./revenue-analyst start --store s3://bucket/prefix

A daemon started elsewhere with the same store waits out a killed holder’s lease, restores the data and carries on.

serve apps can also require a credential on agent routes: static API keys from code, serve.toml or SERVE_API_KEYS, OIDC, or OAuth 2.0 introspection. The agent card stays public. At boot, serve resumes sessions active in the last seven days and re-runs an unfinished tool call only when that is safe.

See Keep the data in a bucket, Authentication and #4515, #4518, #4485.

This fits Slack channels and conversations with several people, where the agent should answer only when it has something to add. The setting is on the agent in the web UI, the API, agent packages, the Framework builder and the CLI, and it replaces Slack’s per-endpoint reply mode. Every sent message is recorded as a conversation.message event.

See Explicit communication and #4447.

Budget summaries also carry the budget ID, so a client can tell which of several stacked budgets stopped the turn. To continue, top up the exhausted budget with the budget API or CLI.

See Budgets and #4477.

In the UI: an organization administrator turns on Voice in Settings → Features. Then add a voice channel from the agent’s Integrations tab and use its talk button. Speech uses an OpenAI provider.

In code: everruns::voice puts a voice channel in front of any Framework session, and serve’s voice feature gives every top-level agent browser calls.

See Voice, Framework voice and #4394, #4402, #4406.

• Codex session reliability: Legacy Codex sessions keep host observability metadata and the session HTTP header while omitting unsupported request-body metadata, fixing ChatGPT sessions that failed with HTTP 400 (#4464).

• Agents home and a New agent page with an agent builder
• Fix ChatGPT plan sessions failing with HTTP 400
• See and revoke AI clients connected to Everruns
• Models and providers move into one registry page
• Voice channels let people call an agent

See the GitHub release and the full release notes for every pull request in this version.

• Agents home: Enable the Agents home feature for run activity, setup alerts and a shared Channels view (#4399).
• Agent scripts: Save and manage reusable shell scripts on agents through the API, CLI and MCP, then run them with tools scripts using the caller’s tools and identity (#4390, #4400).

• Save shell scripts on agents and run them from triggers

See the GitHub release and the full release notes for every pull request in this version.

• Tools in Shell: Call visible tools from a Bashkit script, load deferred MCP servers, and pause for approvals before the script starts (#4385, #4386, #4388).
• Browser tools: Provider-neutral Browserless browser tools and a native Claude browser toolset adapter (#4375, #4376).
• Coordinator threads: Coordinate child threads in Framework applications and Platform Chat (#4377, #4372).

• Browser tool for agents on Browserless
• Coordinator threads in Platform Chat and the Framework

See the GitHub release and the full release notes for every pull request in this version.

Everruns now leaves that metadata out of the request body while keeping the session HTTP header and host observability metadata. No configuration change is needed; existing ChatGPT connections work after upgrading.

See ChatGPT plan and #4464.

• A2A 1.0 HTTP+JSON and PACT: A2A channels serve the HTTP+JSON binding, the PACT Identity profile, and PACT Delegated sign-in, step-up and receipts (#4322, #4328, #4350, #4352)
• AgentID sign-in: Public chat and channels accept AgentID, and agents sign in with their AgentMail inbox (#4334, #4326, #4337)
• MCP in conversations: Users manage their own MCP servers, connect them mid-conversation, and deferred servers load tools on demand (#4315, #4343, #4345, #4349)
• Claude Haiku 5.5 and Daytona desktop computer use (#4359, #4365)

• Agent versions are retired in favor of change history
• A2A channels serve HTTP+JSON and the PACT profiles
• Sign in to channels and Public Chat with AgentID
• Claude Haiku 5.5 model added for Anthropic providers

Before upgrading, read the breaking changes: Agent versions are retired in favor of change history.

See the GitHub release and the full release notes for every pull request in this version.

In code: the Framework’s everruns::coordination capability runs threads as sessions of the same engine. Threads wake the coordinator with automatic updates, and coordination::threads returns each thread’s status, such as NeedsYou, Working or ReadyForReview.

See Chat and threads, Coordination and #4372, #4377.

See the GitHub release and the full release notes for every pull request in this version.

Its profile has tiered pricing, up to 128K output tokens, always-on adaptive thinking with medium effort by default, preserved thinking, tool search and the computer toolset. Select it in an agent’s model picker or reference claude-haiku-5-5 wherever a model ID is accepted.

See Anthropic and #4359.

• Agents get avatars from presets or an uploaded image
• Sandbox Templates and a Sandboxes fleet view
• Change history with reasons and restore for every entity

See the GitHub release and the full release notes for every pull request in this version.

