Automatically compact long conversations when approaching context window
limits (80% of model context or 180k tokens, whichever is first).

Changes:
- Add enableContextCompaction setting (default: enabled) in Settings >
AI
- Add database fields to track compaction state and backup paths
- Create compaction handler that generates structured summaries via LLM
- Store original messages in app data directory before compaction
- Show visual indicator using dyad-status tag when compaction occurs
- Integrate compaction checks into local_agent_handler
<!-- devin-review-badge-begin -->

---

<a href="https://app.devin.ai/review/dyad-sh/dyad/pull/2515"
target="_blank">
  <picture>
<source media="(prefers-color-scheme: dark)"
srcset="https://static.devin.ai/assets/gh-open-in-devin-review-dark.svg?v=1">
<img
src="https://static.devin.ai/assets/gh-open-in-devin-review-light.svg?v=1"
alt="Open with Devin">
  </picture>
</a>
<!-- devin-review-badge-end -->

<!-- CURSOR_SUMMARY -->
---

> [!NOTE]
> **Medium Risk**
> Touches core `local-agent` chat streaming and adds new
persistence/migration paths plus LLM-driven summarization, so
regressions could affect message ordering/history sent to models. Scoped
behind a setting (default on) with added unit/E2E tests, reducing risk.
> 
> **Overview**
> Adds **automatic context compaction** for `local-agent` chats: when
token usage crosses a threshold, the chat is marked for compaction and
the next turn generates an LLM summary, stores an XML transcript backup
under `.dyad/chats/<id>/compaction-*.md`, and inserts a
`is_compaction_summary` assistant message with a new `<dyad-compaction>`
indicator.
> 
> Introduces a user-facing toggle `enableContextCompaction` (default on)
wired into settings schema/defaults/search and Settings UI, plus new DB
fields on `chats` (`pending_compaction`, `compacted_at`,
`compaction_backup_path`) and `messages` (`is_compaction_summary`).
Updates local-agent streaming to (a) run pending compaction before
processing, (b) send only post-compaction history to the LLM, and (c)
update the correct placeholder message by ID to avoid overwriting the
compaction summary; includes new unit/E2E coverage and test-fixture
support for streaming usage data/snapshot scrubbing.
> 
> <sup>Written by [Cursor
Bugbot](https://cursor.com/dashboard?tab=bugbot) for commit
df1cdaacbe3e9f3fb47473a3e6965e302a75fe7a. This will update automatically
on new commits. Configure
[here](https://cursor.com/dashboard?tab=bugbot).</sup>
<!-- /CURSOR_SUMMARY -->

<!-- This is an auto-generated description by cubic. -->
---
## Summary by cubic
Adds automatic context compaction for local-agent mode to summarize long
conversations when token usage reaches 80% of the model’s context window
or 180k tokens. Original messages are preserved and backed up, a
structured summary is inserted, and a dyad-compaction indicator is
shown.

- **New Features**
- New toggle in Settings > AI: enableContextCompaction (default on) with
UI switch and searchable setting.
- DB fields to track compaction and backups (chats: compacted_at,
compaction_backup_path, pending_compaction; messages:
is_compaction_summary).
- Compaction handler backs up LLM-visible messages to
.dyad/chats/<chatId>/compaction-*.md, generates structured summaries
with a system prompt, inserts a summary message, and emits
chat:compaction:complete.
- Local-agent flow: pending compaction runs at the start of the next
turn; LLM history includes only post-compaction messages; streaming
updates target the placeholder message by ID to avoid overwriting the
summary.
- Token threshold checks integrated into local_agent_handler (min(80% of
context window, 180k)).

<sup>Written for commit df1cdaacbe3e9f3fb47473a3e6965e302a75fe7a.
Summary will update on new commits.</sup>

<!-- End of auto-generated description by cubic. -->

---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>