cancels)…
│ ✓ Summary applied (23 → 9 msgs, 4.2 MB → 1.8 MB)
```
Press **ESC** or **Ctrl+C** to abort Level 2 (LLM). Cancellation propagates through context and preserves history intact (no blind fallthrough to Level 3).
With an instruction, you tell the AI **what to preserve**:
```bash theme={"system"}
/compact preserve the auth module architecture and errors found
```
```bash theme={"system"}
/compact keep all file paths and refactoring decisions
```
Guided mode sends your instruction as a directive to the summarizer AI, ensuring that information you consider critical is retained during compaction.
### How It Works
```mermaid theme={"system"}
flowchart TD
A["/compact instruction"] --> B{Instruction provided?}
B -->|Yes| C[Guided Compact]
B -->|No| D[Auto Compact - 3 levels]
C --> E[Split: system + middle + recent]
E --> F[LLM summarizes middle with instruction]
F --> G[Rebuild: system + summary + recent]
D --> H[Level 1: Trim]
H --> I{Fits budget?}
I -->|Yes| J[Done]
I -->|No| K[Level 2: Summarization]
K --> L{Fits budget?}
L -->|Yes| J
L -->|No| M[Level 3: Truncation]
M --> J
```
### Message Preservation
In both modes:
* **System messages** are always fully preserved
* **Recent messages** (4 in guided, 10 in automatic) are kept verbatim
* Only the **middle block** is summarized or removed
* Metadata indicates the message is a summary (`IsSummary: true`)
### Example Output
After compaction, a summary message replaces the middle messages:
```text theme={"system"}
[GUIDED COMPACT — covering 42 earlier messages | instruction: "preserve auth architecture"]
## Files Read
- pkg/auth/handler.go (120 lines) - HTTP handlers for authentication
- pkg/auth/middleware.go (85 lines) - JWT validation middleware
## Files Modified
- pkg/auth/handler.go:45-60 - Added refresh token endpoint
## Key Decisions
- JWT tokens expire in 15 minutes with refresh token rotation
- Middleware validates at gateway level, not per-service
## Errors & Resolutions
- "token expired" panic in middleware → added nil check on claims
```
***
## /rewind — Go Back in Time
`/rewind` lets you restore the conversation to an earlier point, undoing unwanted messages and responses.
### How It Works
ChatCLI automatically saves **checkpoints** of the conversation before each LLM call. You can revert to any of those points.
```bash theme={"system"}
/rewind
```
This displays an interactive menu:
```text theme={"system"}
REWIND — Select a checkpoint to restore
─────────────────────────────────────────
[1] 14:32:05 28 msgs implement the refresh token endpoint
[2] 14:28:12 24 msgs fix the JWT middleware bug
[3] 14:25:01 20 msgs analyze the authentication module
[4] 14:20:33 16 msgs (start)
Select [1-4] or (q)uit:
```
Select the checkpoint number to restore. The history is reverted to the exact state at that point, including all messages and context.
### Keyboard Shortcut: Esc+Esc
Press **Esc** twice quickly (less than 500ms between presses) to open the rewind menu directly from the prompt, without typing `/rewind`.
### Limits
* A maximum of **20 checkpoints** are kept (oldest ones are discarded)
* Checkpoints exist only in the current session (not persisted to disk)
* When rewinding, checkpoints after the restored point are removed
***
## Unified History
ChatCLI uses a **single history array** shared across all modes (chat, agent, coder). This means:
* Switching between `/agent`, `/coder`, and chat **preserves all context**
* The AI doesn't "forget" what was done in another mode
* `/compact` and `/rewind` operate on the complete history, regardless of mode
Sessions saved with older ChatCLI versions that used separate per-mode histories are automatically converted when loaded — histories are merged in chronological order.
***
## When to Use
When ChatCLI warns that the context is large, or when you notice degraded responses.
When you know exactly which information is critical and must not be lost during compaction.
When the AI took a wrong path, generated incorrect code, or you want to try a different approach.
Quick shortcut for rewind without leaving your typing flow.
***
## Next Steps
Save and reuse conversations across projects.
Customize the AI and maintain long-term context across sessions.