• Vendor driver crates moved into everruns-drivers, one feature per vendor (0.35).
• Provider, capability and model-profile APIs moved into everruns-contracts (0.37).
• everruns-platform became everruns-capabilities (0.39).
• Engine, host, builtins, MCP and AG-UI became everruns-core features (0.40), and their shims were removed in 0.41.

everruns-drivers = { version = "0.41", features = ["openai"] }

No capability was removed, and published shim versions are not yanked.

See the upgrade notes. PRs #4035, #4077, #4081, #4083, #4084.

• Framework crates consolidated into everruns-core and the facade

Before upgrading, read the breaking changes: Framework crates consolidated into everruns-core and the facade.

See the GitHub release and the full release notes for every pull request in this version.

Everruns 0.40.0 is a maintenance release with fixes and internal changes and no user-facing entry of its own.

See the GitHub release and the full release notes for every pull request in this version.

• Chat threads keep conversations and ongoing work beside Chat
• Slack permission issues appear in Health with recovery steps

See the GitHub release and the full release notes for every pull request in this version.

• Personal ChatGPT plan connections: Self-hosted deployments can enable personal account sign-in and discover its available models; connections stay scoped to the signed-in user (#4136).
• Stock subagents for custom hosts: Library hosts can run the stock subagent and handoff tools through a store of portable runtime definitions, without constructing control-plane records (#4078).

• Connect a personal ChatGPT plan as a model provider
• Conversation becomes the default harness and Generic is deprecated
• Feature flag grade preview renamed to internal for operators
• Playground for testing agents, and Chat as one permanent conversation

Before upgrading, read the breaking changes: Conversation becomes the default harness and Generic is deprecated, Feature flag grade preview renamed to internal for operators.

See the GitHub release and the full release notes for every pull request in this version.

• Approval history in context: Chats show approval requests and recorded consent; organization admins and owners can filter existing spoken-consent audit history (#4110).
• Agent channels at a glance: Compact agent cards show transport and live, draft or disabled status, with links to Integrations (#4109).
• Navigation keeps your place: Sidebar selection updates as navigation starts, and agent tabs survive refresh through their URL (#4108, #4111).
• Slack installation returns to setup: Valid OAuth callbacks return to the owning agent’s endpoint editor; invalid callbacks use the agents page (#4107).

• Approval history in chat and a new Approvals audit page

See the GitHub release and the full release notes for every pull request in this version.

• Choose an Agent Environment: Author named Bashkit or Daytona profiles and select the profile when starting a Chat or Playground session.
• Slack installation repair: Add to Slack works for draft endpoints, requests declared bot scopes, and shows compact errors with retry.
• Session storage permissions: Listing stored values or secret names requires session view; denied callers receive no private data.

See the GitHub release and the full release notes for every pull request in this version.

• Safer sandbox tools: HTTP egress fails closed on empty DNS pins, and forged session secrets cannot redirect Docker tools (#4058, #4061).
• Interactive questions in chat: AskUser runs through interactive turns so agents can ask for input while working (#4044).
• Platform Chat for every organization: the shell-powered Platform Chat harness becomes the default for every organization (#4060).

• Computer use capability lets agents operate a browser screen

See the GitHub release and the full release notes for every pull request in this version.

Generic is deprecated but unchanged. Selectors hide it behind Show deprecated. On upgrade, agents that inherited Generic as the default are pinned to it, so their behavior does not change. A custom harness whose name collides with a new built-in keeps its ID and definition, but its slug moves to <name>-custom; update callers that look it up by name.

In code: Harness::base(), conversation(), worker_base() and worker() build the same presets.

See Harnesses and the upgrade notes. PR #4133.

The new everruns-serve-agentcore crate hosts a serve app on Amazon Bedrock AgentCore Runtime. It answers GET /ping and POST /invocations, which take an AG-UI RunAgentInput or {"prompt": ...} and stream AG-UI events, plus a /ws WebSocket transport, with serve’s /v1 API alongside.

The session log and agent workspace live on AgentCore session storage. Bedrock models use the runtime’s execution role when AWS_REGION is set, and parked approvals and ask_user questions survive a restart.

See Serve on AgentCore. PRs #3990, #4012.

The celld example adds the app, a Dockerfile and the celld worker: a Durable Object per name that journals mutating requests, snapshots the data directory when idle, and restores and replays when its container is fresh. An end-to-end test kills serve mid-turn, and then the node itself.

See Serve on celld. PR #4019.

It ships its own idempotent PostgreSQL schema through PostgresWorkflowEventStore::migrate and SCHEMA_SQL, tested to match what the server migrations build. Timers and child workflows run as engine tasks: ActivityOptions::start_delay holds a task until it is due, and parents get on_child_workflow_completed and on_child_workflow_failed.

cargo add everruns-durable

See Durable execution. PRs #3994, #4008.

• Execution kernel in a JavaScript isolate: the engine’s execution kernel runs inside a JavaScript isolate, with a celld engine cell (#4024)
• Durable serve apps on celld: new everruns-serve-celld hosting target snapshots and restores a serve app across process and node loss (#4019)
• Workers AI model discovery: the Cloudflare driver lists Workers AI models, and Chat Completions streams keep tokens sent as JSON numbers (#4021)

• Cloudflare AI Gateway and Vercel AI Gateway model providers
• Run an Everruns serve app durably on celld

See the GitHub release and the full release notes for every pull request in this version.

• everruns-durable workflow engine published on crates.io
• Host Everruns serve apps on Amazon Bedrock AgentCore Runtime

See the GitHub release and the full release notes for every pull request in this version.

• AG-UI 1.0: A new everruns-ag-ui crate carries the AG-UI 1.0 wire types (#3964), and the AG-UI endpoint moves to protocol 1.0 with interrupts, resume, token usage, frontend tools, subagents and run metadata (#3965, #3967, #3972, #3979). Any framework session serves AG-UI behind the ag-ui feature, with a durable thread store and a ready-made axum handler (#3976, #3998, #4000), everruns-serve exposes an AG-UI channel (#3978), and a consumer pipeline and SSE client ship with an upstream conformance corpus (#3982).
• A2A 1.0: The A2A endpoint speaks A2A 1.0 and returns task results (#3999), with ListTasks, SubscribeToTask (#4010) and push notifications (#4013); everruns-serve serves every agent over A2A behind an a2a feature (#4014).
• Amazon Bedrock AgentCore hosting: The new everruns-serve-agentcore crate hosts agents on AgentCore Runtime, with Bedrock through the execution role, WebSocket transport and durable approvals (#3990, #4012).
• OpenAI Agents API backend: An opt-in durable OpenAI Agents API backend (#3957) projects events, usage and cost (#3961), supports session lifecycle and portability (#3963), and enforces Everruns policy at tool boundaries (#3959).
• Native computer use with hard approval: Native OpenAI and Anthropic computer tools run behind a hard tool-approval gate (#3968, #3956).
• Slack workspaces per organization: Organizations connect Slack workspaces and pick one per agent (#3977).
• Agent version pinning: Endpoints and triggers can pin a specific agent version (#3975).
• More ways to reach models: Cloudflare and Vercel AI Gateway drivers (#3996), an opt-in Responses WebSocket transport for OpenAI (#3981), and OpenAI background responses that re-attach after a worker restart (#3969).
• GitHub and MCP triggers: PR reviewer and security scanner templates on the GitHub App (#3980), and agents that wake on inbound MCP events (#3970).
• everruns-durable on crates.io: The durable workflow engine fires timers, runs child workflows and ships with a bundled PostgreSQL schema (#3994, #4008).

• PR reviewer and security scanner agent templates for GitHub
• MCP Events add session webhooks and triggers on MCP server events
• OpenAI structured output and background mode for long reasoning
• Connect several Slack workspaces per organization
• A2A 1.0 endpoint with task results, streaming and push notifications
• AG-UI 1.0 endpoint and AG-UI streaming for Framework sessions

See the GitHub release and the full release notes for every pull request in this version.

Two gateway providers are available. Cloudflare reaches many upstream providers and Cloudflare’s own Workers AI models through one account token and account ID. Vercel AI Gateway uses its OpenAI-compatible Responses endpoint and syncs its model list.

The Cloudflare provider lists Workers AI text-generation models with their context window and tool support. Third-party model IDs on Cloudflare are still added by hand, because no endpoint lists what an account can bill.

This also fixes a token dropped from Cloudflare streams when the model emitted a purely numeric token.

See Cloudflare and Vercel. PRs #3996, #4021.

In code: the ag-ui feature lets any Rust host stream a Framework session to an AG-UI client without the Everruns server. Session::ag_ui(input) returns the run as AG-UI 1.0 events, and ask_user questions and tool approvals become interrupts. A durable thread store keeps history between runs, and the ag-ui-axum feature adds a ready-made route:

cargo add everruns --features ag-ui-axum

See AG-UI channels and Serve AG-UI. PRs #3965, #3967, #3972, #3979, #3976, #3998, #4000.

In code: the a2a feature of everruns-serve answers A2A 1.0 for every top-level agent at POST /v1/e/{agent}/a2a, the same paths an Everruns A2A endpoint uses.

