Skip to main content
Evaluate Sandbox, Harness version/image digest, model endpoint/protocol, SDK, scenario, and date together. Catalog capabilities do not establish acceptance for every intersection. This page maps repository evidence; it is not live service status or account health.

Choose the behavior you need

Fresh documentation journeys: 2026-11-20

The tutorial programs now have independent Python and TypeScript evidence for Vercel + built-in Pi 1.5.2 + the configured Messages endpoint using deepseek-ai/DeepSeek-V4-Flash. These records include source-file hashes; compare them before assuming a later edited program has the same verification. Endpoint/account details and resource identities remain in private local evidence, so the model name alone is not a reproducible credential or universal endpoint guarantee. Together these are four tutorial journeys and six Runs for this one configured combination, not four or six distinct compatibility combinations. They are separate from the historical MiniMax and idle-recovery matrices below. Their successful deletion records do not settle earlier unrelated unknown resources.

Evidence snapshots

Counts below were derived from JSON rows.status during this documentation update. They describe each file’s recorded timestamp. The first counts combinations; the second counts SDK rows. Do not add them into a total of passed combinations. Regenerate the table if the source files change.

Read an individual record

  1. Identity: inspect builtin version or full OCI digest/protocol/command. Built-in and OCI OpenCode are distinct distributions.
  2. Model: check actual endpoint, model ID, and Messages/Chat Completions/Responses. A shared vendor does not imply equivalent protocols.
  3. Client: inspect independent Python/TypeScript results, installed versions, and scenarios. A Python Harness does not prove Python SDK acceptance.
  4. Branches: check task, upload, Tools, approval, replay, artifacts, continuation, and recovery separately. Limited contracts such as Aider must not receive credit for untested Tools/artifacts.
  5. Cleanup: separate successful Session teardown, earlier failed-attempt resources, and account-wide resource audits.
The task matrix retains attempts. A final passed row can still include earlier failures or unknown creations. Do not erase history or infer no global orphan resources from passing rows. Keep exact resource IDs in private evidence for reconciliation; public documentation links sanitized summaries.

Status and evidence types

Contract/simulation tests prove deterministic logic; live tasks record remote behavior for that run. Registry implemented/live_resource tags do not substitute for exact end-to-end records. Documentation structure checks do not verify examples or cloud behavior.

A concrete temporal change

The old MiniMax matrix’s unsupported E2B OCI rows describe that snapshot. Later E2B Session-owned Template preparation gained scoped OpenCode evidence: Python/Chat and TypeScript/Messages, file byte hashes, and native Session ID checks. That does not cover Responses or every SDK/protocol intersection. The increment retains exact-output failures alongside later passes with clarified prompts; earlier failures are not retroactively passed. Its resource audit covers that increment, excluding separately retained historical operations. Apply the same discipline to other provider versions and recovery changes.

Establish deployment-specific evidence

Use the compatibility runner guide and acceptance journey, pinning combinations and independently running both SDKs. These commands allocate resources and consume model usage; do not run them implicitly as ordinary documentation checks. Record date, control-plane revision, SDK versions, provider/template, Harness version/digest, model endpoint/protocol, actual command, result, failure branches, and cleanup. Passed means observed; unexecuted work is planned; uncertainty stays unknown. Never log credentials. Retain resource IDs in private local evidence and reconcile uncertain creation/writes. This page indexes the fresh tutorial evidence and historical matrices above; it is not a separate cloud execution. Next: choose a combination, Sandbox capabilities, model protocols.

Recompute the snapshot counts

Run from the repository root. This reads only sanitized local evidence and makes no provider/model requests. It reproduces the status counts above; it does not rerun acceptance or merge the two matrices into a common denominator.
For new or changed combinations, follow tests and evidence and the relevant Sandbox, Harness or model guide.