agentHost: batch AHP log writes and skip URI deep clone

[roblourens]

Two perf fixes for AhpJsonlLogger, the per-message JSONL transport logger used by AHP-based agent host connections.

Why

CPU traces of the agents window under heavy AHP traffic showed ~23% of wall time in MajorGC (3.3 s of a 14.16 s trace) on the sessions/agents renderer. The hot stack was:

VSBuffer.set / concat
 FileService.writeFile
 AhpJsonlLogger._appendLine

main writeFile IPC. Each call allocated a fresh VSBuffer, marshalled it through the IPC channel proxy, and got a reply buffer back. With sustained rates of 1300 messages/s the GC simply couldn't keep up.500

A second cost was _replaceUris, which deep-cloned every JSON-RPC message before JSON.stringify so it could rewrite URI instances to once per log line.strings

What changed

1. Batched writes via natural backpressure. log() now appends to a _pending: VSBuffer[] and schedules a single drain on the write queue. While a writeFile is in flight, all subsequent log calls accumulate and land in the next drain via VSBuffer. capped at 1 MiB per write so the concat itself doesn't create new GC pressure. Rotation is still checked per-entry inside the batch. flush()and ordering semantics are preserved via a_drainScheduledinvariant: pending entries always have a drain on the queue, soawait this._queue is sufficient. **Zero added batching only happens when writes are already backed up.latency** concat

2. Replaced _replaceUris deep clone with a JSON.stringify replacer. URI.prototype.toJSON() already stamps its output with $mid: MarshalledId.Uri. The new replacer detects that marker (guarded by isUriComponents to avoid false positives on user payloads that happen to carry $mid: 1) and rewrites to the canonical URI string via URI.revive(...).toString(). The previous comment in the code warned against using a replacer because toJSON runs but that's actually what makes this approach work: we use the post-toJSON marker as our signal.first

Tests

Added 7 tests on top of the existing 2:

• Coalescence: 50 sync log() calls produce all 50 lines in order but far fewer writes
• Flush ordering across interleaved log/flush/log calls
• URI replacer: top-level URI, nested in arrays/objects, raw UriComponents from a prior toJSON(), URI-shaped objects without $mid (left alone), and $mid: 1 on a non-UriComponents payload (left alone)

All 10 AhpJsonlLogger tests pass and the existing localAhpJsonlLogging tests are unaffected.

Notes for reviewers

emit) with no application-level handlers above, which suggests the inbound load is mostly *replies* to outbound requests rather than a stream of notifications. Identifying the producer would benefit from per-method counters in AhpJsonlLogger.log`.

• The 1 MiB batch cap and the per-entry rotation check are a single oversized message can still rotate mid-batch, matching previous behavior.independent

(Written by Copilot)

[Copilot]

Pull request overview

This PR improves AhpJsonlLogger performance for high-volume Agent Host Protocol traffic by reducing per-message write overhead and avoiding a full deep clone for URI serialization.

Changes:

• Batches pending JSONL log buffers into capped writes while preserving ordering and rotation behavior.
• Replaces URI deep-cloning with a JSON.stringify replacer that detects marshalled URI components.
• Adds tests for batching, flush ordering, and URI serialization behavior.

Show a summary per file

File	Description
src/vs/platform/agentHost/common/ahpJsonlLogger.ts	Implements batched writes and URI-aware JSON serialization.
src/vs/platform/agentHost/test/common/ahpJsonlLogger.test.ts	Adds coverage for batching, flush ordering, and URI stringification cases.

Copilot's findings

• Files reviewed: 2/2 changed files
• Comments generated: 1