Expand LLMSwarmResult.Synthesis into a structured PSCustomObject

[ocalvo]

Problem

Invoke-LLMSwarm currently returns:

Synthesis = $synth.Content   # raw string

(see Public/Invoke-LLMSwarm.ps1:138)

Consumers get an opaque blob. There is no programmatic way to access the synthesizers reasoning, conclusions, per-task references, or token breakdown without string-parsing.

Proposed shape

Replace the single string with a [PSCustomObject] with sub-properties:

Synthesis = [PSCustomObject]@{
    PSTypeName   = "PwrCortex.SwarmSynthesis"
    Summary      = <short one-paragraph conclusion>
    Content      = <full synthesizer response text>
    KeyFindings  = <List[string]>       # extracted bullet points
    References   = <List[SwarmTaskRef]> # which Tasks[i] each finding came from
    Provider     = <string>
    Model        = <string>
    InputTokens  = <int>
    OutputTokens = <int>
    TotalTokens  = <int>
    ElapsedSec   = <double>
    StopReason   = <string>
}

Where SwarmTaskRef is @{ TaskIndex; TaskPrompt; Excerpt }.

Implementation

1. In Invoke-LLMSwarm, instead of storing $synth.Content, build the structured object from the [LLMResponse] returned by the synthesis call.
2. Ask the synthesizer to return its conclusions as a JSON block (or parse bullet lines) so KeyFindings/References can be populated reliably. Fallback: empty lists, Content always populated.
3. Update CLAUDE.md and Invoke-LLMSwarms .OUTPUTS section.
4. Update Enter-LLMChat:192 which passes $swResult.Synthesis into chat history — needs $swResult.Synthesis.Content.

Acceptance

• (Invoke-LLMSwarm ...).Synthesis.KeyFindings returns a list.
• (Invoke-LLMSwarm ...).Synthesis.References[0].TaskIndex resolves.
• Backwards compatibility is explicitly broken (Synthesis was a string, now an object). Release as feat!: / major bump per release-please rules.

Related

• Sibling issue: Replace Write-Host console UI with .format.ps1xml views