fix(opencode): reduce memory usage during prompting with lazy boundary scan and context windowing

[BYK]

Issue for this PR

Closes #18136

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Two targeted optimizations to reduce peak RSS during prompting from ~4-8GB down to ~1.2GB:

1. Lazy compaction boundary scan (filterCompactedLazy)

The prompt loop calls filterCompacted(stream(sessionID)) which streams ALL messages newest→oldest, loading parts for every message. For compacted sessions, most of those parts are discarded once the boundary is found.

New approach: probe the newest 50 message infos (1 DB query, no parts). If a compaction summary is detected, use a two-phase scan — info-only scan to find the boundary, then hydrate parts only for messages after it. If no compaction summary is found, fall back to the original single-pass filterCompacted(stream()) to avoid wasted info-only queries.

2. Context-window message windowing

toModelMessages was called with ALL messages (e.g., 7,704 for a long session), creating ModelMessage wrapper objects for every one. These flow through 4-5 copy layers (toModelMessages → convertToModelMessages → ProviderTransform.message → convertToLanguageModelPrompt), each creating ~60MB of wrapper objects.

Now the prompt loop estimates which messages from the tail fit in the LLM context window (model.limit.context × 4 chars/token) and only passes those to toModelMessages. For a 7,704-message session where ~200 fit, this cuts the conversion pipeline from ~300MB to ~10MB.

3. Prompt loop caching

The conversation is loaded once before the loop. On normal tool-call iterations, only the latest 200-message page is fetched and merged into the cache. Full reload only happens after compaction.

How did you verify your code works?

• Monitored RSS with /proc/<PID>/status every 30s during active prompting
• Before: peak 4.8GB, idle 1GB
• After: peak 1.2GB, idle ~580MB
• All session tests pass (118 pass, 4 skip, 0 fail)
• Tested with both compacted and uncompacted sessions

Screenshots / recordings

Memory monitoring (30s intervals) after fix:

time,rss_mb,hwm_mb
20:46:02,942,1236    ← active prompting
20:48:02,1020,1236   ← peak during tool calls
20:50:02,606,1236    ← settled after activity
20:55:32,568,1236    ← stable idle

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Hey! Your PR title perf(session): reduce memory usage during prompting with lazy boundary scan and context windowing doesn't follow conventional commit format.

Please update it to start with one of:

• feat: or feat(scope): new feature
• fix: or fix(scope): bug fix
• docs: or docs(scope): documentation changes
• chore: or chore(scope): maintenance tasks
• refactor: or refactor(scope): code refactoring
• test: or test(scope): adding or updating tests

Where scope is the package name (e.g., app, desktop, opencode).

See CONTRIBUTING.md for details.