Repository navigation
·
0 commits
to main
since this release
- Added PAC-aware system proxy routing support in the runtime, improving proxy handling in environments that use proxy auto-configuration.
- The SDK now handles tokenizer warmup and process telemetry internally.
- Remote session export refuses disabled or unresolved organization storage policy; SDK results expose export-start state and refusal warnings.
- Defer allocator reclamation until disposed session data is released on opted-in Linux/glibc hosts, and preserve requests made during trim cooldown.
- Fixed a runtime bug by adding the IPv6 unspecified address to the outbound request blocklist.
- When streamer mode is enabled,
models.listnow masks hidden model names. - Fixed custom-agent list and reload returning incomplete catalogs during overlapping reloads or provider hydration, and stale catalogs after credential invalidation.
- Fixed a bug where prompts could be altered when
/afteror/everytext included the character U+0130. - Fixed a runtime bug where
isAbortErrorcould throw when an error message was not a string. - Session metadata and listing reuse unchanged log summaries and recover after transient log-read failures.
- Built-in browser activation respects user preferences and managed policy, including after extension reconciliation.
- Removed the unsupported prePRDescription and postResult variants from generated HookType contracts.
- Preserve provider metadata in SDK model catalogs so provider information remains available when listing models.
- Added
ruleIndexto thesession.fusion_resolvedevent payload. - Improved MCP OAuth compatibility by respecting advertised account-selection support and retrying a specific rejected selection once.
- Clarified Copilot SDK documentation for pending permission recovery.
- Sandbox deny rules support glob patterns, including hidden files and home-relative patterns, without changing literal grants.
- SDK sessions now honor the host's
COPILOT_MODEL_FAMILYsetting when selecting the model family. - SDK clients now receive warnings related to live Marketplace plugins.
- Runtime now loads
.agents,.claudeskills, and.claudeagents from added directories. - Fixed an issue where
session.mcp.startServercould ignore the provided configuration; it now honors the config passed by the caller. - Runtime now applies
enabledPluginsfrom trusted non-repository folders during load, improving plugin activation consistency in Copilot SDK workflows. - Fixed a runtime issue where the grep tool now accepts flag names even when provided without a leading dash.
- Preserve ACP context tier when switching models to keep context behavior consistent across model changes.
- Local-model adoption and cleanup no longer open console windows on Windows.
- Fixed MCP OAuth URL validation to reject backslash-obfuscated dot segments.
- Improved runtime reliability: config key writes that are initially refused are now retried so updates are less likely to be lost.
- Fix missing skill instructions when explicitly approving a tool permission after restarting the runtime.
- Bugfix: INFO-level runtime logs no longer include MCP server peer information.
- Fixed an issue where attachments could be dropped when a request was retried after rate limiting.
- Normalized Responses MCP namespace names by trimming leading and trailing separators, keeping names that session history already used.
- Azure OpenAI BYOK sessions on GPT-5.6/GPT-6 no longer send
prompt_cache_breakpoint, restoring cross-turn prompt cache reuse. - Host tools that share the name of a built-in tool the session is not currently offering (for example
exit_plan_modeoutside plan mode) now run their handler instead of failing with "does not exist". - External tool calls no longer time out automatically after 30 minutes, allowing long-running human-in-the-loop workflows.
- Fix intermittent custom-agent discovery failures when initial authentication completes during an SDK session's first agent scan.
- Fixed a bug where LSP
rootUrivalues resolving outside the repository Git root were accepted; they are now rejected. - Fix a race where an SDK connection closing could temporarily prevent another client from acquiring an idle Azure ML model family.
- assistant.reasoning events now include the turnId, apiCallId and model of the model call that produced them.
- Model-switch confirmations optionally report longContextLimit when the target model's higher context tier can fit the conversation.
- Fixed a race that could leave shell or subagent completion notifications queued after an active turn finishes.
- Fixed
session.provider.addto accept whole-number token limits provided as floats (for example,1000.0). - Forked sessions now keep the source session's SQL todo list (todos and todo_deps).
- Fixed a bug where retried model calls could fail due to missing message IDs in streamed ACP chunks.
- Runtime now captures image inputs and multimodal tool results in OpenTelemetry telemetry.
- Fixed a Windows runtime bug in sandboxed child environments by standardizing PATH handling for more reliable command execution.
- Fixed an issue where MCP OAuth tokens could fail to renew after being rejected during an idle liveness probe.
- Managed settings:
managedSettings.resolve/composereturn the organization support URL as typedvalues.communication.support.url(experimental), so clients can link users to their org's help page. - Fix Python model capability token and vision limit serialization and preserve provider context-window configuration.
- Fix assistant refusal:null replay for strict OpenAI-compatible BYOK providers.
- Added experimental support for MCP skills in the runtime.
