# CLI 0.0.4-microsoft-preview.22

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

**Known limitation: full FINACCONT-BT Discovery is not yet qualified.** The live
continuation advanced beyond repeated artifact reads, but component generation
still exceeded the isolated test's ten-minute response deadline. This release
contains verified host fixes, not a claim of successful end-to-end migration.

## Changes

- Honor the configured global context ceiling unless a phase explicitly overrides
  it. Use verified tokenizer units consistently for prompts and retained evidence.
- Factor identical repeated artifact objects into a lossless shared dictionary.
  Preserve all source fields, hashes and exact audit records. Complete paginated
  documents are assembled only after contiguous coverage and SHA-256 verification.
- Allocate Ollama runtime context for the complete measured prompt plus output and
  template reserves, growing up to the configured ceiling. No smaller model is
  selected and no evidence is discarded to fit a smaller runtime allocation.
- Keep changing archive bookkeeping after source evidence for prompt-prefix reuse.
- Accept a whole Mermaid code fence through envelope normalization, followed by
  the original pinned grammar. Invalid syntax returns feedback before storage;
  custom validation and storage failures still stop the stage.
- Strengthen bounded-write guidance. Allow one shared recovery for output-limit
  stops or timeouts with partial output; incomplete responses are audited and
  never executed. Silent timeouts and repeated failed recovery remain fatal.

Microsoft instructions, coverage/finalization checks, model selection, production
deadlines and tool-call limits remain unchanged. No customer YAML, activation,
entitlements, source files or Azure resources are modified by installation.

## Verification and remaining work

- 122 core, 243 CLI and 24 release/installer regressions pass (389 total);
  CLI lint and production build pass.
- Captured FINACCONT evidence round-trips exactly: five orchestrations and their
  binding. Shared representation reduces the document from 50701 to 20761 Qwen
  tokens without summarizing away source content.
- A live seeded-read continuation with the same local `qwen3-coder:30b` accepted
  model-generated metadata. Its complete architecture candidate passed the
  pinned grammar after the fence normalization fix.
- A later continuation replayed those real writes through normal validation,
  then timed out on two component responses, including bounded recovery. The
  model ignored chunk guidance. Neither partial response executed; source and
  grouping bytes stayed unchanged. This diagnostic used a 600-second limit,
  shorter than the application's default 3600-second absolute limit.
- Reliable bounded component output and a successful complete live Discovery
  remain outstanding. Planning, conversion and runtime business equivalence
  are not established by these continuation or fixture tests.

## PowerShell update

```powershell
Invoke-WebRequest 'https://tools.lambogenius.com/lambogenius-integration-migration/0.0.4-microsoft-preview.22/install-cli.ps1' -OutFile install-cli-preview22.ps1
# Review the downloaded script before running it.
& .\install-cli-preview22.ps1 -SkipDependencies
li.cmd --version
```

Expected version: `0.0.4-microsoft-preview.22`. Omit `-SkipDependencies` to install
missing prerequisites. Preserve previous reports/output and select a fresh
generated workspace: the engine fingerprint changed, so do not force old
Discovery or Conversion checkpoints to resume. Keep your chosen model and
supported context settings. Installation does not reset license counters.