feat: keep Copilot BYOK agent turns alive in server Agent Host

[khanhkit]

Fixes #8046

Motivation

code-server is a remote IDE, so long-running agent work should be owned by a server-side lifecycle rather than by a browser tab. For direct-capable Copilot Chat BYOK providers, this change lets the remote Agent Host own model execution so an active turn is not tied to renderer/tab lifetime.

Design

• Direct-capable BYOK providers are synchronized to Agent Host as bounded, ephemeral descriptors and registered directly with the server-side Copilot SDK.
• Provider configurations that cannot be represented faithfully stay on the existing renderer bridge rather than being silently normalized.
• Credential-bearing syncEphemeralByokModels data stays process-memory-only and is centrally redacted before AHP JSONL serialization.
• Model capability metadata (vision, tool calling, reasoning effort) is preserved.
• Agent Host Changes-pill defaults follow isolation when there is no provider override: folder -> Session Changes, worktree -> Branch Changes.
• Native Copilot authentication semantics are not bypassed; this change is for the direct BYOK path.

Files changed

• patches/copilot-background-byok.diff
• patches/copilot-agenthost-session-changes.diff
• patches/series

Current VS Code 1.141 verification

Current PR head: 73f71d62f4b26c04dd50b595a3837231a1b5b6aa

Pinned VS Code gitlink: 2a59476c9bfcb90b3ddc372c36762471b7dfad1c (VS Code 1.141)

Executed on the 1.141 port:

• full code-server patch series clean apply: 31/31;
• regenerated feature patches reproduce 22/22 target files byte-for-byte from the validated scratch tree;
• copilotSessionLauncher.test: 109/109 pass;
• byokLmBridgeRegistry.test: 11/11 pass;
• protocolServerHandler.test: 175/175 pass;
• ahpJsonlLogger.test: 14/14 pass, including the credential-redaction sentinel;
• browser AgentHostByokLmHandler: 22/22 pass;
• browser AgentHostSessionInputPills: 22/22 pass, including folder -> Session Changes.

Exact-head isolated ARM64 release verification is now GREEN for 73f71d62f4b26c04dd50b595a3837231a1b5b6aa: full build:vscode completed (including core-ci desktop/server/server-web bundles), code-server build completed, release materialized successfully, and a fresh out-of-container checksum verification passed 5897/5897 files. The built runtime reports 73f71d62f4b26c04dd50b595a3837231a1b5b6aa with Code 1.141.0.

GitHub Actions status

Upstream Build run 37796020383 was created for exact head 73f71d62f4b26c04dd50b595a3837231a1b5b6aa, but GitHub currently reports action_required with zero jobs created. The workflow is waiting for the repository's fork-workflow approval; no CI test has executed or failed yet.

Historical note

The earlier VS Code 1.140 candidate 95c0eb85d4265925751949f729851901e8e9db0d passed its isolated build, fork CI, and live background/detach acceptance. Those results are historical evidence only and are not presented as exact-head verification for this 1.141 port.

Upstream coordination

A code-server maintainer suggested upstreaming this behavior directly to VS Code rather than maintaining the server-ownership change as a code-server patch.

VS Code already tracks the exact remote-BYOK ownership/continuity problem in microsoft/vscode#340335, under the broader BYOK plan in microsoft/vscode#340086. The VS Code direction is Agent Host-owned model-provider plugins rather than a direct port of chatLanguageModels.json. The validated implementation and evidence from this PR have been posted to that issue for reference: microsoft/vscode#340335 (comment).

This PR remains open as the exact tested reference. I am not opening a competing VS Code implementation until the owner confirms which pieces should be adapted to the plugin-provider direction.

[code-asher]

Thank you for the PR! The reasoning makes sense, but this looks like a significant change. Could we propose this change upstream directly rather than try maintaining it as a patch here? I think the VS Code maintainers would find it valuable as well, since presumably Codespaces has the same problems, or anyone running regular VS Code in the web.

[khanhkit]

Exact-head VS Code 1.141 verification update:

• candidate: 73f71d62f4b26c04dd50b595a3837231a1b5b6aa
• full build:vscode: PASS (core-ci, desktop/server/server-web bundles)
• code-server build + release: PASS
• fresh release checksum verification: 5897/5897
• runtime identity: exact candidate with Code 1.141.0

Upstream Actions run 37796020383 is still action_required with zero jobs, so it is awaiting repository approval for the fork workflow; no upstream CI test has failed.