See A2A and Framework A2A. PRs #3999, #4010, #4013, #4014.

Calls at xhigh and max reasoning effort on OpenAI and Azure run as background responses. If the stream drops or the worker restarts, Everruns picks the response back up instead of paying for it twice. A cancelled turn cancels the response, and the openai/background driver option overrides the policy.

See OpenAI. PRs #3932, #3969.

• GitHub pull request triggers and per-agent GitHub Apps: Agents connect their own GitHub App by clicks and act on pull requests, with PR triggers, a Connect GitHub card, PR tools, and a PR summarizer guide (#3935, #3943).
• OpenAI Responses hosted tools and remote MCP: The OpenAI Responses driver gains hosted web search (#3918, #3923), hosted code interpreter, shell and file search (#3928), remote MCP with approval and registered-server credentials (#3930), and structured output with background mode (#3932).
• Agent trigger event pipeline: A new event pipeline with filters, dedupe, and per-subject sessions (#3915), plus outbound MCP Events for session webhooks (#3931).
• MCP Apps for ChatGPT and Codex: Everruns serves MCP Apps views for ChatGPT and Codex (#3927), answering form-mode elicitation through ask_user (#3929).
• Pluggable decision drivers: Decision drivers route behind one router (#3921), with the OpenAI Decisions API available as a preview driver (#3924).
• Provider-neutral computer use: A provider-neutral computer tool runs on Browserless (#3922).
• New models: Claude Sonnet 5.5 joins as the favorite Sonnet (#3902) and GPT-6.1 Sol arrives with fast and ultrafast speed tiers (#3925); the sunset Claude Sonnet 4.5 profile is dropped (#3952).
• Rebuilt agent workspace: The agent page is rebuilt as a single view-and-edit workspace (#3917).

• Agent page rebuilt as one view-and-edit workspace
• Everruns MCP server renders as an MCP App in ChatGPT and Codex
• Claude Sonnet 5.5 and GPT-6.1 Sol with fast and ultrafast tiers
• GitHub pull request triggers with a per-agent GitHub App
• OpenAI hosted code interpreter, shell, file search and remote MCP

See the GitHub release and the full release notes for every pull request in this version.

GPT-6.1 Sol joins the OpenAI models. Its controls.speed setting accepts fast and ultrafast. The OpenAI drivers reject a speed tier the model does not list before sending, and usage is priced at the tier that served the call.

Anthropic sunset Claude Sonnet 4.5, so its model profile and catalog entry are removed.

See Anthropic and OpenAI. PRs #3902, #3925, #3952.

• Per-organization Slack app provisioning: Each organization can provision its own Slack app rather than sharing one deployment-wide install (#3830), with concurrent installs for one endpoint serialized so a race can no longer create orphan apps or exhaust the config token (#3843).
• Agent app framework over Engine: An experimental framework for building agent apps directly on the Engine (#3802), with engine-level event listeners (#3804) and pluggable observability integrations (#3806).
• Server-side context compaction: Anthropic conversations compact context server-side, keeping long agent runs within budget (#3831).
• Ask User through tasks: The Ask User capability surfaces through the tasks channel, extending it beyond Platform Chat and MCP (#3828).
• Server hardening: Platform Chat question-answers and tool-results now require the session owner (#3842), unrequested reasoning summaries stay hidden from responses (#3840), feature-flag revocations take effect immediately (#3820), retained observer event memory is bounded (#3837), and native endpoint sessions are reused rather than rebuilt (#3814).

• Ask User questions reach MCP, A2A and Tasks clients
• Framework engines get event listeners and OpenTelemetry export

See the GitHub release and the full release notes for every pull request in this version.

An Engine now delivers every event from every session it owns, including resumed and spawned ones, to the listeners registered on it. You no longer poll Session::events() per session. Each listener drains a bounded ordered queue, and shutdown flushes them within one deadline.

The everruns crate ships two listeners behind independent features, otel and braintrust:

let engine = Engine::builder()
    .observe(OpenTelemetry::from_env())
    .build();

Message content is not exported by default.

See Observability. PRs #3804, #3806.

• Ask User over MCP and A2A: The Ask User capability now reaches agents beyond Platform Chat: it projects as form-mode elicitation over /mcp (#3781) and is carried over the inbound A2A channel as a DataPart (#3780).
• Always-thinking model reliability: Always-thinking models no longer return 400s or empty answers (#3782), preserved thinking keeps the conversation append-only (#3783), append-only clear_at reminders land (#3787), and hosted tool-search replay order is preserved (#3789).
• Server resilience under load: Background sweeps run isolated with their own database pool (#3786, #3779), and the server returns a graceful 503 on database pool exhaustion instead of failing hard (#3788).

