Skip to main content
This page records dated, user-visible changes in the current source workspace. It is not a list of published releases. release-policy.json currently identifies developer-preview preparation and does not approve publication; repository/package ownership and release reviews are still pending. Python and TypeScript package manifests currently say 0.1.0; that number alone does not distinguish every development snapshot or establish registry publication. Record the exact source/distribution fingerprint, lockfile, SDK artifacts and database migration inventory for your installation. A locally built distribution has release.json metadata. Follow upgrade and migration before changing an installation. Dates below identify evidence checkpoints, not a promise that a new version is available from a public download channel.

Identify the snapshot you are adopting

Match a change to the source or distribution you actually install. The two SDK package versions can remain 0.1.0 across different source changes; upgrading by that number alone cannot establish which changes you have. A local distribution’s release.json contains format, sha256 and built_at, produced by the local builder. Its fingerprint identifies the built contents; it is not a signed publication, Git revision or compatibility guarantee. Keep the source revision and any local modifications with that record, and retain the matching lockfile and SDK artifacts. For example, the direct-Session change below is backed by migration 016, which drops the NOT NULL constraints on sessions.agent_id and sessions.agent_revision. A newer SDK installed against an older control plane does not apply that database migration. Use the upgrade procedure to align the deployment, then verify the behavior through the API.

Documentation revisions — 2026-11-20

The first-Session tutorial and project tutorial now link independent Python and TypeScript execution records. Both used Vercel Sandbox, built-in Pi 1.5.2 and the configured Messages-compatible DeepSeek-V4-Flash endpoint. These are documentation-example verification checkpoints, not a new product release or expanded support for other combinations.
  • First-Session record: each SDK completed one Run, verified the expected artifact, closed its Session, confirmed its deletion operation and downloaded identical retained bytes.
  • Project record: each SDK completed two Runs in the same Session, kept the fixed tests unchanged, exercised filtering and Reset in the browser, retained exports and confirmed the old preview was denied after closure.
Follow the tutorials’ current prerequisites and recovery instructions. Compare the example files with the program_sha256 fields before applying these records to modified examples. These records do not establish a clean installation, external Tool authorization/approval, every provider or deployment-wide orphan cleanup. The coverage ledger continues to track unfinished documentation work separately.

2026-11-19 — E2B public OCI preparation

Impact. E2B can prepare public OCI Harness images as Session-owned templates, separately tracking template/build operations before compute creation. This adds a build stage and cleanup responsibility. It does not enable private registry images or a cross-Session template cache. Built-in Harness Sessions continue using their configured template. Action. Check the E2B prerequisites and OCI registration. Choose a compatible immutable Harness definition and a new Session. Allow preparation to finish; do not submit another build merely because a response is lost. Retain template/build operation references until both compute and owned template cleanup are confirmed. Session SDK shape is unchanged by this increment. Verify. Inspect initialization, preparation state, actual execution and cleanup. The dated evidence covers a specific OpenCode digest, installed Python with Chat Completions and TypeScript with Anthropic Messages, actual file hashes/native state across pause, MCP, artifacts and cleanup. It does not certify every SDK/protocol cross-product, Responses or every E2B Harness. Earlier failed attempts remain failures. The increment’s cleanup audit excludes separately parked historical operations.

2026-11-19 — Harness pause/wake and Playground presentation

Impact. Serverless verification distinguishes actual file/conversation restoration per Harness and provider. Console presents SDK and REST API as sibling representations and corrects dark-theme application icon treatment. Switching a code representation preserves the draft and does not submit a task. Action. Read serverless lifecycle and idle/wake. Review the exact combination before expecting memory or filesystem restoration. Use a new Session when changing a fixed Harness definition. Do not interpret provider standby as confirmed deletion, or a filesystem restore as preservation of every running process. Verify. Follow-up execution must retain the intended file bytes and conversation state and finish with confirmed cleanup. The dated serverless record documents actual failures, fixes and UI review; verified combinations explains matrix scope. The earlier matrix’s OCI exclusions describe its snapshot; the separate E2B increment above adds scoped evidence afterward. Do not rewrite older rows as if they tested the newer implementation.

