# CLI 0.0.4-microsoft-preview.15

Microsoft migration engine preview. Stable `latest` remains 0.0.3.

## Grouping-loop fixes

- Keep the newest complete tool result in the prompt when it fits, instead of archiving it at half the available history budget. The grouping overview retains every selected artifact ID and the complete Microsoft connection graph; redundant paths remain available through detail/list tools.
- Keep host-confirmed grouping progress visible, so the model is instructed to submit groups rather than restart detection.
- When grouping validation fails, identify every unaccounted selected artifact ID and explain how to correct the submission. Orchestrations still must belong to exactly one flow, and caller/callee chains must stay together.
- Keep an outstanding validation correction visible across intervening evidence reads, even if its original call moves to the exact-result archive.
- A host-only `lambogenius_correctFlowGroups` tool lets the model explicitly amend individual artifact memberships in its latest rejected submission. Revision checks prevent stale changes; missing and conflicting memberships are listed together. Corrections still pass through the normal grouping validator and storage path.
- Stop three consecutive repeated request/result pairs with an explicit `MODEL_TOOL_LOOP` requirement. Alternating already-seen tools is also detected. Different pages, corrected inputs and changed results remain valid progress.
- Retain all executed calls in the audit and report the affected tool and next action. No success is inferred from a model response alone.

These are Lambogenius host-control improvements around the pinned Microsoft tools and instructions. They do not replace Microsoft parsing/grouping rules, truncate required evidence, or relax finalization checks.

## Install or update (PowerShell)

```powershell
Invoke-WebRequest 'https://tools.lambogenius.com/lambogenius-integration-migration/0.0.4-microsoft-preview.15/install-cli.ps1' -OutFile install-cli-preview15.ps1
# Review the downloaded script before running it.
& .\install-cli-preview15.ps1 -SkipDependencies
li.cmd --version
```

Expected version: `0.0.4-microsoft-preview.15`. Omit `-SkipDependencies` if prerequisites need installation. Use `li.cmd` to avoid PowerShell's built-in `li` alias.

The engine fingerprint changes. Preserve previous outputs and use a fresh generated workspace; do not edit checkpoint fingerprints to force resume. Installation does not reset entitlement counters or extend the license term.

## Regression verification

- 78 Microsoft-core tests, 212 CLI tests, and 24 installer/release tests passed; CLI compilation and lint passed.
- FINACCONT-BT scripted regressions preserved all 98 selected artifacts and 5 orchestrations. The complete grouping response was directly available within the 55,296-byte ceiling, and a deliberately repeated detection call stopped after four calls instead of 100. Customer source bytes were unchanged.
- Bundled startup/license display, standalone precheck, stage failure diagnostics, and machine-readable output smoke checks passed. The packaged bundle matched the smoke-tested bundle byte-for-byte.

### Local-model recovery qualification

An initial local `qwen3-coder:30b` run produced a real FINACCONT-BT grouping submission with omitted and conflicting memberships. That exact rejected submission was replayed into the final host, after its real detection tool ran. The local model then corrected it using the targeted correction tool in **one real model call** (11,727 input + 902 output tokens). Host validation accepted coverage of all 98 selected artifacts and five orchestrations. The largest prompt was 49,123 bytes against a 55,296-byte ceiling, and source bytes were unchanged.

This was a live **recovery** qualification with two scripted setup calls, not a fresh end-to-end CLI migration. Earlier development runs also reproduced an alternating-tool loop, which the new guard stopped after five calls, and a three-minute local-response timeout. The final streaming harness allows ten minutes per response with a 12-call cap. No universal model-success guarantee is implied.

These transport and grouping checks do not establish full discovery, conversion, runtime behavioral equivalence, or deployment success. No customer generated output, global installation, license counters, or Azure resources are changed by these tests.