• Claude always-thinking models no longer return 400s or empty answers
• Ask User lets agents pause and ask you questions mid-run

See the GitHub release and the full release notes for every pull request in this version.

In code: implement the AskUser trait to answer questions from your own interface when you embed the Framework.

See Ask User and Answer an agent’s questions. PRs #3735, #3738, #3747, #3757, #3790.

• Ask User: Agents can pause mid-run to ask a person a question: an inline UI card in Platform Chat (#3735), in-process and hosted strategies (#3738), expiry that resolves to declared defaults (#3757), and credential collection that never enters the log or agent context (#3747). Enabled on the default chat harnesses (#3737).
• Host Shell: A new host_shell capability runs the bash tool as real child processes on the machine hosting the agent, bounded by a kernel policy (Seatbelt and Landlock). Behind the host-shell feature on everruns or everruns-host; see the host shell example (#3719).
• Slack one-click install: Install the Slack app in one click, creating the app and taking its credentials straight from OAuth (#3724).
• New models: Claude Opus 5.5 lands as the recommended Opus (#3769), and GPT-6 Sol and Luna arrive with GPT-6 Luna as the default chat and utility model (#3771, #3774).
• Agent endpoint budget controls: Set spend budgets on agent endpoints from the UI (#3719).

• Agent endpoints and triggers replace Apps in Everruns
• Slack agents ask approval with buttons and show task progress
• Set spend budgets on agent endpoints from the Integrations tab
• Host Shell gives Framework agents a kernel-contained Bash
• MCP catalog presets, acts-as identity and personal connections
• Install an Everruns agent in Slack with one click
• Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna models added

Before upgrading, read the breaking changes: Agent endpoints and triggers replace Apps in Everruns.

See the GitHub release and the full release notes for every pull request in this version.

• A request that ends on an assistant turn (prefill) fails fast with a configuration error that names the fix, instead of a provider 400 and a retry cycle.
• An explicit none reasoning effort maps to low on models where thinking cannot be turned off. Before, the driver sent nothing and the API ran at its own default.
• Each request in a session replays the previous one and only appends to it, so preserved thinking stays valid across turns.

See Anthropic. PRs #3782, #3783.

HostCompute::contained(options) builds the contained environment, and a ShellApprovalGate decides whether each command runs at all; with no gate the answer is no. A missing kernel primitive is a launch error, never a silent fall back to an uncontained shell.

See Host Shell and the Host Shell Agent example. #3715

• Platform Chat v2 on a unified command line: One command line serves both people and agents, with Platform Chat rebuilt on top of it (#3680).
• Soft approval: A soft-approval capability lands and is on by default for Generic and Platform Chat, so an agent can proceed with a lightweight, reversible approval step (#3655).
• Typed judgments and a guardrail engine: Typed judgments become a capability, backing a guardrail engine for classifying and gating agent behavior (#3644).
• Scoped MCP catalogs and acts-as: MCP gains scoped catalog presets and acts-as credential semantics, with endpoint auth stored separately (#3638, #3635).

• Framework direct model calls with Model::complete
• One everruns command line for the CLI, MCP and agent shells
• Guardrails can use typed judgments as their check engine
• Soft approval: agents pause before destructive or outward actions

See the GitHub release and the full release notes for every pull request in this version.

It is enabled at normal on the Generic and Platform Chat harnesses. The approver is resolved on the server from the authenticated message, never from what the model says, and recorded as agent.approval.requested and agent.approval.granted in the audit log.

See Soft approval. #3655

everruns agents list
everruns sessions --help

In code: a Framework application exposing its own command tree can name its root token with CliCommandSource::root instead of everruns.

A Platform Chat v2 harness built on this command line is available behind an organization opt-in flag. See CLI and CLI reference. #3680, #3490, #3496

let model = Model::new("gpt-6.1-sol", OpenAI::from_env()?);
let answer = model.complete("Name the three primary colors.").await?;

Each driver now declares its own credential variables, following the vendor’s SDK, for example GEMINI_API_KEY or GOOGLE_API_KEY, the AWS variables for Bedrock, and LLAMA_API_KEY for Meta. Provider model catalogs are also exposed on the facade crate.

See Direct model calls and Models and providers. #3642, #3668

• Agent exposure management: An agent’s endpoints and triggers now have one home, and a new cross-agent Exposures view lists exposure across all agents. Publish moves to the endpoint, and an agent’s exposure can be suspended (#3624, #3627, #3621).
• Endpoint-attributed sessions and budgets: Sessions and budgets now attribute to the endpoint they arrive through, and Slack and FCP endpoint budgets bind and enforce their limits. Previously configured endpoint budgets that were inert will begin enforcing (#3623, #3625).
• Channels re-parent from App to Agent: Channels move from the retired App abstraction to the agent, with grandfathered apps synthesized into agents automatically (#3618, #3598).

