agentHost/claude: Phase 5 + Phase 6 — IAgent provider, sendMessage (re-land) by TylerLeonhardt · Pull Request #314533 · microsoft/vscode

TylerLeonhardt · 2026-05-05T20:10:09Z

Summary

Re-submits the Claude IAgent provider (Phase 5 + Phase 6) for the Agent Host. The original Phase 6 PR (#314216) was reverted in #314358; this PR re-lands that work plus follow-ups: PR-review fixes, a stale-dep cleanup, the SDK bump from 0.2.112 → 0.2.128, and — most importantly — a switch from bundling the SDK to loading it from a user-specified path at runtime.

Implements phase5-plan.md and phase6-plan.md of the Claude-in-Agent-Host roadmap. Tools remain denied (canUseTool: 'deny') — Phase 7 owns tool wiring.

Commits in this stack

agentHost/claude: post-Phase-4 cleanup
agentHost/claude: lock Phase 5 implementation plan
agentHost/claude: Phase 5 — IAgent provider skeleton
agentHost/claude: Phase 6 — sendMessage, single-turn, no tools
agentHost/claude: address PR review (listSessions resilience, dispose abort) — new since agentHost/claude: Phase 6 — sendMessage, single-turn, no tools #314216
Drop stale @anthropic-ai/sandbox-runtime dep from merge resolution — new
Bump @anthropic-ai/claude-agent-sdk 0.2.112 → 0.2.128 — new
agentHost/claude: address PR review — new
agentHost/claude: don't bundle claude-agent-sdk; load it from a user-specified path — new

Phase 6 highlights

Materialization race safety. ClaudeAgent.sendMessage routes through _sessionSequencer so concurrent first sends collapse into one materialize + N ordered sends — no double-fork of the SDK subprocess.
Two abort gates in _materializeProvisional. Post-sdk.startup() and post-_writeCustomizationDirectory() checks ensure that a disposeSession landing mid-materialize disposes the produced WarmQuery instead of leaking a subprocess into _sessions.
Session ↔ mapper split. ClaudeAgentSession owns the prompt iterator + per-turn deferreds; mapSDKMessageToAgentSignals is a pure reducer with state owned by the session — independently testable, no shared mutable state.
IClaudeAgentSdkService.startup() added alongside listSessions() so the agent can be tested with a fake SDK without spawning a real subprocess.
Reducer ordering invariant. SessionResponsePart is allocated on content_block_start before any delta arrives, so the protocol's "part-must-precede-delta" invariant holds by construction.

Post-revert follow-ups

PR-review fixes (3a4bcadb6cd):

listSessions wraps _sdkService.listSessions() in try/catch. AgentService.listSessions fans out across providers via Promise.all; an SDK dynamic-import failure would otherwise nuke every other provider's session list. Now logs and returns [].
dispose aborts _provisionalSessions AbortControllers before super.dispose(). Previously a racing first sendMessage parked inside _writeCustomizationDirectory could pass the materialize abort gates and call _sessions.set on a disposed DisposableMap, orphaning the WarmQuery. Aborting first triggers the existing post-customization-write abort gate, which asyncDisposes the WarmQuery.

Don't bundle the SDK; load it from a user-specified path (8b298274182):

After the SDK bump, the new SDK ships ~200MB of native claude executables (eight optional platform-specific packages: @anthropic-ai/claude-agent-sdk-{darwin,win32,linux}-{x64,arm64}{,-musl}). Bundling that into VS Code triggered a chain of build issues (universal-app cross-copy, DMG sizing, asar-unpack, cglicenses) and grew the install footprint significantly. After review, we don't bundle the SDK at all and load it from a user-installed location instead.

Note: The chat.agentHost.claudeAgent.path setting is experimental and intended for local testing only while Phase 6+ is under development. The setting is gated to non-stable builds (product.quality !== 'stable') and tagged experimental/advanced. The long-term delivery mechanism for the Claude SDK will be the Extension Marketplace — once that's wired up, this setting will be removed.

