[anycodes]

Description

Starting a template no longer selects one workflow or operation. Specification 5 starts the complete sandbox, preserves every declared service, and exposes launch_template / TemplateSession in Python. Source building remains owned by the CLI. The matching client/server release candidate is 0.4.0a1; it has not been published.

Refs #107

Type of Change

• Refactoring
• Tests
• Documentation

Changes

• Remove launch_workflow and workflow_id; reject v4 runtime manifests with migration guidance. Preserve historical receipts and their old contracts without treating them as v5 verification.
• Start once, execute readiness, expose all endpoints and check SandboxServer protocol v2 health/discovery. Other protocols remain extensible and use author readiness; they are not falsely marked protocol-verified.
• Retain owned-resource journals and cancellation-safe cleanup across every resolved client. Propagate template environment/resources, and use one effective application token.
• Add the implemented Express process-start mode to its metadata, verified against two real HTTP services.
• Update bilingual docs, packaged skills, candidate version/dependency pin, changelog and engineering records.

Testing

• Shared launch vectors: 17 pass; final launcher regressions: 21 pass.
• Both packages lint/type checks pass. Four candidate wheel/sdist artifacts, twine, clean paired installation and typed markers pass.
• Real Express candidate startup: 5 pass, covering base/interpreter-image shell usage, provider name/object, and two HTTP services started together with readiness and owned deletion.
• Full CLI acceptance is not uniformly green: a configured run passed in 376.63 s; an additional run with explicitly mapped Express environment variables failed in 355.17 s at files.rw (upload exit 2, download exit 4, no downloaded body). Other operations and both deletion calls passed. The fixture lost file stderr and per-step timings, so the root cause is unconfirmed. Explicit create timeout is 300 s, but the longer passing run prevents a timeout conclusion. Diagnostics are tracked separately in test: preserve timed diagnostics for CLI file-transfer acceptance #109; a separate focused owned-sandbox file check passed in 180.946 s: upload/download exit 0, exact byte match, deletion and absence confirmed. That success does not explain the earlier failure.
• Full cloud suite: 124 pass, 9 explicit skips, 2 slow deselections. Its missing-ABX_BIN CLI case failed initially, then passed separately with the built 0.5.0 CLI, including create/exec/files/repeated deletion.
• Local full server: 878 pass, 2 environment-intercepted malformed-HTTP failures; local full client: 2260 pass, 20 skips, 2 environment HTTP502 engine failures. No product or test bypass was introduced.
• First implementation head ff07316 passed all 11 remote checks, including client/server Python 3.10/3.11/3.12 and real Node engine integration.
• Final implementation head 66c86f2 passed all 11 remote checks.

Checklist / Definition of Done

• Regression tests and real-component evidence recorded.
• Paired versions and required client/server dependency are 0.4.0a1.
• English and Chinese docs, CHANGELOG and .agents/ records updated.
• Repository health: 106 pass, 1 warning, 0 failures; no secrets or workstation paths added.
• Final-head CI and cross-repository release coordination complete.

Evidence: .agents/evidence/verify/2026-10-25-unified-template-cloud.md and .agents/evidence/verify/2026-10-25-unified-template-cli-diagnostic.md. Final publication-documentation head 97888a7 passed all 11 required checks. The coordinated release retains the CLI diagnostic limitation linked above. No new full-catalog v5 matrix, interpreter API validation, PyPI publication or merge is claimed.