- Python: Close owned runtime stdin during force_stop() so Windows launcher descendants can exit, without blocking the asyncio event loop.
- Fix extension lifecycle state after a failed startup recovers through reload
- Python: keep native runtime callbacks alive until their connection is confirmed closed, including after reader failures.
- Align SDK OpenTelemetry capture settings with the standard GenAI environment variable while retaining legacy compatibility; usage events now expose the provider-returned response model separately from the requested model.
- Fixed a client build performance issue by memoizing trust-anchor filtering to avoid re-filtering root certificates.
- Standalone tool execution can recover after a session abort without starting another model turn.
- Node.js:
sendAndWaitnow honors timeouts above 2^31-1 ms, includingInfinity, instead of timing out after 1 ms. session.agent.listnow returns the session's built-in agents whenincludeBuiltInAgentsis true.- Fix plugin activation during managed-policy refresh and delivery of late background shell completion notifications.
- Fixed a runtime bug where an LSP query that received a
ContentModifiedresponse is now resent automatically. - Fixed a runtime OAuth callback issue by automatically retrying with a new MCP callback port if the previously saved port is refused.
- Fixed Windows managed-sandbox admission probes attempting to launch a shell builtin without a shell.
- Improved runtime behavior: the built-in GitHub MCP connection is now preserved across reloads when credentials are unchanged.
- Sampling requests from MCP servers are now cancelled immediately when the host has not registered interest in
sampling.requested, instead of waiting indefinitely. - Improved prompt-batch handling: if a user rejects one call, sibling prompts in the same batch now settle instead of staying pending.
- .NET: Export tool callback activities through the github.copilot source so instrumented tool spans have observable parents.
- On GNU/Linux, the runtime now prefers the operating system’s OpenSSL library.
- Updated sandbox backends with Linux socket and io_uring restrictions and macOS path and terminal fixes.
- Expose optional original tool-call IDs on preToolUse, postToolUse, and postToolUseFailure across all six SDKs, plus existing W3C tracing metadata in TypeScript, Python, Go, Rust, and Java.
- Return an error when sending with an unavailable model instead of silently choosing another model.
- Autopilot now stops automatically after an accepted plan is completed.
- On macOS console hosts, OneAuth sign-in dialogs now open in the foreground.
- Fix cached MCP OAuth renewal incorrectly requiring permissions from the original authorization request when the approved grant is narrower.
- Fixed a bug where a custom agent with the same name as a built-in agent could be silently shadowed; a warning is now surfaced.
- Fixed hooks.discover reporting an unsupported-effect error and omitting --plugin-dir hooks on the default runtime.
- Enabled session telemetry for eligible Bring Your Own Key (BYOK) sessions.
- Fix sandbox networking so private MCP destinations follow outbound and host rules independently of loopback permission.
- Add
mcpResultMetatotool.execution_completeresults so hosts can restore an MCP App tool result's_metaafter resume. - Improved error reporting by surfacing structured Policy Helper failures in the runtime.
- Go: A panic in a slash-command or elicitation handler no longer crashes the process; the command fails with the panic value and the elicitation is cancelled.
- Source selections now reach the model consistently in immediate steering and batched messages, just as they do when sent while idle.
- Go: A request no longer blocks forever when the connection to a CLI the SDK started closes while that CLI keeps running; it fails with "connection closed" after up to a second.
- Plan and autopilot tools now stay in the tool list after the first send in those modes, even with a replace-mode system message, so switching
agentModeper send no longer breaks prompt-cache reuse. - Fixed a runtime issue so installed plugins continue to work after moving the Copilot home directory.
- Sandbox denial suggestions recognize Go-style filesystem and Unix-socket errors when corroborated by the sandbox policy.
- Improved session continuity in the runtime: long-running sessions can now be resumed from linked event journals.
- Fixed a runtime bug where telemetry uploads are now paused while the collector is unreachable.
- Updated runtime approval behavior: shell scripts that fail to parse now require full-script approval before execution.
- Fixed Azure OpenAI BYOK session resume failures after an assistant returned no text or tool calls.
- A failed
tool.execution_completeevent now includesresult.contentwith the text sent to the model when it differs fromerror.message. - Auto metadata from
models.listandsession.model.listnow reportsproviderId: "copilot"for token-authenticated clients, and omits it for catalogs served through a request handler. - Adds experimental host-owned Microsoft Entra acquisition callbacks in client options across all six SDKs.
- Go: A panic in a CopilotRequestHandler transport, response body or WebSocket handler no longer crashes the process; the model request fails with the panic value.
- Added a new SDK callback to support Kerberos authentication in runtime integrations.
- Rust: A panic in an ExtensionLaunchProvider now answers the runtime with an error instead of delaying session creation by 15 s.