@anthropic-ai/claude-agent-sdk is moved from dependencies to devDependencies. import type keeps typechecking; the package and its eight platform sub-packages are not installed in production.
New string setting chat.agentHost.claudeAgent.path replaces the boolean chat.agentHost.claudeAgent.enabled. When the value is non-empty, the Claude provider registers; otherwise it does not.
The setting is forwarded to the agent host via VSCODE_AGENT_HOST_CLAUDE_SDK_PATH (replacing VSCODE_AGENT_HOST_ENABLE_CLAUDE). agentHostServerMain exposes --claude-sdk-path <path> in place of --enable-claude-agent.
ClaudeAgentSdkService._loadSdk() reads the env var and dynamic-imports from it. If the path is a directory, the package's main entry is resolved from package.json (exports['.'] / main) before the import — Node ESM does not support directory imports of file:// URLs.
All bundling-related plumbing is removed: build/lib/claudeAgentSdk.ts, the gulpfile filters in gulpfile.{vscode,reh}.ts, the @anthropic-ai/claude-agent-sdk-* asar-unpack glob, the universal-app cross-copy + filesToSkip + x64ArchFiles entries, the verify-macho skip patterns, the DMG size = '2g' bump, and the nine cglicenses.json entries.

Tests

45 unit tests in claudeAgent.test.ts covering materialize lifecycle (provisional/abort/double-materialize), prompt iterator pumping, dispose ↔ shutdown ↔ send race matrix, the C1 dispose-during-customization-write regression, mapper output for text + thinking + result, the SDK loader's cache-and-log-once contract, plus the two new regressions added in 3a4bcadb6cd.
2 proxy-backed integration tests in claudeAgent.integrationTest.ts: real ClaudeProxyService (loopback HTTP server) + stubbed Copilot CAPI streaming canned MessageStreamEvents — exercises the full agent → proxy → CAPI → SSE → agent pipeline including the ANTHROPIC_AUTH_TOKEN nonce contract.

Out of scope (deferred per roadmap)

Tools beyond canUseTool: 'deny' (Phase 7)
Fork via config.fork (Phase 6.5 — needs message-UUID lookup)
Mode dropdown / branch UI (Phase 7+)
Transcript reconstruction in getSessionMessages (Phase 13)
Subagent + permission-request signals (Phase 7)

- roadmap.md: mark Phase 4 as DONE, link merged PR #313780. - phase4-plan.md: record live-system smoke completion in §7.8; disabled-gate run skipped (covered by unit tests + env-var guard). - claudeAgent.test.ts: drop gratuitous 'as unknown as' cast in the CCAModel fixture (literal already matches CCAModelBilling exactly; plan §7.4 forbids unsafe casts in tests).

Handoff plan for Phase 5 (replace 7 throwing stubs in claudeAgent.ts). Locked against post-PR-#313841 reality (provisional sessions, onDidMaterializeSession, 30s empty-session GC) and the IAgent contract on origin/main. Decisions captured: - Non-fork createSession is synchronous and in-memory; fork deferred to Phase 6 (throws TODO). - IClaudeAgentSdkService surface mirrors IAgent (no dir parameter on listSessions); SDK loader caches resolved module, retries on failure, logs once. - listSessions joins SDK enumeration with workbench session DB metadata via ISessionDataService; per-entry try/catch resilience. - shutdown() routes per-session teardown through the same SequencerByKey<string> used by disposeSession() so concurrent shutdown/disposeSession cannot double-dispose a wrapper in Phase 6. - 14 unit tests defined (12 lifecycle + 2 resolved-config), including log-once contract and shutdown/disposeSession race guard.

Lands the ClaudeAgent IAgent provider behind the 'chat.agentHost.claudeAgent.enabled' setting (env gate VSCODE_AGENT_HOST_ENABLE_CLAUDE=1). Pins @anthropic-ai/claude-agent-sdk@0.2.112 in workspace + remote/. Implemented in this phase: * createSession - non-fork, in-memory wrapper only. Honors config.session for restore. The fork path and SDK session creation are deferred to Phase 6. * listSessions - SDK is source of truth; per-session DB read is a best-effort overlay (failure never excludes an entry). * disposeSession / shutdown - routed through a per-session SequencerByKey to serialize teardown. * getDescriptor, getProtectedResources, models, onDidSessionProgress, setClientCustomizations, setClientTools, onClientToolCallComplete, setCustomizationEnabled, authenticate, respondTo*Request - minimal Phase-5 wiring. Stubbed for Phase 6 (throw async 'TODO: Phase 6'): sendMessage, abortSession, changeModel, getSessionMessages, plus the createSession fork path. Tests: 29 unit tests in claudeAgent.test.ts cover the createSession restore-id path, listSessions overlay resilience, dispose serialization, and stub surfaces. Note: provisional / onDidMaterializeSession is intentionally omitted in Phase 5 (see plan section 3.3.1) - the workbench needs an immediate sessionAdded until the agent has real materialization work, which arrives in Phase 6 alongside SDK query() startup.