2026-11-18 — Inspect historical Session initialization

Impact. Console and both SDKs can read a Session’s effective initialization snapshot: pinned definition, source Agent revision where applicable, environment names and explicit redaction metadata. It describes historical initialization, not current Runtime health or a credential-bearing replay specification. Action. Update the installed SDK if it lacks Python session_initialization or TypeScript sessionInitialization. Use the snapshot when diagnosing which configuration a Session actually received. Do not reconstruct credentials from it or expect renaming a connection to change historical initialization. Verify. Read the same Session through both clients and compare snapshots; confirm foreign-tenant access is denied and secret values remain absent. The evidence includes installed-package tests, read-only inspection of an existing cloud Session and bilingual browser review. That inspection was not a new cloud execution. See initialization.

2026-11-18 — Remote MCP, Feishu/Lark and Responses integration

Impact. Remote MCP adds discovered, selected Tool schemas through the shared broker. Feishu/Lark adds application credential exchange and fixed operations. OpenAI Responses is a distinct model protocol; it is not automatically implied by a working Chat Completions connection. Action. Follow MCP, application accounts and model protocols. Revalidate changed MCP definitions before new use. Select Responses explicitly with a compatible Harness. Existing immutable OCI registrations keep their definitions; expanded compatibility metadata does not mutate them. Register a new version and create a new Session where needed. Verify. Check selected operations, authorization/approval and the actual model protocol. The original increment separates real DeepWiki MCP cloud journeys from Feishu/Lark deterministic coverage and real CLI binaries against fixture Responses servers. Later matrix evidence may expand specific combinations; the initial record alone does not prove real Lark accounts or a live Responses model passed. Arbitrary MCP OAuth onboarding is not provided by that change.

2026-11-17 — Direct Session entry; saved Agent optional

Impact. Applications can create Sessions using Harness, Sandbox/model connection IDs, optional Tools and environment directly. Saving an AgentDefinition is optional. The Console entry emphasizes Sessions, runtime connections and Tool connections. Existing saved-Agent requests remain supported. Migration action. After taking a matching offline backup and settling existing resources, apply the intended source update using the upgrade procedure. Migration 016 makes Agent references nullable while preserving existing data. Update application readers to handle null agent_id / Agent revision for direct Sessions. Remove any application-only requirement that users must save an Agent before starting; keep the reusable-Agent path where it serves a real need. Do not rewrite old Session definitions. Verification example. Create a direct Session through each SDK without creating an Agent; submit a real task, verify the final result, query its nullable Agent reference, then close and confirm provider deletion. Test optional Tool binding separately. The dated checkpoint records independent Vercel/Pi journeys with an existing SiliconFlow connection, one with Exa, plus package/API/browser checks. It reused credentials and did not rerun every provider, authorization or full Web scenario. This is a source-change migration example, not a fabricated named release.

Documentation revisions — 2026-11-19

The bilingual product documentation adds task-oriented start/build/integration/concept/operation/reference/contribution pages, an Agent reading guide, and a coverage ledger. Drafts and incomplete assets remain identified. Documentation edits alone do not add runtime capabilities, create verified combinations or publish the site. The structural checker validates inventory and selected Markdown structure/link targets. It does not certify prose, executable examples or real-cloud behavior. Keep documentation verification separate from the product-change evidence linked above.

Before adopting any change

Use the upgrade procedure and backup/restore guide: record current versions, settle Runs and resource operations, retain matching old code/backup, apply migrations, and verify the affected SDK/user journeys. Do not edit applied migration checksums or run old code against a newer database to force rollback. Restoring a matching older backup loses later database writes and does not reverse external side effects. Next: upgrades, verified combinations, known failure recovery.