- Improved error messages when a stdio MCP server fails to start by indicating whether the command or working directory is missing.
- Honor client telemetry opt-out for OneAuth authentication diagnostics and preserve inherited environment settings in Rust in-process clients.
- Configure OneAuth MATS application identity with optional entra.matsAppName; omitting it disables MATS without disabling authentication.
- Fixed a runtime issue where connector servers could fail to start after reconcile in sessions that have no MCP servers.
- Added session.permissions.configure fields authenticatedAuthority, for owner-authenticated permission grants, and hostPolicyPermissionHandler, so a host settles its own tools in Assisted Autopilot; judge denials now tell the agent the reason.
- .NET: The default CopilotRequestHandler no longer cuts off model calls whose response takes more than 100 s to start; the runtime's request timeout governs the wait instead.
- Improved runtime reliability: session events are now persisted in publication order before shutdown.
- BYOK providers with
modelProvider: "azure_openai"(for example Azure OpenAI behind a proxy or custom domain) now omit explicit prompt-cache breakpoints, like Azure OpenAI hosts. - Rust: A panic in a CopilotRequestHandler, WebSocket handler or forwarder response hook now fails the model request instead of leaving the turn waiting for a response that never comes.
- Update Node.js dependencies and generated TypeScript map typings to include undefined for missing entries.
- Added built-in cross-provider model fallback support to improve resilience when a selected model is unavailable.
- Sessions using Gemini or Kimi models through Copilot now downscale or remove image attachments before sending a request over the 5 MiB limit, instead of failing with HTTP 413 and forcing a compaction.
- Fixed a runtime bug where
/envincorrectly showed an SDK client’s built-in plugins as installed. - All SDKs: a
capabilities.changedevent now updates only the capabilities it names; capabilities it leaves out are no longer dropped fromsession.capabilities. Java no longer fails to deliver such an event to subscribers. - Failed Connectors can recover during MCP reload after their account credentials change.
- Rust: A panic in a slash-command handler now fails the command with "command handler panicked" instead of leaving it pending.
- The standalone runtime now requires the user's consent before an extension that skips tool permission prompts, registers hooks, handles permission requests or observes them loads; a stored approval or allow-all grants it.
- Improved MCP connection behavior in the runtime by attempting HTTP MCP discovery first, then falling back to legacy discovery.
- Fixed canvas tools (
open_canvas,list_canvas_capabilities,invoke_canvas_action) failing with "Session not found" when called from a task subagent. - Fix: when a session filesystem provider fails a
viewstat with an error other than ENOENT, the tool now reports that error instead of "Path does not exist". - Fixed coauthor preference inheritance when creating subagents.
- Forward host editor identity from the SDK runtime to Copilot API requests.
- Added SDK support to list the configured model for legacy single-provider sessions.
- Fixed a bug where file edit operations could incorrectly replace non-UTF-8 characters throughout an entire file.
- Fixed a bug where SDK model switch requests could affect the wrong session; model switches now apply only to the session they were issued in.
- Fixed a runtime event-publication deadlock that could stall sessions sharing a process.
- Shell tool calls emit
session.background_tasks_changedonly when the background task list changes, instead of once per internal step. - Fixed a bug where resuming a session could crash if the session log contained a deeply nested record.
- Fixed a bug where calling
session.resumewith non-object parameters could abort the server. - Fixed a bug where refusal fallback could retry requests against a model disabled by an organization’s policy.
- Session-summary failure diagnostics omit private parent directories.
- Start model requests without waiting for host event acknowledgements, while preserving stop-hook ordering and cancellation.
- Fixed a bug where rewind now reports agent edits even when the edited file exceeds the capture size limit.
- Reduced memory growth in non-streaming sessions by limiting retained transient events.
- Session event-log reads now return promptly when a session disconnects instead of waiting for the long-poll timeout.
- Java:
ToolInvocation.getArgumentsAsnow deserializesjava.timefields such asInstantandLocalDateinstead of throwingIllegalArgumentException. - Node.js: generated types now carry the schema description as JSDoc on properties that reference a shared type, such as
HistoryRewindResult.outcome. - Fixed:
tool.execution_complete.result.citableSourcesis populated again, so sources a tool returns stay citable after a session is resumed. - Fixed a runtime bug where whole-log event snapshot refresh now returns the full history in a single native buffer.
- Fixed a bug so custom agents are now rejected when their name is empty or duplicates another agent name.
- Fixed
web_fetchto decode web pages using the page’s declared character set instead of always assuming UTF-8. - Fixed a runtime-supervised AHP host joining an application's session unregistering that session's skill provider.
- Add
session.permissions.clearResumeEscalationto drop a resumed session's allow-all or assisted posture for the current run without recording it. - Fixed a runtime issue where an unreachable MCP registry could be reported as a policy denial, improving error reporting accuracy.
