# CLI 0.0.4-microsoft-preview.14

Released September 9, 2026. Microsoft migration engine preview; stable `latest` remains 0.0.3.

## Microsoft stage context-budget fix

- Flow grouping returns a compact inventory and the complete Microsoft connection graph instead of embedding every artifact's full parsed IR.
- Full artifact details and source text remain available through detail and paginated read tools.
- All model-driven Microsoft stages use bounded prompt history. Oversized results and earlier calls remain available through exact archived-result reads, with optional JSON-field selection and pagination.
- Full tool inputs and results are retained in the stage audit. Partial previews are explicitly labeled; they do not establish that evidence is absent.
- Context limits also account for UTF-8 bytes. Required Microsoft instructions remain complete, and progress messages explain when results are archived.
- Microsoft parsers, graph-building logic, orchestration coverage, and host finalization checks are unchanged.

## Install or update (PowerShell)

```powershell
Invoke-WebRequest 'https://tools.lambogenius.com/lambogenius-integration-migration/0.0.4-microsoft-preview.14/install-cli.ps1' -OutFile install-cli-preview14.ps1
# Review the downloaded script before running it.
& .\install-cli-preview14.ps1 -SkipDependencies
li.cmd --version
```

Expected version: `0.0.4-microsoft-preview.14`. Omit `-SkipDependencies` if prerequisites also need installation. Use `li.cmd` in PowerShell to avoid its built-in `li` alias.

This changes the engine fingerprint. Preserve previous generated work and use a fresh generated workspace for the new run; do not edit checkpoint fingerprints to force reuse. Installation does not reset license counters or extend the license term.

## Verification and limits

70 Microsoft-core tests and 211 CLI tests passed, including a regression matching the reported 388,852-character result and 55,296-byte prompt limit.

A read-only FINACCONT-BT regression preserved all 98 selected artifacts, 5 orchestrations and 42 Microsoft graph connections. Its grouping response dropped from 361,471 to 27,738 characters; its largest scripted prompt was 47,974 bytes. Full detailed IR remained retrievable, and omitting an orchestration was still rejected.

The source regression used scripted requests with zero real model calls. It verifies context transport and coverage enforcement, not end-to-end customer conversion or semantic equivalence. No customer source/output, license activation, metering, or cloud resources were changed by the regression. Required instructions must still fit within the configured model context, and the tool-round limit still applies.