0.52.2 - 2026-08-14
- [chat commands]
chat starts OAuth-bound conversations with shared Harnesses published through tenant login discovery aliases, with sessions isolated by verified user identity.
- [eval commands] eval backend resolves Coze or TEA through the AgentKit evaluation gateway by default and provides
--cache-refresh and --verbose for inspecting backend coordinates and troubleshooting TEA requests.
- [eval commands] dataset, evaluator, and target commands expand TEA evaluation support with dataset updates and versions, evaluator template details, draft updates and version submission, plus source evaluation target and target-version listing.
- [eval commands]
eval run supports explicit TEA dataset versions, evaluator versions, source target types, and target versions, and returns the TEA run ID with the experiment ID.
- [Harness] Harness configuration supports private networking, capacity, shared database sessions, and managed model egress for shared OAuth Harnesses, and verifies identity, image, and scaling bindings after deployment.
- [invoke commands]
invoke run no longer reuses the login session id_token automatically for custom_jwt Runtimes; generic invocation now requires an explicit OAuth token through --headers, while shared OAuth Harness chat uses chat.
- [Sandbox commands]
sandbox create provides --llm-shield-app-id for enabling LLM Shield on Skill sandbox tools.
0.52.1 - 2026-08-12
- [Harness Sidecar] Harness Sidecar commands provide
harness sidecar catalog and harness sidecar resolve for printing the Product Component Catalog and resolving profiles plus component switches into a verifiable Sidecar plan before release.
- [Harness Sidecar] Release configuration supports managed
harness_sidecar config, provides runtime descriptions, create-time tags, and Code Pipeline workspace / pipeline pinning fields, and derives Auto TOS buckets from the cloud provider, region, and account.
- [Harness Sidecar]
release provides --json for managed Harness Sidecar releases, emitting newline-delimited build, deploy, runtime-id, and final-result events and verifying the Sidecar, model proxy, and required MCP gateway after release.
- [Harness Sidecar] Release flow now matches existing Runtimes by exact name. For managed Harness Sidecar releases, states that require recreation now require explicit reconciliation instead of automatically deleting an existing runtime.
- [Sandbox commands]
sandbox config --list now prints the current .agentkit/sandbox.yaml file instead of listing inferred default model, network, and tool settings.
0.52.0 - 2026-08-11
- [Auth commands] The CLI can now read
/.well-known/agentkit-cli discovery documents from custom login addresses that include a single tenant path segment, with stricter validation for remote addresses, discovery fields, regions, cloud providers, and STS coordinates. When logging in through a remote discovery document, the CLI saves the profile only after login succeeds.
- [Auth commands]
auth admin publish now accepts existing --issuer, --client-id, --role-trn, and --provider-trn values, so cross-account identity setups can publish only the discovery document; it also supports --domain for returning a custom login domain whose CNAME and HTTPS certificate are already configured.
- [Auth commands] Auth Admin can use valid STS credentials from the current CLI SSO profile for management operations, and uses built-in service endpoints for those operations so project or user-wide endpoint overrides do not affect privileged setup.
- [Installation] The install flow keeps the standalone CLI ahead of same-named commands installed through pip, pipx, venv, or Conda in zsh, bash, or fish, and uninstall restores original commands and shell configuration entries that are still managed by the installer.
- [Configuration] Config commands validate region values, accepting only lowercase, hyphen-separated region codes so invalid regions cannot enter service endpoints or request signing.
0.51.2 - 2026-08-11
- [Model Gateway commands] Model Gateway supports gateway activation, model provider and consumer management, and generation of
curl, OpenAI, Anthropic, or AgentKit calling examples from authorized models.
- [invoke commands]
invoke run now follows the Python SDK invocation contract, resolving targets from agentkit.yaml, Runtime IDs, or endpoints, accepting JSON payloads and headers, and selecting /invoke, ADK /run_sse, or A2A automatically; it can also print raw streaming events or show reasoning content when needed.
- [harness commands]
harness invoke directly invokes a deployed Harness with per-request overrides for the model, system prompt, tools, skills, runtime backend, A2A registry, and maximum LLM calls, with either invoke or run_sse transport.
- [Auth commands] Identity-only mode lets
login --identity-only and auth profile set --identity-only store only the OIDC session. For matching custom_jwt Harness invocations, the CLI can forward that user identity without creating management credentials.
- [Auth commands] Auth Admin generated SSO policies now include permissions for listing sandbox sessions, restoring session snapshots, and querying tool types, so sandbox-related SSO workflows have the required access.
- [Sandbox commands] Creation, initialization, and preview flows are now more predictable:
sandbox create --envs can inject repeatable environment variables, sandbox init writes the current built-in base image into generated Dockerfiles, and sandbox web opens the sandbox home directory in the remote browser. Initial .agentkit/sandbox.yaml files no longer persist inferred model defaults, and the Model Square default model is now deepseek-v4-flash-ga-260731.
- [migrate commands] Dify migration extracts Workflow and Chatflow nodes, edges, variables, features, and dependencies before migration while redacting sensitive fields.
0.51.1 - 2026-08-06
- Fixed missing bundled assets in standalone distributions so lifecycle commands and
sandbox dashboard can load templates and web terminal assets correctly after installation.
0.51.0 - 2026-08-05
0.50.9 - 2026-08-03
- Added global
--provider, the top-level config command, and region selection for init, Harness, and migration.
- Completed BytePlus project switching, MCP service mapping, build images, and migration.
0.50.8 - 2026-07-31
- Added
deploy config --provider for selecting the cloud provider when generating agentkit.yaml.
- Switched BytePlus deployments to the corresponding base images.
0.50.7 - 2026-07-31
0.50.6 - 2026-07-31
0.50.5 - 2026-07-31
0.50.4 - 2026-07-30
- Relaxed wording checks for migration plans to improve compatibility across source projects.
0.50.3 - 2026-07-30
0.50.2 - 2026-07-29
- Added BytePlus resource routing to deployment configuration, establishing the base for the complete multi-cloud deployment flow.
0.50.1 - 2026-07-29
- Changed Runtime defaults to 2 vCPU, 4 GiB of memory, and one minimum instance; see
agentkit.yaml.
- Added
migrate --project for selecting the project used by remote migration.
0.50.0 - 2026-07-28
- Added Auth Admin credential management.
- Added Skill Space references and request-level overrides to Harness.
- Released the revised sandbox commands for configuration, creation, terminal, file, and invocation workflows.
0.46.7 - 2026-07-22
- Added remote Dify and arbitrary-source support to migration.
- Added per-resource region and project overrides plus Runtime private-network settings to
agentkit.yaml.
0.46.6 - 2026-07-19
- Published the AgentKit CLI command reference and the
agentkit.yaml-based build and deployment documentation.
Last modified on September 19, 2026