Implements the Phase 6 plan: provisional sessions materialize on first sendMessage, route a single-turn prompt through the Anthropic Claude Agent SDK's WarmQuery, and stream SDKMessages back as protocol AgentSignals via a pure mapSDKMessageToAgentSignals reducer. Tools remain denied (canUseTool: 'deny'); fork moves to Phase 6.5; Plan Mode UI moves to Phase 7. Highlights: - ClaudeAgent.sendMessage routes through _sessionSequencer to collapse concurrent first sends into one materialize + N ordered sends. - _materializeProvisional has two abort gates (post-startup + post-customizationDirectory write) so disposeSession landing mid-materialize cannot leak a WarmQuery subprocess. - ClaudeAgentSession owns the prompt iterator + per-turn deferreds; mapSDKMessageToAgentSignals is a pure reducer with state owned by the session. - IClaudeAgentSdkService gains startup() alongside listSessions(). Tests: 43 unit + 2 proxy-backed integration. Council-review fixes (C1 dispose race, C2 missing integration test, S1 cwd-less ratification) included.

… abort) Two Copilot-reviewer comments on #314216: 1. listSessions: wrap _sdkService.listSessions() in try/catch. AgentService.listSessions fans out across providers via Promise.all; an SDK dynamic-import failure would otherwise nuke every other provider's session list. Now logs and returns []. 2. dispose: abort _provisionalSessions AbortControllers before super.dispose(). Previously a racing first sendMessage parked inside _writeCustomizationDirectory could pass the materialize abort gates and call _sessions.set on a disposed DisposableMap, orphaning the WarmQuery. Aborting first triggers the existing post-customization-write abort gate, which asyncDisposes the WarmQuery. Tests: 2 new regressions (listSessions empty on SDK throw; agent.dispose() during racing materialize disposes the WarmQuery). 45/45 unit + 2/2 integration pass.

The new SDK no longer vendors native binaries inside the main package. It now ships a ~200MB `claude` executable per platform via 8 optional platform-specific packages, mirroring the @github/copilot pattern: @anthropic-ai/claude-agent-sdk-{darwin,win32}-{x64,arm64} @anthropic-ai/claude-agent-sdk-linux-{x64,arm64}{,-musl} The SDK loader picks the right package at runtime via process.platform /process.arch (and tries -musl first on linux). To strip off-target packages from the build output: - build/lib/claudeAgentSdk.ts mirrors build/lib/copilot.ts - gulpfile.vscode.ts and gulpfile.reh.ts apply the filter alongside the existing copilot one - gulpfile.vscode.ts asar-unpacks @anthropic-ai/claude-agent-sdk-* so the executable stays on disk (asar would break exec permissions) - alpine-{arch} maps to linux-{arch}-musl (claude is statically linked against libc and must match the host) cglicenses.json gets 8 new entries mirroring the parent SDK's "© Anthropic PBC. All rights reserved." text. The new SDK Query interface adds a `readFile` method; FakeQuery and RoundTripQuery test doubles get matching stubs. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Copilot

Pull request overview

Re-lands the Claude agent-host provider’s Phase 5/6 work: provisional Claude sessions now materialize on first send, route prompts through the Claude SDK, and stream session signals back through the agent-host protocol. It also wires the new SDK dependency into agent-host packaging/build flows and updates the phase/smoke documentation around the Claude rollout.

Changes:

Add the Claude SDK service, session wrapper, prompt resolver, event mapper, and sendMessage/materialization flow in the agent-host Claude provider.
Register the Claude SDK service in both agent-host entrypoints and package the platform-specific Claude SDK binaries in desktop/REH builds.
Add proxy-backed integration coverage and refresh the Phase 4/5/6 roadmap, plan, and smoke-test documentation.