- Fixed a bug in Python SDK code generation that could cause a crash when processing numeric constant values.
- Java:
PermissionRequestResultKind.hashCodenow agrees withequalsfor kinds that differ only by non-ASCII case, so such kinds work asHashMapkeys andHashSetmembers. - Fixed
session.idleafter an abort sometimes missingaborted: truewhen the previous turn's idle was still being delivered. - Fixed a bug where
getBuiltinDescriptorsresponses could violate their result contract. - Python and Go: an empty MCP server map (
mcp_servers={}, an emptyMCPServersmap) passed when creating or resuming a session is now sent instead of dropped, so resuming with it removes the session's MCP servers, as in the other SDKs. - Go:
embeddedcli.Path()now checks the bundled CLI againstCliHashbefore installing it, and an install that fails partway no longer leaves a partial file that blocks later installs. - Fixed a bug in
truncateSummarieswherekeepCountkept the oldest summaries instead of the newest as documented. - Fixed a bug where plugin installs and updates could conflict when multiple processes accessed the marketplace at the same time.
- Rust: the
suppress_resume_eventdocs now say it skips thesession.resumeevent; it never changed what happens when the session does not exist. - Report plugin activation warnings from the managed policy that applies to a session, even when the client attaches while that policy is still publishing.
- Fixed a bug where
session.metadata.snapshotcould omit the session's client name. - Node.js, Go, .NET, Java, Python and Rust: the
useLoggedInUserdocs now say that environment variable tokens such asCOPILOT_GITHUB_TOKENare ignored when it isfalse; pass the token throughgitHubTokeninstead. - Fixed a
web_fetchbug where pages hitting the 10 MiB limit inside a script, style, or comment were not reliably reported to the model as truncated. - Python: the default WebSocket forwarder relays upstream messages up to 64 MiB instead of failing above 1 MiB.
- An MCP server disabled before the session's first MCP load, including one started with a supplied config, now stays disabled through that load instead of restarting with its tools offered to the model.
- Improved
write_agenterror messages to identify conflicting selectors and explain how to proceed. - Fixed a runtime bug where local plugins with identical directory names could incorrectly share a cache directory.
- Python: session events received over a TCP connection are now routed with the same session-map locking as stdio.
- Rust: a JSON-RPC frame whose header declares more memory than the host can reserve no longer aborts the process; the body is buffered as it arrives.
- Fixed a bug in
file_searchwhere negation patterns after a wildcard could incorrectly exclude matching files. - Go: a session event handler that is busy, or that makes an RPC call, no longer stops the client from reading its connection, so other calls and sessions keep working while it runs.
- Fixed a bug where MCP requests could hang or behave unexpectedly on oversized stdio responses; requests now fail immediately when a response exceeds the 16 MiB line limit.
- Node.js: a workflow's
ctx.log()andctx.phase()now throw aTypeErrorwhen given a value that is not a string, instead of failing a laterctx.step()orctx.agent()with an unrelated error and dropping progress lines. - Added opt-in session management tools in the runtime.
- Node.js: concurrent
listModels()calls on one client now share a single model list request instead of each sending its own. - Fixed a runtime issue where fallback handling could incorrectly enforce a prompt limit for models with unknown token limits.
- Python and .NET:
ToolSetnow rejects a tool name with a trailing newline, as the other SDKs do, instead of building a filter entry that matches no tool. - Saving or resetting a setting with
user.settings.setnow takes effect when settings.json still holds an older spelling of the same key (for examplecopy_on_select), instead of reporting success while the old value stays in effect. - Fixed a Java SDK issue where concurrent
listModels()calls could each trigger a separate model-list fetch. - Fixed: server-level
plugins.*methods called without a session now register the retiredcopilot-pluginsmarketplace for users with plugins installed from it, soplugins.marketplaces.list, browse and update find it. - Runtime now preserves currently registered plugin hooks while trust admission is queued.
- Local MCP servers are now started once per session after the first credential is resolved.
- Node.js: the default WebSocket request forwarder now sends the request headers, including authorization, on the upstream handshake.
- .NET: hook outputs whose
ModifiedArgsorModifiedResultis an anonymous object or a caller-defined class now reach the runtime when reflection-based serialization is enabled, instead of failing the hook call. account.loginnow rejects aloginthat does not match the token's user, and stores the token under the user's own login spelling.- Fixed a runtime bug where an unreadable managed MDM sandbox allowlist could incorrectly allow all hosts.
- .NET: a request whose error response the client cannot read now fails instead of never completing, and the connection keeps working.
- Java:
CopilotSession.compact()now compacts the session instead of failing with "Unhandled method session.compaction.compact". - Fixed an issue where ACP message chunks could lose the metadata for the subagent that produced them.
- Fixed a runtime issue by adding
COPILOT_PROVIDER_VISION, enabling BYOK tool images to reach the model. - Go: a tool result with empty
TextResultForLLMis now sent to the model as empty text instead of a Go struct dump of the whole result. - Fixed an issue where a crashed session could remain marked as in use if its process ID was later reused.
- Fixed a runtime issue where export events could be emitted while a deferred Mission Control creation was still in progress.
- Fixed a bug where plugin rules using comma-separated glob patterns were not applied.
- Fixed a runtime issue where skills skipped during loading from
skillDirectorieswere not reported. - Java: a slash-command or tool handler whose future fails now reports its exception message without the exception class name in front of it.
- Reuse valid saved MCP OAuth credentials on anonymous connections and initial challenges without requiring another sign-in.
- Fixed a runtime bug where an unknown
run_dynamic_workflowname did not list the registered workflows. - Fixed a runtime issue where tools could fail for the remainder of the first turn if ambient
ghlogin completed after the turn started. - Fixed: a
sessions.openhandoff of a remote session no longer fails when the local working tree has uncommitted changes; they stay in place and thecheck-changesprogress step reports them. - Improved Copilot SDK getting-started docs by ensuring the telemetry TypeScript and Go examples compile.
- Added SDK access to provider status, permission capabilities, and the autopilot stay preference for clients.
- Fixed doubled backslashes in Node.js extension examples in the SDK documentation.
- Fixed a runtime bug where an LSP server that crashed during startup could remain incorrectly marked as "starting".
- Fixed an SDK bug where
max_output_tokenscould be omitted from listed model limits in the Python, Go, .NET, and Java SDKs. - Fixed workflow registration validation to reject unknown fields and non-object values in workflow limits.
- session.abort() now delivers the ephemeral cancellation session.info notice, as interruptMainTurn already did
- Fixed
session_historyso searches and reads operate on message text instead of JSON-encoded content. - The bundled opt-in plugins Browser Use (
browser-use) and Computer Use (computer-use), when registered throughbuiltinPluginDirectories, now stay inactive until enabled (enabledPluginsorplugins.enable), as in the interactive CLI. - .NET SDK docs: four more README examples are compiled by the docs check.
- Rust SDK docs: README examples now compile as doctests, including a corrected elicitation handler example.
- Extensions no longer request elevated access after their connection closes during permission initialization.
- Concurrent deferred repository-hook loads no longer rebuild already loaded hooks or report a nonzero count twice.
- SDK clients now receive
session.mcp_server_status_changedfor MCP servers declared by a custom agent, tagged with the sub-agent'sagentId, including a server that fails to start. - Fixed a runtime bug so tool output spill files are now created with owner-only permissions.
- Fixed a runtime bug where a directory grant could be kept when the current working directory moved outside it under allow-all mode.
- Resuming an Autopilot objective no longer starts a duplicate continuation if its prior run already continued during persistence.
- Fixed a runtime bug where non-vision models are now correctly informed when a tool image is not provided.
- Fixed a Windows plugin uninstall issue where users could see a bare "Access is denied" error when another program had a plugin file open.
- Managed-extension launch diagnostics are now retained only when launch fails.
- Windows clients now continue proxy discovery when DHCP returns an empty or whitespace-only PAC URL.
- Large pastes saved through session.workspaces.saveLargePaste in the same millisecond no longer overwrite each other.
- Fixed a runtime bug where the working directory was not restored after a change was rejected.
- Fixed an issue where workflow run details could fail to report activity for running agents.
- Fixed a bug where
session.abort()could hang indefinitely when an MCP server did not respond. - Fixed remote HTTP MCP servers failing to connect through proxies that require Kerberos (Negotiate) authentication.
- Fixed a runtime bug where hooks could not detect sessions started by another hook, which could cause
agentStophook chains to never terminate. - Sessions now emit a durable
session.warning(warningType: "steering_not_delivered", withoriginatingMessageId) for each steering message that had not been delivered when the session shut down. - Improved the Node.js SDK README by updating Image Support and Streaming examples to type-check correctly.
- Fixed a bug where a subagent started in the same turn as a plan approval could end up with no tools when the session had an explicit tool allow-list.
- Add experimental typed catalog.recommend APIs across all six SDKs, including typed product-validation refusals.
- Fix Rust SDK test compilation under consumers that deny clippy::unwrap_used.
- Runtime now preserves active prefixes during late skill reloads, preventing prefix loss after reloads.
- User-input and exit-plan-mode handlers now receive the runtime request ID of the pending request (
requestId). Breaking (Rust):UserInputHandler::handleandExitPlanModeHandler::handletake a new optionalrequest_idparameter aftersession_id. - Fix WebSocket forwarder recursively closing an upstream that is still connecting.
- Fixed an issue where a
preToolUsehook returningnullformodifiedArgscould incorrectly fail the tool call. - Java: A tool defined without a handler is now declaration-only, as in the other SDKs: its calls stay pending for
session.tools.handlePendingToolCallinstead of failing with a NullPointerException message. - Rust: a tool handler that returns
Errnow fails the tool call like in the other SDKs, so the model sees a generic failure message instead of the error text. Return a failureToolResultto show the model specific failure text. - Fixed a runtime queue bug where
insertAtandupdateTextnow accept empty display text, allowing a running item to be hidden. - Python and Go: an empty custom agent list (
custom_agents=[], an emptyCustomAgentsslice) passed when creating or resuming a session is now sent instead of dropped, so resuming with it removes the session's custom agents, as in the other SDKs. - Clarified the CLI help text for
--secret-env-varsto more accurately describe what the flag protects. - Python and Go: an empty tools list (
tools=[], an emptyToolsslice) passed when creating or resuming a session is now sent instead of dropped, so resuming with it removes the session's external tools, as in the other SDKs. - Node.js: a workflow run's
ctx.signalis now aborted when the workflow body settles, soctx.parallel()andctx.pipeline()branches still running after a failed run stop at their next workflow call and signal-aware work is cancelled. - Fixed a bug where
instructions.discovercould omit repository instruction files that are not valid UTF-8. handlePendingCommandnow returnssuccess: falsewhen the command request was already resolved or does not exist, like the other pending-request responses.- Fixed:
sessions.listwithmetadataLimitno longer returns empty, unnamed sessions past the limit; they were already omitted within it. session.abortnow emits the ephemeralsession.infocancellation note ("Operation cancelled by user") that an in-process cancellation already emitted.- Java:
CopilotClient.getState()now returnsDISCONNECTEDonce the connection to the runtime has closed, for example after the runtime process exits, instead of continuing to reportCONNECTED. - tasks.getProgress now reports recentOutput as an empty string for a shell task that has printed nothing, instead of the text "(no output yet)".
- Fixed an issue where telemetry warnings raised during cold resume were not consistently preserved in the session log.
- Fixed a bug where model request header error messages could include the rejected credential.
- Fixed a runtime bug where the root agent span could be left unnamed when the selected agent defines a
sessionStarthook. - Fixed a bug where OpenTelemetry traces could incorrectly report the provider as "github" after switching to a Windows ML model.
- Python: handlers registered with
session.on()are now called in the order they were registered, matching the other SDKs. - Fixed an MCP OAuth login issue by retrying the login probe without a protocol version, improving authentication reliability.
- Fixed a runtime bug so MCP OAuth discovery transport failures are reported correctly instead of being shown as missing metadata.
session.resumenow rejects areasoningEffortthe model is not entitled to, instead of accepting it.- Improved runtime reliability by preserving extension tool readiness during concurrent resume operations.
- Fixed a prompt hook bug where empty replacements could clear the prompt; dropped replacements are now reported.
- Fixed an MCP OAuth bug where bare-origin resource values are now sent with a trailing slash.
- Document Agent Plugins authoring and clarify SDK plugin loading options.
- Background task completions that arrive after an Autopilot run stops are now delivered with the next message and announced via a
session.infoevent with infoTypebackground_completion_deferred, instead of being discarded. - Fixed a bug where BYOK Chat Completions requests were sent with Copilot sampling defaults.
- Fixed an issue where a SQL tool query that never completes could hang the session even after pressing Ctrl+C.
- Fix native transport initialization on Linux OpenSSL builds missing unused optional algorithms and report unavailable required crypto capabilities without leaving model turns processing indefinitely.
- Align MCP tool telemetry with GenAI semantic conventions, including standard method/stdio transport attributes and structured failure reporting.
- SDK sessions now stay in autopilot mode by default after an autopilot objective completes or pauses, matching the interactive CLI.
- Improved reliability by automatically retrying Fusion plan requests after transient failures (HTTP 429/5xx or dropped connections).
- Fixed a runtime bug where a tool result arriving before its corresponding tool call could be duplicated after the call.
- Provider-registry (BYOK) models whose
modelId(orid) exactly names a built-in model now get that model's editing tools, search tool and system prompt. - Fixed a runtime warning issue where a single oversized request could emit repeated warnings for each stripped attachment.
- When a session's next request no longer fits in the model's context even after trimming older history, the turn now fails with the prompt-budget error instead of dropping the tool calls it just made and making them again.
- Fix MCP OAuth credential coordination and bounded recovery when a newly adopted cached token is rejected.
- On Windows, built-in PowerShell tool commands that run
cls,clearorClear-Hostno longer fail with "The handle is invalid". - CLI: Explicit
--modelselections now take precedence over a custom agent’s configured model. - Tool definitions passed to
session.tools.setor session creation now keep their schema's key order, so re-sending unchanged tools no longer breaks prompt caching. - Runtime preserves current skill catalogs during extension reconciliation so unchanged skills remain available to the first model request.
- Fixed a Unix issue where BYOK API-key command helpers could continue running after timing out.
- Runtime now preserves the inventory of disabled built-in plugins across refreshes.
- Enabled context-management tools by default for MAI-code-2 models in the Copilot runtime.
- All SDKs: permissions.paths.list() returns additionalDirectories and the explicit session additional directory grants (experimental).
extensions.discover()now includes extensions from enabled plugins installed from a local directory marketplace, matching what sessions load.- Fixed a bug where
copilotUserlookup could retry after timing out. - Fixed an issue in the Copilot runtime where LSP
fileExtensionskeys are now accepted with or without a leading dot. - Java:
@CopilotToolmethods in a class nested inside another type now compile and register; a private nested tool class gets a processor error instead of uncompilable generated code. - Fixed a bug where MCP, LSP, and plugin manifest files saved with a UTF-8 BOM were incorrectly rejected.
- Improved the Node.js SDK README by updating Permission Handling examples to type-check correctly.
- Runtime now prevents block compaction when buffers are exhausted and surfaces history truncation to callers.
- Custom subagents with private MCP connections avoid starting inherited servers excluded by their resolved tool list.
- Session metadata from getSessionMetadata and listSessions now includes clientName, the client that created the session, and lastClientName, the client that last ran it, when known.
- Rust: Breaking: SessionMetadata adds client_name and last_client_name; code that builds SessionMetadata with a struct literal or destructures all of its fields must handle them.
- Fixed a bug where
stop_bashcould incorrectly report a non-existent shell ID as successfully stopped. - Fix Kerberos auth callback handling: .NET callbacks no longer block the JSON-RPC reader, Go releases startup waiters when a start fails, and Python answers cancelled callbacks and cancels in-flight callbacks when the connection closes
- Node.js: a request whose params cannot be serialized no longer marks the client disconnected, so a later
start()no longer launches a second runtime. - Improve Claude token estimates used for session context budgeting.
- Sandbox blockedHosts now also match visible TLS SNI in CONNECT tunnels.
- instructions.getDiscoveryPaths, skills.getDiscoveryPaths and agents.getDiscoveryPaths now return absolute paths when a projectPaths entry is relative.
- Fixed a runtime bug where settings merges could overwrite an unreadable
settings.json; merges now leave that file unchanged. - Local workspace diffs use the session's current working directory instead of stale persisted workspace metadata.
- Sandboxed Hugo commands receive Hugo and Go cache grants when developer-tool access is enabled.
- Shell safety warnings are no longer duplicated as standalone session.info events; permission-request warnings and tool errors are unchanged.
- Added validation for reasoning effort values before they are accepted or persisted.
- Improved runtime caching so MCP call results are preserved across unrelated snapshot refreshes.
- Added SDK contracts to expose GitHub services through the runtime.
- Keep credential assignments redacted across detached-shell output pages.
- Added optional causal provenance to inter-agent session events.
- Clarified errors when a managed sandbox cannot be enforced.
- For authenticated github.com sessions, the runtime now automatically uses a default secret-scanning endpoint.
- Eligible WorkIQ OAuth requests use the host's Entra identity with M365 inference, including manual configuration and token renewal, while preserving explicit opt-outs.
- Expose Team marketplaces through plugin APIs.
- Fix access to bundled search binaries in sandboxed SDK sessions when developer-tool access is enabled.
- Switching a concrete-model session to Auto now reconciles unavailable retained tier preferences before validation; explicit tier requests remain strictly validated.
- Fusion model switches accept a null reasoning-effort reset while continuing to reject unsupported effort values.
- Sessions created with
customAgentsandenableConfigDiscoveryoragentDirectoriesnow list discovered agents, such as.github/agentsfiles, from the firstagent.list()instead of only afteragent.reload(). - Requests can fall back to direct networking when system proxy or PAC resolution fails.
- Fixed a bug where hook configuration files encoded with a UTF-8 BOM were incorrectly ignored.
- Fixed a bug where LSP files matched by an empty extension mapping did not use the configured languageId; the configured languageId is now sent.
- Java: The deprecated
parentToolCallIdproperty of the assistant message, message delta, usage, tool execution start and tool execution complete event data is now marked@Deprecated, as in the other SDKs. - Honor HTTP 429 Retry-After timing using existing request policies without adding queue-wide or worker-stop behavior.