• The Everruns UI supports dark mode with a theme setting

See the GitHub release and the full release notes for every pull request in this version.

• Slack agent surface: Agents now run natively in Slack. The agent surface is enabled end to end: replies stream token-by-token into the pane, a stop button cancels the running turn, agent status and the thread title show while a turn runs, thread context is persisted and fed to the agent, and replies post as markdown blocks with correlation metadata (#3560, #3564, #3568, #3581, #3567, #3563).
• Session environments: A session now declares where its commands run and what they may touch, giving agents an explicit execution environment (#3599).
• Endpoint-ID ingress routing: Ingress routes mount on endpoint IDs, with reserved routing tag prefixes to keep session routing unambiguous (#3593, #3589).
• Agent-owned exposure replaces Apps: The App abstraction is retired from the product surface in favor of agent-owned exposure (#3590, #3585).

• Sessions show where their commands run and what they may touch
• Everruns agents run in Slack’s agent pane

See the GitHub release and the full release notes for every pull request in this version.

In the UI: the session Workspace tab gains an Environment panel.

In the API: GET /v1/sessions/{id}/environment and GET /v1/environment-targets are read-only views of the same data. The feature is enabled wherever sandboxes are.

In code: the Framework’s Environment gains named compute and containment members, with HostCompute behind the host-compute feature.

See Workspaces and environments. #3599

• PDF input with file attachments: Agents can now accept PDF documents as file attachments alongside text input (#3535).
• Muse image model: A new Muse image-generation model with a configurable fallback (#3523).
• Quick-connect provider grid: Connecting an LLM provider is now a one-click grid in the UI (#3526).
• Native async tool execution on durable workers: Long-running tool calls execute natively on durable workers instead of blocking a turn (#3466).
• Synchronous wait-for-completion: A non-streaming wait-for-completion path spanning drivers, the sessions API, and the framework (#3541).
• Agent discovery documents: The server serves agent discovery documents, and create_agent/session/message descriptions surface their argument constraints (#3533, #3530).

• Everruns servers publish an MCP server card and auth.md
• Connect an LLM provider with one pasted key, checked before saving
• Image generation adds the Muse image model with a fallback
• Session schedules no longer fire twice on multi-replica deployments
• Attach PDF files to agent messages
• Send a message and wait for the turn result in one request

See the GitHub release and the full release notes for every pull request in this version.

In the API: POST /v1/sessions/{id}/messages?wait=true returns 200 with the turn result when the turn completes or fails, or 202 with partial messages when timeout_ms (default 120 s, at most 600 s) expires. Without wait it still returns 201 immediately, so existing clients are unaffected.

In code: RunOptions::timeout bounds Session::run_with; when it expires the in-process turn is cancelled and its outcome returned.

See Sessions and Stream events. #3541

platform_management is retired. Its 14 hand-written tools are replaced by the platform capability, which derives the same surface from the command catalog. Agents that still list platform_management should switch to platform.

See Platform Management (removed) and Platform. #3513

• Hierarchical AGENTS.md as conversation context: Agent instructions now resolve AGENTS.md files hierarchically and are injected as the leading user-role message of each turn, keeping untrusted workspace instructions below harness safety instructions in the instruction hierarchy (#3501).
• Platform Chat for every user: Each user gets a precreated, pinned Platform Chat thread, and the Platform Chat capability set is widened so it can drive more of the control plane (#3508, #3507).
• Capability deprecation lifecycle: Capabilities can now be marked Deprecated (still functional, surfaces warn) or Retired (inert and hidden), and the platform_management capability was retired via the new lifecycle (#3513).
• Unified everruns command tree for MCP and CLI: A noun-verb command tree is now exposed over MCP, and a host can name the root token of its own command tree from the CLI (#3490, #3496).
• Safer provider onboarding and model management: Provider API keys are validated before being stored, a configured key reliably produces a usable org, discovered models are adopted instead of conflicting on explicit create, and disabled models can be re-enabled (#3516, #3515, #3517, #3512).

• AGENTS.md files resolve hierarchically as conversation context
• Every Everruns user starts in a pinned Platform Chat thread
• platform_management capability retired in favor of platform

Before upgrading, read the breaking changes: platform_management capability retired in favor of platform.

See the GitHub release and the full release notes for every pull request in this version.

