perf(server): ACP item ordinals no longer copy a session-wide map per item

[juliusmarminge]

The ACP V2 adapter (used by Grok, Devin, Antigravity and the registry agents) has the same item-ordinal problem as Claude (#13864) and Cursor (#13866). Every item ordinal of the session sat in one Map that was copied on each new item and never pruned. Only a rollback's runtime replacement cleared it. A long session paid O(items) memory and O(items²) copying.

What changed

• Root items: the ordinal map moves onto the turn (ActiveAcpTurn.itemOrdinals). It is mutated in place and goes away when the turn settles. The per-turn counter map is gone, because the map's size is the counter.
• Subagent child items (a child session's messages and tool calls) used the same session map as a dedupe cache, while their ordinals came from the subagent's own nextChildOrdinal. They now keep that cache on the subagent (childItemOrdinals). The subagent record already carries over into later turns (carryoverSubagents), so a child update that arrives after the root turn settled still finds its ordinal.
• The rollback reset of the two session maps is gone, because the maps are gone. Rollback already refuses while a turn is active.

Why per-turn is safe for root items

Every resolveItemOrdinal call passes the active turn. A subagent keeps its own turnItemOrdinal on its record (existing?.turnItemOrdinal ?? resolveItemOrdinal(...)), including across carryover. Before the fix I instrumented resolveItemOrdinal to flag any hit on an ordinal a different turn had allocated, then ran all replay suites plus the ACP, registry and Grok adapter tests: 260 ACP allocations, 3 cross-turn hits. All 3 came from the mock ACP agent in AcpAdapterV2.test.ts ("waits for immediate allow and deny permission responses…"), which hardcodes toolCallId: "tool-call-1" on every prompt. A real agent does not reuse a tool call id across prompts. Even for that mock, the only effect of the fix is that turn 2's tool gets a turn-2 ordinal instead of reusing turn 1's. The test does not assert ordinals, and it passes unchanged.

Benchmark

This is the same resolver shape as Claude and Cursor. A scratch, uncommitted script ran it before and after, in 10 turns, resolving each item twice:

N items	before	after
1,000	60 ms	1 ms
10,000	7.6 s	4 ms
50,000	253 s	37 ms

A 50k-entry map retains about 1.8 MiB for the life of the session. After the fix, root-item memory is bounded by the largest single turn, and child-item memory by the subagent's lifetime.

Verification

• vp test run with AcpAdapterV2.test.ts, AcpRegistryAdapterV2.test.ts, GrokAdapterV2.test.ts, AntigravityAdapterV2.test.ts, OrchestratorReplayFixtures.integration.test.ts (which includes the Grok and registry replay fixtures: background bash, background subagent, monitor, subagent lineage) and OrchestratorReplayRecovery.integration.test.ts: 6 files, 244 tests pass. The replay fixtures are unchanged.
• vp exec tsc --noEmit -p . in apps/server: no errors.
• vp lint on the touched file: only the existing unused NodePath import warning.
• Not run: repo-wide checks, live ACP agents.

Model: Claude Opus 5.5 (Claude Code)

🤖 Generated with Claude Code

---

[macroscopeapp]

Approvability

Verdict: Approved at 9d8ca68

Macroscope's review found this PR approvable — This is a localized ACP adapter optimization that bounds ordinal-map lifetime and removes quadratic map copying while preserving per-turn and per-subagent ordinal allocation. It introduces no new capability, schema, product-default, security, billing, or static-analysis changes.

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

[macroscopeapp]

Scoping ordinals to each turn changes the result when a native item ID recurs in a later turn: it now receives that turn's ordinal instead of the earlier turn's. Could you add a focused adapter regression test covering reused IDs across turns and stable ordinals on repeated updates? This backend behavior change needs test coverage.

^{Posted via Macroscope — Effect Service Conventions}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ No successful main baseline artifact is available yet. This run establishes the initial measurement.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	—	4.9 KiB	—	6.8 KiB	✅
Codex	Thread snapshot wire	—	3.7 KiB	—	4.9 KiB	✅
Codex	Live turn WebSocket wire	—	1.2 KiB	—	2.0 KiB	✅
Codex	Live turn WebSocket decoded	—	20.4 KiB	—	29.3 KiB	✅
Codex	Live turn messages	—	2	—	8	✅
Claude	Total thread wire	—	4.9 KiB	—	6.8 KiB	✅
Claude	Thread snapshot wire	—	3.7 KiB	—	4.9 KiB	✅
Claude	Live turn WebSocket wire	—	1.2 KiB	—	2.0 KiB	✅
Claude	Live turn WebSocket decoded	—	20.8 KiB	—	29.3 KiB	✅
Claude	Live turn messages	—	2	—	8	✅

Baseline: unavailable · PR result: 9d8ca68 · 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: 106.1 KiB
• Claude decoded thread snapshot: 106.4 KiB

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