[khanhkit]

Yes — I found an existing VS Code issue that tracks the same problem directly: microsoft/vscode#340335 ("BYOK Support for Copilot Agent Host Harness"). It explicitly calls out both the Extension Host dependency and continuity across remote clients.

I added the validated code-server/web reproduction, the 1.140 live detach result, and the exact VS Code 1.141 test/build evidence there:

microsoft/vscode#340335 (comment)

Their proposed long-term direction is Plugin-provided configurable model providers, which is broader than the direct-descriptor bridge in this PR. I am leaving this PR open as a tested reference rather than opening a competing VS Code PR blindly; I can adapt/split the implementation once the VS Code owner confirms which pieces fit that direction.

[khanhkit]

Upstream coordination update and recommendation after a deeper revalidation of the VS Code/Copilot Agent Host path:

Current upstream status

I followed the upstream direction from microsoft/vscode#340335 rather than opening a competing VS Code PR blindly. The intended long-term architecture is now clear: BYOK should be provided by GitHub Copilot Plugin model providers owned by Agent Host, with provider plugins running outside individual session scope and serving models across sessions/clients.

The blocker is that this architecture is not publicly implementable yet from the VS Code side:

• microsoft/vscode#340335 remains open; its model-provider-plugin tasks are still unchecked.
• Current VS Code main revalidated at e3c2316227ee9a4b24449b779409e3ff4d0126a9 still has no modelProviders plugin-manifest contribution.
• VS Code currently still pins @github/copilot-sdk@1.0.19-preview.0.
• Public SDK stable 1.0.19, prerelease 1.0.20-preview.2, and current public SDK main still expose provider discovery/configuration through session.providers.*; there is no public global/out-of-session provider catalog/lifecycle API that implements the #340335 design.
• The newer SDK bump work and managed-plugin lifecycle work I inspected do not add that missing global model-provider contract. They therefore do not provide a safe migration target for this patch yet.
• The existing renderer BYOK bridge remains client/renderer-owned, which is exactly the continuity problem #340335 is intended to solve for remote clients.

So there is currently no minimal upstream VS Code patch I can submit that both solves the remote lifetime problem and follows the architecture the VS Code owners selected. Porting this PR's direct-provider implementation upstream unchanged would instead pre-empt/duplicate the provider-plugin architecture.

Downstream status

This PR remains the exact tested implementation for code-server:

• head: 73f71d62f4b26c04dd50b595a3837231a1b5b6aa
• GitHub currently reports the PR as mergeable.
• full VS Code 1.141 patch stack: 31/31 clean apply
• feature target reproduction: 22/22 files byte-identical
• targeted tests: launcher 109/109, registry 11/11, protocol 175/175, JSONL/redaction 14/14, browser BYOK 22/22, Changes pill 22/22
• exact-head isolated ARM64 build:vscode / core-ci desktop/server/server-web: PASS
• code-server build + release: PASS
• fresh release verification: 5897/5897 checksums, runtime reports exact 73f71d62 / Code 1.141.0

The repository Actions run 37796020383 is still action_required with zero jobs created. This remains a maintainer/admin fork-workflow approval gate, not a test failure; no upstream CI job has executed yet.

Recommendation

Given that the upstream replacement architecture is specified but the required public runtime/plugin contract is not available yet, I recommend merging this PR as a temporary code-server compatibility patch, after the repository's normal maintainer-required CI/approval gate is allowed to execute.

That gives code-server users the server-owned BYOK continuity behavior now instead of leaving the bug open for an indeterminate upstream-runtime dependency. I would keep the downstream patch narrowly scoped and avoid growing it into a parallel provider architecture while #340335 is being implemented.

Retirement commitment

This patch is intended to be temporary, not a permanent code-server fork of Copilot provider architecture.

Once VS Code/Copilot Agent Host ships the #340335-equivalent capability for the environments code-server consumes — specifically an Agent Host-owned, global/out-of-session plugin model-provider lifecycle with the required local/remote configuration path — I will revalidate behavior against code-server's remote/browser-detach acceptance case and remove/simplify the downstream patches in favor of the upstream implementation.

In other words: merge/retain this compatibility layer while the upstream environment is unavailable; retire it once upstream provides equivalent remote BYOK ownership/continuity and that path is verified in code-server.