This keeps untrusted repository instructions below harness safety instructions and keeps the system prompt stable for caching. Resolution is bounded to 16 files, 32 KiB each and 128 KiB in total.

See Agent instructions and Use AGENTS.md. #3501

• ~3.5x faster agent startup: Agent startup-to-ready latency cut by roughly 3.5x (#3486).
• MCP URL-mode elicitation with a consent pause: MCP elicitation now works in URL mode on both sides, with an explicit consent pause surfaced in the UI (#3354).
• Gen-AI and OpenInference trace conventions: Observability now emits Gen-AI agent and OpenInference trace conventions for standard agent tracing (#3353).
• Expanded model catalog: Added OpenAI GPT-6 Astra, Claude Fable 5.1, and Muse Spark 1.3 profiles with live-matrix coverage (#3363, #3351, #3352).
• Broad provider wire-contract hardening: A large sweep of driver and provider fixes plus expanded contract tests across Anthropic, OpenAI, Gemini, OpenRouter, Bedrock, MAI, Meta, and Fireworks, tightening streaming, usage, reasoning, and request-shaping behavior.

• First message of a new chat no longer drops or shows out of order
• Everruns CLI plugins, skills and knowledge-bases commands work again
• Organization default model can fall back to the platform default
• Only organization admins can rename an organization
• Tool calls are validated against their schema before running
• OpenTelemetry traces follow Gen-AI and OpenInference conventions
• MCP URL elicitation pauses for consent in the Everruns UI
• GPT-6 Astra, Claude Fable 5.1 and Muse Spark 1.3 model profiles
• Framework agents start about 3.5 times faster

See the GitHub release and the full release notes for every pull request in this version.

Two costs left the startup path: provider HTTP clients are now resolved per request instead of in each driver constructor, and a session’s event broadcast channel is allocated when the first subscriber arrives rather than with the session. A subscriber still never receives events sent before it subscribed. No code changes are needed.

See Agents. #3486

• OpenAI GPT-6 Astra (gpt-6-astra), seeded as an enabled favorite. Its reasoning effort accepts a new max tier above xhigh.
• Claude Fable 5.1 (claude-fable-5-1 and its 1M-context variant), seeded as enabled favorites with adaptive thinking.
• Muse Spark 1.3 and its Contributor tier on the Meta Model API. Muse Spark 1.2 stays selectable.

New organizations get them on the Models page automatically.

See Providers, Meta and Anthropic. #3363, #3351, #3352

Choose the vocabularies with EVERRUNS_TRACE_CONVENTIONS (default gen_ai,openinference):

# OpenInference only, for example for Arize Phoenix
export EVERRUNS_TRACE_CONVENTIONS=openinference

See OpenTelemetry and Framework observability. #3353

For OpenAI Chat Completions and Open Responses, compatible tools are now sent with strict schemas. A schema that cannot be converted safely is sent non-strict, as before.

See Tool Call Repair and #3336.

Replay to providers is also more accurate: Anthropic thinking blocks keep their own signatures, including redacted_thinking; OpenAI reasoning items keep the ids the provider issued; and the Gemini driver now requests, parses and replays thoughts.

See Event reference, #3276 and #3279.

Highlights

• Reasoning modeled as ordered artifacts: Model reasoning is now captured as ordered artifacts with a published classification, so reasoning steps replay in the order they were produced and their kind is visible to consumers (#3276).
• Bounded batch file reads: The filesystem capability gained bounded batch reads, letting agents fetch several files in a single guarded call instead of one round-trip per file (#3275).

• Agents read several workspace files in one read_many_files call
• Global chat session endpoints removed from the Everruns API
• Messages API reports reasoning in order and the message phase

Before upgrading, read the breaking changes: Global chat session endpoints removed from the Everruns API.

See the GitHub release and the full release notes for every pull request in this version.

Models that emit one tool call per response no longer need one model round per file. In the PR’s tests, three independent reads took 2 model calls instead of 4 and about half the input tokens. Loop detection recognizes repeated batch reads.

See File System and #3275.

Highlights

• Chat with harnesses, archive chats, and generated run summaries: Sessions gained interactive chats against harnesses, the ability to archive chats and toggle showing all of them, and automatically generated run summaries, with count badges on session detail tabs (#3262, #3265, #3260, #3266).
• Connection-level custom provider headers: Providers can now carry custom headers at the connection level, with cache diagnostics to see how they resolve; header values are redacted in diagnostics (#3232, #3240).
• Leaner supply chain and smaller binaries: Owning 114 small third-party contracts in-tree removed those dependencies from the build, and stripping plus CLI dispatch boxing shrank shipped binaries by ~22% (#3245, #3227).

