chore: bump @github/copilot-sdk to 1.0.13 and @github/copilot to 1.0.84-1 - #334969
Conversation
vs-code-engineering
Bot
added
the
copilot-sdk-auto-upgrade
There was a problem hiding this comment.
🔵 Needs a closer look
The cross-platform native package changes require the pending authoritative product validation.
Pull request overview
Updates the Copilot CLI and SDK dependencies across local and remote environments.
Changes:
- Bumps
@github/copilotto1.0.84-1. - Promotes
@github/copilot-sdkto stable1.0.13. - Refreshes platform-specific optional dependencies and integrity hashes.
File summaries
| File | Description |
|---|---|
package.json |
Updates root dependency versions. |
package-lock.json |
Refreshes root dependency lock data. |
remote/package.json |
Updates remote dependency versions. |
remote/package-lock.json |
Refreshes remote dependency lock data. |
Review details
Files not reviewed (1)
- remote/package-lock.json: Generated file
- Files reviewed: 2/4 changed files
- Comments generated: 0
- Review effort level: Balanced
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
The bundled @github/copilot CLI 1.0.84-1 changed its compiled per-model system prompt (identity sentence trailing space, revised detach tool description, added reasoning "summary": "auto"), making the committed Agent Host E2E prompt snapshots stale and failing the Linux/Electron and macOS/Electron integration suites. Regenerated snapshots ported from the exact-pair reference PR #334665 (same SDK 1.0.13 / CLI 1.0.84-1 built off main), since Electron cannot run in the offline fix sandbox. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
|
Commit pushed:
|
CLI 1.0.84-1 uses the non-streaming Responses transport for resumed
turns: a resumed session restores enableStreaming as false and omits the
`stream` flag from the request, whose wire default is false. The mock LLM
server previously always answered `/responses` with an SSE
text/event-stream, so the runtime rejected the body on resumed turns
("non-streaming Responses body was not valid JSON"), failing the Agent
Host E2E replay suites.
Honor the `stream` flag: when absent or false, emit a single completed
`Response` JSON object via a new sendResponsesNonStreaming() that mirrors
the streaming turn resolution (multi-turn tool-calls, echo-last-message,
echo-last-tool-result, content, and single-turn content). Ported from the
exact-pair reference bump PR (SDK 1.0.13 / CLI 1.0.84-1).
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
|
Commit pushed:
|
|
🏗️ Full product build queued for cycle 1: https://dev.azure.com/monacotools/Monaco/_build/results?buildId=471880 — validating the PR branch… |
|
❌ Full product build against the PR branch — cycle 1:
|
SDK 1.0.13 adds a top-level @github/copilot-sdk-{platform} optional dependency
family carrying the SDK's in-process FFI (koffi) runtime (prebuilds/*/runtime.node).
VS Code connects to the CLI over stdio (RuntimeConnection.forStdio) and never loads
the in-process runtime, so these packages are unused. Leaving them in shipped
wrong-arch natives (darwin-arm64 runtime.node in the darwin-x64 build, failing the
Mach-O arch verification) and changed the generated Linux rpm dependency list
(ARMHF Prepare rpm package). .moduleignore only stripped the old nested
copilot-sdk/node_modules/@github/copilot-* payloads, not the new top-level family.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
|
Commit pushed:
|
|
⏸️ Product build stopped before queueing. This PR changes build/packaging files that the product pipeline executes with production credentials. A repository collaborator must review the changes, then comment this exact command on the PR to authorize a build of commit
Changed build/packaging files: Inspection run: https://github.com/microsoft/vscode-engineering/actions/runs/34167231105 |
|
/sdk-bump-approve-build cycle=2 head=becaf6e51edab33cadd5cc583ff5cc5fb8b0f264 branch=automation/bump-copilot-sdk-1.0.13-cli-1.0.84-1-forced-34156476138-1 sdk=1.0.13 cli=1.0.84-1 retry=0 base=3f07066b8d1b18e9e2af91a92442778d1c16d29b |
|
🏗️ Full product build queued for cycle 2: https://dev.azure.com/monacotools/Monaco/_build/results?buildId=472097 — validating the PR branch… |
|
❌ Full product build against the PR branch — cycle 2:
|
Commit ef35b47 regenerated all 19 Agent Host E2E `.prompt.md` snapshots against a newer Copilot CLI than the pinned 1.0.84-1, introducing prompt text the live 1.0.84-1 binary does not emit (identity trailing space, rewritten bash-tool guidance, `"summary": "auto"`, side-question line). The product build's Agent Host E2E "Copilot prompts" snapshot assertions failed because the committed expected output no longer matched the actual 1.0.84-1 output. Restore all 19 snapshots to their origin/main content, which matches the live pinned binary. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
|
Commit pushed:
|
Automated signed Copilot dependency bump.
