Edit tool persists full-file before/after snapshots per call; unbounded part-table growth makes large sessions slow at every prompt #40816

Summary

The edit tool stores the entire file content twice (before + after) in each tool part's metadata.filediff, and the session prompt pipeline hydrates every part of the session on every prompt (stream() pages through all messages + hydrate() loads all parts). On long-lived sessions this compounds into hundreds of MB of DB data and a multi-second, memory-spiking load per prompt.

Evidence — two real sessions (Windows, opencode 1.18.13)

Single edit part anatomy (largest example):

filediff: { file, before: 214,358 chars, after: 214,698 chars, additions, deletions }
total part JSON: 513,383 bytes   ← for ONE small edit to a ~210 KB file

The full-file snapshot is stored with no patch string — the compact unified diff exists in metadata.diff (a few KB) but the two full copies are persisted regardless.

Per-prompt load: packages/opencode/src/session/message-v2.ts — stream() (line 469) pages ALL messages (50/page loop, ~155 queries for session A) and hydrate() (line 98) does inArray(PartTable.message_id, ids) over all parts, then filterCompacted() converts everything to model messages. This happens before any plugin messages.transform compression. Observed: ~1.9k CPU-seconds + ~1.2 GB RSS on a session with 7.7k messages while idle-ish; spike visible during prompt assembly.

Why it didn't used to be an issue

Growth curve: a few hundred messages hydrate in ms. At 8k messages × 32k parts (with ~190 MB of edit snapshots), the same load takes seconds and allocates hundreds of MB transiently. Nothing "breaks" — it degrades linearly with session size, and edit-heavy sessions hit the pain point first.

Suggested fix directions

1. packages/opencode/src/tool/edit.ts (~line 181): cap or omit the full-file before/after in metadata.filediff for large files (e.g. store only the unified diff/patch, or truncate snapshots beyond a size threshold). The compact diff already carries the change; patch is what Snapshot.FileDiff's schema (packages/schema/src/file-diff.ts) actually models as required-ish.
2. packages/opencode/src/session/message-v2.ts stream()/hydrate(): bound the prompt-build load (e.g. hydrate only messages within the context budget, or skip hydrating tool-part metadata for old messages). Compaction already exists but only trims the model-visible tail — stream() still reads the whole history.
3. Consider a storage hygiene pass: old tool parts' metadata.filediff could be dropped by compaction (compaction.prune marks but doesn't reclaim).

Repro

• Long edit-heavy session (thousands of messages), resume it, prompt once; watch RSS + CPU during prompt assembly.
• opencode.db part table: SELECT json_extract(data,'$.type') t, COUNT(*), SUM(LENGTH(data)) FROM part GROUP BY t — tool dwarfs everything; edit-typed parts dominate.

Note: this was diagnosed on 1.18.13; the same edit.ts filediff producer and message-v2.ts hydrate path are present on dev.