• Everruns CLI, server and worker binaries are about 20% smaller
• Start a chat with any harness, not only an agent
• Provider connections carry custom headers and cache diagnostics
• Archive finished chats and show them again on demand
• Session pages show a generated run summary and tab counts

See the GitHub release and the full release notes for every pull request in this version.

runtime.register_capability(Arc::new(capability))?;
runtime.is_capability_registered("ext:name");

Registration only makes the id resolvable. Sessions still enable it with activate_capability, so sessions that never activate it see no change. A duplicate id or a colliding alias is rejected and the existing capability stays. Readers always see a consistent registry snapshot.

See Advanced capabilities and #3224.

Highlights

• Register capabilities on a running runtime: A host holding Arc<InProcessRuntime> can now make a capability discovered after composition resolvable without rebuilding the runtime, via register_capability/is_capability_registered. Registration stays separate from activation, and readers always see a consistent registry snapshot (#3224).
• A2A delegation is now opt-in: Outbound A2A delegation moved behind a new a2a Cargo feature, off by default, so the standard build no longer drags a second HTTP/TLS stack into consumers that only run the local runtime. Hosts that delegate to remote A2A agents enable the a2a feature explicitly (#3221).

• A2A delegation in the everruns crate is now an opt-in feature
• Register a capability on a running Framework runtime

Before upgrading, read the breaking changes: A2A delegation in the everruns crate is now an opt-in feature.

See the GitHub release and the full release notes for every pull request in this version.

Highlights

• First-class sandbox and checkpoint records: Sandboxes and checkpoints are now first-class records, giving durable runs an explicit, addressable execution substrate (#3197).

See the GitHub release and the full release notes for every pull request in this version.

everruns = { version = "*", features = ["local", "a2a"] }

Builds without it skip the A2A client and its HTTP and TLS dependencies, and the A2A delegation capability is not registered.

See A2A, A2A agent delegation and #3221.

Highlights

• The 0.18 neutral-kernel boundary: everruns-core is now a neutral execution kernel. Hosted platform records, sessions, agents and agent versions, harnesses, evals and observers, connectors, system email, feature flags, and OAuth/credential infrastructure, moved out to everruns-platform (and sibling crates), and the logic-free everruns-runtime compatibility crate was removed. REST/gRPC shapes and stored schema are unchanged; see the 0.18 migration guide and the Breaking notes below (#3160, #3101).
• Engine-owned sessions and unified execution: Engines now own the session lifecycle, and a single execution path serves both in-process and durable runs (#3179, #3173, #3167).
• Multi-head workspaces: Framework workspaces now support multiple heads (#3166).
• Metered embedding spend: Retrieval and knowledge-index embedding spend is now metered and ledgered against the organization (#3149, #3163).
• Chats as core functionality: Chats is always present in navigation and search for every organization with no feature opt-in, and now supports chat pinning.

• Chats is on for every organization, with pinned chats
• Session recordings split into Transcript, Timeline and Events
• Knowledge index embedding spend is metered and ledgered
• Framework workspaces gain multiple isolated Git heads
• everruns-core 0.18 becomes a neutral execution kernel

Before upgrading, read the breaking changes: everruns-core 0.18 becomes a neutral execution kernel.

See the GitHub release and the full release notes for every pull request in this version.

REST and gRPC shapes and the stored schema are unchanged; code that imports from everruns_core needs new paths. Self-hosted servers need no change.

See Upgrade notes: 0.18, #3160, #3101 and #3179.

Several crates named here were later consolidated: everruns-platform became everruns-capabilities in 0.39, and everruns-host became an everruns-core feature in 0.40. See Framework crates consolidated into everruns-core and the facade.

With the local feature, a local Git-worktree provider creates each head as a worktree from a revision. Heads are isolated by default; opt into a shared head explicitly. Fork, checkpoint, archive and destroy are explicit calls, and dropping a value never deletes a worktree. The provider interface is open for other backends.

See Workspaces and environments and #3166.

Highlights

• Live, steerable sessions: Sessions are now live and steerable, with canonical session events, session history and resume, and a session work and wake API (#3095, #3090, #3089, #3083).
• Sessions as the operational list: The Sessions surface was rebuilt as the operational list, recording session source and offering a filterable list with facet counts (#3078, #3073).
• Advanced capability authoring: New advanced capability authoring API with typed lifecycle hooks, a workspace access policy, unified capability configuration, and separated provider and model configuration (#3084, #3086, #3085, #3096, #3093).
• Daytona workspace recovery: Sandboxes recover Daytona workspaces after instance loss (#3092).

See the GitHub release and the full release notes for every pull request in this version.