Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

The following comment was made by an LLM, it may be inaccurate:

Potential Duplicate Found:

• PR fix(session): avoid full history hydration after compaction #31638: fix(session): avoid full history hydration after compaction
  • fix(session): avoid full history hydration after compaction #31638

This PR appears to address the same issue—avoiding full history hydration after compaction. Both PRs focus on optimizing memory usage by limiting what gets hydrated from prompt history during session compaction iterations. You should verify whether PR #31638 is already merged or if it needs to be reconciled with the current PR #39930.

Addressed the follow-up comments:

• linked existing issue perf(core): avoid hydrating discarded compacted session parts #35570 with Closes #35570
• replaced the description with the required PR template
• reviewed fix(session): avoid full history hydration after compaction #31638 and documented the implementation difference: fix(session): avoid full history hydration after compaction #31638 scans hydrated pages, while this branch discovers the boundary from message rows before hydrating retained parts
• rebased onto current dev and renamed the commit/PR scope from opencode to session

Thanks for updating your PR! It now meets our contributing guidelines. 👍

Validated this fix against two real production sessions (Windows, opencode 1.18.13). The two-phase walk works exactly as proposed — here's the measured impact and one edge case worth deciding on:

Measured impact (replicated the walk against the live DBs)

Both sessions currently hydrate all parts on every prompt (stream() → filterCompactedEffect), which is the observed multi-second block + RSS spike. The row-walk + boundary-stop cuts that to <1,300 parts per prompt.

One edge case to decide: empty compaction marker behavior

Both sessions carry a compaction part with tail_start_id: null (empty marker, 0-char text — they're compaction-managed by a plugin, not opencode's native compaction). In the new walk:

if (!compaction.tail_start_id) break pages   // ← stops the walk HERE

This is a behavior change from the old path: filterCompacted(stream(...)) with a null tail_start_id returned the full history (tailIndex = -1 → return result), whereas the new walk breaks at the marker and truncates the hydrated set to messages newer than it. For the two sessions above that means ~7,800 / ~7,900 older messages become unavailable to the model even though they previously fit (up to the context window). The PR description states "incomplete compactions … unchanged" — this path changes them.

Options:

• continue instead of break on null tail_start_id — preserves old behavior for incomplete compactions (no history truncation), at the cost of walking to the session start for those sessions.
• Keep break but document it — for plugin-compaction-managed sessions (magic-context et al.) the older history is re-injected by the plugin anyway, so the truncation is invisible; but for plain opencode sessions with a stale/failed compaction marker it would silently shorten available history.

Also worth noting: the null-marker case in old code already skipped compaction reordering (returned everything), so either choice is defensible — but the "incomplete compactions unchanged" claim in the description only holds with the continue variant.

Otherwise the implementation is clean: messageRows shares the pagination core with page(), the boundary detection correctly keys on completed user messages, and the batched hydrate (rows.slice(index, index + size)) keeps the fix's memory bounded. LGTM.

@Qiiks Addressed in commit 4ff112ab19.

• Treat a missing tail_start_id as an incomplete compaction in both filterCompacted and the paginated pre-hydration walk, so the scan continues and the full history is preserved.
• Added a regression with 60 messages after the marker, covering the 50-row pagination boundary.
• Updated the bounded-hydration fixture to use an explicit tail, preserving its original assertion that older discarded parts are not hydrated.

Verification:

• bun test test/session/messages-pagination.test.ts: 53 passed.
• bun test test/session/prompt.test.ts test/session/compaction.test.ts: 108 passed, 2 skipped.
• bun typecheck: passed.
• Prettier check: passed.

The package-wide bun test run was also attempted but could not complete because of unrelated environment failures: missing zod resolution in plugin tests and a v2 PTY test timeout.

Corrected the follow-up in 69d3227b52 after re-checking native compaction semantics.

tail_start_id is intentionally absent when a completed compaction summarizes the full history (SessionCompaction.select returns no tail when keep.start === 0). Treating every missing tail as incomplete made the row walk hydrate history older than a valid completed compaction, defeating the bounded-hydration change and potentially reintroducing context overflow.

The update now:

• stops at completed no-tail compactions in both the iterable and paginated paths
• keeps full history only when the compaction is genuinely incomplete (no completed summary)
• restores the invalid-old-part regression so a completed full compaction proves older parts are not hydrated

Verification: 162 passed, 2 skipped across the targeted session tests; bun typecheck passed; repository pre-push typecheck passed.

Known persisted-data limitation: a plugin marker paired with a completed-looking summary but no tail is indistinguishable from a valid native full-summary compaction, so it follows native completed-compaction semantics.