perf(server): stop caching unused OpenCode tool parts

[t3dotgg]

OpenCode retained tool input and output in a session map even though no reader used those cached parts.

Skip tool parts in that map. Keep text, reasoning, usage tracking, and emitted tool events unchanged.

Validation:

• All 108 adapter tests pass, including tool lifecycle events and late assistant metadata.
• Server typecheck, targeted lint, and formatting pass.
• A real-adapter check with 40 completed tools retained 20 MiB of tool output before this change and zero tool parts after it. The session stayed active. Every tool result and the late assistant text still emitted.

Part of the performance audit, item S6. Other session maps remain separate work.

Created with GPT-6 Astra (preview) in Codex.

---

Note

Low Risk
Targeted memory optimization with no change to emitted events; covered by an expanded adapter test suite.

Overview
OpenCode session handling no longer stores tool message parts in partById. On message.part.updated, only non-tool parts are cached; tool lifecycle still uses the incoming part, so item.started / item.updated / item.completed and text deltas behave the same while large tool I/O is not retained in memory.

A new adapter test covers tool states (pending → running → completed/error) arriving before late assistant message.updated role metadata, asserting event order and payloads still match expectations.

^{Reviewed by Cursor Bugbot for commit 5c02db1. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Stop caching unused OpenCode tool parts in OpenCodeAdapter

Changes message.part.updated handling in OpenCodeAdapter.ts so incoming tool parts are no longer stored in context.partById; non-tool parts continue to be cached. The original part is still passed to downstream message-role and usage-processing logic. Adds a test covering pending, running, completed, and error tool-part lifecycle updates.

^{Macroscope summarized 5c02db1.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.6 KiB	13.8 KiB	+186 B (+1.3%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.0 KiB	7.0 KiB	+3 B (+0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.6 KiB	6.7 KiB	+183 B (+2.7%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	57.1 KiB	58.5 KiB	+1.4 KiB (+2.5%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.6 KiB	13.6 KiB	−17 B (−0.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.0 KiB	7.0 KiB	0 B (0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.5 KiB	−17 B (−0.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	57.9 KiB	57.8 KiB	−44 B (−0.1%)	66.4 KiB	✅
Claude	Live turn messages	10	9	−1 (−10.0%)	21	✅

Baseline: c7bf311 · PR result: 5c02db1 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 113.8 KiB
• Claude decoded thread snapshot: 114.5 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[macroscopeapp]

Approvability

Verdict: Approved at 5c02db1

Macroscope's review found this PR approvable — This is a small, self-contained OpenCode memory optimization that stops retaining large tool parts while preserving emitted lifecycle events and non-tool processing. The added adapter coverage exercises the affected ordering and lifecycle behavior, with no product-default, schema, security, or deployment impact.

^{You can add or adjust custom eligibility rules. Learn more.}