Skip to content

feat: consume v1 resources in sdk and managed agents runtime - #164

Draft
anycodes wants to merge 7 commits into
feat/agents-api-contract-20261112from
feat/template-v1-runtime-20261112
Draft

anycodes wants to merge 7 commits into
feat/agents-api-contract-20261112from
feat/template-v1-runtime-20261112

Conversation

@anycodes

@anycodes anycodes commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor

Description

Use the unreleased v1 resource contract for prepared template launches and Managed Agents API configuration. Forward supported Sandbox instance settings and keep worker-only custom environment values ephemeral; preserve ownership, borrowed-volume behavior and private bridge authentication.

Refs Alibaba-Agent-Sandbox/awesome-templates#309
Stacked on the narrow API input fix: #163
Contract: Alibaba-Agent-Sandbox/awesome-templates#310
Consumers: Alibaba-Agent-Sandbox/agent-sandbox#249 and Alibaba-Agent-Sandbox/awesome-templates#311

Type of Change

  • New feature / refactoring
  • Bug fix
  • Tests and bilingual documentation

Changes

  • Pin the actual canonical schema, normalizer and Core YAML reader. Literal/already-resolved prepared Python launches do not require a CLI subprocess; unresolved expressions still require the appropriate resolver.
  • Preserve timeout precedence, environment/metadata/volume inputs and safe provider validation; expose declared extra ports through real audience-aware application endpoints without inventing HTTP services.
  • Derive Managed settings from one v1 source and inject custom values only into the owning Worker process. Reserve runtime/model/auth fields and keep secret values out of state and plans.
  • Return truthful public Session integer timestamps, required_actions and vault_ids. This does not invent unsupported official Agent fields or an exec-server remote_url.
  • Update client/server dependencies, full owner records, schemas and English/Chinese guides.

Testing

  • Full v1 make test at the earlier 107984 consumer checkpoint: Server880 and Client2596 passed,20 existing skips.
  • Full v1 make ci at that checkpoint: Server880 and Client2586 passed,20 existing skips/10 deselected; lint/type/build/twine and fresh-wheel checks passed. Tests ran inside a real PID namespace to isolate unrelated same-UID processes without weakening reuse safety assertions.
  • Final Session projection delta:95 Managed tests passed. Final Core reader delta:290 relevant tests passed; static checks, both wheel builds and independent installed-wheel/Core-vector/pip checks passed.
  • Official OpenAI SDK3.49.0 against a real loopback Managed HTTP service and SQLite: default create/retrieve/list decoding passed; strict response validation still has documented Agent/harness/environment differences. No blanket strict-SDK compatibility claim.
  • Final c4d43a2 remote CI:all13 checks passed. The prior b988 Markdown links failure was traced to active guides still describing the removed root server field;14 bilingual pages now use actual independent policy loading and tested examples. Runtime/dependency/script bytes are unchanged by that documentation follow-up.
  • Final Server image built from immutable b988 source and ran actual non-root HTTP/SQLite projection, installed Core reader vectors and pip check successfully. A separate real Qwen two-turn task in the unchanged Guest passed downloaded-code/offline-original-test checks. This local Docker/model result does not establish FC/cloudSandbox E2E.

Limitations / Definition of Done

  • Regression coverage, bilingual documentation and proposed records included
  • Package/source verification recorded with exact revisions
  • Latest-head remote CI (13 successes)
  • Required final review and cross-repository acceptance
  • Final Server image build/run verification (older c550 evidence remains a preserved baseline)
  • Real FC deploy/update, old-Session routing, both Agents' real tool/file tasks and independent cleanup

Keep Draft. Real managed E2E is BLOCKED pending an existing PostgreSQL configuration and a Codex-compatible Responses tool-round model. No new cloud resources were created by the local checks. The local broad server and Managed profile remain distinct; this PR does not imply all OpenAI capabilities, published packages or generic Serving lifecycle are delivered.

Owner operation records

Necessary continuation/verification operations now live in this repository's indexed .agents/:
owner runbook and record-placement evidence.
The follow-up changes only .agents/; it preserves exact source scopes, pending FC/Codex/topic
gates and historical failures. Small original probe sources/results are retained with their
owner where needed. Credentials, raw sensitive logs, caches and distribution binaries are excluded.
Record links, source hashes, JSON/Python syntax, existing owner health/hygiene and documentation
checks passed. These are local record-placement checks, not a repeated model/cloud execution;
the new head still has its own CI. Delivery309 remains open.

@anycodes anycodes changed the title feat: consume v1 template and managed instance configuration feat: consume v1 resources in sdk and managed agents runtime Nov 12, 2026

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant