Skip to content

v2.0.2-preview.2

Pre-release
Pre-release

Choose a tag to compare

@copilot-sdk-mirror-writer copilot-sdk-mirror-writer released this 20 Nov 21:52
· 3 commits to main since this release
  • Fixed MCP prompt listing and retrieval, resource listing and reading, and resource template listing failing immediately after session creation with configured, enabled MCP servers. These calls now start the session's MCP servers without requiring a prior server-list call or user message.
  • onErrorOccurred hooks now receive a failed model call's message as plain text instead of a JSON-quoted string.
  • Fixed sandboxed Git operations on Windows with bin-only MinGit installations by allowing read-only access to default configuration, CA data, and templates while preserving sandbox restrictions.
  • Sandboxed commands no longer start when a denied-path glob matches more than 100 paths; the error names the glob to narrow.
  • Reduced slowdowns in long-running sessions where sub-agents execute many synchronous shell commands.
  • Allow user.settings.loadForEdit to read a specified configuration directory and preserve object-shaped params for parameterless Go calls.
  • Allow conditional user.settings.update calls to preserve concurrent settings changes, compare equivalent JSON numbers consistently, retain exact Go and Node settings integers, and accept Node negative-zero conditions.
  • Keep Node settings reads JSON-compatible by default and add explicit exact numeric options for bigint results; reject unsafe ordinary responses without closing the connection. Preserve exact Go settings integers and export UserSettingsJsonValue and Node settings response types.
  • Preserve comparison-sensitive float64 settings values across persistence and Go input routes, retain standard float32 JSON precision, and reject Node bigint conditions and settings writes outside the runtime's exact integer range.
  • Reject oversized Python condition integers, Go settings integer tokens, and Node raw JSON integers before transmission to prevent stale conditional writes.
  • Reject oversized Python settings integers before writes reach the runtime, without changing supported values or caller data.
  • Reject oversized integer tokens in public Java and .NET settings writes before transmission, without changing supported values or invoking custom serializers twice.
  • Reject oversized Python integers before settings transformations and validation, so normalization cannot round an invalid value into a stale conditional match.
  • Preserve the conditional applied outcome in Node unsafe settings response errors, and distinguish completed updates from response decoding failures.
  • Preserve represented float32 integer values in Go conditional settings updates without changing ordinary fractional precision or caller data.
  • Adds sandbox.userPolicy.seatbelt.extraMachLookups to configure exact macOS Mach service lookups without enabling Keychain access. Patterns are rejected, and managed policy restricts allowed names.
  • Fix resuming native resident sessions through the legacy JavaScript host when no session facade exists.
  • Rust: Reclaim idle session initialization entries on long-lived clients while preserving failed-cleanup fencing.
  • Fix TypeScript 7 source compilation of the Node.js SDK shell authentication callback.
  • Cap Gemini Carbon and Argon CAPI generation at 48,000 output tokens to reduce long-generation deadline collisions.
  • Fixed synchronous sub-agents cancelled through session.rpc.tasks.cancel before responding reporting "completed but produced no response" instead of a cancellation message in tool results, events, and failure hooks.
  • Fixed missing text results from scheduled built-in slash commands: SDK clients now receive non-blank output, including /chronicle reindex summaries, as ephemeral session.info events with infoType: "slash_command".
  • Fixed providers.config.updateModel deleting a saved capability set to null that providers.config.read does not return, and providers.config.read failing for every saved model when one has a boolean adaptive_thinking (that value is now left out of the read and kept by updates).
  • Stopping background agents now closes their pending permission requests, including those of cancelled descendants, with a permission.completed result of kind "cancelled" and reason "subagent_cancelled".
  • session.cancelAllBackgroundAgents and session.tasks.cancel now wait for pending permission requests of stopped agents to be settled before returning.
  • SDK sessions can receive independent, default-off automatic tool-search rollouts for MAI above 30 total tools and Gemini above 50 in default context or 150 in long_context, including built-ins. The old combined rollout flag is retired; explicit SDK tool-search settings continue to take precedence.
  • Answering a permission request for a tool call whose arguments nest 125 levels deep no longer fails with "recursion limit exceeded" and leaves the turn waiting.
  • Fixed prompts and progress text after unterminated OSC escape sequences being hidden while shell commands run, affecting read_bash results and bash calls that time out before completion.
  • Fixed plugin language servers failing to start when command or arguments use ${COPILOT_PLUGIN_ROOT}. The placeholder now also resolves to the plugin installation directory in rootUri and initializationOptions.
  • Reduced SDK session creation and resume latency by skipping unnecessary Git lookups for lifecycle diagnostics, without changing diagnostics persistence behavior.
  • Managed settings can make the built-in Browser Use plugin unavailable with features.browserUse: false.
  • Sandboxed local MCP server startup failures now report the underlying reason without the internal GenericFailure prefix in failed server statuses received by SDK clients.
  • .NET and Rust SDK builds at the 0.0.0-dev placeholder version no longer try to download a nonexistent CLI release; they skip the download and explain how to point at a local CLI.
  • Subagent lifecycle events now include the originating interaction ID.
  • Inline blob image attachments are now labelled as inline images so models no longer look for a file with the attachment's display name on disk.
  • Rewinding to a session's last turn after its runtime was force-stopped or killed while idle now restores the files that turn changed.
  • Clarified Node.js SDK tool failure handling: thrown errors are hidden from the model; return resultType: "failure" with textResultForLlm to provide a model-visible explanation.
  • Default Windows sandbox package queries to packageQuery, with explicit opt-out and managed constraints.
  • listModels now reports defaultReasoningEffort for reasoning-capable models.
  • Python: session.rpc.tools.execute is now annotated with the raw tool result it returns (str or dict) instead of ToolResultExpanded.
  • Go: ForceStop now ends a Start that is connecting or waiting for the runtime's replies on every transport, including a client connected to an external server through URIConnection, instead of waiting for it.
  • Node.js: GetAuthStatusResponse.authType now includes "account", "token-provider" and "copilot-api-token", which getAuthStatus() can return.
  • Python: define_tool no longer advertises an integer parameter default outside the 64-bit ranges, which the runtime would round before the model sees it; the handler still receives the exact default.
  • .NET: a postToolUse hook that returns a ToolResultObject as ModifiedResult no longer erases the images, session log and telemetry of the tool result that it did not set.
  • Node.js, Python: listSessions, getLastSessionId, getSessionMetadata, getForegroundSessionId and setForegroundSessionId called while start() is running now wait for it and then use the new connection, or fail with the start's error, instead of failing with Client not connected or sending their request before the connection handshake.
  • infiniteSessions.backgroundCompactionThreshold now controls when models that use Session Memory rollover (such as mai-code-2) start a fresh context window.
  • Declarative sessionEnd hooks from the user hooks directory, settings, repositories and plugins no longer run for a session that never received a prompt. SDK callbacks, explicit hookDirectories and managed-policy hooks still run.
  • Scheduled /rubber-duck in a session without the rubber-duck agent now reports it unavailable, and runs an SDK command of the same name, instead of sending the rubber-duck prompt.
  • Experimental user.settings.diagnoseProblems now reports unrecognized nested settings that have no effect with kind "unknown-nested-key" and a dotted path in rawKey.
  • includedBuiltinSkills with a non-string entry is now rejected with an error instead of silently loading every built-in skill.
  • Go: Client.Start now returns its context's error when the context is done while it waits for another call to start or stop the client.
  • The .NET, Python and Java SDKs now answer a hook or user-input request whose handler fails with a cancellation or an Error, instead of leaving the turn waiting forever.
  • Python, Node.js: delete_session / deleteSession called while start() is running now waits for it and then deletes on the new connection, or raises the start's error, instead of failing with Client not connected or (Python) sending the request before the connection handshake.
  • Calling session.name.setAuto with the session's current name no longer rewrites the session's workspace file or emits session.title_changed; it still returns applied: false.
  • Fixed SDK sessions hanging and send_and_wait timing out when an extension keeps running after its join is refused. Extensions can retry before their original 30-second startup deadline; otherwise, they are stopped and reported as failed so prompts can proceed.
  • Fixed session creation and resume failing with "Session plugin inputs changed during reload" when plugin preparation updated the session's plugins while it started.
  • Fixed forked sessions running scheduled prompts inherited from their source session, including when existing forks are resumed. Inherited schedules remain in history but are inactive; schedules created in the fork and source session are unaffected.
  • A file attachment that is a named pipe, socket or device with an image extension is now skipped instead of stalling the turn, holding the session lock and hanging client.stop().
  • Java: a session event handler that is busy, or that waits for an RPC result, no longer stops the client from reading its connection, so other calls and sessions keep working while it runs.
  • A BYOK model named after a built-in compatibility profile, such as gpt-4, now honours a declared maxPromptTokens above that profile's context window instead of refusing every prompt.
  • Fixed session.history.truncate to preserve accumulated usage in session.usage.getMetrics, including AI credits, request and token counts, usage breakdowns, and code-change totals. Truncating history no longer reduces usage counted toward session AI credit limits.
  • .NET: a tool parameter default that the runtime would round before the model sees it, such as a decimal outside the 64-bit integer range or with more digits than a double holds, is no longer advertised; the handler still receives the exact default.
  • Fixed a runtime issue where invoke_canvas_action could fail when batched with open_canvas for the same instance.
  • Go: tools defined with DefineTool receive integer arguments beyond 2^53 exactly, and a uint64 argument near its maximum no longer fails the call.
  • Client-defined tools called without a required argument now fail with an error naming the missing arguments instead of invoking the handler.
  • createSession now rejects a reasoning effort the model does not support, instead of failing at the first send.
  • Improved Go SDK setup documentation and missing-runtime errors with bundler commands and guidance for setting COPILOT_CLI_PATH.
  • On Linux, model requests now fail immediately with the certificate error when an endpoint's TLS certificate is untrusted, expired, or issued for another host, instead of retrying.
  • Python and Go: an empty plugin directory list (plugin_directories=[], an empty PluginDirectories slice) passed when creating or resuming a session is now sent instead of dropped, so resuming with it unloads the plugins the session mounted from plugin directories, as in the other SDKs.
  • Fixed stale MCP tool cache buildup in SDK hosts that create sessions across many workspaces: cache writes now remove entries older than 14 days, including those for workspaces never reopened.
  • Python: a postToolUse hook can return a ToolResult as modifiedResult, and a canvas action can return a ToolResult. Before, the hook's result was dropped and the canvas action failed.
  • Fixed a bug where MCP tools were not reported when their names clashed with model-facing names.
  • Python: CopilotClient.stop() now runs each open session's on_session_end hook; before, the session was unregistered before it was closed, so the hook was refused.
  • Python: handlers registered with session.on(...) now receive the events sent while CopilotClient.stop() closes their session, such as session.shutdown.
  • Fixed prompts-only and resources-only MCP servers being disconnected when they do not advertise tools and return “Method not found” for the initial tool listing. Prompts and resources remain accessible, and session.mcp.listTools returns an empty list.
  • Fixed concurrent attempts to stop the same background shell incorrectly failing with "Unknown shell ID"; subsequent stops now report that the shell has already stopped.
  • Fixed plugin hooks, including plugin custom-agent hooks, failing to run when their exec path contains ${COPILOT_PLUGIN_ROOT}.
  • Fixed custom-provider getEndpoint results to report the provider type, wire API, and transport actually used by the session, preventing parsing errors in typed SDK bindings.
  • Java: tool arguments with an Optional or OptionalInt record component or bean field now decode: a sent value arrives as Optional.of(value) instead of failing the call, and a left-out or null value as an empty Optional instead of null. This covers @CopilotTool record parameters, Param.of lambda tools and ToolInvocation.getArgumentsAs.
  • Java: tool results that contain Optional values now serialize instead of failing; an empty Optional is written as null.
  • Java: copilot-sdk-java now depends on com.fasterxml.jackson.datatype:jackson-datatype-jdk8.
  • Fixed GitHub MCP reconnections using the previous account's token after switching from an MCP-disabled account to an MCP-enabled account.
  • After an account switch, affected MCP servers now wait for the current account's credentials before reconnecting; hosts that do not explicitly configure GitHub credentials restore them on the next turn.
  • Java: an @CopilotTool method parameter declared @CopilotToolParam(required = false) as Integer, Long, Double, Float, Short or Byte now receives null when the model leaves it out, instead of the tool call failing with a NullPointerException.
  • The built-in web_fetch tool adds its continuation note when max_length is fractional, instead of returning part of the page with no sign that more remains.
  • Update the Python SDK's OpenTelemetry dependencies to address a security advisory.
  • Fix Fusion classification of pasted text while preserving timeline labels and file references.
  • Add session-status to gitHubFetch.executeOperation and optional authOwnerId to account.getWebUrl.
  • Windows Git operations recognize Git batch shims through PATH and PATHEXT.
  • Fixed intermittent session creation failures when repository hook loading overlaps with a hook rebuild.
  • Rust: terminate Unix CLI process groups during startup cancellation and client teardown to avoid orphaned wrapper descendants.
  • Fixed session.usage.getMetrics losing AI credit status, user request counts, and provider/model breakdowns after truncating history before the first prompt and resuming the session.
  • Fixed /exit and /quit commands (including print variants) queued through commands.enqueue staying pending when the active turn finishes during command admission.
  • Assisted Permissions in Autopilot now explains automatic review failures and exclusions in denial feedback, including unavailable review models and review size limits.
  • Incomplete permission-recovery history is now reported as a skipped automatic review rather than a review failure; explicit approval remains required.
  • Go: an extension launch profile returned by an ExtensionLaunchProvider with nil Args or Env is now sent with empty collections, so the runtime launches it instead of rejecting it.
  • Go: a session filesystem directory listing that is empty (nil) or failed is now sent with an empty entries array, so the runtime no longer rejects it and CreateSession no longer fails.
  • Fixed session idle timeouts closing sessions while attached async commands, background agents, user-requested shell commands, or pending background notifications are active. Idle cleanup resumes after work finishes and the timeout elapses again; detached commands still do not keep sessions alive.
  • Deleting a repository or local settings override now leaves missing settings files absent, without creating directories or lock files.
  • Fixed unnecessary restarts of MCP servers started by session.mcp.callTool or MCP Apps reads before initial MCP loading, preserving in-memory state across session.mcp.list and the first turn when server configuration is unchanged.
  • Node: in empty mode, a session option explicitly set to undefined (such as mcpOAuthTokenStorage) now keeps its empty-mode default instead of falling back to the runtime default.
  • Fixed /chronicle cost-tips omitting earlier model usage from SDK sessions resumed in the CLI.
  • Fixed session.rpc.mcp.removeGitHub() incorrectly returning { removed: false } after stopping the built-in GitHub MCP server when using the CLI's JavaScript server mode.
  • When the runtime rejects a message before its turn starts (for example, an unavailable model), the session.error now carries the message's originatingMessageId, so a client attached to an AHP host gets the error at once instead of a generic failure two minutes later.
  • Java: a LocalDateTime or LocalTime tool parameter now accepts a value with a UTC offset (2026-11-15T10:20:30+01:00, 10:20:30Z) and receives the date and time as written, instead of the tool call failing.
  • Rust: events held for a session ID that is not registered yet are now also limited to about 16 MiB, so a client subscribed to a session it never registers no longer keeps hundreds of megabytes of event payloads.
  • Fixed subagent.started, subagent.configured, subagent.completed, and task resolvedModel reporting a required custom-agent model or an explicit task model that the allowed-model policy blocks, instead of the policy fallback the sub-agent runs on. Lifecycle events now report modelSelectionSource as runtime_policy while preserving explicitModelOverride; configured token limits and reasoning effort reflect the fallback without changing execution behavior.
  • Fixed out-of-order delivery of session.info, session.warning, and session.error events from concurrent session.log calls on the same session, so logged events arrive in recorded history order without blocking other sessions.
  • .NET: CopilotClient.Dispose() no longer hangs when called on a UI thread (WinForms, WPF, MAUI) or under another single-threaded SynchronizationContext or TaskScheduler.
  • Go: a postToolUse hook that returns a ToolResult with neither ResultType nor Error set as ModifiedResult no longer marks the tool call as failed; it is sent as a success, as a tool handler's result is.
  • Fixed concurrent agents.discover, skills.discover and instructions.discover calls without a session waiting up to 30 seconds, instead of at most 10, when the Team customizations lookup is slow or unreachable.
  • OpenTelemetry exports now include available enduser.pseudo.id on all session-owned spans, including chat and tool spans. user.name is also included when capture.identity is enabled (disabled by default). New spans reflect refreshed session identity; existing spans retain their original attribution.
  • Node.js: forceStop() now kills a CLI process that is still running after stop() timed out waiting for it to exit, and start() kills such a process before it starts a new one.
  • A turn no longer waits for a running background sub-agent's undelivered events before it ends.
  • Deferred-tool reminder preservation is available behind the default-off stable prompt-prefix experiment; control keeps existing behavior.
  • Prompt-prefix preservation is available behind the default-off stable prompt-prefix experiment; control keeps existing behavior.
  • Go: a generated RPC method whose required list is nil now sends an empty array instead of null, so the runtime no longer rejects the call (for example, session.RPC.Tools.Set with a nil list now removes every tool).
  • Fixed duplicate delivery and redundant instruction_discovered notifications for unscoped rules already included in the system prompt when on-demand discovery accesses files under an in-repository instructionDirectories directory. Scoped rule discovery and session.instructions.getSources remain unchanged.
  • Lifecycle diagnostics now report model catalog refresh progress, completion outcomes, and which refresh is blocking a waiting request.
  • Fixed out-of-order delivery of session.context_changed events from concurrent session.metadata.recordContextChange calls on the same session, including bursts mixed with session.log calls, so the events arrive in recorded history order.
  • Warnings for failed cleanup after cancelling a local-model download now include redacted error details and elapsed cleanup time; cleanup behavior is unchanged.
  • Fixed a race during concurrent embedded SDK shutdown that could cause cleanup errors or close the host channel prematurely.
  • Prompt-prefix preservation remains behind the default-off stable prompt-prefix experiment; both arms recheck the current external tool and approved contract after permission resolution.
  • Fusion routing again prioritizes displayPrompt over prompt and no longer reads attached text or saved-paste files for classification.
  • Go: Requests now return an error instead of panicking if the system randomness source fails while generating a request ID.
  • Added gemini-4-hydrogen to supported model types, with low, medium (default), high, and xhigh reasoning efforts when available for your account.
  • Added Gemini 4 Hydrogen model limits: a 936,000-token prompt budget, 1,000,000-token context window, and 64,000-token output limit.
  • Fixed an issue where subscribing to a republished session could incorrectly return "not found" after a subscription was rejected while the session was unpublished.
  • Fixed an issue where failed usage recovery could interfere with subsequent session storage operations.
  • Fixed security review summaries using outdated reporting guidance when aiSecretDetection is toggled back to a previously used setting within a session.
  • Fixed subagent session-limit notifications to reflect session-wide usage and the current AI credit limit, preventing repeated false limit-reached warnings after the limit is raised or removed.
  • Fixed tracking of host-customized system prompt changes on session resume when another changed input is omitted from the prompt. Affects only the STABLE_PROMPT_PREFIX experiment, which remains disabled by default.
  • Add session marketplace search and non-mutating update probes, plus optional plugin and MCP discovery context and typed memory bridge configurations.
  • Add a guarded marketplace migration method for launchers that use the configured runtime profile.
  • Fixed search snippets and HTTP error previews to omit characters split by truncation instead of adding replacement characters.
  • Oversized HTTP error previews now stay within 512 bytes, excluding the ellipsis, even when the response contains invalid UTF-8.
  • Sub-agent streaming events no longer fall minutes behind on the session's event stream when several sub-agents stream at once.
  • Sessions with client-side programmatic tool calling enabled now forward successful child image results to the model.
  • Fixed shell permission checks so reads of ~ paths outside approved directories, reads through ~user, ~+, ~- or unset path variables, PowerShell FileSystem:: paths and backtick escapes, and globs or brace lists that reach outside approved directories through a workspace symlink require path approval instead of being silently allowed. Requests are denied when approval is required but prompting is unavailable.
  • Strengthened assisted permission checks with GPT-6 Luna and Claude Haiku 5.5 to avoid treating an agent's claims about user requests or approvals as user authorization.
  • Added credential-minting and credential-exposure reason codes to permission approval evaluations and Assisted Permissions exclusion reasons.
  • Assisted Permissions now requires explicit approval for detected SSH login-key installations, deploy-key registrations, and authorization-file changes such as sudoers edits; Autopilot denies these requests.
  • Automation-started turns now require explicit approval for detected commands that pass on live credentials from credential stores; Autopilot denies these requests.
  • Fixed premature unknown-tool warnings for MCP tool allowlist and exclusion entries while MCP tools are still loading or configured servers are not connected.
  • Fixed misleading managed-settings warnings for legacy permissions.model values superseded by a top-level model setting.
  • Go: a generated RPC method now sends an empty array instead of null for a nil required list inside a nested value, such as PermissionRulesSet.Approved or AuthenticatedAuthorityGrant.Statements, so the runtime no longer rejects the call.
  • Control-dispatched steering messages now replace earlier undelivered steering messages instead of being rejected. Messages already in flight are unaffected; only the latest pending message is delivered afterward.
  • Fixed history truncation and conversation-only rewind causing subsequent model requests to use outdated working-directory or added-directory context.
  • Fixed plugins.marketplaces.remove and session.plugins.marketplaces.remove being blocked by missing plugins or unreadable directories in unrelated local marketplaces, while preserving dependent-plugin protections.
  • Fixed SDK session startup and resume hanging during managed-settings refresh when an allowed MCP server stalls during initialization.
  • MCP servers restored after a managed-settings refresh now reconnect in the background; refresh completion no longer guarantees server readiness, but turns still wait for pending connections.
  • Fixed a runtime bug where context attribution did not accept memory message sources.
  • Rust SDK now requires serde_json 1.0.129 or newer.
  • Windows filesystem error messages now preserve .. segments in verbatim paths and display paths containing NUL characters without truncation; file access behavior is unchanged.
  • In-process runtime hosting contains unwinding Rust panics at request, turn, queue, and reverse-request boundaries and reports failures through existing error contracts.
  • Browser authentication now uses 43-character OAuth state values for GitHub (previously 22) and 43-character PKCE verifiers for Entra (previously 86).
  • Reduced session creation overhead when using shared-process hosting.
  • Python: an empty disabled_skills list passed to resume_session is now sent instead of dropped, so resuming a loaded session with it re-enables the skills disabled earlier, as in the other SDKs.
  • Go: empty DisabledSkills and ExcludedTools slices passed to ResumeSession are now sent instead of dropped, so resuming a loaded session with them re-enables the disabled skills and the excluded tools, as in the other SDKs.
  • BYOK provider HTTP 403 errors identified as quota, budget, billing, limit, or policy refusals now include the provider's message and code, when available, in session.error instead of credential guidance; errorType remains authentication.
  • Java: a GitHubTokenProvider or BearerTokenProvider that throws a non-fatal Error (such as AssertionError) from getToken is now answered with an error, so createSession or the send fails with the provider's message instead of waiting forever.
  • Fixed MCP tool progressMessage formatting in tool.execution_progress events: whole numbers no longer include trailing .0, and values such as 0.000001 use decimal notation.
  • Fixed session resume failures after passing a non-boolean continuePendingWork or creating a native-host session with malformed githubMcpToolConfig. Previously affected sessions can resume with their history; invalid values are ignored on resume while valid configuration members are preserved.
  • Session paths, plugin discovery, and MCP discovery now use the account home directory when HOME (Unix) or USERPROFILE (Windows) is unset or empty. Explicit configuration overrides remain unchanged.
  • Windows home-directory discovery no longer falls back to HOMEDRIVE/HOMEPATH or HOME; Unix marketplace discovery now prefers HOME over USERPROFILE. Affected configurations may resolve to different locations; existing files are not migrated.
  • Java: a PermissionHandler that throws a non-fatal Error (such as AssertionError) is now answered with a denial, so the tool call ends and the turn completes instead of waiting until the session is aborted.
  • Fixed sessions with enableHostGitOperations disabled, including SDK clients in mode: "empty", still running Git probes during prompt construction and including pull request template instructions in the system prompt.
  • cancelSamplingExecution now stops a sampling call cancelled right after it started; before, the cancel returned cancelled: true but the model request was still sent and the call returned the answer.
  • Fixed session instruction loading and discovery to include .claude/CLAUDE.md at the repository root and working directory, while continuing to exclude ~/.claude/CLAUDE.md.
  • Instruction discovery paths now include .claude/CLAUDE.md at the repository root, except when the root is the home directory.
  • Disposing an SDK host now attempts to close Web PubSub connections gracefully without delaying shutdown.
  • Fixed a hang when resuming sessions with continuePendingWork after a disconnect dropped a tool result. Affected tool calls are now marked as interrupted, even when no handler is registered on resume.
  • A custom agent's source-qualified tools entries (builtin:, mcp: and custom: names and wildcards) now take effect when the agent runs as a sub-agent through the task tool, as they already did for a selected agent.
  • Fixed concurrent sessions.pruneOld calls reporting sessions deleted by another call and emitting duplicate session.deleted events.
  • Executing /session prune now emits a transient session.info progress event with infoType session_prune_progress before pruning, except for dry runs or invalid arguments.
  • Fixed session resume failures caused by malformed extension_context attachments; invalid entries are now omitted from resumed messages.
  • Fixed malformed extension_context attachments suppressing valid attachments in the same message's model prompt.
  • Go: Client.Stop() on a TCP connection now sends SIGTERM before killing the CLI process, so a runtime started by the installed copilot command exits and Stop no longer returns timed out waiting for CLI process to exit after kill.
  • IPv4-mapped IPv6 loopback addresses are now accepted for local Ollama endpoints, HTTP hook localhost exceptions, local transports, and loopback proxy bypasses; existing opt-in requirements remain unchanged.
  • Catalog HTTP URLs now accept IPv6 loopback ([::1]); catalog and metrics HTTP URLs also accept the full 127.0.0.0/8 range and IPv4-mapped loopback addresses.
  • Plaintext HTTP proxy bypasses no longer treat localhost.. as loopback; localhost. remains supported.
  • Fixed an issue where final telemetry could be dropped during shutdown while idle search daemons were stopping.
  • Plugins installed with session.options.update({ installedPlugins }) now start their MCP servers, and removed plugins' servers stop.
  • Fixed workspace diffs and added-line counts for untracked files to respect Git's text/binary attributes and diff driver binary settings.
  • Remove the unused experimental brokerAvailable account-enumeration input, including Python's broker_available constructor slot; host-owned Entra token providers remain supported.
  • Java: a system-message section transform callback that throws a non-fatal Error (such as AssertionError) now keeps its section, so the turn completes instead of hanging the session.
  • Java: an MCP auth handler that throws a non-fatal Error (such as AssertionError) is now answered cancelled, so the server moves to needs-auth instead of staying pending for the rest of the session.
  • Remote session export now resumes from what Mission Control already stores instead of re-uploading history after a crash or an idle resume.
  • Node.js: session events that arrive before a session registers (after sessions.connect, on rejoin, or during cloud create) are no longer dropped; requests already answered by another client are shown to observers but not re-prompted, and held events never outlive their connection or a deleted session.
  • Preserve asynchronous MCP task results after recoverable polling failures and interrupted blocking reads.
  • Go: an empty ExcludedBuiltInAgents slice passed to ResumeSession is now sent instead of dropped, so resuming a loaded session with it clears the excluded built-in agents, as in the other SDKs.
  • Sandboxed traffic now routes and authenticates through upstream proxies with the client's network callbacks (resolveProxy, authenticateNegotiate, authenticateBasic), as the runtime's own requests do.
  • Fixed self-paced schedules marked with origin "user" to invoke leading skills on every tick, including skills with disable-model-invocation enabled.
  • Add shared sandbox configuration, settings, certificate, and live session controls with generated clients in all six SDKs.
  • Stop automatically retrying additional biological, cybersecurity-access, and image-content policy refusals from model providers.
  • Language server connections now use the last duplicate Content-Length header, reject colon-less header lines and invalid header names, and tolerate invalid UTF-8 bytes in header values. Protocol-conforming traffic is unchanged.
  • Fixed missing subagent.completed events with cancelled: true for synchronous sub-agents stopped through their parent, preventing SDK hosts from showing them as still running.
  • BYOK sessions for known models (e.g. modelId: "gpt-6.1-sol") now default to the model's standard-tier prompt limit instead of 128k when maxPromptTokens is omitted, and expand to the native limit with contextTier: "long_context".
  • Added optional managedControls and SandboxManagedControl types to sandbox.decision policy_resolved events for enterprise-enforced sandbox controls. Rust source-breaking change: exhaustive SandboxDecisionDataPolicyResolved literals must add managed_controls: None or use ..Default::default(); destructuring patterns must bind the new field or use ... Optional JSON decoding and sandbox enforcement are unchanged.
  • Fixed Bash and PowerShell calls rejected because a shell is busy disrupting the running command's output stream. Shell output and partial results now retain the original toolCallId, and output sequence numbers continue without restarting.
  • Assisted Permissions now reuses eligible human path approvals for the same read, avoiding duplicate approval prompts without granting future access.
  • Honor smaller advertised model/provider output limits when choosing implicit Anthropic token budgets.
  • Clarify session errorCode documentation without changing its optional, extensible string contract.
  • Session resume events record whether the resume kept Copilot Responses reasoning (responsesReasoningKept) and whether the strip override decided it (responsesReasoningOverridden).
  • Correct rounding in session continuation-cost and credit displays to match JavaScript; for example, 0.995 now displays as 0.99.
  • Fix MCP progress rounding near half-percent boundaries and use JavaScript-compatible number formatting in tool output and protocol error messages.
  • Preserve cold-resumed system prompts when tools are initialized before the first turn.
  • Fixed plugins.marketplaces.remove to return an "is not registered" error for marketplaces declared only in repository settings, rather than a misleading removed: false result with dependentPlugins when forced removal cannot succeed.
  • Session sandbox config accepts allowTempStorage: false to withhold the automatic temp-storage grant; SDK hosts can observe the setting through OpenTelemetry.
  • Fixed MCP OAuth sign-in failing when a server grants all required scopes but omits the automatically requested optional offline_access scope. Compatible saved registrations are reused without requesting the declined scope again.
  • Runtime shutdown no longer starts another search-cache maintenance pass after cancellation.
  • Fixed explicit Gemini and MAI generic tool-search environment flags being overridden by experiment assignments.
  • Rewinding a turn with file restore now also restores files that a shell command wrote after changing directory in the same command.
  • Preserve workspace edits made after tool initialization and before the first turn of a resumed session.
  • Fixed tool calls hanging on named pipes (FIFOs) when organization content-based exclusion rules apply. These paths now receive an immediate organization content policy denial instead of being read.
  • Cancelling a pending resumeSession call in the Go, Python, Rust and Java SDKs now cancels the resume in the runtime.
  • Fixed file-edit results being unnecessarily replaced by large-output notices when post-edit syntax checks are enabled. Edit summaries now stay inline when space permits, with oversized syntax-check notes saved separately.
  • Preserve client tool-search references in tool completion events so session resume can replay the original search result.
  • Native extensions can declare slash commands for request-driven activation.
  • Sandboxed Windows PowerShell sessions now receive scratch-file guidance that points PowerShell commands to $env:TEMP when that directory exists and temp storage is not forbidden. The sandbox policy still decides what is writable.
  • .NET: Reduce outgoing JSON-RPC framing allocations by serializing directly into pooled buffers.
  • .NET: Remove settings-specific integer validation from the JSON-RPC transport.
  • Fixed model-initiated MCP tool calls failing when servers return tasks for tools without required task support. When both sides enable MCP Tasks, these calls now run as cancellable background tasks and deliver results to the model; direct session.mcp.callTool behavior is unchanged.
  • Subagent declaration resolution preserves Copilot model policy filtering after refreshing its catalog.
  • The .NET SDK preserves supported nested experiment data and tool metadata in create and resume requests when reusing JSON-RPC framing buffers.
  • Fixed secret redaction corrupting JSON text embedded in event and callback string values, preserving parseability across nested JSON encoding while continuing to redact credentials.
  • Fixed an unnecessary delay of up to a minute before the agent receives completion notifications for overlapping background shell commands.
  • Fix cold resume with continuePendingWork: true to continue the model after interrupted ask_user or elicitation requests, completing recorded requests to dismiss stale prompts without replaying interrupted tools.
  • Ensure recovered turns emit assistant.turn_end exactly once after all tool calls settle, including parallel tools that complete later.
  • Remote server discovery now retains servers when liveness checks are inconclusive and correctly excludes exited Windows processes.
  • Local-model server reuse now rejects invalid Unix process IDs and requires successful process-query access on Windows.
  • A custom provider header with a short value, such as "1", no longer turns every occurrence of that text in runtime logs and error messages into asterisks.
  • SDK sessions now back off repeated forced compactions that do not meaningfully reduce conversation size after context-limit errors, reducing unnecessary summarization calls and delays while continuing to retry with truncation.
  • Extensions can no longer change the session sandbox through session RPC; joinSession with permissions.sandbox is rejected.
  • Fixed PowerShell commands using a leading & incorrectly treating native switches such as /d and /c as file paths for permission checks, while preserving checks for actual paths and explicit path arguments.
  • File attachments in public event reads and live notifications now remain filesystem references.
  • Fixed inflated token estimates for JPEG attachments with large metadata, preventing unnecessary auto-compaction and prompt-budget errors.
  • Honor the effective host Git policy and workspace filesystem when resolving roots for hook initialization and plugin-hook reloads, preserving the established hook project root when Git is disabled.
  • Fixed web_fetch rejecting public IPv4 hosts on IPv6-only networks using the well-known NAT64 prefix (64:ff9b::/96). Restricted IPv4 destinations remain blocked, and the local-use NAT64 prefix (64:ff9b:1::/48) is now blocked.
  • Added mai-code-2 to SupportedModel and SupportedModelsTuple, with default reasoning effort set to medium. MAI-Code-2 is not selected automatically as the default model.
  • Fixed session.provider.getEndpoint for sessions configured with providers and models: it now reports the provider that serves the requested model instead of an authentication error or the Copilot API endpoint.
  • The sessionLimits.maxAiCredits documentation now states the minimum of 30 AI Credits that session create, resume and options update already enforce.
  • Preserve the first top-level validation-detail message from compatible provider errors without retaining the surrounding validation context.
  • Subsequent SDK sessions avoid redundant Git discovery after automatic plugin activation reconciliation; working-directory changes, explicit reloads, and plugin content changes still refresh Git discovery.
  • Permission decisions accept optional selectedOptionId for AHP confirmation options.
  • Add runtime-owned file logging and text or JSON redaction through the SDK.
  • Fixed MCP uninstallation failing after a remote installation's configuration entry is deleted; cleanup now removes installation-owned secrets while preserving unrelated configuration and shared authentication.
  • Session history restores blob attachment bytes while file attachments remain path references.
  • Fixed a race that could prevent re-enabled plugins from reinstalling and restoring their skills when activation changes during startup.
  • Restored support for legacy @github/copilot-sdk and @github/copilot-sdk/extension imports in ESM and CommonJS extensions alongside the @microsoft/copilot-sdk equivalents.
  • .NET: Log handled I/O failures when sending JSON-RPC responses and cancellation notifications through the configured ILogger.
  • Fixed .NET SDK background readers retaining request-scoped AsyncLocal values for the session or connection lifetime, while preserving event handlers' own context across awaits.
  • Added mai-code-2-flash to SupportedModel and SupportedModelsTuple and supported it for session limits when available to your account. Automatic model selection is unchanged.
  • Fix Azure default-chain token caches leaking between independent in-process SDK runtimes.
  • Fix first-use MCP Apps tool and resource requests for user-configured servers in fresh and resumed sessions without blocking concurrent session joins during startup.
  • Added per-result large-output control across all six SDKs: set skipLargeOutputProcessing to true to bypass automatic spill-to-file processing. False or unset preserves the session's normal policy.
  • Java ToolResultObject record patterns must include the new skipLargeOutputProcessing component; existing constructors remain available.