- Fixed an issue where setting
session.mcp.disableafter the first MCP load could be lost on the next reload. - Unlisted GPT models now use GPT-6.1 defaults, and unlisted Claude models use Sonnet 5.5 defaults with a 48K output cap and no configured default refusal fallback.
- Java previews with the same base SDK version now update one rolling Maven snapshot coordinate.
- Fixed plugin uninstall behavior so listed plugins can be removed even when their files are missing.
- Sessions created on the standalone runtime host now return before their extensions start, so an extension's startup consent prompt reaches the client.
- Fixed a runtime bug where continuations with no usage were incorrectly reported as prompt-cache breaks.
- Fixed premature extension readiness reporting that could cause an unrelated extension to restart during a managed-policy refresh.
- A session whose events cannot be persisted now emits a
session.error(errorType: "persistence") when persistence starts failing, repeats it every 10th failed attempt while it lasts, and emits onesession.info(infoType: "persistence") when it recovers, instead of asession.erroron every retry. - Fixed an issue where failed HTTP MCP tool calls could display an internal type name in the error message.
- Sessions created or resumed with a stored Efficiency Auto preference that the account's Auto metadata no longer lists now start and route on the provider default, instead of failing every turn.
- Java and Python SDKs now preserve experimental markers on RPC properties.
- Explicitly reap POSIX stdio MCP child processes during normal awaited shutdown.
- Preserve application-defined request metadata in MCP App tool calls.
- Fixed extensions being stuck behind a pending managed policy after the first turn when the session used a token for the signed-in account.
- A leading
~or~/inskillDirectories,agentDirectories,instructionDirectoriesorhookDirectoriesnow resolves to the home directory instead of a folder named~under the working directory. - Added a warning when a top-level SKILL.md in a skills directory hides its skills.
- Node.js, Python, Go, .NET and Rust:
stop()no longer waits forever on a runtime that stops responding while sessions are attached; session detach now shares the 10-second shutdown timeout and a detach that does not finish is reported as a stop error. - Add agent attribution to lifecycle progress diagnostics
- Fixed a runtime issue where
preToolUsehooks were not executed concurrently for parallel tool calls. - Resuming a session on the same connection now replaces the slash commands that connection registered before: a new
commandslist replaces them andcommands: []removes them, instead of leaving commands listed that can no longer run. - Auto model selection no longer falls back to the legacy Auto API when
/autofails; the turn retries once and then fails with a clear error. - Fix a standalone session startup race between initial managed-policy publication and repository-hook loading.
- Fixed a bug where LSP rename applied its WorkspaceEdit as a single change.
- Fixed a runtime bug where abort requests could hang if the host did not respond to tool hook or bearer token callbacks.
session.rpc.shell.execnow honours a timeout above 4294967295 ms (capped there) or a fractional one (rounded up) instead of using the 30-second default, and rejects a negative or non-numeric timeout with an invalid-params error.- Fixed
web_fetchto decode an XML document in the encoding its XML declaration names when the server sends no charset, instead of as UTF-8. createSessionandresumeSessionnow reject aworkingDirectorythat is relative (including~) or is not a directory, instead of silently using the runtime's own directory.- Fixed empty input hints in session.commands.list for built-in commands that declare only choices.
session.options.updatenow rejects aworkingDirectorythat does not exist, is not a directory, or cannot be entered, and leaves the session in its current directory.- Fixed duplicate session.created and session.deleted events from the CLI SDK server.
- Fixed: sign-in flows started with
session.accounts.login.beginare cancelled when their session is disconnected, so they no longer block other sessions from signing in. - session.send, sendAndWait and sendMessages now reject a mode other than "enqueue" or "immediate" with an error, instead of silently queueing the message as "enqueue".
- Fixed customizations.reload incorrectly reporting unloadable custom agent files as reloaded.
- Protected sessions can read their selected extension SDK's authoring references without additional sandbox path grants.
- Coalesce renewable GitHub CLI credential acquisition and retry a transient helper timeout once without switching accounts.
- Fixed an issue where ACP clients could not leave autopilot in sessions started with
--yolo. - Improved telemetry privacy by redacting credentials from engine message telemetry.
- Python: Rename the internal copilot.generated package to copilot._generated; import protocol types through copilot.rpc or copilot.session_events.
- Renamed public SDK identity branding to "Microsoft Copilot SDK".
- Breaking: Group client and session options across all six SDKs and create/resume/open/update RPCs, keep configDirectory and enableConfigDiscovery at the session root, retain explicit spawned-versus-attached runtime connection factories with grouped ownership settings, and require SDK protocol 4.
- Replace singular inference.provider with named provider connections and model rows, adding default-provider alias routing, explicit CAPI disablement, and consistent resume policy across SDKs.
- Managed settings can make the built-in Browser Use plugin unavailable with
features.browserUse: false.