Skip to content

CHANGELOG.md

Latest commit

 

History

History
55 lines (32 loc) · 5.3 KB

CHANGELOG.md

File metadata and controls

55 lines (32 loc) · 5.3 KB

Changelog

All notable changes to opencode-magic-context are documented here. This project adheres to Semantic Versioning.

[Unreleased] — v3.3.1: Tag-Owner Identity Fix

Fixed

  • Tool-call collision bug: when two assistant turns within a single session reused the same OpenCode-generated tool callID (e.g. both invoked read:32), the runtime keyed both invocations to the SAME tag row by messageId == callId. Dropping the first turn's tag silently propagated to the second turn's content, corrupting the conversation. The fix adds composite identity for tool tags: each row is now uniquely identified by (session_id, message_id, tool_owner_message_id) where tool_owner_message_id is the assistant message hosting the invocation.

    This was the bug class behind several user-visible symptoms:

    • "My recent search results just disappeared after I edited a file."
    • Reasoning preservation reverting on a fresh assistant turn even though no ctx_reduce was issued.
    • Heuristic dedup silently merging two semantically distinct tool invocations.

  • Drop queue cross-compartment matching: compartment-runner-drop-queue no longer queues drops for tool tags whose owner lies outside the compartment range. Pre-fix it matched by bare callId, so a callId reused outside the compartment matched an in-compartment tag by string equality and got wrongly dropped.

  • Heuristic dedup cross-owner false positives: applyHeuristicCleanup now keys both the tag-side index AND the fingerprint-side map by composite (ownerMsgId, callId), with the fingerprint VALUE also including ownerMsgId. Cross-owner pairs with same (toolName, args) now produce DISTINCT fingerprints and are NOT merged. Within-same-owner duplicates (Pi parallel-tool-calls shape) still group correctly.

  • message.removed cleanup cascade: deleteTagsByMessageId now also deletes tool tags whose tool_owner_message_id matches the removed message id. Pre-fix, removing an assistant message left its tool tag rows orphaned in the DB.

Added

  • Schema migration v10 (migrations.ts): adds tool_owner_message_id (TEXT NULL) column, partial UNIQUE index idx_tags_tool_composite, and partial lookup index idx_tags_tool_null_owner to back lazy adoption.

  • Layer B backfill (tool-owner-backfill.ts): one-shot lease-based backfill that reads OpenCode's session DB and populates tool_owner_message_id on legacy NULL-owner rows. Validated against the user's real 370MB DB (17.86s for 3,284 sessions).

  • Tagger composite identity API (tagger.ts):

    • assignToolTag(sessionId, callId, ownerMsgId, ...) — allocate or reuse a tool tag scoped by composite identity.
    • getToolTag(sessionId, callId, ownerMsgId) — composite-identity lookup.
    • bindToolTag(sessionId, callId, ownerMsgId, tagNumber) — recovery binding.
    • assignTag / getTag are narrowed to non-tool types (message / file) — TS forbids passing "tool" to them.

  • Lazy adoption fallback: when a transform pass observes a tool call whose composite key has no match in the in-memory map, the tagger queries getNullOwnerToolTag for legacy orphans and atomically claims one via adoptNullOwnerToolTag (NULL guard ensures first claim wins). This handles unbackfilled NULL-owner rows incrementally.

  • End-to-end collision-repro test (packages/e2e-tests/tests/tag-owner-collision.test.ts): drives a real OpenCode + magic-context plugin pair through the bug-class scenario and verifies the schema, indexes, and DB-level invariants hold.

  • Microbenchmark for nearest-prior owner derivation (packages/plugin/scripts/benchmark-nearest-prior.ts): documents plan Test #45's exit criterion (0.0455 ms avg on a 30k-tag session — 10× under the 0.5 ms budget).

Changed

  • Tag-transcript Pi pipeline (tag-transcript.ts): removed the outer db.transaction() wrapper. Per-tag SAVEPOINTs inside assignToolTag / assignTag already provide the atomicity needed; the outer wrapper was a cache-bust amplifier that rolled back ALL tag inserts in a pass on a single late UNIQUE collision while leaving in-memory message mutations and §N§ prefixes already applied.

  • Drop queue API (read-session-chunk.ts: getRawSessionTagKeysThrough): now returns RawSessionTagKeys with messageFileKeys: Set<string> and toolObservations: Map<string, Set<string>> instead of one collapsed Set<string>. The split allows tool tags to be matched by composite identity while message/file tags continue using globally-unique content ids.

Compatibility

  • Pre-v3.3.1 sessions: rows written before this version have tool_owner_message_id = NULL. The Layer B backfill populates them from OpenCode's session DB on plugin upgrade. Sessions for which OpenCode's DB is unavailable (foreign-harness sessions, deleted DB) fall back to lazy adoption — orphans are converted to non-NULL on the next observation. Drop queue and heuristic cleanup gracefully degrade to bare-callId match for unbackfilled rows (plan §Risk #20).

  • Cache stability: composite-key migration deterministically produces the same §N§ prefix across passes for the same observation. Anthropic prompt-cache prefix stability is preserved on defer passes (verified via existing cache-stability.test.ts and the new collision-repro tests).

For prior versions, see git history.