  • Fix .NET SDK compilation after the Microsoft.Copilot namespace rename.
  • Fixed resumed sessions with stable prompt prefixes to reflect artifact renames, additions, and deletions made while stopped, while preserving the saved system prompt.
  • Fixed permission denials without feedback during a foreground sub-agent's initial turn to stop the main agent's turn, preventing it from continuing and requesting the same access again.
  • With sandbox.allowDevToolAccess enabled, sandboxed Rust commands can now write the entire Cargo home, including configuration, credentials, and installed binaries, removing sandbox write restrictions that blocked cargo install. Relocated CARGO_HOME paths are supported when set in the Copilot process environment; explicit read-only and denied paths still take precedence.
  • Instruction discovery now returns distinct session-compatible IDs for repository and working-directory rules with matching paths.
  • Fusion Single turns now stream ordinary text, reasoning, and tool events while preserving standard Fusion resolution and completion metadata.
  • Rust source break: experimental SessionFusionCompletedData adds error_message, failure_reason, and reasoning_effort (optional strings). Update exhaustive literals with None for absent values or ..Default::default(), and exhaustive patterns with the new fields or ... Missing JSON fields still deserialize as None.
  • Added the experimental session.sensitivity_label_observation event, reporting per-request outcomes only when Microsoft 365 supplies valid sensitivity information.
  • Exclude the 16 reported eMudhra CA keys from macOS/Linux runtime trust anchors, alongside the existing ECC G3 exclusion.
  • Added gpt-6.2-sol to supported SDK model types, selectable when advertised by the service; automatic defaults are unchanged.
  • Enabled custom tools, web search, explicit prompt caching, and reasoning-effort and tool updates for gpt-6.2-sol sessions.
  • Added support for selecting GPT-6.2 Astra (gpt-6.2-astra) in SDK sessions when available, without changing automatic defaults.
  • Extended optional GPT-6 Luna compaction to GPT-6.2 Astra sessions; explicitly configured compaction models still take precedence.
  • Update SDK dependencies.
  • Resuming a session compacted with history.compact rebuilds the same summary message the live session sent, instead of rewording its first line.
  • Reduce redundant usage-checkpoint writes during AHP host shutdown.
  • plugins.install now returns the reason in warnings when a plugin's Agent Plugins manifest is rejected, and plugins.update and plugins.updateAll results gain an optional warnings field carrying the same manifest findings.
  • Fixed missing completion events when nested agents are cancelled during concurrent cleanup.
  • Fixed an issue where reselecting an agent could incorrectly reject valid MCP tool calls with a catalog-change error while its MCP server was connecting.
  • Preserve newer AHP session catalogs when older refresh responses arrive late.
  • The initiator field of assistant.usage and model.call_failure events is now documented as "user" for a request that carries a user prompt, "agent" for agent-driven requests, or the delegating caller such as "sub-agent"; it was wrongly described as absent for user-initiated calls.
  • Nested synchronous agents report canceled completion events when their ancestor stops.
  • Fixed Bash shell permission checks to resolve $PWD using the command's working directory, preventing incorrect denials and missed approval requests for paths outside that directory.
  • Fixed a cancellation race that could cause SDK client shutdown to time out while background search discovery was running.
  • Faster first hook or shell command on Windows: the runtime no longer launches PowerShell just to read its version.
  • SDKs can select Assisted Permissions without a flag opt-in.
  • Fixed unscoped /chronicle search, tips, standup, and improve to use local session-history guidance when cloud credentials are unavailable, including offline and unsigned-in BYOK sessions.
  • Session-history queries that fail during local fallback now report the underlying query error and guidance for retrying with local SQLite syntax.
  • Restored connected traces for asynchronous tool callbacks on .NET 8.
  • Session error events report the provider's exact error code and type for every provider, a typed dlpBlock for prompt_dlp_blocked and inline_restricted, and the failing turnId.
  • Fixed live session deletion emitting an invalid shutdown type. Shutdown handlers now receive session.shutdown with shutdownType: "routine", preventing decoding errors such as Python's ValueError.
  • Deleted sessions now report routine instead of destroy in shutdown usage telemetry and the github.copilot.shutdown_type OpenTelemetry attribute, making them indistinguishable from other routine shutdowns by that field.
  • Fixed commands.changed events dropping commands from extensions that have not yet started when model availability changes.
  • .NET SDK now ships net10.0 and netstandard2.0 assemblies; the net8.0-specific assembly is no longer included. .NET 8/9 consumers must rebuild dependent libraries compiled against that asset if they call SDK init-only setters.
  • Restore support for older option formats in existing Node SDK extensions and direct RPC clients, without weakening v2 TypeScript types.
  • Added an optional reason field to the permissions.setMode result, set to managed_policy when managed policy refuses the mode.
  • Enable hybrid prompt caching by default for canonical GPT model versions 5.6 and newer, including future generations.
  • Fixed stale abortability after remote AHP sessions receive a normal idle notification.
  • Stop instructing models to volunteer model-switch acknowledgments; explicit model-identity guidance remains unchanged.
  • AHP clients can display live tool output through non-PTY terminal subscriptions while preserving ordinary tool results.
  • Preserve captured tool output, structured exit/truncation metadata, and terminal identities across cancellation, replay, restart, cache eviction, and chat moves, within retention limits.
  • Publish interrupted tool results before cancellation or error events after captured output is durably acknowledged.
  • Reject chat moves when the destination cannot retain the conversation's tool output.
  • Keep connections responsive while follow-up chat actions wait for pending cancellation to settle.
  • Add Backend hooks for transient chat unloading, delivery-aware dispatch observation, and revalidation of dequeued work; remove_chat now signals permanent retirement.
  • AHP clients receive PTY title updates, with rapid changes coalesced and private titles withheld until session access is authorized.
  • Permission handlers that return approved, approved-for-session or approved-for-location now work instead of failing tools with "unexpected user permission response".
  • Security review fine-tuned model scanning now uses the session's GitHub credential (gitHubToken/gitHubTokenProvider) when the client has no signed-in account.
  • Fixed connections to slow-starting modern stdio MCP servers that reject the legacy initialization fallback.
  • Keep resident skill provider callbacks available while a provider-free resume is pending.