Show a summary per file

File	Description
`src/vs/platform/agentHost/test/node/claudeAgent.integrationTest.ts`	Adds proxy-backed integration tests for Claude agent materialization and proxy auth.
`src/vs/platform/agentHost/node/claude/smoke.md`	Expands the manual smoke plan for Phase 6 sendMessage/materialization behavior.
`src/vs/platform/agentHost/node/claude/roadmap.md`	Marks Phase 4 as completed in the Claude roadmap.
`src/vs/platform/agentHost/node/claude/phase6-plan.md`	Adds the detailed Phase 6 implementation plan.
`src/vs/platform/agentHost/node/claude/phase5-plan.md`	Adds the detailed Phase 5 implementation plan.
`src/vs/platform/agentHost/node/claude/phase4-plan.md`	Records the completed Phase 4 smoke evidence.
`src/vs/platform/agentHost/node/claude/claudePromptResolver.ts`	Introduces prompt/attachment-to-Claude-content-block conversion.
`src/vs/platform/agentHost/node/claude/claudeMapSessionEvents.ts`	Maps Claude SDK stream/result messages into agent-host session actions.
`src/vs/platform/agentHost/node/claude/claudeAgentSession.ts`	Adds the per-session query owner, prompt queue, and SDK message pump.
`src/vs/platform/agentHost/node/claude/claudeAgentSdkService.ts`	Adds the lazy Claude SDK wrapper with `listSessions()` and `startup()`.
`src/vs/platform/agentHost/node/claude/claudeAgent.ts`	Implements provisional session creation, materialization, config schema, session listing, and real `sendMessage`.
`src/vs/platform/agentHost/node/agentHostServerMain.ts`	Registers the Claude SDK service in the standalone agent-host server entrypoint.
`src/vs/platform/agentHost/node/agentHostMain.ts`	Registers the Claude SDK service in the regular agent-host entrypoint.
`src/vs/platform/agentHost/common/claudeSessionConfigKeys.ts`	Adds Claude-specific session config keys/types.
`remote/package.json`	Adds the Claude SDK dependency to the REH package manifest.
`package.json`	Adds the Claude SDK dependency to the root product manifest.
`package-lock.json`	Locks the Claude SDK and its platform-specific optional packages.
`eslint.config.js`	Allows importing the Claude SDK from the agent-host code.
`cglicenses.json`	Adds license overrides for the new Claude SDK platform packages.
`build/lib/claudeAgentSdk.ts`	Adds build helpers for keeping only the target Claude SDK platform package.
`build/gulpfile.vscode.ts`	Applies Claude SDK filtering/unpacking in desktop packaging.
`build/gulpfile.reh.ts`	Applies Claude SDK filtering in REH packaging.

Copilot's findings

Files not reviewed (1)

remote/package-lock.json: Language not supported

Files reviewed: 22/24 changed files
Comments generated: 4

…universal app Mirror the existing @github/copilot pattern in create-universal-app.ts. Each darwin arch build only contains its own platform package (npm only installs the optionalDependency matching the host CPU), so the universal merger fails with files unique to one side. Cross-copy node_modules/@anthropic-ai/claude-agent-sdk-{darwin-x64,darwin-arm64}/ between the two builds, skip them from the equality comparison, and tag the per-arch `claude` executable as arch-specific so the merger keeps both. Also extend verify-macho's skip list to ignore the single-arch `claude` binaries inside the universal app.

# Conflicts: # package-lock.json

- claudeAgent: rewrite `rgPath` from node_modules.asar → node_modules.asar.unpacked before putting it on the Claude subprocess PATH. Mirrors copilotAgent.ts and the workbench search engine helpers; without this, packaged builds advertise a path that doesn't exist on disk. - cglicenses.json: add "// Reason: …" justification comments to the parent @anthropic-ai/claude-agent-sdk entry and each of the 8 new platform sub-packages, matching the file's convention.

The universal macOS app now carries platform-specific binaries for both x64 and arm64 — @github/copilot-darwin-* (~128MB each) and the new @anthropic-ai/claude-agent-sdk-darwin-* (~207MB each) — so the source filesystem the DMG is built from has crossed 1GB. dmgbuild fails with `No space left on device` when ditto can't fit the app inside the volume. Output DMG is LZMA-compressed (format = 'ULMO') so this only changes the build-time staging size, not the shipped artifact size.