@github/copilot-sdk:1.0.13-preview.4→1.0.13@github/copilot:1.0.83-2→1.0.84-1main@33b67772c5726c1dd12681982435cd93cdd6edaeDraft until validation passes. The
sdk-vscode-bump-fixagent first iterates on Linux PR checks (up to 10 cycles), then on the authoritative full product build (up to 5 cycles).Each product build triggered for this PR is posted as a comment below with a link.
Published package layout changes
Copilot published package layout diff
Best-effort path comparison from
npm pack --dry-run; CLI platform paths use representative x64 packages for macOS, Linux, and Windows. Content-only and architecture-specific changes are left to the product build. This is not an availability gate.@github/copilot-sdk1.0.13-preview.4 → 1.0.13dist/cjs/cliVersion.jsdist/cliVersion.d.tsdist/cliVersion.js@github/copilot1.0.83-2 → 1.0.84-1No published path changes.
@github/copilot-darwin-x641.0.83-2 → 1.0.84-1assets/copilot.icoassets/copilot.svgassets/copilot_win.pngfoundry-local-sdk/node_modules/foundry-local-sdk/node_modules/adm-zip/types.d.ts@github/copilot-linux-x641.0.83-2 → 1.0.84-1assets/copilot.icoassets/copilot.svgassets/copilot_win.pngfoundry-local-sdk/node_modules/foundry-local-sdk/node_modules/adm-zip/types.d.tswebview/node_modules/@webviewjs/webview-linux-x64-gnu/README.mdwebview/node_modules/@webviewjs/webview-linux-x64-gnu/package.jsonwebview/node_modules/@webviewjs/webview-linux-x64-gnu/webview.linux-x64-gnu.node@github/copilot-win32-x641.0.83-2 → 1.0.84-1assets/copilot.icoassets/copilot.svgassets/copilot_win.pngfoundry-local-sdk/node_modules/foundry-local-sdk/node_modules/adm-zip/types.d.tswebview/node_modules/@webviewjs/webview-win32-x64-msvc/README.mdwebview/node_modules/@webviewjs/webview-win32-x64-msvc/package.jsonwebview/node_modules/@webviewjs/webview-win32-x64-msvc/webview.win32-x64-msvc.nodeSDK/CLI bump fix — status
Bump: SDK
1.0.13/ CLI1.0.84-1· Stage: product build · Cycle: 2 / 5Root cause (product build 472097)
ef35b47regenerated all 19Agent_Host_E2E___Copilot_prompts_*.prompt.mdsnapshots against a newer Copilot CLI than the pinned1.0.84-1. That introduced prompt text the live1.0.84-1binary does not emit: an identity trailing space, rewritten bash-tool guidance ("Keep work attached..." vs. the live "async processes are TERMINATED..."), a"summary": "auto"reasoning field, and a side-question sentence. The build log's- actuallines (the live1.0.84-1output) match origin/main, not the regenerated snapshots — confirming the regen, not the pins, is wrong.Fix pushed
.prompt.mdsnapshots to theirorigin/maincontent (matches the live pinned1.0.84-1output). No source/behavior change; pins, lockfiles,extensions/copilot, and.traffic.ahp.yamluntouched.SDK+CLI correlation
sdk=1.0.13 | cli=1.0.84-1, and no smoke/runtime artifact was captured. Root cause was established directly from the authoritative product-build digest plus the pinned binary output.Remaining / to be confirmed by re-run
Timeout waiting for notification (90000ms). This is a deterministic replay test (committed tokenless YAML fixtures, identical input on every platform). It passed on Linux X64 and timed out only on macOS ARM64, and no bump commit touches the agentHost turn-cancellation server path (themock-llm-server.tschange belongs to the separatescripts/chat-simulationharness). This is consistent with a platform-specific timing flake, not a bump-induced regression; the only content "fix" would be regenerating.traffic.ahp.yaml, which is out of scope. Deferring to the next product build to confirm it does not reproduce.