  • Fix native extension reload readiness, external reload overrides, and queued reload diagnostics.
  • Removed retired Claude Opus 5 and GPT-5.6 Terra from the built-in supported-model catalog and the SupportedModel and SupportedModelsTuple types.
  • GPT-6 Astra sessions now use Claude Opus 5.5 instead of Claude Opus 5 for the complementary rubber-duck critic.
  • ACP clients can now provide session-scoped stdio MCP servers through session/new and session/load, preserving command arguments and literal per-session environment values. Both omitted transport types and explicit type: "stdio" are supported.
  • Allow skills.config.setSkillDisabled to select an explicit configuration directory.
  • Session shutdown waits for in-flight local SQLite work to release the session database before acknowledging close.
  • Assisted Permissions no longer depend on the AUTO_APPROVAL feature flag; an explicit AUTO_APPROVAL=false is ignored with a deprecation warning.
  • AssistedApprovalAssessment with status: "failed" can now carry an optional message explaining the failure, such as the judge model's error or how to make a judge model available.
  • Restores the deprecated Rust SDK per-session with_remote_session builder for create and resume options.
  • Prevent startup hook reloads from failing while managed plugin policy becomes available.
  • A sandboxed shell command that fails with a Rust "failed to create directory" permission refusal for a path the sandbox policy denies can now raise an optional per-command sandbox-bypass permission request (requestSandboxBypass).
  • Fixed unsandboxed extension hosts on macOS and Linux remaining alive and holding listening ports after their parent runtime exits.
  • Stopping an in-process client now waits for cancelled local SQLite work to finish.
  • Fixed plugin agents remaining missing in existing sessions when a policy is re-enabled after an interrupted policy withdrawal.
  • Sending a message with agentMode: "plan" now blocks workspace writes outside the plan file until the plan is approved, as plan mode does.
  • Host-only agent discovery paths no longer fail when the process working directory disappears.
  • Fixed local Windows shell commands in SDK sessions being incorrectly blocked from reading UNC network shares when Content Exclusion rules are present; normal shell permissions and sandboxing still apply.
  • session.extensions.reconcile now succeeds when an enabled extension fails to start, and reports that extension with status failed; stop, cleanup and policy failures still return errors.
  • Read-only tools from non-built-in MCP servers now go through the permission handler instead of being auto-approved.
  • Fixed MCP HTTP session loss detection stalling when a server leaves a timed-out probe's cancellation unanswered.
  • Added gemini-4-neon to SupportedModel and SupportedModelsTuple for accounts with model access, with a 1,048,576-token context window, 65,536-token output limit, and Low through Extra High reasoning; Medium is the default.
  • Gemini 4 Neon now preserves Extra High reasoning instead of downgrading it to High when the existing Interactions transport is enabled.
  • Newer OpenAI models inherit image token estimation profiles from the nearest eligible predecessor unless explicitly overridden; GPT-6.2 Sol and Astra use profiles validated against live Responses usage.
  • Sandboxed agents are now guided to use device-code or non-interactive sign-in when browser sign-in cannot complete, and report the failure if no alternative succeeds.
  • Restoring files when rewinding a session now also reverts a file the rewound turns edited through a symlink they created or re-pointed.
  • When enabled, stronger search-subagent guidance favors delegated codebase exploration. Guidance remains stable across session resumes, and built-in tool descriptors reflect it.
  • Sessions offering the search subagent now emit a session.info event recording whether exploration guidance is enabled.
  • Replace experimental account enumerate/get/set APIs with concrete operations and descriptive account/provider types.
  • Sped up language-server file watcher registration in large Windows workspaces.
  • Add experimental session managed-settings refresh to fetch fresh policy and await tracked enforcement without restarting the session.
  • Automatic and manual compaction now consistently use medium reasoning effort for eligible GPT requests at high, xhigh, or max effort. Explicit compaction reasoning settings still take precedence; session effort, conversation history, and other models are unchanged.
  • Adds plugin provenance for providers contributed by declarative plugin provider adapters.
  • Fixed session.abort() cancelling pending extension capability and environment-access consent requests. Requests now remain answerable while the extension stays connected, allowing approval after a turn is stopped.
  • Include destination hostnames in sandbox URL-policy denial errors for host-rule and outbound-access blocks.
  • Node.js: Fixed graceful shutdown when stop cancels a pending startup.
  • All SDKs: Fixed runtime.shutdown to return its documented null result after cleanup.
  • Staged HydraFusion phases now stream provisional assistant.message_start, assistant.message_delta, assistant.reasoning_delta, assistant.tool_call_delta and assistant.reasoning events that carry fusion attribution without a commitId.
  • Prevent repeated sandbox directory-grant requests for files in the same patch after the host accepts a shared grant.
  • Reduce native session setup overhead and avoid unnecessary response and event copies in the Rust SDK.
  • A session without an account now reports a send as an authentication error with errorCode: "no_account" and remediation: "sign_in", and session.model.list and session.quota.refresh rejections carry the same fields in their error data.
  • Fixed GitHub CLI credential selection in sandboxed shells for repository and hostname flags, aliases, extensions, option values, and --.
  • GitHub CLI commands that only display built-in help no longer trigger credential lookups.
  • When information-flow controls are enabled, improved GitHub CLI repository targeting and stopped treating help-only commands as writes.
  • A resumed session keeps the permission grants its earlier prompts made for the rest of the session, instead of asking again.
  • With a session AI credit limit, the cost meter after a successful tool result is no longer labelled as postToolUse hook guidance in the text sent to the model and in tool.execution_complete result content.
  • Add a runtime RPC to activate a built-in plugin's sessionStart hooks without creating a session.
  • Clarified that a permission.completed result of denied-interactively-by-user does not mean a person decided; read decisionSource.
  • Sandbox sessions return to direct Azure Web PubSub connections, which no longer inherit the host’s HTTP proxy settings or trusted root certificates.
  • Preserve pending TCP shutdown responses while rejecting new connections during shutdown.
  • Close SQLite connections opened by session database work after cancellation.
  • Return a coherent skill catalog when a local catalog changes during skill loading.
  • Fix MCP credentials accumulating in the secret filter after their sessions end.
  • Add secrets.removeFilterValues and release client-added filter values on disconnect.