Sessions & Context

OpenOrca manages conversation history through sessions and monitors context window usage to prevent exceeding your model's limits.

Sessions

Auto-Save

By default (autoSave: true), OpenOrca saves your conversation when you exit. The session is stored as a JSON file in ~/.openorca/sessions/.

Resuming Sessions from the Command Line

Instead of manually loading sessions with /session load, you can resume directly on startup:

openorca --continue          # Resume the most recent session
openorca -c                  # Short form

openorca --resume abc123     # Resume a specific session by ID
openorca -r abc123           # Short form

Managing Sessions

> /session list                # List saved sessions (shows up to 20)
> /session save                # Save current session
> /session save My Refactor    # Save with a custom title
> /session load abc123         # Load a session by ID
> /session delete abc123       # Delete a session
> /rename My New Title         # Rename the current session

The session list shows: ID, title, last updated timestamp, and message count.

Session Storage

Sessions are JSON files at ~/.openorca/sessions/{id}.json. Each contains:

• Session ID and title
• Creation and update timestamps
• Full message history (roles, text content, tool calls, tool results)

The maxSessions config setting (default: 100) limits how many sessions are kept. Oldest sessions are pruned when the limit is exceeded.

Context Window

Understanding Context

The context window is the total number of tokens your model can process at once — this includes the system prompt, all conversation history, and the next response. When you exceed it, the model starts losing information or producing errors.

Checking Usage

Use /context (or /ctx) to see current context window usage:

• Estimated tokens — approximate token count of the full conversation
• Context window — your configured window size (from contextWindowSize)
• Usage percentage — color-coded: green (<70%), yellow (70-90%), red (>90%)
• Auto-compact threshold — when auto-compaction kicks in
• Message counts — breakdown by role (system, user, assistant, tool)
• Visual bar — [#####-----] showing usage at a glance

Setting the Right Window Size

Set contextWindowSize in your config to match your model's actual context window:

Model	Typical Context Window
Mistral 7B	8192 or 32768
Llama 3 8B	8192
DeepSeek R1 7B	16384 or 32768
Qwen 2.5 7B	32768

Check your model's card in LM Studio for the exact value.

Compaction

When conversations grow long, compaction summarizes older messages to free up context space.

Manual Compaction

> /compact                              # Compact with default summarization
> /compact focus on the database schema  # Direct the summary to preserve specific context

What happens:

1. Messages older than the last N turns are collected (N = compactPreserveLastN, default 4)
2. These messages are sent to the LLM with a summarization prompt
3. The summary replaces the older messages
4. Recent turns are preserved exactly as-is
5. You see a before/after count: Compacted: 24 messages -> 6 messages (~3200 -> ~800 tokens)

Auto-Compaction

When autoCompactEnabled is true (default), OpenOrca checks context usage at the start of each agent loop iteration. If usage exceeds autoCompactThreshold (default: 80%) of contextWindowSize, it automatically runs compaction.

You'll see Compacting conversation... in the output when this happens.

Compaction Settings

Setting	Default	Description
autoCompactEnabled	true	Enable/disable auto-compaction
autoCompactThreshold	0.8	Trigger at this percentage of context window
compactPreserveLastN	4	Recent turns to keep verbatim

Rewind

Remove recent conversation turns without compacting:

> /rewind        # Remove the last turn (1 user message + 1 assistant response)
> /rewind 3      # Remove the last 3 turns

This is useful when the model went down a wrong path and you want to try a different approach.

Best Practices

1. Set the right context window — match contextWindowSize to your model's actual limit
2. Use /context regularly — especially during long sessions
3. Compact with focus instructions — tell it what to preserve when context is precious
4. Use /rewind for mistakes — cheaper than compacting, keeps more context
5. Start fresh for new tasks — use /clear when switching to an unrelated task
6. Save before long sessions — use /session save as a checkpoint you can return to