…specified path Move `@anthropic-ai/claude-agent-sdk` from `dependencies` to `devDependencies` so the ~200MB-per-arch platform binaries are no longer shipped with VS Code. The SDK becomes opt-in and externally-installed. User-facing surface: - Replace boolean setting `chat.agentHost.claudeAgent.enabled` with string setting `chat.agentHost.claudeAgent.path`. When the setting is non-empty, the Claude provider registers; when empty (the default), it does not. - The setting value is forwarded to the agent host via the `VSCODE_AGENT_HOST_CLAUDE_SDK_PATH` env var (replacing the previous `VSCODE_AGENT_HOST_ENABLE_CLAUDE` flag). - `agentHostServerMain` exposes a `--claude-sdk-path <path>` CLI flag in place of the previous `--enable-claude-agent` flag. Runtime loader: - `ClaudeAgentSdkService._loadSdk()` now reads the env var and dynamic-imports from there. If the path is a directory, the package's main entry is resolved from `package.json` (`exports['.']` / `main`) before the import — Node ESM does not support directory imports of `file://` URLs. Build/packaging cleanup (no longer needed once the SDK is gone from production deps): - Drop `build/lib/claudeAgentSdk.ts` and its callers in `gulpfile.{vscode,reh}.ts`. - Drop the `@anthropic-ai/claude-agent-sdk-*` glob from the asar-unpack list in `gulpfile.vscode.ts`. - Revert universal-app cross-copy + filesToSkip + x64ArchFiles entries in `build/darwin/create-universal-app.ts` and the corresponding skip patterns in `build/darwin/verify-macho.ts`. - Revert DMG volume size from 2g back to 1g in `build/darwin/dmg-settings.py.template` (was bumped earlier in this branch to fit the bundled SDK; no longer needed). - Remove the 9 `@anthropic-ai/claude-agent-sdk*` entries from `cglicenses.json` (no longer shipped, no manifest to license). Type imports of `@anthropic-ai/claude-agent-sdk` continue to work via the devDependency, so source code that does `import type` from the package still typechecks.

TylerLeonhardt and others added 7 commits May 5, 2026 10:47

Drop stale @anthropic-ai/sandbox-runtime dep from merge resolution

096a18f

Copilot AI review requested due to automatic review settings May 5, 2026 20:10

Copilot started reviewing on behalf of TylerLeonhardt May 5, 2026 20:11 View session

microsoft deleted a comment from github-actions Bot May 5, 2026

Copilot AI reviewed May 5, 2026

View reviewed changes

Comment thread src/vs/platform/agentHost/node/claude/claudeAgentSession.ts

Comment thread cglicenses.json Outdated

Comment thread build/lib/claudeAgentSdk.ts Outdated

Comment thread src/vs/platform/agentHost/node/claude/claudeAgent.ts Outdated

vs-code-engineering Bot assigned TylerLeonhardt May 5, 2026

austindyoung mentioned this pull request May 5, 2026

[mirror] microsoft/vscode#314533 agentHost/claude: Phase 5 + Phase 6 — IAgent provider, sendMessage (re-land) austindyoung/vscode-driftfence-mirror-fork#1971

Open

roblourens previously approved these changes May 6, 2026

View reviewed changes

TylerLeonhardt dismissed roblourens’s stale review via 3a823b2 May 6, 2026 16:21

TylerLeonhardt added 3 commits May 6, 2026 11:00

Merge remote-tracking branch 'origin/main' into tyler/claude-ahp-phase6

1bc6b0f

# Conflicts: # package-lock.json

TylerLeonhardt force-pushed the tyler/claude-ahp-phase6 branch from 3c602a0 to 8b29827 Compare May 6, 2026 21:19

TylerLeonhardt force-pushed the tyler/claude-ahp-phase6 branch from 8b29827 to d8fa2b6 Compare May 6, 2026 22:01

roblourens previously approved these changes May 6, 2026

View reviewed changes

revert cglicenses change

ff1a040

TylerLeonhardt dismissed roblourens’s stale review via ff1a040 May 6, 2026 22:43

dmitrivMS approved these changes May 6, 2026

View reviewed changes

TylerLeonhardt merged commit c9b1f88 into main May 6, 2026
26 checks passed

TylerLeonhardt deleted the tyler/claude-ahp-phase6 branch May 6, 2026 23:48

vs-code-engineering Bot added this to the 1.120.0 milestone May 6, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

agentHost/claude: Phase 5 + Phase 6 — IAgent provider, sendMessage (re-land)#314533

agentHost/claude: Phase 5 + Phase 6 — IAgent provider, sendMessage (re-land)#314533
TylerLeonhardt merged 13 commits intomainfrom
tyler/claude-ahp-phase6

TylerLeonhardt commented May 5, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

TylerLeonhardt commented May 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Commits in this stack

Phase 6 highlights

Post-revert follow-ups

Tests

Out of scope (deferred per roadmap)

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Copilot's findings

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

TylerLeonhardt commented May 5, 2026 •

edited

Loading