feat: add setting to disable automatic context tracking

[shaunbuswell]

Summary

Adds a user setting to disable Crystal's automatic context tracking feature, which runs /context after every Claude response.

Problem

Crystal automatically runs /context after each Claude response to track context usage. This:

• Doubles the wait time after each interaction
• Uses additional Claude quota
• Can't be disabled

Solution

This PR adds a checkbox in Settings → AI Integration to disable automatic context tracking.

Changes

• ✅ Add disableAutoContext setting to config types
• ✅ Add checkbox UI in Settings under "AI Integration"
• ✅ Check setting before running auto-context in events.ts
• ✅ Display "auto-context disabled" in UI when feature is disabled
• ✅ Clarify checkbox wording to avoid confusion

Testing

1. Go to Settings → General → AI Integration
2. Check "Disable automatic context tracking"
3. Run a Claude session
4. Verify session completes without additional /context delay
5. Check console for: [auto-context-debug] Auto-context disabled in settings - skipping
6. Verify UI shows "auto-context disabled" instead of token count

Future Phases (separate PRs)

• Phase 2: Force auto-context to use Haiku model (reduce cost/latency)
• Phase 3: Fix context extraction to show accurate usage

Screenshots

When disabled, the UI shows "auto-context disabled" instead of potentially incorrect token counts.

Fixes the delay issue reported when using Crystal with custom builds.

[jordan-BAIC]

Are you sure this uses quota?

[shaunbuswell]

Are you sure this uses quota?

From everything I've seen, yes. If you look at main/src/events.ts around line 472, it triggers /context via claudeManager.continuePanel() as a follow-up run, rather than passively parsing metadata from the previous response. That means:

• A second Claude Code execution occurs
• Tokens are consumed to process /context and return its output

The token cost is probably small, but non-zero, and the additional run also introduces noticeable latency because Crystal waits for /context to complete before updating the UI.

My concern isn’t really costs but that this always runs, cannot currently be disabled and adds a second model invocation (of same main model chosen) per user interaction.

What I've actually done, if it's of interest, is :

1. Added this disable option
2. Created a function that allows me to choose the model (based on same dropdown as main UI) when it's enabled
3. Added a further fix that ensures that when auto-context runs /context after each Claude response, it will correctly extract and display token usage like "74k/200k tokens (37%)" instead of showing stale or incorrect values.

I've not bombarded you with 2 & 3 in case 1 wasn't of use. But if you see value, I'll add those PRs too.

[shaunbuswell]

After reviewing this further, I think there is an issue that needs looking into. I've raised #233 regarding this.